Prediction of part-load behaviour of natural gas combined cycles by applying thermodynamic and artificial neural network models

The prediction of power plant performances has traditionally relied on complex thermodynamic models, which incorporate numerous assumptions and operating parameters. As a result, evaluating their energy performance at design or part-load conditions demands significant computational resources to solve complex systems of non-linear equations. Machine learning approaches offer a potential solution to reduce this computational burden. This study predicts the part-load behaviour of combined-cycle gas turbines using a method that integrates thermodynamic and artificial neural network models. The approach involves the random generation of input variables, representing power plant operating conditions. Output variables, including energy and economic performance indicators, are then evaluated through rigorous thermodynamic simulation. Artificial neural networks are trained and validated using these datasets, and their ability to replicate thermodynamic models is assessed using statistical performance metrics. The method is applied to a three-pressure and reheat combined-cycle gas turbine, evaluating part-load performance under varying ambient conditions and considering a part-load strategy based on inlet guide vane and turbine inlet temperature variations. The analysis also evaluates the economic revenues from integrating combined-cycle gas turbines into the Italian day-ahead electricity market, considering ambient condition profiles typical of winter and summer days for a hypothetical power station site. The results demonstrate the model’s capability to define optimal part-load strategies and identify bidding strategies that maximize profits in interconnected electricity and gas markets.