Short-term wind power prediction based on a new hybrid model

The global energy crisis and the integration of intermittent wind power into electrical grids have necessitated accurate short-term forecasting, which has become crucial for mitigating grid imbalances and compensating for limitations. To address this issue, this study investigates the performance of a novel hybrid prediction model, enhanced through data analysis and neural network refinements used to reduce wind power forecasting errors. This technique employs feature engineering, including statistical and temporal pattern analysis, to extract new input features from raw data and process missing values or outliers. Architectural refinements include self-attention mechanisms within dilated convolution and the decoder-free Transformer design, optimized to capture complex temporal dependencies efficiently. A novel hybrid framework integrating customized models leverages feature engineering to enhance forecasting accuracy. The resulting model, validated on datasets from 14 geographically diverse wind farms, significantly reduces prediction errors. Specifically, feature engineering alone boosted accuracy by at least 5.44% (RMSE), while the final ensemble model, integrating the strengths of individual models, achieved a 7.89% RMSE ranking score improvement in generalization performance compared to the next best single model. These results demonstrate the effectiveness of the proposed technique for reliable short-term wind power forecasting across varied terrains, supporting its use for improved operational planning and grid management. • A hybrid neural network model improves short-term wind power prediction accuracy. • Feature engineering improves data quality, reducing errors by at least 5.44%. • The ensemble model increases generalizability by 7.89%, compared to single models. • Validation tests with multi-region datasets confirm robust model applicability.