Integrated process unit control and plantwide optimisation

The objective of Dynamic Real-Time-Optimisation is to determine optimal operating trajectories for a process plant based on (possibly changing) economic and operational objectives. These trajectories comprise manipulated and control variables and are passed on to the model predictive control module, which takes care, that the plant follows the specified path of operation. Since dynamic optimisation is still not state-of-the-art for applications of industrial relevance, algorithms and software tools have to fulfil the key requirements to deliver a solution accurately, robustly and sufficiently fast. Within the INCOOP project, significant progress has been made towards improvements of algorithms to accomplish these goals. Two principle dynamic optimisation methods were considered: the so called sequential (at RWTH-LPT) and simultaneous (at CMU) approaches. Algorithmic improvements for the sequential approach method include a new, efficient sensitivity integration method and an automated grid adaptation strategy, which increase the robustness and computational performance especially for large-scale process model. For the simultaneous method, algorithmic improvements such as an advanced interior point and filter line-search method have been introduced. In parallel MDC has extended their optimisation code to better exploit the structure of D-RTO problems. These extensions include removing limits on problem size, dealing with sparsity, improved use of dual-space and warm start. These techniques also have been implemented into software tools, as mentioned in the section 7 Dynamic real-time optimisation software tools. The algorithms provide potential for future enhancements and future (real-time) applications.

The objective of this work has been to design and implement dynamic data reconciliation and state reconstruction techniques. These techniques focus on two fields of application: - To provide the model predictive controller the latest estimate of the dynamic process variables including the states and the disturbances; - To enable model based "measurement" of process variables for which no direct information is available from sensor systems.

A key objective of INCOOP has been the application of real-time dynamic optimisation techniques to industrial processes. Therefore the computational performance of the algorithms is a critical issue. For most industrial problems, a major part of the computation time is spent in the solution of the describing models - usually a system of differential-algebraic equations (DAEs) - and not only rather in the optimisation algorithm itself. While the inexorable increase in computing power hardware will contribute to a steady improvement in application performance this is insufficient. It is also necessary to speed-up the numerical solution of the DAEs model by simplification (reduction) of the model equations. Three strategies for model reduction have been considered: Proper Orthogonal Decomposition (POD), the automatic elimination of redundant trivial equations and physical insight based model reduction.

The newly developed techniques for dynamic optimisation have been implemented as software tools. The sequential and the simultaneous approach prototype software tools (DyOS at RWTHLPT and DynoPC at CMU) have been developed and successfully applied to the end-users benchmark problems. The application showed a significant potential for improvement of the plant operation. It was considered neither technically nor commercially viable to develop an entirely new software product as a part of INCOOP. Instead it was decided to augment MDCs existing steady-state product, RTO+®, with the capability to prototype dynamic RTO. Therefore, MDCs software developments focused on the following: - An OPC Client Interface. - Extension of RTO+® to enable D-RTO to be prototyped. - Example systems for a continuous and batch optimisation. -- Series of stirred tank reactors -- Batch fermentation

The objective has been to develop high performance non-linear model predictive control (NL-MPC) techniques that enable broad bandwidth, high performance process control along optimal trajectories, obeying constraints. The control systems have to support close tracking of the dynamic trajectories obtained from the Dynamic Real-Time Optimisation. Furthermore the control systems have to suppress the effects of disturbances on process behaviour and have to handle the impact of model inaccuracies on closed loop process responses. Accurate, on-line accurate tracking of the trajectory delivered by the dynamic real-time optimiser requires that the HP-NLMPC module respect the dynamic constraints of the process. The HPNLMPC module solves on-line a constrained optimisation problem and determines an optimal control input over a fixed future time-horizon, based on the predicted future behaviour of the process and on the desired reference trajectory. A number of HP-NLMPC schemes have been investigated, implemented and thoroughly tested. Given the initial status of the process, estimates of disturbances and the reference trajectories, the optimiser in the HP-NLMPC module produces the manipulated variable such that input and output trajectories follow the reference trajectories as close as possible subject to the constraints imposed in the optimisation. The MPC module is operating in delta-mode, which means that the differences between actual trajectories and reference trajectories are treated, in a well-defined sense, within the algorithm.

To support the field tests of the newly developed high performance non-linear model predictive control and state estimation techniques a prototype test environment has been built on the basis of the IPCOS data server, IPCOS INCA tools, gPROMS and Matlab. The prototype software tools have been designed to enable sufficiently robust testing in an industrial production environment.

In this work three successive generations of hybrid models have been developed (first-principles models extended with some elements of empirical models) for the selected benchmarking plants to mimic their dynamic behaviour. Each model covers a pre-specified window of operation. The successive model generations will represent increasing levels of performance. The modelling environment chosen for INCOOP is gPROMS because of its state-of the-art functionality for simulation of complex dynamic systems. The key results of this work are: - Four generations of dynamic model of a large scale petro-chemical plant. The latest model has been validated against available plant data. - A dynamic model of a polymerization plant. - Improved methods for: -- Model development: a model synthesis method has been developed that is better suited for large-scale problems than existing techniques. -- Model validation: wavelets have been used to validate the dynamic model for the different frequencies present in the plant.

Livrables

Partager cette page Partager cette page sur les réseaux sociaux

Télécharger Télécharger le contenu de la page