基于MATLAB的BP神经网络在空气质量预报中的应用

BP neural network and its application to the forecast of the air quality based on the MATLAB

选用上海市环境监测中心发布的PM_(10)空气污染指数(PM_(10)API)数据和中国气象局公布的有关气象资料,分别建立了未改进BP网络模型和采用"提前终止法"泛化改进的BP网络模型,用于预报上海市19个区县的PM_(10)A PI指数。结果表明,改进BP网络模型预测值和实际值的线性回归显著,其预测相对误差为1.76%～29.45%,预测效果优于未改进BP网络模型,预测精度较高,推广能力较强。本研究为空气质量日报和预报提供了一种全新的思路和方法。

This paper is aimed at offering a detailed introduction to an unimproved BP model and the ＂early termination＂ generalized improved one for the purpose of forecasting the PM10API of 19 districts in Shanghai. So as to make decisions and prevention measures, it is necessary to present the likely trend of air pollution to the public and provide more timely, accurate and comprehensive information on the air quality. As a matter of fact, the concentration of atmospheric pollutants has been becoming a predominant factor that affects the quality of people＇s living standard. Thus, it has become an indispensible part of our meteorological forecast to inform the public of intensity of the pollutants discharge and their effect on the quality of their daily life. And, here, the API is actually the calculation results on the concentration of atmospheric pollutants. Based on the above reasons, this paper has chosen PM10API data from the Shanghai Environmental Monitoring Center and the relevant meteorological data issued by the China National Meteorological Administration. The seven factors are to be collected on the day before, including the day＇s average atmospheric pressure, average temperature, average relative humidity, the mean total cloud cover, the mean wind speed, in addition to the rainfall, along with the accounting for the basic forecasting details of the PM10API, though the intraday PM10API belong to the output information and data. Then the unimproved BP model and the ＂early termination＂ generalized improved model were used to forecast the PM10API of 19 districts in Shanghai. We have also regularized the data concerned so as to speed up the convergence of the training network and prevent the problem of ＂over-fit＂ in calculation. When the BP model was improved by the method of ＂early termination＂, the actual process of training can be temporarily stopped in advance and will let it be to turn to the best training stage. Therefore, our study shows that the linear regression model between the values predicted and the values observed can be expected to get approved by significant testing and the forecast results of the improved BP model are more accurate than that by the unimproved BP one. The forecast relative error of the improved BP model proves between 1.76% and 29.45%, which indicates the higher precision. This study provides some new ideas and methods for daily reporting and forecasting of our air quality.