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 Stage-B 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. This makes the data directly consumable by return-conditioned behaviour-cloning models (e.g. Decision Transformers).

Files

File	Description
dt_trajectories_full.parquet	Conservative teacher — SDP energy dispatch with deg_cost_per_mwh=50. FCAS-heavy, lower energy cycling.
dt_trajectories_aggressive.parquet	Aggressive teacher — SDP energy dispatch with deg_cost_per_mwh=20. Energy-arbitrage-heavy (+57% energy dispatch vs conservative), slightly lower FCAS.

Together the two files bracket the FCAS↔energy trade-off spectrum; training on their concatenation lets return-to-go (RTG) conditioning select the operating point at inference.

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)

Dataset composition

• 320 episodes / 3.13M rows per file
• 5 regions × 2 horizons (short ≈ 14-day, medium ≈ 8-week) × 4 battery sizes × 8 episodes each — perfectly 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 the honest SDP/MPC executor:

1. A seasonal time-of-day RRP forecast is built from training-era data only (the predictable diurnal/monthly price structure; spikes are deliberately excluded).
2. A stochastic dynamic program (AEMOSDPSolver) plans the energy dispatch over the episode (backward induction over the forecast, degradation-aware via a rainflow model + linear throughput surrogate).
3. A greedy current-price FCAS bidder allocates the residual headroom to the highest-priced raise/lower service at each step.

No future (realized) prices are used — the teacher is honest (non-clairvoyant) and degradation-aware. See src/aemo_sdp_executor.py and 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_full.parquet")
print(df.height)  # 3,133,440

Decision Transformer training

The schema matches the TrajectoryDataset used by the energydecision repo's scripts/pretrain_decision_transformer.py. Train a 9-dim full_fcas Decision Transformer (8×768, mixed Tanh/Sigmoid action head) on dt_trajectories_full for a conservative-FCAS-heavy clone, or on the concatenation of both files for RTG-selectable FCAS↔energy behaviour.

Related

• Source repo: mrvictoru/energydecision
• Pretrained Decision Transformers: energydecision-dt-v2
• Older AEMO corpus: AEMO_simulated_trade