Power Load Forecasting Based On Regional Power Consumption Unit Energy Similarity And Individual Feature Mining

Regional load forecasting technology based on machine learning and power big data is an important content of intelligent power consumption and a key step in intelligent demand response.Regional load has many influencing factors and strong volatility.When faced with large load changes,it is difficult to achieve accurate forecasting by only considering the load timing characteristics for mining;At the same time,the load forecasting of individual units can help to accurately push demand response strategies and promote the smooth operation of regional power systems.However,the uncertainty of individual unit power consumption behavior is large,and the power consumption laws are difficult to capture.Moreover,the large number of regional units leads to high modeling costs.In response to this,this thesis studies accurate forecasting strategies for regional power system loads and individual unit loads.The main research contents include the following two points:(1)Power load forecasting based on similar power consumption units and graph convolution neural networks: Firstly,cluster the regional power consumption units according to the historical data of the regional load to obtain groups of similar power consumption units of each category;Then,a self-supervised learning method is designed for each cluster of unit group to mine similar association relationships between units,obtain a similar power consumption matrix,and construct a similar unit graph for each cluster based on the matrix;Finally,a graph convolution neural network is used to mine load spatial features based on similar power consumption unit diagrams,and a long short-term memory network is used to extract temporal features from the extracted features to achieve regional power load forecasting.The proposed algorithm is applied to the load forecasting of UCI apartment load data and Irish residential smart meter data,and the experimental results show the effectiveness of the algorithm in regional power system load forecasting.(2)Accurate power load forecasting based on individual energy consumption behavior feature mining and model sharing: Firstly,aiming at the characteristics of high randomness and high prediction difficulty of individual energy consumption behavior in regional units,clustering based on the power characteristics of power consumption units is conducted to obtain similar power consumption unit groups for classification prediction;Then,for each cluster of unit group,the graph convolution network is used to obtain similar unit aggregation features,and the full connect neural network is used to obtain external features.The two types of features are combined with the original load characteristics.The shared long short-term memory network layer is used to sequentially extract individual unit features to achieve preliminary load prediction for individual units within the cluster;Finally,a full connect neural network is designed separately for each individual unit and joint prediction is performed to complete the final load prediction for each individual unit.The proposed algorithm is applied to residential load forecasting for Irish smart meter data,and experimental results verify that the algorithm can effectively achieve efficient and accurate load forecasting for individual regional units.This research focuses on solving various shortcomings in regional power system load forecasting.Firstly,a prediction model framework based on clustering algorithm and spatiotemporal feature mining is proposed for regional power system integrated forecasting.Secondly,in the face of individual load forecasting for regional units,deep learning is considered to obtain and fuse multiple features of units to express load characteristics,A load forecasting method based on shared model feature mining and individual model joint forecasting is proposed,and the feasibility and effectiveness of the algorithm in this thesis is verified by an example application.The thesis contains 25 figures,10 tables,and 86 references.