Toward real-time junction temperature estimation with a physics-informed attention LSTM

Accurate real-time junction temperature estimation for automotive power electronics is essential to guarantee reliability under dynamic operating conditions. Purely datadriven recurrent networks, such as Long Short-Term Memory (LSTM), can capture complex transients but, without physical constraints, may yield biased or poorly calibrated predictions; conversely, purely physics-based lumped models can be robust yet miss nonlinearities and operating-context variability. We propose a physics-informed LSTM that enforces steady-state and transient energy-balance constraints derived from a lumped 1-RC thermal model. The network is optimized through Bayesian hyperparameter search. On proprietary automotive drive cycles, the physics-informed model delivers substantial gains: relative to a plain LSTM, the mean absolute error is reduced by 18.6%; relative to a calibrated 1-RC ODE, the mean squared error is lowered by 56.6%. Predictive intervals obtained via Monte Carlo dropout are well calibrated: for nominal 95% intervals, empirical coverage is 98.25% for the LSTM and 99.54% for the physicsinformed model, with comparable sharpness. These results show that embedding lightweight physics in sequence models improves prediction reliability without additional computational overhead, making the approach suitable for onboard deployment.