Transferable Household Electricity Scenario Prediction

As an important component of power grid planning,electricity load forecasting has significant implications for the safe and efficient operation of the power system.Accurate electricity load forecasting,particularly for household electricity load forecasting shortly,is of certain significance for demand response in the power system.The intermittent and fluctuating nature of residential electricity load,as well as the problem of insufficient accumulated data for some households,can affect the accuracy of residential electricity load forecasting.This article proposes a scenario prediction method based on multiple sets of household electricity load data and combines model transfer to conduct relevant research.A flow-based generative model approach is adopted to establish a load scenario prediction model in the source domain household.The flow-based generative model has strong processing capabilities for continuous value vectors.Scene prediction can accurately describe the random load variation with a few load scenarios,and describe the uncertainty of power load in the form of prediction interval or quantile,and generate a set of trusted time series data to provide more possibilities for generating the final forecast results.To make the flow-based generative model more dependent on historical load data,the structure of the affine coupling layer is improved to enhance the model’s predictive performance.The experimental results show that the prediction interval generated by the improved flow generation model can better cover the real load valueTo fully utilize the readily available source domain household electricity load data,grey correlation analysis is introduced to select source domain households that are similar to the target domain households based on the correlation between the load data of the source and target domains.By setting different correlation threshold values between the target domain households and the source domain households,the source domain households with correlation below the threshold are discarded,while those above the threshold are used to establish a multi-source domain household dataset.A pre-training model was established using multi-source domain household electricity load data,and the model transfer was performed to establish a scenario prediction model.In the target domain household,the partial single-step flow network structure parameters of the pre-training model were fine-tuned and trained using the target domain household electricity load data,and a daily electricity load prediction model under small sample scenarios was constructed.The experimental results analyzed and compared the predictive performance of the target domain model with and without model transfer,as well as the predictive results of single-source domain transfer and multi-source domain transfer.The experimental results show that the model transfer based on multi-source domain can significantly improve the prediction performance of target domain model under small samples.In summary,the scene prediction method based on flow generation models can effectively reduce the impact of randomness and fluctuation of household electricity load on prediction accuracy.Model transfer can improve the prediction performance of household electricity load prediction models in small sample data scenarios,and the establishment of multi-source domain households can further improve the prediction performance of the target domain model after model transfer.