基于DKDE与改进mRMR特征选择的短期光伏出力预测

With the growth of the installed capacity of the photovoltaic (PV) power generation, its proportion in energy consumption keeps increasing. Accurate forecasting of PV power generation is of great significance to the development plan and dispatching operation of power systems. At present, there are still few researches on feature selection in PV forecasting. Unreasonable feature selection often leads to loss of information, and it is difficult to effectively mine the mapping relationship between meteorological parameters and the power output, which results in low forecasting accuracy. Therefore, this paper proposes a feature selection method for PV forecasting based on improved mutual information calculation and improved max-relevance and min-redundancy (mRMR). Aiming at the problem that it is difficult to directly calculate the correlation and mutual information of continuous random variables, based on the theory of diffusion kernel density estimation (DKDE), an interval division method based on the probability density is proposed and applied to the discretization of variables, which improves the ability of mutual information to represent actual limited data sets. Then, the incremental search process of the traditional mRMR is improved, and an improved mRMR algorithm is proposed, which can select multiple feature subsets in parallel. The XGBoost (eXtreme gradient boosting) algorithm is applied to each feature subset to construct the weather information and PV power forecasting model. Finally, the effectiveness and accuracy of the proposed method are verified by the measured data of an actual photovoltaic power station.