Data-driven optimal energy management for a wind-solar-diesel-battery-reverse osmosis hybrid energy system using a deep reinforcement learning approach

• Published in: Energy conversion and management 2021-01, Vol.227, p.113608, Article 113608
• Main Authors: Zhang, Guozhou, Hu, Weihao, Cao, Di, Liu, Wen, Huang, Rui, Huang, Qi, Chen, Zhe, Blaabjerg, Frede
• Format: Article
• Language: English
• Subjects: Algorithms; Arid regions; Arid zones; Cost reduction; Deep learning; Deep reinforcement learning; Diesel; Diesel fuels; Energy; Energy management; Energy policy; Entropy (Information theory); Fossil fuels; Hybrid energy system; Hybrid systems; Information entropy theory; Machine learning; Optimal control; Reinforcement; Renewable energy; Reverse osmosis; Solar energy; Uncertainty; Water shortages; Wind
• ISSN: 0196-8904; 1879-2227
• DOI: 10.1016/j.enconman.2020.113608

•The dynamic energy management of hybrid energy system is studied.•Uncertainty of renewable energy, demand side and price are considered.•A data-driven deep reinforcement learning method is applied.•The flexibility adjustment of battery and water tank is achieved. Significant dependence on fossil fuels and freshwater shortage are common problems in remote and arid regions. In this context, the operation of a wind-solar-diesel-battery-reverse osmosis hybrid energy system has become a suitable option to solve this problem. However, owing to the uncertainties of renewable energy availability and load demand, it is a challenge for operators to develop an energy management scheme for such a system. This study aims to determine a real-time dynamic energy management strategy considering the uncertainties of the system. To this end, the energy management of a hybrid energy system is presented as an optimal control objective, and multi-targets are considered along with constraints. The information entropy theory is introduced to calculate the weight factor for the trade-off between different targets. Then, a deep reinforcement learning algorithm is adopted to solve this problem and obtain the optimal control policy. Finally, the proposed method is applied to a typical hybrid energy system, and numerous data are applied to train an agent to obtain the optimal energy management policy. Simulation results demonstrate that a well-trained agent can provide a better control policy and reduce costs by up to 14.17% in comparison with other methods.