Neural Network and Multiple Regression Models for PM10 Prediction in Athens: A Comparative Assessment

Abstract

Particulate atmospheric pollution in urban areas is considered to have significant impact on human health. Therefore, the ability to make accurate predictions of particulate ambient concentrations is important to improve public awareness and air quality management. This study examines the possibility of using neural network methods as tools for daily average particulate matter with aerodynamic diameter <10 µm (PM₁₀) concentration forecasting, providing an alternative to statistical models widely used up to this day. Based on a data inventory, in a fixed central site in Athens, Greece, ranging over a two-year period, and using mainly meteorological variables as inputs, neural network models and multiple linear regression models were developed and evaluated. Comparison statistics used indicate that the neural network approach has an edge over regression models, expressed both in terms of prediction error (root mean square error values lower by 8.2–9.4%) and of episodic prediction ability (false alarm rate values lower by 7–13%). The results demonstrate that artificial neural networks (ANNs), if properly trained and formed, can provide adequate solutions to particulate pollution prognostic demands.     

Impacts of Meteorological Factors on PM10: Artificial Neural Networks (ANN) and Multiple Linear Regression (MLR) Approaches

In this study, prediction capacities of multi-linear regression (MLR) and artificial neural networks (ANN) onto coarse particulate matter (PM₁₀) concentrations were investigated. Different meteorological factors on particulate pollution were chosen for operating variables in the model analyses. Two different regions (urban and industrial) were identified in the region of Kocaeli, Turkey. All data sets were obtained from air quality monitoring network of the Ministry of Environment and Urban Planning, and 120 data sets were used in the MLR and ANN models. Regression equations explained the effects of the meteorological factors in MLR analyses. In the ANN model, backpropagation network with two hidden layers has achieved the best prediction efficiency. Determination coefficients and error values were examined for each model. ANN models displayed more accurate results compared to MLR.     

Model for forecasting expressway fine particulate matter and carbon monoxide concentration: application of regression and neural network models

The Borman Expressway is a heavily traveled 16-mi segment of the Interstate 80/94 freeway through Northwestern Indiana. The Lake and Porter counties through which this expressway passes are designated as particulate matter < 2.5 microm (PM2.5) and ozone 8-hr standard nonattainment areas. The Purdue University air quality group has been collecting PM2.5, carbon monoxide (CO), wind speed, wind direction, pressure, and temperature data since September 1999. In this work, regression and neural network models were developed for forecasting hourly PM2.5 and CO concentrations. Time series of PM2.5 and CO concentrations, traffic data, and meteorological parameters were used for developing the neural network and regression models. The models were compared using a number of statistical quality indicators. Both models had reasonable accuracy in predicting hourly PM2.5 concentration with coefficient of determination -0.80, root mean square error (RMSE) <4 microg/m3, and index of agreement (IA) > 0.90. For CO prediction, both models showed moderate forecasting performance with a coefficient of determination -0.55, RMSE < 0.50 ppm, and IA -0.85. These models are computationally less cumbersome and require less number of predictors as compared with the deterministic models. The availability of real time PM2.5 and CO forecasts will help highway managers to identify air pollution episodic events beforehand and to determine mitigation strategies.     

Two-days ahead prediction of daily maximum concentrations of SO2, O3, PM10, NO2, CO in the urban area of Palermo,Italy

Artificial neural networks are functional alternative techniques in modelling the intricate vehicular exhaust emission dispersion phenomenon. Pollutant predictions are notoriously complex when using either deterministic or stochastic models, which explains why this model was developed using a neural network. Neural networks have the ability to learn about non-linear relationships between the used variables. In this paper a recurrent neural network (Elman model) based forecaster for the prediction of daily maximum concentrations of SO₂, O₃, PM10, NO₂, CO in the city of Palermo is proposed. The effectiveness of the presented forecaster was tested using a time series recorded between 1 January 2003 to 31 December 2004 in eight monitoring stations in urban area of Palermo (Italy). Experimental trials show that the developed and tuned model is appropriate, giving small values of root mean square error (RMSE) , mean absolute error (MAE) and mean square error (MSE). In addition, the related correlation coefficient ranges from 0.72 to 0.97 for each forecasted pollutant, underlying a small difference between the forecasted and the measured values. The above results make the proposed forecaster a powerful tool for pollution management systems.     

Evaluation of artificial neural networks for fine particulate pollution (PM10 and PM2.5) forecasting

Multi-layer perceptron (MLP) artificial neural network (ANN) models are compared with traditional multiple regression (MLR) models for daily maximum and average O3 and particulate matter (PM10 and PM2.5) forecasting. MLP particulate forecasting models show little if any improvement over MLR models and exhibit less skill than do O3 forecasting models. Meteorological variables (precipitation, wind, and temperature), persistence, and co-pollutant data are shown to be useful PM predictors. If MLP approaches are adopted for PM forecasting, training methods that improve extreme value prediction are recommended.     

空气中PM_(10)浓度的BP神经网络预报研究

建立了某市PM10浓度预报的分段BP神经网络模型,经验证,所建立的BP预报模型,预测精度比较高,PM10日平均浓度误差大多在-0．010～0．010mg／m^3范围内,相对误差在-20％～20％,表明BP神经网络对PM10的浓度预报是一种有效的工具。     

A Pilot Indoor-Outdoor Study of Organic Particulate Matter and Particulate Mutagenicity

Abstract

Many large metropolitan areas experience elevated concentrations of ground-level ozone pollution during the summertime “smog season”. Local environmental or health agencies often need to make daily air pollution forecasts for public advisories and for input into decisions regarding abatement measures and air quality management. Such forecasts are usually based on statistical relationships between weather conditions and ambient air pollution concentrations. Multivariate linear regression models have been widely used for this purpose, and well-specified regressions can provide reasonable results. However, pollution-weather relationships are typically complex and nonlinear—especially for ozone—properties that might be better captured by neural networks. This study investigates the potential for using neural networks to forecast ozone pollution, as compared to traditional regression models. Multiple regression models and neural networks are examined for a range of cities under different climate and ozone regimes, enabling a comparative study of the two approaches. Model comparison statistics indicate that neural network techniques are somewhat (but not dramatically) better than regression models for daily ozone prediction, and that all types of models are sensitive to different weather-ozone regimes and the role of persistence in aiding predictions.     

QSAR modeling with the electrotopological state indices: predicting the toxicity of organic chemicals

A quantitative structure-activity relationship model, based on the atom-type electrotopological state (E-state) indices, for the prediction of toxicity to fathead minnow for a diverse set of 140 organic chemicals is presented. Multiple linear regression and artificial neural network techniques were employed in the modeling of experimental toxicity (-logLC(50)) values ranging from 0.85 to 6.09. For the training set of 130 organic compounds a linear regression model with r(2)=0.84 and s=0.36 was obtained with 14 atom-type E-state indices. For the test set of 10 compounds, the corresponding statistics were r(2)=0.83 and s=0.47, respectively. Neural networks gave a significant improvement using the same set of parameters, and the standard deviations were s=0.31 for the training set and s=0.30 for the test set when an artificial neural network with five neurons in the hidden layer was used. The results clearly show that accurate models can be rapidly calculated for the prediction of toxicity for a diverse set of organic chemicals using easily calculated parameters.     

Prediction of dissolved oxygen concentration in hypoxic river systems using support vector machine: a case study of Wen-Rui Tang River,China

Accurate quantification of dissolved oxygen (DO) is critically important for managing water resources and controlling pollution. Artificial intelligence (AI) models have been successfully applied for modeling DO content in aquatic ecosystems with limited data. However, the efficacy of these AI models in predicting DO levels in the hypoxic river systems having multiple pollution sources and complicated pollutants behaviors is unclear. Given this dilemma, we developed a promising AI model, known as support vector machine (SVM), to predict the DO concentration in a hypoxic river in southeastern China. Four different calibration models, specifically, multiple linear regression, back propagation neural network, general regression neural network, and SVM, were established, and their prediction accuracy was systemically investigated and compared. A total of 11 hydro-chemical variables were used as model inputs. These variables were measured bimonthly at eight sampling sites along the rural-suburban-urban portion of Wen-Rui Tang River from 2004 to 2008. The performances of the established models were assessed through the mean square error (MSE), determination coefficient (R ²), and Nash-Sutcliffe (NS) model efficiency. The results indicated that the SVM model was superior to other models in predicting DO concentration in Wen-Rui Tang River. For SVM, the MSE, R ², and NS values for the testing subset were 0.9416 mg/L, 0.8646, and 0.8763, respectively. Sensitivity analysis showed that ammonium-nitrogen was the most significant input variable of the proposal SVM model. Overall, these results demonstrated that the proposed SVM model can efficiently predict water quality, especially for highly impaired and hypoxic river systems.     

Deep learning architecture for air quality predictions

With the rapid development of urbanization and industrialization, many developing countries are suffering from heavy air pollution. Governments and citizens have expressed increasing concern regarding air pollution because it affects human health and sustainable development worldwide. Current air quality prediction methods mainly use shallow models; however, these methods produce unsatisfactory results, which inspired us to investigate methods of predicting air quality based on deep architecture models. In this paper, a novel spatiotemporal deep learning (STDL)-based air quality prediction method that inherently considers spatial and temporal correlations is proposed. A stacked autoencoder (SAE) model is used to extract inherent air quality features, and it is trained in a greedy layer-wise manner. Compared with traditional time series prediction models, our model can predict the air quality of all stations simultaneously and shows the temporal stability in all seasons. Moreover, a comparison with the spatiotemporal artificial neural network (STANN), auto regression moving average (ARMA), and support vector regression (SVR) models demonstrates that the proposed method of performing air quality predictions has a superior performance.     

Modelling of dissolved oxygen content using artificial neural networks: Danube River, North Serbia, case study

The aims of this study are to create an artificial neural network (ANN) model using non-specific water quality parameters and to examine the accuracy of three different ANN architectures: General Regression Neural Network (GRNN), Backpropagation Neural Network (BPNN) and Recurrent Neural Network (RNN), for prediction of dissolved oxygen (DO) concentration in the Danube River. The neural network model has been developed using measured data collected from the Bezdan monitoring station on the Danube River. The input variables used for the ANN model are water flow, temperature, pH and electrical conductivity. The model was trained and validated using available data from 2004 to 2008 and tested using the data from 2009. The order of performance for the created architectures based on their comparison with the test data is RNN > GRNN > BPNN. The ANN results are compared with multiple linear regression (MLR) model using multiple statistical indicators. The comparison of the RNN model with the MLR model indicates that the RNN model performs much better, since all predictions of the RNN model for the test data were within the error of less than ±10 %. In case of the MLR, only 55 % of predictions were within the error of less than ±10 %. The developed RNN model can be used as a tool for the prediction of DO in river waters.     

Forecasting summertime surface-level ozone concentrations in the Lower Fraser Valley of British Columbia: an ensemble neural network approach

Empirical models for predicting daily maximum hourly average ozone concentrations were developed for 10 monitoring stations in the Lower Fraser Valley (LFV) of British Columbia. According to data from 1991 to 1996, ensemble neural network models increased explained variance an average of 7% over multiple linear regression models using the same input variables. Without modification, all models performed poorly on days when the observed peak ozone concentration exceeded 82 parts per billion, the National Ambient Air Quality Objective. When numbers of extreme events in training data were increased using a histogram equalization process, models were able to forecast exceedances with improved accuracy. Modified generalized additive model (GAM) plots and associated measures of input variable importance and interaction were generated for a subset of the trained models and used to investigate relationships between input variables and ozone levels. The neural network models displayed a high degree of interaction among inputs, and it is likely the ability of these model types to account for interactions, rather than the nonlinearity of individual input variables, that explains their improved forecast skill. Inspection of GAM-style plots indicated that the relative importance of input variables in the ensemble neural network models varied with geographic location within the LFV. Four distinct groups of stations were identified, and rankings of inputs within the groups were generally consistent with physical intuition and results of prior studies.     

Prediction of sulfur dioxide concentrations at a site near downtown Santiago,Chile

We have analyzed the possibility to predict hourly averages of sulfur dioxide concentrations in the atmosphere at a site not far from the downtown area in the city of Santiago, Chile. We have compared the forecasts produced assuming persistence, linear regressions and feed forward neural networks. The effect of meteorological conditions is included by using forecasted values of temperature, relative humidity and wind speed at the time of the intended prediction as inputs to the different models. The best predictions for hourly averages are obtained with a three-layer neural network that has hourly averages of sulfur dioxide concentrations every 6 h on the previous day plus the actual values of the meteorological variables as input. Training the network with 1995 data, error in 8 h in advance prediction for 1996 data is of the order of 30%.     

Comparing universal kriging and land-use regression for predicting concentrations of gaseous oxides of nitrogen (NOx) for the Multi-Ethnic Study of Atherosclerosis and Air Pollution (MESA Air)

BACKGROUND: Epidemiological studies that assess the health effects of long-term exposure to ambient air pollution are used to inform public policy. These studies rely on exposure models that use data collected from pollution monitoring sites to predict exposures at subject locations. Land use regression (LUR) and universal kriging (UK) have been suggested as potential prediction methods. We evaluate these approaches on a dataset including measurements from three seasons in Los Angeles, CA. METHODS: The measurements of gaseous oxides of nitrogen (NOx) used in this study are from a "snapshot" sampling campaign that is part of the Multi-Ethnic Study of Atherosclerosis and Air Pollution (MESA Air). The measurements in Los Angeles were collected during three two-week periods in the summer, autumn, and winter, each with about 150 sites. The design included clusters of monitors on either side of busy roads to capture near-field gradients of traffic-related pollution. LUR and UK prediction models were created using geographic information system (GIS)-based covariates. Selection of covariates was based on 10-fold cross-validated (CV) R(2) and root mean square error (RMSE). Since UK requires specialized software, a computationally simpler two-step procedure was also employed to approximate fitting the UK model using readily available regression and GIS software. RESULTS: UK models consistently performed as well as or better than the analogous LUR models. The best CV R(2) values for season-specific UK models predicting log(NOx) were 0.75, 0.72, and 0.74 (CV RMSE 0.20, 0.17, and 0.15) for summer, autumn, and winter, respectively. The best CV R(2) values for season-specific LUR models predicting log(NOx) were 0.74, 0.60, and 0.67 (CV RMSE 0.20, 0.20, and 0.17). The two-stage approximation to UK also performed better than LUR and nearly as well as the full UK model with CV R(2) values 0.75, 0.70, and 0.70 (CV RMSE 0.20, 0.17, and 0.17) for summer, autumn, and winter, respectively. CONCLUSION: High quality LUR and UK prediction models for NOx in Los Angeles were developed for the three seasons based on data collected for MESA Air. In our study, UK consistently outperformed LUR. Similarly, the 2-step approach was more effective than the LUR models, with performance equal to or slightly worse than UK.     

Prediction of hourly O3 concentrations using support vector regression algorithms

In this paper we present an application of the Support Vector Regression algorithm (SVMr) to the prediction of hourly ozone values in Madrid urban area. In order to improve the training capacity of SVMrs, we have used a recently proposed approach, based on reductions of the SVMr hyper-parameters search space. Using the modified SVMr, we study different influences which may modify the ozone prediction, such as previous ozone measurements in a given station, measurements in neighbors stations, and the influence of meteorologic variables. We use statistical tests to verify the significance of incorporating different variables into the SVMr. A comparison with the results obtained using a neural network (multi-layer perceptron) is also carried out. This study has been carried out in 5 different stations of the air pollution monitoring network of Madrid, so the conclusions raised are backed by real data. The final result of the work is a robust and powerful software for tropospheric ozone prediction in Madrid. Also, the prediction tool based on SVMr is flexible enough to incorporate any other prediction variable, such as city models, or traffic patters, which may improve the prediction obtained with the SVMr.     

An investigation of distributed lag models in the context of air pollution and mortality time series analysis

In particulate air pollution mortality time series studies, the particulate air pollution exposure measure used is typically the current day's or the previous day's air pollution concentration or a multi-day moving average air pollution concentration. Distributed lag models (DLMs) that allow for differential air pollution effects that are spread over multiple days are seen as an improvement over using a single- or multi-day moving average air pollution exposure measure. However, at the current time, the statistical properties of DLMs as a measure of air pollution exposure have not been investigated. In this paper, a simulation study is used to investigate the performance of DLMs as a measure of air pollution exposure in comparison with single- and multi-day moving average air pollution exposure measures under various forms for the true effect of air pollution on mortality. The simulation study shows that DLMs offer a more robust measure of the effect of air pollution on mortality and avoid the potential for a large negative bias compared with single- or multi-day moving average air pollution exposure measures. This is important information. In many U.S. cities, particulate air pollution concentrations are observed only once every six days, meaning it is often only possible to use single-day particulate air pollution exposure measures. The results from this paper will help quantify the magnitude of the negative bias that can result from using single-day exposure measures. The implications of this work for future air pollution mortality time series studies are discussed. The data used in this paper are concurrent daily time series of mortality, weather, and particulate air pollution from Cook County, IL, for the period 1987-1994.     

Forecasting air quality time series using deep learning

This paper presents one of the first applications of deep learning (DL) techniques to predict air pollution time series. Air quality management relies extensively on time series data captured at air monitoring stations as the basis of identifying population exposure to airborne pollutants and determining compliance with local ambient air standards. In this paper, 8 hr averaged surface ozone (O₃) concentrations were predicted using deep learning consisting of a recurrent neural network (RNN) with long short-term memory (LSTM). Hourly air quality and meteorological data were used to train and forecast values up to 72 hours with low error rates. The LSTM was able to forecast the duration of continuous O₃ exceedances as well. Prior to training the network, the dataset was reviewed for missing data and outliers. Missing data were imputed using a novel technique that averaged gaps less than eight time steps with incremental steps based on first-order differences of neighboring time periods. Data were then used to train decision trees to evaluate input feature importance over different time prediction horizons. The number of features used to train the LSTM model was reduced from 25 features to 5 features, resulting in improved accuracy as measured by Mean Absolute Error (MAE). Parameter sensitivity analysis identified look-back nodes associated with the RNN proved to be a significant source of error if not aligned with the prediction horizon. Overall, MAE's less than 2 were calculated for predictions out to 72 hours.

Implications: Novel deep learning techniques were used to train an 8-hour averaged ozone forecast model. Missing data and outliers within the captured data set were replaced using a new imputation method that generated calculated values closer to the expected value based on the time and season. Decision trees were used to identify input variables with the greatest importance. The methods presented in this paper allow air managers to forecast long range air pollution concentration while only monitoring key parameters and without transforming the data set in its entirety, thus allowing real time inputs and continuous prediction.     

Receptor modeling application framework for particle source apportionment

Receptor models infer contributions from particulate matter (PM) source types using multivariate measurements of particle chemical and physical properties. Receptor models complement source models that estimate concentrations from emissions inventories and transport meteorology. Enrichment factor, chemical mass balance, multiple linear regression, eigenvector. edge detection, neural network, aerosol evolution, and aerosol equilibrium models have all been used to solve particulate air quality problems, and more than 500 citations of their theory and application document these uses. While elements, ions, and carbons were often used to apportion TSP, PM10, and PM2.5 among many source types, many of these components have been reduced in source emissions such that more complex measurements of carbon fractions, specific organic compounds, single particle characteristics, and isotopic abundances now need to be measured in source and receptor samples. Compliance monitoring networks are not usually designed to obtain data for the observables, locations, and time periods that allow receptor models to be applied. Measurements from existing networks can be used to form conceptual models that allow the needed monitoring network to be optimized. The framework for using receptor models to solve air quality problems consists of: (1) formulating a conceptual model; (2) identifying potential sources; (3) characterizing source emissions; (4) obtaining and analyzing ambient PM samples for major components and source markers; (5) confirming source types with multivariate receptor models; (6) quantifying source contributions with the chemical mass balance; (7) estimating profile changes and the limiting precursor gases for secondary aerosols; and (8) reconciling receptor modeling results with source models, emissions inventories, and receptor data analyses.     

Prediction of ozone concentrations in Oporto city with statistical approaches

The performance of three statistical methods: time-series, multiple linear regression and feedforward artificial neural networks models were compared to predict the daily mean ozone concentrations. The study here reported was based on data from one urban site with traffic influences and one rural background site. The studies were performed for the year 2002 and the respective four trimesters separately. In the multiple linear regression and feedforward artificial neural network models, the concentrations of ozone, the concentrations of its precursors (nitrogen oxides) and some meteorological variables for one and two days before the prediction day were used as predictors. For the application of these models in the validation step, the inputs of ozone concentration for one and two days before were replaced by the ozone concentrations predicted by the models. The results showed that time-series modelling was not profitable. In the development step, similar performances were obtained with multiple linear regression and feedforward artificial neural network. Better performance indexes were achieved with feedforward artificial neural network models in validation step. Concluding, feedforward artificial neural network models were more efficient to predict ozone concentrations.