Multi‐Location Temporal Forecasting of Near‐Surface Temperature and Relative Humidity Across Diverse Climatic Regions of India Using a Stacked Ensemble Framework

ABSTRACT Accurate short‐term forecasting of near‐surface air temperature and relative humidity is important for climate monitoring, urban planning, energy management, agriculture and other weather‐sensitive socio‐economic activities. However, the strong climatic variability across India presents challenges for developing forecasting frameworks capable of generalising across multiple geographic regions while maintaining predictive accuracy. This study proposes a multi‐location temporal forecasting framework based on a stacked ensemble machine learning approach for predicting near‐surface temperature and relative humidity across multiple Indian cities representing diverse climatic conditions. The proposed forecasting architecture integrates Extreme Gradient Boosting (XGBoost), Light Gradient Boosting Machine (LightGBM), and CatBoost as base learners. Predictions generated by these models were combined using Ridge regression as a meta‐learner within a stacked ensemble framework. The framework was implemented within a multi‐output regression structure to enable simultaneous prediction of temperature and relative humidity. Model performance was evaluated using the coefficient of determination ( R 2 ), root mean squared error (RMSE) and mean absolute error (MAE) under chronological held‐out testing and time‐series‐aware validation. On the pooled evaluation of 8742 held‐out observations, the stacked ensemble achieved R 2 values of 0.933 for temperature and 0.787 for relative humidity, with MAE values of 0.802°C and 6.320°C percentage points, respectively. Per‐city persistence achieved a comparable temperature R 2 of 0.933 and a lower temperature MAE of 0.783°C, while ARIMA achieved a higher humidity R 2 of 0.797 and a lower humidity MAE of 5.699 percentage points. Five‐fold walk‐forward validation yielded mean R 2 values of 0.935 ± 0.018 for temperature and 0.898 ± 0.024 for humidity. Leave‐one‐city‐out validation assessed performance at unseen locations. Recursive multi‐step forecasting at 6‐, 24‐, 48‐ and 96‐h horizons showed progressive forecast degradation with increasing lead time, particularly beyond 24 h. SHAP‐based interpretability analysis identified lagged temperature as an influential predictor, highlighting the importance of temporal persistence in short‐term meteorological forecasting. Overall, the results show that the stacked ensemble provides useful short‐term predictive skill, while the walk‐forward, persistence, LOCO and multi‐step evaluations place clear limits on claims of temporal and cross‐location generalisation.