Short-term Power Load Forecasting of Deep Learning Based on Similar Daily Clustering and Time Correlation
To address the problems of insufficient feature extraction and high noise of load data,a multi-branch combined power load forcasting method based on multi-factor similar daily clustering,time correlation analysis and two-layer decomposition for noise reduction is proposed.Firstly,pearson correlation coefficient(PCC)and maximum information coefficient(MIC)methods are used to comprehensively analyze the linear and nonlinear correlations of daily load influencing factors,and to enhance the selecting of important features.The selected high correlation meteorological factors,date factors and 24-hour daily load data are dimensionally reduced by principal component analysis(PCA)method,and then subjected to K-medoids similar daily clustering.Secondly,multi-dimensional analysis and multi-feature extraction are carried out on the load,weather and date datas of similar days in each cluster,and multi-feature extraction matrix blocks are constructed to enhance the periodic rule and spatio-temporal characteristics of the data.Combining variational mode decomposition(VMD)and empirical wavelet transform(EWT)algorithms to extract the multi-scale fluctuations rule of the original data,increase the details of the data,and reduce the nonlinear degree of the data at the same time.The residual gated convolution module of different input branches in the combined forecasting model is used to fully explore the local correlation of the data,extract the local short-term dependence,and obtain high-dimensional features.The bidirectional long short-term memory(BiLSTM)network with parallel input branches is used to extract the time series features of the data and explore the long-term dependency relationship.Finally,different types of features are integrated and strengthened comprehensively to achieve short-term power load forecasting.The experimental results show that the proposed method of multi-feature extraction and multi-model combination can obtain higher forecasting accuracy in short-term power load single-step forecasting.In multi-step forecasting,the forecasting method can greatly improve the forecasting accuracy.The proposed method has excellent overall forecasting performance.
power load forecastingsimilar daily clusteringtime correlation analysisgated convolutional networkself-attention mechanism
万方数据
维普
