A collaborative learning strategy for model-free control of an array of wave energy converters

Ocean energy is an abundant but relatively unexploited renewable energy source which has the potential to become one of the key players in the upscaling of global renewable energy production for the near future. Despite this huge potential, ocean energy technology and especially wave energy technology is still considered to be immature compared to other renewable energy technologies. One of the main goals to be achieved is to reduce the levelized cost of electricity (LCOE) coming from wave energy converter devices in order to make them economically competitive with respect to other more established renewable energy sources. To achieve this, one of the main areas of focus in recent years has been to develop and optimise control strategies to improve the efficiency of the energy conversion process. The main challenges that wave energy converters (WECs) face, stem from the irregular reciprocating nature of the energy source, making the design of the control strategy, the WEC itself, and any modelling of the WEC-wave interaction extremely challenging. A large portion of the most popular control strategies adopted on wave energy converters rely on a model-based control strategy to determine the optimal control action to be taken. The control action is usually optimised over a predefined range of wave conditions which can be grouped within a single statistical description of the current sea state by using parameters such as the significant wave height Hs and the wave energy period Te. Although these control strategies may give good results, they are inevitably affected by modelling errors and uncertainties, together with a control which is not optimised on a wave-by-wave basis. In this thesis, a model free control strategy for an array of heaving point absorbers is explored. A model free control approach was chosen since it allows to neglect the device model and wave interaction modelling, which in turn allows to avoid modelling errors and to directly develop the control strategy using data obtained while at sea. The proposed strategy involves an initial online optimisation of the control parameters of a reactive control law using genetic algorithms to map the point absorber array to a population of individuals within a metaheuristic optimization framework. This allows the single point absorbers to collaborate and learn form one another to reach the common goal of finding the optimal control parameters for each discrete sea state encountered. After the initial sea-state-based model free collaborative optimisation has reached satisfactory results, a secondary mechanism based on machine learning through neural networks is used to try and learn interdependencies between the discrete sea states and the relative optimal control parameters to achieve a continuous control, no longer dependant on a statistical description of the sea state but based on direct force measurements on the heaving point absorbers. In this framework, the data collected during the initial optimization using genetic algorithms is then used to train the neural networks so to output a continuous control command. Preliminary simulation results show that an array of point absorbers using a genetic algorithm based collaborative optimisation is able to achieve control parameters close to the theoretical optimal ones within only a few days from deployment, while the neural networks show comparable performance, indicating that with further tweaking of the learning procedure, superior results may be obtainable.