Distributed Predictive Control of Large-Scale Neural Network Models Identified through Physics-Based Learning Procedures. Application to a Cogeneration Plant.
Motivated by increasing applications to energy and manufacturing processes, this PhD thesis project focuses on developing novel control solutions for controlling large-scale plants. Traditional centralized control approaches face challenges in controlling such systems since they suffer from robustness, scalability, and computational complexity issues. Moreover, the application of distributed control approaches commonly developed for linear systems might be restrictive due to the complexity and nonlinearity inherent in these processes.
Exploiting the increasing sensorization of industrial plants, our proposed methodology aims to integrate recurrent neural networks (RNNs) modelling with distributed model predictive control approaches. The key idea lies in training physics-based RNNs, i.e. imparting system-specific knowledge during training, with the aim to not only improve model accuracy and interpretability but also to facilitate the development of distributed control algorithms. Firstly, unlike conventional distributed approaches that decompose a centralized plant model in multiple subsystems, we aim to identify subsystems directly from the data. Secondly, during the training phase, our idea is to develop a model capable of mimicking the sparse architecture of the plant, while also imparting to each subsystem model a structure suitable for control design. This could involve, for instance, to tackle in each model the disturbances generated by the influence of dynamic neighbouring subsystems. Then, we can develop distributed control algorithms by leveraging the intrinsic structure and properties of the neural network. To facilitate the application of the approach to real scenarios, we consider a dataset suitably augmented with artificial signals introduced based on the physical knowledge of the plant to supply for unmeasured variables.
To validate and demonstrate the effectiveness of the proposed approach, the developed algorithms will be tested in controlling an existing cogeneration power plant dedicated to the combined generation of electrical energy, heat, and steam.
Back to Current Students