Deciding When to Use a Personalized Model for Load Forecasting

Load forecasting plays a vital role in achieving supply-demand balance in power systems and lays the foundation for economic dispatch, demand response, etc. Conventionally, a global forecasting model will be constructed for the whole dataset. The problem of non-independently and identically distributed (non-i.i.d) data causes performance degradation of the global model. Many studies suggest building personalized models to account for heterogeneous distributions. However, there lacks a discussion about how much benefit can be achieved by the personalization over the global model and when it is necessary to conduct the personalization. To this end, we investigate when to use a personalized model for load forecasting. To answer this question, we first compare the generalization bound of the personalized model to the global model, demonstrating the potential for personalization to enhance performance. Then, we propose a Cluster-Oriented Representations Encoder (CORE) model to map the input feature space into representation space with clustering structure. The mixture distributions of the global dataset can be recognized to construct sub-representation datasets to lay the foundation for personalization. Subsequently, we quantify the personalization gain achieved by fine-tuning the global model with mathematical derivation. Finally, we decide whether to use the personalized model at each timestep and develop an adaptive personalization strategy for load forecasting. Comprehensive case studies have been carried out to validate the efficacy of the proposed adaptive personalization strategy in both single-load and multi-load forecasting and demonstrate the necessity to balance the global and personalized modeling approaches.