Reinforcement Learning for Managing and Predicting Gas Consumption of CGS: A Novel Approach to Optimizing the Gas Distribution Network

Efficient city gate station (CGS) management is critical in ensuring cost-effectiveness and network stability in natural gas distribution networks. However, CGS performance faces challenges such as high operational costs, pressure fluctuations, and seasonal demand variations. Traditional forecasting methodologies, including ARIMA and econometric models, struggle to handle nonlinear consumption patterns, while Deep Learning (DL) algorithms like LSTM and GRU, despite offering improved accuracy, lack adaptability in dynamic environments. To address these limitations, this study proposes a hybrid Reinforcement Learning (RL) framework that integrates State-Action-Reward-State-Action (SARSA), Deep Q-Networks (DQN), and Policy Gradient (PG) algorithms. By defining gas distribution as a Markov Decision Process (MDP), our model employs DQN for value-based decision-making and PG for policy optimization, enhancing both accuracy and adaptability. Additionally, real-time forecasting models are incorporated to improve demand prediction and response efficiency. Experimental results validate the effectiveness of our RL-based model, achieving: A) ۵۵% reduction in forecasting errors compared to traditional statistical models, B) ۸۸% decrease in pressure fluctuations, ensuring network stability, C) ۹% cost savings through optimized gas distribution strategies. These findings highlight the transformative potential of RL in gas distribution management, improving efficiency, adaptability, and resilience in dynamic conditions and regulatory constraints. The proposed framework introduces a scalable and intelligent decision-making platform for next-generation energy infrastructure.

کلیدواژه ها:

Reinforcement Learning ، Deep Q-Networks ، Natural Gas Optimization ، City Gate Stations ، SARSA ، Policy Gradient ، Markov Decision Process