Artificial Intelligence-Powered Energy Management of Reverse Osmosis Desalination Plants

  • Author / Creator
    Soleimanzade, Mohammad Amin
  • The rapid increase in global water and energy demand due to industrialization and population growth is a pressing challenge humankind faces today. Recent estimates indicate that due to population growth and reduction of water supplies, 40% of the global population is struggling with water scarcity, and a 20% increase is predicted for this number by 2025. Furthermore, The global energy demand expanded from 5000 million tons of oil equivalent in 1971 to 11700 million tons of oil equivalent in 2010, and it is predicted that it will increase up to 33% by 2030. The exponential increase in energy demand has exacerbated the greenhouse gas emissions as fossil fuels are mainly used to supply the required energy. The deployment of renewable energy sources, such as solar and wind power, to increase energy supply and diminish the adverse environmental effects of fossil fuels is considered an efficient solution to these problems. Among renewable energy sources, the exploitation of solar power generation has received significant attention and is considered one of the most promising options. However, the intermittent nature of photovoltaic (PV) power brings a huge challenge to PV-powered microgrids and desalination systems. Hence, it is essential to design advanced control techniques capable of coping with this challenge to optimize the performance of PV-driven desalination systems in terms of water production and energy consumption.
    This thesis proposes two artificial intelligence-powered energy management systems for a hybrid grid-connected PV-reverse osmosis-pressure retarded osmosis desalination plant. In the first part of the thesis, an intelligent energy management system (IEMS) is developed to maximize the total water production and contaminant removal efficiency while keeping the grid’s supplied power as low as possible. To promote the performance of the IEMS, the prediction of PV solar power is performed by three deep neural networks based on convolutional neural networks and long short-term memory neural networks. These networks are designed to perform 5-hour-ahead PV power forecasting, and the model with the smallest error is selected. The IEMS employs the particle swarm optimization (PSO) algorithm to find the optimum solutions of the system for each time step. Four performance indices are defined through which the IEMS is evaluated. The proposed technique results are compared with two benchmark methods, one of which is similar to the IEMS; however, it does not incorporate the PV power predictions. The superiority of the IEMS over the first benchmark is demonstrated by studying three scenarios: two successive sunny days, two successive cloudy days, and 10 days of operation. Moreover, the simulations are executed for different forecast horizons to investigate the effects of this parameter on the optimization results. The impacts of the best-found forecaster errors are also explored by repeating the simulations with the actual PV power data. Finally, the optimization is performed by two other stochastic algorithms: grey wolf optimizer (GWO) and genetic algorithm (GA). It is found that PSO outperforms GWO and GA for solving this optimization problem.
    The second part of this thesis proposes a novel deep reinforcement learning-accelerated energy management system for the desalination plant mentioned above. The energy management problem is formulated as a partially observable Markov decision process, and the soft actor-critic (SAC) algorithm is employed as the core of the energy management system. We introduce 1-dimensional convolutional neural networks (1-D CNNs) to the actor, critic, and value function networks of the SAC algorithm to address the partial observability dilemma involved in PV-powered energy systems. The superiority of the proposed CNN-SAC model is verified by comparing its learning performance and simulation results with those of four state-of-the-art deep reinforcement learning algorithms: Deep deterministic policy gradient (DDPG), proximal policy optimization (PPO), twin delayed DDPG (TD3), and vanilla SAC. The results show that the CNN-SAC model outperforms the benchmark methods in terms of effective solar energy exploitation and power scheduling. By conducting ablation studies, the critical contribution of the introduced 1-D CNN is demonstrated, and we highlight the significance of providing historical PV data for substantial performance enhancement. The average and standard deviation of evaluation scores obtained during the last stages of training reveal that the 1-D CNN significantly improves the final performance and stability of the SAC algorithm.

  • Subjects / Keywords
  • Graduation date
    Spring 2022
  • Type of Item
  • Degree
    Master of Science
  • DOI
  • License
    This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.