Categorical Spatial Interpolation of Solar Irradiance for Regional Distributed Photovoltaic Power Forecasting

High-spatial-resolution (HSR) solar irradiance forecast data is important for regional distributed photovoltaic (PV) power forecasting. Distributed PV sites are widely geographically distributed and the cost of obtaining HSR irradiance forecast data with full regional coverage is very high. Improving spatial resolution through interpolation based on limited sparse irradiance forecast reference points is a low-cost and feasible approach. However, existing irradiance spatial interpolation methods would produce large errors under variable weather conditions (e.g., cloudy) with limited reference points, because local abrupt changes in irradiance usually occur under such weather conditions due to cloud movement. This paper proposes a categorical spatial interpolation method to improve the accuracy of distributed PV power forecasting with limited sparse irradiance forecast data. The method first converts satellite-derived shortwave radiation images into binary irradiance maps through adaptive thresholding, then trains a 3D U-net model to predict future irradiance category probabilities for each grid. These probability maps dynamically govern two parallel interpolation processes: one using reference stations in predicted clear-sky regions, another utilizing stations in cloudy zones, with final irradiance values determined by probability-weighted fusion. Case studies on a real dataset verify the effectiveness and superiority of the proposed method.