Computationally Efficient Optimization of Long Term Energy Storage Using Machine Learning

Abstract

Energy storage can be charged when energy is cheap and discharged when it is expensive to make an energy system more profitable or used to make the plant operation more efficient to reduce CO2 emissions. To optimize long term energy storage with conventional methods a long time horizon must be used. When the long term energy storage is combined with a complex energy system the computational cost becomes large when using conventional methods. To reduce the time horizon, an algorithm will be used to decide the state of charge of the long term energy storage at the end of the day. This algorithm is trained using machine learning with data of the optimal state of charge obtained by running computationally heavy long time mixed integer linear programming ahead of time. Then a one-day or week mixed integer linear programming optimization will be done for the production planning. The seasonal patterns of the long term energy storage can then be captured while giving the plant operator a simple one-day or week production plan. A case study will be done with a combined heat and power plant system with 4 boilers, a long-term thermal storage, and a hydrogen storage system. Using this method the complexities of a multi energy system with long term energy storage can be captured while doing day ahead production planning