Physics-ML Synergy
In general, the algorithms (situation awareness, optimal control, etc) designed for system operation (and planning) can be implemented by either model-driven or data-driven methods. Each category has pros and cons. Taking the electrical power grid as an example, neither type of methods can fulfill the requirements in speed, accuracy, and operational efficiency when faced with the emerging threat induced by cyberattacks, climate changes, upgrades in energy resources and power electronics.
Model-driven approaches rely on physical models that describe the system. While recognized for their accuracy, they can be slow due to nonlinearities, and vulnerable to modern cyberattacks that leverage system physics. State-of-the-art data-driven models are physics-informed machine learning (ML) techniques which incorporate domain knowledge into ML for specific applications. Although providing fast alternatives, some lack safety guarantees and can still be inaccurate on unseen data when they fail to generalize. Therefore, the first question we ask is "Can we have a Fast AND Accurate tool?"
The second question is about operational efficiency. At any time during system operation, what to follow from the current situation? We always want an automatic and reliable answer. However, this may not be available. For example, at some time, some modern anomaly on the electrical power grid is detected by some advanced ML security product. Anomalous data patterns are observed on several sensors, and we get the alarm. However, from this alarm, we are unable to know exactly what the root cause is, and on which specific device the anomaly has occurred. This often necessitates human intervention in the anomaly mitigation process, sometimes leaving it unknown what to follow from the current situation. Therefore, we ask an important question: "Can we advance operational efficiency by providing information that can be seamlessly integrated into the system operation pipeline to automatic and reliable decision-making?"
The answer is yes! My research is exploring Physics-ML Synergy, a framework that make model-driven and data-driven approaches interconnect and collaborate adaptively to maximize the overall benefits.
Model-driven approaches rely on physical models that describe the system. While recognized for their accuracy, they can be slow due to nonlinearities, and vulnerable to modern cyberattacks that leverage system physics. State-of-the-art data-driven models are physics-informed machine learning (ML) techniques which incorporate domain knowledge into ML for specific applications. Although providing fast alternatives, some lack safety guarantees and can still be inaccurate on unseen data when they fail to generalize. Therefore, the first question we ask is "Can we have a Fast AND Accurate tool?"
The second question is about operational efficiency. At any time during system operation, what to follow from the current situation? We always want an automatic and reliable answer. However, this may not be available. For example, at some time, some modern anomaly on the electrical power grid is detected by some advanced ML security product. Anomalous data patterns are observed on several sensors, and we get the alarm. However, from this alarm, we are unable to know exactly what the root cause is, and on which specific device the anomaly has occurred. This often necessitates human intervention in the anomaly mitigation process, sometimes leaving it unknown what to follow from the current situation. Therefore, we ask an important question: "Can we advance operational efficiency by providing information that can be seamlessly integrated into the system operation pipeline to automatic and reliable decision-making?"
The answer is yes! My research is exploring Physics-ML Synergy, a framework that make model-driven and data-driven approaches interconnect and collaborate adaptively to maximize the overall benefits.
SynSA-Grid: Physics-ML Synergy for Situation Awareness of Power Grids
(High-level idea and preliminary results, Nov 2023)
|
In this work, Physics-ML synergy motivates a collaborative architecture to interconnect model-driven system identification and a time-series ML model to advance the overall situation awareness.
Two state-of-the-art components are brought into the hybrid design.
|
|
What can we achieve with a synergy design? Let's see it from a toy example.
We are faced with a variety of anomalies: traditional random bad data, topology error, and modern cyberattack where attackers modify data to mislead system identification. When using the model-driven system identification model (ckt-GSE) alone, it provides estimates on state variables (bus voltages of power grid) and graph structures, but the solution is very inaccurate whenever the cyberattack happens. Whereas, after 1 interconnection loop in the physics-ML synergy, the augmented system identification gives accurate solutions. When using the ML component (DynWatch anomaly detector) alone, we get no other information about the system's state except some anomaly scores. This lacks operational efficiency. By comparison, the physics-ML synergy enables root cause diagnosis, figuring out the specific type and location of anomalies. The use of ckt-GSE in the synergy also produces the accurate topology after correcting topology errors. Takeaway # 1: Physics-ML synergy enables an all-in-one situation awareness product that brings together system identification, anomaly detection and root cause diagnosis capabilities into one integrated model to advance operational efficiency. |
Given some (steady-state) time-series data collected from a power grid whose graph structure (topology) changes every 60 ticks.
Inaccuracy of ckt-GSE, quantified by root mean squared error (RMSE) of bus voltage values.
Physics-ML Synergy enables accurate state solution (after 1 interconnection loop).
Anomaly scores produced by DynWatch.
Physics-ML synergy enables root cause diagnosis, to figure out which type the anomaly is.
|
How did we make this happen? Here, let's talk about the intuitive understanding without diving into math.
Intuitively, we designed an adaptive interconnection scheme for the two components to exchange information and augment each other. Specifically, the ckt-GSE which is a convex constrained optimization problem runs system identification for every time tick, and this creates a time-series of state solutions. The state series are then given to DynWatch for time-series processing, so that at any time t, a state distribution (q in the figure) is predicted based on previous data. This distribution can be fed back to ckt-GSE, to augment system identification as prior knowledge.
Takeaway # 2: Physics-ML synergy preserves numerical stability as the augmented system identification remains a convex programming problem. This enables the entire framework to be Fast and Scalable.
Importantly, the feeding of prior knowledge to ckt-GSE is guided by the uncertainty quantification which measures the trustworthiness of that prediction. Similarly, the updated domain knowledge produced by system identification are fed to ML upon a rigorous verification. Thus, the interconnection between components is carefully decided on the trustworthiness of information, so that the synergy framework not only enables connection but also decides the right time to interconnect so that each component feed reliable information to other(s).
Takeaway # 2: Physics-ML synergy incorporates ML into the system operation pipeline through adaptive interconnection to reduce the risks of using ML.
Intuitively, we designed an adaptive interconnection scheme for the two components to exchange information and augment each other. Specifically, the ckt-GSE which is a convex constrained optimization problem runs system identification for every time tick, and this creates a time-series of state solutions. The state series are then given to DynWatch for time-series processing, so that at any time t, a state distribution (q in the figure) is predicted based on previous data. This distribution can be fed back to ckt-GSE, to augment system identification as prior knowledge.
Takeaway # 2: Physics-ML synergy preserves numerical stability as the augmented system identification remains a convex programming problem. This enables the entire framework to be Fast and Scalable.
Importantly, the feeding of prior knowledge to ckt-GSE is guided by the uncertainty quantification which measures the trustworthiness of that prediction. Similarly, the updated domain knowledge produced by system identification are fed to ML upon a rigorous verification. Thus, the interconnection between components is carefully decided on the trustworthiness of information, so that the synergy framework not only enables connection but also decides the right time to interconnect so that each component feed reliable information to other(s).
Takeaway # 2: Physics-ML synergy incorporates ML into the system operation pipeline through adaptive interconnection to reduce the risks of using ML.
