Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage...

38
Detection of Unregistered Electric Vehicles Using a Recurrent Neural Network By: Ben Groene Strategies for Intelligent Low- Voltage Network Monitoring

Transcript of Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage...

Page 1: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Detection of Unregistered Electric Vehicles Using a Recurrent Neural Network By: Ben Groene

Strategies for Intelligent Low-Voltage Network Monitoring

Page 2: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

1. Introduction

2. Experimental Method

3. Model

4. Results

5. Conclusion & Future Development Needs

Agenda

2

Page 3: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

3

1. Introduction

Page 4: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• Germany • Goal: Reach 1 million EVs on the road by 2020

• Reality: 1 million EVs projected to be reached but not until 2021 or 2022

Projection of Accumulated New EV Registrations through 2025 in Germany

1.1 Electromobility in Germany

4 Source: Nationale Plattform Elektromobilitaet

Page 5: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• EV penetration expected to rise → Distribution system not designed for high EV penetration

• Most distribution systems lack monitoring devices → Distribution systems should be monitored for existing and future problems caused by EVs

• Proposed solution → Detection of EV charging events using a neural

network!

1.2 Problem Statement

5

Page 6: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

1. Program initial machine learning model using Keras machine learning framework in Python

2. Prepare background data

3. Add Artificial EV charging events to background data

4. Train model what EV charging looks like • Supervised Learning – Tell model at which timesteps there is EV charging and

at which timesteps there is not.

5. Optimize model

6. Given a sequence of previously unseen measurement data, predict at

which timesteps an EV was connected

1.3 Summary of Task

6

Page 7: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

0 = No EV charging, 1 = EV charging

1.4 Prediction of Electric Vehicles - Intuition

7

Page 8: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

8

2. Experimental Method

Page 9: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• Data obtained from PMU • 500-millisecond current data in Germany

• Measured at low-voltage substation

• Feeder characteristics (as of 2016): • 400 kVA transformer supplied four feeders with a total of:

189 households

21 PV systems (160 kW installed capacity)

• PMU installed on feeder with highest PV penetration

2.1 Background Data

9

Page 10: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• In most cases, EVs charge to full power within 10 seconds, which can be seen in the 2 EV charging examples below

2.1 Background Data

10 Source: Fachhochschule Bingen, 2013

Page 11: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• Data converted to 10-second resolution • Simplify structure of charging events (current spike)

• Reduce time, computational and memory requirements

• End of charging often tapered or stepped • Difficult to detect

• Data transformed into delta values • Takes away impact of actual current level

2.1 Background Data

11

Page 12: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

2.1 Background Data

12

• A month of data after converting to 10s delta values

• Many delta values in EV charging range

Page 13: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• Artificial EV charging events in range 15-33 A inserted into background data • 2 charging events per day → two random times, two random amplitudes in

above range

2.2 Artificial EV Injection

13

Page 14: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• Separate data into three

different sets 1. Training set (60%)

2. Cross-validation (CV) set (20%)

3. Test set (20%)

2.3 The Three Data Sets

14

Page 15: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

15

3. Model

Page 16: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• Bidirectional Recurrent Neural Network (BRNN)? • State-of-the-art technique for dealing with sequential data such as time-series data

• Temporal dependencies within the time-series can be learned

• A single training example is an entire sequence (1 day in this case)

• Standard NNs cannot deal with sequential data effectively

3.1 Sequence Models and BRNNs

16

Page 17: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• RNNs have trouble remembering information over long sequences • Known as vanishing gradient problem

Solution: Long Short-Term Memory (LSTM) cells

• LSTM cells have a special architecture which allows them to remember information over long sequences • State-of-the-art for time-series data

• Available for use via Keras machine learning framework

3.1 Sequence Models and BRNNs

17

Page 18: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

3.1 Sequence Models and BRNNs

18

Source: Medium

• Example of BRNN • Input = Sequence (8,640 timesteps,

delta values)

• Output = Sequence (8,640 timesteps, 0…1)

• Each timestep is sent through a separate neural network with a unique output

• Each neural network shares information with previous timestep and next timestep

Page 19: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

19

4. Results

Page 20: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

4.1 Metrics

20

Page 21: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

4.2 Final Model & Performance

21

Final Model Precision Recall Test Set 0.9428 0.8545

85.5% of charging events detected with a false positive rate of 5.7%!

Page 22: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

4.2 Final Model & Performance

22

• Lower charging events more difficult to detect • Still able to detect over 70% of 15 A charging events

Page 23: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• Measurement data was available for one other feeder which was used to test transferability of the model

• A switching event took place on the feeder after the measurement data was taken for this project → New data was used as another transferability test as well as to test robustness of model • Test Case #1: EV charging events injected into new data exactly as before

• Test Case #2: EV charging events injected with charging events in the range of 12-16 A instead of 15-33A

4.3 Transferability

23

Page 24: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Test Case #2 (12-16 A) Precision Recall Test Set 0.9778 0.7458

Alternative Feeder Precision Recall Test Set 1.0000 0.9857

4.3 Transferability

24

Test Case #1 Precision Recall Test Set 1.0000 0.9828

Page 25: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

25

5. Conclusion & Future Development Needs

Page 26: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• The model was able to detect approx. 85.5% of EV charging events in the test set with a false positive rate (1-precision) of only 5.7%

• Initial transferability tests show signs that the model will also perform well on other feeders

• Model could detect events of lower charging amplitudes with only a slightly lower detection rate • Even when the model is trained only on events in the range 15-33 A

5.1 Results Summary

26

Page 27: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

• Further transferability testing

• More data should be obtained to test transferability

• Uncertainty how model will perform on actual EV charging events • Pilot projects

• Changes in EV charging characteristics • Controlled charging methods

5.2 Future Development Needs

27

Page 28: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

28

Thank you for your attention!

Page 29: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

29

ANNEX

Page 30: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Annex 1: Smart Meter Rollout

30 Source: Deloitte, 2016

Page 31: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Annex 2: Testing Data Preprocessing Methodology with Real Charging Event

31

Page 32: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Annex 3: Bias vs. Variance

32

• The more complex the model is → the more likely the model is to overfit the training data & vice versa

Source: cv-tricks.com

Page 33: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Annex 4: Standard Neural Network Example

33

Page 34: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Annex 5: Long Short-Term Memory (LSTM)

34

Source: adapted from, Colah.github.io, 2015

Page 35: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Annex 6: Transferability - Test Case #3

35

Page 36: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Annex 7: Final Model & Performance

36

Page 37: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Annex 8: Final Model & Performance

37

Page 38: Strategies for Intelligent Low- Voltage Network Monitoring€¦ · • Measured at low-voltage substation • Feeder characteristics (as of 2016): • 400 kVA transformer supplied

Annex 9: Final Model & Performance

38

Final Model Precision Recall Training Set 0.9672 0.9213 Test Set 0.9428 0.8545

85.5% of charging events detected with a false positive rate of 5.7%!