(374a) Novel Plant Modelling Paradigm Eliminates Plant Level Nonlinearities, Incorporates Naturally Hybrid Models, and Corresponding Software Architecture

Current plant wide models fall represent streams either as bulk material or energy flows (planning or scheduling) or mole flows and fractions (design, rating, or RTO). Disconnect between these two paradigms means that the models developed for RTO can not be used for planning or scheduling due to their complexity and convergence difficulties, while planning models are not sufficiently accurate for optimization of plant operations.

Properties of materials (e.g. C5 to C10 hydrocarbons density, heat capacity, latent heat) expressed per unit mass are of similar order of magnitude. They change very little if material composition changes. Hence, bulk properties of streams in a plant can be accurately calculated by local approximation around some base conditions. In an existing plant, a stream leaving some equipment stays at approximately the same base conditions with some minor temperature and pressure variations. Therefore, if the properties at the base conditions are known, local approximations can compute accurately the update properties, provided mass units are used to measure quantity of materials.

First principles models carry stream mole fractions in order to be able to calculate stream properties and phase fractions at given temperature and pressure. This paradigm has its roots in Cavett’s [1] three flashes with recycle problem. For an operating plant, the properties at the base conditions can be calculated from rigorous thermodynamics or can be determined experimentally. This leads to a conclusion that we do not need mole fractions to represent streams in an existing plant. If base condition properties and their local approximations are not sufficiently accurate (which is hardly ever the case), then the ode models can calculate bulk stream properties by invoking rigorous thermodynamics. Therefore, the stream properties are calculated by approximate thermodynamic and updated if needed via rigorous thermodynamic calculations.

One should note that in many instances, e.g. iron ore smelting, food production, and others, rigorous thermodynamic calculations are not possible and that one has the rely on the bulk properties determined experimentally. Proposed modelling paradigm incorporates such situations in a straightforward manner.

Consequently, one can use simplified thermodynamics and component mass flows to describe the streams, thereby eliminating nonlinear equations associated with stream flash calculations or stream mixing. Such flowsheet-level abstraction is also well suited to incorporation of data driven, hybrid models.

This plant modelling paradigm, based on bulk and component mass flows and approximate thermodynamics has been implemented in Python as an extension of PYOMO [2].

Different stream classes can be used in different parts of the network. Stream classes are defined in system tables, which makes it possible to instantiate the software in domain specific incarnations, e.g. pulp and paper, mining, petrochemicals, refining, etc. making it possible to model e.g. systems comprised of electricity generation units, petrochemical plants, pipelines, and local distribution systems. Such models are becoming more and more important, as we move towards optimizing plants that consume or produce energy and need to include electricity price variations and CO2 footprint in the objective functions or in the constraints.

New node models are added to the library by model writer coding a Python function that generates equations in Pyomo compliant syntax. If a node is described by a data-driven model, it’s surrogate model suitable for incorporation into an equation-oriented paradigm is included as the node model.

Implementation in Python makes machine learning/AI tools natively available in the same environment, opening avenues for straight forward inclusion of data driven node models into the network.

A case study of modelling a blue hydrogen plant comprised of electricity generation unit, air separation unit, natural gas reforming, and hydrogen distribution network is presented. Stream thermodynamic properties at the base conditions have been computed by Aspen Plus [3] plant model.

Different model abstractions are created by selecting corresponding sets of equations, e.g. mass balances, or energy balances, or mass and energy balances, or etc.

Comparison shows that the hybrid plant model is on par with Aspen Plus model in terms of accuracy. The key advantage is that the hybrid plant model can be used for multi-period optimization (24+ periods) with very short execution times, which is not achievable by using AspenPlus.

References

[1] Cavett, R. H., “Application of Numerical Methods to the Convergence of Simulated Processes Involving Recycle Loops”, American Petroleum Institute, 43, 57 (1963).

[2] Hart, W.E., Laird, C.D., Watson, J.P., Woodruff, D.L., Hackebeil, G.A., Nicholson, B.L., Siirola, J.D., 2017. Pyomo — Optimization Modeling in Python, Springer

[3] Aspen Plus, https://www.aspentech.com/en/products/engineering/aspen-plus