Learning-Based Model Predictive Control for a Wind-Powered Fresh Water Production Plant with Integrated Electricity Production

Abstract

Growth of the population, industry and agriculture increases the demand for fresh water. However, with only a small portion of the water on earth being suitable for direct human use and consumption, this growth makes water scarcity an increasing problem. Seawater desalination is an effective method to counter such water stress. Reverse Osmosis (RO) is an efficient desalination method with relatively low energy consumption, where a semipermeable membrane rejects dissolved contaminants present in the feed water. Delft Offshore Turbine (DOT) aims to produce a seawater hydraulic wind turbine that performs this RO process directly from wind energy. However, there are some challenges with this new concept of how to regulate the production of fresh water and electricity optimally. The main challenge in the optimisation of this system lies in the control of the electricity production, which is pressure-based, and control of the fresh water production, which is flow-based, while having a direct influence on each other, as well as that these controllers need a fast direct response to wind fluctuations. An interesting control approach that can handle such multivariate systems is Model Predictive Control (MPC). MPC optimises the overall performance of a system, while handling input and output constraints. A drawback, however, is that the performance highly depends on the accuracy of the system model. This work investigates the use of machine learning to extend the MPC framework, where in addition to a nominal system model, a Gaussian Process (GP) is used to learn unmodeled system dynamics, making it possible to use approximate system models. Tests done in simulation and on a physical setup have showed increased control performance compared to MPC by successful learning of the system online making up for unknown system dynamics or uncertainties. Further research may be done on optimising the GPs by reducing computational costs while maintaining information about the system, and better exploit the given uncertainty of the GPs.