<![CDATA[This research evaluates multi-horizon forecasting of daily mean temperature (TAVG) and maximum temperature (TMAX) at the Malang/Karangploso Climatology Station WMO 96943 using daily meteorological observations from 2014–2023. The predictors include humidity, rainfall, atmospheric pressure, wind variables, weather conditions, and sunshine duration. Forecasts are generated for H+1, H+3, H+7, and H+14 using a chronological train, validation, and test split. The study compares GRU, LSTM, Hybrid Gated LSTM-GRU, XGBoost, and Weighted Ensemble models against Persistence and Climatology baselines. To prevent data leakage, preprocessing includes missing-date handling, placeholder correction, training-set-based imputation, lag and rolling feature construction, and input normalization. Optuna is used to tune recurrent models, while ensemble weights are optimized on the validation set. Model performance is assessed using MAE, RMSE, and R², with the lowest test RMSE as the main selection criterion. Results show that the Weighted Ensemble achieves the best aggregate performance, with mean MAE of 0.796, mean RMSE of 1.007, and mean R² of 0.472. GRU is the strongest individual model, with mean RMSE of 1.022. However, the best model varies by target and horizon. Weighted Ensemble leads in five of eight scenarios for TAVG H+1, TAVG H+3, TAVG H+14, TMAX H+1 and TMAX H+14, Hybrid Gated performs best for TAVG H+7, and GRU is superior for TMAX H+3 and H+7. ]]>
Copyrights © 2026
