b"Deep reinforcement learningAn intelligent system can learn and control a complex integrated and decision analytics forenergy system and can improve operations and profit. integrated energy systems T his project developed a novel deep reinforcement learning approach that can manage an integrated energy system consisting of several different components. The approach demonstrated that it could produce a stable and efficient system that gains significantly better rewards than traditional methods. We demonstrated how a nuclear power plant can couple with elements such as hydrogen generation, solar, and thermal and electrical storage to make the best plant decisions and dynamically adjust to forces such as energy markets. PROJECT NUMBER: 21A1050-073FP In the process, the project prepared a novel control and learning environment and tested several different learning systems. It created numerous novel models of TOTAL APPROVED AMOUNT:energy systems, and it created a framework to integrate models that were created $1,030,000 over 3 years from different sources using functional mock-up interfaces and functional mock-up PRINCIPAL INVESTIGATOR:units. The project also created methods to optimize the system configurations and Victor Walker improve efficiency. CO-INVESTIGATORS: The project created a framework to integrate to a hardware-in-the-loop system Congjian Wang, INL and demonstrated the resulting deep reinforcement learning control system using Tyler Westover, INL physical hardware that maintained a stable grid environment while improving Zonggen Yi, INL rewards. The deep reinforcement learning algorithm's rewards were calculated Ahmad Javaid, University of Toledo based on variables such as electricity and hydrogen selling prices within a stable Michael Heben, University of Toledo grid environment. The control achieved a substantial 30-50% reward improvement Raghav Khanna, University of Toledo when factoring in hydrogen generation and selling, compared to control without considering hydrogen. And the deep reinforcement learning control showed between 5-30% improvement in modeled reward over traditional control.80"