Vehicle-level multi-strategy benchmarking of NOx emission forecasting using integrated onboard and meteorological data

Nitrogen oxide (NOx) emissions from heavy-duty vehicles pose significant environmental and health risks, especially in urban regions with high traffic density. This study presents a comparative multi-strategy forecasting framework to forecast NOx emissions using real-world fleet data from Brazil, enriched with interpolated meteorological variables. A preprocessing pipeline integrates onboard sensor readings, trip metadata, and weather station data through temporal and spatial interpolation, including kriging. Four regression models: linear regression, random forest, XGBoost, and LSTM, were evaluated under single-step, recursive, multi-output, and direct per-horizon forecasting strategies. Results show that forecasting performance depends jointly on the forecasting strategy and vehicle category. The multi-output strategy provided the most balanced long-horizon behaviour for LSTM in the examined case, reducing relative errors compared with less stable long-horizon alternatives, although variance explanation remained strategy-dependent. In its best single-step configuration, LSTM achieved a MAPE of 10.28%. Adjusted average speed emerged as the most relevant predictor, whereas meteorological variables improved performance only in selected settings. By integrating continuous onboard fleet data from Brazil with spatio-temporal meteorological interpolation via kriging, this study provides a reproducible benchmarking framework for vehicle-level NOx forecasting under real-world multi-horizon conditions.