Short-term Photovoltaic Power Prediction Based On Ensemble Learning

The over-dependence on fossil energy and the increasingly serious environmental pollution force all countries in the world to shift their development focus to renewable clean energy,thus realizing the transformation of energy structure.Photovoltaic power generation has many advantages,such as rich resources,convenient use,clean and pollution-free,and has a broad space for development.However,photovoltaic power generation is susceptible to weather,showing intermittency,seasonality,volatility and uncontrollability day and night,causing a non-negligible impact on the power grid and affecting the stable operation of the power grid.High-precision photovoltaic forecasting is an effective measure to deal with this problem,which has practical significance for power grid economic dispatch and energy market decision-making.In recent years,with the rapid development of machine learning,support vector machines,Artificial Neural Networks,random forest,etc.have been widely used in photovoltaic power prediction,and achieved good results.However,there are still large limitations in the prediction performance of a single model,and ensemble learning is an effective method to improve the prediction accuracy and enhance the generalization ability of the model.In this paper,based on Ensemble learning,a dual channel feature processing feature extraction hybrid model(BOA-LSTM-RBF)and an adaptive upsampling Ensemble learning model(TCN-LGBM-ensemble)based on lightweight Gradient accelerator(LGBM)and time convolution neural network(TCN)were constructed respectively,and Short-term photovoltaic power is predicted at different plants.The details are as follows:(1)The influencing factors of photovoltaic power generation are explored in multiple dimensions,through visual analysis,gray correlation analysis,and Pearson correlation analysis.(2)The data set realizes the transformation of high-quality data through missing value filling,outlier processing,and feature screening.(3)The BOA-LSTM-RBF model realizes photovoltaic prediction by combining Long and short term memory neural network(LSTM)and radial basis function neural network(RBF).The time series features and highly nonlinear discrete features in the data are extracted using LSTM and RBF,respectively.The Bayesian Optimization Algorithm(BOA)optimizes the number of neurons in the hidden layer and the training batch to further improve the accuracy.The experimental results of three different photovoltaic stations show that the proposed model has significant advantages over LSTM,BOA optimized LSTM(BOA-LSTM)and long and short time memory radial basis function neural network(LSTM-RBF)in prediction accuracy and versatility.The model can not only achieve excellent prediction results in the case of stable power changes,but also quickly respond to weather changes even in the case of severe weather fluctuations,ensuring excellent prediction performance.(4)The TCN-LGBM-Ensemble model combines the powerful massive data processing capabilities of LGBM with the temporal and spatial feature processing capabilities of TCN,and expand the extracted information through the upsampling,Finally,LGBM classification is used to dynamically weight the candidate values,high-precision and high-generalization photovoltaic power prediction can be achieved.Compared with the LGBM model,the TCN model,and the TCN-LGBM-mean model using the conventional integrated strategy averaging method as the LGBM and TCN integration strategy,the MAE of the TCN-LGBM-Ensemble model decreased by 1.51%～25.1%,the MSE is decreased by 0.98%～14%,and the RMSE is decreased by 0.6%～7.7%.In different seasons and different weather types,TCN-LGBM-Ensemble shows higher prediction accuracy and generalization ability.Based on a large number of actual photovoltaic power data,the effectiveness and excellent performance of the proposed BOA-LSTM-RBF model and TCN-LGBM-Ensemble model are verified through comparative analysis of various neural network models,which can provide high-precision,reliable and highly generalized short-term photovoltaic power prediction for photovoltaic power stations.