Short-term Photovoltaic Power Interval Forecasting Based On Ensemble Learning

The growing demand for energy and the over-exploitation of fossil energy sources have led to increasingly serious energy and environmental problems,and solar energy has received widespread attention from governments and academics around the world because of its clean,pollution-free and inexhaustible characteristics.Photovoltaic power generation,as a type of solar power generation,converts light energy into electricity to achieve clean and pollution-free power generation.However,photovoltaic power generation exhibits certain intermittency and volatility,making it difficult to be grid-connected.Accurate forecasting of PV power generation helps ensure stable operation of the grid.Traditional point forecasting techniques cannot provide sufficient uncertainty information for the forecast results,and therefore have certain limitations.The interval prediction can obtain the upper and lower limits of PV power generation at a certain confidence level.The interval contains sufficient uncertainty information,which is important for grid-connected PV power generation and power system dispatching.With the development of big data and artificial intelligence technology,machine learning methods have taken an important position in PV power generation forecasting.In the most of existing researches,the machine learning methods has achieved higher accuracy.However,the accuracy and robustness of a single machine learning model is limited,while ensemble learning that combines multiple base learners into a strong learner can obtain higher accuracy and stronger robustness.In addition,the tree-based ensemble model cannot guarantee the variety of base learners.In this thesis,an ensemble model based on stacking method is proposed,which ensures the variety of the base model and achieves high accuracy interval prediction of PV power generation.In this thesis,the PV power point forecasting model and interval forecasting model are constructed respectively.The interval forecasting model is built on the results of point prediction research.The interval forecasting model is based on quantile regression,and quantile long-short term memory(QLSTM)and quantile convolutional neural network(QCNN)are built as the base learners of the ensemble model,so that the ensemble model has the advantages of extracting temporal correlation by long-short term memory(LTSM)and the ability of extracting spatial features by convolutional neural network(CNN)at the same time.The ensemble model stacked QLSTM-QCNN-GA-LGBM for short-term PV power interval forecasting is composed by using light gradient boosters(LGBM)as the meta-learner of the model,to adjust the weights of the base model dynamically,and the hyperparameters of which is optimized by genetic algorithm(GA).The main contents and results of this thesis are as follows:(1)Improve the quality of input data set through data preprocessing technology.Which can also improve the speed of the model and reduces the memory usage when the model is running.(2)Feature selecting was carried out through Pearson correlation coefficient analysis in case the negative impact of weakly correlated features to the results.(3).A stacked LSTM-CNN-GA-XGB ensemble model based on the stacking ensemble method which uses extreme gradient boosting(XGBoost)as meta-learner is established to achieve high accuracy short-term PV point forecasting.This model combines the capabilities of LSTM for capturing the dependency of historical PV data and CNN for extracting the spatial features.In this study,a real dataset is used to discuss the accuracy and robustness of the proposed ensemble model by calculating the annual and seasonal values of MAE,RMSE,and R~2.The MAE value decreases by over 16.0%and 9.2%and the RMSE value decreases over 15.3%and 4.0%than LSTM and CNN models respectively.Furthermore,by comparing the proposed model against the results of state-of-art literatures,the ensemble model has the lowest MAE and RMSE values,and the R~2 value is also higher than most of the models in the literatures,according to which there is a conclusion that the proposed model is competitive in the field of PV power forecasting.(4)Based on the results of the point forecasting method,a new model stacked QLSTM-QCNN-GA-LGBM is constructed to achieve high accuracy short-term PV power interval prediction.A KNN model is used to fill the missing values to improve the quality of the input data.Moreover,the XGBoost is changed into LGBM as the meta learner of the stacking ensemble model,and the quantile regression is used as the loss function of each base learner,to construct a new stacking model named stacked QLSTM-QCNN-GA-LGBM for short-term PV interval forecasting.The results show that the PICP,PINAW and AIS values of the forecasting interval obtained by the ensemble model are better than the single deep learning model,and the ensemble model obtains stronger robustness by analyzing the seasonal performance of each model.In this thesis,a real dataset is studied to investigate the accuracy and robustness of the stacked QLSTM-QCNN-GA-LGBM model based on stacking method for short-term interval forecasting of PV power generation.The results show that the ensemble model outperforms the single deep learning model,and the interval forecasting model can provide more uncertainty information compared with the point forecasting model,which is helpful to make correct decisions in scheduling of the power system and ensure the stable operation of the grid.