Neuro-fuzzy network for PM2.5 prediction

Abstract

This paper proposes a PM2.5 prediction system based on neuro-fuzzy neural networks which can be trained through historical recorded information. Time series training data are employed to forecast the PM2.5 values in the air in the future. Because of the uncertainty of the involved impact factors, fuzzy elements are added to the forecasting system. Our prediction system is a four-layer fuzzy neural network, consisting of the input layer, fuzzy layer, inference layer, and output layer. First of all, training data are partitioned into fuzzy clusters whose membership functions are characterized by the learned means and variances. Fuzzy rules are then extracted and constructed. Next, least squares optimization and gradient descent backpropagation are applied to refine the parameters of the fuzzy rules. The output of the system, indicating the forecast PM2.5, is derived through the fuzzy inference process. Experimental results are shown to demonstrate the effectiveness of the proposed forecasting system.