Online transfer learning-based residential demand response potential forecasting for load aggregator

Abstract

Accurate demand response (DR) potential forecasting is the basis for load aggregators (LA) to make optimal bidding strategies in DR market trading. LAs usually face practical challenges when they perform forecasts for those new customers who have no historical response data. Transfer learning provides a promising solution to this problem by leveraging knowledge acquired from other existing contracted customers. However, traditional transfer learning methods are trained offline and cannot make use of the latest response information of these new customers, which may result in large forecasting errors since the response behavior of new customers usually dynamically changes. To address the above issues, this paper proposes an online transfer learning-based DR potential forecasting framework, in which two forecasting models are established. The first one is built using the historical data of existing customers and this model is then transferred to the target domain (i.e., new customers) by parameter sharing and fine-tuning. The second model is built using the local response data of new customers, which gradually accumulates with the increasing participation of DR events. These two models are combined by an adaptive ensemble framework based on their online performances, thus enabling it to dynamically track the changes in new customers' response behavior. Case studies on a real-world dataset validate the effectiveness of the proposed framework.