Research Of Regional Grid Load Forecasting Based On Deep Learning Neural Network

Load forecasting is the key to the development of the grid,as well as its operation and maintenance.The accuracy of forecasting is the prerequisite for ensuring the stable operation of the grid and the stable development of the regional economy.However,insufficient model accuracy has become a major factor influencing load forecasting results in reality.Therefore,it is of great theoretical significance and engineering value to achieve more accurate and efficient load forecasting through the study of improving model accuracy.The content of this article is as follows.1.The concept and method of power load and its forecasting are described,as well as the model structure and solving process of BP neural network.The basic principles of deep learning algorithms are introduced,which can train multi-layer network structures through a large number of sample data so that the fundamental characteristics of the data set can be obtained.Thereby the accuracy of data processing can be effectively improved.For the situation that the previous neural network cannot meet the training of deep learning.A layer-by-layer training method is used to divide the training process into two processes,which are the bottom-up unsupervised learning process and the up-bottom supervised learning process,to make the performance of the network more optimized.The three basic methods of deep learning of DBN,CNN and AE are deeply analyzed,as well as their calculation process.2.Based on the single-day average power load data,single-day average temperature data,and holiday and non-holiday load average data changes within two years in Weifang area,and the correlation between the date,holidays,non-holidays and temperature and the power load is found though Matlab simulation.These factors above are influencing factors of load forecasting.Considering the fact that accurate temperature data within the month to be measured is difficult to obtain,a BP network model is constructed and trained to predict the temperature in a specific time period.The network architecture and training process of DBN model are described,working day and holiday information are processed by binary encoding method.Taking the number of network layers,input layer nodes and hidden layer nodes as the starting point,the DBN model is constructed based on the selected parameters and used for load forecasting according to the influence of temperature,and the results are analyzed.3.Considering that current computer technology can not meet the demand STLF for the sake of power load data being complex and large,as well as the low prediction accuracy of DBN model,a load forecasting method of deep learning algorithm based on Spark parallel computing framework is proposed after in-depth analysis of parallel data analysis based on the Spark parallel computing framework and parallel model based on deep learning,which can improve operating efficiency.The K-means clustering algorithm in the parallel computing environment of the Spark parallel computing framework is used to classify the input historical load data to achieve data feature extraction.The information entropy theory is used to calculate the number of hidden layer nodes of the deep learning model.The DBN model is used for load classification,and the result is used as the input element of the deep learning load predictor.The SAE deep learning algorithm is selected for STLF,the KL divergence is used to sparsely constrain the objective function,and the parameter of SAE is optimized by the minimum value of the error function.Thereby the efficiency and accuracy of STLF are improved.The load data in Weifang area is used as sample for simulation,load forecasting results of three models of BP network,DBN and deep learning based on Spark parallel computing framework are compared,thus the effectiveness of the proposed method is verified.In summary,the efficiency of the load forecasting method of deep learning algorithm based on Spark parallel computing framework is improved through data parallelism and model parallelism.The increase in load forecasting data sets makes greater use of multiple relevant parameters of load data to effectively extract the features of massive data and improve the efficiency and accuracy of STLF.