Bidding Behavior Analysis in Joint Electricity and Carbon Market by Hybrid Experimental Learning

Abstract

It is important to model the causality behind social behaviors in the electricity market. Existing methods, including theoretical models and economic experiments, are difficult to be applied in practice. In this paper, a data-driven approach, named Hybrid Experimental Learning (HEL) which combined machine learning and experimental economics, is proposed to model social behavior. With the historical and experiment data, HEL applies a machine learning generative model for understanding social behaviors. The output of the generative model is fed into a causal estimator to explain the causality. According to the pilot spot market rule and potential carbon market rule in Guangdong, bidding strategies of generators are generated by Wasserstein generative adversarial networks (WGAN). In addition, an instrumental variable (IV) method based on the local surrogate model is applied for the causal inference between the inputs and outputs of WGANs for coal generators, which can be described as their bidding mechanisms. The effectiveness is verified on bidding strategies simulation in a simplified Guangdong power system.