The landlord's question was six years ago (today, 202 1). I wonder if the landlord is still concerned about this topic now? Neural network self-tuning PID is definitely effective. At present, neural network self-tuning PID mainly faces three problems: one is the choice of initial value, and unreasonable initial value is easy to make the closed-loop system unstable; Second, neural network self-tuning PID itself needs more artificial parameters, PID control itself only needs three artificial parameters, while neural network self-tuning PID needs four (three initial values and one learning factor), which makes neural network self-tuning PID more troublesome than traditional PID algorithm; Third, there is a lack of complete proof of theoretical stability. The online updating rule of neural network self-tuning PID is stable for a long time and is widely cited and applied. However, the stability proof of closed-loop system based on neural network self-tuning PID has not been well solved, which limits the popularization of neural network self-tuning PID to some extent.

I personally have done some such research. If you are interested, you can refer to one of my journal papers.

Data-driven tracking control of discrete nonlinear systems based on LM and relay feedback PID neural network