Year: 2026
Type: article
Abstract: Reinforcement learning may improve strategic bidding in electricity markets, yet many studies rely on stylized or synthetic environments. This study evaluates reinforcement learning using bid-level data from the Kazakhstan Electricity and Power Market Operator (KOREM) covering July 2023 November 2025. Objectives include reconstructing endogenous pay-as-cleared prices through merit order clearing, modeling one strategic generator alongside passive historical participants, and evaluating out-of-sample performance. Marginal costs are approximated by the 10th percentile of each generator’s historical bid prices. Double Q-learning replaces classical Q-learning to address observed instability associated with overestimation bias, while bounded markup decisions and realistic market constraints ensure economically plausible bidding. Evaluation using a chronological 80/20 train–test split indicates an approximately 0.37% profit improvement relative to cost-based bidding, without distorting clearing price distributions or intraday patterns. Stabilized reinforcement learning can modestly enhance strategic bidding in an empirically grounded electricity market while preserving fundamental price formation properties.
Source: НАУКА ИНТЕЛЛЕКТУАЛЬНЫХ СИСТЕМ
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7225113003