Research on PM2.5 Concentration Prediction Model Based on Weighted KNN-BP Neural Network

Abstract

Using the weighted KNN-BP neural network method determined by the membership function, a dynamic real-time prediction model for PM_2.5 concentration is established. The model takes the concentrations of six pollutants (PM_2.5, PM₁₀, NO₂, CO, O₃, SO₂) over the past hour, along with six meteorological conditions (weather phenomena, temperature, pressure, humidity, wind speed, wind direction), as well as the day of the week and the hour of the day at the prediction time as dimensions for the KNN instance. Three nearest neighbors are selected to determine the membership weight of each neighbor variable based on the obtained Euclidean distance. Finally, all neighbor dimensions are used as input layer data for the BP neural network, outputting the predicted PM_2.5 concentration for the next hour. This method avoids the issue of traditional BP neural networks not reflecting the impact of historical time window data on current predictions. Prediction experiments were conducted on data from the monitoring station in Dongcheng District, Beijing, from 2014-05-01T00:00 to 2014-09-10T23:00. The results show that the weighted KNN-BP neural network prediction model has the lowest prediction error compared to other methods, and exhibits the best stability, making it an effective method for real-time prediction of PM_2.5 concentration.