AN AI-DRIVEN HYBRID LSTM-RF FRAMEWORK FOR DECOUPLING ENVIRONMENTAL AND ANTHROPOGENIC IMPACTS ON AIR QUALITY

Abstract

Evaluating atmospheric emission policies requires accurately separating anthropogenic activities from meteorological influences. Traditional statistical approaches often fail to capture the non-linear, high-dimensional dynamics of atmospheric environments. This research introduces a hybrid machine learning framework integrating an LSTM neural network and Random Forest meteorological normalization to decouple these overlapping effects. A dynamic inversion model was constructed using 2021 hourly observations of PM_2.5, PM10, temperature, humidity, and wind speed from Jinan's Guodian station. The LSTM captured long-term temporal dependencies, achieving a coefficient of determination exceeding 0.80 across seasonal variations. To isolate meteorological impacts, the RF model employed a 14-day random resampling technique, calculating the conditional expectation of pollutant concentrations to marginalize weather randomness. Results revealed distinct seasonal contribution rates: meteorological factors contributed 47.36% to PM_2.5 concentrations during severe December pollution events, driven by stagnant conditions and high humidity. Conversely, this contribution dropped to 24.87% in February, indicating primarily anthropogenic origins. Feature analysis identified humidity and temperature as the primary meteorological drivers of particulate accumulation. This framework provides an interpretable methodology to evaluate the actual efficacy of emission interventions under varying weather scenarios.