Research On Short-term Residential Load Forecasting Method Based On LSTM Combination Model

As an important part of smart grid,load forecasting provides information support for power grid and power supply construction planning,and provides data basis for power grid dispatching and power consumption decision-making of power consumers.It is one of the important tasks to ensure the safe operation of the power system,energy saving and emission reduction,and improve efficiency.With the further development of smart grid function,the object of load forecasting is continuously subdivided,coupled with the improvement of power grid communication system and the popularization of information acquisition equipment,the forecasting object extends from system level load to residential level load.Compared with system level load,residential level load has higher nonlinearity,and the increase of time-varying load demand in distribution network makes the complex residential level load data more difficult to predict.How to improve the accuracy of residential load forecasting has become the focus of many scholars.With the growth of computer computing power,some data processing methods that are difficult to realize can be applied to load forecasting.In recent years,the rise of deep learning provides more methods for load forecasting.Its powerful nonlinear processing ability makes it gradually replace the traditional load forecasting model and become the mainstream.Aiming at the problem of high nonlinearity and difficult prediction of residential load,a multi-objective optimal combination model based on long short-term memory(LSTM)network is proposed in this paper.Based on LSTM,the model constructs a deep learning network,supplemented by multi feature input to improve the accuracy of the network.Multi objective evolutionary algorithm based on decomposition(MOEA/D)is used to optimize the parameters of LSTM network,and multiple high-quality sub models are obtained.The combined output strategy based on deep belief network(DBN)is used to combine the outputs of multiple sub models to form combined prediction,so as to improve the generalization ability of the prediction model.The specific work contents are as follows:(1)Summarize the characteristics of residential load.Based on the residential user load data provided by the smart grid smart city(SGSC)project in Australia,this paper compares the residential load with the system load,highlights the characteristics of the residential load,and explores the key optimization direction of the prediction model.(2)Build the data preprocessing framework based on multi feature input.Because there are many factors that affect the change of residential load,in order to make the prediction model fully mine the potential correlation in the data,this paper uses the multi feature input composed of power consumption component,time component,week component and holiday component.We hope to increase the hidden knowledge that the prediction model can learn by adding the dimension of input data,so as to improve the prediction accuracy of the model.Based on the data provided by SGSC,the experiment is carried out by combining the multi feature input framework with LSTM network.Compared with the traditional prediction model and deep learning model,it is proved that the multi feature input framework can improve the overall accuracy of the prediction model.(3)The network optimization strategy based on MOEA/D is proposed.In order to improve the overall performance of the network,MOEA/D algorithm is used to optimize the parameters of the network.Taking the accuracy and difference of LSTM network as the target space and the super parameters of LSTM network as the decision space,the network is optimized.In this adaptive way,the network super parameters can automatically adapt to the load data,avoiding the error caused by artificially setting parameters or setting fixed parameters.At the same time,in order to make the MOEA/D algorithm more consistent with the deep learning framework,this paper improves the MOEA/D algorithm,and illustrates the effectiveness of the improvement through comparative experiments.(4)The combined strategy based on DBN is proposed.The DBN network is used to combine the prediction results of multiple sub models into the final prediction results.Through this method similar to ensemble learning,we can overcome the unstable factors in the prediction of a single network,so as to alleviate the effect of data over fitting and improve the generalization ability of the model.This paper compares the multi-objective optimization model based on long term and short-term memory network with the existing multiple intelligent prediction models to illustrates the advanced nature of the model.