Abstract:In order to address the problem of low accuracy in photovoltaic power forecasting caused by the difficulty of traditional single models in effectively analyzing the volatility patterns in historical data, this paper introduces a combination of the categorical boosting algorithm, kernel density estimation, and ultra-short-term photovoltaic power forecasting.Firstly, feature engineering is applied to extract the characteristic vectors of radiation, temperature and humidity related to the modeling, and the statistical model of photovoltaic power probability distribution is created.Secondly, based on the power distribution characteristics and CatBoost algorithm, the power prediction model of PV station is proposed.Finally, the mentioned model algorithm is applied to the actual site calculation to verify its effectiveness, and the comparison with the existing common prediction algorithms shows that the model proposed in this paper can effectively improve the prediction performance.Compared with the traditional models SVR, DTR, KNN, LSTM and Light GBM, the prediction error of SRMSE is decreased by 27.59%, 8.69%, 16.21%, 23.33% and 12.56%, respectively.