energydecision-dt-v2-sdp

AEMO SDP-Teacher Decision Transformer (planner-distilled, physics-v2)

The shipped / preferred model of the energydecision benchmark. A standalone Decision Transformer distilled from an honest (non-clairvoyant), degradation-aware SDP-planning teacher — no solver in the control loop (the J_t(soc) prompt is precomputed) — that beats the PPO reference on all four identity surfaces and remains strong under the market-impact benchmark.

2026-09 update (physics-v2): this model was retrained on the corrected degradation/efficiency physics (round-trip efficiency 0.80, calendar aging, corrected rainflow C-rate units) and a wear-calibrated teacher (--deg-calibration 0.12, aligning the planner's per-step wear with the environment's realized per-cycle wear). See Provenance below. It supersedes the earlier aemo_dt_sdp_jtsoc_fullcorpus.pt.

Files in this repo

(Removed 2026-09: the behaviour-cloning-era files aemo_dt_fcas_model.pt and aemo_dt_fcas_best_checkpoint.pt.)

Results (physics-v2 shipped model, aemo_dt_sdp_jtsoc_v2cal.pt, rtg_mode="auto")

All four identity surfaces beat PPO with paired-difference 95% CIs excluding zero. Methodology, the physics fixes, and the full experiment ladder: report.md §8.2.10/§8.2.11 and docs/known_issues.md.

Provenance: what changed in physics-v2

The model card's numbers above are from the corrected pipeline. Prior household and AEMO numbers produced before these fixes are historical.

Architecture

Note on return_scale: 23,284.83 is the correct value to use at inference, but it does not reproduce exactly as p90(|J_t|)/10 of the released corpus (recomputed ≈21,067). The checkpoint sidecar is authoritative; the discrepancy is documented in the paper's provenance appendix.

Action space: dim 0 energy dispatch in [-1, 1]; dims 1–8 FCAS bids in [0, 1] (RAISEREG, LOWERREG, RAISE6SEC, LOWER6SEC, RAISE60SEC, LOWER60SEC, RAISE5MIN, LOWER5MIN). State space: 18-dim normalized market observations (time features, RRP, demand, 8 FCAS prices, generation mix, SOC).

Training data

• Shipped: dt_trajectories_jtsoc_v2cal_conservative.parquet — 320 episodes, 3.13M rows, honest SDP teacher with wear calibration and per-episode J_t(soc) cost-to-go RTG. Hosted at mrvictoru/AEMO_simulated_trade_sdp.
• Generated by scripts/generate_sdp_dt_trajectories.py (--deg-calibration 0.12, --rtg-mode j_t_soc); method in report.md §4.3/§5.4.
• Corpus-generation dependency (not an inference dependency): the hierarchical teacher uses a separate SOC-waypoint DT (mrvictoru/energydecision-waypoint-dt) to propose SOC targets; its label generator is scripts/generate_soc_waypoint_targets.py in the GitHub repo. The shipped policy does not need this model at inference.

Usage

import json
import torch
from huggingface_hub import hf_hub_download
from decision_transformer import DecisionTransformer
from decision import AEMOAgent

repo = "mrvictoru/energydecision-dt-v2-sdp"
ckpt = hf_hub_download(repo, "aemo_dt_sdp_jtsoc_v2cal.pt")
meta = json.loads(hf_hub_download(repo, "aemo_dt_sdp_jtsoc_v2cal.pt.meta.json"))

model_kwargs = dict(
    state_dim=meta["model"]["state_dim"], act_dim=meta["model"]["act_dim"],
    h_dim=meta["model"]["h_dim"], context_len=meta["model"]["context_len"],
    n_block=8, n_heads=12, n_kv_heads=6, drop_p=0.15,
    max_timestep=2016, rope_enabled=True, qk_norm=True, tie_weights=True,
    action_head_mode="mixed",
)
model = DecisionTransformer(**model_kwargs)
model.load_from_checkpoint(ckpt)             # raw state_dict
model.return_scale = meta["return_scale"]    # 23284.83
model.eval()

agent = AEMOAgent(env, algorithm="dt", model=model, rtg_mode="auto")

rtg_mode="auto" resolves to state-dependent j_t_soc prompting on price-taking (identity) surfaces and a conservative constant-RTG fallback under a market-impact model.

Inference prerequisites (what else you need)

1. Source code. The DecisionTransformer model, AEMOAgent, and the AEMOBatteryTradingEnv live in the energydecision repo, not in this model repo.
2. Seasonal RRP profiles (bundled here). For rtg_mode="j_t_soc" / "auto" on identity surfaces, the agent builds a J_t(soc) cost-to-go table from a seasonal price forecast, cached to data/aemo_sdp/seasonal_rrp_<REGION>.json. The five profiles are included in this repo under data/aemo_sdp/ — copy them to data/aemo_sdp/ in the repo checkout (otherwise the j_t_soc path has to rebuild them from raw AEMO parquets, which the simulation environment fetches separately). Building the table runs one backward-induction pass per episode (not per step); because it depends only on the seasonal forecast and the SOC grid, it can be precomputed or cached, after which each step is a constant-time lookup.
3. Alternative without profiles. rtg_mode="constant" (with a prompt value) needs no seasonal profile, but it is not the shipped setting and under-performs on identity surfaces.

Intended use

Research into offline RL / sequence modeling for energy markets; simulation of multi-market BESS dispatch in Australia's NEM; baseline for planner-distilled offline RL. Not intended for live trading without further validation (sim-to-real is open), risk management, and regulatory compliance.

Citation

@misc{energydecision-dt-v2-sdp,
  title = {AEMO SDP-Teacher Decision Transformer (planner-distilled, physics-v2)},
  author = {mrvictoru},
  year = {2026},
  url = {https://huggingface.co/mrvictoru/energydecision-dt-v2-sdp}
}

Related

• Dataset: mrvictoru/AEMO_simulated_trade_sdp
• Modern v2 (behaviour-cloning era, superseded): energydecision-dt-v2
• Impact-aware variant: energydecision-dt-v2-impact
• Source repo: mrvictoru/energydecision