LLM-Enhanced Trading Decision Framework with Multi-Scale Memory for Electricity Markets

AI-based methods have advanced electricity market trading, yet most time-series and deep reinforcement learning (DRL) models struggle to incorporate unstructured information such as news and policies. This paper introduces a large language model (LLM)-driven trading framework that integrates multi-level memory - short-, mid-, long-term, and reflective - to emulate human decision processes under complex market signals. The framework dynamically evaluates the semantic relevance of news across temporal scales and models delayed effects in relation to market behavior. Experiments on the Australian electricity market show that our approach outperforms traditional strategies in both interpretability and predictive performance. These findings highlight the critical role of textual data in improving market foresight and adaptive trading under information-rich conditions.