Data-driven Power Loss Identification through Physics-Based Thermal Model Backpropagation

TL;DR

This paper tackles the challenge of estimating power losses in power-electronic devices for digital twins when direct loss measurements are unavailable. It introduces a cascaded, physics-informed hybrid framework that backpropagates through a reduced-order thermal model to correct a nominal power-loss model using only temperature data. Two neural architectures, a bootstrapped feedforward network and a recurrent network, are studied with normalization and physics-guided losses to ensure stability and physical consistency. Results on simulated and real data show substantial reductions in both temperature estimation error from $7.2\pm6.8^{\circ}\mathrm{C}$ to $0.3\pm0.3^{\circ}\mathrm{C}$ and power-loss error from $5.4\pm6.6\mathrm{W}$ to $0.2\pm0.3\mathrm{W}$, enabling real-time, sensor-constrained deployments.

Abstract

Digital twins for power electronics require accurate power losses whose direct measurements are often impractical or impossible in real-world applications. This paper presents a novel hybrid framework that combines physics-based thermal modeling with data-driven techniques to identify and correct power losses accurately using only temperature measurements. Our approach leverages a cascaded architecture where a neural network learns to correct the outputs of a nominal power loss model by backpropagating through a reduced-order thermal model. We explore two neural architectures, a bootstrapped feedforward network, and a recurrent neural network, demonstrating that the bootstrapped feedforward approach achieves superior performance while maintaining computational efficiency for real-time applications. Between the interconnection, we included normalization strategies and physics-guided training loss functions to preserve stability and ensure physical consistency. Experimental results show that our hybrid model reduces both temperature estimation errors (from 7.2+-6.8°C to 0.3+-0.3°C) and power loss prediction errors (from 5.4+-6.6W to 0.2+-0.3W) compared to traditional physics-based approaches, even in the presence of thermal model uncertainties. This methodology allows us to accurately estimate power losses without direct measurements, making it particularly helpful for real-time industrial applications where sensor placement is hindered by cost and physical limitations.

Figures (7)

• Figure 1: Power Loss Model and Thermal Model interconnected in cascade for temperature estimation.
• Figure 2: Data-Driven correction configuration that acts directly to correct the nominal estimated power losses.
• Figure 3: Distribution representation to select the simulated, unknown, real parameters from different distributions.
• Figure 4: Hybrid model validation Losses for all the scenarios and configurations. The results show how the configuration with bootstrap can reach better performance with respect to a traditional in any scenario.
• Figure 5: hybrid model comparison. Top: temperature and power loss estimation accuracy between physics-based and hybrid models. Bottom: absolute error distributions, showing accurate outperforming noisy by $58\%$ in temperature estimation and $82\%$ for power losses.