Solar Irradiance Prediction Based On Seasonal Classification And Bidirectional Long Short-term Memory Neural Network Hybrid Model

With the exploitation and use of traditional fossil energy,fossil energy is increasingly depleted and the ecological environment has been repeatedly destroyed.At the same time,mankind’s demand for energy is increasing day by day.If it is still over-reliant on fossil energy,it will lead to energy crises in various countries around the world.Competition will also threaten the peace and stability of the world,so the exploration of new energy is imperative.Solar energy has been widely concerned because of its unlimited,environmental protection and pollution-free.In recent years,the installed capacity of photovoltaics has been increasing rapidly,and the share of photovoltaic power generation in the power grid has also been increasing year by year.Because solar energy is affected by many factors which has strong randomness and instability.It is difficult to predict,which has a greater impact on photovoltaic power generation,grid connection,and grid stability and security.The power of photovoltaic power generation is most directly and significantly affected by the solar irradiance on the ground.Therefore,the efficient and accurate prediction of solar irradiance is of great significance to photovoltaic power generation and the national new energy strategy.The solar irradiance is affected by factors such as geography,climate and weather,which has strong seasonal periodicity,day and night periodicity and obvious characteristics,such as randomness and instability.In recent years,with the rapid development of deep learning,long and short-term memory neural network models(LSTM)have been widely used in solar irradiance prediction and achieved good results.However,the prediction results of the LSTM model have obvious lag and many traditional prediction methods don’t consider the influence of the season on the predicted value of solar irradiance,which make the prediction accuracy of the model unsatisfactory.this paper proposes a bidirectional long short-term memory network model which based on seasonal classification(CS-BILSTM)and use batch normalization(BN)algorithm optimization it.Establishing BN-CS-BILSTM model for Predict the value of solar irradiance.The specific research content is as follows:1)The characteristics and influencing factors of solar irradiance are discussed.Preproceing the data set to improve data quality.2)The data set is classified by seasons,and the correlation analysis and correlation coefficient matrix are used to screen out several influencing factors with the highest correlation with solar irradiance in each season.3)A comparative analysis of the advantages and disadvantages of long and LSTM and bidirectional long short-term memory network model(BILSTM)neural network in predicting solar irradiance,and confirming that this paper uses BILSTM neural network model for solar irradiance prediction.4)Establishing a bidirectional long and short-term memory neural network model based on season classification which extract the past and future time series characteristics of the data set and use the high correlation influencing factors extracted from each season as the input of each season’s corresponding model.Training the structural parameters of the BILSTM neural network model.5)Compared with the LSTM and BILSTM neural network models without seasonal classification,the results show that the average value of the prediction error MSE of the BILSTM model in the four seasons is about 0.641,which is about 18.8% less than the LSTM model.CS-BILSTM reduces about 4.1% compared with BILSTM model.6)Aiming at the problem that the neural network model is prone to reduce the prediction accuracy and increase the time cost due to the random change of the input value distribution during the training process.This paper uses the BN algorithm to further optimize the CS-BILSTM model and construct the BN-CS-BILSTM model is compared with the CS-BILSTM model.The prediction results of the BN-CS-BILSTM model in the four seasons are only 0.582,which is about 5.3% less than the CS-BILSTM model while the training time cost has been reduced by about 11.7%.The results show that the BN-CS-BILSTM model is superior to other models in forecast accuracy and time cost in various weather and seasons.Further improving the prediction accuracy of solar irradiance is of great significance to the power dispatching of relevant departments and maintaining the stability of the power grid.The BILSTM model overcome the shortcoming of some traditional prediction models that cannot consider the forward and backward information of time series data at the same time.The CS-BILSTM model based on seasonal classification improves the prediction accuracy of each season,making the model for the overall solar irradiance the prediction accuracy has been further improved.The BN-CS-BILSTM model better solves the impact of the random change of the input value distribution on the model prediction accuracy and convergence time,which improves the model’s performance Practicality.