Ultra-short-term wind power prediction for enhanced reliability considering error distribution characteristics and guided correction

Existing studies on wind power prediction primarily focus on improving prediction accuracy while neglecting the reliability of the predictions. Considering that overestimation in wind power prediction may compromise the secure operation of the power grid, this study proposes an ultra-short-term wind power prediction method that incorporates the characteristics of prediction error distribution and applies guided correction. To enhance prediction accuracy, the method integrates a TCN (temporal convolutional network) with a self-attention mechanism to construct a multi-perspective feature-enhanced attention module, improving the model's ability to capture both local temporal dependencies and global feature information. Furthermore, by incorporating asymmetric coefficients and bias terms into the loss function, a reliability-oriented prediction strategy is developed, enabling guided modulation of the prediction outputs to improve reliability. The bias term affects prediction accuracy and reliability roughly 10 times more than the asymmetry coefficient. The relationship between prediction accuracy and reliability is systematically analyzed. Reliability analysis indicates that improvements in reliability within a certain range have minimal impact on prediction accuracy. Under a balanced scheme, the probability of predicted values being lower than actual values increased by 63.58%. Finally, comparative analyses demonstrate the effectiveness and advantages of the proposed prediction model.