Regime-Aware Gated Ensemble for Monthly Temperature Forecasting in Nineveh Governorate

Abstract

Accurate monthly temperature forecasting is important for planning in semi-arid regions, where agricultural scheduling, energy management, and heat-risk preparedness depend on reliable seasonal outlooks. This study develops a regime-aware forecasting framework for monthly maximum (MAX) and minimum (MIN) temperature prediction in Nineveh Governorate, Iraq, across four sites: Al-Hamdaniah, Baaj, Mosul, and Rabiaa. Forecasting is formulated as a mixture-of-experts problem in which three complementary experts SARIMA, a stacking ensemble (ARIMA + LSTM + tree-based learner), and a reverse hybrid (LSTM→SARIMA)  are combined through a gating network. The gate assigns horizon-dependent weights using regime-relevant diagnostics computed from historical data only, together with a forecast-horizon indicator. To avoid temporal leakage, the framework employs chronological train/validation/test splitting, validation-origin training for the gate, and multi-origin rolling evaluation over an 18-month forecasting horizon. The results show that SARIMA is the strongest practical model on average for MAX series, whereas ETS (Holt–Winters exponential smoothing) is highly competitive for MIN series. The proposed gated model yields modest but repeatable improvements over a simple equal-weight ensemble in several site–series cases, with statistical support in a subset of comparisons. Overall, the findings support a practical forecasting workflow in which strong seasonal statistical baselines serve as default models, while regime-aware gating improves robustness when model suitability varies across conditions.