MOGMO-based Spatio-temporal Wind Power Forecasting Framework. • A spatio-temporal model is built using UK wind farm data from 2021 to 2023. • Lasso-based VAR improves forecast accuracy in high-dimensional wind datasets. • VAR models outperform ARIMA and baseline by 2.7% in RMSE performance. • Forecast uncertainty is quantified with 95% prediction intervals. • Spatial correlations among wind farms enhance prediction reliability. Accurate short-term wind power forecasting is crucial for ensuring the efficient and cost-effective operation of power systems and energy markets. Traditional models predominantly depend on past outputs from individual wind farms, often neglecting spatial dependencies and the need for quantifying forecast uncertainty. This study addresses these limitations by incorporating both temporal and geographical information to enhance forecasting accuracy. Using persistent forecasts as a benchmark, we apply ARIMA and VAR models—enhanced with Lasso regularization to manage high-dimensionality and prevent overfitting. We also investigate data transformations (logit and Box-Cox) to address non-constant variance and boundary issues. To address trade-offs among prediction accuracy, uncertainty quantification, and model simplicity, we employ a multi-objective optimization framework based on the Geometric Mean Optimizer (MOGMO). Each wind farm’s optimal model is selected based on performance across these criteria. The results demonstrate that the VAR model applied to normalized, non-transformed data yields the most accurate forecasts, surpassing the benchmark by approximately 2.7%. Prediction intervals are also constructed to quantify forecast uncertainty, achieving high coverage and acceptable interval widths, thereby demonstrating the robustness of the proposed framework.
Yuqi Huang (Sun,) studied this question.