Algorithms for Industrial Demand-Side Management Under Uncertainty
Lead Research Organisation:
Imperial College London
Department Name: Computing
Abstract
Uncertainties are present in many energy-related process design (e.g., how should a process be configured?) and operational (e.g., what is the best production schedule for a day/week?) optimisation problems of current industrial interest. The efficiency of an energy-intensive hydrogen production plant can be greatly improved by optimising the steam-methane reformer, but design decisions regarding the reformer are subject to uncertain catalyst performance. Likewise, an electricity-intensive air separation unit can derive economic savings and reduce peak power demand by engaging in demand-response; however, deciding optimal production schedules relies on uncertain forecasts of electricity supply and product demand. Regrettably, state-of-the-art software is not suitable for decision-making under these uncertain conditions, severely limiting the benefits of industrial demand-side management (DSM) towards national energy efficiency. Here, DSM refers to measures of improving the energy system at the side of consumption, ranging from reducing overall demand by increasing process efficiencies to smarter consumption patterns through demand response operation. Demand response (DR) operation aims to increase the systemic integration of volatile renewable energy sources by matching consumption to the short-term and long-term (daily to seasonal) fluctuations in supply.
Motivated by the above, this interdisciplinary project will introduce Algorithms for Industrial Demand-Side Management Under Uncertainty. The potential of curtailing carbon emissions through improving the efficiency of energy-intensive process industries is massive, with industrial entities comprising 17% of total energy consumption in the United Kingdom in 2017. DR operation in the electricity-intensive process industries further reduces carbon emissions by synchronising demand with renewable-based generation. Therefore, a complete DSM decision-making toolkit must consider uncertainty in both design and operational decisions of process systems. In modern environments, these tools must also be computationally scalable, synergise with the abundant available data, and accompany decisions with rationale. The proposed scientific advances have numerous immediate applications: optimising energy efficiency in manufacturing, balancing the power grid through DR, and mitigating negative effects of disturbances.
The primary observation of the proposed research is that modern markets and environments dictate a deviation from the accepted paradigm of deterministic (i.e., no uncertainty is modeled), local (i.e., risks sub-optimal decision-making) optimisation. The process industries require a new generation of decision-making algorithms that can solve, and re-solve, large-scale optimisation problems to global optimality, often in an online or recurring fashion. The proposed research introduces DSM technologies that: (1) automatically decompose process models for global optimisation, (2) exploit historical operating data for planning and scheduling, and (3) produce explainable results for user-friendly re-optimisation. The fellowship will be held at the Department of Computing at Imperial College, which has an outstanding reputation and provides an ideal environment for the proposed software advances. Imperial is also the birthplace of the field of process systems engineering (PSE) and thus is a premier forum for applied PSE research. By providing freely available software tools, we will contribute to the forefront of PSE, as well as relevant related domains of optimisation theory, data science, and artificial intelligence. Finally, promoting the algorithmic advancements by releasing and contributing to open-source software will spur new academic and industrial applications in computational decision-making for energy efficiency.
Motivated by the above, this interdisciplinary project will introduce Algorithms for Industrial Demand-Side Management Under Uncertainty. The potential of curtailing carbon emissions through improving the efficiency of energy-intensive process industries is massive, with industrial entities comprising 17% of total energy consumption in the United Kingdom in 2017. DR operation in the electricity-intensive process industries further reduces carbon emissions by synchronising demand with renewable-based generation. Therefore, a complete DSM decision-making toolkit must consider uncertainty in both design and operational decisions of process systems. In modern environments, these tools must also be computationally scalable, synergise with the abundant available data, and accompany decisions with rationale. The proposed scientific advances have numerous immediate applications: optimising energy efficiency in manufacturing, balancing the power grid through DR, and mitigating negative effects of disturbances.
The primary observation of the proposed research is that modern markets and environments dictate a deviation from the accepted paradigm of deterministic (i.e., no uncertainty is modeled), local (i.e., risks sub-optimal decision-making) optimisation. The process industries require a new generation of decision-making algorithms that can solve, and re-solve, large-scale optimisation problems to global optimality, often in an online or recurring fashion. The proposed research introduces DSM technologies that: (1) automatically decompose process models for global optimisation, (2) exploit historical operating data for planning and scheduling, and (3) produce explainable results for user-friendly re-optimisation. The fellowship will be held at the Department of Computing at Imperial College, which has an outstanding reputation and provides an ideal environment for the proposed software advances. Imperial is also the birthplace of the field of process systems engineering (PSE) and thus is a premier forum for applied PSE research. By providing freely available software tools, we will contribute to the forefront of PSE, as well as relevant related domains of optimisation theory, data science, and artificial intelligence. Finally, promoting the algorithmic advancements by releasing and contributing to open-source software will spur new academic and industrial applications in computational decision-making for energy efficiency.
Planned Impact
In addition to the academy, the proposed research has a high potential to impact a wide variety of interested parties, both in the UK and internationally. These parties may include, but are not limited to, process industries, software developers, policy makers, and the broader public.
In the process industries, two major impediments to the adoption of advanced optimization technology are difficulty in formulating large optimisation problems, and the reluctance of plant operators to implement any optimal solution they do not understand. The proposed research address both of these themes, and I will catalyse industrial partnerships by i) seeking industrial case studies, ii) tailoring the proposed decision-making tools to industrial problems, iii) releasing a finalised version of the algorithm(s) as open-source software. For example, the fellowship partners with a software vendor supporting the process industries, as well as a global industrial operator of energy-intensive processes. I expect that the decision-making tools introduced in this fellowship will help process industries participate in demand-side management programs, lowering manufacturing costs and increasing the penetration of renewable-based electricity generation.
In terms of software developers, I will collaborate with commercial software entities to implement the novel decision-making capabilities in one or more widely adopted software package(s). In addition, my codes will be made open-source, well-documented, and reusable, allowing elements to feed into future open-source and/or commercial optimization software packages. The fellowship will introduce new algorithms that will advance the state-of-the-art in optimization software, encouraging more software developers to enter this domain. The compiled suites of test problems will allow software developers to better align their offerings with industrial demand-side management problems.
For policy makers, the tools will be open-source and available to policy and decision-makers in the international community, allowing them to identify when, and to what extent, demand-side management policies are possible in the process industries. This could help guide policy makers in selecting regulatory structures for relevant industries.
To reach the broader public, I use online forums to educate a broad community on the developments and open-source software from this proposal. To educate the next generation of users, I will also promote developed tools on the educational sections of websites in the areas of chemical and process systems engineering. I will attract undergraduate and graduate researchers to the work by offering scholarships to participate in educational workshops that I will hold at UK academic hubs.
In the process industries, two major impediments to the adoption of advanced optimization technology are difficulty in formulating large optimisation problems, and the reluctance of plant operators to implement any optimal solution they do not understand. The proposed research address both of these themes, and I will catalyse industrial partnerships by i) seeking industrial case studies, ii) tailoring the proposed decision-making tools to industrial problems, iii) releasing a finalised version of the algorithm(s) as open-source software. For example, the fellowship partners with a software vendor supporting the process industries, as well as a global industrial operator of energy-intensive processes. I expect that the decision-making tools introduced in this fellowship will help process industries participate in demand-side management programs, lowering manufacturing costs and increasing the penetration of renewable-based electricity generation.
In terms of software developers, I will collaborate with commercial software entities to implement the novel decision-making capabilities in one or more widely adopted software package(s). In addition, my codes will be made open-source, well-documented, and reusable, allowing elements to feed into future open-source and/or commercial optimization software packages. The fellowship will introduce new algorithms that will advance the state-of-the-art in optimization software, encouraging more software developers to enter this domain. The compiled suites of test problems will allow software developers to better align their offerings with industrial demand-side management problems.
For policy makers, the tools will be open-source and available to policy and decision-makers in the international community, allowing them to identify when, and to what extent, demand-side management policies are possible in the process industries. This could help guide policy makers in selecting regulatory structures for relevant industries.
To reach the broader public, I use online forums to educate a broad community on the developments and open-source software from this proposal. To educate the next generation of users, I will also promote developed tools on the educational sections of websites in the areas of chemical and process systems engineering. I will attract undergraduate and graduate researchers to the work by offering scholarships to participate in educational workshops that I will hold at UK academic hubs.
Publications
Bermúdez J
(2023)
33rd European Symposium on Computer Aided Process Engineering
Burtea R
(2023)
33rd European Symposium on Computer Aided Process Engineering
Ceccon F
(2022)
OMLT: Optimization & Machine Learning Toolkit
in Journal of Machine Learning Research (JMLR)
Folch J
(2023)
Combining multi-fidelity modelling and asynchronous batch Bayesian Optimization
in Computers & Chemical Engineering
Folch JP
(2022)
SnAKe: Bayesian Optimization with Pathwise Exploration
Thebelt A
(2022)
Multi-objective constrained optimization for energy applications via tree ensembles
in Applied Energy
Thebelt A
(2022)
Maximizing information from chemical engineering data sets: Applications to machine learning
in Chemical Engineering Science
Tsay C
(2021)
Sobolev trained neural network surrogate models for optimization
in Computers & Chemical Engineering
| Description | The work has fallen predominantly in Aims 2&3 of the fellowship proposal: hybridising data-driven modelling in optimisation, as well as optimisation methods and software. Work funded through this award has achieved new optimisation and modelling methodologies for deploying neural-network surrogate models in optimisation, i.e., using neural network models to replace (part of) an optimisation model. This can potentially expand the scale of process optimisation problems that can be tackled computationally, and we have released a generic open-source tool called OMLT: Optimisation and Machine Learning Toolkit. Side projects have included deploying optimisation over surrogate models in black-box optimisation contexts, such as in Bayesian optimisation or reinforcement learning. We have also briefly studied other surrogate model forms, such as tree-kernel Gaussian processes. |
| Exploitation Route | Notably, OMLT, our open source tool resulting from this funding, allows other users to easily deploy the surrogate modelling techniques developed by this project, as well as other strategies from the literature, which we have also implemented. OMLT receives approximately 20k downloads per month (according to PyPI Stats; accessed March 2023), suggesting there are a wide variety of use cases for these methodologies. |
| Sectors | Chemicals,Digital/Communication/Information Technologies (including Software),Energy,Manufacturing, including Industrial Biotechology |
| Description | Optimisation methodologies have been developed alongside and and incorporated into commercial (e.g., Gurobi) and open-source (e.g., pyomo.GDP) optimisation tools. In turn, this translates to promoting the use of hybrid/data-driven models in optimisation and promotes the nascent research area at the intersection of optimisation and machine learning. |
| First Year Of Impact | 2022 |
| Sector | Chemicals,Digital/Communication/Information Technologies (including Software),Energy |
| Description | The Alan Turing Institute Post-Doctoral Enrichment Awards |
| Amount | £2,000 (GBP) |
| Organisation | Alan Turing Institute |
| Sector | Academic/University |
| Country | United Kingdom |
| Start | 03/2022 |
| End | 10/2022 |
| Title | OMLT: Optimization and Machine Learning Toolkit |
| Description | OMLT is a Python package for representing machine learning models (neural networks and gradient-boosted trees) within the Pyomo optimization environment. The package provides various optimization formulations for machine learning models (such as full-space, reduced-space, and MILP) as well as an interface to import sequential Keras and general ONNX models. |
| Type Of Technology | Software |
| Year Produced | 2022 |
| Open Source License? | Yes |
| Impact | OMLT is an open source software package incorporating surrogate models, which have been trained using machine learning, into larger optimisation problems. Computer science applications include maximizing a neural acquisition function and verifying neural networks. Engineering applications include the use of machine learning models to replace complicated constraints in larger design/operations problems. OMLT 1.0 supports GBTs through an ONNX (https://github.com/onnx/onnx) interface and NNs through both ONNX and Keras interfaces. We discuss the advances in optimisation technology that made OMLT possible and show how OMLT seamlessly integrates with the python-based algebraic modeling language Pyomo (http://www.pyomo.org). The literature often presents different optimization formulations as competitors, but in OMLT, competing formulations become alternatives: users can select the best for a specific application. We provide examples including neural network verification, autothermal reformer optimization, and Bayesian optimization. |
| URL | https://github.com/cog-imperial/OMLT |