What question did this study set out to answer?

This study aims to enhance short-term wind power forecasting with a model that addresses nonlinearity and uncertainty.

May 16, 2026Open Access

Trustworthy Wind Power Forecasting Based on Inverted Transformer with Variable-Wise Interaction and Evidential Learning

Puntos clave

This study aims to enhance short-term wind power forecasting with a model that addresses nonlinearity and uncertainty.
Proposed FoM-EiT model integrates cyclic feature encoding and inverted Transformer for variable-wise learning.
Conducted experiments on four real-world wind farms to test forecasting accuracy and uncertainty measures.
Utilized statistical tests to compare model performance against existing baselines.
Achieved R2 values of 0.6342, 0.8211, 0.7844, and 0.9161 for point forecasting across different farms.
FoM-EiT's Prediction Interval Coverage Probability reached values of 0.9492 to 0.9709, indicating high reliability.
Statistical tests confirmed significant improvements over conventional methods in both forecasting accuracy and uncertainty quantification.

Resumen

The inherent nonlinearity and uncertainty of wind power generation pose significant challenges to the security, stability, and economic operation of power grids. Therefore, accurate and reliable wind power forecasting is crucial for seamless grid integration and effective risk assessment. Existing forecasting models often focus on improving point-prediction accuracy while overlooking effective multivariate dependency modeling and reliable uncertainty quantification, limiting both the informativeness and reliability of their forecasts. This study proposes a Fractional-order Momentum optimized Evidential iTransformer (FoM-EiT) for short-term wind power forecasting from multivariate time series. The proposed model integrates cyclic feature encoding for periodic variables, an inverted Transformer for variable-wise interaction learning, and an evidential output head that jointly produces point forecasts and uncertainty estimates from a shared representation. The proposed fractional-order momentum (FoM) optimization accumulates gradient history over an extended window, thereby smoothing oscillations caused by gradient competition and stabilizing the joint training process. Experiments on four real-world wind farms from different geographical regions show that FoM-EiT achieves competitive point forecasting performance, with R2 values of 0.6342, 0.8211, 0.7844, and 0.9161, and the Wilcoxon signed-rank test indicates that its advantages over the baselines are statistically significant in the vast majority of comparisons. For uncertainty quantification, FoM-EiT achieves Prediction Interval Coverage Probability (PICP) values of 0.9492, 0.9682, 0.9709, and 0.9498, while the Winkler score results further show that its prediction intervals outperform the conformal prediction and quantile regression baselines in terms of overall interval quality. These results indicate that FoM-EiT provides both accurate forecasts and trustworthy uncertainty information, making it a practical tool for dispatch, reserve allocation, and risk-aware short-term power system operation.

Leer artículo completoexternamente

Me gusta

Guardar

Ver artículo completo