Generalizable Temperature Nowcasting with Physics-Constrained RNNs for Predictive Maintenance of Wind Turbine Components

TL;DR

This work tackles predictive maintenance for wind turbine gearbox bearings under partial system knowledge by proposing a physics-constrained recurrent neural network (PC-RNN) for bearing temperature nowcasting. The model integrates a physics-based heat transfer relation with learnable coefficients $\lambda_1,\lambda_2,\lambda_3$, and imposes a physics loss based on an Euler discretization to enforce plausibility. Compared to a baseline RNN and a Linear surrogate, PC-RNN demonstrates improved generalization to unseen wind farms and operates without reliance on wind speed forecasts, which reduces a major source of uncertainty. The approach is broadly applicable to other systems with partially known physics and can enhance predictive maintenance in low-data regimes.

Abstract

Machine learning plays an important role in the operation of current wind energy production systems. One central application is predictive maintenance to increase efficiency and lower electricity costs by reducing downtimes. Integrating physics-based knowledge in neural networks to enforce their physical plausibilty is a promising method to improve current approaches, but incomplete system information often impedes their application in real world scenarios. We describe a simple and efficient way for physics-constrained deep learning-based predictive maintenance for wind turbine gearbox bearings with partial system knowledge. The approach is based on temperature nowcasting constrained by physics, where unknown system coefficients are treated as learnable neural network parameters. Results show improved generalization performance to unseen environments compared to a baseline neural network, which is especially important in low data scenarios often encountered in real-world applications.

Figures (5)

• Figure 1: Physics-constrained recurrent neural network (PC-RNN).
• Figure 2: Generalization performance for models trained on Plant B. Error ranges $\pm 1$ SE.
• Figure 3: Feature densities for the individual datasets from different wind plants.
• Figure 4: Example of PC-RNN $T_t^b$ prediction (left) and standardized model gradients from the PC-RNN physics component (Equation \ref{['eq:dT_neuralnetwork']}) with learned coefficients (right). Grey areas show time steps where model gradient and true gradient have different sign.
• Figure 5: Model generalization performance for training set size of 6 WTGs per dataset. The grey areas mark results for unseen WTGs from training datasets (in-plant generalization).