Gao, Yuan
(2022)
Application of search algorithms and machine learning on aircraft electrical drive systems: design, optimization and control.
PhD thesis, University of Nottingham.
Abstract
The More-Electric Aircraft (MEA) has been introduced to lower the aircraft weight, improve the aircraft operation efficiency and reduce the environmental impact of aviation. Within the MEA concept, there is an evolutionary trend towards more-electric configurations by replacing the conventional mechanical, pneumatic, and hydraulic systems. To this end, there is a requirement of increasing number of on-board electrical drive systems. Likewise, within hybrid/all-electric propulsion concepts, electrical drives play a pivotal role in handling the drivetrain electromechanical power conversions. However, the great challenges concerning the wide applications of aircraft electrical drives are in achieving high power density and high efficiency during their design, optimization and control.
The main target of this PhD thesis is to develop the hybrid application of optimization and machine learning (ML) algorithms on the aircraft electrical drive systems. This hybrid approach considers three subsystems within the aircraft electrical drive system: an input DC filter, a two-level converter and an AC load (three-phase motor or linear loads). The novel contribution of this work is in combining search algorithms and ML to investigate the optimization-based design and control topics of an aircraft electrical drive system, which comprises of that three subsystems.
There are three optimization-based design cases and one control case in this thesis. The first part presents the design and optimization of the input passive filter components. Instead of a stand-alone capacitor, an LC filter is applied at the front-end DC interface because it can save the filter volume/weight with smaller required passive values. Sizing models of inductor and capacitor are given to serve as the basis in this optimization-based design topic. More importantly, a novel algorithm category is introduced to solve the filter Multi-Objective Optimization Problem (MOOP). Two categories are search algorithms and surrogate (ML) algorithms, both of which are tried and finally compared in the case study. Moreover, three filter designs were selected from the optimization results for the experimental validation where, an inverter-based constant power load (CPL) was used as a particular example tested with these three different LC filters.
Moreover, an ML-based inverter control topic is also studied, also in the subsystem level. The main motivation is to address the heavy computation burden of model predictive control (MPC) strategy. After collecting large number of samples from the original MPC operation, a ML model can be trained to imitate how the control algorithm performs during the real-time application. The theory basis of this ML imitation approach is that the MPC process is completely deterministic: for the same set of inputs (i.e. circuit measurements) and a given cost function, the control outputs will always be the same. Two different imitation controllers are designed for the converter case, both of which are tested and compared by the simulation and experiment results.
The second optimization-based design case is to investigate the ML aided MOOP of motor, the third component/subsystem in the drive system. The same with the first optimization case, there are two objectives: mass and power losses; however, the utilization of ML is totally different. In this case, neural network, a well-known ML technique, is trained as a surrogate model mapping from design variables to the dedicated correction factors for the motor MOOP. The main motivation is that analytical motor models usually prone to loss the prediction accuracy in a large design space, compared to the finite-element analysis (FEA). With the help of the trained NN, the analytical predictions can be extremely close to the FEA outputs in the whole design space. Therefore, the analytical base models and the well-trained NN(s) can then become an excellent surrogate for the motor global optimization, which would be extremely fast and accurate. Finally, this NN-aided approach is compared with the conventional FEA-based optimization (using search algorithms).
Finally, the third optimization-based design case is to study the optimization problem of an entire motor drive system which comprises a two-level converter and a three-phase motor. Therefore, this study is for the system-level optimization. Though there is only one objective in this optimization problem, the NN application as a correction surrogate model is similar with the second case. In order to get the target data for the NN surrogate training, motor FEA model and circuit model are jointly simulated in a loop (on separated platforms but both called by the codes on MATLAB). After the data collection of small number of designs, the desired NN can be trained extremely fast whose performance is validated by other samples (rather than the training data). At last, based on the large number of samples in the design space, the analytical base models and the well-trained NN can generate the global optimization of the motor drive system, also in the short time. For the training of each NN, only if the training performance is good and/or the predictions are close to a new data set, the NN can be referred as a well-trained NN and then used in following optimization.
In this thesis, the key role is not focusing on the algorithms themselves, but on their applications. All the used search and machine learning algorithms have been developed in the previous studies. However, this thesis first proposes a simple and clear categorization for the commonly used algorithms to discuss their differences and connections. Based on that, this thesis then develops how to use these algorithms to address optimization and control problems for an MEA electrical drive system. Therefore, the main contribution of this thesis should be the application of different types of algorithms on the aircraft drive systems.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
![[img]](/style/images/fileicons/application_pdf.png)