Data-driven control methods of dual-active-bridge-based grid-connected battery energy storage system

• Main Author: Zeng, Yu
• Other Authors: Josep Pou
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering
• Online Access: https://hdl.handle.net/10356/171350
• Institution: Nanyang Technological University

Carbon dioxide emissions cause global warming and a series of environmental problems. To avoid the worst effects of climate change and protect a habitable planet, replacing fossil fuels with renewable energy sources (RESs) would dramatically reduce carbon emissions. With the large integration of RESs, microgrids (MGs) have become a hot topic. However, the intermitted nature of RESs may cause the oscillations of power and voltage. Meanwhile, the intrinsic issues of the grid interactive devices and environment may cause the instability problems of power grids. Battery energy storage system (BESS) is a promising solution, which can help buffer the power oscillations and maintain the power balance within the grid. Bidirectional direct current (dc)-dc converters have a promising future to integrate the RESs and BESSs into the future microgrid. Among all the dc-dc converters, dual active bridge (DAB)-based converters have attracted much attention due to soft-switching and galvanic isolation. The integration of RESs and BESSs into MGs will bring many challenges to the control of DAB-based converters. The motivation and objectives are using artificial intelligence (AI)-aided methods to solve the existing control challenges in the grid-connected DAB-BESSs. Firstly, an artificial neural network-based active disturbance rejection control (ANN-ADRC) method is proposed to regulate a constant output voltage quickly and accurately under different operating conditions. The ADRC controller is designed based on the small-signal modeling of the DAB-ESSs. Feedforward compensation and uncertainty estimations of the extended state observer help to improve the dynamic performance and to reduce the number of current sensors. After satisfying the conditions of stability analysis, the parameters of the ADRC controller are selected automatically via ANN. The ANN is trained with two inputs (ADRC controller parameters) and two outputs (performance indicators of the ADRC controller). The well-trained ANN can be used as a surrogate model to obtain the optimal solution of the objective function easily and quickly. The proposed ANN-ADRC algorithm is validated to achieve fast dynamic performance under various operating conditions. The aim of this section is to propose an artificial neural network based active disturbance rejection control (ANN-ADRC) method to realize the automatic selection of parameters of the ADRC controller and improve overall performance under real-time hardware experiments. The stability analysis for the ADRC framework is also implemented in consideration of the parameter variations. The proposed method can also save load-current sensors. Then, a multi-agent soft actor-critic (MASAC) enabled multi-agent deep reinforcement learning (MA-DRL) algorithm for output current sharing of the input-series output-parallel DAB (ISOP-DAB) converter is proposed. Compared with tha previous work, the proposed method is expected to balance power among different submodules (SMs) and regulate output voltage. The modular converter is divided into different SMs, which are treated as DRL agents of Markov games. Further, all agents are executed decentralized to provide online control decisions after collaborative training. The proposed MASAC algorithm can show optimal dynamic performance. Next, an MA-DRL-based autonomous input voltage sharing (IVS) control and triple-phase-shift (TPS) modulation method for ISOP-DAB converters are presented to solve the three main challenges: (i) uncertainties of the dc microgrid, (ii) power balance problem, and (iii) inductor current stress minimization of the DAB converter. Specifically, the control and modulation problem of the ISOP-DAB converter is formed as a Markov game with several DRL agents. Subsequently, the MA twin-delayed deep deterministic policy gradient (TD3) algorithm is applied to train the DRL agents in an offline manner. After the training process, the multiple agents can provide online control decisions for the ISOP-DAB converter to balance the IVS, and minimize the current stress among different SMs. Without accurate model information, the proposed method can adaptively obtain the optimal modulation variable combinations in a stochastic and uncertain environment. Finally, a decentralized uniform control method is proposed for the ISOP-DAB converter feeding constant power loads. The MA-DRL-based method is proposed to coordinate multiple control objectives. For the offline training stage, each DRL agent will supervise an SM of the ISOP-DAB converter, receive the states, and output the actions in real-time. Using the TPS modulation, the optimal phase-shift ratio combinations are learned by SAC agents. The parameters of neural networks in agents will be updated to maximize the reward. After the training process is completed, the multiple agents can provide online control decisions for the ISOP-DAB converter with only the local information of each SM. Without accurate model information, the proposed method can adaptively balance the power, and optimize the efficiency of the ISOP-DAB converter in a stochastic and uncertain environment. Hot-swap experiments can be implemented because of the reliability and modularity improvement brought by the decentralized configuration.