Implementing ultra-short-term wind power forecasting without information leakage through cascade decomposition and attention mechanism

The depletion of fossil fuels and environmental pollution have increasingly led to the recognition of wind power as a significant sustainable energy source. However, the intermittent and unstable nature of wind energy underscores the critical importance of accurate wind power forecasting for maintaining the stability of power systems. This paper aims to achieve precise forecasting of ultra-short-term wind power generation by proposing an innovative and practical method utilizing a novel self-sustaining cascading rolling mechanism. Initially, employing a rigorous data partitioning approach to ensure the independence of the training and testing datasets, and determining a rolling decomposition window of 192 time steps through preliminary experiments. Subsequently, the decomposition window was gradually shifted backward along the temporal axis, applying the ICEEMDAN algorithm independently within each window to eliminate any possibility of information leakage. Finally, a TCN-BiLSTM-Attention forecasting model was constructed, which accepts the multiple components obtained from the rolling decomposition as input, allowing for accurate predictions of wind power fluctuations over various forecasting horizons ranging from 15 min to 1 h. The effectiveness of the hybrid algorithm was validated through comprehensive experiments. Thanks to the resolution of the information leakage issue, this hybrid method can be implemented in a simulated online context.