Knowledge-inspired data-driven prediction of overheating risks in flexible thermal-power plants

Mechanism-data-integrated methods are promising technologies for safe and flexible operation of power stations, which play an important role in compensating for the renewable energy intermittency and fluctuation. As an attempt in this direction, this work is devoted to the tube overheating problem in the boiler with the aim of developing effective predictive methods. First, we obtain insights from the real incidents with excessive metal temperatures. With data collected over a six-month period from a 350-MW cogeneration unit operating in a flexible mode, we quantified 230 overheat events using steam-pressure-dependent permissible limits. Power law distributions with tails are revealed for the severity of overheating, and three types of events can be classified. Among them, the ‘moderate-type’ overheating can be accompanied with simultaneous overheating of multiple neighboring tubes, and is thus recognized as the primary target for the real-time prediction. Our data-driven model rests on the long-short-term memory neutral network and gives satisfactory outputs under normal operating conditions with a mean absolute error of 3.40 °C. However, the original model fails to reach the extreme values whilst tube overheating due to severe dataset imbalance as only 0.012% of the data correspond to overheating. Finally, we devise an additional strategy making use of the change in the predictive capability of the model. The integrated method successfully predicts all overheat events of a tube over two minutes in advance, and the false alert is kept at a minimum level.