Control-oriented, physics-inspired data-driven Modeling and simulation of the clinker production Pyro process

Official URL: https://risc01.sabanciuniv.edu/record=b3390573

The cement industry is one of the critical components of modern society, playing a vital role in infrastructure and building construction. However, it is meanwhile one of the most energy-intensive and pollutant industries in the world. On the other hand, most current cement production plants are controlled and operated manually making them non-optimal. Implementing novel controllers can help to solve these problems. Model Predictive Controllers (MPC) have shown tremendous potential in this regard since they provide optimal controllers and can consider constraints on inputs and process variables. The essential part of MPC is its dynamic predictive model. While efforts have been made to model the cement production process, there is still a lack of suitable models for implementing MPC in cement production processes. Traditionally, physics-based models have been considered for modeling the cement production process. However, these models are typically complex with a huge number of parameters and computationally time-consuming, making them inapplicable to MPC. Recently, data-driven methods such as system identification and machine learning models have been developed for cement production process modeling. In spite of this, the majority of the literature did not take into account the physical principles and internal dynamics of the process. Further, they did not discuss the performance of their models and directly applied them to MPC. This thesis aims to develop essential predictive models for implementing MPC for the cement production pyro process by considering both the process’s internal dynamics and data-driven methods. Moreover, we investigated the models performance comprehensively with a particular focus on long-term predictions which is essential for the successful implementation of MPC. The first step was to develop a simple and linear control-oriented model suitable for the MPC. Therefore, system identification methods were the focus. Moreover, the first principles of mass and energy conservation laws are considered to discover internal dynamics and inter-component relations in the process. As a result of incorporating these physics insights into systems identification models, a gray-box model has been developed. Next, more sophisticated simulation models were developed to represent the real plant in MPC implementation, since it was not possible to implement the designed controller in the real plant due to high risks and costs. For this purpose, machine learning models particularly sequence modeling machine learning models such as recurrent neural networks, and transformers are used. The selected ML models are modified and implemented on data from the cement production process at the Akcansa Cimento Plant. Both the control-oriented models and simulation models were used for various prediction tasks on the collected data. Results show that in control-oriented models, the gray-box model performs better than the black-box model in validation data, especially for long-term predictions. This depicts the benefit of considering the internal dynamics and inter-components of the process and integrating them into the data-driven model. The results also reveal that the suggested simulation ML models are capable of modeling the cement production process and predicting its future states. Among the proposed ML models, the transformer outperforms others as it exploits the attention mechanism which overcomes RNN problems and can capture long-term dependencies. It should be noted, however, that the selection of a suitable model is dependent upon the objective task, the problem, and the available data. Accordingly, selecting the right model for their intended use would be the user’s responsibility. Lastly, the developed models can be used to design and implement model predictive controllers for the cement production process in the future.