Pretrained models can violate round-trip energy consistency at grades

[CubicsYang]

For synthetic 1-mile links at a fixed speed, I evaluated an uphill link at +grade_pct, a downhill link at -grade_pct, and a flat link at 0% with the direct routee.powertrain v2 API. For each condition, I calculated

round_trip_excess = E(+g) + E(-g) - 2 * E(0)

If the predictions represent energy required for the two directed links, the potential-energy contribution should cancel over the matched pair and losses should make this quantity non-negative. I instead observe negative values for both a BEV and a hybrid model, including a large negative region for Model Y at 15–20 mph.

I am opening this as a question / possible physical-consistency issue rather than asserting that all negative downhill predictions are incorrect. A
negative downhill BEV prediction may be a valid representation of regeneration; the concern is the negative result after pairing equal and opposite grades at identical distance and speed.

Environment and pinned models

• routee-powertrain==2.0.2
• Python conda environment: routee-powertrain
• Model Y: tesla/model_y_bev/2022/rf_base_fe510e40/v1
  (electric_kwh, kilowatt-hour)
• RAV4 Hybrid LE: toyota/rav4_hybrid_le_hev/2022/rf_base_fe510e40/v1
  (fuel_gge, gallons gasoline)
• Direct model features: speed_mph, grade_pct, distance_mi
• Inputs: distance = 1 mi; speeds = 15, 20, …, 45 mph; positive grades =
  0.25% to 15.00% in 0.25-point increments.

This test calls routee.powertrain.load_model(...).predict(...) directly.

Here is the Jupyter notebook

routee_powertrain_grade_asymmetry_minimal_repro.ipynb

[nreinicke]

Thanks for opening and for providing a detailed description and example. I was able to reproduce your findings and agree that this is an issue with the physical practicality of these predictions. I'm going to dive in deeper today and will let you know more as I work through it.