← 論文一覧に戻る

深層強化学習に基づくCO2圧入・貯留動的スキーム最適化の新手法

A New Approach for CO2 Flooding and Storage Dynamic Scheme Optimization Based on Deep Reinforcement Learning (原題)

Peng Dong, Ren-Wei Liu, Xin Lu, Bei-Chen Zhao, Fan Zhang, Lu-Lu Cai

ADIPEC2026-11-02#CCUSOrigin: CN経営インパクト: コスト削減対象セクター: power
DOI: 10.2118/236252-ms
原典: https://doi.org/10.2118/236252-ms

🤖 gxceed AI 要約

日本語

CO2圧入・貯留（CCUS）において、増油回収（EOR）と貯留効率を両立させる圧入・生産スキームを深層強化学習（DRL）で動的最適化する手法を提案。二段階の深層学習サロゲートモデルにより30年スキームの評価時間を9.17分から1.53秒に短縮し、TD3アルゴリズムを遺伝的アルゴリズムと比較。炭素税を考慮したNPVは2.67億ドルに収束し、NPV7.71%増、累積産油量11.55%増、貯留効率9.5ポイント改善を達成した。

English

This study proposes a deep reinforcement learning (DRL) method to dynamically optimize CO2 flooding and storage schemes, maximizing NPV while accounting for a carbon tax. A two-level deep-learning surrogate cuts 30-year scheme evaluation from 9.17 min to 1.53s. The TD3 agent outperforms a genetic algorithm, raising NPV by 7.71%, cumulative oil by 11.55%, and storage efficiency by 9.5 points, while increasing trapped and dissolved CO2 storage.

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

日本はCCS/CCUSをGX推進戦略の重点領域と位置づけ、炭素税・カーボンプライシング導入を議論中。炭素税を織り込んだCCUS事業の経済最適化は、国内CCS事業化検討やGX経済移行債の対象事業評価に示唆を与える。

In the global GX context

As carbon pricing expands under EU ETS and emerging Asian schemes, integrating carbon tax into CCUS project optimization speaks directly to transition finance and the bankability of carbon capture business models globally.

👥 読者別の含意

🔬研究者:DRLとサロゲートモデルを組み合わせたCCUS最適化の定量ベンチマークとして参照価値が高い。

🏢実務担当者:炭素税を織り込んだCCUS/EOR事業の投資回収・NPV評価の設計に活用できる。

🏛政策担当者:炭素税水準がCCUS事業の経済性と貯留効率に与える影響を政策設計に反映する材料となる。

📄 Abstract（原文）

As a win-win technology, CO2 flooding and storage can enhance oil recovery (EOR) while reducing greenhouse gas emissions. However, collaboratively optimizing injection-production schemes remain challenging. This study proposes a dynamic optimization method using deep reinforcement learning (DRL) to maximize net present value (NPV) while considering carbon tax. The problem is formulated as a Markov decision process. An agent observes reservoir states and annually adjusts injection-production parameters. To support the numerous evaluations required by DRL, a two-level deep-learning surrogate model is trained using reservoir numerical simulation samples, reducing the evaluation time of a 30-year scheme from 9.17 min to 1.53s. The TD3 algorithm is employed and benchmarked against a genetic algorithm (GA). With an oil price of $460/ton, CO2 cost of $90/ton, and carbon tax of $60/ton, the reward converges after 1500 episodes, and NPV stabilizes at $267 million. During the first 15 years, the injection rate increases while bottom-hole pressure decreases, indicating an emphasis on EOR. During the final 15 years, the injection rate remains stable while bottom-hole pressure rises, improving storage efficiency. Compared with GA, DRL increases NPV by 7.71%, cumulative oil production by 11.55%, and CO2 storage efficiency by 9.50 percentage points. Trapped storage increases from 17100 to 216900 tons, and dissolved storage rises from 59100 to 66500 tons, reducing free CO2 and improving storage safety. The proposed method therefore simultaneously improves economic performance, EOR, storage efficiency, and security, providing an effective approach for designing collaborative CO2 flooding and storage schemes.

🔗 Provenance — このレコードを発見したソース

🔔 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。