Short Term Prediction of PM2.5 Concentration Based on PCA-LSTM-Random Forest

Abstract

As the primary atmospheric pollutant, accurate short-term PM2.5 prediction is critical for air quality early warning and control. Existing methods are limited by high-dimensional feature redundancy, complex time-series dependence, and nonlinear correlations. This study constructs a PCA-LSTM-random forest integrated model using 2023–2025 daily air quality and meteorological data in Haidian District, Beijing, for 1–3-day PM2.5 forecasting. Forty-dimensional time-series features are built from 10 core variables and their 1–3-day lags; PCA reduces dimensions to 15 principal components (cumulative variance contribution rate 85.16%) to eliminate multicollinearity. LSTM extracts 128-dimensional deep time-series features, which are fused with 10 original physical variables. Random forest fits nonlinear correlations, with global optimization of 1,296 hyperparameter combinations via grid search and time-series cross-validation. Verified against single random forest and pure LSTM, the model achieves trainingset R 2 = 0.993, 1–3-day RMSE = 4.21, 5.87, 7.53 μg/m 3 , and qualification rate >93%, outperforming baselines and supporting high-precision regional air quality early warning.