Short-Term Photovoltaic Power Forecasting Based On Machine Learning

Under the major demand of carbon peak and carbon neutrality,the development of photovoltaic power generation is an significant way to achieve energy sustainable development.With the increasing penetration rate of photovoltaic power generation in the power grid,the intermittency and volatility of the power grid operation has brought more and more prominent security problems.Reliable photovoltaic power forecasting can effectively reduce the negative impact brought by the uncertainty of photovoltaic power generation,which is an important techniques to ensure the absorption of new energy and the stability and safety of grid operation.Due to the powerful non-linear data processing capability,machine learning models have been widely used in the field of photovoltaic power forecasting.It is difficult for a single machine learning model to meet the requirements of the power system for high accuracy photovoltaic power forecasting.This paper focuses on the combined photovoltaic power forecasting model based on machine learning,and the following main works have been carried out in this paper.(1)A photovoltaic power forecasting method based on wavelet packet transform and gated recurrent unit network is designed.The box-plot method is utilized to process the outliers of photovoltaic data,and Pearson correlation coefficient is utilized to select the important meteorological factors.The wavelet packet transform technique is utilized to two-layer decompose and single-branch reconstruct the photovoltaic power into a series of sub-signals,which are used with important meteorological factors to construct the inputs for each gated recurrent unit network respectively.The output of each network is integrated to obtain the final predicted photovoltaic power value.By fully mining the frequency information hidden in the photovoltaic power,the proposed method can achieve accurate photovoltaic power forecasting.(2)A photovoltaic power forecasting method based on similar day selection method is established.The similar day index based on the trend factor and volatility factor of the global horizontal radiation and the composite meteorological factor is designed.The weight of each matching coefficient is determined by Levy flight beetle antennae search algorithm,and the similar days of the current forecast day are selected by comparing the total similarity index value of each historical day.The inputs of each gated recurrent unit consist of each similar day power sub-signal and meteorological data respectively.The network structure of gated recurrent unit is optimized by Levy flight beetle antennae search algorithm.The proposed similar day selection method eliminates the historical day sample data which is significantly different from the forecast day,improves the quality of training data,and further improves the accuracy of photovoltaic power forecasting.(3)The simulation and comparison experiment based on the real dataset of photovoltaic power plant is carried out.Comparing with other signal decomposition techniques and machine learning models,the effectiveness of wavelet packet decomposition and gated recurrent unit in extracting frequency features accurately and improving the efficiency of network training is verified.Twelve prediction days belonging to different weather and different seasons in the dataset are selected for simulation experiments.The results show that the photovoltaic power forecasting method based on similar day selection has significant advantages in terms of prediction accuracy and generalization ability.