CAformer-based temperature prediction for traction motors of EMUs

Traction motor temperature prediction is essential for monitoring the operational condition of electric multiple units. From a computational perspective, temperature sequences exhibit complex long-range dependencies as well as strong local temporal correlations, posing significant challenges for conventional models. Moreover, autoregressive forecasting strategies tend to suffer from error accumulation in long-horizon prediction tasks. To address these issues, this study proposes a novel data-driven framework termed CAformer. Unlike existing convolution-augmented Transformer models that rely on simple module stacking, CAformer introduces a causality-aware convolution–attention coupling mechanism, in which causal convolution is tightly integrated into the self-attention process to jointly model local temporal dynamics and global dependencies in a unified manner. In addition, a non-autoregressive prediction paradigm with implicit temporal dependency preservation is developed to mitigate cumulative prediction errors while maintaining temporal consistency across the prediction horizon. Furthermore, a learnable representation mechanism is adopted to enhance the adaptability of temporal feature encoding under complex operating conditions. Experiments conducted on real operational data collected from CR300BF electric multiple units demonstrate that the proposed model consistently outperforms several representative baseline models, including recurrent neural network, gated recurrent unit, Transformer, Informer, and PatchTST, achieving lower mean absolute error, root mean square error, and maximum absolute error. The results further show that CAformer is capable of accurately capturing both fine-grained fluctuations and large-scale variations in temperature sequences.