Applications of artificial intelligence in the circuit and modulation design of DC-DC converters

• Main Author: Li, Xinze
• Other Authors: Mao Kezhi
• Format: Thesis-Doctor of Philosophy
• Language: English
• Published: Nanyang Technological University 2023
• Subjects: Engineering::Electrical and electronic engineering::Power electronics; Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: https://hdl.handle.net/10356/165390
• Institution: Nanyang Technological University

Threatened by the global warming and the depletion of fossil fuel, renewable energy such as solar energy and clean transportation solutions like electric vehicles are attracting more attentions nowadays, which appeal for a future with more power converters. Among all types of power converters, DC-DC converters are the key enablers for DC voltage and power regulation. To realize an optimally performed DC-DC converter, the circuit parameter aspect and modulation aspect of DC-DC converters should be designed with great care. However, the conventional approaches for the circuit parameter and modulation design of DC-DC converters suffer from two nontrivial challenges: heavy manpower burden and low design accuracy. The issue of heavy manpower burden is primarily attributable to the massive human-dependence in the analysis and deduction process of design objectives. The issue of low design accuracy is mainly caused by the model simplification and mathematical approximations. Fortunately, the advanced artificial intelligence (AI) techniques such as neural network (NN), metaheuristic algorithm (MHA), and fuzzy inference system (FIS), can be applied to solve those challenges. Consequently, this thesis focuses on the applications of AI techniques in the circuit and modulation design of DC-DC converters to realize high-level automation while maintaining high design accuracy. To highlight the objectives and motivation of this thesis, Chapter 2 gives a comprehensive literature review targeting the circuit parameter design and modulation design of DC-DC converters. In the beginning, Chapter 2 first introduces the DC-DC converters, followed by a detailed review of circuit parameter design approaches for DC-DC converters. Subsequently, the modulation approaches for DC-DC converters are discussed, and this chapter ends with the existing applications of AI techniques in the circuit parameter and modulation design of DC-DC converters. Chapter 2 indirectly emphasizes that the proposed AI-based automated design approaches in this thesis are of great significance to the circuit and modulation design of DC-DC converters. Chapter 3 and Chapter 4 aim at the circuit parameter design approaches for DC-DC converters. In Chapter 3, to achieve a good holistic design of a synchronous Buck converter, efficiency, cutoff frequency, and power density are optimized at the same time. The optimization of the three conflicted design objectives is intrinsically a multi-objective design problem, which is solved by the novel coevolving archived multi-objective simulated annealing (coevolving-AMOSA) algorithm. With the proposed coevolving-AMOSA algorithm, a Pareto frontier with better uniformity and completeness can be obtained. Based on the obtained Pareto frontier, considering various application scenarios, three optimal designs are given. 100 W hardware experiments have been done to verify the three design cases. Since the approach in Chapter 3 still relies on human efforts for the deduction of design objectives, Chapter 4 proposes an AI-based design (AI-D) method, which can achieve full automation in both the deduction process and the optimization process. The proposed AI-D approach is composed of four stages, briefly discussed as follows. First, design specifications are given. Second, lookup tables that contain the practical features of components are built, which make the simulation closer to the reality and thus increase design accuracy. Third, simulation is built and run to generate training data for NN, and the trained NNs serve as the surrogate models for design objectives. Fourth, genetic algorithm interacts with NNs to search for the optimal design. To reflect how the proposed AI-D approach can be applied, given the application scenario of the accessory load supply system in electric vehicle, an efficiency-oriented synchronous Buck converter with constraints on size and ripples is provided. Detailed hardware experiments validate the effectiveness and high design accuracy of the proposed AI-D methodology. Chapter 5 and Chapter 6 study the modulation design approaches for DC-DC converters. To facilitate the automated design idea in Chapter 4 to modulation design, Chapter 5 puts forward an AI-based triple phase shift modulation (AI-TPSM) for dual active bridge converters. In the proposed AI-TPSM, NN, MHA, and FIS are adopted in the deduction stage, optimization stage, and online realization stage, respectively. With the integration of NN and simulation, the deduction process of current stress is automated. With an MHA, the optimal modulation variables under various operating conditions are automatically found. The proposed FIS-based control diagram realizes the online modulation with continuous values, which solves the problem of discreteness in lookup-table-based online modulation approaches. With the proposed AI-TPSM, the optimal current stress can be achieved over the whole load range and voltage range. 1 kW hardware experiments comprehensively validate the optimal current stress and high efficiency performance in steady state, and verify the fast dynamic response under voltage and load steps. Except for single modulation strategy, hybrid modulation is also studied in-depth in this thesis. In Chapter 6, hybrid extended phase shift (HEPS) modulation is considered to reach high modulation performance while keeping simple implementation. The proposed approach in Chapter 6 is also a fully automated modulation design approach, but has several adjustments compared with the method in Chapter 5. First, compared with the current stress considered in Chapter 5, the optimization objectives in Chapter 6 are efficiency and full-range ZVS operation. Moreover, an advanced ensemble learning technique, extreme gradient boosting (XGBoost), is adopted to learn the data-driven surrogate models of efficiency and ZVS performance. In addition, a novel PSO algorithm is proposed to achieve fast convergence speed during optimization process. The proposed AI-based HEPS modulation can realize optimal efficiency and all-switch ZVS operation over entire operating ranges, which have been experimentally validated. Moreover, hardware experiments also validate the satisfactory dynamic performance. Finally, the conclusion of this thesis is summarized. Five potential future research directions are highlighted, including: AI-based automated design considering more design parameters, more design objectives, and more power converters; Temporal behavior modeling of power converters; More complicated modulation strategies; Reinforcement learning for real-time modulation and circuit topology design; Applications of other advanced AI techniques in power electronics.