Machine Learning for

Power Systems: From

Pure Data-Driven to
Physics-Informed Methods
Prof. Nanpeng Yu
Department of Electrical and Computer Engineering
Department of Computer Science
(cooperating faculty)
Director of Energy, Economics, and Environment
Research Center
[email protected]
951.827.3688
Team Members
Energy, Economics, and Environment (E3) Research Center
Director: Dr. Nanpeng Yu.
Over 40 affiliated faculty in CSE, ECE, STAT and SPP
Project Scientists and Postdoctoral Researchers
Dr. Koji Yamashita, Dr. Mikhail Bragin, Dr. Yinglun Li, and Dr. Zuzhao Ye
PhD and MS Students
Yuanbin Cheng, Osten Anderson, Jingtao Qin, Joseph Brown, Maojie Tang, Zhixuan Tan,
Shaorong Zhang, and Jeffrey Chan
Alumni PhD Students
Dr. Brandon Foggo Dr. Wei Wang Dr. Yuanqi Gao Dr. Jie Shi
Machine Learning Senior Data Scientist Machine Learning Postdoctoral
Scientist at Habitat at Walmart Labs Engineer at Lucid Researcher at
Energy Motors Cornell

Dr. Wenyu Wang Dr. Farzana Kabir Dr. Yinglun Li Dr. Zuzhao Ye
Senior Engineer at EV Grid Infrastructure Senior ML Scientist Founder of
Quanta Technology Specialist at California at Alibaba AmpTrans
Energy Commission International Digital
Commercial Group 2
Research Projects and Sponsors
Over $16 Million of R&D Funding in the past 10 years
Federal government, state agencies, private companies
Research Tracks
Physics-informed Machine Learning for Power Systems
Scalable Optimization in Critical Infrastructure Systems
Transportation and Building Electrification
Decarbonization Planning
Energy Efficient Data Center

3
Outline
Volume, Variety, Velocity & Value of Big Data in Power Systems
Applications of Machine Learning in Power Systems
Transmission system, distribution system, end-use customers
Motivation for Physics-informed Methods in Power Systems
Leverage Unique Properties of Data from Power Systems
Low rank and sparsity, high and low entropy

Integrate Machine Learning Models with Physical Models of

Power Systems
Renewable resource model, power flow model
Topology of power network, system control model
Power system dynamic model

Summary and Discussions 4

Volume and Variety
Variety of Big Data in Power Systems
Supervisory control and data acquisition (SCADA) system
Phasor Measurement Unit (PMU), Micro-PMU
Digital Fault Recorder, Equipment Monitors
Census Data, Advanced Metering Infrastructure (AMI)
RF
Neighborhood
Area Mesh Wide-Area Network

Weather Station, Electricity Market Network

Wireless
Network

Meter Data
Geographical Information System Cell Relay Management
System

Volume of Big Data in Power Systems

PMUs
Over 2,500 PMU (10,000 measurement channels) in the U.S.

Roughly 450 TB of data generated annually (60Hz)

AMI
Over 1.1 billion smart meter installations worldwide as of 2022.
5
Over 2 petabytes of smart meter data.
Velocity and Value
Velocity of Data in Power Systems
Sampling Frequency
AMI Data: Sampling frequency increases from once a month to 1 reading every 15 mins
PMU: Mature technology (30 – 240 Hz), continuous point on wave (1920 – 61,400 Hz)

Bottleneck in Communication Systems of Distribution Network

Limited bandwidth for ZigBee network
Most of the utilities in the US receives smart meter data with ~24 hour delay

Edge Computing Trend

Itron and Landis+Gyr extend edge computing capability of smart meters
Centralized → distributed / decentralized monitoring, computing, control and decision making

Value of Big Data Analytics in Power Systems

According to GTM Research, electric utilities around the world spent over $3.8
billion on data analytics solutions in 2020.
Analysis from Indigo Advisory group indicates that the market for AI in the energy
sector could be worth $13 billion as of the end of 2023. 6
 Applications of Machine Learning in
Transmission Systems and Electricity Market
Electricity Market Applications
Price & Load Forecasting, Algorithmic Trading
Learn Power System Dynamics
Event Detection & Classification

Generator Tripping

Accelerate UC & OPF

Model Validation &

Equipment Monitoring
Identify Faulty Equipment in Parameter Estimation
Substations

State Estimation 7
Linear State Estimation
 Applications of Machine Learning for
Power Distribution Systems and End Use Customers
Spatio-temporal Forecasting
Electric Load / DERs – Short-Term / Long-Term
Anomaly Detection
Electricity Theft, Unauthorized System Monitoring
Solar Interconnection State Estimation & Visualization

Distribution System Controls

Deep Reinforcement Learning

Network Topology and

Parameter Identification
Transformer-to-customer, Phase
connectivity, Impedance estimation
Equipment Monitoring Customer Behavior Analysis
Predictive Maintenance Customer segmentation, nonintrusive 8
Online Diagnosis load monitoring, demand response
Physics-Informed ML for Power Systems: Motivation
Purely Data-driven ML algorithms: widely adopted by researchers and
practitioners to solve a myriad of problems in power systems.
Big data: e.g. load & price forecasting, predictive maintenance of transformers
Struggle to deal with system monitoring, sequential decision-making, large-scale
optimization, and control problems in power systems.
Technical Challenges
Accuracy, generalization capability, sample efficiency, safety and interpretability
Physics-informed ML Algorithms
Synergistic combination of machine learning & physical model or information
Embed domain knowledge, unique data property, system model in ML algorithms
Introduce inductive bias, improve explainability, and generalize to unforeseen
scenarios from a finite training dataset
9
 Low Rank and Sparsity Properties of
Power System Measurement Data

Rank of the matrix: the number of linearly independent rows or columns in the matrix
Low Rank and Sparsity: Voltage Event Detection Using
Optimization with Structured Sparsity Inducing Norms
Key Observations
Voltage related events trigged by system
faults are often regional events
The 𝑋 − 𝐿 during voltage event periods
have row-sparse structure
Rows of residual matrix correspond to
PMUs highly impacted by the event

Main Idea
Decompose the streaming PMU data matrix
𝑋 into
A low-rank matrix 𝐿, a row-sparse event-
pattern matrix 𝑆, and a noise matrix 𝐺
Extract anomaly features from 𝐿 & 𝑆
Use clustering algorithm to identify power X. Kong, B, Foggo, and N. Yu, “Online Voltage Event
system voltage events Detection Using Optimization with Structured Sparsity-
Inducing Norms,“ IEEE Transactions on Power
Systems, vol. 37, no. 5, Sep. 2022. 11
 Decompose PMU Data Matrix with Proximal Bilateral
Random Projection (PBRP) to Detect Events

Solution Approach: Coordinate Descent

Residual PMU data matrices during voltage events have distinctive sparsity structure
Computationally efficient PBRP algorithm is proposed to decompose PMU data matrices
Online voltage event detection algo. shows better accuracy & scalability on PMU data (Eastern
Interconnection) 12
 Low and High Entropy Power System
Measurements

Information entropy of data: the average amount of information conveyed by an

event, when considering all possible outcomes
𝐻 𝑥 ≐ −෍ 𝑝 𝑥 log 𝑝(𝑥)
𝑥∈𝒳
 High & Low Entropy of Dataset:
Information Loading-based Regularization
Background
Abstract Representation of Deep Neural Network based Classifier

Main Idea
Control the amount of information compression between the input layer and the last hidden
layer of a deep neural network
Balance memorization and generalization
Low entropy High entropy
input feature space input feature space
e.g. Vol. mag. in distribution e.g. PMU data during power
system (4–11 bits) system events ( > 60 bits)

Algorithm
Augment the typical cross-entropy loss function with estimated mutual information between
the input layer and the hidden representation

B. Foggo, N. Yu, J. Shi and Y. Gao, "Information Losses in Neural Classifiers from Sampling," IEEE Transactions on
Neural Networks and Learning Systems, vol. 31, no. 10, pp. 4073-4083, 2020. DOI: 10.1109/TNNLS.2019.2952029. 14
Phase Connectivity Identification
Very few electric utility companies have completely accurate phase connectivity
information in GIS!
Validated using real-world distribution circuits data from SCE and PG&E.

Learned Representations
on Circuit V

Phase Identification Accuracy

B. Foggo and N. Yu, "Improving Supervised Phase Identification Through the Theory of Information
Losses," IEEE Transactions on Smart Grid, vol. 11, no. 3, pp. 2337-2346, 2020. 15
System Event Classification with PMU Data
Formulated as a classification problem
Normal operation condition, line event, generator event, oscillation event

Learned Representation

F1 Score on Testing Dataset from Eastern Interconnection

J. Shi, B. Foggo, and N. Yu, "Power System Event Identification based on Deep Neural Network with Information Loading," 16
IEEE Transactions on Power Systems, vol. 36, no. 6, pp. 5622-5632, Nov. 2021.
 Renewable Energy Resource Model

Solar PV System Performance Model: used to understand and predict energy or power
output from PV systems under a wide range of environmental, design, and site.
 Physical Solar PV Performance Model
Estimation of Behind-the-Meter Solar Generation
Net Metering: 𝑛𝑒𝑡 𝑙𝑜𝑎𝑑 = 𝑙𝑜𝑎𝑑 − 𝑠𝑜𝑙𝑎𝑟 𝑔𝑒𝑛𝑒𝑟𝑎𝑡𝑖𝑜𝑛

Mixture Hidden Markov Model for Load Modeling

𝐿𝑛𝑡 =𝑎𝑧𝑛𝑡 + 𝑏𝑛 + 𝑐𝑛,𝑧𝑛𝑡 𝑥𝑡 + 𝜖𝑛𝑡 , 𝑛 = 1 … 𝑁,
𝑡 = 1 … 𝑇, 𝑏𝑛 ~𝑁(0, 𝜎 2 ), 𝜖𝑛𝑡 ~𝑁(0, 𝜆2𝑧𝑛𝑡 )

𝑇 𝑔𝑡 𝜽𝑆 : solar PV generation at time 𝑡 based on the physical solar PV

2
argmin ෍ 𝑆𝑡 − 𝑔𝑡 𝜽𝑆 system performance model with the technical parameters 𝜽𝑆 .
𝜽𝑆
𝑡=1
𝑃𝑎𝑐 = 𝑔 𝜽𝑆 = 𝜂 𝜂𝑛𝑜𝑚 , 𝑃𝑑𝑐 𝑃𝑑𝑐
subject 𝜽𝑆,𝑚𝑖𝑛 ≤ 𝜽𝑆 ≤ 𝜽𝑆,𝑚𝑎𝑥
𝐸𝑡𝑟 𝜃𝑡 , 𝜃𝑎𝑧
𝑃𝑑𝑐 = 𝑔′ 𝑃𝑑𝑐0 , 𝜃𝑡 , 𝜃𝑎𝑧 , 𝑙 = (1 − 𝑙) × 𝑃𝑑𝑐0 1 + 𝛾 𝑇𝑐 𝜃𝑡 , 𝜃𝑎𝑧 − 𝑇0
𝐸0
18
Testing Results

Validated using real-world smart meter and solar PV

generation data from Austin, Texas.
The MHMM follows the actual load much more closely than
the benchmark algorithms during the low load periods.
The synergistic combination of physical solar PV system
performance model and the statistical community load model
yields more accurate solar PV generation estimation.
The MHMM allows sharing of info. across individual
customers, which leads to more accurate load and solar PV
generation estimates.
F. Kabir, N. Yu, W. Yao, R. Yang, and Y. Zhang, "Joint Estimation of Behind-the-Meter Solar Generation in a
Community," IEEE Transactions on Sustainable Energy, vol. 12, no. 1, pp. 682-694, Jan. 2021. 19
Steady State Model of Power Systems

Power Flow Model: Linearized or Nonlinear Model of Power Flow

 Power Flow Model:
Physics-informed Electricity Theft Detection
Key Idea
There exists approximate linear models between
power and voltage magnitudes to distribution
secondaries.
A large discrepancy between estimated & measured
power consumption data indicates potential theft.
𝑛𝑐 𝝌
𝜷𝒊
𝑦𝑖 𝑡 = 𝒙(𝒕)𝑻 ෍ 𝑦𝑗 (𝑡) 𝑦 + 𝜖𝑖′ (𝑡)
𝑗=1 𝛽𝑖

Where 𝒙 𝒕 = [1, 𝑣1 𝑡 , 𝑣2 𝑡 , ⋯ , 𝑣𝑛𝑐 𝑡 ]𝑇 , 𝑦𝑖 𝑡 = 𝑝𝑖 (𝑡)

(𝑒)
Let 𝑦෤𝑖 and 𝑦෤𝑖 denote the out-of-sample residual time series for the energy thief
(𝑒) 𝑦
Lemma 1. 𝑦෤𝑖 − 𝑦෤𝑖 = − ෍ 𝛽𝑗 𝑦𝑖𝑠 Lemma 1 & Lemma 3 combine to show that a thief’s residuals
𝑗≠𝑖 will become negative once he or she begins to steal power.
Lemma 2. ෍ 𝑦෤𝑗
(𝑒)
= ෍ 𝑦෤𝑗 = 0 Lemma 2 shows that the residuals of the other customers will
𝑗 𝑗
raise to balance their sum.

Lemma 3. ∀𝛿 > 0, there exists a training data window length 𝑇 > 0 such that for each 𝑗: ℙ 𝛽𝑗𝑦 ≥ −𝛿 > 1 − 𝛿

Yuanqi Gao, Brandon Foggo, and Nanpeng Yu, "A Physically Inspired Data-driven Model for Electricity Theft
Detection with Smart Meter Data," IEEE Transactions on Industrial Informatics, vol. 15, no. 9, 2019. 21
Testing Results with Real-world Smart Meter Data
12 KV circuit from Southern California Edison, 6
months of smart meter data from 980 customers
and 190 transformers.
The average electric load consumed by the
customer is 1.6 kWh.
The mean of the estimation residual is -0.01 kWh
and its standard deviation is 0.1 kWh.

Anomaly score for customer 𝑘 is much higher than

that of all other customers in the feeder

Anomaly score increases monotonically with the

amount of stolen electricity.
In all cases, customer 𝑘’s anomaly score will
surpass the 95th percentile of all customers if it
steals more than 32 kWh in two weeks or 0.19
kWh per hour.
In cases 1-3, if customer 𝑘 steals more than 0.38
kWh of power per hour, then its anomaly score will
the largest of all customers. 22
 Physics-informed Phase Connectivity Identification in
Power Distribution System
The time difference version of the physical model of distribution network
෥ 𝑡 = 𝑋෥
𝒗 ෠ 𝑇𝒑
𝒗𝑟𝑒𝑓 𝑡 + 𝑋𝐾𝑋 ෥ 𝑡 + 𝑋 𝐿෠ 𝑋 𝑇 𝒒
෥ 𝑡 + 𝒏(𝑡)

𝒗 𝑡 }𝑇𝑡=1 , given 𝒙, {෥
The likelihood of observing {෥ 𝒑 𝑡 }𝑇𝑡=1 and {෥
𝒒 𝑡 }𝑇𝑡=1 is
𝑇
𝑃𝑟𝑜𝑏 ෥ 𝑡
𝒗 𝑡=1 𝒑 𝑡 }𝑇𝑡=1 , {෥
{෥ 𝒒 𝑡 }𝑇𝑡=1 ; 𝒙)
𝑇
−
Σ𝑁 2 1 𝑇 −1
= 𝑀𝑇 × exp {− 2 ෍ [෥ ෥(𝑡, 𝒙)]𝑇 Σ𝑁
𝒗 𝑡 −𝒗 [෥ ෥(𝑡, 𝒙)]}
𝒗 𝑡 −𝒗
𝑡=1
(2𝜋) 2

Lemma 1. Let 𝒙∗ be the correct phase connection. If the following two conditions are satisfied, then
as 𝑇 → ∞, 𝒙∗ is a global optimizer of 𝑓 𝒙 .
෥𝑟𝑒𝑓 𝑡𝑙 , 𝒑
1. 𝒏 𝑡𝑘 is i.i.d. and independent of 𝒗 ෥ 𝑡𝑙 , for ∀𝑡𝑘 , 𝑡𝑙 ∈ 𝑍 + .
෥ 𝑡𝑙 , and 𝒒
෥𝑟𝑒𝑓 𝑡𝑘 , 𝒑
2. 𝒗 ෥ 𝑡𝑘 , and 𝒒 ෥𝑟𝑒𝑓 𝑡𝑙 , 𝒑
෥ 𝑡𝑘 are independent of 𝒗 ෥ 𝑡𝑙 , for ∀𝑡𝑘 , 𝑡𝑙 ∈ 𝑍 + , 𝑡𝑘 ≠ 𝑡𝑙 .
෥ 𝑡𝑙 , and 𝒒
Directly minimizing 𝑓(𝒙) is difficult.
Key Idea: phase identification problem → maximum marginal likelihood estimation (MMLE)
problem.
W. Wang and N. Yu, "Maximum Marginal Likelihood Estimation of Phase Connections in Power Distribution 23
Systems," IEEE Transactions on Power Systems, vol. 35, no. 5, pp.3906-3917, Sep. 2020.
Numerical Results with Smart Meter Data
Number of Loads per Phase in the IEEE Test Circuits
Feeder A B C AB BC CA ABC Total
37-bus 5 5 6 3 2 2 2 25
123-bus 18 17 17 9 9 10 5 85
342-bus 30 38 31 35 31 33 10 208

Smart Meter Data Phase Identification Accuracy of Different Methods with 90 days of Meter Data

Length: 90 days of hourly average

Meter 37-Bus 123-Bus 342-Bus
real power consumption data Method
Class Feeder Feeder Feeder
Provided by FortisBC
Correlation- 0.1% 100% 98.75% 81.82%
Measurement noise ~ 0.1 and based Approach 0.2% 100% 97.5% 81.31%
0.2 accuracy class smart
Clustering- 0.1% 100% 100% 93.43%
meters established in ANSI.
based Approach 0.2% 100% 98.75% 91.41%
Voltage measurements
MMLE-based 0.1% 100% 100% 100%
rounded to the nearest 1 V for
Algorithm 0.2% 100% 100% 100%
the primary loads & 0.1 V for
the secondary loads.
The MMLE-based algorithm outperforms the correlation and clustering-based approaches.
24
The improvement in accuracy increases as the complexity of the distribution feeder increases.
 Graphical Model of Power Network

Graph Model for Power Systems: Nodes → Vertices, Power Lines → Branches
Graph Model: Physics-informed Graphical Learning for
Distribution Line Parameter Estimation
Key Idea: Embed physical
equations of power flow in the
graphical learning model
Inspired by graphical neural
network (GNN)
Difference between physics-
informed GL and GNN
Leverage 3Φ power flow-based
physical transition fcn. to replace
the deep neural networks in GNN.
Key Step: Derive the gradient of
voltage magnitude loss function
w.r.t. line segment’s resistance
and reactance parameters with
an iterative method.
Estimate distribution line parameters with SGD considering prior estimates of line parameters
and physical constraints.
Improve computation efficiency with grid partition scheme and fast forward/backward function.
Wenyu Wang and Nanpeng Yu, "Estimate Three-phase Distribution Line Parameters with Physics-informed 26
Graphical Learning Method," IEEE Transactions on Power Systems, vol. 37, no. 5, pp. 3577-3591, Sep. 2022,
Fast GL and Numerical Study Results
MADR Improvement of Parameter Estimation Methods
in the Test Feeder (Avg. / Choose Optimal Value)

15 days of smart meter data from 13.2 kV

distribution circuit in National Grid
177 line sections, 491 loads, 23 solar PV systems Avg. Runtime (Seconds) of Main Functions of Parameter
Estimation Methods in Sub-networks of different sizes
Feeder partitioned into 10-subnetworks (parallel
estimation)
Fast Graphical Learning has significantly higher
MADR improvement (30% improvement in MADR)
Approximately 10 times reduction in computation
time
Wenyu Wang, Nanpeng Yu, and Yue Zhao, "Fast Graphical Learning Method for Parameter 27
Estimation in Large-Scale Distribution Networks," IEEE SmartGridComm, 2022.
Solve Large-Scale MIP Problem with Generic GL
Solving unit commitment (UC) or security constrained UC (SCUC) problems is crucial to market operations
The UC or SCUC problem boils down to solving large-scale mixed integer programing (MIP) problems.
State-of-the-art commercial solvers (e.g. Gurobi, CPLEX) use sophisticated branch and bound algorithms

Used to solve large-scale MIP problems (e.g. UC)

Drawbacks
Scalability: the number of constraints in the bipartite
graph increases exponentially with the number of
nodes in a power system
Nair, Vinod, et al. "Solving mixed integer programs using
neural networks." arXiv preprint arXiv:2012.13349 (2020). 28
Google Deep Mind
 Accelerate UC Solution with Physics-informed GL
Learn to branch and dive for optimization problems on large-scale networks
Instead of adopting bi-partite graph, leveraging power network to perform physics-informed GL

Outperforms commercial
MIP solver (e.g. Gurobi
and CPLEX) on large-
scale UCR problems
(6515-bus)
Better scalability,
optimality and
interpretability
29
J. Qin & N. Yu, "Solve Large-scale Unit Commitment Problems by Physics-informed Graph Learning" arXiv preprint arXiv:2311.15216
 System Control Model

System Control Model: Model-based Controller or Heuristic Control Policy from

Human Operator
Control Model: Batch Constrained Reinforcement Learning-
based Distribution System Control Overall Framework
Motivation
Costly & time consuming to learn optimal control
strategy by directly interacting with physical network.
Learning from finite historical dataset lead to large
extrapolation errors.
Solution: Batch-constrained soft actor-critic
algorithm
Key Idea: train a control policy
Maximize the total discounted return
Minimize dissimilarity between learned control policy &
behavior policy of the batch data

Weekly Operational Cost of test feeder

Y. Gao, W. Wang, J. Shi, and N. Yu, "Batch-constrained Reinforcement Learning for Dynamic Distribution Network 31
Reconfiguration," IEEE Transactions on Smart Grid, vol. 11, no. 6, pp. 5357-5369, 2020.
 Dynamic Model of Power Systems

Dynamic Model of Power System: Differential and Algebraic Equations (DAEs),

Hamiltonian and Nearly Hamiltonian System
Dynamic Model: Generator Dynamic Parameter Estimation
with Physics-based Neural Ordinary Differential Equations
Key Ideas
Convert the forward solver of ODEs for power system dynamics into physics-informed neural networks.
Calculate the loss function based on the difference between dynamic simulation results from the neural
networks and pseudo PMU measurements.

Calculate the gradient of

the loss function w.r.t.
dynamic parameters based
on the neural ODE
technique and the adjoint
method.
Update the dynamic
parameters iteratively with
a quasi-Newton Method.

X. Kong, K. Yamashita, B. Foggo, and N. Yu, "Dynamic Parameter Estimation with Physics-based Neural Ordinary 33
Differential Equations" IEEE PES GM, 2022.
Numerical Study Results Generate noisy PMU measurements from dynamic
WECC 3-machine 9-bus system simulation data.
A single transmission line is disconnected at 5s. The
simulation time is 10s.
Two disturbance scenarios (between nodes 5 and 7,
between nodes 8 and 9)
When the PMU data length is 3s, our proposed algorithm
achieves the lowest relative estimation error.
The physics-based neural ODE algorithm outperforms the
baseline algorithm in terms of estimation accuracy for
most of the unknown parameters

Physics-based neural ODE algorithm has much

shorter computation time than the baseline algorithm.

When the data length is 3s, the running time of our model
is just 4.82 seconds, which is nearly 8 times faster than
the baseline model.
Mini-batch scheme of neural network training shortens
34
model running time.
 Learning Dynamic System with Hamiltonian Neural Networks
Hamiltonian mechanics: can predict the motion of a energy-conserved system.
𝑇 𝑇
State variables: generalized position 𝐪 = 𝑞1 , 𝑞2 , ⋯ , 𝑞𝑛 momentum 𝐩 = 𝑝1 , 𝑝2 , ⋯ , 𝑝𝑛
𝐪 and 𝐩 correspond to voltage angle 𝛿 and angular speed 𝜔 in power system dynamic model

Model single machine infinite bus system as a Hamiltonian: 𝑚1 𝛿ሷ + 𝑑1 𝛿ሶ + 𝐵12 𝑉1 𝑉2 sin(𝛿) − 𝑃1 = 0

Can we learn the Hamiltonian or Lagrangian function (energy conservation law) directly?

𝑑𝒒 𝑑𝒑
𝐻(𝒒, 𝒑) ,
𝑑𝑡 𝑑𝑡
Legendre transform
Learn Hamiltonian Learn ODEs
The Hamiltonian can be regarded
as the energy function

[1] Greydanus S, Dzamba M, Yosinski J. Hamiltonian neural networks. Advances in neural information processing systems, 2019.
[2] Cranmer M, Greydanus S, Hoyer S, et al. Lagrangian neural networks. arXiv preprint arXiv:2003.04630, 2020.
35
 Learning Power System Dynamics: Nearly Hamiltonian NN
Can we formulate the SMIB system as a Hamiltonian function?
𝑚1 𝛿ሷ + 𝑑1 𝛿ሶ + 𝐵12 𝑉1 𝑉2 sin(𝛿) − 𝑃1 = 0
Unfortunately, the answer is No. If the damping coefficient 𝑑1 is positive, then the SMIB system
is a dissipative system instead of a conservative system.
Hamiltonian System Nearly Hamiltonian System
Hamiltonian differential equations Hamiltonian differential equations

Total energy is conserved: Total energy decreases:

Shaorong Zhang and

Nanpeng Yu, "Learning
Power System
Dynamics with Nearly-
Hamiltonian Neural
Networks," IEEE PES
General Meeting, 2023.

36
Summary
Key: synergistic combine ML algorithms with physical domain knowledge

Drastic improvement in accuracy, generalization capability, sample

efficiency, safety and interpretability

Types of power system domain knowledge or physical model

Unique data properties: low rank, high/low entropy

Steady state models: renewable resource model, topology, power flow model

Dynamic models: dynamic simulation and control models, Hamiltonian function

How to combine physical model with machine learning model?

Iterative fitting the physical and data-driven models
Embed physics model in machine learning algorithms
Data preprocessing, loss function, neural architecture
Discussions
Importance of Collaborating with Electric Utilities
Learn about current practice and realistic challenges
Work on solutions and datasets that could address real-world problems
Smart meter data for phase connectivity identification
PMU data from all three U.S. interconnections for system monitoring

Real-world data is the best source for validating your ML algorithms

Simulation data does not possess the same property as real-world data
Algorithm trained with simulation data will struggle to solve real-world problems
Reproducibility: Dataset and Open-Source Software
Learning Power System Dynamics with Neural ODE and NHNN
https://github.com/szhan311/Neural_ODE_Power_System_Dynamic, https://github.com/szhan311/Nearly-
Hamiltonian-Neural-Network

RL-based Control in Distribution System

A RL-based VVC Dataset and Testing Environment. https://github.com/yg-smile/RL VVC dataset

Repository of synthetic PMU data generated from U.S. power grid data
pmuBAGE: https://github.com/NanpengYu/pmuBAGE
Recent Publications
Theoretical Machine Learning
B. Foggo, N. Yu, J. Shi and Y. Gao, "Information Losses in Neural Classifiers from Sampling," IEEE Transactions on Neural Networks
and Learning Systems, vol. 31, no. 10, pp. 4073-4083, 2020.
Brandon Foggo and Nanpeng Yu, "Analyzing Data Selection Techniques with Tools from the Theory of Information Losses," IEEE
International Conference on Big Data, pp. 1-10, 2021.
Brandon Foggo and Nanpeng Yu, "On the Maximum Mutual Information Capacity of Neural Architectures," ICML Workshop on Neural
Compression: From Information Theory to Applications, 2023.

Machine Learning for Power Transmission Systems

Shaorong Zhang and Nanpeng Yu, "Learning Power System Dynamics with Nearly-Hamiltonian Neural Networks," IEEE PES GM, 2023.
X. Kong, B, Foggo, and N. Yu, “Online Voltage Event Detection Using Optimization with Structured Sparsity-Inducing Norms,“ IEEE
Transactions on Power Systems, vol. 37, no. 5, Sep. 2022.
J. Shi, B. Foggo, and N. Yu, "Power System Event Identification based on Deep Neural Network with Information Loading," IEEE
Transactions on Power Systems, vol. 36, no. 6, pp. 5622-5632, Nov. 2021.
Shaorong Zhang, Koji Yamashita, and Nanpeng Yu, "Learning Power System Dynamics with Neural Ordinary Differential Equations,"
IEEE PES General Meeting, 2024.
Yinglun Li, Nanpeng Yu, and Wei Wang, "Machine Learning-Driven Virtual Bidding with Electricity Market Efficiency Analysis," IEEE
Transactions on Power Systems, vol. 37, no. 1, pp. 354-364, Jan. 2022.
Yuanbin Cheng, Nanpeng Yu, Brandon Foggo, and Koji Yamashita, "Online Power System Event Detection via Bidirectional Generative
Adversarial Networks," to appear in IEEE Transactions on Power Systems, 2022.
Brandon Foggo and Nanpeng Yu, "Online PMU Missing Value Replacement via Event-Participation Decomposition," IEEE Transactions
on Power Systems, vol. 37, no. 1, pp. 488-496, Jan. 2022.
Xianghao Kong, Koji Yamashita, Brandon Foggo, and Nanpeng Yu, "Dynamic Parameter Estimation with Physics-based Neural Ordinary
Differential Equations" IEEE PES General Meeting, 2022.
Jie Shi, Koji Yamashita, and Nanpeng Yu, "Power System Event Identification with Transfer Learning Using Large-scale Real-world
Synchrophasor Data in the United States," IEEE ISGT North America, 2022.
Yuanbin Cheng, Koji Yamashita, and Nanpeng Yu, "Adversarial Attacks on Deep Neural Network-based Power System Event
Classification Models," IEEE ISGT Asia, 2022. 39
Recent Publications
Machine Learning for Power Distribution Systems and DERs
Farzana Kabir, Yuanqi Gao, Wenyu Wang and Nanpeng Yu, "Deep Reinforcement Learning-based Two-Timescale Volt-VAR Control
with Degradation-aware Smart Inverters in Power Distribution Systems," Applied Energy, vol. 335, April 2023.
Wei Wang, Nanpeng Yu, Yuanqi Gao, and Jie Shi, "Safe Off-Policy Deep Reinforcement Learning Algorithm for Volt-VAR Control
Problems in Power Distribution Systems," IEEE Transactions on Smart Grid, vol. 11, no. 4, pp. 3008-3018, 2020.
Daner Hu, Zhenhui Ye, Yuanqi Gao, Yonggang Peng, and Nanpeng Yu, "Multi-agent Deep Reinforcement Learning for Voltage Control
with Coordinated Active and Reactive Power Optimization," IEEE Transactions on Smart Grid, vol. 13, no. 6, pp. 4873-4866, Nov. 2022.
Zuzhao Ye, Yuanqi Gao, and Nanpeng Yu, "Learning to Operate an Electric Vehicle Charging Station Considering Vehicle-grid
Integration," IEEE Transactions on Smart Grid, vol. 13, no. 4, pp. 3038-3048, 2022.
Wenyu Wang and Nanpeng Yu, "Maximum Marginal Likelihood Estimation of Phase Connections in Power Distribution Systems," IEEE
Transactions on Power Systems, vol. 35, no. 5, pp.3906-3917, Sep, 2020.
Yuanqi Gao, Brandon Foggo, and Nanpeng Yu "A Physically Inspired Data-Driven Model for Electricity Theft Detection with Smart Meter
Data," IEEE Transactions on Industrial Informatics, vol. 15, no. 9, pp. 5076-5088, 2019.
Brandon Foggo and Nanpeng Yu, "Improving Supervised Phase Identification Through the Theory of Information Losses," IEEE
Transactions on Smart Grid, vol. 11, pp. 2337-2346, 2020.
Jie Shi, Yuanqi Gao, Wei Wang, Nanpeng Yu, and Petros Ioannou, "Operating Electric Vehicle Fleet for Ride-Hailing Services with
Reinforcement Learning," IEEE Transactions on Intelligent Transportation Systems, vol. 21, no. 11, pp. 4822-4834, Nov. 2020.
Yuanqi Gao and Nanpeng Yu, "Model-Augmented Safe Reinforcement Learning for Volt-VAR Control in Power Distribution
Networks," Applied Energy, vol. 313, May 2022.
Wenyu Wang and Nanpeng Yu, "Estimate Three-phase Distribution Line Parameters with Physics-informed Graphical Learning
Method," IEEE Transactions on Power Systems, vol. 37, no. 5, pp. 3577-3591, Sep. 2022.
Yuanqi Gao, Wei Wang, and Nanpeng Yu, "Consensus Multi-agent Reinforcement Learning for Volt-VAR Control in Power Distribution
Networks," IEEE Transactions on Smart Grid, vol. 12, no. 4, pp. 3594-3604, July, 2021.
Yuanqi Gao, Wei Wang, Jie Shi, and Nanpeng Yu, "Batch-constrained Reinforcement Learning for Dynamic Distribution Network
Reconfiguration," IEEE Transactions on Smart Grid, vol. 11, no. 6, pp. 5357-5369, 2020.
Farzana Kabir, Nanpeng Yu, Weixin Yao, Rui Yang, and Yingchen Zhang, "Joint Estimation of Behind-the-Meter Solar Generation in a
Community," IEEE Transactions on Sustainable Energy, vol. 12, no. 1, pp. 682-694, Jan. 2021. 40
 Contact Information:
Nanpeng Yu, [email protected]
Lab Website: https://intra.ece.ucr.edu/~nyu/

Thank You

Questions?

Collaborating Companies: SCE, PG&E, National Grid, Fortis BC, Con Edison, NYPA,
ComEd, Exelon, EPRI, RPU, ISO-NE and CAISO.
Funding Support: DOE, NSF, CEC, NYSERDA, EPRI, DEED, UCOP, NREL & Utilities.