Fog computing enabled air quality monitoring and prediction leveraging deep learning in IoT

Abstract

With the rapid industrialization and urbanization worldwide, air quality levels are deteriorating at an unprecedented rate and posing a substantial threat to humans and the environment. This brings the concern to effectively monitor and forecast air quality levels in real-time. Conventional air quality monitoring stations are built based on centralized architectures involving high latency, communication technologies demanding high power, sensors involving high costs and decision making with moderate accuracy. To address the limitations of the existing systems, we propose a smart and distinct Air Quality Monitoring and Forecasting system embracing Fog Computing with IoT and Deep Learning (DL). The system is a three-layered architecture with the Sensing layer first, Fog Computing layer in between, and Cloud Computing layer at the end. Fog Computing is a powerful new generation paradigm that brings storage, computation, and networking at the edge of the IoT network and reduce network latency. A DL based BiLSTM (Bidirectional Long Short-Term Memory) model is deployed in the Fog Computing layer. The proposed system aims at real-time monitoring and accurate air quality forecasting to support decision making and aid timely prevention and control of pollutant emissions by alerting the stakeholders when a dangerous Air Quality Index (AQI) is expected. Experimental results show that the BiLSTM model has a better predictive performance considering the meteorological parameters than the baseline models in terms of MAE and RMSE. A proof of concept realizing the proposed system is elaborated in the paper.

Keywords

Air quality monitoring air quality prediction fog computing deep learning bi-directional long short-term memory

1 Introduction

India currently has around 731 manual and 150 real-time air quality monitoring stations across 70 cities. With the number of real-time monitoring stations being relatively less in the current scenario as per the analysis of CPCB (Central Pollution Control Board), the Indian Government is planning to double the number of real-time monitoring stations by 2024. This also has motivated to develop a real-time system embracing state-of-the-art technologies.

Most of the existing works of air quality monitoring and prediction are based on centralized architectures involving high latency, communication technologies demanding high power, sensors involving high costs, and decision-making with moderate accuracy. To address these limitations, we propose a distinct Air Quality Monitoring and Forecasting system embracing Fog Computing with IoT, LPWAN, Deep Learning (DL), and Cloud Computing.

Owing to the proliferation of IoT applications, sending a large volume of data from end devices to the cloud layer for storage, analysis, and decision making would be infeasible due to the latencies from round trip propagation delay, bandwidth constraints, and computation overhead. To address these limitations, a new computing paradigm namely Fog Computing [4], introduced by Cisco aims to provide services at the edge of the IoT network. Instead of directly transferring the data to the cloud, Fog Computing enables storage, computation, decision making, and data management in the proximity of data sources to reduce the network traffic and the computational load of Cloud Computing. Fog computing supports low latency, distributed computing, location awareness, minimized bandwidth consumption and improves real-time decision-making. Thus, the proposed system utilizes the Fog Computing architecture with IoT, where the sensor node undertakes the data acquisition task, transmits it to the Fog Computing layer using LPWAN technology. Low-Power Wide-Area Network (LPWAN) technologies are a promising solution to the distributed Fog Computing nodes in IoT networks offering long-range, low-cost, and low-power consumption. Previous research show that utilizing deep learning technologies to forecast air quality is more accurate than using typical machine learning methods. To guarantee good prediction accuracy and low response latency, a DL based BiLSTM method is employed on the fog node to accurately forecast the pollutants PM_2.5, PM₁₀, CO, NO₂, SO₂, 0₃ responsible for determining the Air Quality Index (AQI). The results reveal that the BiLSTM method has good predictive performance compared with existing baseline Machine Learning (ML) and Deep Learning (DL) methods, namely Decision Tree Regression (DTR), Long Short-Term Memory (LSTM), Convolutional LSTM (Conv-LSTM), Multilayer Perceptron (MLP), Gated Recurrent Units (GRU) and Bidirectional GRU (BiGRU). Furthermore, the model considers the meteorological parameters as it plays a vital role influencing the pollutant values. Accurate forecasting is crucial to provide early warnings and enhance pollution management system. Cloud Computing offers long-term storage of data collected from the Fog layer and supports computation-intensive tasks and responds to client requests. The paper elaborates on the proposed fog-enabled architecture and present the proof of concept.

The main contributions of the paper are as follows:

Proposed a Fog Computing enabled IoT architecture for Air Quality monitoring and forecasting.

Designed a real-time Air Quality monitoring system to acquire the air pollutants and meteorological data.

Developed a smart fog gateway and identified a suitable model to accurately forecast air quality and deploy on the fog node.

Developed a user interface using cloud services to display the air quality trends to the users.

The rest of the paper is organized as follows. Section 2 discusses the related works of air quality monitoring and forecasting in the recent years. Section 3 presents in detail the functions and construction methods of the proposed Fog Computing enabled system. Section 4 details the system implementation and compares the results of the model’s prediction performance. Finally, Section 6 provides the conclusion of this work.

2 Related studies

The related works and technologies discussed in this section are carefully analyzed and considered in the work to solve the issues of the existing IoT-based air quality monitoring and prediction methods.

2.1. Air quality monitoring and related requirements

Air Quality monitoring system is classified as indoor and outdoor monitoring based on the location of the event’s occurrence. Outdoor refers to air pollution in open and industrial areas. On the other hand, indoor denote air pollution in a closed space like a home, office, shopping store, etc. With different scenarios and pollutant nature in outdoor and indoor monitoring, the requirements of the monitoring system vary as mentioned in Table 1. Our work focuses on outdoor monitoring.

Table 1
Requirements of Air pollution monitoring

Monitoring Development, installation, Energy consumption Accuracy Response time Cost

system and maintenance

Indoor Easy Low Medium Medium Low

Outdoor Medium Simple High Medium Medium

Industrial Medium Medium Very High Fast Medium

Monitoring	Development, installation,	Energy consumption	Accuracy	Response time	Cost
Indoor	Easy	Low	Medium	Medium	Low
Outdoor	Medium	Simple	High	Medium	Medium
Industrial	Medium	Medium	Very High	Fast	Medium

2.1.1 Air Quality Index (AQI)

As a standard for measuring and expressing air quality, government agencies have established a quantitative tool called Air Quality Index (AQI). AQI describes the degree of air pollution and provides easy-to-read and understandable information to the public [5]. AQI emphasizes the seriousness of air pollution, and the health impacts it poses, especially to vulnerable groups like children, the elderly, and people who are suffering from cardiovascular or respiratory disorders. AQI is determined by measuring six major pollutants, i.e., Ozone (O₃), Carbon Monoxide (CO), Sulphur Dioxide (SO₂), Nitrogen Dioxide (NO₂), Particulate Matter 2.5 (PM_2.5), Particulate Matter 10 (PM₁₀). AQI transforms the pollutants value into a single index value, categorizing the pollution situation into one among the six categories. Each country has its own method for assessing and defining the AQI levels. The description categories of air quality adopted for India along with the associated health impacts are depicted in Table 2. AQI values are represented in six quantiles ranging from good to severe. The six classifications are as Good, Satisfactory, Moderate, Poor, Very poor and Severe. Higher the AQI prompts an increase in the contamination range with higher concentrations of some pollutants and greater is the health concern.

Table 2
AQI levels and remarks

AQI Range AQI Category Indicative Color Remarks

0-50 Good Green Satisfactory - Minimal impact or no risk

51-100 Satisfactory Yellow Acceptable - moderate health concern to small population who are sensitive pollution

101-200 Moderately Polluted Orange Unhealthy for sensitive populations

201-300 Poor Red Everyone may start to get health effects. Serious effects to sensitive populations.

301-400 Very poor Purple Serious health impacts

401-500 Severe Maroon Entire crowd is likely to be affected

AQI Range	AQI Category	Indicative Color	Remarks
0-50	Good	Green	Satisfactory - Minimal impact or no risk
51-100	Satisfactory	Yellow	Acceptable - moderate health concern to small population who are sensitive pollution
101-200	Moderately Polluted	Orange	Unhealthy for sensitive populations
201-300	Poor	Red	Everyone may start to get health effects. Serious effects to sensitive populations.
301-400	Very poor	Purple	Serious health impacts
401-500	Severe	Maroon	Entire crowd is likely to be affected

Individual AQI for each pollutant’s concentration as I₁, I_2 …,I_n is calculated with the linear interpolation equation as follows. $I_{n} = \frac{(I_{high} - I_{low})}{(C_{high} - C_{low})} * (C - I_{low}) + I_{low}$ (1)

c _high - Breakpoint concentration greater or equal to given concentration.

c _low - Breakpoint concentration smaller or equal to given concentration.

I_high - AQI value corresponding to c _high

I_low - AQI value corresponding to c _low

C - Concentration of the pollutant.

The highest of all values is the overall AQI of the location at a given point in time i.e., Overall AQI (IAQI)=Max {I₁, I₂ . . . , I_n}

2.2 Fog Computing in IoT

With IoT increasingly becoming a part of our daily life, there is a discernible increase in the smart objects generating large volumes of data. However, sending all the data directly to the centralized cloud poses challenges, including latencies from round trip propagation delay, bandwidth constraints, and computation overhead. To address some of the short comings of cloud computing, the Fog Computing [6] concept is introduced to bring cloud services to the edge of the IoT network. Fog Computing is a distributed computing paradigm that enables local storage and computation, real-time decision making, and data management closer to the data sources. Fog Computing layer is an intermediary layer between IoT devices and the cloud layer. The devices in the Fog Computing layer are called fog nodes. The fog nodes have limited resources in terms of processing, computation, memory size, network, and storage like gateways, switches, routers, base stations, smartphones, etc., The capacity of fog nodes is smaller and less powerful compared to cloud servers and typically higher than that of end devices but still constrained. The fog nodes that are at one or two hops away from IoT devices filters out irrelevant data, partially or completely processes the sensor data on the fog node and minimizes the amount of data sent to cloud platforms for processing, analysis, and storage. Fog Computing significantly reduces the service latency, enables better utilization of resources, minimizes bandwidth consumption, and offers improved levels of services and outcomes for IoT deployments.

2.2.1 Characterization of fog computing

Security: A level of protection is provided to ensure safety and facilitate trusted transactions. FC nodes store sensitive data before transmission of data to the cloud layer.

Cognition: Fog layer nodes are aware of the end-users’ objectives. FC addresses the cognition by providing solutions based on contextual requirements.

Agility: FC delivers greater business agility and offers opportunities for developers to develop applications with the right tools and deploy them. It plays a vital role in the growth of mobile devices and other new services.

Latency: The fog layer performs filtering, processing, analytics, and other time-sensitive tasks on fog nodes like gateways, routers located in the proximity of the end-devices. This meets the need of IoT applications with stringent time requirements.

Efficiency: Fog because of the vicinity, is tightly integrated with the end-devices, supporting enhanced performance and efficiency. These advantages serve as enablers of new services and business models for networking and computing.

Few research that investigated Fog Computing are as follows: Santos et al. [7] investigated Fog Computing for anomaly detection in a smart city. The system achieves a faster response time on implementing the approaches on the fog layer compared with the traditional centralized cloud layer. Nicholas et al. [8] developed a smart fog-enabled gateway and enhanced the end-to-end interaction between the end device and the cloud. Tuli et al. [9] showed that the experimental results of Fog Bus settings enhanced the latency, energy, network, and computing infrastructures CPU usage. Dutta et al. [10] proposed a fog-based IoT setup, and the results proved to be advantageous in terms of delay over the cloud-only setting. Furthermore, embracing Fog Computing for aggregation and compression of data reduced the amount of data sent to the cloud.

With the growth of networked devices, Fog Computing has gained emphasis in recent times making IoT systems more efficient and scalable. The related works above reveal that introducing Fog Computing in IoT applications offers benefits in terms of bandwidth, latency, local storage, and processing. With this consideration, Fog Computing is embraced in the proposed system.

2.3 Low Power Wide Area Network (LPWAN) technologies

Widely installed short-range connectivity like Wi-Fi, Zigbee, and Bluetooth are not suitable for long-range performance scenarios. Although 5 G solutions based on cellular technology can provide extensive coverage, it consumes a lot of energy. Whereas, Low Power Wide Area Networks (LPWAN) have gained popularity with the characteristics of low power, long-range, and low-cost communication [11]. The LPWAN technologies did not even exist as early as the beginning of 2013. Nevertheless, with the rapid growth of the IoT industry and its requirements, LPWAN has gained emphasis in recent days and is one of the fast-growing fields of IoT networks. Table 3 summarizes the characteristics and differences of popular LPWAN technologies like Sigfox, LoRa, DASH7, and Narrow Band IoT (NB- IoT) that emerged in both licensed and unlicensed markets.

Table 3
Comparison of LPWAN technologies

Technology LoRa WAN SigFox NB-IoT DASH7 Wi-Fi LTE-m

Frequency Band 433/868/78 0/915 MHz (Unlicensed ISM) 868 MHz/902 MHz (Unlicensed ISM) Cellular (Licensed Band) 433.92 /868/915MHz (Unlicensed) 2.4-GHz / 900-Mhz (Unlicensed ISM) Cellular (Licensed Band)

Range 2- 5 km (urban) 15 km(suburban) 45 km (rural) 3 - 10 km(urban) 30 –50 km (rural) 2-5 km 1–2 km 1 km 2 –5 km

Data Rate 50 kbps 300 bps 250 kbps 167 kbps <347 Mbps 1Mbps

Bidirectional Yes Limited Yes Yes Yes Yes

Power Efficiency Very high Very high Medium High High Medium

Proprietary No Yes No No No No

Security AES 128 bit, Data Encryption at 3 Levels VPN+SSL encryption 2048-bit RSA encryption+128-bit SSL encryption AES 128-bit shared key encryption WPA/WPA2 128-bit SSL encryption

Mobility Yes No Yes Yes Yes Yes

Standardization LoRa Alliance Sigfox Company 3GPP DASH7 Alliance Wi-Fi Alliance 3GPP

Technology	LoRa WAN	SigFox	NB-IoT	DASH7	Wi-Fi	LTE-m
Frequency Band	433/868/78 0/915 MHz (Unlicensed ISM)	868 MHz/902 MHz (Unlicensed ISM)	Cellular (Licensed Band)	433.92 /868/915MHz (Unlicensed)	2.4-GHz / 900-Mhz (Unlicensed ISM)	Cellular (Licensed Band)
Range	2- 5 km (urban) 15 km(suburban) 45 km (rural)	3 - 10 km(urban) 30 –50 km (rural)	2-5 km	1–2 km	1 km	2 –5 km
Data Rate	50 kbps	300 bps	250 kbps	167 kbps	<347 Mbps	1Mbps
Bidirectional	Yes	Limited	Yes	Yes	Yes	Yes
Power Efficiency	Very high	Very high	Medium	High	High	Medium
Proprietary	No	Yes	No	No	No	No
Security	AES 128 bit, Data Encryption at 3 Levels	VPN+SSL encryption	2048-bit RSA encryption+128-bit SSL encryption	AES 128-bit shared key encryption	WPA/WPA2	128-bit SSL encryption
Mobility	Yes	No	Yes	Yes	Yes	Yes
Standardization	LoRa Alliance	Sigfox Company	3GPP	DASH7 Alliance	Wi-Fi Alliance	3GPP

Mekki et al. [12] and Sinha et al. [13] discussed and compared LoRa, Sigfox, and NB-IoT. The studies disclose that LoRa and Sigfox have very good characteristics in terms of capacity, battery, and cost. Meanwhile, Tan et al. [14] suggested NB-IoT has advantages in terms of QoS, latency, range, and performance. But the resources of the NB-IoT spectrum are very limited, costly, consume higher power, and lower battery life compared with LoRa. Osman et al. [15] mentioned that Sigfox has a long communication range of 50 km that reduced the number of transmissions to ensure energy-saving features. Sigfox aims for a more global solution, while LoRa enables developers to construct a secure and private network. LoRa and DASH7 use the 868 MHz band. Although both use the same frequency band, the maximum communication range of DASH7 is lower in comparison with LoRa. Furthermore, DASH7 uses Gaussian Frequency-Shift Keying (GFSK) modulation, where LoRa uses a proprietary spread spectrum modulation owned by Semtech. In the comparison of LPWAN solutions, LoRa is considered a suitable option for deploying city-scale applications.

LoRa is an emerging LPWAN solution to transfer and receive data over 2-20 kilometers with low costs. The name LoRa derives from its long-range functionality. Although it supports very low data rates, it consumes less power resulting in the increased battery life of the sensor nodes. It uses a spread spectrum approach that helps against interference or noise. Tests in urban areas have shown that a communication range of 3 km can be achieved in noisy environments. LoRa WAN protocol is designed with very good built-in security features based on AES cryptography in comparison with other IoT solutions.

Few works featured LoRa under real-world settings. Carlsson et al. [16] explored the performance of LoRa and further discussed its capabilities, performance, and limitations in urban and rural settings. El Chall et al. [17] presented an analysis on LoRa and determined its transmission range is 8 km in urban areas. A recent study by Basford et al. [18] on the long-term implementation of LPWAN technologies concluded that LoRa is the most viable choice in a smart city monitoring framework. A similar study by Sanchez-Gomez et al. [19] reached the same conclusion, suggesting LoRa for several smart city use-cases. This section provided an overview and a comparative analysis of the essential parameters of LPWAN technologies and identify that LoRa is an efficient communication technology to adapt and enhance the performance of the proposed Air Quality Monitoring system.

2.4 Air quality monitoring systems

The urge to monitor Air Quality levels effectively stems from the dangers that poor air quality brings to humans and the environment. Real-time Air Quality monitoring plays a significant role in laying a foundation for accurate air quality forecasting. We review the related works of air quality monitoring systems that considers primary air quality attributes.

Lai X et al. [20] implemented a low-cost IoT based air quality monitoring system. The system monitored the concentration of six air pollutants and achieved immediate predictions. The hardware part of the system included Raspberry Pi, sensor network, and Wi-Fi and the software part included the cloud system. They employed ML based models on the edge device to avoid transmission delays caused by bandwidth and network connection limitations. Furthermore, they mentioned including the external environmental factors in the future work to enhance the accuracy of the system. Duangsuwan, Sarun, et al. [21] focused on the development of smart sensors to track air pollution in Bangkok, Thailand. The parameters monitored include PM₁₀, CO, CO₂, noise level, and O₃. The sensor node uses the Linkit Smart 7688 microcontroller. The processed data on the sensor node is shared with the next layer using NB-IoT, and AQI is determined for the collected data. The main limitation is that the resources of the NB-IoT spectrum are costly. Kim SH, et al. [22] proposed an IoT-based atmospheric monitoring system to effectively observe air quality levels. The developed system was installed in Changwon National University and the Haman industrial complex area. The system uses the LTE network, and the scaling can cause the network cost to increase.

Santos et al. [23] presented an edge computing-based anomaly detection approach for Air Quality Monitoring in smart cities. The implementation was performed on Antwerp’s City of Things platform, where a set of sensors were integrated, and mounted on BPost’s delivery cars to gather real-time air quality information. They deployed the ML based unsupervised anomaly detection models on the fog node to send timely alerts to the citizens when unusual Air Quality levels are detected. Furthermore, the results verified that using Fog Computing over the centralized Cloud Computing architecture improved the response time of the application. Wang D et al. [24] designed an air quality monitoring system to achieve a home weather station allowing the user to access the data on his mobile. A tree-based network is created with the Zigbee, and the data forwarded from sub nodes collecting the sensor data is transmitted to the LM3S8962 gateway, and finally uploaded to the server. Toma C et al. [25] provided a solution for pollution monitoring in a smart city. The parameters collected by the system included CO₂, CO, NH₄, CH₄, toluene, temperature, humidity, pressure, altitude, and noise decibels. They discussed the technologies and platforms used within the solution and claimed their system offered benefits in terms of security, reliability, availability, and scalability. Senthilkumar et al. [26] proposed an embedded-based intelligent air pollution monitoring system. The embedded system collects the data and forwards it to the fog node. The functionality of the system was demonstrated on experimentation under various settings. However, they failed to use LPWAN technology to communicate with fog nodes. Rebeiro-Hargrave A et al. [27] designed and developed a Mega Sense Cyber Physical System to monitor urban air quality levels in real-time. The system aggregated the collected data and presented privacy-aware maps and graphs. The citizens of Helsinki received history profiles, personalized advice, and pollution hot spot maps through the mobile application. Their future work aims to integrate the system into mobile platforms such as drones. Baiocchi et al. [28] explored IoT concepts with smart systems to track real-time pollution around the City of Tacoma. It demonstrates how edge devices can ease cloud. They embraced IoT concepts, edge computing, and Microsoft Azure Framework, and demonstrated the effectiveness of the system. Furthermore, they discussed the pros and cons of integrating edge computing and cloud computing capabilities in the system.

Although the existing works as summarized in Table 4 propose mature and robust air quality monitoring systems, majority of the works investigate a very few pollutant parameters, fail to embrace gateways as the fog node, ignore the meteorological parameters that have a direct correlation with pollutants concentration, and do not effectively integrate LPWAN technologies and Fog Computing to meet the requirements of remote connections, low latency and real-time decision making. Hence, the proposed system considers these limitations and effectively integrate Fog Computing, LoRa, and Cloud Computing technologies to develop a smart fog enabled Air Quality monitoring system as in Table 5 assuring measurement accuracy with minimum cost.

Table 4
Summary of air quality monitoring systems

S. No Paper Edge / Fog Computing LPWAN Computing Cloud Parameters and Hardware Limitations

1 Lai X et al. 2019 [20] ✓ X X Parameters: SO₂, NO₂, CO, O₃, PM_2.5, and PM₁₀. Other Hardware: ESP8266 Edge node: Raspberry Pi 3 Model B Sensors: ZH03A Laser dust module, SGA-700 Intelligent Gas Sensor •Ignores environmental factors like wind speed that are likely to influence the concentration of the pollutants.

2 Duangsuwan S et al., 2018 [21] X ✓ X Parameters: PM₁₀, CO, CO₂, noise level, and O₃. Communication Technology: NB-IoT. Monitored duration: Oct. 7 to Oct. 13, 2017. •The resources of the NB-IoT spectrum are very limited and costly.

•Missed out on meteorological parameters.

•Monitoring duration is less

3 Kim SH et al., 2017 [22] X X ✓ Parameters: PM₁₀, PM_2.5, Ozone, SO₂, NO₂. Sensors: MQ7, MQ131, GSNT11, 2SH12, PM2007 Sampling duration: 30 minutes Location: Korea •The devices are connected via LTE network and consume more bandwidth.

•Failed to use the gateway as the fog node.

•On scaling the cost of the network tend to increase.

•Miss out on the valid data when the sampling rate is high

4 Santos et al., 2018 [23] ✓ ✓ X Parameters: PM₁, PM_2.5, PM₁₀. Communication Technologies: Lora WAN, Sigfox and DASH7 •Fail to mention the embedded board considered as the fog node and the sensors deployed.

•Only a limited number of pollutants are measured.

5 Wang D et al., 2016. [24] X X X Parameters: PM_2.5,Temperature, Humidity MCU: STC12C5608 Sensors: DS18B20, GP2Y1010AU0F Gateway: LM3S8962 Connectivity: Zigbee •The gateway is not used as the fog node.

6 Toma C et al., 2019 [25] ✓ X ✓ Parameters: Temperature, atmospheric pressure, altitude, humidity level, CO₂, CO, methane, and ammonium. Sensors: SNS-MQ13, MQ9, MQ135, BMP280 Sensor node: Raspberry Pi •Fail to monitor parameters responsible for determining AQI.

8 Senthilkumar R et al., 2020 [26] ✓ X ✓ Parameters: PM₁₀, PM_2.5, CO, NO₂, SO₂, and O₃. Sensors: GP2Y1014AU0F, DSM501, MQ-7, GSNT11, SO2-AF, MiCS2610-11, DHT11. Communication module: IEEE 802.15.4 K •Low-data rate wireless personal area network covering only 10 m.

•Not described about the end device and fog node used in the work.

9 Rebeiro-Hargrave A et al., 2020 [27] X X ✓ Parameters: Temperature, relative humidity, and wind speed, PM_2.5 and PM₁₀, NO₂, CO and O₃. •Failed to use Fog Computing layer.

•Fails to store data on edge devices

•The system does insert run time configuration updates.

10 Baiocchi O et al., 2019 [28] ✓ X ✓ Sensors: Grove Dust sensor and a Grove Air Quality Sensor. Fog node: Raspberry Pi Cloud: Microsoft Azure Framework •Data is sent only if it breaches a certain threshold.

•The sensors are more suitable only for indoor monitoring.

S. No	Paper	Edge / Fog Computing	LPWAN Computing	Cloud	Parameters and Hardware	Limitations
1	Lai X et al. 2019 [20]	✓	X	X	Parameters: SO₂, NO₂, CO, O₃, PM_2.5, and PM₁₀. Other Hardware: ESP8266 Edge node: Raspberry Pi 3 Model B Sensors: ZH03A Laser dust module, SGA-700 Intelligent Gas Sensor	•Ignores environmental factors like wind speed that are likely to influence the concentration of the pollutants.
2	Duangsuwan S et al., 2018 [21]	X	✓	X	Parameters: PM₁₀, CO, CO₂, noise level, and O₃. Communication Technology: NB-IoT. Monitored duration: Oct. 7 to Oct. 13, 2017.	•The resources of the NB-IoT spectrum are very limited and costly.
						•Missed out on meteorological parameters.
						•Monitoring duration is less
3	Kim SH et al., 2017 [22]	X	X	✓	Parameters: PM₁₀, PM_2.5, Ozone, SO₂, NO₂. Sensors: MQ7, MQ131, GSNT11, 2SH12, PM2007 Sampling duration: 30 minutes Location: Korea	•The devices are connected via LTE network and consume more bandwidth.
						•Failed to use the gateway as the fog node.
						•On scaling the cost of the network tend to increase.
						•Miss out on the valid data when the sampling rate is high
4	Santos et al., 2018 [23]	✓	✓	X	Parameters: PM₁, PM_2.5, PM₁₀. Communication Technologies: Lora WAN, Sigfox and DASH7	•Fail to mention the embedded board considered as the fog node and the sensors deployed.
						•Only a limited number of pollutants are measured.
5	Wang D et al., 2016. [24]	X	X	X	Parameters: PM_2.5,Temperature, Humidity MCU: STC12C5608 Sensors: DS18B20, GP2Y1010AU0F Gateway: LM3S8962 Connectivity: Zigbee	•The gateway is not used as the fog node.
6	Toma C et al., 2019 [25]	✓	X	✓	Parameters: Temperature, atmospheric pressure, altitude, humidity level, CO₂, CO, methane, and ammonium. Sensors: SNS-MQ13, MQ9, MQ135, BMP280 Sensor node: Raspberry Pi	•Fail to monitor parameters responsible for determining AQI.
8	Senthilkumar R et al., 2020 [26]	✓	X	✓	Parameters: PM₁₀, PM_2.5, CO, NO₂, SO₂, and O₃. Sensors: GP2Y1014AU0F, DSM501, MQ-7, GSNT11, SO2-AF, MiCS2610-11, DHT11. Communication module: IEEE 802.15.4 K	•Low-data rate wireless personal area network covering only 10 m.
						•Not described about the end device and fog node used in the work.
9	Rebeiro-Hargrave A et al., 2020 [27]	X	X	✓	Parameters: Temperature, relative humidity, and wind speed, PM_2.5 and PM₁₀, NO₂, CO and O₃.	•Failed to use Fog Computing layer.
						•Fails to store data on edge devices
						•The system does insert run time configuration updates.
10	Baiocchi O et al., 2019 [28]	✓	X	✓	Sensors: Grove Dust sensor and a Grove Air Quality Sensor. Fog node: Raspberry Pi Cloud: Microsoft Azure Framework	•Data is sent only if it breaches a certain threshold.
						•The sensors are more suitable only for indoor monitoring.

Table 5

Proposed system

S. No	Paper Title	Fog Computing	LPWAN	Cloud Computing	Parameters and Hardware	Advantages
1	Proposed System	✓	✓	✓	Parameters: PM_2.5, PM₁₀, CO, NO₂, SO₂,0₃, Temperature, Relative Humidity, Pressure, Solar Radiation, Wind speed and wind direction.	•Aims to address the limitations of the previous works by leveraging the benefits of the Fog Computing, LPWAN, and Cloud Computing technologies.
					Sensors: MICS-4514, MQ7, SDS011, MQ136, BME280, MQ131, NEO-6M, PXV7002DP, DRF1276DM. Fog Computing Layer: Raspberry Pi Model 3B+	•Uses gateway as the fog node.
						•Monitors the meteorological parameters along with the pollutants.
					Cloud: AWS	•Real-time low cost and decentralized air pollution monitoring solution.

2.5 Air quality prediction model

Air quality forecasting or prediction refers to estimating the concentration of pollutants for time ahead based on real-time data and trends in historical data. Predicting air pollutants concentration has gained attention in recent research. It provides important information to the public, helps reduce emissions and enhance pollution management. To make a good prediction, fitting an appropriate model is important. However, accurate prediction is still a challenge because of the complex trends in the data. The common methods for air quality prediction encompass traditional statistical models and data-driven methods such as Machine Learning and Deep Learning methods.

Yu R et al. [29] proposed a Random Forest approach to predict the AQI for urban sensing systems. The model forecasts the AQI of Shenyang based on the data from air quality monitoring stations. The results of the model outperformed single Decision Tree (DT), Naïve Bayes, Logistic Regression, and ANN. However, the amount of data used for modelling is limited. Taneja S et al. [30] discussed the prediction of pollutants trends for future years using Linear Regression and Multilayer Perceptron (MLP) methods. They predict that pollutants NO2 and O3 are likely to increase, while SO2 and CO levels would follow past trends and PM₁₀ is likely to increase drastically. The main limitation is that the models look at only the linear relationships between the data, easily affected by outliers and fails to model non-linear patterns in the data. The meteorological parameters are ignored while modelling the data. Ameer S et al. [31] performed air quality prediction using four regression techniques and presented a comparison analysis to find the best model to accurately forecast air quality in terms of data size and processing time. The techniques included DTR, Random Forest Regression (RFR), MLP regression and Gradient Boosting Regression GBR). The results indicated that RFR is the most effective methodology for predicting pollution for data sets of various sizes, locations, and features. Furthermore, RFR had the lowest error rate and performed well in detecting the peak values among the four approaches. Bisht M et al. [32] proposed an intelligent air pollution prediction system using Extreme Learning Machine (ELM) to predict the AQI for five pollutants. It is identified that the AQI values predicted by ELM is close to the actual values for five out of seven days in the test set. ELM resulted in much faster learning compared to backpropagation based FFNN. ELM-based prediction was found to have greater accuracy than the existing SAFAR system. Rekhi JK et al. [33] analyzed the time series air pollution dataset of New Delhi and forecasted monthly future values of two major pollutants SO2 and NO2 using ARIMA. However, ARIMA fails to effectively capture the sudden changes in exhaust emissions. Zheng H et al., 2018 [34] developed a novel multiple kernel learning-based approach with support vector classifier as the base learner for air quality forecasting. The proposed method forecasted severe air pollution and outperformed other models including Random Forest, LSTM, Multilayer Perceptron and ARIMA. Leong WC et al. 2020 [35] proposed a support vector machine to model the air pollution index. The study investigated the kernel functions model parameters. SVM using radial basis function (RBF) kernel function effectively solved the problem of complex air pollution index modeling. However, SVM algorithm is not suitable for large datasets and does not perform well when the data set has more noise. Bai Y, et al., 2016 [36] constructed a W-BPNN model using wavelet technique and back propagation neural network (BPNN) to forecast daily air pollutants (NO₂, PM₁₀, and SO₂) concentrations. The results revealed that the W-BPNN model had better prediction performance for the three pollutants than mono-BPNN model. However, the work neglected data pre-processing and optimization of parameters, leading to some bias in modelling. Yan R et al. [37] developed a multiple-hour and multiple-site forecasting model using CNN-LSTM. The CNN-LSTM exhibited better performance compared with the BPNN and the CNN. Li X et al. [38] presented a novel long short-term memory neural network extended (LSTME) model to predict air pollutant concentrations. The model captured long time dependencies and automatically determined the optimum time lags. LSTME outperformed the spatiotemporal deep learning (STDL) model, the time delay neural network (TDNN) model, ARMA, SVR, and the traditional LSTM. Compared with the RNN model, the LSTME exhibited better prediction performance. The experiments revealed that DL-based models exhibited better prediction performance compared to traditional shallow models such as SVR, ARMA and TDNN. Based on the related works summarized as in Table 6, most studies consider only one or a small number of pollutants while modelling the data and few works ignore the meteorological variables that have a direct correlation with the pollutants. Furthermore, the works that use statistical models like SMA, WMA, and ARIMA for short term forecasting exhibit simplicity, good performance, interpretability and are computationally less expensive but they fail to capture the non-linear patterns. Furthermore, researchers investigate ML-based approaches that overcome the non-linear limitations and uncertainties to achieve better accuracy. For Indian context, many air quality forecasting works have employed various traditional ML and DL methods. Besides the ML algorithms, DL methods shows higher prediction accuracy. Although the traditional ML models are widely employed, they are not suitable to capture long-term dependencies in time series data. Whereas DL uses the structure of ANN with capabilities of non-linear mapping, robustness, self-adaption and modeling complex relationships between inputs and output variables to effectively capture the trends of air quality. The main research gap is that the works fail to employ forecasting models with good predictive performance on the Fog nodes to support real-time decision-making and low latency. Hence, the proposed system employs a DL based model with good predictive performance on the fog node considering all the six primary pollutants and meteorological parameters to predict air quality accurately.

Table 6
Summary of Air Quality Prediction models

S.No Author (s) Method Input variables Predicted variables Dataset Comparison models Evaluation metrics Limitations

1 Yu R et al. 2016 [29] Random Forest PM 10, PM 2.5, CO, NO2, SO2, O3, timestamp, temperature, humidity, barometric pressure, wind speed and visibility AQI Shenyang, May 2015 - June 2015 Naïve Bayes, Logistic Regression, DT and ANN Relative Absolute Error (RAE) Limited amounts of data.

2 Taneja S et al. 2016 [30] Linear Regression and MLP PM_2.5, SO₂, CO, PM₁₀, NO₂, O₃ PM2.5, SO₂, CO, PM₁₀, NO₂, O₃ New Delhi, 2011–2015 Fitting curve Looks at only the linear relationships between dependent and independent variables. Easily affected by outliers Ignores the influence of meteorological parameters. Unable to model non-linear patterns.

3 Ameer S et al. 2019 [31] Random Forest regression PM_2.5, Dew, pressure, temperature, and wind speed PM_2.5 Guangzhou, Chengdu, Beijing, Shanghai and Shenyang, Jan 2010 to Dec 2015 DTR, RFR, MLP and GBR MAE, RMSE Unable to predict trends, i.e., they do not extrapolate

4 Bisht M et al. 2018. [32] Multi-variable ELM regression PM₁₀, PM_2.5, NO₂, CO, O₃, relative humidity, wind speed, temperature, rainfall and wind direction PM₁₀, PM_2.5, NO₂ and O₃. Delhi University monitoring station, February, and March 2016 SAFAR RMSE ELM is fast to train, but it does not encode more than 1 layer of abstraction, so it is not “deep”.

5 Rekhi JK et al., 2020 [33] ARIMA SO₂ and NO₂ SO₂ and NO₂ Delhi, January 2009 to December 2015. - MSE, RMSE, MAE There will be sudden increase in vehicle exhaust emissions in the morning and evening. ARIMA cannot effectively predict this change in points.

6 Zheng H et al., 2018 [34] Multiple kernel learning (MKL) model O₃, NO₂ SO₂, PM_2.5, PM₁₀, temperature, dew, pressure, humidity, wind speed and wind direction PM 2.5 and air quality health index (AQHI) Hong Kong February 2013 -January 2015 and Beijing, January 2010 - December 2014 ARIMA, RF, MLP, LSTM and SVM MSE, Accuracy Good for Short-term and extreme pollution forecasting.

7 Leong WC et al., [35] SVM SO₂, PM₁₀, O₃, CO, NO₂ Air Pollution Index (API) Perak and Penang State of Malaysia, 2009 - 2014. - Sum of square of error (SSE), mean of sum of square of error (MSSE) and coefficient of determination (R2) SVM is not suitable for big datasets and does not perform well if the data set has more noise.

8 Bai Y et al., 2016 [36] Wavelet -BPNN PM₁₀, SO₂, NO₂, wind speed, temperature, pressure, mean relative humidity, visibility, precipitation, and illumination PM₁₀, SO₂, and NO₂ Chongqing, China. January 2011 to December 2011. BPNN MAPE, RMSE and correlation coefficient (CC) Neglects data pre-processing and optimization of parameters, leading to some bias.

9 Yan R et al., 2021 [37] CNN-LSTM PM_2.5, PM₁₀, SO₂, NO₂, O₃, and CO, precipitation, temperature, dew point, pressure, wind speeds, and wind directions AQI Beijing 2015 - 2016 CNN, BPNN, LSTM RMSE, IA Fails to compare with other DL based models.

10 Li X et al., 2017 [38] LSTM Extended PM_2.5, temperature, humidity, wind speed and visibility PM_2.5 Beijing City from Jan 2014 –May 2016 SVR, ARMA, LSTM, TDNN RMSE, MAE, MAPE Only a few parameters are considered in modelling. Fails to compare with ML models.

11 Proposed System Bidirectional LSTM PM_2.5, PM₁₀, CO, NO₂, SO₂, 0₃, Temperature, Relative Humidity, Pressure, Solar Radiation, Wind speed and wind direction. PM_2.5, PM₁₀, CO, NO₂, SO₂, 0₃ CPCB dataset, Coimbatore, India, 2019- 2020 DTR, MLP, LSTM, Conv-LSTM, BiGRU RMSE, MAE –

S.No	Author (s)	Method	Input variables	Predicted variables	Dataset	Comparison models	Evaluation metrics	Limitations
1	Yu R et al. 2016 [29]	Random Forest	PM 10, PM 2.5, CO, NO2, SO2, O3, timestamp, temperature, humidity, barometric pressure, wind speed and visibility	AQI	Shenyang, May 2015 - June 2015	Naïve Bayes, Logistic Regression, DT and ANN	Relative Absolute Error (RAE)	Limited amounts of data.
2	Taneja S et al. 2016 [30]	Linear Regression and MLP	PM_2.5, SO₂, CO, PM₁₀, NO₂, O₃	PM2.5, SO₂, CO, PM₁₀, NO₂, O₃	New Delhi, 2011–2015		Fitting curve	Looks at only the linear relationships between dependent and independent variables. Easily affected by outliers Ignores the influence of meteorological parameters. Unable to model non-linear patterns.
3	Ameer S et al. 2019 [31]	Random Forest regression	PM_2.5, Dew, pressure, temperature, and wind speed	PM_2.5	Guangzhou, Chengdu, Beijing, Shanghai and Shenyang, Jan 2010 to Dec 2015	DTR, RFR, MLP and GBR	MAE, RMSE	Unable to predict trends, i.e., they do not extrapolate
4	Bisht M et al. 2018. [32]	Multi-variable ELM regression	PM₁₀, PM_2.5, NO₂, CO, O₃, relative humidity, wind speed, temperature, rainfall and wind direction	PM₁₀, PM_2.5, NO₂ and O₃.	Delhi University monitoring station, February, and March 2016	SAFAR	RMSE	ELM is fast to train, but it does not encode more than 1 layer of abstraction, so it is not “deep”.
5	Rekhi JK et al., 2020 [33]	ARIMA	SO₂ and NO₂	SO₂ and NO₂	Delhi, January 2009 to December 2015.	-	MSE, RMSE, MAE	There will be sudden increase in vehicle exhaust emissions in the morning and evening. ARIMA cannot effectively predict this change in points.
6	Zheng H et al., 2018 [34]	Multiple kernel learning (MKL) model	O₃, NO₂ SO₂, PM_2.5, PM₁₀, temperature, dew, pressure, humidity, wind speed and wind direction	PM 2.5 and air quality health index (AQHI)	Hong Kong February 2013 -January 2015 and Beijing, January 2010 - December 2014	ARIMA, RF, MLP, LSTM and SVM	MSE, Accuracy	Good for Short-term and extreme pollution forecasting.
7	Leong WC et al., [35]	SVM	SO₂, PM₁₀, O₃, CO, NO₂	Air Pollution Index (API)	Perak and Penang State of Malaysia, 2009 - 2014.	-	Sum of square of error (SSE), mean of sum of square of error (MSSE) and coefficient of determination (R2)	SVM is not suitable for big datasets and does not perform well if the data set has more noise.
8	Bai Y et al., 2016 [36]	Wavelet -BPNN	PM₁₀, SO₂, NO₂, wind speed, temperature, pressure, mean relative humidity, visibility, precipitation, and illumination	PM₁₀, SO₂, and NO₂	Chongqing, China. January 2011 to December 2011.	BPNN	MAPE, RMSE and correlation coefficient (CC)	Neglects data pre-processing and optimization of parameters, leading to some bias.
9	Yan R et al., 2021 [37]	CNN-LSTM	PM_2.5, PM₁₀, SO₂, NO₂, O₃, and CO, precipitation, temperature, dew point, pressure, wind speeds, and wind directions	AQI	Beijing 2015 - 2016	CNN, BPNN, LSTM	RMSE, IA	Fails to compare with other DL based models.
10	Li X et al., 2017 [38]	LSTM Extended	PM_2.5, temperature, humidity, wind speed and visibility	PM_2.5	Beijing City from Jan 2014 –May 2016	SVR, ARMA, LSTM, TDNN	RMSE, MAE, MAPE	Only a few parameters are considered in modelling. Fails to compare with ML models.
11	Proposed System	Bidirectional LSTM	PM_2.5, PM₁₀, CO, NO₂, SO₂, 0₃, Temperature, Relative Humidity, Pressure, Solar Radiation, Wind speed and wind direction.	PM_2.5, PM₁₀, CO, NO₂, SO₂, 0₃	CPCB dataset, Coimbatore, India, 2019- 2020	DTR, MLP, LSTM, Conv-LSTM, BiGRU	RMSE, MAE	–

3 Proposed fog based air pollution monitoring and prediction system

The proposed Fog Computing based Air Quality Monitoring and Prediction (AQMP) system is a three-layered architecture with the Sensing layer first, Fog Computing layer in between, and Cloud Computing layer at the end. It integrates multiple modules to effectively monitor, store and predict air quality in real-time. All the layers communicate via LoRa and Wi-Fi. An overall view of the system architecture is depicted in Fig. 1, and the description of the proposed system is explained below.

Fig. 1

Architecture of the proposed air quality monitoring prediction system.

3.1 Sensing layer

The first layer is the sensing layer, comprising the resource-constrained sensor nodes to acquire the data from the environment and transmit it to the Fog Computing layer. The components of this layer are elucidated below.

Controller unit: Arduino Mega 2560, a cost-effective, resource-constrained, low-power microcontroller is the controlling node. It interfaces the sensors and performs necessary processing on the raw data.

Height of senor node: Referring to the guidelines of CPCB, the system will be deployed at 3-10 m from the ground. This is to avoid bias from a specific source that could influence the results.

Nature of monitoring: The proposed system is static i.e.) in fixed infrastructure. The main limitation of the dynamic system is that the temporal coverage is incomplete. Thus, a static system is chosen to provide a holistic view by continuously capturing the data.

Sensor array: An array of sensors interfaced to the controller unit will measure the real-time prominent air pollutants including CO, NO₂, SO₂, PM_2.5, PM₁₀, and O₃ and meteorological parameters like temperature, pressure, humidity, wind direction, and wind speed. Technical advancements have enabled low-cost sensors, less in weight and size, consuming low power. Their availability in the market has facilitated the deployment of several monitoring systems with acceptable data accuracy sufficient to capture the air quality trends. The sensors used to measure the parameters in the work as in Table 7, are selected based on the range, lifetime, and compatibility with the microcontroller. The unified unit μg/m³ is chosen for converting the measured pollutant values, except for CO measured in mg/m³.

Table 7
Sensors to monitor air pollutants and meteorological parameters

S. No Parameters Sensor model

1 PM 2.5 (Particulate matter 2.5) SDS011

2 NO2 (Nitrogen dioxide) MICS-4514

3 SO2 (Sulphur dioxide) MQ-136

4 CO (Carbon monoxide) MQ-7

5 PM 10 SDS011

6 Ozone MQ-131

7 Ambient Temperature BME280

8 Relative Humidity BME280

9 Pressure BME280

10 Wind speed MPXV7002DP

11 Wind Direction MPXV7002DP

S. No	Parameters	Sensor model
1	PM 2.5 (Particulate matter 2.5)	SDS011
2	NO2 (Nitrogen dioxide)	MICS-4514
3	SO2 (Sulphur dioxide)	MQ-136
4	CO (Carbon monoxide)	MQ-7
5	PM 10	SDS011
6	Ozone	MQ-131
7	Ambient Temperature	BME280
8	Relative Humidity	BME280
9	Pressure	BME280
10	Wind speed	MPXV7002DP
11	Wind Direction	MPXV7002DP

Pre-calibration: The characteristics of low-cost sensors vary between sensors in the manufacturing lot. Hence, every sensor is individually pre-calibrated to obtain acceptable accuracy while estimating the gas capacity. Many of the existing works overlooks the calibration which affects the accuracy of the captured values. Algorithm 1 presents the steps for pre-calibration of sensors.

Algorithm 1: Pre-Calibration of Gas Sensors
1: R_g estimation (R_g - Resistance of the sensor in the pure air)
2: R_p estimation (R_p - Resistance of the sensor in presence of
a specific gas)
3: Analog interpret sensor pin
4: Get various samples and calculate the mean (M)
5: R_p = M/pure air factor
6: Extrapolate coefficients x and y
7: Estimate ppm, ppm=x $\times {(\frac{R_{g}}{R_{p}})}^{y}$

Deployment: The deployment consists of two phases including testing and actual deployment.

Sampling Rate: The sensor node periodically read the parameter data of sensors at an interval of 1 minute.

Power module: The module is powered by a lithium-powered battery to provide safe and stable power.

Communication module: The raw data acquired by sensor nodes placed is pre-calibrated, then along with the sensor ID, timestamp, it is transmitted to the Fog Computing layer for processing at a single hop using the LoRa module fitted to the Arduino. The system uses Dorji DRF1276DM, a special LoRa board is embedded with the Semtech SX1276 LoRa chip for communication. This module has the following features: ultra-long range spread spectrum, high interference immunity, operates at 868/915 MHz ISM frequency band, urban range of 2-5 km urban, low data rate transfer, low power consumption,+20 dBm power amplifier, and star networking ability.

3.2 Fog computing layer

The Fog Computing layer uses Raspberry Pi 3 Model B + running Raspbian OS, called the fog node [39, 40]. This development board has the features of low cost, low power consumption, adequate storage, good processing, networking, and computing capabilities to serve as the fog node. There are many expensive industrial standard boards with greater computation power that can serve as the edge node. But a cost-effective Raspberry Pi is preferred as the fog node in many IoT applications.

Raspberry Pi 3 Model B has BCM2837 Quad-Core (4x ARM Cortex-A53, 1.2 GHz), 4 USB ports, 40 GPIO pins. Raspberry Pi does not have internal flash memory, so it uses an SD card for storage. The GPIO pins on the board support only digital inputs, so an ADC is required to connect analog sensors. R Pi with built-in quad-core1.2 GHz 64-bit CPU that offers good processing and computing capabilities and thus preferred as the fog node in many IoT applications. The nodes of the previous layer share the data with the nearest fog nodes without directly uploading it to the cloud. Data is received on R Pi (fog node) by the LoRa receiver.

3.2.1 Fog computing services

The services to manage the resources and the ingested data on the fog node in the proposed system are explained below.

Authentication server – This module authenticates the nodes joining and leaving the fog node.

Daemons : The processes run continuously to listen to (LoRa or other) messages to collect the incoming data and handle the requests.

Configuration management : It tunes the associated nodes on encountering certain events of atmospheric risks.

Interoperability : This module maps the data collected from multiple nodes to a common data standard. Due to its simplicity and compatibility with various programming languages, the JSON standard is selected. The common format before pre-processing and sharing the data reduces overhead, consumption of resources, and bandwidth.

Publisher : Publisher sends the locally processed and filtered data to the Cloud Computing platform.

Log : The logging module keeps track of the events of fog nodes events to enable history tracking and to identify bugs and faults.

Event responses : Intelligent reactions and warnings are triggered, when the AQI breaches the threshold.

Data summarization : It converts the voluminous sensor data into meaningful summaries. The results are concise, interpretable, capturing the normal and abnormal trends in data. The system becomes computationally less expensive, highly condensed to meet bandwidth requirements. The summaries are sent to the cloud while retaining sufficient data to represent the trends in the concentration.

Processing module : It filters, pre-processes, forecasts pollutants for future time steps and provide early warning in the occurrence of dangerous pollution levels.

Fog enables the processing and operation of data closer to the source. It interacts with the opposite two layers. The services rendered by the fog layer above makes the fog node a smart fog gateway. Finally, after processing and prediction the fog node, the results and monitoring data are sent to the cloud using MQTT protocol.

3.3 Cloud computing layer

The Cloud Computing layer is the end layer, supporting long-term storage and complex processing of air pollution data collected over sustained periods.

This layer acquires the data from the fog computing system, stores it in the database, and provides feedback results and data visualization to the clients. The work chooses the AWS platform, as it provides secure IoT services. The APIs are made available to share data with third parties.

4 System implementation

A proof-of-concept of the proposed system entailing the hardware and software implementation is detailed in this section. The hardware module mainly includes the Sensing Node and the Fog Computing device. The software component includes the prediction model and the cloud. The summary of tools and technologies to develop the proposed system is described in Table 8.

Table 8
Summary of the tools and technologies used to develop Fog Computing based Air Quality Monitoring and Prediction system

Layer Tools and Technologies

Terminal device layer MICS-4514, MQ7, SDS011, MQ136, BME280, MQ131, NEO-6M, PXV7002DP, Arduino Mega 2560

Communication Layer LoRa- DRF1276DM

Fog Computing Layer Raspberry Pi Model 3B+

Air Quality Forecasting BiLSTM model

Application Layer MQTT

Cloud Computing Layer AWS

Layer	Tools and Technologies
Terminal device layer	MICS-4514, MQ7, SDS011, MQ136, BME280, MQ131, NEO-6M, PXV7002DP, Arduino Mega 2560
Communication Layer	LoRa- DRF1276DM
Fog Computing Layer	Raspberry Pi Model 3B+
Air Quality Forecasting	BiLSTM model
Application Layer	MQTT
Cloud Computing Layer	AWS

4.1 Hardware implementation

A customized PCB is developed to monitor and predict air quality in real-time. The components of the hardware and other configurations are explained in Section 3.1. The hardware of the system necessarily includes the Sensor node that acquires the data and the Fog Computing node that processes the data. The system is tested successfully in the lab environment as in Fig. 2.

Fig. 2

Air quality monitoring and prediction system.

4.2 Software implementation

The data acquired from the monitoring system is transmitted to the Fog Computing layer using LoRa. Further, a suitable air quality prediction model is identified and deployed on the fog node. The air quality data is further transmitted asnd stored in the AWS cloud. A detailed description of this is presented below:

4.2.1 Air quality prediction model

The intelligence is introduced on the Fog node by deploying a Deep Learning model to forecast pollutants concentration for future time steps. Initially, the data gathered from the Central Pollution Control Board (CPCB) monitoring station [41], situated at SIDCO Kurichi in Coimbatore city is used train the data with the model and deploy it on the Fog node. The CPCB dataset is presented in 15-minute intervals from 00 : 00 on June 1, 2019, to 23 : 59 on April 30, 2020. The investigative variables acquired from the location include PM_2.5, PM₁₀, CO, NO₂, SO₂, O₃, Temperature, Pressure, Relative Humidity, Solar Radiation, Wind speed, and Wind direction. The work considers the meteorological parameters as they are responsible for the dispersion and transformation of pollutants in the atmosphere and have a direct correlation with the pollutant’s values.

To analyze the basic and statistical characteristics of six pollutants such as minimum, maximum, extremum, mean, median, standard deviation, skewness, and kurtosis, a descriptive analysis of the data is presented in Table 9. PM_2.5, PM₁₀, SO₂, NO₂, and O₃ have a a dimension of μg/m³, and CO has a dimension of mg/m³.

Table 9
Descriptive Statistics of the exploratory variables in the dataset

Pollutant Min Max Mean Median Standard Deviation Skewness Kurtosis

PM_2.5 0.01 103.09 23.62 24.77 12.82 1.15 2.10

PM₁₀ 0.31 355.12 32.25 32.25 20.055 3.58 0.67

NO₂ 0.01 158.25 8.67 20.77 22.77 1.35 1.37

CO 0.01 4.43 0.92 0.92 0.53 0.30 –0.31

SO2 0.01 46.21 8.67 9.03 3.10 1.07 6.68

Ozone 0.02 147.2 30.50 21.78 1.31 1.31 1.65

SR 22.32 1095.87 199.32 39.01 256.78 1.21 0.169

RH 5.33 99.04 65.25 72.21 17.16 –0.87 0.001

WS 0.83 32.08 9.19 9.11 4.73 0.81 0.49

WD 2.12 353.6 169.120 198.85 87.01 –0.159 –1.6577

Pollutant	Min	Max	Mean	Median	Standard Deviation	Skewness	Kurtosis
PM_2.5	0.01	103.09	23.62	24.77	12.82	1.15	2.10
PM₁₀	0.31	355.12	32.25	32.25	20.055	3.58	0.67
NO₂	0.01	158.25	8.67	20.77	22.77	1.35	1.37
CO	0.01	4.43	0.92	0.92	0.53	0.30	–0.31
SO2	0.01	46.21	8.67	9.03	3.10	1.07	6.68
Ozone	0.02	147.2	30.50	21.78	1.31	1.31	1.65
SR	22.32	1095.87	199.32	39.01	256.78	1.21	0.169
RH	5.33	99.04	65.25	72.21	17.16	–0.87	0.001
WS	0.83	32.08	9.19	9.11	4.73	0.81	0.49
WD	2.12	353.6	169.120	198.85	87.01	–0.159	–1.6577

4.2.1.1. Experimental parameters

Appropriate features that show a strong relationship should be chosen as they are important in forecasting the dependent variables. The lack of relevant variables could prevent the model from approximating the dynamics between the predictors and the prediction variables. If there are numerous input parameters, there would be a need for a dimensionality reduction method. Based on the findings from the literature review, we choose an optimal number of influential parameters for air quality modeling. The parameters include PM_2.5, PM₁₀, CO, NO₂, SO₂, 03, Temperature, Relative Humidity, Pressure, Solar Radiation, Wind speed, and Wind direction.

4.2.1.2. Data pre-processing

Due to the time-dependency, time-series data are subjected to a few missing points while being recorded because of sensor failure or network problems in the monitoring station. Data pre-processing is a key step in developing ANN models as it improves the representation of the collected data. The incorrect or incomplete data may cause discontinuity and introduce significant bias while modelling the data. In our work, outliers are eliminated, and the data is cleaned by filling up the missing values with the average value of the data in the same period. Min-max scaling is performed to linearly transform each of the attributes of data in the range 0-1. It helps to avoid model sensitivity to extreme values of the data. The scaled data is then fed into the network for prediction. The results are re-scaled to get the actual value. A minimum data pre-processing is sufficient for Deep Learning.

4.2.1.3 Model deployment

Multivariate Bidirectional LSTM model is deployed on the fog node to predict future trends of the primary air pollutants concentration. Long Short-Term Memory (LSTM) is a form of artificial Recurrent Neural Network (RNN) used to model the sequential relationship between past, present, and future data [42]. The LSTM models are effective, particularly when it comes to the preservation of long short-term memory. Furthermore, LSTM effectively overcomes the gradient explosion and disappearance problems seen in RNN with concept of memory cells and controlling gates. LSTM is a special kind of RNN, capable of learning long-term temporal dependencies across time steps and predict the sequences. The generalization ability of LSTM allows it to exhibit superior performances over the traditional NN. Each LSTM unit has an input gate, a forget gate, an output gate, and a memory cell. The memory cells in LSTMs allow them to accumulate steps over prediction sequences, allowing them to do well in long-term tasks [43]. The model includes processing input, recurrent and output layers at each time step. In bidirectional LSTM, the hidden layer is added in the reverse direction of the data flow. Bidirectional LSTM model runs the inputs in forward and backward directions, one from past to future and the other from future to past. Therefore, BiLSTM is more effective than unidirectional LSTM.

The architecture of a typical LSTM layer is briefly described as follows.

–Input Layer: By concatenating pollutant concentrations and meteorological factors, this layer creates dense embedding. $X_{t}^{e} = (x_{t}, m)$ (2) where the function f produces a vector $X_{t}^{e}$ for the given input vectors x_t pollution concentrations, and m (meteorological data). The vector of the input layer is passed to the recurrent layer.

–Recurrent layer: This layer generates hidden representations of vector $X_{t}^{e}$ . The hidden states h_t are computed using input from previous time step h_t-1 and input $X_{t}^{e}$ . The hidden layers are internally connected. The hidden state is updated using $h_{t} = φ (X_{t}^{e}, h_{t - 1})$ (3) where Φ is a memory cell of LSTM.

–Output Layer: This is the final layer that outputs pollutant concentration values of future time step considered with y_t.

LSTM comprises multiple LSTM units. Each LSTM unit has an input gate, a forget gate, an output gate, and a memory cell. Inputs to the LSTM cell at any time step are the current input x_t, previous hidden state h_t1, and previous memory state c_t-1. The LSTM architecture used in this paper is defined by the equations below.

–Memory cell (m_t): $m_{t} = f_{t} * m_{t - 1} + i_{t} * c_{t}$ (4) $c_{t} = \tanh (W_{m} x_{t} - W_{m} h_{t - 1} + b_{m})$ (5) where t denotes the current time step, t - 1 denotes the previous time step, σ is the logistic sigmoid function, W represents the weight matrices, h_t–1 represents the hidden state at step t-1, x_t represents input at step t, and b represents bias vector.

–Input gate (i_t): a gate that determines which values to update the memory state from the input. $i_{t} = σ (W_{i} x_{t} - W_{i} h_{t - 1} + b_{i})$ (6) –Forget gate (ft): a gate that decides to reset the old memory. $f_{t} = σ (W_{f} x_{t} - W_{f} h_{t - 1} + b_{f})$ (7) –Output gate (o_t): a gate that decides what to output based on the memory of the input and block. $o_{t} = σ (W_{o} x_{t} - W_{o} h_{t - 1} + b_{o})$ (8) Hidden cell state (h_t): h_t is given by the following. $h_{t} = o_{t} * tanh (m_{t})$ (9)

The overview of the LSTM cell is shown in Fig. 3 where x_t, h_t, and m_t represents the input vector of the model at time t, LSTM output at time t, and memory cell state respectively. The symbol * represents the element-wise multiplication. The tanh is the hyperbolic tangent function and σ is the logistic sigmoid function. i_t, o_t, and f_t are the input, output, and forgot gates respectively. c_t is the input update value b_i, b_o, b_m, and b_f are the bias vectors of each gate, W are the weight matrices.

Fig. 3

Overview of LSTM cell at time t.

The bidirectional structure is applied to the LSTM to make a Bidirectional LSTM (BiLSTM) as shown in Fig. 4.

Fig. 4

Bidirectional LSTM structure.

In BiLSTM, the two hidden sequences forward $\vec{h_{t}}$ and backward are used to calculate the output sequence y_t by iterating from time t = 1 to t = T and t = T to t = 1 respectively. The forward, backward and output sequences are represented in Equations (10), (11) and (12) as follows: $\vec{h_{t}} = H (W_{x \vec{h}} x_{t} + W_{\vec{h} \vec{h}} \vec{h_{t - 1}} + b_{\vec{h}})$ (10) $\vec{h_{t}} = H (W_{x {\overset{\leftarrow}{h_{t}}}^{x_{t}}} + W_{\overset{\leftarrow}{h} \overset{\leftarrow}{h}} {\overset{\leftarrow}{h}}_{t + 1} + b_{\overset{\leftarrow}{h}})$ (11) $y_{t} = (W_{\vec{h} y} {\vec{h}}_{t} + W_{\overset{\leftarrow}{h} y} {\overset{\leftarrow}{h}}_{t} + b_{y})$ (12)

By using BiLSTM, trends in air quality data can be effectively captured from past and future dependency information.

The time series Air Quality dataset is split into train and test set constituting 80% and 20% instances. The training set data is transformed into time-series samples using a sliding window approach as in Algorithm 1. The data, window size, and step length are the inputs to the algorithm that outputs X and Y used to train the models.

The number of samples in a batch is the window size for each epoch in training. The training model takes sequence of the data with the chosen window size as the input and generates the output for the future time steps in multiple batches. The training data input and output is split based on the sliding window procedure as in the Algorithm 2,

Algorithm 2 Splitting time serial data
Input: data, win_size, step_length
Output: X, Y
1. function sliding_window (data, win_size, step_length)
2. d ← length(data)
3. win_num ← d - win_size - step_length + 1;
4. for i from 1 to win_num do
5. window ← data (i:i + win_size - 1);
6. X ← [X; window]
7. step ← data (i + win_size: i + win_size +
step_length - 1);
8. Y ← [Y; step];
9. end for
10. return X, Y;
11. end function

where data is one-by-d matrix, d is the length of the data, win_size is the number of observations in the sliding window, step_length is the number of steps ahead to forecast, X is the training input, and Y is the training output.

Based on this BiLSTM model is trained, fitted, and deployed on the smart fog gateway. Further, based on the real-time data acquired from the developed air quality monitoring system at the same location, the future hours pollutant values are predicted by the BiLSTM model deployed on the smart fog gateway and AQI is determined. AQI is then categorized into one of the situations as in Table 2. If the AQI category is very poor or severe, the early warning module on the fog node issues warnings to the stakeholders. Also, the fog node configures the end node to sample data more frequently in the occurrence of such events to gain a detailed inference.

4.2.1.4 Performance evaluation metrics

To evaluate the accuracy of the forecasting models the statistical metrics namely, RMSE and MAE are considered. It comprehensively investigates the difference between the actual and predicted values. The equations defining the metrics are listed below.

–Root Mean Square Error (RMSE): RMSE is a statistical parameter to determine the accuracy of a model. Smaller the RMSE indicates a lower error. $RMSE = \sqrt{\frac{1}{n} \sum_{j = 1}^{n} {(y_{j} - {\hat{y}}_{j})}^{2}}$ (13)

–Mean Absolute Error (MAE): MAE is another statistical measure to measure the average performance of the model. Smaller values of MAE are desired. $MAE = \frac{1}{n} \sum_{j = 1}^{n} | y_{j} - {\hat{y}}_{j} |$ (14) where y_j is the actual value, $\hat{y}$ is the predicted value, and n denotes the total number of samples.

4.2.2 Cloud services

The air quality data transmitted from the Fog Computing layer using MQTT protocol is fetched by AWS IoT Core and the data is stored in the AWS Dynamo DB, a No SQL database. Further a Lambda function is created using the AWS Lambda console and HTTP API is created using the API Gateway console. Whenever the user requests HTTP API, API Gateway forwards the request to the Lambda function, fetches data from Dynamo DB, and returns the data to the front-end of the website as highlighted in Fig. 5 and presents the Air Quality data to the users.

4.3 Result and analysis

Fig. 5

Air quality monitoring and prediction system.

We conducted a series of experiments to evaluate the prediction approach proposed in the paper. First, the influence of meteorological features on air quality prediction is validated. Second, the prediction performance of the Bidirectional LSTM model in the proposed system is verified by a comparative analysis with the baseline models for predicting six pollutants as presented in Table 10. The baseline models include LSTM, Convolutional LSTM (Conv-LSTM), Multilayer Perceptron (MLP), Decision Tree Regression (DTR), Gated Recurrent Units (GRU) and Bidirectional GRU (BiGRU). The prediction performance of the models is validated against MAE and RMSE to compare the actual and predicted values. The parameters of the models are optimized using grid search. By using BiLSTM, trends in air quality data can be effectively captured from past and future dependency information.

Table 10

Prediction performance of models without considering meteorological parameters

S. No	MODELS	METRICS	PM_2.5	PM₁₀	NO₂	SO₂	CO	Ozone	Total	Average
1	LSTM	RMSE	8.053360	14.85228	14.23015	1.905714	0.261958	13.43559	52.73905	8.789842
		MAE	6.175676	9.502236	10.584	1.319576	0.177388	10.2347	37.99357	6.332261
2	BiLSTM	RMSE	7.381752	15.0811	15.13218	1.92881	0.231522	14.13347	53.88884	8.981474
		MAE	5.424083	9.702279	11.4445	1.295015	0.161847	11.71465	39.74237	6.623728
3	Conv-LSTM	RMSE	8.240373	13.75506	13.79699	1.606259	0.143464	13.42197	50.96412	8.49402
		MAE	6.117047	9.480564	9.081684	1.125045	0.204779	10.14848	36.15759	6.026266
4	MLP	RMSE	7.533423	19.90301	12.98253	2.21557	0.249687	11.19589	54.08012	9.013353
		MAE	5.761191	13.6882	9.662147	1.408238	0.168592	8.677899	39.36627	6.561045
5	DTR	RMSE	8.26282	15.37592	19.08861	1.48287	0.124136	13.35934	57.69369	9.615615
		MAE	5.88988	9.155012	15.06026	1.154313	0.083284	9.72129	41.06404	6.844007
6	GRU	RMSE	8.771185	14.08342	12.99094	1.663069	0.21886	16.08324	53.81072	8.968453
		MAE	6.919886	10.91078	8.159612	1.203389	0.15402	12.52108	39.86876	6.644794
7	BiGRU	RMSE	8.104585	11.09623	12.14566	2.205417	0.317763	21.0798	54.94945	9.158241
		MAE	5.523153	7.314919	9.438648	1.363604	0.185659	14.17586	38.00184	6.333641

The hyperparameters play a vital role in the performance of the Deep Learning models. The BiLSTM network is designed with 200 neurons in the first hidden layer and six neurons in the output layer for predicting pollutant values. With the same configuration, over-fitting issues occur when neurons exceeded 200. The models are trained for 100 epochs. To avoid overfitting, using a dropout of 0.2 between the layers provided good results. Adam’s version of the stochastic gradient descent is chosen as the optimizer which is a good option for RNN. Deciding the window size is important because a smaller window size cannot provide sufficient information for model inputs, while a higher window size increases the computation complexity and the problem of learning important features. To minimize the computing costs, we chose the window size as 194 (2 days data). There is no specified rule to choose the best value of hyperparameters and is identified by the trial-and-error method. We also aim to determine and optimize the prediction model’s hyperparameters to achieve the best accuracy. The models are tested with multiple network configurations to find the best network with more accurate results.

The influence of meteorological parameters on pollutants is validated by comparing the results of the prediction model with and without considering the meteorological parameters as in Tables 10 and 11 respectively. On comparing the results of Tables 10 and 11, the average RMSE and MAE values of pollutants for Table 10 that does not consider meteorological parameters are significantly higher for all the models compared to Table 11 that considers the meteorological parameters. This implies that models trained with the dataset considering the meteorological parameters as in Table 11 have good prediction performance, confirming the influence of meteorological parameters on the pollutants prediction results.

Table 11

Prediction performance of different models considering the meteorological parameters

S. No	MODELS	METRICS	PM_2.5	PM₁₀	NO₂	SO₂	CO	Ozone	Total	Average
1	LSTM	RMSE	7.543121	14.462348	12.987323	1.936662	0.247156	11.824361	49.000971	8.166828
		MAE	5.428765	11.627475	9.342756	1.356253	0.169923	8.753752	36.678924	6.113154
2	BiLSTM	RMSE	6.735464	10.33456	12.457381	1.802735	0.210876	11.984627	43.525643	7.254273
		MAE	5.123332	6.452841	8.729476	1.310002	0.165768	9.065432	30.846851	5.141141
3	Conv-LSTM	RMSE	7.947834	15.847663	10.345634	1.783543	0.200342	11.432568	47.557584	7.926264
		MAE	5.910354	11.345734	7.785435	1.334562	0.130364	9.067281	35.57373	5.928955
4	MLP	RMSE	6.904227	18.874956	13.245273	2.296284	0.273645	12.736452	54.330837	9.055139
		MAE	5.384284	13.728347	10.200345	1.592374	0.138487	9.128467	40.172304	6.695384
5	DTR	RMSE	7.273345	14.237639	21.274767	1.774658	0.274662	12.274647	57.109718	9.518286
		MAE	4.634749	7.746256	15.726454	1.274656	0.147636	8.354638	37.884389	6.314064
6	GRU	RMSE	8.738556	13.526489	11.842905	1.726475	0.216385	15.123645	51.174455	8.529075
		MAE	6.537473	9.637485	8.742953	1.458356	0.152837	12.027493	38.556597	6.426099
7	BiGRU	RMSE	7.124653	16.285737	10.784676	1.937583	0.127562	9.152746	45.412957	7.568826
		MAE	5.276382	13.287482	6.924662	1.326237	0.117465	6.279293	33.211521	5.535253

Further, analyzing the results of Tables 10, 11, it is observed that the ML-based DTR model has larger MAE and RMSE values reflecting its poor predictive performance. Among the variants of RNN (LSTM, BiLSTM, Conv-LSTM, GRU, and BiGRU) show better results compared to the Feed Forward network model (MLP) and DTR. Furthermore, comparing the variants of RNN, BiLSTM has a lower total error value of MAE and RMSE and performs well on all pollutant factors except CO and Ozone. BiGRU has the second smallest average RMSE and MAE values for CO and Ozone, indicating good performance. In addition, Conv LSTM also has a performance closer BiGRU. The comparison results suggest RNN models are more suitable for air quality forecasting. The fitting curve representing the patterns of predicted values of the BiLSTM model over the actual values is displayed in Fig. 6. The orange line indicates the predicted values while the blue line represents the actual values of the data. The trends of the predicted values are reflected closer to the actual values of the pollutants in the BiLSTM model which helps in determining AQI effectively. The results show that BiLSTM not only captures the local trends, but also the long-term temporal dependencies of pollutant concentrations and meteorological parameters from the data. BiLSTM accurately predicts the pollutant concentration fluctuations and delivers stable performance. To sum up the comparison, BiLSTM exhibits superior performance over the baseline models as indicated by RMSE and MAE values in Table 11. Our forecasting model comparison results are consistent with those of the related studies that DL based method is more suitable for air quality prediction.

Fig. 6

BLSTM forecast results vs actual values of six primary pollutants.

The analysis of the results of Table 11 reveals that the BiLSTM model is a suitable model and achieves more accurate air quality predictions and by integrating historical time air quality data and meteorological data into the model. Further, the BiLSTM model is deployed on the fog node and the air quality predictions are made for future time steps and users are warned if AQI breaches the threshold. Further, the historical data is stored in the cloud and the clients access the real-time data through API calls.

5 Conclusion

In this paper, an Air Quality Monitoring and Prediction System is developed leveraging Fog Computing and Deep Learning. A low-cost real-time air quality monitoring system acquires the primary pollutants along with the meteorological parameters and transmits them to the fog node using LoRa for local processing and accurate forecasting to provide early warnings. The paper also reviewed and analyzed the state-of-the-art air quality prediction methods to tackle the highly dynamic pollution forecasting problem that plays a vital role in making informed decisions, supporting early warning and control mechanisms to mitigate the poor air quality. Further, to determine an accurate model, seven models are compared, and the results verify that the performance of BiLSTM outperformed other methods to accurately forecast future hour’s pollutants concentration by considering the meteorological parameters. The BiLSTM model is then deployed on the fog node to make future hours air quality predictions. Thus, the fog node is embedded with intelligence making it a smart fog gateway. The smart fog gateway forwards the real-time air quality data to the cloud and users access the data through cloud services. The prediction of air quality on smart fog gateway helps to determine future hours AQI in real-time and help the city planning committee in road traffic signal coordination and promote the usage of public transport for commutation. In turn, the proposed Air Quality Monitoring and Prediction System aims to bring environmental, economic, and social benefits by leveraging state-of-the-art technologies to promote a sustainable smart city. In the future, we plan to optimize Deep Learning methods on Fog Computing devices for long-term predictions.

References

Hashem

I.A.T.

, Yaqoob

, Anuar

N.B.

, Mokhtar

, Gani

and Khan

S.U.

, The rise of “big data” on cloud computing: Review and open research issues, Information Systems 47 (2015), 98–115.

Jiang

X.-Q.

, Mei

X.-D.

and Feng

, Air pollution and chronic airway diseases: what should people know and do? Journal of Thoracic Disease 8(1) (2016), 31.

https://www.technologyreview/2019/05/08/135467/satellites-will-let-us-track-air-pollution-from-every-powerplant-in-the-world/’.

, Dhelim

, Ning

and Qiu

, Survey on fog computing:architecture, key technologies, applications and open issues, , Journal of Network and Computer Applications 98 (2017), 27–42.

CPCB (2015) National Air Quality Index. pp. 1–55.

Bonomi

, Milito

, Zhu

, Addepalli

Fog computing and its role in the internet of things, in: Proceedings of the first edition of the MCC workshop on Mobile cloud Computing (2012), 13–16.

Santos

, Leroux

, Wauters

, Volckaert

, TurckAnomaly

F.D.

Anomaly detection for smart city applications over 5g low power wide area networks, in: NOMS 2018-2018 IEEE/IFIP Network Operations and Management Symposium (2018), 1–209.

Constant

, Borthakur

, Abtahi

, Dubey

, Mankodiya

Fog-assisted wiot: A smart fog gateway for end-to-end analytics in wearable internet of things, arXiv preprint arXiv:1701.08680 (2017).

Tuli

, Mahmud

, Tuli

and Buyya

, Fogbus: A blockchain-based lightweight framework for edge and fog computing, Journal of Systems and Software 154 (2019), 22–36.

10.

Dutta

, Roy

IoT-fog-cloud based architecture for smart city: Prototype of a smart building, in: 2017 7th International Conference on Cloud Computing, Data Science Engineering-Confluence (2017), 237–242.

11.

Gia

T.N.

, Queralta

J.P.

, Westerlund

Exploiting LoRa, edge, and fog computing for traffic monitoring in smart cities, Elsevier (2020), 347–371.

12.

Mekki

, Bajic

, Chaxel

and Meyer

, A comparative study ofLPWAN technologies for large-scale IoT deployment, ICTExpress 5(1) (2019), 1–7.

13.

Sinha

R.S.

, Wei

and Hwang

S.-H.

, A survey on LPWA technology:LoRa and NB-IoT, Ict Express 3(1) (2017), 14–21.

14.

Tan

Comparison of LoRa and NB-IoTin Terms of Power Consumption (2020).

15.

Osman

N.I.

, Abbas

E.B.

Simulation and modelling of LoRa and sigfox low power wide area network technologies, in: 2018 International Conference on Computer, Control, Electrical, and Electronics Engineering (ICC-CEEE) (2018), 1–5.

16.

Carlsson

, Kuzminykh

, Franksson

, Liljegren

, Measuring a LoRa network: performance, possibilities and limitations, Springer (2018), 116–128.

17.

Chall

R.E.

, Lahoud

and Helou

M.E.

, LoRaWAN network: Radiopropagation models and performance evaluation in variousenvironments in Lebanon, IEEE Internet of Things Journal 6(2) (2019), 2366–2378.

18.

Basford

P.J.

, Bulot

F.M.

, Apetroaie-Cristea

and Cox

S.J.

, and S.J.Ossont, LoRaWAN for smart city IoT deployments: A long termevaluation, Sensors 20(3) (2020), 648.

19.

Sanchez-Gomez

, Gallego-Madrid

, Sanchez-Iborra

Performance study of LoRaWAN for smart-city applications, in: 2019 IEEE 2nd 5GWorld Forum (5GWF) (2019), 58–62.

20.

Lai

, Yang

, Wang

and Chen

, IoT implementation of Kalmanfilter to improve accuracy of air quality monitoring and prediction, Applied Sciences 9(9) (2019), 1831.

21.

Duangsuwan

, Takarn

, Nujankaew

, Jamjareeg-Ulgarn

A study of air pollution smart sensors lpwan via nb-iot for thailand smart cities 4.0, in: 2018 10th International Conference on Knowledge and Smart Technology (KST) (2018), 206–209.

22.

Kim

S.H.

, Jeong

J.M.

, Hwang

M.T.

, Kang

C.S.

Development of an IoT-based atmospheric environment monitoring system, in: 2017 International Conference on Information and Communication Technology Convergence (ICTC) (2017), 861–863.

23.

Santos

, Leroux

, Wauters

, Volckaert

, TurckAnomaly

F.D.

Anomaly detection for smart city applications over 5g low power wide area networks, in: NOMS 2018-2018 IEEE/IFIP Network Operations and Management Symposium (2018), 1–9.

24.

Wang

, Jiang

, Dan

Design of air quality monitoring system based on Internet of Things, in: 2016 10th International Conference on Software, Knowledge, Information Management & Applications (SKIMA) (2016), 418–423.

25.

Toma

, Alexandru

, Popa

and Zamfiroiu

, IoT solution forsmart cities’ pollution monitoring and the security challenges, Sensors 19(15) (2019), 3401.

26.

Senthilkumar

, Venkatakrishnan

and Balaji

, Intelligent basednovel embedded system based IoT enabled air pollution monitoringsystem, , Microprocessors and Microsystems 77 (2020), 103172.

27.

Rebeiro-Hargrave

, Motlagh

N.H.

, Varjonen

, Lager-systemfor

Lager-system for real-time urban air quality monitoring, in: 2020 15th IEEE Conference on Industrial Electronics and Applications (ICIEA) (2020), 1–6.

28.

Baiocchi

, Al-Masri

, Biondi

, Jeyaraman

, Pospisil

, Boyer

, Souza

C.P.D.

Air Pollution Detection System Using Edge Computing (2019).

29.

, Yang

, Han

and Move

O.A.

, RAQ-A random forestapproach for predicting air quality in urban sensing systems, Sensors 16(1) (2016), 86.

30.

Taneja

, Sharma

, Oberoi

, Navoria

Predicting trends in air pollution in Delhi using data mining, in: 2016 1st India international conference on information processing (IICIP) (2016), 1–6.

31.

Ameer

, Shah

M.A.

, Khan

, Song

, Maple

, Islam

S.U.

and Asghar

M.N.

, Comparative analysis of machine learning techniques forpredicting air quality in smart cities, , IEEE Access 7 (2019), 128325–128338.

32.

Bisht

, Seeja

K.R.

Air pollution prediction using Extreme Learning Machine: A case study on Delhi (India), in: Proceedings of First International Conference on Smart System, Innovations and Computing (2018), 181–189.

33.

Rekhi

J.K.

, Nagrath

, Jain

Forecasting Air Quality of Delhi Using ARIMA Model, Springer (2020), 315–325.

34.

Zheng Li

, Lu

and Ruan

, A multiple kernel learning approach for air quality prediction, , Advances in Meteorology 2018 (2018).

35.

Leong

W.C.

, Kelani

R.O.

and Ahmad

, Prediction of air pollutionindex (API) using support vector machine (SVM), Journal ofEnvironmental Chemical Engineering 8(3) (2020), 103208.

36.

Bai

, Li

, Wang

, Xie

and Li

, Air pollutantsconcentrations forecasting using back propagation neural networkbased on wavelet decomposition with meteorological conditions, Atmospheric Pollution Research 7(3) (2016), 557–566.

37.

Yan

, Liao

, Yang

, Sun

, Nong

and Li

, Multi-hour andmulti-site air quality index forecasting in Beijing using CNN, LSTM,CNN-LSTM, and spatiotemporal clustering, , Expert Systems withApplications 169 (2021), 114513.

38.

, Peng

, Yao

, Cui

, Hu

, You

and Chi

, Longshort-term memory neural network for air pollutant concentrationpredictions: Method development and evaluation, , EnvironmentalPollution 231 (2017), 997–1004.

39.

Bharathi

P.D.

, Ananthanarayanan

, Sivakumar

P.B.

Computing-Based Environmental Monitoring Using Nordic Thingy: 52 and Raspberry Pi, Springer (2020), 269–279.

40.

Prakash

, Darshaun

K.G.

, Yaazhlene

, Ganesh

M.V.

and Vasudha

, Fog computing: Issues, challenges and future directions, International Journal of Electrical and Computer Engineering 7(6) (2017), 3669.

41.

Online: https://app.cpcbccr.com/ccr/#/caaqm-dashboard-all/caaqm-landing/data.

42.

Ameer

, Shah

M.A.

, Khan

, Song

, Maple

, Islam

S.U.

and Asghar

M.N.

, Comparative analysis of machine learning techniques for predicting air quality in smart cities, IEEE Access 7 (2019), 128325–128338.

43.

Athira

, Geetha

, Vinayakumar

and Soman

K.P.

, Deepairnet: Applying recurrent networks for air quality prediction, , Procedia Computer Science 132 (2018), 1394–1403.

Fog computing enabled air quality monitoring and prediction leveraging deep learning in IoT

Abstract

Keywords

1 Introduction

2 Related studies

2.1. Air quality monitoring and related requirements

Table 1 Requirements of Air pollution monitoring Monitoring Development, installation, Energy consumption Accuracy Response time Cost system and maintenance Indoor Easy Low Medium Medium Low Outdoor Medium Simple High Medium Medium Industrial Medium Medium Very High Fast Medium

2.2.1 Characterization of fog computing

2.3 Low Power Wide Area Network (LPWAN) technologies

3.2.1 Fog computing services

3.3 Cloud computing layer

4 System implementation

4.2.1 Air quality prediction model

4.3 Result and analysis

References

Table 1
Requirements of Air pollution monitoring

Monitoring Development, installation, Energy consumption Accuracy Response time Cost

system and maintenance

Indoor Easy Low Medium Medium Low

Outdoor Medium Simple High Medium Medium

Industrial Medium Medium Very High Fast Medium