Improving Air Quality Forecasts with LSTM and SHAP Explainability: A Case Study in Jakarta

Accurate air-quality forecasting is essential for public-health advisories in large tropical megacities such as Jakarta. This study develops an explainable deep-learning pipeline to predict Indonesia’s Air Pollution Standard Index (ISPU) at the DKI-5 station using daily data from 2017–2021. After handling missing values and integrating meteorological variables, all features were min–max normalized and framed with a lag window of five days. A stacked LSTM (128 and 64 units, dropout 0.2, Adam optimizer, MSE loss) was trained with an 80/20 train–test split. Model performance was assessed using MAE, RMSE, and R2R^2R2. To open the “black box,” SHAP was applied to quantify each feature’s contribution to the predictions. Results show stable convergence of training and validation losses and good generalization. The best configuration achieved MAE ≈ 7.96, RMSE ≈ 10.26, and R2≈ 0.52 on the test set. SHAP analysis indicates that PM10_lag1 is the most influential predictor, followed by wind speed (ff_avg_lag1), relative humidity (RH_avg_lag1), and average temperature (Tavg_lag1), confirming the joint role of recent pollutant levels and meteorology in driving ISPU variability. Compared with a previous LSTM configuration on the same site, the proposed model lowers RMSE by ≈25%, evidencing a more accurate and reliable ISPU forecast while providing transparent feature attributions. The proposed LSTM–SHAP framework offers an interpretable decision-support tool for air-quality management in Jakarta.