跳到主要內容 :::

年度	112	專案性質	實驗性質
專案類別	模場試驗	研究主題	整治
申請機構	國立臺灣大學	申請系所	環境工程學研究所
專案主持人	駱尚廉	職等/職稱	特聘教授
專案中文名稱	以人工智慧結合遠端監控優化場址之現地整治成效
中文關鍵字	人工智慧,整治成效優化,遠端監控,機器學習
專案英文名稱	Optimization of remediation performance with artificial Intelligence and remote sensing Technology
英文關鍵字	Artificial Intelligence, machine learning, remote sensing, remediation performance optimization
執行金額	3,163,000元
執行期間	2024/11/1 至 2026/12/31
計畫中文摘要	污染場址採取現地整治工法進行整治作業時，場址各類數據的取得與分析為優化整治成效中關鍵性的因素，而這些數據的變異性將影響到場址概念模式的修正、整治系統的調整與列車型整治方案的決策，因此，本研究將以LNAPL污染類型的場址採取現地地下水抽提、土壤氣體抽提、雙相抽提、地下水注氣為主的工法者為研究對象，蒐集10-20處於同類型工法的污染場址整治成功的案例，以人工智慧中機器學習的方法，對於蒐集數據進行訓練並建立對於場址概念模式、整治成效的預測模型，進而達到從整治過程的數據決策出修定場址概念模式、整治系統操作參數調整與整治成效評估指標；其中，指標之決定將納入生命週期評估的參數。為了驗證前述訓練的模型，本研究於挑選的模場中安裝整治系統中重要參數的遠端監控設備，其中，遠端監控設備將包括井內水壓、氣壓、整治設備流量、真空度、尾氣濃度等傳感器取得設備運轉過程中的各項即時參數，並搭配必要的現場取樣比對與蒐集場址其他定期監測數據，導入建立的模型內評估模型的準確性，同時也繼續作為機器學習數據的一部分，持續驗證模型，以最終確認模型的可行性。 本研究整合場址內土壤氣體、土壤及地下水頂空濃度等現場快篩數據利用RBF方法建立場址分層污染分布模型，並對應抽提井所在網格分層污染分布作為輸入參數，加入地下水水位、整治設備操作參數對應抽除系統尾氣濃度，操作MLP、RBFN及KRR三種演算法建立預測模型，初步得到KRR的模型效能最佳，以回歸模型R2值來看，訓練集、驗證集與測試集的效能達到0.87、0.75及0.76；利用SVM分類機針對尾氣濃度是否達到尾氣處理設備採用觸媒處理(>=750ppmV)以及是否濃度偏低應採取替代方案的分類上(<150ppmV)，最佳模型準確度與AUC分別達到0.95,0.91及0.88、0.95，顯示在分類上已經具有一定預測能力；遠端監控系統已於現場安裝完成，後續將持續收集場址資料並導入邊緣計算器以利預測模型於現場應用，進而達到提高整治操作效率並降低整治成本，提高節能減碳，進而達到永續韌性的整治目標。。
計畫英文摘要	When implementing on-site remediation using in situ methods for contaminated sites, the acquisition and analysis of various types of data on the site are crucial factors for optimizing the effectiveness of the remediation. The variability of these data will affect the adjustment of the site conceptual model, the tuning of the remediation system, and the decision-making for the train-type remediation scheme. Therefore, this study will use machine learning techniques to train and establish predictive models for the conceptual model and remediation performance with 10-20 successfully remediated sites, which are LNAPL (Light Non-Aqueous Phase Liquid) contaminated sites with groundwater extraction, soil gas extraction, dual phase extraction, and air sparging remediation technology implemented. Furtherly, this approach aims to achieve data-driven decision-making in the remediation process, including the revision of site conceptual models, adjustment of operating parameters for the remediation system, and evaluation of indicators for remediation effectiveness. The determination of these indicators will incorporate parameters from lifecycle assessments. To validate the trained models, the study will install remote sensing devices for important parameters of the remediation system at selected pilot sites. The remote sensing devices will include sensors for in-well water pressure, air pressure, flow rates of remediation equipment, vacuum levels, exhaust gas concentrations, and other real-time parameters during the operation of the equipment. These parameters will be compared and correlated with necessary on-site sampling and other regularly collected monitoring data to assess the accuracy of the models. Additionally, these data will continue to be part of the machine learning process to continuously validate the models, ultimately confirming their feasibility. This study integrates on-site rapid screening data, including soil gas, soil, and groundwater headspace concentrations, to establish a layered contamination distribution model of the site using the Radial Basis Function (RBF) method. The model uses the layered contamination distribution within the grid corresponding to extraction wells as input parameters, along with groundwater level, operational parameters of remediation equipment, and exhaust gas concentrations from the extraction system as the output. It employs three algorithms—MLP, RBFN, and KRR—to build predictive models, with preliminary results showing that the KRR model performs best. In terms of regression model R² values, the training, validation, and testing sets achieve scores of 0.87, 0.75, and 0.76, respectively. An SVM classifier is used to determine whether the exhaust gas concentration meets the threshold for catalytic treatment (>=750 ppmV) or whether alternative measures are required due to low concentrations (<150 ppmV). The best classification model achieves accuracy and AUC of 0.95, 0.91 for the 750ppmV classification , and 0.88, and 0.95 for 150ppmV, indicating substantial predictive capability. A remote monitoring system has been installed on-site, and data collection will continue to feed into an edge computer, enabling on-site application of the predictive model. This approach aims to enhance remediation operation efficiency, reduce remediation costs, improve energy conservation and carbon reduction, and ultimately achieve the goals of sustainable and resilient remediation.