Research on Structure of Shallow BPNN-PID

With the advantages of simple structure, mature technology and powerful self-learning ability, BPNN has been popularly used in tuning parameters of PID algorithm. However, there has a little theoretical guidance and experimental verification for building the structure of BPNN in BP-PID. In this paper, the structure of BPNN, activation function of hidden layer, and the initial weights are studied and simulated through grid search. The experimental results show that a three-layer BPNN, whose layer neuron number is 5 or 9 with an activation function is Tanh or ELU can be the most useful structure than other choices. Moreover, during the tracing process, the exported NN weights used for fine-tuning further training, when the tracing error is equal to or less than 0.001, can shorten the whole training time and improve the fitting ability of BPNN. This paper can be a theoretical guidance for further research of NN-PID area.