Model-Free Control System Architecture for the ISWEC

In the field of renewable energy, one of the most promising technologies is wave energy. The amount of energy potentially available from the motion of waves in seas and oceans is high. Only a small part of this energy is currently converted and exploited. Systems used to extract wave energy are called Wave Energy Converters. One of these devices, the ISWEC (Inertial Sea Wave Energy Converter), uses the gyroscopic effect of a rotating flywheel to convert the energy of sea waves motion into electrical energy. This system is composed of an anchored hull capable of orienting itself with respect to the incident wave and that contains inside a gyroscopic platform. The mechanical energy collected by the gyroscopic system is then converted through the PTO system. The wave resource is variable in time and space. Its characteristics can be expressed through various parameters, which affect the amount of energy available and the way the system will be stimulated. These parameters define the sea state. It is therefore clear that the control system of these converters must be able to adapt its action to the sea state changes to maximize the extracted energy. The control system that is currently used on the ISWEC generates a torque that follows a proportional derivative control law, whose parameters are computed using a lookup table built offline on the basis of simulations made with a model of the system. The flywheel angular velocity is also controlled by means of a lookup table. This control system is unfortunately affected by uncertainties due to the differences between the theoretical model on which it is based and the actual model. Also, the proportional control parameters do not change with the variation in time of the real system but are chosen in the design phase based on simulations of the system with ideal waves. For these reasons, the conversion process efficiency is affected, and the extractable energy reduced. For the above motivations, to improve its productivity, a new control system is pursued in this work. The aimed controller is model-free, and it is able to tune itself, adapting over time, and finding an optimal action on the basis only of the previous values of inputs and outputs. The development of this control system is based on the concept of building a metamodel and optimizing it. A metamodel is a model built through real experiences (replaced in this work by simulations) that is used to map the different inputs (sea state, stiffness, damping and flywheel speed) with the obtained outputs (average extracted power within a proper amount of minutes). To build this metamodel, different strategies based on regression trees, ensemble of trees, neural networks and radial basis functions have been adopted and compared. An exploration strategy has been elaborated for the development of the metamodel. This strategy is composed of a first stage of pre-learning and a second one of optimization (following the concept of exploration/exploitation). In the first phase the aim of the control system is the exploration of the map and of some of its possible combinations. In the second phase, once enough experience has been obtained and the metamodel has been built, the control action is optimized to obtain the maximum power based on the current sea-state. The control system thus developed is then studied in terms of productivity and its losses, developed during the first phase of pre-learning, are also studied.