Machine learning for digital twins to predict responsiveness of cyber-physical energy systems

Cyber-Physical Systems are becoming more autonomous, interconnected, complex and adaptive, and are expected to operate in highly dynamic environments. This is especially challenging for energy ecosystems that are increasingly difficult to control and maintain as the number of participating manufacturers and users grows. Digital Twins help analyze and predict these systems in the form of digital reflections that operate in parallel with the physical system. In this paper, we use Machine Learning to improve the predictive power of Digital Twins for Cyber-Physical Energy Systems. Specifically, we use a Temporal Convolutional Neural Network model to learn the temporal patterns in the system and predict its responsiveness to specific power setpoint instructions. Real-life data from ten batteries were used to predict the behavior over time. Compared to the baseline model that uses the prior probability of response and the average response rate within the configured time window, the model predicts the batteries' responsiveness more accurately. The more temporal information is used as input for prediction, the better the model performs in both precision and recall. The results show that this compensates for the lack of information when fewer metrics are used. The use of Machine Learning for Digital Twins can help maintain a heterogeneous energy ecosystem, while minimizing the need to acquire or disclose detailed information.