21
Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems Ron Snijders (TNO), Paolo Pileggi (TNO), Jeroen Broekhuijsen (TNO), Jacques Verriet (TNO / ESI), Marco Wiering (University of Groningen), Koek Kok (TNO / TU Eindhoven) 8 th Workshop on Modeling and Simulation of Cyber-Physical Energy Systems Virtual Workshop April 21 2020

Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

  • Upload
    others

  • View
    0

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Ron Snijders (TNO), Paolo Pileggi (TNO), Jeroen Broekhuijsen (TNO), Jacques Verriet (TNO / ESI), Marco Wiering (University of Groningen),Koek Kok (TNO / TU Eindhoven)

8th Workshop on Modeling and Simulation of Cyber-Physical Energy Systems Virtual Workshop April 21 2020

Page 2: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Outline

2 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

• Introduction

• Digital Twin

• Energy System Pilot

• Response Prediction Model

• Experiments

• Conclusions

Page 3: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Introduction

3 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

• Goals:

1. Use Machine Learning to improve the predictive power of Digital Twins for Cyber-Physical Energy Systems

2. Experimental validation using battery data from a energy smart grid pilot in the Dutch town of Heerhugowaard.

Page 4: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Digital Twin in the CPSoS Context

4 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Monitor

Execute

Digital Thread

Physicalsystem

DigitaltwinAnalyze

Plan

Page 5: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

The Heerhugowaard Energy Smart Grid Pilot

5 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

For more information:https://flexible-energy.eu/http://flexiblepower.github.io/

Page 6: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Anomaly Detection using Expert System

6 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Page 7: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Predict Responsiveness using ML

7 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Page 8: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Response Prediction Model

8 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

• Aim:

• Given historical metric data (power setpoint, power, current,

voltage and state-of charge),

• predict whether the battery is likely (1) or unlikely (0) to respond

to a negative power setpoint (i.e. it should provide power).

Page 9: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Power Request Events

Goal: Will it respond or not?

9 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Page 10: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

• Using a Temporal Convolutional Network (TCN), a form of Deep Learning:• Outperforms recurrent neural networks (LSTM/GRU) on many standard benchmarking datasets.

• https://arxiv.org/pdf/1803.01271.pdf• https://arxiv.org/pdf/1803.01271.pdf

Method

10 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Page 11: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

• The TCN is extended with an extra hidden dense layer of 20 neurons.

• Makes use of standard techniques in Deep Learning, such as:• The Adam algorithm is used for stochastic gradient decent.• The Rectified Linear Unit (ReLU) is used as the activation function throughout the

network.• With the exception of the output neuron, which uses a sigmoid function.

• Binary cross-entropy is used as the loss function.

• Using the Python Deep Learning library Keras as the implementation and an existing TCN implementation as the basis for our implementation.

• https://github.com/philipperemy/keras-tcn

Method

11 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Page 12: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Experiments

12 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

• The dataset used, consists of:• 8 months of data collected from 10 different batteries. • A total of 972 power request events.

• The data is noisy, misses data and contains changing usage patterns:

Page 13: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Experiments

13 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

• Baseline method used for comparison:• Based on the mean historical response of the battery within the time window or the complete lifetime of the battery.

Page 14: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Experiments

14 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

• Evaluation method used:

• The first 85% was used to train and optimize a single model.

• The remaining 15% (the holdout set) was used to report and evaluate the final performance.

• Evaluation based on accuracy, precision, recall, F1 score and the Matthews Correlation Coefficient (MCC).

Page 15: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Experiments

15 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

• The following conditions were evaluated:

• Window size of either 1, 4, 8, 16, 32 and 64 hours.

• Two different sets of metrics: • TCN-all: Using all metrics (power, power setpoint, current, voltage, state-of-charge).

• TCN-pwr: Using only power and power setpoint.

• All variations of the TCN model were trained for 300 epochs.

Page 16: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Results

16 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Page 17: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Results

17 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Page 18: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Conclusions

18 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

• Machine Learning can help, by way of the Digital Twin, to better utilize existing Cyber-Physical Energy Systems.

• Using Machine Learning helps to predict whether batteries will respond to the instructions they receive more reliably.

• The method benefits from using more temporal data, compensating for the potential lack of information.

• Useful to improve management of heterogeneous energy ecosystems composed of non-transparent systems.

Page 19: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Future Research

19 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Page 20: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Acknowledgements

The authors like to thank: • All parties involved in the Heerhugowaard flexible energy system for generating and providing the real-life battery data.

• TNO colleagues Coen van Leeuwen, Wilco Wijbrandi and Mente Konsman in specific for their insightful discussions and advice.

20 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems

Page 21: Responsiveness of Cyber-Physical Energy Systems Machine ...palensky.org/mscpes/2020/MSCPES_2020_slides_5.pdf•Evaluation method used: •The first 85% was used to train and optimize

Thank You!

Any Questions?

Please contact:

Ron [email protected]

www.tno.nl

21 | Machine Learning for Digital Twins to Predict Responsiveness of Cyber-Physical Energy Systems