mrvictoru/AEMO_simulated_trade_sdp

AEMO SDP-Teacher Trajectories

Offline trajectories for Decision Transformer training, generated by replaying the honest SDP/MPC executor on historical Australian NEM (AEMO) market data. These are the teacher trajectories from the energydecision research codebase.

Each row is a single 5-minute market interval with a self-consistent (normalized observation, 9-dim action, reward) triple: the action is what the honest optimal planner dispatched, the reward is what it earned, and the observation is the environment state under that action.

2026-09 update (physics-v2): the shipped corpus was regenerated with the corrected degradation/efficiency physics and a wear-calibrated teacher (planner per-step wear matched to the environment's realized per-cycle wear; --deg-calibration 0.12). It supersedes dt_trajectories_jtsoc_combined.parquet.

Files

File	Episodes	Rows	RTG	Status
dt_trajectories_jtsoc_v2cal_conservative.parquet	320	3.13M	rtg_value = J_t(soc) cost-to-go	✅ Training data of the shipped physics-v2 model
dt_trajectories_jtsoc_combined.parquet	640	6.27M	rtg_value = J_t(soc)	Historical (v1 Stage C training data; pre-physics-fix)

(Removed 2026-09: the Stage B discounted-return corpora dt_trajectories_full.parquet and dt_trajectories_aggressive.parquet, which paired with the deleted Stage B model weights.)

What is different about the physics-v2 corpus?

1. Wear-calibrated teacher. The planner's per-step wear estimate previously over-counted the environment's realized (per-closed-cycle) wear by ~8.4× on identical actions, making the distilled policy over-conservative. A deg_calibration factor was added and the corpus regenerated with --deg-calibration 0.12, so the teacher's planned wear matches the environment's accounting. Net effect vs the pre-fix corpus: energy dispatch ≈ −27%, FCAS bidding ≈ +14%, and the retrained standalone DT beats the previous shipped checkpoint on 3/4 identity surfaces.
2. Corrected shared physics. Round-trip efficiency (0.80), calendar aging, and corrected rainflow C-rate units are active in both the teacher and the environment.

The shipped model trained on this corpus is energydecision-dt-v2-sdp (aemo_dt_sdp_jtsoc_v2cal.pt). Train on it with --rtg-source j_t_soc (consumed via the rtg_value column).

Schema

Column	Type	Description
step	int	Step index within the episode
norm_observation	list[float] (18)	Normalized observation: time (hour/day sin/cos), is_peak, RRP, TOTALDEMAND, 8× FCAS prices, solar/wind fraction, SOC fraction
raw_observation	list[float] (18)	Raw (unnormalized) observation, same layout
action	list[float] (9)	full_fcas action: [energy_dispatch, 8× FCAS bids]; energy ∈ [-1,1] (charge +, discharge −), FCAS ∈ [0,1]
reward	float	Per-step env reward (profit in $, divided by 1000)
info	dict	Env step info (revenue/degradation breakdown)
episode_id	str	{REGION}__{horizon}__{battery}__epNNN
source_policy	str	Always sdp_teacher
region	str	NSW1, QLD1, SA1, TAS1, or VIC1
battery	str	medium_1c (10 MWh), fast_375c (8 MWh), large_07c (50 MWh), or small_05c (2 MWh)
rtg_value	float	jtsoc_* files only — J_t(soc) cost-to-go at this step (state-dependent remaining value under the seasonal forecast)

Dataset composition

• Shipped physics-v2 corpus: 320 episodes / 3.13M rows
• 5 regions × 2 horizons (short ≈ 14-day, medium ≈ 8-week) × 4 battery sizes × 8 episodes each — balanced
• Time range: 2021–2023 (training-era, pre-eval)

How it was generated

Each episode is a random window sliced from the region's cached preprocessed 5-minute AEMO data (processed_{REGION}_2021-2023.parquet), replayed through a hierarchical planning teacher. The teacher is a three-part pipeline, not a standalone SDP solver:

1. SOC-waypoint DT predicts a coarse target-SOC trajectory (K = 8 waypoints). This is a separate model, hosted at mrvictoru/energydecision-waypoint-dt and required to reproduce this corpus. It was trained on SOC-waypoint labels derived from perfect-foresight Oracle-LP paths, generated by scripts/generate_soc_waypoint_targets.py (GitHub; the ~1.3 GB label parquet is regenerated from the repo, not hosted here).
2. Honest SDP executor (AEMOSDPSolver in src/aemo_sdp_executor.py) plans energy dispatch between the waypoints, under:
   • a seasonal time-of-day RRP forecast built from training-era data only (the predictable diurnal/monthly structure; spikes excluded);
   • a degradation-aware per-step wear estimate scaled by deg_calibration (corrected per-step wear; see known-issues B4).
3. Greedy current-price FCAS bidder allocates residual headroom to the highest-priced raise/lower service each step.

The free-terminal recursion additionally yields the J_t(soc) cost-to-go table, sampled per step into rtg_value.

No future (realized) prices are used at generation time — the executor is honest (non-clairvoyant); perfect foresight appears only in the waypoint-DT training labels. The full pipeline is scripts/generate_sdp_dt_trajectories.py in the energydecision repo.

Usage

import polars as pl

df = pl.read_parquet("hf://datasets/mrvictoru/AEMO_simulated_trade_sdp/dt_trajectories_jtsoc_v2cal_conservative.parquet")
print(df.height)  # 3,133,440

Decision Transformer training

The schema matches the TrajectoryDataset used by scripts/pretrain_decision_transformer.py. Train a 9-dim full_fcas Decision Transformer (8×768, mixed Tanh/Sigmoid action head) on the physics-v2 corpus with --rtg-source j_t_soc → the shipped configuration.

Related

• Shipped model: energydecision-dt-v2-sdp
• Source repo: mrvictoru/energydecision
• Older AEMO corpus (behaviour-cloning era): AEMO_simulated_trade