Abstract
Measuring the turbine inlet temperature (TIT) in gas turbines is daunting due to the exceptionally high temperatures involved. Using thermocouples is not a cost-effective solution for directly measuring TIT, and their durability over extended periods is also a challenge. Online, near real-time accurate prediction of TIT and its close-loop feedback control is critical due to its significant impact on the rate of power generation and on the lifespan of the hot gas path components in gas turbines. Any increase in TIT can cause a reduction in their lifespan, while a decrease can result in a power loss for the customer. To compute TIT, mass and energy balance approaches are used iteratively online in near real-time environment, which relies on continuous measurement of critical signals from gas turbines. These signals include ambient conditions, gas turbine (GT) output, fuel flow, compressor outlet temperature, turbine exit temperature etc.
In a multi-shaft combined cycle power plant, gas turbine power and steam turbine power are measured separately, simplifying the prediction of gas turbine inlet temperature (TIT). However, in a single-shaft power plant, where both turbines’ power output is combined, predicting TIT becomes challenging. Accurate TIT prediction is crucial for effective gas turbine control in single-shaft combined cycle units. Therefore, a rigorous validation methodology using gas turbine fleet data is necessary to ensure efficient operation.
A new approach for predicting TIT in single-shaft power plants has been developed and validated using data from a multi-shaft plant. The paper describes the methodology, called the Heuristic Transfer function, which utilizes prior domain information for TIT prediction and its validation process using fleet data. Despite using the Heuristic Transfer function, the accuracy and variance of the predicted TIT over time did not meet acceptable thresholds.
To address the concerns of prediction inaccuracy and non-stationarity over time, a novel approach is proposed in the paper. It involves combining the baseline heuristic transfer function with a stochastic model to achieve accurate and time-invariant predictions. The stochastic model is developed using residuals from the heuristic transfer function, providing a more robust and reliable TIT prediction method.