<![CDATA[Accurate daily temperature forecasts were needed to support agricultural planning, disaster-risk management, and urban planning in Indonesia. This study evaluated whether a weighted ensemble of BigLSTM, BigGRU, and a Transformer improved one-day-ahead temperature forecasts for six Java cities: Jakarta, Bandung, Semarang, Yogyakarta, Surabaya, and Malang. The models used daily NASA POWER reanalysis data (2000–2026) with 12 meteorological, calendar, and lag features. To avoid temporal leakage, we used a fixed final holdout (from 1 January 2024), expanding-window validation, causal gap filling, and five random seeds per model. Inverse-validation-loss ensemble weights were compared with equal weights. Benchmarks were persistence, seasonal-naive, climatology, a harmonic-regression ARIMA model, and gradient boosting. On the holdout, the ensemble attained a mean root mean squared error of 0.506 °C, 13.8% lower than persistence. Diebold–Mariano and block-bootstrap tests showed significant improvement over persistence, seasonal-naive, climatology, and the ARIMA model in all six cities after Holm adjustment, but not over gradient boosting (0.521 °C) or the Transformer alone (0.503 °C). Inverse-loss weighting did not outperform equal weighting, and member residuals were highly correlated (0.967–0.998). Leave-one-out ablation showed that the gain of the ensemble over its members came mainly from the Transformer. Ljung–Box tests indicated residual autocorrelation only in Bandung. The ensemble was reliable but not uniquely superior for next-day city temperature forecasting.]]>
