The realization of high-precision of the concentration of PM2.5 is of great significance for guiding people's traffic travel, taking targeted intervention measures, and establishing an early warning and forecasting system. However, due to various factors, it is difficult to achieve high-precision prediction. In order to improve the prediction accuracy of PM2.5, this paper proposes a deep learning prediction model based on Lasso for dual feature extraction and Multilayer Perceptron (MLP). Firstly, on the basis of data analysis and processing, the number of leading steps is selected by LASSO, and the hyperparameter is optimized based on Akaike Information Criterion (AIC) and Bayesian Information Criterion (BIC) (hereinafter referred to as "LASSO-LarsIC"); Secondly, on the basis of the selection of leading steps, LASSO-LarsIC is used to extract features of different influencing factors in different leading steps, 76.92% of different influencing factors at the same leading step are eliminated, which reduces the dimension of input variables and improves the generalization ability greatly; Finally, the extracted optimal feature set is used as the input variable, and MLP is applied to predict the concentration of PM2.5 with 1 hour in advance. The verification results show that the prediction accuracy deviation in this paper is 4.85μg/m3, and improved by more than 30% before and after feature extraction, the effect of feature selection is obvious.
|