Deep Temporal Learning Versus Electrical Integration for Regenerative-Braking Energy and Peak-Power Classes

Regenerative-braking energy is determined by battery-pack current, voltage, and event duration, yet temporal networks have been proposed to estimate it from compressed event sequences. We audited four archived architectures (1D CNN, LSTM, GRU, and TCN) on 8181 events from one publicly available electric-vehicle dataset and reanalyzed their validation design. Events were represented by 32 samples of current, voltage, state of charge, temperature, and elapsed time. The original three-repeat comparison changed the grouped holdout between repeats and used globally derived energy-class thresholds; one extreme event moved between development and test. A direct 32-point trapezoidal-integration baseline, calculated from the identical network input, achieved R2 = 0.999879, RMSE = 0.554 Wh, and MAE = 0.147 Wh over all events. On fixed, non-overlapping recording-group and strictly forward whole-group tests, its RMSEs were 0.302 and 0.414 Wh, respectively. We additionally evaluate paired four- versus five-channel neural models, group-level dependence, distribution shift, and energy-magnitude versus peak-power classifications. The results show that this full-event, post-event task is largely an electrical-integration problem; high energy-derived classification scores should not be interpreted as recognition of an independent braking state. These single-vehicle, offline results do not establish cross-platform transfer or real-time control feasibility.