From Forecast Uncertainty to Economic Value: Stochastic and Robust MILP for Demand-Side Energy Management
Year: 2026
Type: software
Abstract: Training Randomness in Architecture Ablations — v1.0.0 Reproducibility package accompanying the manuscript: Training Randomness in Architecture Ablations: A Crossed Seed–Window Analysis with Preregistered Replication Contents crossed seed–window analysis materials; REML and modified Hartung–Knapp meta-analysis; dataset-clustered sensitivity analysis; simulation-based power and coverage analysis; preregistered ECL replication materials; OpenML-CC18 feature-identity ablation on 17 classification tasks; 17 per-task CC18 prediction artifacts; configurations, manifests, checksums, and compact derived result tables; manuscript and Supplement. OpenML-CC18 analysis The non-energy benchmark evaluates a feature-identity ablation in a compact FT-Transformer. Balanced error is the primary metric, with cross-entropy as a secondary sensitivity metric. Dataset-level effects are pooled using REML with modified Hartung–Knapp inference. Reproducibility The CC18 pipeline begins with: src/cc18/00_preflight_check.py and continues through dataset selection, model training, per-dataset analysis, sparsity diagnostics, and REML–Hartung–Knapp pooling. The 17 per-task NPZ files are included in: artifacts/cc18_per_example_artifacts.zip File integrity can be verified using: CHECKSUMS.sha256 Licensing Source code: MIT License. Author-created configurations, manifests, documentation, compact derived results, and generated CC18 artifacts: CC BY 4.0. Source datasets remain subject to their original licenses and terms. The manuscript is not covered by the repository licenses pending the journal publishing agreement.
Language: English
Topic: Smart Grid Energy Management
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203451209