EnergyDecision-DT-V2-Forecast: Forecast-Conditioned Decision Transformer for AEMO FCAS Battery Trading

Model Description

EnergyDecision-DT-V2-Forecast is a ForecastDecisionTransformer — an extension of the modern v2 Decision Transformer that conditions on 48-step TTM price forecasts alongside the historical observation/action sequence. It predicts optimal battery dispatch and FCAS bids using both past market data and predicted future prices (RRP, demand, 4 FCAS services).

This is the first model in the benchmark to use explicit price forecasts as conditioning tokens, removing the information bottleneck of pure history conditioning.

Key Features

• Forecast token prefix: 48 TTM-generated price forecasts prepended as a learned prefix. History tokens attend to forecast tokens via the existing causal mask.
• RoPE enabled: Rotary Position Embeddings for better position-aware sequence modeling.
• Modern architecture: Grouped-Query Attention (6 KV heads, 12 Q heads), QK-Norm, SwiGLU FFN, RMSNorm pre-norm, weight tying.
• Action Space (9-dim): Energy dispatch + 8 FCAS contingency bids.
• State Space (18-dim): Normalized market observations.
• Context Length: 210 history + 48 forecast = 258 timesteps total.

Intended Use

This model is intended for:

• Research into forecast-conditioned offline RL for energy markets
• Simulation of battery trading strategies using both historical data and price forecasts
• Baseline for comparing forecast-enhanced transformers against standard Decision Transformers and RL baselines

It is not intended for live trading without further validation, risk management, and regulatory compliance.

Training Data

• Source: AEMO simulated trade dataset
• Datasets used: aemo_fcas_dataset.parquet (FCAS v2, 76.9M rows) + aemo_sdp_trajectories.parquet (SDP, 2.1M rows) + aemo_fcas_grpo_dataset.parquet (GRPO, 11.9M rows)
• Forecast conditioning: ttm_forecasts.npz — TTM (Granite TTM-R3) 48-step normalize price forecasts for 6 channels (RRP, demand, 4 FCAS services), aligned via episode_start column
• Source policies: SB3(PPO, A2C, etc), GRPO-DT, SDP energy-only

Model Architecture

ForecastDecisionTransformer(
  (embed_return):      Linear(1 -> 768)
  (embed_state):       Linear(18 -> 768)
  (embed_action):      Linear(9 -> 768)
  (embed_timestep):    Embedding(100000 -> 768)
  (embed_forecast_type): Embedding(2 -> 768)  # NEW: history(0) vs forecast(1)
  (embed_ln):          RMSNorm(768)
  (transformer): 8x ModernBlock(
    (norm1): RMSNorm(768)
    (attn): CausalSelfAttention(
      q_proj: Linear(768 -> 768)   # 12 Q heads × 64 head_dim
      k_proj: Linear(768 -> 384)   # 6 KV heads × 64 head_dim
      v_proj: Linear(768 -> 384)   # 6 KV heads × 64 head_dim
      out_proj: Linear(768 -> 768)
      qk_norm: RMSNorm(64) per head
      n_rep: 2 (GQA)
    )
    (norm2): RMSNorm(768)
    (ffn): SwiGLU(768 -> 3072 -> 768, dropout=0.15)
  )
  (ln_f): RMSNorm(768)
  (pred_act):   Linear(768 -> 9)  -> Tanh   [tied with embed_act]
  (pred_state): Linear(768 -> 18)         [tied with embed_state]
  (pred_return): Linear(768 -> 1)         [tied with embed_return]
)

Forecast token processing: At each forward pass, the model receives forecast_states (48 timesteps × 18-dim) and forecast_rtgs. These are embedded with type=1 (forecast) and prepended before the history tokens. The model's causal attention allows history tokens to attend to all forecast tokens + prior history.

Hyperparameters

Parameter	Value
Blocks	8
Hidden dim	768
Attention heads (Q)	12
KV heads (GQA)	6
History context length	210
Forecast length	48
Total sequence length	774 (210+48 × 3 interleaved)
Dropout	0.15
QK-Norm	✅ Enabled
Weight tying	✅ Enabled
RoPE	✅ Enabled
State dim	18
Action dim	9
Discount factor	0.95
Return scale	2.0
Loss weights	action=0.999, state=0.002, return=0.0001

Training Results

Loss Curves

Epoch,Train Loss,Val Loss 1,0.041705,0.007797 2,0.006725,0.005105 3,0.005220,0.004579 4,0.004722,0.004228

Epoch	Train Loss	Val Loss
1	0.041705	0.007797
2	0.006725	0.005105
3	0.005220	0.004579
4	0.004722	0.004228

Best validation loss: 0.004228 (epoch 4). Global steps completed: 43,648.

Training Configuration

Parameter	Value
Optimizer	AdamW (lr=3e-5, weight_decay=1e-4)
Batch size	64
Epochs	4
Scheduler	StepLR (gamma=0.9, step=1 epoch)
Gradient clip norm	1.0
Mixed precision	AMP (torch.cuda.amp)

RTG Calibration

The forecast DT uses forecast_states as conditioning, which changes the optimal RTG prompt. For the best dispatch-matched profit, calibrate RTG on your evaluation surface. Note that the forecast DT's RTG response differs from the standard DT — preliminary testing suggests rtg_value=0.0 remains effective, but full calibration sweeps are ongoing.

Usage

Installation

The model is loaded with the ForecastDecisionTransformer class from the energydecision repository:

git clone https://github.com/mrvictoru/energydecision.git
cd energydecision

Loading the Model

import torch
from forecast_decision_transformer import ForecastDecisionTransformer

model = ForecastDecisionTransformer(
    state_dim=18,
    act_dim=9,
    n_block=8,
    h_dim=768,
    context_len=210,
    forecast_len=48,
    n_heads=12,
    drop_p=0.15,
    max_timestep=100000,
    rope_enabled=True,
    n_kv_heads=6,
    qk_norm=True,
    tie_weights=True,
)

state_dict = torch.hub.load_state_dict_from_url(
    "https://huggingface.co/mrvictoru/energydecision-dt-v2-forecast/resolve/main/forecast_dt_model.pt",
    map_location="cpu"
)
if "model_state_dict" in state_dict:
    model.load_state_dict(state_dict["model_state_dict"])
else:
    model.load_state_dict(state_dict)
model.eval()

Inference with Forecast Tokens

import torch

# History inputs (batch, context_len, dim)
states = torch.zeros(1, 210, 18)
actions = torch.zeros(1, 210, 9)
rtgs = torch.full((1, 210, 1), 0.0)  # RTG prompt — use 0.0
timesteps = torch.arange(210).unsqueeze(0)
mask = torch.ones(1, 210, dtype=torch.bool)

# Forecast inputs (batch, forecast_len, dim) — from TTM model
forecast_states = torch.randn(1, 48, 18)  # predicted market states
forecast_rtgs = torch.randn(1, 48, 1)     # forecast RTGs
forecast_timesteps = torch.arange(210, 210 + 48).unsqueeze(0)

with torch.no_grad():
    ret, state, action = model(
        states, actions, rtgs, timesteps, mask,
        forecast_states=forecast_states,
        forecast_rtgs=forecast_rtgs,
        forecast_timesteps=forecast_timesteps,
    )
    # action[0, -1] is the next-step action (9-dim)

Using with TTM Forecasts at Inference

In a real deployment, the TTM forecasts would be generated on-the-fly from the AEMO context window:

# Forward: TTM takes last 512 steps of prices, predicts next 48
prices_context = aemo_data.iloc[current_step - 512:current_step]
ttm_forecast = ttm_model(past_values=prices_context)  # predicted prices
forecast_states = assemble_observation(ttm_forecast)   # build 18-dim from predictions

Training Infrastructure

Resource	Details
Hardware	CUDA-capable GPU
Framework	PyTorch, transformers, granite-tsfm
Precision	Mixed precision (AMP)
Checkpointing	Every 500 batches + best model on validation loss
Source Code	github.com/mrvictoru/energydecision

Remarks

• Forecast quality: The TTM forecasts are generated by a fine-tuned ibm-granite/granite-timeseries-ttm-r3 model (512-48-dec-512-r3 revision). See the dataset page for details on the forecast data.
• No perfect foresight: The model trains on the same imperfect TTM forecasts it receives at inference — no train-inference gap.
• Architecture gap: Whether the forecast conditioning provides meaningful improvement over the standard DT is an open research question. The forecast DT is a step towards incorporating forward-looking information that the standard DT (210-step context window) cannot access.

Standard Tier Leaderboard (Jul 2026)

Oct 2024 · 5 NEM regions · 144h · medium_1c · full_fcas · best RTG per model

Model	Profit/ep	FCAS/ep	Deg/ep	Best RTG
Modern v2	$4,991	$4,836	$229	10.0
Dispatch Dalrymple North	$4,660	$2,287	$1,020	—
Forecast DT	$4,564	$3,663	$270	50.0
Phase C GRPO (mod v2)	$4,322	$2,508	$1,058	10.0
Phase 1 GRPO (legacy)	$2,678	$2,914	$384	50.0
PPO reference	$2,353	$2,192	$236	—

Key findings:

• RTG calibration (0–100) revealed every DT variant gains 5–75%. The original 0.0–2.0 range was far too narrow.
• Modern v2 peaks at RTG=10 ($4,991/ep). Forecast DT peaks at RTG=50 ($4,564/ep).
• The forecast DT is a well-implemented negative result — explicit TTM price forecasts do not beat the implicit 210-step context window.
• FCAS TTM forecasts have near-zero correlation (~0.01–0.07), limiting any forecast-conditioned approach.