An Overview
Our lab’s core product is an optimization-based toolbox focused on the diagnosis and mitigation of high-impact threats (e.g. blackouts) in power systems, enhancing resilience across all stages of grid disturbance. Its key enabler, the SparseAct algorithm, identifies dominant sources of failure and recommends fast-acting, effective responses at a minimal set of targeted locations. Our work also synthesizes advanced power system modeling with artificial intelligence (AI) to improve the speed and accuracy of threat detection, analysis, and response.
The Exploiting of Sparsity
Sparsity is the special structure in vectors or matrices that only a small fraction of non-zero values exist – carrying essential information; whereas most other entries are zero – ignorable without losing significant accuracy. Our research is organized around exploiting the sparsity structures in system, data, and threats.
SparseAct: Make A Power System Resilient
|
Sparse Diagnosis & Recommendation for A Collapsed System
Given a single scenario of extreme event or contingency, power grid may result in system collapse (i.e., blackout). In this project, we investigate proactive corrective action using sparse optimization find only a few identified locations whose compensation can mitigate failures. Find out the basic idea here. And the sparse idea has been extended to distribution grid (see related works here, here, and here) |
Enforcing sparse diagnosis iteratively, and finally identify 1 key location to fix the collapse of case2383wp
|
Pinpoint Critical Deficiencies to Reduce Instability Risk
(1) A localized compensation can stabilize the entire system frequency!
When disturbance (like generator outage) occurs, droop response automatically adjusts power output to bring system to a new steady state. What might make a system unstable or collapse after droop response? Our work develops a frequency-aware sparse optimization to identify related system vulnerabilities. See our work arXiv:2511.07553
(1) A localized compensation can stabilize the entire system frequency!
When disturbance (like generator outage) occurs, droop response automatically adjusts power output to bring system to a new steady state. What might make a system unstable or collapse after droop response? Our work develops a frequency-aware sparse optimization to identify related system vulnerabilities. See our work arXiv:2511.07553
The new steady state after a generator outage will have a dangerous frequency drop of >0.57Hz!
|
Compensate at node 13, then frequency will stay in the desirable 60±0.3Hz bound.
|
(2) Sparse siting, sizing and type selection of new resources to stabilize frequency and voltage jointly
Left: droop control and local dynamic voltage regulators responded to a sudden generation outage, and the system ends up having 11 under voltage buses. (right) siting and sizing of 7 new dynamic VAR resources have been identified. With these new infrastructures added, the system will no longer have voltage violations following the same contingency!
SparseAct: Co-studying Multiple Scenarios
Infrastructure planning for growing stress: persistent siting, adaptive sizing
We propose a new concept of "persistency" and connect it with the siting and sizing of new resources under growing demand stress. See our work of multi-period sparse optimization: arXiv preprint arXiv:2510.14045
We propose a new concept of "persistency" and connect it with the siting and sizing of new resources under growing demand stress. See our work of multi-period sparse optimization: arXiv preprint arXiv:2510.14045
case2383wp: given a sequence of blackouts induced by load factor (LF) ranging from 1.35 to 1.44, we identify a growing set of vulnerability locations responsible for these failures.
Sparse inspection for cyber-resilient system monitoring
1. Time-series AI: sparse learning for dynamic time-series anomaly detection
Given sensors placed on dynamic graphs (system structure changes due to on/off line switching), how to detect anomalies
from the time-series data? A key observation underlying my work is that, for a time moment t, only a sparse subset of historical data points are most relevant to the present time t, due to data distribution shifts induced by system changes. Our method, DynWatch, exploits the sparse structure in data coherency to find out a sparse set of relevant historical data via sparse temporal weighting. The method achieves real-time processing at the millisecond scale for 60K-node power systems (75% the size of the Eastern Interconnection) per time tick per sensor. Find out more details here.
2. Accurately estimating system states and topology while identifying random bad data and wrong line statuses
In real time when facing bad measurement devices and wrong switch statuses, the resulting anomalous data and topology errors prevents us from knowing the correct system conditions like voltage profile, line outages, etc. In this project, we investigate robust state estimator that can pinpoint these errors and retain accurate voltage and topology information. Find out more details here, here, and here.
3. Combinging state estimation + AI for further robustness against targeted interactive false data
See arXiv preprint arXiv:2510.14043
Given sensors placed on dynamic graphs (system structure changes due to on/off line switching), how to detect anomalies
from the time-series data? A key observation underlying my work is that, for a time moment t, only a sparse subset of historical data points are most relevant to the present time t, due to data distribution shifts induced by system changes. Our method, DynWatch, exploits the sparse structure in data coherency to find out a sparse set of relevant historical data via sparse temporal weighting. The method achieves real-time processing at the millisecond scale for 60K-node power systems (75% the size of the Eastern Interconnection) per time tick per sensor. Find out more details here.
2. Accurately estimating system states and topology while identifying random bad data and wrong line statuses
In real time when facing bad measurement devices and wrong switch statuses, the resulting anomalous data and topology errors prevents us from knowing the correct system conditions like voltage profile, line outages, etc. In this project, we investigate robust state estimator that can pinpoint these errors and retain accurate voltage and topology information. Find out more details here, here, and here.
3. Combinging state estimation + AI for further robustness against targeted interactive false data
See arXiv preprint arXiv:2510.14043
More broadly exploiting sparsity
Sparse Weighting: Adapting Learning Models to Non-IID Data
Power system data distributions shift due to topology changes and operating conditions. We develop sparse temporal reweighting schemes to adapt learning models to these non-IID conditions.
Find an application in time-series anomaly detection here.
Sparse Probabilistic Graphical Models for Power Systems
The power grid’s sparse connectivity can be exploited to build more interpretable and efficient learning models. Our work explores probabilistic graphical models augmented with neural networks to incorporate grid topology and domain knowledge.
Find an application of ML-based simulation warm-starter here.
Power system data distributions shift due to topology changes and operating conditions. We develop sparse temporal reweighting schemes to adapt learning models to these non-IID conditions.
Find an application in time-series anomaly detection here.
Sparse Probabilistic Graphical Models for Power Systems
The power grid’s sparse connectivity can be exploited to build more interpretable and efficient learning models. Our work explores probabilistic graphical models augmented with neural networks to incorporate grid topology and domain knowledge.
Find an application of ML-based simulation warm-starter here.