Abstract：Aiming at the problem of large output voltage fluctuation and long recovery time of three-phase

pulse-width modulation（PWM）rectifiers when the load changes，an improved single-neuron gradient learning

control strategy was proposed. Due to the poor adaptability of the traditional PI controller parameters when the load

changes，a single neuron PI control was adopted in the voltage outer loop，and the gradient descent method was used to adjust the weight parameters online. In order to avoid falling into a local optimal solution during the solution process，a stochastic gradient descent algorithm with restart function（SGDR）was used，and cosine annealing was used to change the learning rate of the weights to improve the convergence performance of the algorithm. Through Matlab and hardware-in-the-loop simulation experiments，the dynamic response performance of the voltage outer loop of the three-phase PWM rectifier under different control algorithms was compared and analyzed. The results show that the three-phase PWM rectifier controlled by the improved single neuron PI algorithm has smaller voltage fluctuation，faster dynamic response and more stable operating state when the load changes.

      Key words：rectifier；voltage outer loop；single neuron；gradient learning；cosine annealing；load perturbation