Recurrent neural network structures for learning control valve behaviour

Control valves are ubiquitous in process control, yet they are rarely explicitly modelled. Here, we propose a theoretical valve model as a recurrent neural network (RNN) cell, allowing its parameters to be learned with gradient descent methods. Further we alter the theoretical model by incorporating a one-dimensional neural network. The models are capable of predicting valve opening from its reference value and can be easily combined with other neural network layers. We compare their performance to two long short-term memory networks (LSTMs) and showcase the performance improvements of our suggested physics-based models. In particular, we present how a gradient descent based learning algorithm finds parameters that lead to improved performance by the original theoretical valve model. The learning experiments are carried out on datasets from two different modes of operation, and we explore whether parameters that are suitable both modes can be found. The results show the benefit of using a physically inspired model for learning, like interpretable parameter values.