Abstract: Particulate matter has a significant impact on human health when the concentration levels of this substance exceed Malaysia Ambient Air Quality Guidelines (MAAQG). This research focused only on particulate matter with an aerodynamic diameter less than 10 μm, namely PM10. Statistical modeling is required to predict future PM10 concentration. The aim of this study is to develop and predict next day, next two-day and next three-day PM10 concentration in a sub-urban area (Seberang Jaya) of Malaysia. This study used daily average monitoring records from 2001 to 2010. Two main models for predicting PM10 concentration were used: feedforward backpropagation and general regression neural network models. The models for the artificial neural network show that feedforward backpropagation is better than the general regression neural network with fewer errors; as much as 5.6% for next day, 3.5% for next two-day and 2.5% for next three-day predictions. These models will help local authorities to take an appropriate course of action to reduce PM10 concentration and could also be used as an early warning system.
INTRODUCTION
The Department of Environment (DOE., 2006) is the government body responsible for monitoring air quality in Malaysia. Department of Environment Malaysia monitors air quality continuously via 52 stations located in urban, sub-urban and industrial areas. These monitoring stations are located in strategic locations to detect any significant change of air quality. The Malaysia Ambient Air Quality Guidelines (MAAQG, DOE., 2011) set out the allowable levels of various air pollutants. The MAAQG set the safe threshold value for PM10 at 150 μg m-3 for a 24 h averaging period and 50 μg m-3 per year.
Both acute and chronic human health impacts may occur when the concentration levels of air pollutants exceed the air quality guidelines (QUARG., 1996; Lee, 2010). Nasir et al. (1998) reported in 1997 (haze episode in Malaysia) the estimated negative effect to health for asthma attacks was 285,277. There were 118,804 cases of bronchitis in children and 3889 cases in adults, in addition to 2003 respiratory-related hospital admissions, 26,864 emergency room visits and 5,000,760 limited activity days.
Thus, particulate matter (PM10) has become a challenge in managing Malaysia’s air quality. Developing PM10 forecasting models is of key importance in PM10 monitoring. Artificial neural network (ANN) models can be used for predicting future PM10 concentrations in urban and sub-urban areas in Malaysia.
Neural network modeling has become popular as a tool for prediction modeling. This study used two types of neural networks, multi-layer perception (Feedforward Backpropagation (FFBP) and radial basis functions (General Regression Neural Network (GRNN). The FFBP model for predicting future PM10 concentrations has not yet been developed in Malaysia. GRNN studies for prediction of PM10 concentration are very few and have also not yet been developed for Malaysia. This research would bring great impact to the air quality studies in Malaysia.
MATERIALS AND METHODS
Site description: Seberang Jaya is located in North-West coast of Peninsular Malaysia. This monitoring station is located in the residential area of Sek Keb Seberang Jaya 2. Vehicular traffic is the main source of emission in Seberang Jaya. Power plants and industrial activities in a 10 km radius of the site can also be considered as possible sources of PM10.
Parameter selection: Seven daily-averaged parameters from the ten years of the monitoring record base (2001-2010) were used in the study in order to gain a better understanding of PM10 variability. All the monitoring records provided by Department of Environment Malaysia followed the quality assurance procedure as recommended by Department of Environment Malaysia. The parameters that were used in this research are particulate matter with an aerodynamic diameter less than 10 μm (PM10, μg m-3), nitrogen dioxide (NO2, ppm), wind speed (WS, km h-1), sulphur dioxide (SO2, ppm), Relative Humidity (RH %), Carbon Monoxide (CO, ppm) and ambient temperature (T, °C). Parameters influencing future PM10 concentrations used by previous researchers are shown in Table 1.
Feedforward backpropagation: Feedforward backpropagation (FFBP) neural networks have been selected because the model minimizes the error between target and obtained values and has in common the learning algorithms of neural networks (Sun et al., 2008). Figure 1 shows the procedure for the development of FFBP models.
| Table 1: | Summary of parameter selection in previous research |
| DoW: Day of the week and ΔT: Tmax-Tmin | |
| Fig. 1: | Procedure for development of FFBP models |
In the first stage, development of the model using FFBP is based on 80% of observations, where future PM10 is considered as a dependent variable and WS, RH, T, SO2, NO2, PM10 and CO as independent variables. Defining neural network architecture is a very important step to enable FFBP models to achieve the optimum result. In this research three layers of neural network were used, i.e. an input layer, a hidden layer and output layer (Fig. 2) (Singh and Deo, 2007).
There are two main phases in FFBP, forward pass phase and backward pass phase. First, in forward pass phase, the training monitoring record set is propagated through the hidden layer and comes out of the neural network through the output layer. The output values are then compared to actual target output values. The residual (error) between output layer and actual values is calculated and propagated back towards the hidden layer.
| Fig. 2: | Architecture of a feedforward backpropagation neural network (FFBP) (Singh and Deo, 2007) |
In this phase, the modifications to the connection strengths are made based on the differences between computed and observed information monitoring records at the output units (Viotti et al., 2002; Cigizoglu and Kisi, 2006). Adjustment to the weighting was made, so that error was reduced with each alteration and the FFBP models approaching the desired outputs (Sun et al., 2008). The advantage of this method lies in the ability of the model to adjust itself during the training phase, by comparing the actual and obtained values with low error rate.
In stage two, initializing weight and stopping criteria is also considered. Initializing weight and bias are normally selected randomly; this research used small random values. Maximum error used as stopping criteria is set at 0.05. In this study the training process was set at maximum 10,000 iterations, or until it reached the maximum error.
Levenberg-Marquardt optimization was used as a network training function to update weight and bias values. This method was chosen because this is the fastest algorithm in the Matlab toolbox and is recommended for supervised learning (MATLAB., 2012).
Stage three is an important in FFBP because the number of neurons in the hidden layer has strong influence on the output. The optimum number of neurons is very important, because too few neurons will contribute to under-fitting while too many neurons will lead to over-fitting. The number of neurons is also related to the complexity of mapping between input and output, compared to the amount of noise and the amount of training monitoring record.
A variety of techniques and methods have been suggested by researchers to determine the number of hidden nodes. A combination of techniques suggested by Heaton (2008) and Charytoniuk and Chen (2000) were used in this study to determine the number of nodes in the hidden layer. After running a trial and error method with a maximum twice input layer, ranking and comparison of all performance indicators in stage five, the best FFBP based on the number of hidden nodes was obtained in stage six. In stage six, tests for overfitting are required using cross validation, as recommended by Sterlin (2007), for the purpose of preventing overfitting in FFBP models. Then, the monitoring records were randomly divided into five folders. Figure 3 illustrates the cross validation technique. This study used 20% for each folder. All five combinations were run using 80% for training and 20% for testing, then the error was calculated for all of the sets. If the values error for training were less than the values error for the test set, this indicated that no overfitting occurred.
Stage seven focused on determining the best activation function for the FFBP model. Activation functions are important because they have a significant impact on the applicability of training algorithms. Kamruzzaman and Aziz (2002) suggested the use of a sigmoid activation function because sigmoid units are easier to train than other activation functions.
| Fig. 3: | Illustration of cross validation technique (Sterlin, 2007) |
| Fig. 4(a-c): | Sigmoid transfer function (MATLAB., 2012) |
Besides that, a small change in the weights will usually produce a change in outputs, making it possible to tell whether that change in the weights is accurate or not. All the sigmoid transfer functions that were used are shown in Fig. 4. They are the tangent sigmoid (Fig. 4a), log sigmoid (Fig. 4b) and linear transfer functions (Fig. 4c).
The FFBP model was obtained and the model was validated based on the performance indicator value using 20% complete observed monitoring records. All the procedures were repeated until all combinations of sigmoid transfer function were completed. Then, all performance indicators were ranked to obtain the best FFBP.
General regression neural network: A general regression neural network is one of the methods for curve fits. This method tries to adjust the values of the function coefficient such that the function closely approximates the available monitoring record. FFBP is based on equations that are connected using weighting factors but GRNN uses statistical distribution to select the equation within the structure. Figure 5 shows the procedure for development of GRNN models.
Development of GRNN models using 80% of the monitoring records is shown in stage one. Stage two was to obtain the GRNN model, using smoothness function (σ) values of 0.1(0.1)0.9. In GRNN the critical step for good prediction is selection of the smoothness parameter (σ), so this study used a trial and error method to get the best smoothness function for predicting future PM10 concentrations.
| Fig. 5: | Procedure for development of GRNN models |
| Fig. 6: | Architecture of general regression neural network (Chen and Yu, 2009) |
Figure 6 shows the architecture of the general regression neural network. The input layer used all the independent variables, i.e. SO2, NO2, CO, RH, T, WS and PM10. The relation between input and output is shown as (Chen and Yu, 2009):
where, wj is weight between the jth pattern and summation layer node X = (x1, x2,..., x7) is dimensional of input vector and φ is Gaussian function.
| Table 2: | Performance indicators for measuring accuracy and errors of model |
| n: Number of observations, Oi: Observed values, Pi: Predicted values, : Mean of the observed values, : Mean of the predicted values, Sobs: Standard deviation of the observed values and Spred: Standard deviation of the predicted values | |
After the GRNN model was obtained in stage two, the model needed to be validated using the remaining 20% of the monitoring record. In stage three, performance indicators were used to measure the accuracy and error of the model.
Ranking and comparison of all the performance indicators at all smoothness functions was conducted in stage four. The best smoothness function was σt found for the range 0.1(0.1)0.9 in stage five. To increase accuracy, further analysis was carried out by addition or subtraction of the best smoothness function in the range of σt(0.01)σt+1 and σt-1(0.01)σt, where, σt is the best smoothness function and σt±1 is the smoothness function after or before the best smoothness function shown in stage six. Performance indicators were compared and the best GRNN models were obtained.
Performance indicators: The model’s performance in both development and validation processes was evaluated by calculating performance indicators. Performance indicators that have been used are, (a) coefficient of determination (R2), Root Mean Square Error (RMSE), Index of Agreement (IA), Normalized Absolute Error (NAE) and Prediction Accuracy (PA). The equations used are given in Table 2 (Yusof et al., 2010; Lu, 2004).
RESULT AND DISCUSSION
The Seberang Jaya monitoring station is located near Seberang Perai monitoring station. However, the Seberang Jaya monitoring station is surrounded by major roads such as Projek Lebuhraya Utara-Selatan Berhad (PLUS) Expressway, Butterworth-Kulim Expressway (BKE) and Federal Road. On 8 August 2005 Seberang Jaya was affected by land and forest fires in the Riau Province of Central Sumatra, Indonesia, with a maximum concentration of 247.583 μg m-3 as shown in Table 3. The average PM10 concentration exceeds than 50 μg m-3.
Feedforward backpropagation models: The main problem faced using the FFBP method was deciding on the FFBP network architecture, such as activation function and number of neurons in hidden layer. Table 4 gives the validation models using different numbers of neurons based on a few performance indicators. The initial value of neurons was five and the number was increased by one unit until optimal value was achieved. Table 4 shows that 12 is the best number of neurons in hidden layer for next-day predictions at Seberang Jaya.
Table 5 shows the result for testing overfitting using cross validation. Table 5 shows that there was not a problem overfitting for Seberang Jaya, because the average result for testing validation (0.1721) was larger than the training set (0.1602).
| Table 3: | Descriptive statistics for Seberang Jaya |
| *S.D: Standard deviation | |
| Table 4: | Validation models using different number of neurons |
| Table 5: | Results for overfitting problem using cross validation |
| Table 6: | Results using different activation functions using original parameters as inputs |
| Table 7: | Summary of FFBP model for next-day |
| *Transfer function A: Transfer function from input layer to hidden layer, **Transfer function B : Transfer function from hidden layer to output layer | |
After determining the best number of neurons (hidden node) in the hidden layer, the best transfer function was obtained. The best transfer functions are obtained using sigmoid transfer functions. This gives the best weights for the activation function. In this study three sigmoid functions were used i.e., linear (purelin), log-sigmoid (logsig) and hyperbolic tangent sigmoid (tansig).
Table 6 shows performance indicators using the three different transfer functions. The best transfer function for Seberang Jaya (next-day) from input to hidden layer is Tangent Sigmoid (tansig) and activation function from hidden layer to output layer is linear transfer function (purelin).
Figure 7 shows the scatter plot of predicted against observed PM10 concentrations in Seberang Jaya. UL and LL are upper and lower 95% confident limits for FFBP models. The results shows that 95.83% of next day (Fig. 7a), 96.03% of next two-day (Fig. 7b) and 97.51% of next three-day (Fig. 7c) monitoring records fall within the confidence limits.
Repeating the procedure revealed the best number of neurons and transfer functions for next two-day and next three-day predictions for Seberang Jaya. Table 7 showed the best model using FFBP for (a) Next-day, (b) Next two-days and (c) Next three-days for all monitoring stations. None of the FFBP models had a problem with over-fitting after being tested using cross validation. For transfer functions from the hidden layer and output layer, linear transfer function (purelin) gives the best result. For the input layer to output layer, tangent sigmoid function is usually chosen as the best transfer function.
General regression neural network: In GRNN the critical step for good prediction is selection of the smoothness parameter (spread). This study used a trial and error method to get the best smoothness function for predicting future PM10 concentrations. Table 8 show performance indicators for the GRNN model using different smoothing factors. Performance indicators were used to select the best smoothing function for predicting PM10 concentration for next-day at Seberang Jaya. Ranking all five performance indicators, the 0.2 smoothing function gives the best GRNN models for PM10 concentration.
| Fig. 7(a-c): | Scatter plot of predicted against observed PM10 concentrations (μg m-3) for FFBP models |
| Table 8: | GRNN result using different smoothing functions for next-day (step one) |
Therefore, 0.2 was used as smoothing function to represent PM10 concentration models in the Seberang Jaya monitoring station for next-day prediction. To obtain a more accurate model, further analysis was carried out taking a smoothing function in the range of 0.2±0.01 i.e., 0.1(0.01)0.2 or 0.2(0.01)0.3. Table 9 shows performance indicators for smoothing functions values 0.2±0.01 and which shows the best smoothing function to be 0.19. This step will improve the accuracy of GRNN models.
| Fig. 8(a-c): | Scatter plot of predicted against observed PM10 concentrations (μg m-3) for GRNN models |
| Table 9: | GRNN result using different smoothing functions for next-day (step two) |
| Table 10: | Summary of GRNN models for all prediction days |
The scatter plots were plotted for the predicted versus observed values of PM10 for next-day, next two-days and next three-days as shown in Fig. 8. The results show that 97.08% of next-day (Fig. 8a), 96.66% of next two-day (Fig. 8b) and 97.07% of next three-day (Fig. 8c) monitoring records fall within the confidence limits.
| Table 11: | Comparison performance indicator for FFBP and GRNN models |
| Table 12: | Comparison of PM10 models using artificial neural network approaches |
Repeating the procedure revealed the best number of neurons and transfer functions for next two-days and next three-days for Seberang Jaya. Table 10 shows the best model using GRNN next-day, next two-days and next three-days. Table 10 shows the range of the smoothing factor for all monitoring stations is from 0.07-0.24.
Comparison of performance: Table 11 showed the comparison of performance indicators between (Multiple Linear Regressions (MLR) and GRNN models. In general, both models show the same results for performance indicators, reflecting greater accuracy in next-day (0.6956) PM10 concentration predictions compared with average accuracy in the next two-day (0.5842) and next three-day predictions (0.4553). Table 11 also shows FFBP models perform better than GRNN models for all three days in predicting PM10 concentration, with fewer errors, as much as 5.6% for next day, 3.5% for next two-day and 2.5% for next three-day predictions. The accuracy of the model appears to be far greater than the GRNN model, with improvements of 8.1% for next day, 11.4% for next two-day and 10.3% for next three-day predictions. The results also showed that FFBP and GRNN could predict future PM10 concentration accurately for the next three-days for this monitoring station.
Table 12 describes the use of ANN in predicting PM10 concentrations. In Table 12, it was found the ANNs are used widely for predicting PM10 concentration, but from the current knowledge none of the FFBP and GRNN models for predicting PM10 concentration have been used in Malaysia. The results of this study show the value of Index of Agreement (IA) is from 0.69-0.94 for next day prediction and coefficients of determination (R2) were between 0.51-0.91 for next day prediction only. These results show similar values of R2 and IA with previous researchers, as shown in Table 12.
CONCLUSION
The FFBP and GRNN approaches have been proven to be effective as techniques to predict future next day, next two-day and next three-day PM10 concentrations in advance for a sub-urban area in Malaysia. The results show that FFBP performs better than GRNN because FFBP has a more complex structure than GRNN. FFBP gives more accurate results when many alterations have been performed to reduce the error between observed and predicted values. GRNN has a simple structure, is less time-consuming than FFBP and is based on curve-fitting prediction. Both FFBP and GRNN give useful information on PM10 status and could be used government bodies to reduce the impact of air pollution in Seberang Jaya, Malaysia.
ACKNOWLEDGMENT
This study was funded by Universiti Teknologi MARA under Grant FRGS/1/2014/STWN01/UITM/02/4. Thank you to Universiti Sains Malaysia and Universiti Teknologi MARA for providing financial support to carry out this study and also thanks to the Department of Environment Malaysia for their support.