Data-driven control of wave energy systems using random forests and deep neural networks

In the field of sustainable energy and alternatives to fossil fuels, wave energy is generating an increasing interest due to its untapped potential. However, the levelised cost of energy of wave energy systems is still not able to compete with other renewable technologies, mainly due to high costs associated with their conversion process. In this context, the development of energy-maximising control strategies plays an important role towards the economic viability of wave energy technology, by optimising the overall energy conversion, hence contributing towards minimising the associated cost of energy. State-of-the-art control systems for wave energy converters are mostly model-based, exploiting control-oriented models of the device to compute the applied control actions with a limited computational burden. Nevertheless, these representations of the system are simplified, and often based upon unrealistic assumptions, such as small motion around the zero equilibrium position, which are inherently invalidated during device operations and that can lead to large uncertainties, resulting in suboptimal power absorption. For these reasons, in this paper, a purely data-driven control strategy is developed. This strategy exploits random forests (RFs) and deep neural networks (DNNs) to gradually learn from real experiences towards an optimal proportional-integral control action. These structures are used as surrogate models (built upon the data coming from past experiences) to converge to the optimal control parameters in a surrogate-optimisation-like manner. To manage the exploration and exploitation needs of controllers based on this approach, a learning strategy is developed and presented. Some considerations are made on the choice of the input features of the surrogate structures, which deeply affect the control strategy learning results. To assess the performances of both the control and learning strategies, one year of operations has been simulated under control settings guided by the proposed data-driven approach, showing also the potential capabilities that the adoption of RFs and DNNs has in learning, even in sea conditions with a limited number of occurrences.