Ultra-short-term multi-step prediction of wind power fusing multivariate time-series relationships and multi-task correction

The expansion of the scale of wind power grid connections has made high-precision power forecasting critical to ensuring grid safety. This study proposes a new framework that addresses the limitations of existing ultra-short-term multi-step prediction methods in capturing temporal dependencies among multiple variables and cumulative forecasting errors across multiple steps. The new framework combines cross-variable temporal fusion convolution with multi-task dynamic error correction. First, a time-series feature extraction mechanism that combines wide receptive field convolutional kernels and attention mechanisms is employed. Next, a time-convolutional network with a cross-variable grouping design is used to capture deep dependencies among the data of multi-source time series and generate an initial four-hour forecast. Finally, a multi-expert gated network is constructed that the network employs a dynamic error correction mechanism to optimize multi-step prediction results. Case studies of wind farms under different scenarios demonstrate that the proposed method significantly improves prediction accuracy and it provides reliable support for real-time grid dispatch decisions.