Design And Implementation Of Power Load Forecasting System Based On Deep Learning

With the rapid growth of power load and the increase of renewable energy,the userside power load data has strong volatility and nonlinearity.Because the load series is extremely volatile and nonlinear,extracting features and training prediction models is more difficult.To address the issues raised above,we propose a power load forecasting model based on deep learning.First,we use variational mode decomposition(VMD)to decompose the highly volatile load sequence into a stationary sequence and a nonstationary sequence with different characteristics.Then a deep bidirectional long shortterm memory neural network(DBi LSTM)is used to efficiently learn nonlinearity and accurately predict temporal sequences.Finally,we designed a power load forecasting system based on the deep learning model following the software engineering design method.The main works in this thesis are as follows:(1)Data preprocessing.On the one hand,there are missing and abnormal values in power load data;on the other hand,the load has strong volatility.First,the data is filled with missing data and abnormal data is processed.Then we use variational mode decomposition to decompose the preprocessed data into stationary and non-stationary sequences with different features.In the VMD process,the number of decomposed modal functions affects the decomposition effect.We adopt the method,the ratio of residual energy after mode decomposition,to determine the number of VMD decomposed modes.And the subsequences of feature decomposition are smoother and more regular by the above method.At the end of the data preprocessing chapter,it is verified through experiments that data preprocessing has a good effect on improving the prediction results.(2)Building deep learning model.The load sequence not only has strong volatility,but its nonlinearity is complex and difficult to train.The train process of load sequence is difficult and complex due to its strong volatility and nonlinearity.Therefore,a new power load forecasting method is proposed to solve this problem and based on deep learning.First,the input of the model is the component sequence decomposed by VMD,and a neural network,deep bidirectional long short-term memory,is used as a prediction model to learn the information in nonlinear sequences.Then,the attention mechanism(AM)is used to assign weights to important information,which makes the prediction results more accurate.Finally,an improved particle swarm optimization(IPSO)algorithm is used to optimize the hyperparameters of the DBi LSTM model since the parameters of the DBi LSTM model are difficult to select.At the end of the chapter,the public dataset is used to test the model and the real dataset to verify the prediction effect,then multiple existing models are compared with our model to highlight the excellent prediction results of it.(3)Load forecasting system.We use the idea of software engineering to carry on the system demand analysis and function design and complete the implementation of the power load forecasting system.Finally,we do system tests on the core functions.