Improved neural network PID controller for regulated power supply

The mathematical model of device's main circuit is established and the closed-loop voltage stabilization control method is analyzed. With the strong non-linearity and variable structures and variable parameters, it is difficult to achieve ideal control effects using the classic PID controller. Artificial neural network was combined with conventional PID regulator to construct a neural network PID controller that did not rely on the precise mathematical model of controlled objects. To attain faster convergence speed of the neural network, the Levenberg-Marquardt algorithm was adopted to calculate the updating quantities of connection weights, to which random disturbances retained in certain probability were applied for speeding up the iterative process out of local minima. The device's main circuit together with neural network PID controller was simulated and the results show that the system has quick responses, strong robustness and smooth adjustment. Testing and validation of such controller were also conducted experimentally using a prototype with voltage rating 660 V and volume rating 400 kVA.