Summary
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 intractable, but also extremely costly. Recently proposed data-driven methods have been shown to handle the sheer complexity and have an impressive performance, achieving higher accuracy while being 250-1000 times faster than traditional methods. However, power systems are safety-critical systems, where data-driven methods will never be applied if they remain a black-box.
This proposal removes the barriers for the application of data-driven approaches in power system problems, proposing methods that exploit the underlying physical properties of power systems. We propose the development of physics-aware verifiable neural networks and a neural network training procedure that can supply by-design guarantees of the neural network prediction accuracy. Accuracy does no longer need to be a statistical metric. Instead, our methods can supply a provable upper bound of the prediction error over the whole input space, that the power system operators can trust. We further show how neural networks can capture non-linear constraints impossible to capture before, and can reduce any non-linear optimization problem to a tractable mixed-integer linear program with verified accuracy, potentially boosting computation speed and tractability. From a power systems context, this enables us to treat power system dynamics and optimization in a unified framework that accurately captures the true feasible region, removes various approximations, and eliminates redispatching costs, saving billions of euros per year. The proposed methods naturally extend beyond power systems, finding application to a wide range of physical safety-critical systems.
This proposal removes the barriers for the application of data-driven approaches in power system problems, proposing methods that exploit the underlying physical properties of power systems. We propose the development of physics-aware verifiable neural networks and a neural network training procedure that can supply by-design guarantees of the neural network prediction accuracy. Accuracy does no longer need to be a statistical metric. Instead, our methods can supply a provable upper bound of the prediction error over the whole input space, that the power system operators can trust. We further show how neural networks can capture non-linear constraints impossible to capture before, and can reduce any non-linear optimization problem to a tractable mixed-integer linear program with verified accuracy, potentially boosting computation speed and tractability. From a power systems context, this enables us to treat power system dynamics and optimization in a unified framework that accurately captures the true feasible region, removes various approximations, and eliminates redispatching costs, saving billions of euros per year. The proposed methods naturally extend beyond power systems, finding application to a wide range of physical safety-critical systems.
Unfold all
/
Fold all
More information & hyperlinks
| Web resources: | https://cordis.europa.eu/project/id/949899 |
| Start date: | 01-05-2022 |
| End date: | 30-04-2027 |
| Total budget - Public funding: | 1 499 183,00 Euro - 1 499 183,00 Euro |
Cordis data
Original description
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 intractable, but also extremely costly. Recently proposed data-driven methods have been shown to handle the sheer complexity and have an impressive performance, achieving higher accuracy while being 250-1000 times faster than traditional methods. However, power systems are safety-critical systems, where data-driven methods will never be applied if they remain a black-box.This proposal removes the barriers for the application of data-driven approaches in power system problems, proposing methods that exploit the underlying physical properties of power systems. We propose the development of physics-aware verifiable neural networks and a neural network training procedure that can supply by-design guarantees of the neural network prediction accuracy. Accuracy does no longer need to be a statistical metric. Instead, our methods can supply a provable upper bound of the prediction error over the whole input space, that the power system operators can trust. We further show how neural networks can capture non-linear constraints impossible to capture before, and can reduce any non-linear optimization problem to a tractable mixed-integer linear program with verified accuracy, potentially boosting computation speed and tractability. From a power systems context, this enables us to treat power system dynamics and optimization in a unified framework that accurately captures the true feasible region, removes various approximations, and eliminates redispatching costs, saving billions of euros per year. The proposed methods naturally extend beyond power systems, finding application to a wide range of physical safety-critical systems.
Status
SIGNEDCall topic
ERC-2020-STGUpdate Date
27-04-2024
Images
No images available.
Geographical location(s)
