Data-efficient digital twin for turbine heat rate of industrial thermal power plant

Developing a robust and efficient data-driven digital twin system for industrial thermal power systems remains challenging due to data drift, change in operating behaviour of the system and ineffective data-sampling issues for data-driven model development. We present data-efficient model training framework that incorporates data sampling from large volumes of asymmetric and controlled data regimes of industrial power systems. Artificial Neural Network (ANN) model is trained on the sampled and representative dataset to predict turbine heat rate (THR) of 660-MW capacity thermal power plant. Later, THR is minimized by a constrained non-linear optimisation technique at 50 %, 75 %, and 100 % capacity discharge of power plant, and the optimisation-based results are validated in the operation of the power plant with the mean absolute percentage errors of 0.79 %, 2.98 % and 0.33 % respectively. The analysis on cost of operation and carbon dioxide (CO2) reduction reveals that minimizing THR through the data-efficient model training and optimization framework can save around 13 million USD with a reduction of 28 kilotonnes (kt) of CO2 per year. Finally, the data-efficient trained ANN model is deployed as a digital twin system for monitoring the THR and is found to be more than 90 % accurate on 19000 min of real-time monitoring window. This research paves the way for data-efficient sampling from the controlled datasets of industrial systems that leads to improved generalisation capacity of the trained machine learning models for their integration in the digital twin systems for monitoring the performance of industrial power systems.