A parallel differential learning ensemble framework based on enhanced feature extraction and anti-information leakage mechanism for ultra-short-term wind speed forecast

Accurate ultra-short-term prediction plays a very important role in maintaining power equipment, preventing accidents, and optimizing dispatch effectiveness. Currently, the decomposition-integration method is widely used in ultra-short-term wind speed prediction. However, most of the existing models ignore the problem of information leakage that occurs during data processing and the effect of discrepancies between multiple decomposition sequences on the prediction results, which poses a great challenge to the accuracy of wind speed prediction. Therefore, this study proposes an improved hybrid wind speed prediction framework based on an improved decomposition method, an anti-information leakage mechanism and an enhanced deep learning algorithm. First, the original sequences are processed using improved singular spectrum analysis (ISSA) to achieve an effective mining of deep features. Second, Transformer is selected to construct the input-output relationship model between the original sequence and the feature components to form an anti-information leakage mechanism. Finally, an enhanced hybrid deep learning model is built using the concept of parallel processing, which can simultaneously process subsequences of different complexity and effectively reduce the prediction error of the model. Simulation experiments are conducted using four sets of data from wind farms located in Liaoning Province, China. The results of the simulations demonstrate that the model performs better in predictions than the benchmark model.