Nowcasting PM2.5 in Beijing Using Synchronous Covariates and Lagged Features: Model Comparison and Variable Selection Stability

Reliable nowcasting of PM2.5 is of practical importance for daily air-quality monitoring and urban management.

PM2.5 concentrations are jointly influenced by emission sources, meteorological conditions, temporal patterns, and station heterogeneity, and the explanatory variables exhibit strong correlations.

This study compares regularized regression methods (Ridge, Lasso, Elastic Net) with deep learning models (MLP and LSTM) for same-hour PM2.5 estimation using hourly observations from 12 monitoring stations in Beijing from March 2013 to February 2017.

Because the feature set includes both PM2.5 lagged terms and synchronous co-pollutant measurements, the task is framed as nowcasting rather than strict forecasting.

Model evaluation employs timestamp-based chronological train/test splitting and TimeSeriesSplit cross-validation.

The MLP achieves the best performance (RMSE = 13.651 ug/m3, R^2 = 0.972), reducing RMSE by approximately 13% relative to all regularized regression models (RMSE approximately 15.6, R^2 approximately 0.964).

The three linear models perform nearly identically.

As an exploratory supplement, the LSTM--constrained by CPU computational limitations to a subsampled training set--underperforms (RMSE = 26.553, R^2 = 0.889) and is not treated as a primary conclusion.

Feature-group ablation shows that lagged terms and synchronous pollutants carry the dominant estimation information.

Variable selection stability analysis under the 1SE rule reveals that Lasso favors sparse selection (4 stable variables) while Elastic Net retains correlated variable groups (11 stable variables); CO, NO2, PM10, and the first-order PM2.5 lag form a robust core set.

Per-station and monthly error analyses further reveal spatial and seasonal heterogeneity in model errors.