A hybrid constrained coral reefs optimization algorithm with machine learning for optimizing multi-reservoir systems operation. (15th May 2021)

• Record Type: Journal Article
• Authors: Emami, Mohammad; Nazif, Sara; Mousavi, Sayed-Farhad; Karami, Hojat; Daccache, Andre
• Is Part Of: Journal of environmental management. Volume 286(2021)
• Publication Date: 2021-05-15
• Subjects: Water resources management -- Artificial intelligence -- Heuristic method -- Decision support tool -- Multi-agent approach -- Particle swarm optimization; Environmental policy -- Periodicals; Environmental management -- Periodicals; Environment -- Periodicals; Ecology -- Periodicals; 363.705
• DOI: 10.1016/j.jenvman.2021.112250
• Languages: English
• ISSNs: 0301-4797

Abstract

Abstract: The continuous growing demand for water, prolonged periods of drought, and climatic uncertainties attributed mainly to climate change mean surface water reservoirs more than ever need to be managed efficiently. Several optimization algorithms have been developed to optimize multi-reservoir systems operation, mostly during severe dry/wet seasons, to mitigate extreme-events consequences. Yet, convergence speed, presence of local optimums, and calculation-cost efficiency are challenging while looking for the global optimum. In this paper, the problem of finding an efficient optimal operation policy in multi-reservoir systems is discussed. The complexity of the long-term operating rules and the reservoirs' upstream and downstream joint-demands projected in recursive constraints make this problem formidable. The original Coral Reefs Optimization (CRO) algorithm, which is a meta-heuristic evolutionary algorithm, and two modified versions have been used to solve this problem. Proposed modifications reduce the calculation cost by narrowing the search space called a constrained-CCRO and adjusting reproduction operators with a reinforcement learning approach, namely the Q-Learning method (i.e., the CCRO-QL algorithm). The modified versions search for the optimum solution in the feasible region instead of the entire problem domain. The models' performance has been evaluated by solving five mathematical benchmark problems and a well-known continuous four-reservoir system (CFr) problem. Obtained results have been compared with those in the literature and the global optimum, which Linear Programming (LP) achieves. The CCRO-QL is shown to be very calculation-cost-effective in locating the global optimum or near-optimal solutions and efficient in terms of convergence, accuracy, and robustness. Highlights: The coral reefs optimization algorithm solved mathematical problems. The proposed hybrid algorithm optimized a benchmark four-reservoir problem. Adjusting crossover operators using reinforcement learning improved performance. Recursive-calculation-cost efficiency improved by search in the feasible region.