Research On Residential Customer Pattern Clustering And Short-Term Load Forecasting Based On Smart Meter Data

To meet the peak demand of residential electricity consumption,the potential solution is demand response.The powerful bi-directional communication mode of smart meter enables efficient load management and accurate demand assessment for utility companies which is an important support for demand response.By setting up the Clustering model,utility companies can realize load management of residential customers.By further realizing incremental load pattern clustering of residential customers,utility companies can better cope with the management pressure of dynamic electricity consumption data.By building short-term residential load forecasting model,utility companies can evaluate the potential of customers to participate in demand response.In this thesis,the pattern clustering,incremental pattern clustering and short-term load forecasting of residential customers based on smart meter data are studied.The specific contents are as follows:According to K-means algorithm performance influenced by the randomness of initial clustering center,and similarity of Residential Load Profiles(RLPs)determined by Self-Organizing Map(SOM)through iterative training,a two-stage Weighted SOM residential load pattern clustering framework is proposed to get Typical Residential Load Profiles(TRLPs)of all customers.In order to better measure clustering performance of the algorithms,a clustering performance evaluation method based on the joint sum of squares error and Davidson-Boding index is proposed.Because the Euclidean distance as a similarity measure cannot accurately distinguish the difference between load profiles and computation complexity is large,a three-stage incremental piecewise slope residential load pattern clustering framework is proposed based on shape of load profiles.The first two stages are static clustering that TRLPs are obtained by piecewise slope clustering.The third stage is dynamic clustering.According to slope similarity,incremental clustering is carried out.According to uncertainty of residential load and non-parallelism of Long-Short Term Memory(LSTM),a short term residential load forecasting framework based on Temporal Convolutional Networks(TCN)is proposed.Through analysis and sorting of load correlation between residential electrical appliances and residence,relative loads are taken as auxiliary input of TCN and the residual layer of TCN is further optimized to have higher prediction accuracy.The validity of proposed methods are verified on real data sets.In conclusion,residential customers pattern clustering can help load management select candidate customers to participate in demand response.The incremental residential customers pattern clustering can grasp the dynamic change of residential customers’ s electricity consumption behavior.Short-term residential load forecasting can evaluate the demand response potential of candidate customers participating in demand response.