A bottom-up short-term residential load forecasting approach based on appliance characteristic analysis and multi-task learning

Abstract Residential load forecasting has been an intractable problem due to the small load scale and strong randomness of residents' power consumption behaviors. The traditional methods only use load data at household-level and cannot adequately consider the power consumption habits of users, so the forecasting effect is limited. To address these challenges, this paper proposes a bottom-up approach considering the load characteristic of the appliance-level. The novelty lies in the following three aspects: 1) considering the working principles and load characteristic of appliances, household appliances are classified into continuous and intermittent load appliances, to achieve more refined load forecasting by aggregating loads of appliances; 2) considering the fluctuation of load curves, the seasonal-trend decomposition procedure based on Loess (STL) is applied to continuous load appliances to obtain more fine-grained and regular load data, which can give full play to the advantage of gated recurrent unit (GRU) for time series prediction; 3) in order to take power consumption behavior of users and correlations of appliances into account, a multi-task learning (MTL) network is designed for intermittent load appliances, and the model can enhance computing efficiency by predicting multiple appliances simultaneously. Case studies show that the proposed method can effectively reduce error of residential load forecasting.