A Principle-Constrained Wind Field Image Generation Framework for Short-Term Wind Power Forecasting

The stochastic and nonstationary wind nature introduces considerable uncertainty in wind power, challenging the power grid management and market clearing. Investigating the spatial-temporal wind field is critical to predicting future wind power variations. However, there is still room for more accurate descriptions of wind field evolution characteristics. In this study, wind field evolution can be segregated into convection, diffusion, circulation, and other unknown processes through modeling by multi-order partial differential equations. Jointly driven by prior knowledge and deep learning, a novel Phycell was proposed to learn temporal dependencies from consecutive wind field images. Thus, a recursive wind field prediction framework was established to obtain multi-step-ahead wind filed images. Furthermore, the wind field forecasts were handled by guided attention to jointly capture the profiles and non-stationarization of future wind power variations. Compared with other state-of-the-art methods, the proposed framework can generate reliable wind field image forecasting results, with average mean absolute error decrease of 9.23%. In addition, accurate wind power forecasting results can be achieved by decreasing normalized mean absolute error of 5.23%. Obvious accuracy improvement and acceptable computational burden indicate the applicability of the proposed method.