Research On Short-term Power Load Forecasting And Load Curve Clustering Based On Machine Learning

Short-term power load forecasting provides valuable reference for power dispatching and power generation planning.With the development of power big data industry,the data volume and data dimensions of the power system continue to grow,and traditional load forecasting methods have therefore exposed more limitations.In recent years,a variety of machine learning methods stand out in load forecasting cases due to their powerful learning capabilities.This paper will use multiple machine learning algorithms to solve the load forecasting problem,and combine actual data to study how to improve the accuracy of short-term power load forecasting,and propose a complete load forecasting system.The main research contents include:(1)This paper introduces the data preprocessing methods,summarizes the influencing factors of the load,and extracts the corresponding features.For the purpose of making the generalization performance of the forecasting model better and speeding up the model training,correlation coefficient matrix and gradient boosting decision tree are applied to select important features and eliminate redundant and invalid features.(2)In order to explore the electricity consumption patterns of different users and improve the model's ability to learn electricity consumption patterns,this paper adopts a forecasting strategy of clustering users and then building a load forecasting model for each cluster of users.According to this strategy,a load curve clustering algorithm based on improved ISODATA is presented,which is called L-ISODATA(Load curve-ISODATA).The ISODATA clustering algorithm is improved from the perspectives of the selection of initial clustering centers and the nonlinear mapping of the kernel method.The selection strategy of initial clustering centers is optimized to accelerate the algorithm convergence speed,and the kernel method is used to capture the features of the load curves in high-dimensional space.A clustering model of electricity users based on the L-ISODATA clustering algorithm is presented,and compared with the traditional K-means algorithm and ISODATA algorithm.(3)Single load forecasting method has limitations,in order to improve the prediction accuracy of the forecasting algorithm in different cases,this paper proposes an integrated load forecasting model.First,a load forecasting model based on Cat Boost is presented.In order to solve the problem that Cat Boost does not consider time series features,the model combines the Cat Boost algorithm with historical load features.Then,a multi-layer LSTM load forecasting model is established.Subsequently,the integrated learning is adopted,and a weighted ridge regression algorithm based on time interval is presented to combine the advantages of the two algorithms to construct an integrated load forecasting model.The model is compared with other models through experiments to prove the superiority of the proposed model.For each cluster of users,an integrated forecasting model is established.The experimental results prove that the prediction accuracy after clustering is better than that of non-clustering when using the same load forecasting model.