Indoor air quality prediction using optimizers: A comparative study

1.1 Background

Poor IAQ is rising as a significant challenge that leaves a relevant impact on the poorest and unprotected individuals worldwide [10]. Its association with human well-being is comparable to sexually transmitted diseases and tobacco use [10, 16]. The Environmental Protection Agency (EPA) in the United States takes the responsibility to develop regulations for maintaining indoor and outdoor air quality. Studies published by EPA also reveal that indoor pollutant levels are approximately one hundred times higher than outdoor pollutant levels, leading to some of the most significant environmental problems affecting general well-being [17].

There is a wide range of primary and secondary pollutants that contribute to poor IAQ; however, PM levels are recognized as major contributing parameters for decayed public health [18]. For routine monitoring and regulatory purposes, ambient PM is divided into two different quantities: PM₁₀ and PM_2.5. They represent mass concentrations of tiny particles collected by automatic systems as per international sampling standards related to ‘inhalable’ and ‘respirable’ particles. M_10–2.5, particles with the aerodynamic diameter >2.5 microns, but≤nominal 10 microns are defined as coarse fraction particles. Particles with diameter <1 micron are represented by PM_1.0. Furthermore, “Ultrafine” particles (UFP) are the particles with diameter < 0.1 microns (100 nm) in size [18].

PM is a multifaceted combination of liquefied and solid biological and mineral particles suspended in the air [19, 20]. PM generally includes dirt, dust, smoke, soot, and liquid droplets that can penetrate the lower airways of humans leading to severe harmful health effects. Indoor exposure to PM in developing countries increases the risk of acute lower respiratory infections [21, 22]. Several studies reveal that excessive exposure to PM levels is closely associated with mortality among young children [23, 24]. At the same time, it is a prominent cause of several chronic health problems, including chronic obstructive pulmonary disease, cardiovascular disease, and lung cancer among adults [25, 26].

Approximately 60,000 premature deaths are reported every year in the United States, and they are directly linked to poor air quality levels. The cost of healthcare facilities for treating diseases associated with air quality reach $150 billion [27]. As per a report presented by the European Environment Agency in the year 2016, air pollution caused 400,000 premature deaths within the European Union (EU). PM levels lead to 412,000 premature deaths in 41 different European countries 374,000 were reported in the EU [28]. Moreover, the cost of dealing with the harmful effects of air pollutants emitted from industrial facilities in 2012 was estimated somewhere around 59 to 189 billion euros in the EU [29]. The highly populated cities in India contribute to the highest pollution levels in the world, and the capital city of Delhi is mentioned on the top of the list. As per a study conducted by CNN in the year 2019, 21 out of the world’s 30 most polluted cities were from India [30]. The statistics of the year 2016 reveal that 140 million people out of 1.38 billion population in are affected by air quality, which is ten times worst as compared to the safe limit set by WHO [31]. Poor IAQ is the second major cause behind the increased mortality rate in India [32]. The 2018 Environment Performance Index Results report ranked India 177th among 180 countries based on 24 performance indicators covering ecosystem vitality and environmental health [33]. Moreover, over 1,00,000 deaths are reported in 29 Indian cities due to rising PM2.5 levels [34]. The adverse effects of ultrafine air particles are further linked to lung health and systemic circulation, where toxic components can cause potential tissue damage and inflammation [35]. It is necessary to notice that PM exposure is ubiquitous, and there is no defined and studied “safe” level. Hence, behavioural modification strategies can play a considerable role in enhanced overall health and well-being. The technology inspired PM monitoring and prediction systems can help policymakers to develop solid strategies for maintaining public health. That is why several researchers in the past few years have contributed their ideas in this direction. Numerous experimental studies from recent years are highlighted in Section 1.2.

1.2 Related work

Ghaemi et al. [36] proposed a Spatio-temporal system designed using the LaSVM-based online algorithm for dynamic prediction of air pollution in Tehran. The authors used meteorological data on carbon monoxide (CO), nitrogen dioxide (NO₂), sulfur dioxide (SO₂), ozone (O₃), and PM₁₀ from few target locations in Tehran. These pollutant concentrations were further used to calculate Air Quality Index (AQI) and then a prediction system was designed to estimate those values based on past monitoring. This system was able to predict AQI values for the next 24 hours with a root mean square error (MSE) of 0.54, the overall accuracy of 0.71, and a coefficient of determination equal to 0.81.

Another air quality prediction system was proposed by Zheng et al. [37] with the Multiple Kernel Learning approach. The four parameters used in this experimental study were O₃, NO₂, SO₂, PM, fine suspended particulates and respirable suspended particulates. The performance of the proposed system was further compared with the Random Forest (RF), classical autoregressive integrated moving average (ARIMA) model, support vector machine (SVM), long short-term memory (LSTM) and multiple-layer perceptron (MLP) neural network.

Soh et al. [38] designed an adaptive deep learning model for predicting air quality using spatial-temporal relations. This model was developed based on two datasets collected from 76 different locations of 23 cities of Taiwan and Beijing. The prediction system was mainly designed for predicting PM_2.5 based on wind direction, wind speed, temperature, humidity, PM_2.5, and PM₁₀ features. The system was able to forecast air quality for 48 hours using a combination of different neural networks, including LMST, Convolution Neural Network (CNN), and Artificial Neural Network (ANN) for extracting spatial-temporal relations.

Zhu et al. [39] proposed an air quality forecasting system using machine learning methods that could provide a prediction for around 24 hours. This study focused on three air pollutants: SO₂, PM_2.5 and O₃. The data was collected from two different weather stations located in the two suburban residential areas of Illinois. The authors in this study presented refined models for predicting IAQ by using advanced regularization techniques for performance optimization.

Tiwari et al. [40] also designed an air pollution level prediction system by considering four major pollutant concentrations: O₃, CO₂, SO₂, and NO₂. The authors of this study presented a feed-forward ANN-based model for predicting IAQ. The three input parameters applied at the input of ANN were Mean, 1^st Hourly Max and 1^st Max Value, whereas the output parameter was AQI. Furthermore, the ARIMA model was used to validate the predicted AQI.

In this experimental analysis, we designed an IoT based monitoring system for measuring essential IAQ parameters, and the prediction system was further developed using Keras optimizers. Although there are several libraries and frameworks available for handling time series analysis, the authors decided to work with Keras due to its ability to handle small datasets more efficiently. This open-source library is also preferred for the simplicity of coding environment, core architecture, and ability to run on the top of relevant frameworks such as TensorFlow, Theano, and R. The details about data collection, pre-processing and system development are described in the further sections.

2.1 Dataset

Chandigarh is a union territory in India and is the capital of two neighbouring states, Punjab and Haryana. The estimated population of the city is 1.06 million with a higher number of vehicles and industrial units that lead to severe air pollution. This experimental analysis is carried out to estimate the state of IAP by measuring pollutant concentrations in the area that is badly affected by poor air quality. The sensor network was installed in the cafeteria of an educational institute (National Institute of Technical Teacher’s Training and Research) in Chandigarh city. The cafeteria stays populated throughout the day with peak traffic between 11:00 am to 11:30 am and 4:00 pm to 4:30 pm. The sensor network was placed close to the cooking area. The data was recorded for 6 months, two readings per hour, from 5th April to 4th September 2019.

As per WHO, the pollution index for Chandigarh city is 51.32. Moreover, air quality and air pollution are moderate, with an estimated value of 54.57 and 45.43, respectively. The PM10 level usually stays around 110, whereas PM2.5 is observed to be 59. However, the PM10 pollution level in the city is reported to be very high around the year. On the other side, the temperature typically varies between 48- and 106-degrees Fahrenheit over the year, and humidity stays within the estimated range of 24% to 97%. But as the vehicle traffic and industrial pollution is increasing with each passing day, the scenario may be more complicated in the coming years [41]. Hence, an early evaluation of air pollution levels and their impact is necessary.

2.2 Requirements:

Hardware requirements:

Arduino Uno

Software requirements:

Arduino IDE

2.3 Methodology

The design of the prediction system is based on the real-time data collected using IoT sensor networks installed in Chandigarh City. The framework of the methodology is given in Fig. 1.

OPEN IN VIEWER

Fig. 1

Framework of the methodology.

The entire process can be categorized into three stages as data collection, data pre-processing and prediction model design. The detailed explanation of the methodology is given below.

2.3.1 Hardware design and data collection:

For analysing the levels of IAP and to design a prediction system for timely indication of critical instances, an IoT sensor network with PM, temperature, and humidity sensor was developed. Real-time data was collected with the help of the Arduino Uno based sensor network that uses the SDS011 sensor for measuring PM_2.5, PM₁₀ and DHT11 for measuring temperature and humidity. Table 1 shows the specifications of the essential hardware used for the cyber-physical system. The data was sent to the ThingSpeak channel for online storage by using the ESP8266 module. The diagram representing a sensor network for measuring pollutant concentrations is shown in Fig. 2.

OPEN IN VIEWER

Table 1

Specifications of sensors used for designing an IAQ monitoring system

Hardware	Parameters	Measuring rate	Accuracy	Power consumption info
SDS011	PM2.5, PM10	0.0–999.9μg/m3	15% and±10μg/m3	Voltage: 5 V Current: 70 mA±10 mA
DHT11	Temperature and Relative Humidity	Temperature: 0–50°C	Temperature: ±2°C	Voltage: 3–5.5 V
		Humidity: 20–90%	Humidity: ±5% RH	Operating current: 0.3 mA (measuring) 60 uA (standby)

OPEN IN VIEWER

Fig. 2

Framework of IoT sensor network for measuring pollutant concentration.

Table 1 shows the specifications of data collection elements used to record real-time measurements from the target monitoring area. However, the accuracy of the sensors installed is a crucial factor to verify the reliability of the monitoring system.

Reliability tests for sensors: DHT11 sensor was considered for this study due to its finest temperature and humidity measurement range with the accuracy between 2 to 5%. This sensor comes in pre-calibrated form from the industry. However, in order to ensure accuracy in measurements, the performance was compared with the standard temperature and humidity measurement meter (Honeywell TM005X) available in the lab.

The decision for selection of SDS011 laser sensor over other low-cost sensors was supported by existing studies conducted by early researchers [42–44]. PM NOVA sensor was also pre-calibrated in the factory; however, the reliability test is required to ensure its accuracy for the IAQ monitoring system. The reliability test was conducted after 24 hours of a pre-heating period that is required to ensure accurate readings in a particular environment. The reliability of the sensor was continuously tested against the monitoring air station operated and maintained by the pollution control board in the city. The performance of the sensor was stable and within the accuracy range specified by the manufacturers. In order to maintain the accuracy of readings, careful sensor placement procedures were followed so that the normal flow of air cannot be blocked towards the chamber of the laser dust sensor.

Specifications of IAQ Gateway: IAQ Gateway consists of Arduino Uno and ESP8266 Wi-Fi module. ESP8266 is a self-contained system on chip with integrated TCP/IP protocol stack. It works on the basis of IEEE 802.11 standards and uses a Wi-Fi channel with 20 to 25 MHz band for communication. This module comes with a self-calibrated radio frequency (RF); hence, there is no need to add external RF parts for operation. The microcontroller uses AT commands over UART channels with a default baud rate of 115200 to communicate with the ESP8266 module.

The sensor network recorded data for four parameters PM_2.5, PM₁₀, temperature, and humidity. The natural deposition process of PM is greatly affected by relative humidity and temperature. Studies reveal that temperature has a negative correlation with PM₁₀, whereas relative humidity has a positive correlation with PM₁₀ [45].

The pollutant concentrations are further used to determine the AQI. It is basically an indicator of pollution levels as defined by EPA in the United States for public use. The formula specified for AQI calculations is given by Equation 1. I = I high - I low C high - C low ( C - C low ) + I low(1)

Where I represent AQI, C is pollutant concentration, C _low C_low is a concentration breakpoint, which is equal to or less than C. C _high C_high is a concentration breakpoint that is equal to or greater than C. I _low I_low and I _high I_high are the index breakpoints corresponding to C _low C_low and C _high C_high, respectively. By using this equation, the AQI for real-time IAQ data of IoT sensor network located in Chandigarh city is calculated. This data was further pre-processed and trained with the help of optimizers. The details are described in the further sections.

2.3.2 Data pre-processing

The collected air pollution data and calculated AQI values are paired together to create the desired data format that can be further processed through the prediction model. The hourly mean values of the recorded parameters were processed. However, the recorded data had several missing values and multiple records at some hours during 6 months of monitoring. To deal with the missing values and inappropriate entries in the dataset, the dataset was pre-processed. Masking approach was used to impute the missing values in the dataset, and the rows having NaN values were dropped to clean the data. After data pre-processing, a dataset containing relevant values of parameters was obtained. For visualization of available multidimensional data Kernel Density Estimation (KDE) algorithm is utilized. It helps to plot the shape of distribution by keeping the density of the observations in one axis and height along the other axis. The matrix plot of the pairwise relationship between parameters (AQI, PM10, PM2.5, Temperature and Humidity) of the available training dataset is shown in Fig. 3.

OPEN IN VIEWER

Fig. 3

Pairwise relationship between AQI, PM10, PM2.5, Temp and Hum parameters of the available training dataset.

In this section, the authors describe the different approaches for IAQ prediction. The selection of the right optimizer and configuration of a set of parameters can help to squeeze even the last bit of accuracy for the prediction model. This experimental analysis is based on seven different optimizers, and their performance is compared based on mean absolute error (MAE), MSE, Root Mean Square Error (RMSE), R² Score (Co-efficient of determination), Mean Absolute Percentage Error (MAPE) and Accuracy. In the section, the optimizers tested by the authors are succinctly explained.

Stochastic Gradient Descent:

Gradient descent (GD) is a fundamental technique for training and optimizing intelligent systems. It basically considers the following equation for minimizing loss function and for enhancing system accuracy: θ = θ - η · ∇ J ( θ ) θ = θ - η · ∇ J ( θ )(3)

Where ‘η’ represents learning rate, ‘∇J (θ)’ is the gradient of loss function-J(θ) w.r.t parameters- ‘θ.’

Stochastic Gradient Descent (SGD) optimizer is known as the simplest form of GD in terms of its behaviour and concept as well [46]. It starts with a small learning rate and follows the gradient on the cost surface. After each iteration, it generates new weights that are better than the old ones obtained in the previous iterations. It mainly includes support for learning rate decay and momentum. In the training network, a random batch of samples is taken for each iteration and values of θ are updated every time.

The mathematical representation for SGD is shown below: θ = θ - η · ∇ J ( θ ; x ( i ) ; y ( i ) )(4)

Where, x⁽ⁱ⁾ and y⁽ⁱ⁾ indicate training examples.

Adagrad:

Adagrad or Adaptive Learning Rate optimizer simply follows the learning rate (η) to adapt as per the input parameters [47]. It works by making small updates for the frequent parameters and big updates for the infrequent parameters. Hence, this optimizer is widely recommended for handling sparse data.

In the previous method, we made updates at once for all parameters θ because all θ_i components were having the same η . But in the case of Adagrad, all parameters θ_i follow different η at every time step t. For Adagrad, we set g t , i θ t + 1 , i = θ t , i - η G t , ii + ɛ . g t , i as the gradient of the loss function with respect to θ_i at time step t. The equation is shown below: θ t + 1 , i = θ t , i - η ∑ i = 1 t g i 2 + ɛ . g t , i(5)

Where ∈ = 10e - 8.

The biggest benefit of Adagrad is that it does not require manual tuning of η. In most cases, the default value is taken as 0.01.

Adadelta:

Adadelta is basically an extension of Adagrad; it works for removing the decaying learning rate problem of the previous algorithm. Instead of using all previously squared gradients, Adadelta restricts the window of all accumulated past gradients to a fixed size w [48].

Instead of storing w previous squared gradients, Adadelta defines the sum of gradients in a recursive manner as the decaying mean of the past squared gradients. The final formula for Adadelta is defined below: Δ θ t = - η E [ g 2 ] t + ɛ . g t(6)

Here the denominator just represents the Room Mean Square (RMS) error criterion of the gradient and can be written as RMS [g]  _t  ; hence, the 5th equation can be rewritten as: Δ θ t = - η RMS [ g ] t . g t(7)

One important thing to note about Adadelta is that one need not put special efforts for setting η.

RMSprop:

Room Mean Squared Propagation (RMSprop) algorithm is also a kind of adaptive learning rate method [49]. Both Adadelta and RMSprop were developed independently in the same duration to resolve the radically diminishing learning rate problem of Adagrad. RMSprop is almost identical to the first update vector of Adadelta and can be given as: Δ θ t = θ t - η E [ g 2 ] t + ɛ . g t(8)

Note that RMSprop also divides the model learning rate by the exponentially decaying average of squared gradients.

Adam:

Adaptive Moment Estimation (Adam) optimizer is recommended for average [50]. This algorithm makes use of adaptive learning rates for each parameter and momentum to ensure fast convergence. While storing the exponentially decaying average of the past squared gradients such as AdaDelta, Adam also keeps track of the exponentially decaying average of the past gradients m(t). The final formula for updating parameter with Adam optimizer is: θ t + 1 = θ t - η v ˆ t + ɛ . m ˆ t(9)

Here v_t and m_t are the uncentered variance and mean of the gradients, respectively.

This optimizer works well in the practical environment with its fast convergence rate and comparatively higher learning speed.

Adamax:

The v _t factor of the Adam optimizer update rule leads to the inversely proportional scaling of the gradient with respect to the ℓ₂ norm of the past gradients as well as the current gradient |g_t|²

Here, the basic equation is represented as: v t = β 2 p v t - 1 + ( 1 - β 2 p ) | g t | p v t(10)

It was further updated by Kigma and Ba [51], to achieve more stable value. In order to avoid confusion with Adam, the new equation was represented by u _t that represents infinity norm-constrained v _t . The new equation can be shown as: u t = β 2 ∞ v t - 1 + ( 1 - β 2 ∞ ) | g t | ∞(11)

Where β₂ represents decay rate whose default value is 0.999. With this improvement in the equation, the final Adamax update rule can be represented as: θ t + 1 = θ t - η u t . m ˆ t θ t + 1 = θ t - η u t . m ˆ t(12)

It is important to mention that u _t is dependent on the max operation; it does not go towards zero as v _t and m ˆ t in Adam. Hence, there is no need to calculate bias correction for u _t .

Nadam:

Generally, Adam is represented as the combination of momentum and RMSprop; on the other side, the Nadam optimizer or Nesterov-Accelerated Adaptive Moment Estimation is the combination of Adam and Nesterov Accelerated Gradient (NAG) [52]. Hence, the final update rule for Nadam can be represented by adding the Nesterov Momentum term to the Adam equation. The new rule can be shown as: θ t + 1 = θ t - η v ˆ t + ɛ . ( β 1 m ˆ t + ( 1 - β 1 ) g t 1 - β 1 t )(13)

Where β₁ is the decay rate with a default value of 0.9.

All these optimizer algorithms present small parametric variations over one another that add the difference to their overall performance in designing prediction system for IAQ.

Following the concept of seven different optimizers discussed in section 2.4, the IAQ prediction model was designed. The main idea was to compare the performance of all these optimizers based on relevant regression metrics so that the most efficient optimizer can be identified to design a real-time, intelligent IAQ monitoring and prediction system. Figure 4 shows the stages of the IAQ prediction model designed for this comparative study. The data obtained from the hardware module is fed to the prediction algorithm. However, instead of putting raw data as it is, the pre-processing stage is added to the system. The first most step at this stage was to clean the dataset by following the masking approach discussed in section 2.3.2. After that, the response label was applied to the AQI parameter and the rest parameters (PM10, PM2.5, Temp and Hum) were considered as input features to train the prediction model. It is crucial to mention at this stage that statistical feature extraction was used to describe the dataset. None of the input parameters were discarded at this stage since several studies report that PM concentrations are greatly affected by temperature and humidity levels [35, 53]. Hence, all four input parameters were used for training to ensure that their overall impact is utilized for designing efficient prediction model. After feature selection and label assignment, the data were normalized by following the technique mentioned in section 2.3.3. The normalized data was then used to train model containing seven different optimizers and then their performance was compared using six different regression metrics, including MSE, MAE, RMSE, R² score, MAPE and Accuracy. Based on these metrics, the best model was selected for designing an intelligent IAQ prediction system that can be used for real-time analysis in the future. The results obtained after training the prediction model with the seven optimizers are discussed in the next section.

OPEN IN VIEWER

Fig. 4

Stages of IAQ prediction model for comparing performance of seven optimizers.