energydecision-waypoint-dt

AEMO SOC-Waypoint Decision Transformer (teacher helper)

A small Decision Transformer that predicts target-SOC waypoints for the hierarchical planning teacher used to build the AEMO_simulated_trade_sdp corpus.

This is NOT the shipped policy. The shipped standalone policy is energydecision-dt-v2-sdp (aemo_dt_sdp_jtsoc_v2cal.pt), which needs no solver in the control loop (its J_t(soc) prompt is precomputed) and not this model at inference. This checkpoint is a corpus-generation input only.

What it does

The teacher is a two-stage hierarchy:

1. This model predicts a coarse target-SOC waypoint trajectory (K = 8 waypoints) over the episode.
2. The honest SDP executor (src/aemo_sdp_executor.py) plans energy dispatch between waypoints under a seasonal price forecast.
3. A greedy FCAS bidder allocates residual headroom to the highest-priced raise/lower service each step.

The resulting (observation, action, reward) triples form the SDP-teacher corpus.

How it was trained

• Labels: data/aemo_dt_soc_oracle/aemo_soc_waypoints.parquet — SOC waypoint targets produced from perfect-foresight Oracle LP paths.
• Label generator (GitHub, not shipped here): scripts/generate_soc_waypoint_targets.py. The label dataset is large (~1.3 GB) and is regenerated from the repo + cached AEMO data, so it is not hosted here.
• Loss: masked MSE on the 8-dimensional sigmoid waypoint output (see the *_loss_surface_manifest.json).

Note on foresight: perfect foresight appears only in the waypoint labels. The corpus-generation executor itself is honest (seasonal forecast, no realized future prices) — see the dataset card "How it was generated".

Files

Architecture

Usage — regenerate the SDP-teacher corpus

# in the energydecision repo checkout
python3 scripts/generate_sdp_dt_trajectories.py \
  --regions NSW1 QLD1 SA1 TAS1 VIC1 \
  --horizons short medium \
  --batteries medium_1c fast_375c large_07c small_05c \
  --episodes-per-slot 8 \
  --model-path <path to>/soc_waypoint_dt_best.pt \
  --deg-cost-per-mwh 50.0 --deg-calibration 0.12 \
  --rtg-mode j_t_soc \
  --out data/aemo_dt_sdp/dt_trajectories_jtsoc_v2cal_conservative.parquet

The default --model-path is models/aemo/dt/soc_waypoint_dt_best.pt; download this checkpoint there (or pass the path explicitly).

Related

• Shipped policy: energydecision-dt-v2-sdp
• Training corpus: AEMO_simulated_trade_sdp
• Source repo (generators, env, executor): mrvictoru/energydecision