1corr Decision-Focused TFT + SRRL

This directory adds a new pipeline without modifying the existing TFT code in 1using.

Main entry:

python 1corr\joint_train.py --rounds 3

The pipeline uses the existing 1using/train_tft1.py checkpoint format, predicts the 96-point TFT horizon, selects the default SRRL time points 0,24,48,72, converts them to the normalized wind/sun/load profile consumed by 1using/3main.py, then alternates:

1. train SRRL on the current TFT profile through 3main.py --profile-npz;
2. evaluate local profile perturbations with the current SRRL policy;
3. fit a differentiable surrogate for reward/cost;
4. update TFT with supervised loss plus surrogate decision loss.

Quick profile-only smoke run:

python 1corr\joint_train.py --dry-run

To run only the TFT/surrogate side with an existing SRRL checkpoint:

python 1corr\joint_train.py --skip-srrl --init-srrl-checkpoint path\to\final_model.pt