Stacking Hybrid Model for Time Series Forecasting

Abstract

As environmental risks continue to grow, the task of forecasting pollution levels is becoming increasingly important. Traditional statistical models such as ARIMA demonstrate high accuracy at background pollution levels but lose forecast quality at peak concentrations that are critical from a public health perspective. Machine learning models, in particular gradient boosting, conversely, perform better at reproducing nonlinear peaks, but degrade over long forecast horizons due to recursive error accumulation. These limitations necessitate the development of hybrid approaches capable of combining the advantages of both model classes.

A hybrid stacking architecture for hourly time series forecasting has been developed and investigated, it combines the structural Prophet model and the LightGBM gradient boosting algorithm. Unlike the traditional residual correction approach, in the proposed architecture the Prophet forecast is used as a deterministic seasonal feature alongside autoregressive and calendar features, on the basis of which LightGBM directly predicts the target value.

The implementation and testing of the proposed model were carried out using real air quality monitoring data obtained from the EcoCity public air quality monitoring network. The statistical significance of the results was confirmed by the Diebold-Mariano test. Comparison with baseline models showed that the hybrid model statistically significantly outperforms the baseline models and demonstrates no critical accuracy failures across all concentration regimes. The results obtained confirm the feasibility of using the proposed model in early warning systems for hazardous air pollution levels.