yzzpz/Charged-checkpoints

Charged checkpoints

Checkpoints for probabilistic forecasting of hourly station-level EV charging volume (Kairos_23m backbone; clean split, seq_l=72; H=24/48/72; crpsΣ main metric).

Code: https://github.com/yzzpz1015/Charged Data: https://huggingface.co/datasets/yzzpz/Charged-data

Contents

Directory	Files	Role
kairos_l1_clean/<CITY>/plain_pre72_fold6.pth	6 cities	① point prediction
kairos_l1_clean/<CITY>/quant_pre72_fold6.pth	6 cities	③ per-station quantile head (main method)
patchtst_quant_clean/<CITY>/patch_quant_pre24_fold6.pth	MEL, SPO	backbone × head ablation
csdi_clean/<CITY>_csdi.pth	5 cities	④ block CSDI
csdi_clean/<CITY>_zidiff.pth	5 cities	⑤ absorbing-state diffusion
timediff_nomix_clean/<CITY>_timediff.pth	MEL, SPO	TimeDiff reference (no future mixup)

Cities: LOA JHB MEL SPO SZH AMS. AMS is reported separately (~95.9% of stations frozen in the test period) and excluded from cross-city averages.

Notes

• The Kairos backbone (mldi-lab/Kairos_23m) is downloaded automatically by the training code.
• result/timediff_clean/ (future-mixup version) is intentionally not included: its sampler collapses (MPIW50 ≈ 0.02) and its metrics are invalid.
• Evaluation protocol and per-city numbers: see the GitHub README.