An Integrated Energy Trading Strategy Management Architecture with Deep Reinforcement Learning: Application in Türkiye

For industrial enterprises, effective energy trading in intraday markets is vital for cost reduction but challenging due to market volatility and the complexity of decision-making. This study proposes a novel integrated energy trading framework utilizing Deep Reinforcement Learning (DRL) to automate and optimize trading strategies in the Intraday Market (IDM). Unlike conventional rule-based or purely simulation-driven RL models, the proposed framework employs a state-augmented Markov Decision Process (MDP) formulation that integrates exogenous System Marginal Price (SMP) interval forecasts directly into the agent’s observation space. Specifically, a Deep Q-Network (DQN) architecture augmented with Prioritized Experience Replay (PER) is developed to handle the high volatility and continuous state space of energy markets, overcoming the limitations of traditional tabular Q-learning. To enhance the agent's decision-making capabilities, a hybrid forecasting mechanism employing LightGBM and CatBoost algorithms is integrated to predict SMP. The proposed system was validated using real-world data from Energy Exchange Istanbul (EXIST). The results demonstrate that the DRL-based agent achieved approximately $1 savings per MWh and reduced monthly operational labor costs by over $6,000 compared to rule-based strategies, achieving a ~158% performance improvement over discrete-state Q-table baselines and proving its efficacy in minimizing energy costs and managing imbalances.