Publication

Abstract : Accurate short-term wind power forecasting is crucial for the effective incorporation of renewable energy into modern power systems. Building upon previous hybrid deep learning approaches, this study presents an advanced, AI-driven framework that extensively compares machine learning and deep learning models for short-term wind power forecasting utilizing SCADA data, including variables such as wind speed, direction, theoretical power, and actual power output at 10 mins intervals. Existing approaches suffer from inadequate preprocessing of noisy SCADA data, lack of anomaly detection mechanisms, and the absence of physics-informed constraints ensuring compliance with aerodynamic principles. Moreover, many models fail to capture complex nonlinear and temporal dependencies, limiting their forecasting reliability. This study proposes an AI-driven forecasting framework that compares ML and DL models such as XGBoost, LightGBM, Random Forest, Transformer variations, and Temporal Fusion Transformer with processed data parameters. Anomaly detection is performed using physics-informed neural networks and long short-term memory networks ensuring physical plausibility of forecasts under Betz’ law. Gaussian noise augmentation, normalization, and dimensionality reduction using principal component analysis, which reduced the input feature space while retaining over 90% of the variance, are used for preprocessing. XGBoost demonstrates the highest forecasting accuracy among all evaluated models, attaining an score of 0.9890, a mean squared error of 0.0015, a root-mean-squared error of 0.0385, and a mean absolute error of 0.0151 following preprocessing. These findings emphasize the efficiency of ensemble learning, especially gradient boosting, in addressing complex nonlinearities present in wind power generation. These results indicate that a data-driven and preprocessing-enhanced methodology significantly improves prediction reliability, thus facilitating more intelligent and sustainable energy systems.