<![CDATA[This study evaluates a CNN-LSTM architecture with scaled dot-product attention for hour-ahead PM2.5 forecasting in Jabodetabek. Due to data constraints, the final model operates on a single univariate input channel - PM2.5's own first-order lag (t-1) - rather than the intended multivariate design; a supplementary experiment that added cyclical calendar attributes (hour, day-of-week, month) to the input did not improve accuracy and was therefore excluded from the final architecture. A three-member ensemble achieves R² = 0.788 and MAE = 9.654 µg/m³ on the held-out test set, outperforming a naive persistence baseline on R² (0.678) and RMSE, though the two are comparable on MAE, indicating the model's advantage lies chiefly in curbing large errors rather than typical-case error. Performance is strong in the dominant high-concentration regime but degrades sharply in the sparsely represented low-to-moderate range, and residual diagnostics indicate the model has not fully captured the series' temporal dependence. These results position the architecture as a working univariate baseline rather than an operationally validated forecaster, and motivate future work incorporating co-pollutant and meteorological data with per-city sequence construction to realize its full multivariate design.]]>
Copyrights © 2026
