Analysis plan for "Donor Choice in Zero-Shot Load-Forecast Transfer: What It Costs, and Why a Pretrained Model Wins"

Written analysis plan for the study "Donor Choice in Zero-Shot Load-Forecast Transfer: What It Costs, and Why a Pretrained Model Wins", submitted to GECOST 2027. The study measures the day-ahead forecast error of a model trained on one Moroccan distribution feeder and applied without retraining to another, across all 272 ordered pairs of 17 smart-metered feeders in four cities. This document records, for each of 15 analyses, the procedure and the reading rules that decide how its outcome is interpreted. The plan covers: the replication gate; the transfer-ratio estimand and dyadic bootstrap; the seasonal-naive baseline; the donor-selection experiment; pooled same-sector models; alternative feeder labels; shape descriptors; the tuned local reference; weather covariates and two discriminating tests of weather sensitivity; second and summer evaluation blocks; few-shot local references; an additive feeder-identity comparison; a permutation test of sector labels; a partition analysis of where pooled transfer beats the naive; a second learner; and a zero-shot foundation-model baseline. Each section records the procedure and the reading rules for one analysis, written before that analysis was run. This record was deposited after the work was completed, so it documents that order rather than independently certifying it. Several of the rules returned negative or inconclusive outcomes, and all are reported in the paper, including the finding that no donor-transfer configuration beats a seasonal naive and that a pretrained model outperforms every configuration measured. Data: https://www.kaggle.com/datasets/malaklahmidi1/moroccan-feeder-load-time-series-30-min-cleaned

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-21
DOI
https://doi.org/10.5281/zenodo.24409084
Primary Topic
Energy Load and Power Forecasting
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Analysis plan for "Donor Choice in Zero-Shot Load-Forecast Transfer: What It Costs, and Why a Pretrained Model Wins"

Fatima Zahra Baayou
1 links
Zenodo (CERN European Organization for Nuclear Research)
Energy Load and Power Forecasting
preprint

Analysis plan for "Donor Choice in Zero-Shot Load-Forecast Transfer: What It Costs, and Why a Pretrained Model Wins"

Fatima Zahra Baayou
preprint en

Abstract

Written analysis plan for the study "Donor Choice in Zero-Shot Load-Forecast Transfer: What It Costs, and Why a Pretrained Model Wins", submitted to GECOST 2027. The study measures the day-ahead forecast error of a model trained on one Moroccan distribution feeder and applied without retraining to another, across all 272 ordered pairs of 17 smart-metered feeders in four cities. This document records, for each of 15 analyses, the procedure and the reading rules that decide how its outcome is interpreted. The plan covers: the replication gate; the transfer-ratio estimand and dyadic bootstrap; the seasonal-naive baseline; the donor-selection experiment; pooled same-sector models; alternative feeder labels; shape descriptors; the tuned local reference; weather covariates and two discriminating tests of weather sensitivity; second and summer evaluation blocks; few-shot local references; an additive feeder-identity comparison; a permutation test of sector labels; a partition analysis of where pooled transfer beats the naive; a second learner; and a zero-shot foundation-model baseline. Each section records the procedure and the reading rules for one analysis, written before that analysis was run. This record was deposited after the work was completed, so it documents that order rather than independently certifying it. Several of the rules returned negative or inconclusive outcomes, and all are reported in the paper, including the finding that no donor-transfer configuration beats a seasonal naive and that a pretrained model outperforms every configuration measured. Data: https://www.kaggle.com/datasets/malaklahmidi1/moroccan-feeder-load-time-series-30-min-cleaned

Zenodo (CERN European Organization for Nuclear Research)
Al Akhawayn University (MA)
Energy Load and Power Forecasting
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon
Analysis plan for "Donor Choice in Zero-Shot Load-Forecast Transfer: What It Costs, and Why a Pretrained Model Wins" — Fatima Zahra Baayou · Zenodo (CERN European Organization for Nuclear Research) (2026) | TGRS Research Map | TGRS