Comparative evaluation of traditional, machine learning, deep learning, and hybrid models for short-term wind speed forecasting in Colombo, Sri Lanka

Short-term wind speed forecasting is critical for optimizing wind-dependent operations in Sri Lanka. However, existing studies lack comprehensive comparisons across different time series forecasting frameworks and insufficiently examine performance differences between univariate and multivariate approaches. This study addresses these gaps by evaluating traditional time series models (ARIMA, ARIMAX, VAR), machine learning models (Random Forest, XGBoost, Support Vector Regression), and deep learning models (Feed Forward Neural Network, Long Short-Term Memory, Convolutional Neural Network) for daily wind speed forecasting in Colombo, Sri Lanka. A residual-correction hybrid framework combining the optimal standalone XGB with deep learning models (XGB-FNN, XGB-LSTM, XGB-CNN) is systematically evaluated for the first time in the Sri Lankan wind speed forecasting context. Meteorological data from January 1, 2013, to December 31, 2023, encompassing ten atmospheric variables, were analyzed. Model hyperparameters were optimized using Grid Search CV, Random Search CV, and Optuna with time series cross-validation. All residual-correction hybrid models used Optuna-tuned XGBoost, which is the optimal standalone model, as the primary model. The XGB-CNN with Optuna-tuned CNN achieved the best numerical performance with MSE of 0.5172, RMSE of 0.7192, MAE of 0.5432, and R 2 of 0.7855. Subsequent Diebold-Mariano testing and Multiple-Comparison Correction via Model Confidence Set confirmed that, although all hybrids showed numerically improved performance over the standalone XGBoost model, these improvements were not statistically significant. This suggests that both the optimal standalone XGBoost model and the proposed hybrid forecasting frameworks constitute robust and reliable approaches for forecasting in the present study. The most influential features identified were minimum, maximum, and average temperatures, precipitation, wind direction, southwest monsoon season, cloud cover, previous-day wind speed, previous-day sea level pressure, current and previous-day relative humidity, and surface pressure change from the previous day.