Periodic Reporting for period 2 - VeriPhIED (Verified physics-aware machine learning to transform non-linear power system stability and optimization)
Reporting period: 2023-11-01 to 2025-04-30
VeriPhIED’s goal is to develop the necessary tools to enable the trustworthy use of Artificial Intelligence (AI) in electric power systems. Measures against global warming require disruptive changes in the electricity sector. Drastically reducing CO2 emissions involves replacing bulk generation units with millions of renewable energy sources, along with a rapid increase of electricity demand. Maintaining the stability of the system with current approaches becomes not only computationally very difficult, but also extremely costly. AI can help. But AI is still considered a black-box. Who would trust a black-box operating an electric grid, where a possible malfunction (blackout) results not just in economic damage of millions of Euros but puts in danger human lives?
VeriPhIED removes the barriers for the application of Machine Learning in power system problems, developing methods that exploit the underlying physical properties of power systems. We propose the development of physics-informed verifiable neural networks and a neural network training procedure that can supply guarantees of the neural network prediction accuracy. Accuracy no longer needs to be a statistical metric. Instead, our methods can supply a provable upper bound of the prediction error that the power system operators can trust.
Our work moves along two axes. First, develop physics-informed neural networks which demonstrate drastically higher computation speeds (10x-100x faster) while maintaining good accuracy, to assess power system stability and perform power system optimization. Second, design neural network verification procedures that can deliver performance guarantees for the neural networks we develop, so that power system operators can trust them. Although our focus in on power systems, our methods can naturally extend beyond them, finding application to a wide range of physical safety-critical systems, such as self-driving cars, autonomous robots, and chemical processes.
VeriPhIED removes the barriers for the application of Machine Learning in power system problems, developing methods that exploit the underlying physical properties of power systems. We propose the development of physics-informed verifiable neural networks and a neural network training procedure that can supply guarantees of the neural network prediction accuracy. Accuracy no longer needs to be a statistical metric. Instead, our methods can supply a provable upper bound of the prediction error that the power system operators can trust.
Our work moves along two axes. First, develop physics-informed neural networks which demonstrate drastically higher computation speeds (10x-100x faster) while maintaining good accuracy, to assess power system stability and perform power system optimization. Second, design neural network verification procedures that can deliver performance guarantees for the neural networks we develop, so that power system operators can trust them. Although our focus in on power systems, our methods can naturally extend beyond them, finding application to a wide range of physical safety-critical systems, such as self-driving cars, autonomous robots, and chemical processes.
Our work during the first 2.5 years of VeriPhIED has revolved around 2 main topics. First, the development of Physics-Informed Neural Networks (PINNs) for power systems. PINNs are unique in the sense that they not only learn from data, but they rather exploit the underlying physical laws to deliver a Neural Network that is more accurate and less dependent on the quality of the training data.
Our team was the first to introduce physics-informed neural networks in the field of power systems. During the project, we worked with the leading offshore wind developer in the world, the Danish company Ørsted. Ørsted is strongly interested in speeding up their simulations as this can accelerate their wind farm design processes and reduce the time to construction. We showed that PINNs can achieve a 100x speedup of these simulations: to be specific, we achieved a computation time of approx. 30 minutes for both training and evaluation of the PINNs, instead of 2 days that conventional tools would have required.
Trying to address the scalability issues we have been facing with Physics-Informed Neural Networks, we put forward a completely new concept: PINNSim, a neural-network based simulator. PINNSim presents a modular and scalable approach to perform power system dynamic simulations with neural networks, at hopefully 10x-100x higher speeds. Each component is modelled by a separate PINN. The novelty comes from the new algorithm that has to integrate all PINNs in a single framework to simulate a whole system. PINNSim is still in a proof-of-concept phase.
Second, we focused on Neural Network verification. Our work has laid the foundation for delivering the performance guarantees of Neural Network that perform classification (e.g. assessing if critical scenarios lead to blackouts or not) or regression (e.g. determining optimal operating points) in power systems. Trying to address scalability issues when treating larger neural networks we exploited the substantial body of work carried out on neural network verification in the field of computer science and the winning open-source algorithm ab-CROWN, which, however, cannot apply in power system problems. Our work so far has tailored the ab-CROWN algorithm so that it can treat successfully power system optimization problems, and it won the best paper award in the HICSS 2025 conference.
Besides Physics-Informed Neural Networks and Neural Network verification, we have worked with Neural Network interpretability, and also developed methods that use neural networks to accelerate power system optimization. Finally, we have also developed an open-source toolbox, with methods that go beyond the state of the art, to efficiently generate data samples for training neural networks for power system stability assessment.
In terms of knowledge transfer and dissemination, we had several activities over the past 2.5 years, with the one standing one probably being the Tutorial Day on “Trustworthy AI for Power Systems”, which we were invited to organize during the 2024 Power System Computation Conference. Distinguished speakers from MIT, GeorgiaTech, Univ. of Washington, Univ. of Vermont, and DTU offered lectures to 200 attendees. All slides and code can be found here: http://www.chatziva.com/pscc2024.html(opens in new window)
Our team was the first to introduce physics-informed neural networks in the field of power systems. During the project, we worked with the leading offshore wind developer in the world, the Danish company Ørsted. Ørsted is strongly interested in speeding up their simulations as this can accelerate their wind farm design processes and reduce the time to construction. We showed that PINNs can achieve a 100x speedup of these simulations: to be specific, we achieved a computation time of approx. 30 minutes for both training and evaluation of the PINNs, instead of 2 days that conventional tools would have required.
Trying to address the scalability issues we have been facing with Physics-Informed Neural Networks, we put forward a completely new concept: PINNSim, a neural-network based simulator. PINNSim presents a modular and scalable approach to perform power system dynamic simulations with neural networks, at hopefully 10x-100x higher speeds. Each component is modelled by a separate PINN. The novelty comes from the new algorithm that has to integrate all PINNs in a single framework to simulate a whole system. PINNSim is still in a proof-of-concept phase.
Second, we focused on Neural Network verification. Our work has laid the foundation for delivering the performance guarantees of Neural Network that perform classification (e.g. assessing if critical scenarios lead to blackouts or not) or regression (e.g. determining optimal operating points) in power systems. Trying to address scalability issues when treating larger neural networks we exploited the substantial body of work carried out on neural network verification in the field of computer science and the winning open-source algorithm ab-CROWN, which, however, cannot apply in power system problems. Our work so far has tailored the ab-CROWN algorithm so that it can treat successfully power system optimization problems, and it won the best paper award in the HICSS 2025 conference.
Besides Physics-Informed Neural Networks and Neural Network verification, we have worked with Neural Network interpretability, and also developed methods that use neural networks to accelerate power system optimization. Finally, we have also developed an open-source toolbox, with methods that go beyond the state of the art, to efficiently generate data samples for training neural networks for power system stability assessment.
In terms of knowledge transfer and dissemination, we had several activities over the past 2.5 years, with the one standing one probably being the Tutorial Day on “Trustworthy AI for Power Systems”, which we were invited to organize during the 2024 Power System Computation Conference. Distinguished speakers from MIT, GeorgiaTech, Univ. of Washington, Univ. of Vermont, and DTU offered lectures to 200 attendees. All slides and code can be found here: http://www.chatziva.com/pscc2024.html(opens in new window)
During the first 2.5 years of the project we have had over 19 peer-reviewed publications which document progress that advances the state of the art across range of topics including: physics informed neural networks, neural network verification for power systems, sampling beyond statistics, and finding feasible solutions to challenging power system optimization problems through the assistance of neural networks. If we were to select two lines of work that stand out, and where we expect to make significant novel contributions in the second half of VeriPhIED, these are: (i) Physics Informed Neural Networks for Power System Dynamic Simulations and (ii) Verification of Physics-Informed Neural Networks treating systems with differential equations.
PINNSim, the concept we put forward on a neural-network-based simulator, can advance the field of power system dynamic simulations (and possibly of simulators of other non-linear dynamic systems) significantly beyond the state of the art. It is a radically different way to simulate power systems, with the potential to achieve simulation speedups of 10x-100x. Our work in the second half of the VeriPhIED project will focus on producing a working prototype, beyond the proof-of-concept we have shown, with the potential to change how power system simulations will be carried out in the future.
Our focus on Neural Network verification for PINNs can also significantly advance the state of the art, not only in the field of power systems but across different fields. So far, there is no method that can verify Neural Networks which treat systems that involve differential equations. Our goal is to unite the strengths of ab-CROWN as a verifier with new approaches that address the challenges posed by differential equations. This will equip power engineers with a tool to extract performance guarantees for an arbitrary NN representing or controlling a power system component; and enable the deployment of trustworthy AI in power systems. If successful, such a tool can find application across fields beyond power systems.
PINNSim, the concept we put forward on a neural-network-based simulator, can advance the field of power system dynamic simulations (and possibly of simulators of other non-linear dynamic systems) significantly beyond the state of the art. It is a radically different way to simulate power systems, with the potential to achieve simulation speedups of 10x-100x. Our work in the second half of the VeriPhIED project will focus on producing a working prototype, beyond the proof-of-concept we have shown, with the potential to change how power system simulations will be carried out in the future.
Our focus on Neural Network verification for PINNs can also significantly advance the state of the art, not only in the field of power systems but across different fields. So far, there is no method that can verify Neural Networks which treat systems that involve differential equations. Our goal is to unite the strengths of ab-CROWN as a verifier with new approaches that address the challenges posed by differential equations. This will equip power engineers with a tool to extract performance guarantees for an arbitrary NN representing or controlling a power system component; and enable the deployment of trustworthy AI in power systems. If successful, such a tool can find application across fields beyond power systems.