Artificial Neural Network Models for Daily PM 10 Air Pollution Index Prediction in the Urban Area of Wuhan,China

Abstract

Dust storm is a critical remote source that causes low air quality in many cities in China. The prediction accuracy of high particulate matter with a diameter <10 μm (PM₁₀) air pollution index (API) event caused by dust storm is low in China. To solve this problem, dust storm data from northern China was first used to tune the Elman-based forecast model to predict the daily PM₁₀ API with a lead time of 1 day. Effectiveness of this forecaster was tested using a time series recorded from September 1, 2001, to December 31, 2007, at six monitoring stations located within the urban area of Wuhan, China. Experimental trials show that the improved Elman model provides low root mean square error values and mean absolute error values in comparison to the standard Elman model. In addition, higher coefficient of determination (r² = 0.62) and accuracy rate (83.33%) values were realized for the improved Elman model in comparison to the standard Elman model (r² = 0.22, accuracy rate = 64.81%) when predicting high PM₁₀ API events caused by dust storms.

Introduction

The particulate matter with a diameter <10 μm(PM₁₀) air pollution index (API) is a quantitative measure to uniformly report on the air quality of different constituents with respect to human health. API is the conversion of PM₁₀, one of the primary pollutants in China, which is measured at sampling stations at a 0:500 scale. An index result of 100 corresponds to the short-term “air quality objective” established under the Air Pollution Control Ordinance. China's Ministry of Environmental Protection classifies air quality standards into five major categories with respect to API values (Table 1): I (clean), II (good), III (low-level pollution), IV (mid-level pollution), and V (high-level pollution). In addition, categories III and IV include two subcategories each: III₁ and III₂ as well as IV₁ and IV₂.

Table 1.

Daily Air Pollution Index and Air Quality Management in China

Air pollution index	Air quality classification	Health influence	Air quality description and management
API ≤50	I (clean)	No	No action is needed.
50 < API ≤100	II (good)	No	No action is needed.
100 < API ≤150	III₁ (low-level pollution)	Symptoms are light but aggravating in healthy people.	People should be cautions when participating in outdoor activities.
150 < API ≤200	III₂	Moreover, symptoms start to appear in healthy people.
200 < API ≤250	IV₁ (mid-level pollution)	Patients with heart disease and pulmonary symptoms are affected notably. An endurance reduction commonly appears in healthy people while exerting themselves outdoors.	People with existing heart or respiratory illnesses are advised to reduce physical exertion and outdoor activities.
250 < API ≤300	IV₂
API ≥300	V (high-level pollution)	Healthy people exhibit obvious and intense symptoms while participating in outdoor activities. Certain diseases appear prematurely.	Air pollution is severe. Consequently, the general public is advised to reduce physical exertion and outdoor activities.

API, air pollution index.

Model forecasting with short lead times tends to be an adequate method to use when planning health warning systems. When an API reaches a high value, the use of forecasting models can best suggest in advance what regulations are to be enforced. This would prevent unnecessary annoyances for urban inhabitants. Since its introductory application in support of ambient pollutant concentration modeling (Boznar et al., 1991), the artificial neural network (ANN) model has been used to forecast a wide range of pollutants and concentration levels at various time scales with very good results (Coulibaly et al., 2001; Kolehmainen et al., 2001; Brunelli et al., 2007).

In the recent years, the application of ANN has been extended for use in the prediction of particulate matter concentrations (Perez and Reyes, 2002; Chaloulakou et al., 2003; Ordieres et al., 2005). For example, better forecast accuracy was achieved when applying neural networks for the prediction of PM₁₀ concentrations in Belgium (Kukkonen et al., 2005) than were other approaches such as linear regressors as well as deterministic physically based models and the use of meteorological predictors that benefit their performance. Hooyberghs et al. (2005) describe the development of a multilayer perceptron neural network to forecast the daily average PM₁₀ concentration of urban areas in Belgium at 1 day ahead. Corani (2005) compared feed-forward neural networks to prune neural networks and lazy learning neural networks. He found no strong differences between the forecasting accuracy of the different models. Nevertheless, lazy learning provides the best performance in terms of indicators related to average prediction goodness (correlation, mean absolute error [MAE], etc.), whereas prune neural network is superior in detecting the alarm exceedance and attention threshold. In some cases, deseasonalized data have been found to improve model prediction accuracy. Lu et al. (2002) have used neural networks to predict hourly respirable suspended particle concentrations in Hong Kong up to 24 h in advance by way of pollutant data used as model input values.

Only a few studies regarding API prediction have been published to date. Lang et al. (2004) combined the advantages of gray model GM(1,1), unbiased GM(1,1), unequal time difference GM(1,1), pGM(1,1), and back-propagation (BP) neural networks to calculate monthly APIs. Jiang et al. (2004) used an improved multilayer perceptron network to formulate an API prediction in Shanghai, China, 1 day in advance. Liu and Pan (2008) adopted the Levenberg-Marquardt algorithm to achieve a higher speed and a lower error rate in conjunction with the BP neural network, and as their results show, ANN outperformed the linear regression models.

Although API prediction was developed in some major cities in China in recent years, the prediction accuracy stayed in a low level, especially the prediction of high PM₁₀ API event caused by dust storm (Tong, 2006).

It has also been suggested that air pollution forecasting should be available at least 24 h in advance to provide crucial aid to authorities where they can enact short-term measures in the case of episodic events (Hubbard and Cobourn, 1998).

A recurrent neural network (Elman model) based forecaster for the prediction of daily maximum PM₁₀ API concentrations in the city of Wuhan is proposed in this study. Experimental trials were aimed to improve the Elman model to improve the prediction accuracy of high PM₁₀ API event caused by dust storm. During this phase, an Elman-based model was tested and improved upon to achieve the task of providing a 1 day in advance daily API forecast. Dust storm data from northern China were first used to tune the Elman-based forecaster to predict the daily PM₁₀ API with a lead time of 1 day. Networks were assembled using a time series recorded from September 1, 2001, to December 31, 2007, at six monitoring stations (St-1 to St-6) located within the city of Wuhan. Model validation was carried out by comparing model prediction values with a different set of “real-world” recorded data that were not used to train the network. To guarantee the robust performance of the model and to test set independent features, a cross validation strategy was used for validation. In addition, both the Elman and improved Elman models were tested and compared in their performance of achieving a 1 day in advance forecast of a high-level API event caused by a dust storm event originating in northern China.

Study Area and Data

Study area

Wuhan (see Fig. 1a) is the capital of Hubei Province located in central China (latitude 29°58′ N to 31°22′ N and longitude 113°41′ E to 1,151°05′ E). The Yangtze River—the world's third longest river—meets its largest tributary, Hanshui, in Wuhan, cutting the city into three parts: Hankou, Wuchang, and Hanyang, which are the three towns of Wuhan. Hankou, located on the northern bank of Yangtze River, is a metropolitan area that hosts governmental organizations, shopping centers, and residential districts. Wuchang, located on the southern bank of Yangtze River, is where tertiary schools, scenic spots, and heavy industrial and high-technology zones are situated. Finally, Hanyang, located on the western plain, is where residential areas, automobile factories, and scenic spots are found. The population of Wuhan is ∼8.58 million, and its total area is 8,494 km². Wuhan has a subtropical humid monsoon climate and experiences hot and humid summers each year. The average daily temperature in July is 37.2°C, and the maximum often exceeds 40°C. With a water body area making up 25.8% of its territory, Wuhan is ranked first among major Chinese cities in water resources. East Lake—China's largest urban lake—covers an area of 33 km². As well as being the political, economic, and cultural center of Hubei Province, Wuhan possesses one of the largest junctions of land, water, and air transportation in China.

FIG. 1.

Wuhan: (a) geography and (b) location of observational stations.

Data

Network training is based upon data measured during a 7-year period from September 1, 2001, to August 31, 2007. Daily PM₁₀ API data were acquired at St-1 to St-6 (Table 2 and Fig. 1b) made available by the Wuhan Environmental Protection Bureau. Meteorological variables of average daily temperature (T [°C]), relative humidity (%), wind speed (m s⁻¹), barometric pressure (P [bar]), rainfall amount (mm), and sunshine duration (h) were monitored at a meteorological station located within the Wuhan Meteorological Bureau (Table 2 and Fig. 1b).

Table 2.

Characterization of the Six Monitoring Station Locations and Basic Descriptive Statistics for the Air Pollution Index Measured from September 2001 to August 2007

Site	Location	Arithmetic mean	Median	Range
St-1	Residential suburban area	88.3	83	20–441
St-2	East Lake Park	76.1	73	14–190
St-3	Residential area in Hanyang	92.5	88	22–365
St-4	Residential area in Hankou	91.0	85.5	15–379
St-5	Commercial area	92.9	88	15–412
St-6	Heavy industrial area	100.6	96	11–397
St-M	Residential area in Hankou

St-1 to St-6, six monitoring stations; St-M, meteorological station.

Dust storms occur often in northern China during springtime. Research has shown that dust storms occurring over northwestern China may extend into the North Pacific Ocean, Japan, and Korea (Duce et al., 1980). Severe dust storms can even stretch across the Pacific Ocean and reach the western coast of North America as was the case on April 18, 1998 (Wang et al., 2003). Major sources of dust storms are the Gobi Desert in Mongolia, northern China, and the Taklimakan Desert in western China (Sun et al., 2001). During certain seasons, severe air pollution in Wuhan was the result of dust storms that developed within northern China (Li et al., 2004; Feng et al., 2010). As a result, dust storms originating in northern China were taken into account in this study. Daily dust storm data were provided by the China Meteorological Data Serving Service System (CMDSSS) (cdc.cma.gov.cn).

Meteorological variables were restricted to ones that can be routinely forecast by Wuhan Meteorological Bureau and CMDSSS.

Meteorological parameter input values used in the development of the models corresponds to the actual time for which the prediction applies in the absence of available data from numerical weather forecasts.

Methodology

The objectives of this article were to develop and test a recurrent neural network model. The following is a brief description of the model.

The Elman neural network used in this study was first proposed by Jeffrey L. Elman (Elman, 1990). For this network, both the outputs of the hidden and output layers are allowed to feedback onto themselves through a buffer layer referred to as the context layer. This feedback allows Elman networks to learn, recognize, and generate temporal as well as spatial patterns. Each hidden neuron is connected to only one context layer neuron by means of a constant weight of one value. Hence, the context layer virtually constitutes a copy of the state of the hidden layer. The number of context neurons, consequently, is equal to the number of hidden neurons. Alternatively, each output layer neuron can be connected to only one neuron of a second context layer by means of a constant weight of one value (Fig. 2). The input, output, and context neurons typically retain linear activation functions, and the hidden neurons retain the logistic activation function. In the experimental trials carried out for this study, however, the most successful operations were obtained using the following activation function: \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6}\begin{document} \begin{align*} a_j(t) = \begin{cases}0.82\times a_j(t-1)+0.8\times{\rm net}_j(t)\times(1-a_j (t-1)) & {\rm if} \ {\rm net}_j(t) > 0 \\ 0.82\times a_j(t-1)+0.8\times{\rm net}_j(t)\times(1+a_j (t-1)) & {\rm if} \ {\rm net}_j(t) \leq 0\end{cases} \tag {1} \end{align*} \end{document}

FIG. 2.

Improved Elman neural network architecture for St-i (i = 1, 2, … , 6). API, air pollution index; T, average daily temperature; RH, relative humidity; WS, wind speed; P, barometric pressure; RF, rainfall amount; SD, sunshine duration; St, station.

where a_j(t) is the activation of unit j in step t and net_j is the input of unit j in step t. The training algorithm is the resilient BP (RProp) (Riedmiller and Braun, 1993). The basic principle of RProp is to eliminate the harmful influence of the size of the partial derivative on the weight step. As a consequence, only the sign of the derivative is considered to indicate the direction of the weight update. Each weight, therefore, has its own adaptive step size Δ_ji: \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6}\begin{document} \begin{align*} \Delta_{ji} (t) = - \Delta_{ji} (t) {\times}{\rm sign} (g_{ji} (t)) , \tag{2} \end{align*} \end{document}

where Δ_ji is calculated as \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6}\begin{document} \begin{align*} \Delta_{ji} = \begin{cases} \eta^ + {\times} \Delta_{ji} (t - 1) \quad {\rm if} \ g_{ji} \cdot g_{ji} (t - 1) > 0 \\ \eta^ - {\times} \Delta_{ji} (t - 1) \quad {\rm if} \ g_{ji} \cdot g_{ji} (t - 1) > 0 \\ \quad \Delta_{ji} (t - 1) \quad \quad \, {\rm otherwise}\end{cases} \tag{3} \end{align*} \end{document}

where g(t) is the gradient function. The values of the parameters used to learn the neural network in this instance are η⁺ = 1.2, η⁻ = 0.5, Δ₀ = 0.07 (fixed starting value for Δ_ji), and Δ_max = 50 (the upper limit of Δ_ji).

Experimental Setup

Based on a previous study (Li et al., 2004, 2007), Elman mode input patterns for each station contain a set of daily values where one value is applied to both the current and the following day and the final value is the specific API station under consideration. There are, therefore, a total of 15 values for each Elman mode input pattern: average temperature (T), relative humidity, wind speed, barometric pressure (P), rainfall amount, sunshine duration from meteorological station, dust storm data from CMDSSS, and the PM₁₀ API from one specific station (St1–St6) as illustrated in Fig. 2.

The dataset, which was used to build the neural network database, constitutes the daily values related to a period from September 1, 2001, to December 31, 2007. Neural model performance was evaluated by applying a cross-validation strategy to test the effectiveness of the tested model for its ability to make accurate predictions. The entire dataset from September 1, 2001, to December 31, 2006, was used as a training set, whereas the 2007 dataset was shared between three subsets, using two of three subsets to complete the training set and the remaining subset applied as test set. As a result, three different training and test sets were used to guarantee robust performance and test set selection independency features for all models that were developed and tuned. All training and test sets are shown in Table 3.

Table 3.

Training and Test Sets Used for Cross-Validation Strategy

	2001	2002			…	2006			2007
	Sept.–Dec.	Jan.–Apr.	May–Aug.	Sept.–Dec.	…	Jan.–Apr.	May–Aug.	Sept.–Dec.	Jan.–Apr.	May–Aug.	Sept.–Dec.
Set 1	^*	^*	^*	^*	…	^*	^*	^*		^*	^*
Set 2	^*	^*	^*	^*	…	^*	^*	^*	^*		^*
Set 3	^*	^*	^*	^*	…	^*	^*	^*	^*	^*

Data from four successive months were cyclically used as the test sets.

Dates marked by asterisks were used for network training.

The data were preprocessed to eliminate instrumental errors by means of replacing holes in the established time series with the value before or after a hole occurred. In addition, each value in the neural network was normalized within the range [0, 1] using the following linear transformation: \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6}\begin{document} \begin{align*} X^{\prime} = (X - V_{\min}) / (V_{\max} - V_{\min}) \tag{4} \end{align*} \end{document}

where X′ is the new normalized value, X the old value, V_max the maximum of the dataset under consideration, and V_min the minimum of the dataset under consideration. The set of normalized values was used as the neural network input.

Evaluation and Discussion

Model comparative evaluation

Trials were carried out with and without the dust storm input to facilitate training and optimization and evaluate the forecasting task for the daily PM₁₀ API.

Accordingly, taking into account the previous description, the training set is comprised of a value of 72 months, whereas the test set is comprised of a value of 4 months. Training epochs were set at 500 for each neural network model.

The aim of the experimental trials was to establish optimized architecture for each model. Model performance was evaluated using the following parameters: MAE, linear correlation coefficient (r), and root mean square error (RMSE).

The optimized topology of the Elman neural network incorporating the dust storm data input consists of 15 input neurons, 25 hidden neurons, and 1 output neuron. Without the dust storm data input, the Elman neural network consists of 13 input neurons, 21 hidden neurons, and 1 output neuron.

Prediction of high PM₁₀ API events caused by dust storm activity

The evaluation of model performance was extended to include the prediction of high PM₁₀ API events caused by dust storm activity in northern China. This task is of particular importance for authorities because successful and timely predictions of high PM₁₀ concentrations can be useful in announcing activity restrictions for the protection of public health when necessary.

The prediction accuracy rate of high PM₁₀ API events caused by dust storm (AHAd) activity was introduced in this study to evaluate the two models: \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6}\begin{document} \begin{align*} {\rm AHAd} = \frac {n_{100}} {N_{100}} {\times} 100\% \tag {5} \end{align*} \end{document}

where AHAd is the accuracy rate of high PM₁₀ API events caused by dust storm activity; n₁₀₀ is the total predicted number of records from six stations in which the PM₁₀ API value exceeded 100 (caused by dust storm activity), and N₁₀₀ (the value is 54 in this study) is the total number of records from six stations (each station has nine records) in which the PM₁₀ API value exceeded 100 (caused by dust storm activity). Figure 3 displays the comparison between the improved Elman and the standard Elman while performing forecasting task on high PM₁₀ API caused by dust storm.

FIG. 3.

Prediction of high PM₁₀ API caused by dust storm by applying the Elman neural network and improved Elman neural network. PM₁₀, particulate matter with a diameter <10 μm.

When taking into account Table 4 and Fig. 4 that compare the predicted and observed values, it can be found that the values of the coefficient of determination (r²) are higher for the improved model as well as AHAd is higher for the improved model (83.33) than for the standard Elman model (64.81). The improved model also outperformed the Elman model in other parameters.

FIG. 4.

Scatter plots of the predicted (ordinate) versus the observed (abscissa) concentrations for both models.

Table 4.

Static Index Applied to Both the Elman Neural Network and Improved Elman Neural Network for High Air Pollution Index Event Predictions

	R	MAE	RMSE	n₁₀₀	AHAd (%)
ELMAN with dust data input	0.79	18.76	22.46	45	83.33
ELMAN without dust data input	0.47	29.60	36.87	35	64.81

AHAd is the accuracy rate of high particulate matter with a diameter <10 μm API events caused by dust storm activity.

MAE, mean absolute error; RMSE, root mean square error.

Conclusion

The aim of this research was to develop a predictive model to forecast peak values of PM₁₀ air pollutants in the urban area of Wuhan, China. The authors have improved upon a predictive model by using an Elman neural network. Each input pattern is composed by 15 daily values (described in the Experimental Setup section). Time series data were recorded from September 1, 2001, to December 31, 2007, within the urban area of Wuhan at six separate stations. Prediction tasks are related to daily PM₁₀ API forecasting. Three statistical indicators (r, MAE, and RMSE) were utilized to estimate the acquired results. Although these acquired experimental results show small differences in MAE and RMSE between the Elman neural network and improved Elman neural network, the coefficient of determination (r² = 0.62) values for the improved model was higher than the standard model (r² = 0.22) when performing forecasting task on high PM₁₀ API caused by dust storm activity in northern China. The improved Elman model also outperformed the standard model in the accuracy rate of high PM₁₀ API events caused by dust storm activity. The improved model's principal benefit is its potential as a tool to predict PM₁₀ API parameters inside cities throughout China, making the proposed forecaster a powerful tool in support of systems designed for high PM₁₀ pollution management.

Footnotes

Acknowledgments

This study was jointly supported by the Wuhan Science and Technology Bureau of Wuhan TSP and UHI program (No. 200860423209) and the Institute of Geodesy and Geophysics, Chinese Academy of Sciences. PM10 data used for this study were provided by the Wuhan Environmental Protection Bureau. The meteorological data were provided by the Wuhan Meteorological Station. The dust storm data were provided by the CMDSSS (). Key science and technology program supported by Wuhan Science and Technology Bureau (No. 200860423209). The authors express their appreciation to three anonymous reviewers for their constructive and insightful comments.

Author Disclosure Statement

No competing financial interests exist.

References

Boznar

, Lesjak

, Mlakar

1991. A neural network-based method for short-term predictions of ambient SO₂ concentrations in highly polluted industrial areas of complex terrain. Atmos. Environ., 27B:221.

Brunelli

, Piazza

, Pignato

, Sorbello

, Vitabile

2007. Two-days ahead prediction of daily maximum concentrations of SO₂, O₃, PM₁₀, NO₂, CO in the urban area of Palermo, Italy. Atmos. Environ., 41:2967.

Chaloulakou

, Grivas

, Spyrellis

2003. Neural network and multiple regression models for PM₁₀ prediction in Athens. A comparative assessment. J. Air Waste Manage. Assoc., 53:1183.

Corani

2005. Air quality prediction in Milan: feed-forward neural networks, pruned neural networks and lazy learning. Ecol Model., 185:513.

Coulibaly

, Anctil

, Aravena

, Bobé

2001. Artificial neural network modeling of water table depth fluctuations. Water Resour. Res., 37:885.

Duce

R.A.

, Unni

C.K.

, Ray

B.J.

, Prospero

J.M.

, Merrill

J.T.

1980. Long-range atmospheric transport of soil dust storm from Asia to the tropical North Pacific: Temporal variability. Science, 209:1522.

Elman

J.L.

1990. Finding structure in time. Cogn. Sci., 14:179.

Feng

, Wu

S.J.

, Du

, Li

X.D.

, Ling

, Xue

H.P.

, Cai

S.M.

2010. Variations of PM₁₀ concentrations in Wuhan, China. Environ. Monit. Assess.[Epub ahead of print]10.1007/s10661-010–1581-6.

Hubbard

M.C.

, Cobourn

W.G.

1998. Development of a regression model to forecast ground-level ozone concentrations, in Louisville, KY. Atmos. Environ., 32:2637.

10.

Jiang

D.H.

, Zhang

, Hu

, Zeng

, Tan

J.G.

, Shao

D.M.

2004. Progress in developing an ANN model for air pollution index forecast. Atmos. Environ., 38:7055.

11.

Kolehmainen

, Martikainen

, Ruuskanen

2001. Neural networks and periodic components used in air quality forecasting. Atmos. Environ., 35:815.

12.

Kukkonen

, Partanen

, Karppinen

, Ruuskanen

, Junninen

, Kolehmainen

, Niska

, Hooyberghs

, Mensink

, Dumont

, Fierens

, Brasseur

2005. A neural network forecast for daily average PM₁₀ concentrations in Belgium. Atmos. Environ., 39:3279.

13.

Lang

, Su

X.H.

, Zhou

X.J.

2004. Prediction for air pollution index based on organic grey neural network model. J. Harbin Inst. Technol., 36:1598(in Chinese).

14.

, Wei

H.M.

, Wei

, Jiang

M.L.

, Wang

F.F.

2007. Daily Change Characteristics of PM10 in Wuhan. Environ. Sci. Technol., 30:39(in Chinese)

15.

, Wei

W.H.

, Wei

, Wang

Z.C.

, Yang

2004. Air pollution-meteorological condition relationship in Wuhan City. J Hubei Meteorol., 23:18(in Chinese)

16.

Liu

Z.Z.

, Pan

2008. Study on air pollution index forecasting in Xiamen city—comparison between time series analysis and neural network. Mind Comput., 2:33.

17.

W.Z.

, Wang

W.J.

, Fan

H.Y.

, Leung

A.Y.T.

, Xu

Z.B.

, Lo

S.M.

, Wong

J.C.K.

2002. Prediction of pollutant levels in Causeway Bay Area of Hong Kong using an improved neural network model. J. Environ. Eng., 128:1146.

18.

Ordieres

J.B.

, Vergara

E.P.

, Capuz

R.S.

, Salazar

R.E.

2005. Neural network prediction model for fine particulate matter (PM2.5) on the US–exico border in El Paso (Texas) and Ciudad Juarez (Chihuahua) Environ. Model Softw., 20:547.

19.

Perez

, Reyes

2002. Prediction of maximum of 24-h average of PM₁₀ concentrations 30 h in advance in Santiago, Chile. Atmos. Environ., 36:4555.

20.

Riedmiller

, Braun

1993. Direct adaptive method for faster BackPROPagation learning: The RPROP algorithm. Proceedings of the IEEE International Conference on Neural Networks 1993 (ICNN 93) IEEE: San Francisco, CA, 586.

21.

Sun

J.M.

, Zhang

M.Y.

, Liu

T.S.

2001. Spatial and temporal characteristics of dust storms in China and its surrounding regions, 1960–1999: relations to source area and climate. J. Geophys. Res., 106:10325.

22.

Tong

Y.C.

2006. Air quality forecast and development in major cities of China. Environ. Monit. China, 22:69(in Chinese).

23.

Wang

S.G.

, Wang

J.Y.

, Zhou

Z.J.

, Shang

K.Z.

, Yang

D.B.

, Zhao

Z.S.

2003. Regional characteristics of dust events in China. J. Geogr. Sci., 13:35.

Artificial Neural Network Models for Daily PM 10 Air Pollution Index Prediction in the Urban Area of Wuhan,China

Abstract

Abstract

Introduction

Study Area and Data

Study area

Data

Methodology

Experimental Setup

Evaluation and Discussion

Model comparative evaluation

Prediction of high PM10 API events caused by dust storm activity

Conclusion

Footnotes

Acknowledgments

Author Disclosure Statement

References

Prediction of high PM₁₀ API events caused by dust storm activity