Short-term Photovoltaic Power Prediction Based On Broad Learning System

Energy is the foundation of economic and social development,and fossil energy has supported the progress of human civilization for hundreds of years.However,excessive reliance on fossil energy will lead to serious damage to the ecological environment and energy depletion.Photovoltaic power generation with rich resources,pollution-free and low cost can effectively solve the energy crisis and environmental pollution problems.The installed capacity of photovoltaic power generation in the world has been increasing,which has effectively alleviated the energy pressure of countries.However,due to the intermittent and volatile nature of photovoltaic power generation,it has brought challenges to the real-time scheduling and stable operation of the grid.Therefore,it is of great significance to effectively predict the output power of photovoltaic power generation for energy development.The power prediction model based on data-driven method is generally divided into physical prediction model,statistical prediction model and machine learning prediction model.Physical prediction model and statistical prediction model are greatly affected by meteorological conditions and data noise,while neural network is outstanding in dealing with nonlinear and complex problems.Although deep neural network has high capability of feature extraction and nonlinear approximation,it has complex structure,slow modeling speed and high computational cost.The broad learning system extracts features through nonlinear transformation,and has good nonlinear approximation ability.At the same time,the output weight of the network is obtained by pseudo-inverse,which greatly reduces the calculation time of the network.At present,broad learning system has been applied in the fields of food safety,construction engineering,biomedicine and education and teaching,but the research on broad learning system in the field of photovoltaic power prediction is relatively small.Based on this,this thesis systematically studies the application of broad learning system in the field of photovoltaic power prediction.The results show that the broad learning system based on genetic algorithm optimization and the broad learning system based on K-means clustering algorithm can achieve high efficiency and high precision prediction of photovoltaic power generation in the short term.The main contents and results of this thesis are as follows:(1)Analyze the current research and existing problems of domestic and foreign scholars on photovoltaic power generation,and explore the feasibility of broad learning system in the field of photovoltaic power generation prediction.(2)Analyze the characteristics and existing laws of photovoltaic data,and explore the impact of various meteorological factors on the output power of photovoltaic power generation.The missing and abnormal values of the original photovoltaic data are processed to improve the data quality.Pearson correlation analysis was used to screen the input characteristics of the model.(3)The broad learning system prediction model(GA-BLS)based on genetic algorithm is established,and the actual photovoltaic data is used for experimental verification.The experimental results show that GA-BLS can effectively predict the output power of photovoltaic power generation in different seasons and different weather conditions.The GA-BLS model is compared with the traditional prediction algorithm convolution neural network(CNN),short-term and short-term memory neural network(LSTM),lightweight gradient lifting machine(LGBM)and support vector machine(SVM).The results show that the average absolute error(MAE),mean square error(MSE)and root mean square error(RMSE)of the GA-BLS prediction results are 0.1175,0.0603 and 0.2455,respectively,which are the lowest of the five models.In addition,GA-BLS also has high computational efficiency while ensuring high prediction accuracy.Compared with CNN,LSTM and SVM,the training time of GA-BLS decreased by 3 orders of magnitude,only 2.3118s.(4)The influence of different weather on photovoltaic output power is different,a broad learning system prediction model based on K-means clustering is proposed.The K-means clustering algorithm is used to cluster the data into three data sets:sunny,cloudy and rainy days,and then the corresponding broad learning system is established for training and prediction.The CNN,LSTM,SVM and single BLS prediction models are established for comparative experiments.The experimental results show that the prediction accuracy of the method proposed in this thesis in sunny,cloudy and rainy weather is also improved compared with the single BLS model,and the prediction error evaluation index MAE is reduced by44.88%,7.45%and 0.76%respectively.Compared with the single BLS model,the prediction effect of the model in this thesis is not significantly improved in rainy days.This may be because the original data set contains less data of rainy weather types,and the model training is not sufficient,resulting in little improvement in prediction accuracy.In addition,the training time of CNN,LSTM and SVM is 2768.0225s,3373.1757s and 20345.2809s respectively,while the total training time of the model in this thesis is only 7.2134s,which greatly reduces the calculation cost.While ensuring high computational efficiency,the prediction accuracy of the model proposed in this thesis is also improved compared with CNN,LSTM and SVM.Under the three weather types of sunny,cloudy and rainy days,the determination coefficient R~2 of the prediction results of the model in this thesis is 98.41%,94.06%and 80.40%respectively,which are the highest among the four models.