価格・炭素排出の不確実性下における最適制御のための決定論的学習:持続可能な水処理のケーススタディ
Decision-Focused Learning for Optimal Control under Price and Carbon Emission Uncertainty: A Case Study in Sustainable Water Treatment (原題)
Oufan Hai
🤖 gxceed AI 要約
日本語
太陽光・蓄電池を備えた水処理施設の最適運用を、電力価格と炭素強度の不確実性下で検討した修士論文。将来価格を予測するMLモデルと最適化を分離する従来のpredict-then-optimizeに対し、決定論的学習(DFL)を適用。7種のDFL手法と時系列予測モデルを比較し、DFLが運用コストを最大3.8%削減することを示した。
English
This thesis applies decision-focused learning (DFL) to receding-horizon control of a solar- and battery-equipped water treatment plant under electricity price and carbon intensity uncertainty. Unlike predict-then-optimize, DFL trains the forecasting model on decision quality. Benchmarking 7 DFL methods and 7 forecasting models, DFL cut operating costs by up to 3.8% versus the two-stage baseline.
Unofficial AI-generated summary based on the public title and abstract. Not an official translation.
📝 gxceed 編集解説 — Why this matters
日本のGX文脈において
電力価格と炭素強度を同時に扱う運用最適化は、再エネ調達や炭素会計と親和性が高く、SSBJ開示におけるScope 2削減策の実務的裏付けとなりうる。日本のエネルギー多消費インフラの脱炭素運用に示唆を与える。
In the global GX context
Demonstrates how ML-based operational control can jointly optimize cost and carbon under market uncertainty, relevant to Scope 2 reduction and transition planning under ISSB/TCFD. Adds empirical evidence on task-specific ML losses for decarbonization operations.
👥 読者別の含意
🔬研究者:MLと最適化を統合するDFLが脱炭素運用制御で有効であることを示す実証ベンチマークとして参考になる。
🏢実務担当者:電力価格・炭素強度の変動下で蓄電池・再エネを運用する施設のコスト・排出同時最適化の手法として応用可能。
🏛政策担当者:炭素強度の高い電力市場における需要側の柔軟な運用がコスト・排出削減に寄与する点を政策設計に反映できる。
📄 Abstract(原文)
Water treatment plants (WTPs) are critical infrastructure but also significant energy consumers, accounting for approximately 2% of energy use in the United States and contributing significantly to greenhouse gas emissions. In this thesis, we consider the problem of optimal control for the sustainable operation of a water treatment plant equipped with solar panels and batteries under uncertainty in electricity prices and carbon emissions. We formulate it as a receding-horizon control problem in which input parameters, such as future prices and carbon intensity, are provided by a time-series forecasting model. The traditional predict-then-optimize approach to solving a decision-making under uncertainty problem decouples training of the machine learning (ML) model from decision-making, which involves solving an optimization problem. As a result, the solver cannot provide feedback to the machine learning model during training, leading to a mismatch between the learning objective and the quality of the decisions. To address the mismatch, this thesis investigates the use of decision-focused learning (DFL) for receding horizon control of a grid-tied wastewater treatment plant. We optimize two separate goals: minimizing the plant's operating costs or carbon emissions. We first show that the structure of the underlying optimization problem depends on the chosen objective. Under some assumptions, we derive and prove a relaxation of our combinatorial optimization problem into a convex one. We then benchmark 7 state-of-the-art decision-focused learning methods designed for different types of optimization problems. We also benchmark 7 time-series forecasting models and select 2 for our experiments. Our results show that decision-focused learning can reduce operating costs by 3.8% and 2.5% with deterministic and stochastic decision-focused learning algorithms, respectively, compared to the traditional two-stage predict-then-optimize approach, underscoring the value of training forecasting models with task-specific loss functions rather than accuracy metrics such as mean squared error.
🔗 Provenance — このレコードを発見したソース
- openalex https://doi.org/10.7939/85931first seen 2026-11-19 05:01:16
🔔 こうした論文の新着を逃したくない方は キーワードアラート に登録(無料・3キーワードまで)。
gxceed は公開メタデータに基づく研究支援データセットです。要約・翻訳・解説は AI 支援で生成されています。 最終的な解釈・検証は利用者が原典資料に基づいて行うことを前提とします。