Research On Short-term Power Load Forecasting Based On Long And Short Term Memory Neural Network

The power sector is related to the economy and development of a country,and it is the supporting sector of a country’s science and technology,economy,military,education and other fields.As the society changes with each passing day the modernization process of rapid development,economic development and residents’ life growing demand for electric power,the power supply reliability of the power sector and power quality put forward higher request,therefore can no longer in the traditional model of electricity power system control,and intelligent control should be used.It is of great engineering value and theoretical significance to realize the intelligent power system to ensure that the power sector can run economically,safely and stably with high efficiency,which can not only maintain the normal operation of the city’s survival function,but also optimize the allocation of resources and relieve the pressure of energy.The realization of intelligent power system depends on the power load forecasting model with high prediction accuracy.The higher the accuracy of power load forecasting,the more powerful the basis for the correct dispatching decision of the power system.However,the traditional power load forecasting method still has the following challenges:(1)Traditional power load forecasting methods have poor adaptability to the historical power load data with nonlinear input-output relationship under the influence of multiple factors.(2)Traditional power load forecasting methods have problems such as unsatisfactory prediction accuracy.Compared with the traditional prediction methods,the neural network method has strong adaptive ability and learning ability,and its flexible structure can well capture the complex interdependence relationship between multiple influencing factors and historical power load data,especially in the feature separation and extraction of data and abstract learning.In order to improve the accuracy of short-term power load prediction and solve the above problems,relying on the strong learning ability and adaptive ability of the artificial neural network,considers many weather factors affecting the power load,and selects the Long Short Term Memory neural network algorithm(LSTM)which is sensitive to the time series data of power load data.In order to improve the prediction accuracy of LSTM,Attentional Mechanism(AM)is added to the original LSTM.At the same time,an improved Attentional Mechanism-Long Short Term neural network(AM-LSTM)model was proposed,and then combined with the power load data of a certain place to optimize and adjust the model parameters.Through experiments and comparison with the prediction results of AM-LSTM,the experimental results show that the improved AMLSTM short-term power load prediction model has better prediction results.In order to further improve the prediction accuracy of the model,the improved AM-LSTM model is used to combine with the Residual Network and Convolutional Neural Network Model(R-CNN)in parallel based on ensemble learning of bagging.The combination forecasting model,which has been proved by experiments to effectively improve the accuracy of short-term load forecasting.The main research work and research results of this paper are as follows,(1)This paper studies the internal variation law of power load data in a certain place for 3 years and the external factors that may affect it,and analyzes which of these external factors,such as temperature,humidity,rainfall,etc.,can most affect the accuracy of load prediction,which provides a theoretical basis for load prediction.What’s more,the maximum temperature,the daily minimum temperature,the daily average temperature and the daily maximum load,the daily average load,the daily minimum load and the daily total load with higher correlation were taken as the characteristics to improve the prediction accuracy of the model.(2)An improved Long Short Term Memory neural network based on Attentional Mechanism model is proposed,which is referred to as the improved AM-LSTM model.Long Short Term Memory(LSTM)algorithm model is sensitive to the order of time series data,so LSTM model is adopted to predict short-term power load.In order to improve the accuracy of the model,attention mechanism is added on the basis of LSTM model.In order to output the state information of LSTM cell better,the Tanh layer through which the cell information passed was changed into the weight-based activation function group based on Tanh function,Sigmod function and Re LU function.The experimental verification showed that the short-term power load prediction accuracy based on the improved AM-LSTM model was higher.(3)A combination prediction model which combining the improved AM-LSTM model with the R-CNN model was proposed.With the development of economy,the factors affecting power load increase and there are some uncertain factors,but single improved AM-LSTM model in dealing with many factors and high dimensional data have some difficulties,in order to overcome these difficulties,to join the R-CNN model.Based on the idea of ensemble learning,the improved AM-LSTM model is combined with the R-CNN model through weighted average strategy.The combined model can improve the prediction accuracy of the short-term power load prediction model.