Nonlinear Model Predictive Control using Neural ODE Replicas of Dynamic Simulators

We propose simulation-based nonlinear model predictive control as a first step towards autonomous decision-making for stable operation of large complex dynamical systems such as chemical plants. The effect of abrupt external disturbances should be quickly eliminated, taking into account such complex dynamic responses, to maintain stable production. In this paper, we propose a control system to eliminate these effects. The system uses engineering models, including dynamic simulators, based on chemical engineering knowledge. Dynamic simulators are generally not differentiable with respect to actions; however, differentiable models are advantageous for fast nonlinear optimization. To take advantage of both reliable dynamic simulators and differentiable models, we introduce neural ordinary differentiable equation models and clone the behaviour of simulators on them. The cloned differentiable neural replica model is then incorporated into a gradient-based nonlinear model predictive control. Evaluation of this method in a real methanol distillation plant confirms that it can significantly remove abrupt heavy rain disturbances compared to existing methods.