Simulated Annealing-Optimized LSTM for Large-Scale Temperature Forecasting Across Türkiye

Accurate temperature prediction is essential for understanding climate variability and hydrological extremes. In this context, Long Short-Term Memory (LSTM) networks have become a widely adopted tool for temperature forecasting; however, their performance strongly depends on hyperparameter selection. This study proposes a combinatorial optimization framework that integrates the Simulated Annealing (SA) algorithm with LSTM networks to enhance long-term temperature forecasting performance. To evaluate the proposed approach, monthly temperature data (1927–2024) from the Turkish State Meteorological Service (MGM) were used. A spatial hold-out strategy (57 training and 24 testing provinces) was employed to assess generalization performance. Model performance was evaluated using MAE, RMSE, R2, and NSE. Results indicate that the SA-LSTM model significantly improves prediction accuracy compared with the conventional LSTM configuration. The optimized model achieved lower prediction errors (MAE = 2.56; RMSE = 3.42) and higher agreement metrics (R2 = 0.856; NSE = 0.848) on the independent testing dataset. These findings demonstrate that combinatorial hyperparameter optimization enhances the robustness and predictive capability of deep learning models for large-scale temperature forecasting and provides a robust and reliable tool for climate and hydrological modeling.