Design and development of a FOC algorithm based on machine learning for the HEPS

The High Energy Photon Source (HEPS) is the fourth-generation synchrotron light source with a beam energy of 6 GeV, developed by the Institute of High Energy Physics. As a essential partment of the HEPS, The Fast Orbit Feedback (FOFB) systemhas been developed to maintain the beam orbit stability. In this work, a neural network-based algrithm has been designed and developed to replace the tradional PID control algrithm of FOFB. In this papaer, we have introduced the design and developmenyt of the MLP neural network, which has been trained using operational data from HEPS, and hyperparameter optimization was conducted to improve performance. The trained model was then quantized to support deployment on FPGA hardware. A laboratory test environment was set up, where BPM data was fed into the neural network, and the corrected values for the corrector magnets were output after computation. Experimental results show that the neural network maintained a control error of approximately 1 mA. These results demonstrate the feasibility of using neural networks as an effective alternative to PID control in FOFB systems.