Multi-Model Forecasting of Shoreline Air Temperature using Machine Learning and Deep Learning: The CidAlmeida Station, Portugal

Abstract

Accurate short-term meteorological forecasting at reservoir shorelines is essential for reservoir management and adaptation planning in Mediterranean climates. This study proposes a multi-stage data-driven framework for time-series prediction and explainable modeling using hourly meteorological data (2018–2023) from the CidAlmeida station of the Alqueva reservoir. Several imputation techniques were used to fill missing data, where Stineman interpolation achieved the lowest errors, while linear interpolation performed best for maximum wind speed, wind direction, and precipitation. A range of statistical, machine learning, deep learning, and hybrid models were evaluated using MAE, RMSE, MAPE, R2, and NSE. XGBoost achieved the best overall performance (RMSE = 0.45°C, MAE = 0.34°C, R2 = 0.9962), followed by Random Forest (RMSE = 0.63°C). Deep-learning models, including LSTM, LSTM-KANs, GRU, and Transformer variants, also demonstrated strong predictive performance with RMSE values below 1°C. In contrast, the statistical models (ARIMA, SARIMA, VAR, and Holt-Winters) produced substantially larger errors, highlighting the superiority of machine learning and deep learning approaches. Seasonal analysis further confirmed the robustness of XGBoost, maintaining RMSE between 0.38–0.42°C across all seasons. SHAP analysis revealed that lagged air temperature is the dominant predictor. Overall, the framework provides an accurate, efficient, and interpretable solution for short-term microclimate prediction.