Temporal pattern-aware temperature forecasting using CatBoost: A hybrid machine learning approach

• Introduced an Adaptive Temporal Drift Encoding (ATDE) mechanism that dynamically models seasonal variability for improved temporal awareness. • Developed a robust and interpretable CatBoost-based forecasting framework optimized through Optuna adaptive hyperparameter search . • Demonstrated strong cross-regional and cross-country generalization across mountainous, coastal, rainforest, and continental climates. • Conducted detailed extreme-weather and high-error analyses to evaluate model robustness under volatile climatic conditions. • Quantified the contribution of temporal feature engineering through an ablation study , showing a 21.6 % MSE reduction and R² improvement of +0.07 . Accurate temperature forecasting remains challenging due to complex and dynamic temporal patterns. This study evaluates CatBoost, LightGBM, and XGBoost within a time-series cross-validation framework enhanced by Adaptive Temporal Drift Encoding (ATDE) and enriched with engineered lagged and temporal features. ATDE is introduced as a dynamic extension of traditional sine cosine encodings that adaptively scale temporal features according to local temperature variability, improving temporal sensitivity and generalization. Experiments on real-world meteorological data from multiple Indonesian weather stations show that CatBoost consistently achieves the lowest prediction errors, with an average RMSE of 0.413°C at the primary site (Citeko). To evaluate robustness, the model was extended to five international datasets (China, India, France, Russia, and Turkey), maintaining stable performance across diverse climatic zones. Comparative evaluation against deep-learning baselines (LSTM and CNN-LSTM) demonstrates that CatBoost + ATDE achieves substantially lower RMSE (< 1°C vs > 2°C) while requiring significantly less computational cost. Hyperparameters were optimized using an Optuna-based adaptive search to ensure rigorous and reproducible tuning. Error analysis highlights increased uncertainty under extreme weather transitions, while ablation studies confirm that adaptive temporal encoding reduces RMSE by up to 15 %. Overall, the proposed lightweight yet interpretable framework underscores the effectiveness of adaptive temporal feature modeling for robust, climate-adaptive, and energy-efficient temperature forecasting.