Research On Voltage Forecasting Of Distribution Networks With High Proportion Distributed Photovoltaics Based On Deep Learning

Under the dual pressure of deteriorating environmental problems and the shortage of traditional energy,the capacity of grid-connected photovoltaic power has increased rapidly,and the proportion of residential distributed photovoltaics(PVs)connected to some low voltage distribution networks has become higher and higher,resulting in voltage fluctuation and limit violation,the problem is also getting worse,and it is difficult to solve it by relying only on traditional voltage regulation.Considering the certain regularity of photovoltaic power,predicting voltage trends from the perspective of data driven provides a new idea for the solution of this problem.Due to the advancement of artificial intelligence technology and its flexible application in power systems,based on an actual project,this paper studies methods of voltage foresacting of distribution networks with high proportion PVs based on deep learning.Firstly,the working principle of the photocoltaic power system and its inverter control strategy are introduced.According to the grid-connected structure of residential distributed PVs and its operation mode,the influences of increasing penetration of distributed grid-connected PVs is analyzed,it mainly includes the problem of voltage fluctuation and limit violation,which makes it difficult for traditional voltage adjustment strategies to work.It points out the necessity of voltage foresacting from the data level.Secondly,in order to provide a reliable sample data set for voltage foresacting,the historical voltage and power data which can reflect the trend of voltage are filled and nondimensionalized;the actual operating voltage data of distribution networks with high proportion of distributed PVs are analyzed for the time features,and the time features at different scales were obtained and processed discretely;For reducing the complexity of the model,the e Xtreme Gradient Boosting(XGBoost)algorithm is used to select features of the input feature vector,and the feature weights of each dimension are used to filter.At the same time,the complete hour feature contains time transition features,maked different feature reduction strategies to provide multiple feature subsets for foresacting models.Thirdly,considering the superiority of the Long Short-Term Memory(LSTM)network in time series foresacting,this paper taked the lead in using LSTM to predict voltage.The feature subsets are constructed by different feature reduction strategies,the model is trained using the back propagation algorithm,and the voltage foresacting results are output.The analysis of case study shows the rationality of different feature reduction strategies,and comparison with traditional foresacting models at different time scales,which verifies the effectiveness of LSTM in voltage foresacting.Finally,in order to further improve the accuracy of voltage foresacting,aiming at the problems of LSTM without considering the data structure characteristics and fully extracting the changing characteristics of historical data,learned form the temporal convolutional network(TCN)proposed in the field of natural language processing,the feasibility of TCN for solving the voltage prediction problem is analyzed from the perspective of model structure and actual situation of the input features;Combining the channel structure of the dilated causal convolutional layer with historical data features,and taking full advantage of time feature,a voltage foresacting model architecture based on TCN is proposed in this paper;through analysis of case study,it can be seen that the TCN model is more sensitive to time features in the selection of feature reduction strategies,in terms of voltage foresacting in different seasons and at different time scales,its overall prediction accuracy has reached a new level compared with LSTM.The rationality and effectiveness of the proposed model architecture in voltage foresacting are explained from the result.