Prediction driven control of a gas turbine–battery hybrid power plant providing an ancillary service

Due to the increasing use of renewable resources, such as photovoltaic or wind power plants, there is a growing need for ancillary services to stabilize the grid balance. In this paper, we study the control of a real-world hybrid power plant providing ancillary services formed by a set of fast aeroderivative gas turbines (AGTs) and a large battery energy storage system. Our paper serves as a feasibility study, demonstrating how forecasting of automatic frequency restoration reserve (aFRR) activation could improve the economic performance of the hybrid power plant before the idea is implemented in the real control system. The control of the hybrid power plant is formulated as an optimization problem, while machine learning assists in making better unit commitment decisions. We focus on the aFRR ancillary service. Considering the required service response time and the time parameters of AGTs, it is the most challenging ancillary service to provide from a control algorithm design standpoint. We compare our control with the control that does not exploit machine learning and the theoretical bounds the control can achieve. The comparison illustrates the positive impact of machine learning on the operational costs of the power plant. Experiments show that the behavior of the control with machine learning is close to the optimal control assuming the full knowledge of the aFRR power required in the future. Furthermore, there is a noticeably reduced number of AGT starts and reduced gas consumption by 72 %–90 % of the possible savings, related to the optimal value of gas consumption obtained through the optimal control.