Accurate modeling of Electric Power Steering (EPS) systems is essential for predicting electrical power demand in modern vehicle electrical and electronic (E/E) architectures. This paper investigates sequential deep learning approaches for data-driven EPS torque prediction on real-world driving data (∼2.2 million samples, 4988 driving sequences), using a sequence-level evaluation protocol with block-bootstrap confidence intervals and permutation tests. Explicit temporal modeling via LSTM significantly outperforms non-sequential MLP baselines, with medium-to-large effect sizes (Cohen’s d = 0.70–1.12, p < 0.001), confirming that sequential processing of vehicle dynamics signals is essential for accurate torque estimation. This paper further evaluates whether attention mechanisms, including Simple, Additive (Bahdanau), and Scaled Dot-Product variants, can improve upon the LSTM baseline. None yields a practically relevant improvement, with effect sizes remaining negligible (|d| < 0.1) for Simple and Additive Attention, while Scaled Dot-Product Attention significantly degrades performance (p < 0.001). Analysis of the learned attention weights reveals two failure modes, recency collapse and uniform collapse, explaining why attention cannot contribute beyond what the LSTM hidden state already encodes for short automotive time series. Additionally, a small LSTM with 85K parameters achieves near-identical performance to medium models with 598K parameters at 3.5× lower inference latency, establishing it as the recommended architecture for deployment in resource-constrained automotive systems.