Adaptive Ensemble of XGBoost and LSTM for Temperature Forecasting

Accurate weather prediction is crucial for numerous applications, ranging from daily decision-making to emergency response and disaster mitigation. This study introduces an adaptive ensemble method for temperature forecasting that integrates two distinct machine learning algorithms. The ensemble framework dynamically adjusts the weights of each individual model based on their performance characteristics, ensuring that the most reliable predictions are prioritized. The method was tested on historical weather data from five major European cities, consistently demonstrating superior performance compared to standalone models. The results show that the adaptive ensemble achieved R2 values exceeding 0.99 across all locations, indicating a high degree of predictive accuracy. Notably, the geographic location of each city significantly influenced the weight allocation within the ensemble, suggesting that spatially-dependent feature interactions play a more dominant role than temporal patterns in determining temperature variations in these regions. These findings highlight the potential of adaptive ensemble strategies in enhancing the robustness and precision of weather forecasting models across diverse climatic and geographical contexts.