Short-term Load Forecasting using Partial Adversarial Adaptation

Accurate short-term load forecasting (STLF) is vital for modern power systems to be safe and efficient. STLF presents difficulties because of the cyclic, non-stationary, seasonal, and nonlinear nature of electricity consumption. With advanced metering infrastructure (AMI) collecting massive power con-sumption data, machine learning (ML) methods have emerged as promising tools to improve STLF accuracy. However, one critical challenge faced by existing ML methods is to deal with the domain shift problem when applying ML models into the industrial scenario. To address the limitation, we propose a novel partial adversarial adaptation enhanced deep forecasting (PADF) framework. Specifically, we identify the outlier samples in source domain using a domain classifier without gradient reversal layer (GRL) and reduce the domain shift between normal target and source domain samples at the same time. We demonstrate the effectiveness of PAFF on real-world large-scale datasets from the state grid corporation of China (SGCC).