Model-Free Predictive Current Control Under Ultralocal Model of FTPMV-RDM Based on Extended State Observer with Optimized Feedback Gains and Precise Input Gains

Traditional model-based predictive current control (MBPCC) is highly dependent on motor parameters. Although model-free predictive current control (MFPCC) based on a ultralocal model extended state observer (ESO) reduces this dependence, the optimality of feedback gains and the accuracy of input gains remain major obstacles to achieving high robustness. To address these issues, this paper proposes a novel MFPCC for a six-phase fault-tolerant permanent magnet vernier rim-driven motor (FTPMV-RDM). The first innovation is a new approach for optimizing the ESO feedback gains using artificial neural networks, which leverage strong data learning and generalization capabilities to improve parameter coverage density and coordinate performance indicators with global optimality through a flexibly designed fitness function. The second is a novel ESO driving strategy that not only uses estimation error to estimate total disturbance, but also employs a high-update-rate prediction error model to solve the system input gain in real time. The third is the introduction of a reward-and-punishment mechanism to adaptively adjust the filter coefficient, ensuring its stability and optimization after input gain estimation. Furthermore, the candidate voltage vector set is screened to mitigate adverse effects of zero-sequence excitation, and the input gain is alternately solved by x-axis and y-axis prediction errors in the x-y harmonic plane, effectively addressing secondary control objectives. Experimental results, compared with traditional MBPCC and MFPCC, confirm the effectiveness of the proposed method.