A reference-model strategy for self-protective smart inverters

Abstract

This dissertation presents novel self-protective methods for grid-interactive inverters. The self-protective methods use reference models to inspect the incoming power setpoints, detect unsafe setpoints, and protect the inverter-based system accordingly. By employing these new self-protective methods based on reference models, grid-interactive inverters can effectively contribute to distributed energy generation within the energy infrastructure, where a supervisory control structure is required for energy management and economic dispatch. In a centralized supervisory control structure, the inverters need to be in contact with aggregators, other energy generation units, or the utility operating center. The communication capability makes a grid-interactive inverter a cyber-physical device. However, the connection of inverters to a communication network exposes the inverters to active attackers who can interfere with the control infrastructure and send malicious setpoints to the local controller. Such malicious setpoints can have harmful consequences, such as uncontrolled power oscillations, voltage sags and swells, equipment damage, and blackouts. This dissertation develops the self-protective methods using steady-state and dynamic reference models. The self-protection strategy inspects incoming power setpoints from the utility operator or third-party aggregators using the reference models before engaging the setpoints to the local controllers. At first, a self-learning feature for a self-protective inverter is developed. Self-protective inverters use the self-learning feature to learn their normal operating region by estimating unknown system parameters and employing known parameters from the measurements. The unknown system (grid) parameters are estimated by injecting a current at a different frequency than the fundamental power frequency. Consequently, two distinct real-time grid parameter estimation techniques, namely the model reference estimation and recursive least square method, have been developed and experimentally validated. Then, analytical steady-state and dynamic reference models are developed to inspect the incoming power setpoints. The self-protection method uses these reference models to learn the safe operating region of the inverters. The steady-state reference model is developed based on steady-state linear and nonlinear operating regions, and the dynamic reference model is formed based on the full-order inverter model. Based on the risk of unsafe operation defined by the steady-state model, if the setpoints fall in a high-risk region, the setpoints are rejected by the self-protection method, and the previously accepted setpoints remain in operation. Following the steady-state model check, the accepted setpoints are further examined using the full-order dynamic model to predict the location of dominant eigenvalues of the system using root-locus studies, thus predicting the dynamic response and making the decision whether the commanded setpoints are safe or unsafe. The performance of the self-protection strategy using the analytical reference models is tested using hardware experiments. The steady-state analytical model demonstrates satisfactory performance, characterized by its mathematical simplicity and ease of implementation in real-time applications. However, despite the promising performance and enhanced capabilities of the full-order dynamic model, the complexity of the full-order model can cause a high computational burden on digital signal processors (DSP) for real-time applications. In order to address this challenge, a stability criterion is developed from a simplified inverter model to determine the stability margin, thus predicting the inverter behavior for commanded setpoints. This criterion serves as a predictive tool for ensuring the safe operation of the inverter. One notable advantage of the stability criterion is its ability to quickly estimate the gain margin of the controller without requiring knowledge of pole locations. Experimental tests are conducted on hardware to validate the effectiveness of the developed stability criterion. While the stability criterion successfully reduces the computational burden, it is important to note that the simplified criterion is inherently less accurate compared to the full-order inverter model. This is due to its incapability to capture the intricacies and dynamics of the system fully. To address this limitation, this dissertation proposes novel hybrid dynamic models by combining the data-driven model with the analytical model. This work presents two forms of hybrid dynamic models using a data-driven neural network platform. The first hybrid model uses analytically evaluated risk factors as one of the inputs of the neural network to predict stable and unstable operation as the output of the neural network. The other hybrid model uses the data-driven neural network and developed stability criterion in parallel to predict the stable and unstable operation. Furthermore, a standard neural network is implemented as a benchmark that considers all relevant information as input and predicts whether the operation of the inverter is safe or unsafe. This benchmark highlights the merits of the developed hybrid dynamic models. The effectiveness of the hybrid dynamic models is thoroughly evaluated through simulation and hardware tests to ensure their practical applicability.