Research Interests

Optimal Design of Renewable Energy Management Systems under Uncertainty

Energy production pathways are under a tremendous change, especially after the strict regulations enforced by the ”Green Deal. Energy production systems typically include energy sources, process components, interconnections, interactions within the system, and various energy outputs supplied by the system. Energy sources usually comprise excess electricity from the grid, electricity from photovoltaics and wind, biomass, and heat. Energy products contain electricity, heat, transportation fuel and chemical feed-stocks. The energy management systems for renewable sources include components for storage of electricity, heat, fuels and chemicals, combined heat and power generation (fuel cells, gas turbines, steam boilers) and hydrogen production (electrolyzers, gasifiers). In addition, both energy supplies and demands are time-dependent and uncertain. Thus, the energy management systems form complex processing networks with interacting flows of material, heat, power and water that need to be integrated to achieve the best economic performance.

In our Tubitak 2232 research project, we apply the principles of decision making under uncertainty to tackle the aforementioned challenges. Combination of the proposed approaches with divide and conquer methods, Benders decompositions, and decentralization and parallelization structures are promising in terms of dealing with computational complexity for both design and control levels. Furthermore, the study of integration of design and control, which still remains scarce in the community, is also an encouraging subject for energy systems in terms of flexibility and economic optimality.

Artificial Intelligence-based Methods Towards Sustainable Chemical and Energy Processes

Physics-informed Machine Learning for Simultaneous Modeling and Optimization

Constructing first-principles models is usually a challenging and time-consuming task due to the complexity of the real-life processes. On the other hand, data-driven modeling, and in particular neural network models often suffer from issues such as overfitting and lack of useful and high-quality data. At the same time, embedding trained machine learning models directly into the optimization problems has become an effective and state-of-the-art approach for surrogate optimization, whose performance can be improved by physics-aware training. It is planned to investigate the impact of such physics-aware methods for processes in chemical engineering systems such as oil and gas, wood-working, paper and pulp and renewable energy integrated grids. Furthermore, physics-informed trained machine learning models are expected to offer many reliable offline and online optimization chances when directly embedded into nonlinear programming frameworks. It is planned to integrate such physics-aware frameworks with piece-wise linear relaxations with current real-time optimization (RTO) and control methods especially in chemical and energy engineering to reduce the sequential effort of first-model-than-optimize to model-and-optimize simultaneously [4]. Furthermore, initial analysis demonstrate that the convex nature of the trained physics-informed machine learning-based optimization methods bring about global solutions for the investigated problems [5,6].

Many model-based control approaches in the literature lack the systematic integration of the effect of uncertainty. Uncertainty might come into play as structural, parametric and demand and supply unpredictability. As a result, the design of corresponding model-based controllers usually call for efficient and scalable stochastic and/or robust optimization algorithms. Fast NMPC, which enables higher optimization frequency but requires much faster on-line computations, might result in much better closed-loop performances, especially for stochastic approaches. Such physics-informed trained machine learning based controllers are expected to be useful in terms of matching the specific needs of optimization and control technologies with scalable implementations.

Finally, such methods are expected to provide better performance when combined with reinforcement-learning based agent training methods which will be an effective method for the management of energy systems under uncertainty in general. Inclusion of uncertainty into the existing decision making (optimization) platforms, and testing for industrial case studies is an ultimately  scarce subject. Theoretical studies suggest that cost improvement up to 20% is possible through stochastic optimization as opposed to deterministic approaches. Industrial comparisons and implementations considering LCA are therefore promising under uncertainty. Current/new case studies are to be extended utilizing general/specific uncertainty metrics both related to minimum cost and desired life-cycle metrics. Than, these finite and potentially numerous case studies will be taken into account simultaneously in the optimization steps for stochastic optimization and control in a systematic fashion. The main aim is to demonstrate and deal with the effects of inevitable uncertainty on LCA studies and to come up with single, unique solutions to respond to all possible uncertainty effects such as green raw material supply shortages or surplus, disruptions or environmental regularizations in different industries such as bio-chemical, renewable energy, pulp, wood and paper etc.

Large-scale Process Modeling, Optimization and Control through AI

Artificial Intelligence is the way to the future. Subtopics of AI intersecting with PSE include modeling, optimization and control, exhibiting certain possibilities to reduce production cost and decrease expenditures. Robust parameter estimation techniques would increase the validity of process models whereas the improvements in decomposition and divide and conquer methods, and complexity reduction algorithms reduce the real-time effort for optimization and control. Real-time application in the face of possible collaborations with industry is vital to answer the open questions in this field.

Many model-based control approaches in the literature lack the systematic integration of the effect of uncertainty. Uncertainty might come into play as structural, parametric and demand and supply unpredictability. As a result, the design of corresponding model based controllers usually call for efficient and scalable stochastic and/or robust optimization algorithms. Fast NMPC, which enables higher optimization frequency but requires much faster on-line computations, might result in much better closed-loop performances, especially for stochastic approaches. Combination of PMP and parsimonious input parameterizations might be efficient for stochastic, mixed-integer NMPC, and moving horizon state estimation. Furthermore, hierarchical, decentralized and distributed structures should be incorporated into stochastic and robust control formulations so as to cope with computational complexity. In addition, real-time implementations should be performed for displaying the advantages of the proposed methods.

We are planning to employ projects with the industry by using process systems engineering approaches such as: combination of first-principles modeling, machine learning, optimum artificial neural network design and surrogate modeling, model order reduction, process optimization and hierarchical and distributed optimizing control through classical MPC and economic MPC.

References

[1] S.J. Qin, T.A. Badgwell, A survey of industrial model predictive control technology, Control Engineering Practice, 11 (2003) 733-764.

[2] Tatar, S. M., Akulker, H., Sildir, H., & Aydin, E. (2022). Optimal design and operation of integrated microgrids under intermittent renewable energy sources coupled with green hydrogen and demand scenarios. International Journal of Hydrogen Energy, 45(65), 27848.

[3] Akulker, H., Aydin, E. (2023). Optimal design and operation of a multi-energy microgrid using mixed-integer nonlinear programming: Impact of carbon cap and trade system and taxing on equipment selections, Applied Energy, 330, 120313.

[4] Sildir, H., & Aydin, E. (2022). A mixed-integer linear programming based training and feature selection method for artificial neural networks using piece-wise linear approximations. Chemical Engineering Science249, 117273.

[5] Koksal, E.S., Aydin, E. Physics Informed Piecewise Linear Neural Networks for Process Optimization. Computers & Chemical Engineering (under review)

[6] Asrav, T., Aydin, E. Physics-Informed Recurrent Neural Networks combined with Hyper-parameter Optimization for Dynamic Process Systems. Computers & Chemical Engineering (under review)