The Hybrid Neural Networks-ARIMA/X Models and ANFIS Model for PM-10 Forecasting: A Case Study of Chiang Mai, Thailand’s High Season

Abstract

It was demonstrated that efficient air quality models are very useful tools in forecasting air pollutants. Consequently, in this paper, the influence of exogenous variable (X) related meteorological parameters and correlated toxic gas together with the significant historical PM-10 values was analyzed to formulate the numerical hybrid PM-10 forecast model during high season (January-April) in Chiang Mai Province, northern Thailand. The hybrid model is divided into two stages, dealing firstly with nonlinear transformation through the multilayer perceptron neural network (MLPNN) and radial basis function neural network (RBFNN), and secondly with statistical estimation of the linearly stationary residuals using an autoregressive integrated moving average (ARIMA) and ARIMAX, and denoted by hMLPNN/RBFNN-ARIMA/X model. On the tradeoff between the accuracy using root mean square error (RMSE) and mean absolute error (MAE), and the reliability through Akaike information criterion(AIC) for PM-10 forecasting, the hMLPNN-ARIMA model was identified as the optimal model whereas the hMLPNN-ARIMAX and the hRBFNN-ARIMA model were identified as the suboptimal model. For further comparison against PM-10 forecast based an adaptive neuro-fuzzy inference system (ANFIS) model, it was indicated that the hMLPNN-ARIMA model provided slightly more accurate but clearly more reliable than that of the ANFIS including the others.