|
Abstract
|
This paper applies SHapley Additive exPlanations (SHAP) — a comprehensive interpretability framework based on Shapley values from cooperative game theory — to analyze a Bidirectional Long ShortTerm Memory (BiLSTM) model trained to predict multivariate PM2.5 levels. Time-series data were collected from three cities in southern Thailand (Surat Thani, Phuket, and Hat Yai) and included air pollutant factors (PM2.5) along with meteorological factors (precipitation, temperature, humidity, wind direction, and wind speed) from 2023 to 2024. The trained model was designed to forecast air pollution levels over a 24-hour period. Correlation analysis was performed using Pearson coefficients, and linear interpolation was applied to address missing values, improving model accuracy. Model performance was evaluated using three quantitative metrics: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE). The findings showed 1) the proposed BiLSTM model can effectively extract features from multivariate time series data and learn features sufficiently; 2) Using SHAP, PM2.5 were strongly influenced by precipitation, humidity, and temperature. Other than expanding the literatures regarding deep learning models in multivariate context, this study provides practical contribution for management and strategy development of weather. These findings enhance the interpretability of AI models for environmental forecasting and clarify how meteorological factors drive PM2.5 variations. Building on this, future work will focus on developing automated AI systems for real-time, cross-border air quality analysis and management.
|