Research On Short-term Power Load Forecasting Methods Based On Deep Learning

Power load forecasting plays a vital role in power system dispatching planning and it is the basis of economic operation of power system.With the rapid development of China’s economy and science and technology,people’s living standards continue to improve,industrial and residential electricity demand is also growing rapidly.In recent years,the types and quantities of load connected to the grid have been increasing,especially the access of a large number of distributed renewable energy and new loads such as electric vehicles,which makes the change of power load in various regions more random,and puts forward higher requirements for the accuracy and timeliness of power load forecasting methods.This paper studies the regional short-term power load forecasting methods,and the work done is as follows:Firstly,this paper analyzes the current research status in the field of regional power load forecasting at home and abroad,and summarizes the factors that affect the change of power load,and improves the methods of traditional power forecast feature engineering.Power load forecasting features usually include internal features such as similar days and historical load sequences,as well as external features such as date,weather,temperature,humidity,and air quality.Based on the actual data set,this paper selects 23-dimensional power load forecasting features,uses Pearson’s correlation analysis method to initially screen the features,retains the features that are highly correlated with load changes to form the initial feature set,and then uses principal component analysis to analyze the initial features.The dimensionality reduction process of the collection is performed to explore the smallest dimensional feature representation that can characterize the load change,thereby improving the generalization ability of the forecasting model.Secondly,this paper analyzes the advantages and disadvantages of traditional power load forecasting methods.Based on the advantages of convolutional neural networks and recurrent neural networks in deep learning,this paper proposes a CNN-BiGRU hybrid neural network prediction model based on the Attention mechanism,and uses the CNN network to extract the relationship between prediction features and load data in high-dimensional space,and then constructs the high-dimensional feature vector of the time series sequence.Input the above results into the BiGRU network and get the prediction results.At the same time,the CBAM attention mechanism is added to the CNN network,and the self-attention mechanism is added after the BiGRU network,so that the model can focus on important information in the load characteristics.The above measures can effectively improve the performance of the model when dealing with timing fitting problems.Finally,based on the load data of a certain city in Zhejiang Province provided by the Ninth Electrical Engineering Mathematical Modeling Competition,this paper conducts the performance comparison experiments of the base model,the CNN network combination model,and the attention mechanism combination model.First of all,comparative experiments are carried out on base models such as random forest,SVR,BP neural network,BiLSTM and BiGRU network.The results show that BiLSTM and BiGRU sequence networks have better prediction accuracy.Secondly,a pair of CNN-BiLSTM and CNN-BiGRU models are built.The performance of the CNN combined model is analyzed,and the result shows that the combination of CNN and sequence network can significantly improve the prediction accuracy of the model.Finally,three groups of attention mechanism models including the model proposed in this paper are built for comparison,and the results show that the proposed model in this paper has significant advantages on accuracy and efficiency of load forecasting.The research results of this paper have certain reference value and practical value for improving the accuracy and timeliness of power load forecasting.