Operationally Valuable Spatiotemporal Wind Power Forecasting for Enhanced Grid Stability and Economic Performance

The inherent stochasticity of wind power presents significant challenges to power grid stability and the economic viability of wind energy assets. This research develops and validates a novel short-term forecasting framework designed to transcend conventional accuracy metrics by producing forecasts that are both statistically robust and operationally valuable. Utilizing high-frequency multisite power data from a regional wind farm cluster, the methodology centers on comprehensive feature engineering to capture critical spatiotemporal dependencies. A diverse suite of machine-learning algorithms, including Elastic Net, gradient-boosted trees, ensembles, and a temporal convolutional network (TCN) baseline, was systematically evaluated. The performance of key models was further enhanced through automated hyperparameter optimization (HPO) using the Optuna framework. The results demonstrate exceptional predictive performance, with a regularized Elastic Net and an optimized Categorical Boosting (CatBoost) model achieving the highest accuracy on unseen test data, yielding coefficient of determination ( R 2 ) values of 0.996 and 0.986, respectively. A critical finding is that these precise forecasts significantly mitigate grid operating expenses, reducing the required spinning reserve capacity to 12.2 MW and limiting real-time balancing costs to 42.4 thousand United States Dollars (kUSD), which generates approximately 157,000 kUSD in annualized savings compared to a naive persistence baseline. Feature importance analysis consistently reveals that the immediate historical state of the system, such as recent power ramps and rolling averages, was the most dominant predictor. The developed framework provides a validated, high-fidelity tool capable of delivering actionable intelligence to grid operators, thereby supporting more secure grid management and improving the financial performance of wind energy integration.