[PatrickHenkel1]

Phase 3 of the physXAI 1.0.0 refactoring, the core feature: single step models are evaluated and trained over a horizon, as in an MPC. Stacked on #72 (base v1/phase-2-expr).

Summary

• Recursive evaluation (physXAI.evaluation.rollout):
  • PreprocessingSingleStep(..., eval_horizon='24h') applies every single step model recursively on windows of the val/test/train data.
  • The model must predict one of its own inputs: Change(T) or T with input T and its lags.
  • RMSE, MAE and bias per step are stored in td.horizon_metrics and {name}_training_data.json, and plotted over the horizon in hours.
• Derived inputs are recomputed in every step: expressions inside the model anyway; data columns such as vT from their expressions. This was impossible with constructed features.
• Implementation: RecursiveCell is a keras RNN cell around the step model and is differentiable. scikit-learn linear regressions are wrapped as a frozen dense layer.
• Training on the multi step error: recursive_training=RecursiveTraining(horizon='24h', first_step_weight=0.25, mode='single_then_multi').
  • Available for all keras models; ModularANN(solver='ols') can serve as the initialization.
  • Only the step model weights are trained, so AgentLib-MPC still receives a single step model.
• Example: executables/bestest_hydronic_heat_pump/TAir_recursive.py (physics terms plus a residual MLP).

Results on the BOPTEST data (test split, 12 h horizon)

	RMSE 1 h	RMSE 6 h	RMSE 12 h	bias 12 h
single step training	0.109 K	0.619 K	0.617 K	0.34 K
+ recursive fine-tuning (6 h)	0.095 K	0.468 K	0.405 K	0.19 K

The single step RMSE is nearly identical for both (0.038 / 0.037). Only 14 days of data, i.e. 4 test windows, so the numbers are indicative.

Fixes

• The legacy multi step evaluation wrote its predictions into the training data.
• BuildContext cached tensors by id() of temporary expressions, which Python reuses (wrong tensors).
• Early stopping on val_loss without validation data; seeds were set at object creation.

Test plan

• pytest: 102 passed. New tests cover exact rollouts (error < 1e-3 over the horizon), comparison with a hand-written recursion including the recomputed vT, an offset showing up as an accumulating bias, lags and absolute outputs, step weights and both training modes.
• All executables run.

🤖 Generated with Claude Code

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has insufficient remaining quota. Try requesting a review at a lower effort level.

An organization owner can choose the repository organization as the billing source for reviews requested by members with a paid Copilot license. Learn about Copilot code review billing.