IEEE PES ISGT NA 2019

Panel Session Monday (2/18)

Artificial Intelligence in Power System
Operations and Planning
Session Chair: Feng Qiu, ANL
Panelists:
Z. Wang (Iowa State University)
A. Santos Xavier (ANL)
D. Deka (LANL)
 1

Data-Driven and Machine Learning-

based Power Distribution Grid
Operation
Zhaoyu Wang
Harpole-Pentair Assistant Professor
Presented by Qianzhi Zhang
Iowa State University

ISGT 2019
Washington DC
 2

Introduction to the Observability Problem

• Challenge: Limited System Observability

Problem: How to Use

the Data to Enhance
System Observability?

• Modernization in Distribution Systems:

 Advanced Metering Infrastructure (AMI)
 Micro Phasor Measurement Units (μPMU)
 Distributed Energy Resources (DER)
 3

Available Utility AMI Data

• General Description:
Total Customers
Utilities Substations Feeders Transformers
Customer with Meters
3 5 27 1726 9118 6631

• Duration: 4 years (2014 - 2018)

• Measurement Type: Smart Meters and SCADA
• Data Time Resolution: 15 Minutes – 1 Hour
• Customer Type:
Residential Commercial Industrial Other
84.67% 14.11% 0.67% 0.55%
 4

Available Utility Data

Smart Meter Measurement Network Topology/Model
Data For Load Monitoring Information
Sample Customer Voltage

Sample Customer Energy Consumption

 5

AMI Data Pre-Processing

 Smart Meter Data Problems:
 Outliers/Bad Data
 Communication Failure
 Missing Data
 Counter-Measures:
 Engineering intuition
(data inconsistency)
 Conventional Statistical Tools
(e.g. Z-score)
 Robust Computation
(e.g. relevance vector machines)
 Anomaly Detection Algorithms
 6

Average Daily Consumption of Sample Customers

Residential
Commercial

Industrial
 7

Data-Driven Load Modeling: Seasonality

of Typical Customer Behavior

 Typical Residential Customer Load Pattern Bank

 Methodology: Data Clustering (Unsupervised Learning)
 8

Data-Driven Load Modeling: Seasonality of Load

Summer

Impact of seasons on the distribution

and time of a typical customer
demand peak

Winter
 9

Data-Driven Load Modeling: Sensitivity

of Load to Day and Type of Customers
Typical
Load
Patterns
on
Weekends

- Industrial - Commercial - Residential

Typical
Load
Patterns
on
Weekdays
 10

Statistical Load Models: Temporal

Correlations

Very Small Correlation

Between Customer Loads’ Correlation between
Smart Meter Time-Series Consumption Time-Series
(90% below 0.27) at Different Time-Scales
 11

Proposed Solution: Multi-Timescale Customer

Consumption Inference From Monthly Billing Data

• Challenge - Near-zero correlation between customer

monthly consumption and hourly load data
• Solution: Extending observability using data of
customers with smart meter to obtain a stage-by-stage
transition process (High Correlation!)
Y. Yuan, K. Dehghanpour, F. Bu, and Z. Wang, “A multi-timescale data-driven approach to enhance distribution system
• observability,” accepted by IEEE Trans. Power Syst., 2019.
 12

Proposed Solution: Multi-Timescale Customer

Consumption Inference From Monthly Billing Data

• Assigning Multi-
Timescale Models
to Different Typical
Load Behavior
Patterns Obtained
From Observed
Loads
 13

Numerical Results: Individual Customer Load

Inferring the hourly demand of an Impact of accurate consumption

unobserved commercial load in one pattern identification on the
month (average estimation error ≈ accuracy of the inference
8.5% of total energy) (industrial load patterns are
close and stable)
 14

Numerical Results: Individual Customer Load

Tracking the typical daily consumption Using inferred load for accurate
pattern of unobserved customers system monitoring (branch
using a Bayesian learning approach current state estimation)
 15

Numerical Results: Feeder Load

Comparison with the state of the art; Accuracy of the multi-timescale

method on a feeder with 80 residential customers (test set)
(Training performed over the data of 6,000 customers)
 16

Conclusions
• Access to Real Utility Data:
 Designing More Accurate and Realistic Load Models
 Discovering Strengths and Shortcomings of Available
Algorithms in the Literature
 Active Collaboration with Industry
• Using Data-Driven Load Inference for Monitoring
Distribution Systems:
 Identifying Temporal Correlations for Load Estimation
 Exploiting Latent Trends in Customer Behavior at Different
Time-Scales for Enhancing Inference Accuracy
 Facilitating Computational Efficiency for Real-Time Application
 17

Joint Structure & Parameter

Estimation in Distribution Grids

• Deep Deka

• Joint Work with:

Sejun Park (KAIST)
Michael Chertkov (LANL)
 18

Distribution Grid

• Final Tier in electricity transfer

• Voltage: High Medium Low

 19

Distribution Grid ‘Operational’ Structure

• Disjoint Trees
• 1 substation per tree (considered root)
• Switches/Relays decide structure

Original Structure Operational Forest

Substation Other nodes

 20

Learning Problems

• Structure Learning
• Configuration of switches

Substation
Load Nodes
Learning Problems

• Structure Learning
• Learning Line Impedances

R, X

Substation
Load Nodes
Learning Problems

• Structure Learning
• Learning Line Impedances
• Incomplete observations

Substation
Load Nodes
Missing Node
Applications of Learning Problems

• Structure:
• Failure/Fault Identification
• Connection/phase verification

• Impedance Estimation:
• Non-intrusive control
• Use in DSO optimization
• Optimizing flows

• Learning with Missing Data:

• Privacy quantification
• Meter Location Selection
 Related Approaches
• Topology/phase estimation
– Agnostic Machine Learning:
 Empirical evidence based (Clustering, Greedy approaches)
 Reno et al, Arya et al, von Mier et al., Rajagopal et al.
– Power Systems based Learning:
 Physics of flow to drive ML framework
 Bolognani et al, Rajagopal et al, Annaswamy et al, Low et al.

Our approach:
 Advantage: Provable results, Missing data extensions
 IEEE TCNS’17, PSCC’18
 Physics
(Power-Systems) Physics Informed
Interpretability

Informed Tuning Machine Learning

(Power System interpretable
but repetitive & off-line,
hand-controlled)

Physics-Free
PMU data-to-
predictions Machine Learning
approaches (automatic, training & execution efficient,
but lacking Power System interpretability)

Speed
 Advantage: Provable results, Missing data extensions
Learning with nodal voltages

• Data: Time-series Nodal voltages at all nodes

• Unobserved: all lines

• Estimate: Operational Topology
• IEEE TCNS’17
 Voltages in Radial Network

• Variance of voltage diff.:

• Minimum along any direction reached at nearest neighbor

𝑐𝑐 a

𝑏𝑏 𝑏𝑏
𝑏𝑏

𝑐𝑐

𝑎𝑎 𝑎𝑎
𝑐𝑐
 Topology Reconstruction

Greedy Topology Learning: (IEEE TCNS’17)

1. Spanning Tree with edge weights given by

Advantages:
• NO additional information needed a
(no impedance/injection statistics)
𝑏𝑏
• Complexity is O(ElogE)
• Works for monotonic flows (gas,water, heating)
• Can be extended to missing nodes
𝑐𝑐
 Topology Learning (No missing nodes)

33-bus test system, Matpower

Reference: 12 KV substation voltage
Extension to missing nodes:
Joint Estimation of Topology & Injections in Distribution
Grids with Missing Nodes, arXiv:1804.04742
Effect of Quantization Effect of Noise

3 decimal point: 1 V resolution

Learning with end-users

• Data: Time-series Nodal voltages and injection samples at leaves

• Unobserved: all intermediate nodes & lines
• Estimate: Operational Topology + Line Impedance
End-user data

• Time-stamped voltage magnitudes (V)

• Time-stamped nodal active & reactive injections (P &Q)

Cross-covariances:
Injection Data assumption
•• Assumption (Now)
(prior work)
: :
Observed are
Injections nodes are uncorrelated/independent
correlated and follow a Graphical Model

15- min detrended injections from households in DiSC data-set (Aalborg Univ.)
Injection Data assumption
• Assumption (Now) :
Injections are correlated and follow a Graphical Model

Graphical Model:
Correlation Conditional Dependence
Learning with end-users

• Data: Time-series Nodal voltages and injection samples at leaves

• Algorithm:

b
Learning with end-users

• Data: Time-series Nodal voltages and injection samples at leaves

• Algorithm:
 Compute effective impedances between leaf pairs

b
Learning with end-users

• Algorithm:
 Compute effective impedances between leaves

 Key: Effective resistances are additive on trees

b
 Learning with end-users

• Algorithm:
 Compute effective impedances between leaves
 Recursive Grouping Algo (Willsky’11) to learn topology & distances
from known effective impedances

Lemma: Given any additive distance 𝑑𝑑,

1. 𝑎𝑎, 𝑏𝑏 are leaf nodes with common parent iff

𝑑𝑑 𝑎𝑎, 𝑐𝑐 − 𝑑𝑑 𝑏𝑏, 𝑐𝑐 = 𝑑𝑑 𝑎𝑎, 𝑐𝑐 ′ − 𝑑𝑑(𝑏𝑏, 𝑐𝑐 ′ ) for all c
𝑐𝑐, 𝑐𝑐 ′ ≠ 𝑎𝑎, 𝑏𝑏
a
2. 𝑎𝑎 is a leaf node and 𝑏𝑏 is its parent iff
𝑑𝑑 𝑎𝑎, 𝑐𝑐 − 𝑑𝑑(𝑏𝑏, 𝑐𝑐) = 𝑑𝑑 𝑎𝑎, 𝑏𝑏 for all 𝑐𝑐 ≠ 𝑎𝑎, 𝑏𝑏 b
 Recursive Grouping Algo

1. 𝑎𝑎, 𝑏𝑏 are leaf nodes with common parent iff

𝑑𝑑 𝑎𝑎, 𝑐𝑐 − 𝑑𝑑 𝑏𝑏, 𝑐𝑐 = 𝑑𝑑 𝑎𝑎, 𝑐𝑐 ′ − 𝑑𝑑(𝑏𝑏, 𝑐𝑐 ′ ) for all 𝑐𝑐, 𝑐𝑐 ′ ≠ 𝑎𝑎, 𝑏𝑏

2. 𝑎𝑎 is a leaf node and 𝑏𝑏 is its parent iff

𝑑𝑑 𝑎𝑎, 𝑐𝑐 − 𝑑𝑑(𝑏𝑏, 𝑐𝑐) = 𝑑𝑑 𝑎𝑎, 𝑏𝑏 for all 𝑐𝑐 ≠ 𝑎𝑎, 𝑏𝑏

1. Learn siblings
 Recursive Grouping Algo

1. 𝑎𝑎, 𝑏𝑏 are leaf nodes with common parent iff

𝑑𝑑 𝑎𝑎, 𝑐𝑐 − 𝑑𝑑 𝑏𝑏, 𝑐𝑐 = 𝑑𝑑 𝑎𝑎, 𝑐𝑐 ′ − 𝑑𝑑(𝑏𝑏, 𝑐𝑐 ′ ) for all 𝑐𝑐, 𝑐𝑐 ′ ≠ 𝑎𝑎, 𝑏𝑏

2. 𝑎𝑎 is a leaf node and 𝑏𝑏 is its parent iff

𝑑𝑑 𝑎𝑎, 𝑐𝑐 − 𝑑𝑑(𝑏𝑏, 𝑐𝑐) = 𝑑𝑑 𝑎𝑎, 𝑏𝑏 for all 𝑐𝑐 ≠ 𝑎𝑎, 𝑏𝑏

2. Introduce parents

3. Update distance
Recursive Grouping Algo

1. Learn siblings
Recursive Grouping Algo

2. Introduce parents

3. Update distance
Recursive Grouping Algo

After Iterations
Learning with end-users

• Algorithm:
 Compute effective impedances between leaves
 Recursive Grouping Algo to learn topology & distances from known
effective impedances

 Threshold for finite samples effects: dynamically selected

 Less observability: Kron reduced network estimated
Effect of Correlated Injection

• Algorithm:
 Compute effective impedances between leaves
 Recursive Grouping Algo to learn topology & distances from known
effective impedances
Effect of Correlated Injection

• Algorithm:
 Compute effective impedances between leaves

 Uncorrelated Injections

– Two equations with 2 unknowns a

– No inversion necessary (2x2 matrix)
b
Effect of Correlated Injection

• Algorithm:
 Compute effective impedances between leaves

 Correlated Injections

– O(V) equations with O(V) unknowns

– Inverse Matrix estimation
Effect of Correlated Injection

• Algorithm:
 Compute effective impedances between leaves

 Correlated Injections

 Computing inverse
 Graphical Model for injections
Effect of Correlated Injection

• Algorithm:
 Compute effective impedances between leaves

 Correlated Injections

 Computing inverse
 Graphical Model for injections
 ML estimate (SPICE) for inverse:
 Sample Complexity

Uncorrelated : Correlated :
For a grid with constant depth 𝑂𝑂( 𝑉𝑉 2 log 𝑉𝑉 /𝜂𝜂 ) samples
and sub-Gaussian complex power recovers the true topology with
injections, 𝑂𝑂( 𝑉𝑉 log 𝑉𝑉 /𝜂𝜂 ) probability 1 − 𝜂𝜂.
samples recovers the true
topology with probability 1 − 𝜂𝜂.
• Steps:
 Bound decay of (covariance of injections with voltages)
 Bounds between empirical and true values of covariances
 Bound distance using Union bounds
• Learning with End-Users in Distribution Grids (PSCC 2018, TCNS under review)
• Learning with Correlated Injections (Asilomar 2018)
Simulations: IEEE 33 bus graphs (Matpower samples)

500 600 700 800 900 1000

Simulations: 30 bus random graphs

500 600 700 800 900 1000 500 600 700 800 900 1000 500 600 700 800 900 1000
Extensions:

• Straightforward:
 Three-phase unbalanced systems (under linearized flows)
 Change detection: Fault localization, nodal change
 Dynamical setting (LTI system) with swing equations
(CDC 2018, PSCC 2018)

• More Work:
 Radial  Loopy grids
Other Physics-Informed AI approaches in grid
• Neural Network features for ARC: Average Rank of Correct faulted line

fault localization

• Data-Driven Arbitrage using

Neural Network scenarios
(poster ISGT)

Generator models using

Reactive power
Neural Networks for
time-series data
OpenIPSL Generator ----- True
(12th order) ----- predicted
 Web: https://cnls.lanl.gov/gridscience
 Email: ansi-info@lanl.gov for post-doc positions and school.

Jan 7 - 11, 2019

Support from:
Linearized Power Flow (LinDistFlow)

• Non-linear AC flow:

• First order expansion:

Invertible linear model:

a

c b
•

wt. Laplacian matrix Incidence matrix

Slack Bus d
 1

Learning to Solve the Security-

Constrained Unit Commitment Problem

Alinson S. Xavier1 Feng Qiu1 Shabbir Ahmed2

1 Argonne National Laboratory, Lemont, IL

2 Georgia Institute of Technology, Atlanta, GA
 2

Security-Constrained Unit Commitment (SCUC)

• Problem: Find cost-efficient schedule for power generation

– Production during each hour must satisfy demand
– Power flows must be within safe limits
– Must withstand loss of a single transmission element
– Other physical, operational and economic constraints
• Widely used in power systems planning and operations:
– Day-Ahead Electricity Market
– Reliability assessment
– System expansion planning
 3

Need for Computational Performance

• Day-Ahead Electricity Market:
– Very short market clearing window (3 or 4 hours)
– Optimization occasionally ends with large optimality gaps
• Expansion Planning:
– High uncertainty; large number of scenarios
• Model Improvements:
– Accurate modeling of enhanced combined-cycle units
– Longer planning horizons, finer time granularity
– Higher resolution for production cost curves
– Reduction in optimality gap
 4

Technical Opportunity
• SCUC is solved repeatedly, multiple times per day:
– Day-ahead market: once per day
– Resource reliability and commitment: twice per day
– Reliability assessment commitment: 24 times per day
• Data only changes slightly from run to run:
– Almost same network topology and set of generators
– Slightly different demands and production costs
– Market operators have large amounts of historical data
• Perfect setting for Machine Learning
 5

Proposed Architecture
Current Approach Proposed AI Approach

Goal: Develop methods and tools that learn from past experience and become
progressively more efficient.
 6

Previous ML Research

• Solve SCUC using Artificial Neural Networks:

– Sasaki, Watanabe, Kubokawa, Yorino & Yokoyama (1992)
– Wang & Shahidehpour (1993)
– Walsh & O’mally (1997)
– Liang & Kang (2000)
• Use ML to enhance Integer Programming:
– Alvarez, Wehenkel, Louveaux (2014)
– Alvarez, Louveaux, Wehenkel (2017)
– Khalil, Dilkina, Nemhauser, Ahmed, Shao (2017)
 7

Our Approach and Results

• Our approach: Design ML models that help existing tools to

find the solution faster, based on statistical data:
– Predict redundant constraints
– Predict good initial feasible solutions
– Predict affine subspaces containing the solution
• Computational experiments:
– Realistic, large-scale instances (6515 buses, 1388 units)
– Uncertain production costs, peak system load, temporal &
geographical load distribution
• Results: 12 times faster, no negative impact on quality
 8

Simplified Formulation
 9

Learning Contingency Constraints

• Transmission constraints have large impact on performance:
– Quadratic number, typically very dense
– Very few are actually binding, but hard to tell in advance
• Transmission Model: Predict which constraints should be
added to the relaxation and which should be omitted
• Training phase:
– Solve problem without any transmission constraints
– Add small subset of most-violated constraints and resolve
– Repeat until no further violations are found
• Test phase: If constraint was necessary for at least 1% of
training cases, add at start, then follow previous algorithm
 10

Learning Initial Feasible Solution

• Primal bounds are still a bottleneck:

– Modern formulations usually yield very strong dual bounds
– Most time is spent finding high-quality primal solutions
• Warm Start Model: Find, among large set of previous
solutions, ones that are likely to work well as warm starts.
• Training phase (instance-based learning):
– Solve each training instance and store its solution
• Test phase:
– Find k training instances closest to the test instance
– Use their k optimal solutions as warm starts
 11

Learning Affine Subspaces: Hints

• Optimal solutions have a number of patterns:
– Some units are operational throughout the day
– Some units are only operational during peak demand
• Affine Subspace Model: Find subspaces (described by a set of
hyperplanes) where the solution is very likely to reside.
• Training phase:
– Consider a fixed set of candidate hyperplanes
– Build supervised models to predict if hyperplane is
satisfied. Discard models with low precision or recall.
• Test Phase:
– Predict which hyperplanes are likely satisfied using
previous models and add them to the relaxation.
 12

Learning Affine Subspaces: Training

• Hyperplanes considered:
– xgt= 0
– xgt = 1
– xgt= xg,t+1
• Classifier: Support Vector Machines
• Training high-quality models:
– Discard hyperplanes with very unbalanced labels
– Measure precision and recall using k-fold cross validation
– Discard models with low recall or precision not significantly
better than dummy predictor
 13

Computational Experiments: Instances

• Computational Environment:
– Java, Python, using pandas and scikit-learn
– IBM ILOG CPLEX 12.8.0 as MIP solver
– Intel Xeon E5-2695v4, 36 cores, 128GB DDR4
• Realistic, large-scale instances (adapted from MATPOWER):
– 9 instances in total
– Up to 6,515 buses
– Up to 1,388 generation units
– Up to 9,037 transmission lines
– Scale comparable to PJM and MISO
 14

Training and Test Sample Generation

• 300 variations of each instance generated for training

• 50 variations generated for testing
• Four sources of uncertainty:
– Peak system-wide load
– Production and start costs
– Geographic load distribution
– Temporal load profile
• Based on public data available from PJM and MISO
 15

Performance: Number of Iterations

 16

Performance: Fixed Variables

 17

Performance: Running Time (Chart)

 18

Performance: Running Time (Table)

 19

Performance: Optimality Gap

 20

Conclusion and Future Research

• Summary:
– Machine-learning methods to accelerate solution of SCUC
– 12 times speedup on large-scale, realistic instances
– No negative impact on solution quality
– https://arxiv.org/abs/1902.01697
• Learning-Enabled Optimization (LEO) for Power Systems:
– More general AI framework for optimization
– Applicable to other routinely-solved problems
– Self-adaptive tools that improve over time
– Enhance existing methods, instead of replacing them
 21

Acknowledgments

• This work is supported by Argonne National Laboratory

Directed Research and Development (LDRD) Swift Program.

• We gratefully acknowledge use of the Bebop cluster in the

Laboratory Computing Resource Center (LCRC) at Argonne
National Laboratory.