Abstract:
Steam turbines are the key elements of thermal power plants. Their efcient functioning is vital for optimizing power production and minimizing operational expenses. But the turbine performance is continuously changing with changing thermodynamic conditions namely main steam pressure, main steam temperature, cold reheat pressure and cold reheat temperature. Traditional industrial control systems such as PID and DCS guarantee the stability of operations, but they are not designed for intelligent optimization of efciency of the operations in the dynamic operating environment. In this work, a physics aware reinforcement learning based supervisory optimization framework is presented for HP steam turbine efciency improvement. The proposed system is a closed loop supervisory architecture combining thermodynamic efciency estimation, reinforcement learning optimization, recommendation evaluation and real-time dashboard visualization. The HP turbine efciency is calculated by thermodynamic relations based on the change in enthalpy of the steam and the calculated efciency is checked with the dataset efciency values for physical consistency and reliability. Three reinforcement learning algorithms (PPO, SAC, and TD3) are trained to provide safe operational recommendations via controlled modifcations in selected thermodynamic operating parameters. Prior to presenting recommendations to the operator, the expected turbine efciency is re-evaluated after each recommended action using the physics-based effciency model. The reinforcement learning models are also compared with No Action and Random Action baseline strategies in terms of efciency improvement and statistical evaluation metrics. A real-time supervisory dashboard is also developed to visualize the current HP turbine efficiency, recommended operational changes, expected optimized efciency, and efciency improvement. The proposed framework also enables real-time data ingestion from an external system that allows live monitoring and recommendation generation in dynamic operating conditions. The system developed serves as a supervisory advisory layer, not interfering directly with the existing industrial PID/DCS infrastructure. The proposed framework demonstrates the potential of the combination of thermodynamic modeling and reinforcement learning to develop intelligent, adaptive, and safe supervisory optimization systems for steam turbine efficiency improvement in modern thermal power plants.
