Hybrid thermodynamic & machine learning model to predict Kawerau geothermal station output live

Abstract

Previously the Mercury Optimisation Engineering team has been able to predict a geothermal station’s MW production to +/-0.5MW for one given time period, using station, steam field and integrated models. The Mercury Generation Optimisation Engineering and Decision Science teams saw an opportunity to create a digital twin to simulate various scenarios and accurately predict the net MW production at Kawerau live.
Using an agile approach, the joint team created a hybrid model that uses both engineering principles and Machine Learning (ML) to predict Mercury’s geothermal flash plant Kawerau’s net MW production live. The new model takes and predicts many inputs, including ambient weather forecasts (humidity and temperature), predicted available fuel, station derates, maintenance at the station and station operating states. The hybrid model has been tuned to be able to predict the net MW output to an accuracy of +/-0.3MW (Kawerau Gross MCR 113.67MW).
The production model provides:
1) A model that provides Operators with optimised set points to maximise generation, without putting the station into constant alarm.
2) A forecasting model that improves MW prediction accuracy. This provides Wholesale and Trading greater accuracy MW inputs for their Waikato River scheme model, particularly pertinent in dry years.
3) A hindcasting tool to help account for lost MW for various operating scenarios, which can be used to better prioritise projects to improve the station’s efficiency and minimise downtime.
The optimised hybrid model helps maximise Mercury’s geothermal baseload electricity generation, helping to address New Zealand’s energy trilemma.