ATMOSPHERIC VISIBILITY PREDICTION BASED ON NEURAL NETWORK AND CHAOTIC TIME SERIES ALGORITHMS

摘要

In recent years, the problem of low visibility problems caused by air pollution has become increasingly serious. Accurate prediction of atmospheric visibility is more related to anthropogenic activities. In this study, visibility, environmental and meteorological data (2017-2018) from the Yinchuan area of China were selected as experimental data, and the LM-BP neural network, RBF neural network and chaotic time series algorithm were each used to predict atmospheric visibility. For the atmospheric visibility models based on the two neural networks, 1032 data elements were selected to establish the nonlinear relationship between meteorological factors and visibility. These models adopted PM2.5, PM10, SO2, Оз, NO2, atmospheric temperature, wind speed, air pressure, humidity and particle average mass concentration in the atmosphere as inputs and atmospheric visibility was used as the output. Moreover, 1047 elements of atmospheric visibility data were selected to establish the prediction model based on the chaotic time series algorithm. Experiments were performed to verify the feasibility of each model. The results showed the predictions obtained from the two models based on neural networks were better than from the chaotic time series algorithm. Although the RBF neural network model achieved better prediction results than the LM-BP model, the time overhead of the RBF prediction was much greater. However, in practical application, if only historical visibility data were used as the sample input, the chaotic time series algorithm could be used to predict visibility.