| Short-term power load forecasting refers to the prediction of load changes in the next one hour to one week,and its purpose is to provide a reference for the daily dispatch and market transactions of the power system to ensure the stable operation of the power system.However,the randomness and uncertainty of short-term power loads make it extremely difficult to predict them accurately.In order to improve the prediction accuracy of short-term power load,this thesis conducts in-depth research on the combined forecasting model from the perspective of deep learning,and the main research content is as follows:(1)In order to solve the problem of insufficient extraction of short-term power load features by traditional single models,this thesis proposes an MGRF residual fusion prediction model based on Multilayer Perceptron(MLP)and Gated Recurrent Unit(GRU),which can pay more attention to the important details of short-term power load,so as to better capture the fluctuation of power load.Firstly,the model uses the nonlinear mapping capability of MLP to extract the backbone features of short-term power loads,and obtains more detailed residual features by separating the backbone features from the original data.Then,the powerful time series processing ability of GRU is used to extract the time series information between the residual features,and finally the features extracted by MLP and GRU are fused to obtain the final predicted value.Experimental results show that compared with single MLP and GRU models,the MGRF residual fusion model not only has higher prediction accuracy,but also converges faster.(2)In order to extract the time series characteristics of short-term power load data more fully,this thesis introduces Variational Mode Decomposition(VMD)on the basis of MGRF residual fusion model to construct a combinatorial prediction model based on deep learning,which can adopt different prediction strategies according to the frequency characteristics of different components of VMD.For the low-frequency components,a DD-MGRF residual fusion model is proposed,which separates the trend information and seasonal factors from the sequence by performing first-order differentiation of the low-frequency components,so as to improve the prediction accuracy of the low-frequency components and accelerate the convergence speed of the model.In order to make better use of the periodicity of the IF component,a PDD-GRU model is proposed,which uses periodic difference to optimize the GRU network,so as to improve the prediction efficiency of the model for the IF component.For the high-frequency components,a bidirectional liquid time-constant networks(BiLTC)is proposed,and the attention mechanism is used to weighted and fuse the forward and reverse time series information to effectively extract the strong nonlinear and nonstationary implicit information in the high-frequency components.In order to obtain the best timing information of different components,multi-scale sliding windows are used to fuse each component to further improve the accuracy and stability of prediction.The experimental results on the real data set of power load in Shaanxi Province show that the prediction accuracy of the proposed deep combination model is 99.939%,which is 0.524%、0.412%、0.148%and 0.234%higher than that of LTC,MGRF,VMD-GRU and VMDMLP-GRU methods.The main contribution of this thesis is to propose the MGRF residual fusion model and the combined prediction model based on deep learning,which can better adapt to the high nonlinearity and non-stationarity of power load data and achieve more accurate short-term power load forecasting,in order to provide reference and reference for short-term power load forecasting. |
