Forecasting the PM-10 using a deep neural network

Abstract

The air pollutants related to PM-10 are increasingly adversely affecting people in upper Northern Thailand, especially during the annual dry season. Due to the highly nonlinear dynamics of PM-10 contributed by various factors, in this study a deep neural network (DNN) has been implemented as a tool forecasting PM-10 for air quality alerts. In its design, the time lags of PM10 and significant meteorology conditions, as well as the well-correlated fire-hotspots as major PM-10 sources in this area, are included in the predictor set. The training hyperparameters were optimized automatically by a genetic algorithm, whereas the DNN’s parameters were fine-tuned using back-propagation algorithm. Besides, regularization based on a dropout technique was employed to prevent over-fitting. In testing the proposed DNN-based PM-10 forecasting model outperformed the others. For oneday ahead forecasting it provides a good up to 85% accuracy.

Keywords

• deep neural network
• pm-10
• genetic algorithm
• dropout
• machine learning