LSTM-GRU Combined Short-term Load Forecasting Based On Improved Dimensionality Reduction Clustering-decomposition Strategy

Improving the accuracy of power load forecasting is of great significance for optimizing generator set configuration and reducing system operating costs.In order to effectively improve the accuracy of load forecasting,this paper introduces the idea of clustering based on main characteristics to achieve horizontal clustering of load time series;Designed a strategy based on dimensionality reduction clustering and VMD decomposition to realize the longitudinal decomposition of the load after clustering;Based on the clustering-decomposition results,a short-term load forecasting method based on LSTM-GRU combination is proposed,and the effectiveness of the proposed method is verified through actual calculation examples.The specific research content is as follows:Firstly,in view of the difference of users’ electricity consumption behavior and the strong randomness of load sequence,the idea of clustering is introduced,and a load clustering method based on main characteristics is proposed.Firstly,the load data is preprocessed,and the singular value decomposition is used to extract the main characteristics of the load,which is used as the input parameter of the clustering;The K-means algorithm based on selecting the number of clusters realizes the horizontal clustering of user load sequences.Validation of a calculation example through the power load data of a certain district shows that the load clustering method based on the extraction of main characteristics can obtain differentiated power consumption characteristics and reduce the difficulty of short-term load forecasting.Secondly,aiming at the problem of large forecast errors caused by the nonlinearity and instability of the load sequence,an improved forecasting strategy based on dimensionality reduction clustering-decomposition is proposed.By studying and comparing various decomposition principles,the variational modal decomposition is applied to the decomposition of the load sequence after the clustering,so that each type of load is decomposed into a single frequency component.Realize the longitudinal decomposition of the load sequence,so as to provide a good data set for load forecasting.Through VMD decomposition of each type of load under a certain station area,the results show that the frequency of the decomposed load sequence is single,and the inherent change law of the load is extracted,so that the influence between different components is greatly reduced.Thirdly,considering that LSTM has a better prediction effect on the time series of low-frequency components,and the GRU neural network has a higher degree of fit for nonlinear high-frequency sequences,this paper proposes the LSTM-GRU combined prediction algorithm.After clustering-decomposition,the high and low frequency components of each load are respectively predicted,then the predicted value of regional users’ power load is obtained by integrating the predicted results of each component.Validation of a calculation example through the power load data of a certain station shows that the proposed method has high prediction accuracy.Finally,in order to further verify the effectiveness of the method proposed in this paper,the actual power load data of a province in Northwest China were used for example verification,and five kinds of comparison schemes were designed to compare the prediction error indexes.The results show that the proposed prediction algorithm has absolute superiority in the prediction accuracy,and can provide a general idea and method for the prediction of random time series.