EnergyDecision-DT-V2-Forecast

⚠️ NEGATIVE RESULT — Superseded by the Stage C standalone DT.

This ForecastDecisionTransformer (48-step TTM forecast tokens + modern v2 backbone) scored $4,564/ep on the standard tier — 8.5% below the modern v2 baseline ($4,991/ep). Explicit TTM price forecasts add no meaningful edge over the implicit 210-step context window.

The shipped model is now the Stage C standalone DT distilled from an honest SDP-planning teacher, which beats PPO on all 4 identity surfaces and passes the impact gate without any explicit forecast tokens:

Surface	Stage C DT	Forecast DT	PPO
Standard Oct	$11,573	$4,564	$2,353
Dispatch-matched	$35,320	—	$22,530

This checkpoint is retained as a documented negative result. See report.md §8.2.8 for the full analysis.

---

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).

Key Features

• Forecast token prefix: 48 TTM-generated price forecasts prepended as a learned prefix
• RoPE enabled: Rotary Position Embeddings for position-aware sequence modeling
• Modern architecture: 8×768 GQA, QK-Norm, SwiGLU, RMSNorm, weight tying
• Action Space (9-dim): Energy dispatch + 8 FCAS contingency bids
• Context Length: 210 history + 48 forecast = 258 timesteps total

Why This Didn't Work

The TTM forecasts have near-zero correlation with FCAS prices (~0.01–0.07). The model simply learns to ignore the forecast tokens and rely on history alone — the same information the standard DT already has.

Usage

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,
)