InduFlexControl

Context

Describe the specific need/challenge tackled by the proposed research in the context of the ‘MOONSHOT’ objective, and its relation to both energy intensive industry and energy sector.

In order to transform our energy system to a sustainable, low carbon and climate-friendly eco-system, flexibility will be an important enabler to host more renewable energy sources (RES). More specific, flexibility will be necessary to support stable and secure grid operation on the one hand and to guarantee system adequacy and resilience on the other hand. Within the range of available flexibility options, power-to-X solutions provide the highest flexibility potential, because of their large energy volumes (hundreds of MWh to tens of GWh), high power ratings (tens to hundreds of MW) and long-term storage potential (weeks, months, seasons).

However, power-to-X solutions have a very high technology and infrastructure cost. Although many components do exist, they are not integrated in the industrial processes. Additionally, the energy-intensive industry has been readily optimising process behaviour in many different aspects but has not yet taken full advantage of its inherent flexibility. A fortiori, designing – or redesigning – industrial processes to incorporate power-to-X flexibility will allow the highest CO2 reduction.

In order to make this happen, a fundamental research question has to be solved: how to unlock and control the flexibility of the energy-intensive processes in a way that is best suited for the overall eco-system? This is non-trivial, mainly because it has to consider three aspects at the same time: 

1. the industrial process characteristics
2. the energy market design
3. the power/energy network configuration, all with their constraints.

Major advances in control algorithms are required to fully exploit the latent flexibility in industrial processes and the power-to-X flexibility potential.

Goals

Describe the scientific goals of the “full project”, including the goals and gains to be achieved at full execution of the research trajectory and expected time to realise this, clearly described in toll gates based on “best guess estimations”. In this toll gate plan, indications of critical elements, criteria, constraints, know how or Proof of Concept should be indicated.

The ultimate goal of the ‘full project’ is to design industrial processes by using advanced control to maximise the value of the unlocked flexibility from the energy-intensive industry and power-to-X technologies (power-to-hydrogen, power-to-heat, power-to-chemicals, power-to-products, etc.) so that value can be generated for the industrial entities and consequently for the eco-system.

It is therefore necessary to proactively investigate the electrical and thermal flexibility in the chemical and steel industry, both in individual processes as within companies and in clusters of plants, across different energy vectors (sector coupling). Appropriate models of processes and utilities (such as steam, cooling, …) for assessing the flexibility potential are needed to allow a techno-economic analysis in the first place. Process models can sometimes be built on the basis of physical principles but are mostly based on many assumptions. However, either due to the unknown or too complex nature of the processes, or for confidentiality reasons, often no process model is available, hindering in depth analysis of electrical and thermal flexibility. Design-for-flexibility is the final goal in which processes are designed or redesigned with the specific goal to be able to unlock and provide flexibility and so take part in the energy transition while remaining cost-efficient.

The scientific goals comprise the design of innovative control, optimization and analysis methods and tools for an early “design for flexibility”-phase in an energy-intensive process industry, while staying tuned to the evolving European energy markets, by assessing the flexibility potential in relation to the future flexibility needs in a high-RES system. Such control allows to create flexibility, assess and unlock the full techno-economic flexibility potential in complicated, integrated multi-energy industrial processes (including power-to-X) for future extremely low greenhouse gas emission scenarios. Additionally, because of the unknown and complex nature of processes, radically new MPC (model predictive control) and deep learning based techniques will be elaborated and applied to energy-intensive industrial processes to reach these goals, as they combine the best of both worlds (model-based approach for robustness and model-free/data-driven approaches to cope with the unknown and complex nature of processes).

The first toll gate is to prove the feasibility of the advanced flexibility control algorithms (combining MPC with deep learning) that take into account the process characteristics, the energy market context and the requirements for integration in the power and heat networks.
In the next toll gates, radically new energy market design, including business models, and power/energy network configurations will be considered, linking with the other task forces of MOT4.

Specify the activities required to reach the first toll gate in more detail. These activities are the aim of the proposed initial SBO as a first step.

The initial SBO will be built around three pillars (energy-intensive processes, energy markets, and energy networks), on top of which the advanced flexibility control framework – which is the core of the project – will be established.

The following highlights the major research questions to be addressed in each of these activities.

Advanced flexibility control based on model-based and data-driven approaches

Many industrial processes deal with a varying quality of input materials (ores, fuels, ...), which significantly influences the energy use of these processes. Lots of historic data are often available (esp. with the ongoing digital transformation of the sector). This data allows characterising the involved uncertainty at the process' input side; however calculating the effect of this variability on the optimal set points for an energy-efficient implementation and on the resulting flexibility of the energy use requires innovative machine learning techniques to model these processes, to forecast their variability and to exploit and control the flexibility from it. This provides ample opportunities towards data-driven modelling, forecasting and control of flexibility.

Model-based control (and specifically model predictive control) has been a cornerstone of automation of energy-intensive processes, because of its responsive power and scalability over a sliding time horizon. Model-based control has demonstrated its ability to be robust. MPC can be optimal as well when the underlying model reflects the actual process which is in our design-for-flexibility particularly challenging. Data-driven control starts from the available data (e.g. reinforcement learning allows to learn a control policy directly from data) while machine learning techniques allow characterisation and forecasting, model building and control.

This project aims to combine the advantages of model-based and data-driven techniques. Such deep learning based approximate MPC opens up the road to automated and expert-free design, tuning and commissioning of model predictive control via a type of plug-and-play approach which is adaptive (and hence maintenance-free). It provides a robust control policy, ideally suited for an industrial environment. Depending on the chosen objective function of the optimisation, decreased CO2 emissions, energy efficiency, optimal costs, more renewables, etc. are all within reach.

The key towards this combination is the encoding of the MPC model structure as a deep learning model that includes control objectives, system models and constraints. Physics-based models can be integrated in deep neural networks by using sequence models (commonly used, e.g., in natural language processing domains). It is based on the hypothesis that it is possible to come up with a one-to-one mapping of an ordinary differential equation in the form of a state space model to the corresponding recurrent neural network. (For instance, the order of a state space model in the ordinary differential equation corresponds with the amount of recurrences in a recurrent neural network.) The result is a scalable and flexible control design with constraints satisfaction properties, and deployment of controllers in the form of neural network policies with low memory and computational footprint with minimum software dependencies, hence operationally reliable and maintenance free.

All of this involves quantifying stochasticity of industrial loads (for the different energy carriers) and (renewable) supplies (for electricity, heat, etc.), and the subsequent development of infrastructure sizing strategies that account for those stochastic variations, tailored in this case to power-to-X solutions. Such characterisation is the core input to the online decision making algorithms (combining model predictive control with machine learning) that offer flexibility to compensate variations and thus cost-effectively exploit existing power grid infrastructure.

The time horizon of these algorithms is quite dynamic, with a focus not only on immediate control for the current or next time-steps but also on a larger time horizon, in order to do an optimal energy-aware planning and flexibility characterization for future time periods. Such energy-aware plans with flexibility created by the industries themselves (hence confidential) can be aggregated bottom-up to enable optimal coordination (e.g., grid and market friendliness) and corresponding flexibility activation or disaggregation decisions.

Such decision making algorithms will inevitably be distributed, as multiple industrial processes and plants should be able to give input to the algorithms without revealing confidential data, bringing in another research challenge.

Incorporating process aspects

The challenge of controlling the flexibility only becomes harder if some redundancy on the supply side allows to easily switch between energy carriers, due to power-to-X technologies (e.g., power-to-steam allows producing steam from green electricity, competing with steam from residual heat, from fossil fuels or from synthetic fuels - all with a different carbon footprint). Many industrial processes are very complex and combine several energy carriers (heat, cold, different gases and chemical fluids, electricity), requiring a coupling over different physical domains, and over different timescales. This complicates the assessment of the opportunities for flexibility and implies barriers for redesigning processes.

Market side

The European energy markets are in evolution, especially concerning flexibility and capacity mechanisms. New services to support the system will emerge and new market mechanisms might emerge that will organize the procurement and delivery of system services by flexibility providers. Next, the changing market setting in a high RES scenario will also fundamentally impact existing energy markets (day-ahead and intraday), having an impact on both the price level and price profile (and hence price volatility). A key question is what the expected economic value of the flexibility in power-to-X and industrial processes will be (assuming flexibility in industrial processes is well characterized), keeping in mind differences in valuing availability and energy provision of flexibility. Elements to take into account to assess the value of flexibility are the evolutions in market places, the evolutions in prices, and the evolutions in other sources of flexibility (next to the flexibility from energy-intensive industries). Fundamental methods and analysis tools need to be developed to assess the value of flexibility and create radically new business models to empower the industries to become the front runners in the energy transition. They should consider suited energy portfolio management (over multiple energy vectors, sector-coupling), market coupling, etc., incorporating the different timescales of the flexibility providers and matching them with the required response times of the market products.

This makes a very clear link to MOT4 TF3 on system integration from a market perspective.

Power and heat network side

The transformation to a low-carbon energy system poses challenges to the grid to guarantee a secure and stable operation. These technical challenges, which are the result of the paradigm shift in the energy production (decrease of conventional generation, increase in RES, and their different geographical spread), are amongst others: frequency control issues, voltage control issues, network congestion management issues, degradation of system adequacy and system resilience. In order to answer to the changing needs of the system in a high RES scenario, new and innovative system services need to be designed, which can be provided by the flexibility sources. The technical specifications of these new services (availability, reliability, ramping speed,…) driven by the technical needs of the grid, will define to what extent flexibility from industry-intensive processes could provide these services.

Additionally, not only the electric power system has to be considered but also other energy grids (the networks for heat at different temperature levels, steam, synthetic fuels, hydrogen, …), requiring a holistic approach and sector-coupling.

Finally, on top of that a high security of supply is required for economic reasons: industrial processes require a very high dependability from the energy supply systems. This complex interaction needs in-depth reliability assessment, taking all types of uncertainties into account.

This makes a very clear link to MOT4 TF3 on system integration from a network/grid perspective.

The disruptive nature of the moonshot project is the ‘design for flexibility’ of the energy-intensive industry thanks to its inherent flexibility and that of power-to-X technologies, ensuring that this design concept takes into account the market and grid evolutions in a high RES scenario. The first toll gate in this moonshot is the ‘controller design for flexibility’. Radically new techniques, integrating model predictive control and deep learning will be at its core. These techniques jointly consider the energy-intensive industrial processes, the energy market context and the power system integration.