連合学習がデータ主権とグローバルサウス横断の協働的エネルギー転換モデリングを両立：インドネシアを軸とした公開データベンチマーク

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:連合学習の実データ規模での有効性と限界（モデル容量・データ量依存）を定量化した基準を提供する。

🏢実務担当者:越境データ共有が難しい途上国インフラ投資のデューデリジェンスや協働モデリングの設計に応用できる。

🏛政策担当者:国家AIラボと共有訓練基盤によるデータ主権を保った国際協調の枠組み設計の根拠となる。

<title>Abstract</title> <p>Meeting Paris-aligned trajectories in the Global South requires rapid, data-intensive coordination of where and how low-carbon energy infrastructure is built, yet the underlying data are fragmented across sovereign jurisdictions reluctant to pool sensitive energy, land and infrastructure information. Wealthy blocs are answering their own version of this problem with shared or sovereign compute at extraordinary cost — a €10-billion European network of publicly funded ‘AI Factories’ and a US$500-billion private buildout in the United States — scales no developing country can replicate alone. Federated learning (FL), in which institutions jointly train a shared model by exchanging model parameters rather than raw data, offers an alternative technical foundation, but its value for cross-country energy modelling in developing regions has not been quantified. Treating 44 Global South countries (16,404 operating facilities) in the Global Energy Monitor Global Integrated Power Tracker as sovereign data silos, we benchmarked four training regimes on a geospatial renewable-versus-fossil classification task. Federated averaging matched centralized pooling of all raw data to three decimal places (AUC 0.895 versus 0.895) while no record left its country of origin, and personalized FL exceeded isolated national training (0.910 versus 0.905; two-sided Wilcoxon p = 0.048), with the advantage larger on average among data-poor silos and widening as training data were artificially restricted. Independent replication on the 2021 WRI Global Power Plant Database (31 silos, 6,545 plants) reproduces the sovereignty result, though there the personalization advantage is statistically indistinguishable from zero. A high-capacity gradient-boosted model trained on pooled data outperforms every federated configuration tested (0.960 versus 0.815 trained locally), a gap driven by training-set volume rather than implicit country identification: the median silo holds 63 training facilities against 11,479 pooled. Aggregation therefore has large and rising value with model capacity, no country's own data suffice for the most capable methods, and extending federation to such models is the decisive next step. These results provide a quantitative foundation for federated architectures — national AI Labs coupled to shared, renewably powered training infrastructure under parity governance — as coordination infrastructure for Global South energy transitions, with Indonesia as the natural first national pilot.</p>

• Research Square https://doi.org/10.21203/rs.3.rs-10722871/v1first seen 2026-11-13 04:29:17