Data-Driven Control And Fault Detection With Applications

Due to the rapid development of solar power techniques, it becomes possible for new energy sources to replace the traditional fossil fuels. In the researches of photovoltaic (PV) power generation, besides the material techniques of photovoltaic cells and the level of technics in production, the control of PV inverter is very important to the performance of the photovoltaic power system. The control performance of PV inverters directly determines the stability, output efficiency, safety etc. In addition, for the demand of safety, the fault detection techenique is also very important in solar power generation systems. For the trend that solar power system being used in distributed generation systems, the distributed generation techniques are needed to be studied. In distributed solar power generation system, taking the intermittence and interactions amang PV units into consideration, distributed control strategies is used to ensure stabilization and optimization of a distributed generation network.For a complex system such as solar power generation systems, it is hard and time-consuming to obtain mathematical models for the traditional model-based methods. To overcome this drawback, with the help of computer technology and digital control technology, data-driven controller design methods have become popular in the field of control theory and engineering. By using input output data, the controller can be directly designed. This type of approaches has unique advantages in theory and obvious value in application.This dissertation focuses on the research of data-driven subspace-based predictive control and data-driven subspace-based adaptive fault detection method, as well as data-driven subspace-based distributed control. The research results were validated by their applications in solar power generation systems. By analyzing and classifying the grid-connected inverters, several issues in solar power generation system have been addressed, such as grid impedance variation for current control, disturbances and system parameters uncertainties for fault detection, and the PVs’interactions for distributed generation system. The concrete contents of this dissertation are given as follows:Data-driven subspace-based predictive controller design is analyzed, which is based on subspace identification method. A data-driven subspace predictor can be directly obtained from the offline input output data. Based on the subspace predictor, combined with the Hx mixed sensitivity control framework, a data-driven subspace-based H∞mixed sensitivity control scheme is proposed. The demands on the tracking control performance and the robust performance are specified in the controller design. By comparing with the traditional PI controller and the subspace-based LQG controller, the simulation results demonstrate the applicability of proposed controller in grid-connected PV systems and the robustness to deal with the impacts of grid impedance variations on inverter current control, especially the inductive variations.If a distribution network with many PVs that constitute a high level of penetration is considered, the intermittence of PVs’ energy will result in some problems. To deal with, a two level control structure is designed:in the first level, a distributed cooperative control is used so that all PVs work in accordance with the same power output ratio and a reference current for each PV is offered; in the second level, to solve the control problem caused by the interactions among PVs, data-driven distributed predictive control is proposed. Both of the control schemes in two levels are verified in numerical and SimPowerSystems simulations.A data-driven fault detection method is proposed to ensure the safety of the grid-connected solar power generation system. The faults are needed to be detected immediately for avoiding further damages of electrical equipments and grid stability. In the proposed data-driven fault detection method, the predictive current from subspace predictor is compared with the measured current. Then an adaptive filter is used to obtain a residual signal to indicate faults from the predictive error. This filter is updated by using online data. The simulation results demonstrate that the proposed fault detection method outperforms traditional algorithms, being able to detect kinds of faults in solar power generation systems.