Industrial utility systems are inherently complex, exhibiting nonlinear, time-dependent behaviour driven by fluctuating demands, operator intervention, and equipment constraints. These non-continuous dynamics challenge traditional steady-state modelling approaches, which often oversimplify system interactions and lead to inaccurate evaluation of heat recovery projects. This paper proposes a generalised, data-driven methodology for developing surrogate models of non-continuous utility systems. Trained on high-resolution historical plant data, the model captures nonlinear, time-dependent system behaviour and is used to predict fuel consumption, cogeneration, boiler steam generation, and equipment-level steam demand across full-year operational periods. The approach is demonstrated using a large pulp and paper mill comprising nine interacting plants and cogeneration. Accurate surrogate models were trained to predict key system responses, including natural gas consumption, swing boiler behaviour, and power generation and cogeneration. The models were then coupled with multi-steady-state simulations to evaluate a heat exchanger network retrofit design. Results show that the standard industry practice of applying a uniform steam cost dramatically overestimates the financial benefits. For the retrofit design, the conventional method predicted NZD 9. 2 million annual savings, whereas the surrogate-model evaluation revealed a net increase in annual operating cost of NZD 0. 61 million due to reduced cogeneration and displacement of the lower-cost biomass in the swing boiler. The findings demonstrate that data-driven surrogate modelling provides a more credible approach for evaluating retrofit designs at non-continuous sites, helping engineering teams avoid economically unfeasible investments.
Hall et al. (Wed,) studied this question.