Digital Twin-Based Architecture for Run-Time Power Modeling for Sensorless Edge Devices

Run-time power and thermal management software are key to optimizing any embedded system to save energy. Power feedback is required to make an informed decision which can be taken either from power sensors or power models for power management. In most of the edge devices, there is a lack of dedicated power sensors as it is costly, and deploying them at scale across multiple devices can be difficult. Power models are the most used solution to estimate power and are trained on seen workloads. However, adaptability is a major concern if an unseen workload comes, resulting in the model’s accuracy drops increasing mean square error. This paper proposes a digital twin-based power model capable of handling unknown workloads without hampering the model’s accuracy removing the need for retraining on the edge. The proposed method has been proved on two power model techniques (i.e. Linear regression and Random forest) on Nvidia’s Jetson Nano platform which makes this generalized. There is a significant improvement in power estimation accuracy for new or unseen workloads. Specifically, when utilizing the linear regression model and the random forest model, the mean squared error (MSE) is reduced to 87% and 94% compared to the state-of-the-art method for the unseen workload respectively. Furthermore, the proposed architecture achieves an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value of 0.99 and a mean average percentage error (MAPE) of 0.12% compared to state-of-the-art that has R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value of 0.92 and MAPE of 2.9% for unseen workloads and saving 20,061 2-input NAND gate equivalent area.