(362d) Optimal Cleaning Scheduling and Control of Heat Exchanger Networks: An Industrial Case Study

Petroleum refining processes are still subject to a large number of inefficiencies that compromise their operation, safety, and profitability. Fouling of heat exchangers is a major source of energy inefficiencies because the deposition of unwanted material over the surfaces decreases the heat transfer rate and increases pressure drops, which are reflected in a higher energy consumption, lower throughputs and eventual shut downs.

Periodic cleaning of the units is an effective mitigation strategy. However, in a large network it is not easy to decide the sequence of cleanings or whether is better to clean a unit or multiple units at a specific time, such that the overall operational cost is minimum and the operation satisfies a large number of constraints. Controlling flows through bypasses and parallel branches is also commonly used. Simultaneous schedule and control optimisation based on quantitative models is in principle very useful but poses very challenging formulation and solution problems. Detailed dynamic models of crude oil fouling can describe the process accurately over wider ranges because of the many variables and factors used to capture the underlying phenomena, and are useful for simulation and control. However semi-empirical fouling models can be more easily tuned using plant measurements, such that they can predict the behaviour of a specific network in the desired range of operating conditions, and used in combinatorial schedule optimisation.

Here, the optimal cleaning scheduling and flow control of a heat exchanger network is formulated as a MINLP problem with realistically complex models, with the objective to minimize operational cost over a long horizon (~years). The main decision variables are the sequence and timing of cleanings of the units of the network, and the flow rates (e.g split fraction) within the network. This problem has two main difficulties: i) the combinatorial nature given by the large number of binary variables, and ii) the large size and large number of nonlinearities that arise from the fouling model and the heat exchanger model. To handle these difficulties, we propose to model the discrete decisions as complementarity constraints, relaxing the binary variables and solving a sequence of relaxations of the original problem until the final solution satisfies all the constraints of the original MINLP problem.

This work presents an industrial case study of the hot end of a refinery preheat train for which measurements are available, including a record of the cleanings implemented during the operation period. The case study is used to demonstrate the capabilities of the proposed approach. First, the plant measurements are used to fit the parameters of two models: i) a distributed model that considers axial and radial variations within the heat exchanger and the fouling deposit, and ii) a lumped model in the axial direction for the heat exchanger, but that considers radial variations in the heat transfer and deposit growth. The lumped model is used within the MINLP formulation, the solution of which utilises a relaxation of complementary constraints to obtain an optimal cleaning schedule and flow distribution for the entire network. Then, the results are compared and validated using the more detailed model. Finally the potential savings of the optimal operation are compared against the savings obtained in the actual operation.