Power Load Forecasting Based On Deep Learning

With the development of technology,the living standard of people are being continuously improved,electricity power is becoming increasingly important for Human beings,No matter daily life or industrial production electricity power is influencing every aspect of human life.Power load forecasting is one of important tasks of electric power department.The routine production and daily dispatch of electric power depend on the forecast result and the accuracy of electric load forecast directly affects the productivity effect of electric department.In recent years,the development of social production demands higher precision of electric forecast,and appropriate forecasting methods are more and more important.Therefore,a variety of forecasting methods emerged.And among these methods,neural network model received much concern.After many years of development,the standard neural network model has derived many branches,among which recurrent neural network developed rapidly and gained great success in NLP.But,because of the shortcoming such as easily over-fitting,gradient disappearing,and the failure in learning long term dependence,the development of recurrent neural network is restricted.After being modified,the LSTMs model was raised on the basis of recurrent neural network and it solved the gradient disappearing problem.In this thesis,we first explain the significance and necessity of studying power load forecasting,and introduce the concept and classification of power load forecasting,list the principal factors which influence power load forecasting.Then,starting with the perceptron,we introduced the history and the origin of LSTMs model,sketched the model theories and merits and demerits of the LSTMs model in each of its developmental stages.Particularly,in this thesis,we emphasis on the network structure,calculation principle,and process of the LSTMs model,combining with graphics.Then we introduce the model gate structure,and explain how the model deal with the gradient disappearing and long term dependence problems.At the end,we conduct an empirical analysis with the historical date of Poland electric power company.In order to evaluate the forecasting result,the time-series model and LSTMs model are established respectively.When comparing these two models,the simple average of time-series model and LSTMs model are considered and the simple seasonal model is regarded as the reference of the both models.The result of comparison proves the applicability and validity of the LSTMs model.In the end,we summarize the insufficiency of the model and propose some possible methods to improve it.