| 摘要: | 本計畫以佶鼎科技股份有限公司廠址作為模場試驗場域，驗證與優化土壤及地下水污染整治技術。調查結果顯示，地下水中銅(Cu)與鎳(Ni)污染分布面積分別為507 及 811.2 平方公尺，另有多口監測井檢出鉛(Pb)濃度超標，顯示場址具多重重金屬污染特性。主要採用抽出處理法(Pump and Treat)控制污染，雖整體濃度呈下降趨勢，仍有局部監測井超標。為提升整治效能，輔以現地土壤淋洗法(In-situ SoilFlushing)，促進土壤中金屬脫附並抽取處理，並已完成模場試驗。場址含水層由多層砂質與黏土交互組成，具高度異質性， 7~8 公尺處之黏土層形成滯留屏障，使污染物分佈不均，抽出與灌注效率受限。此外，地下水流向受降雨與鄰近抽水行為影響，流場變異大，增加整治操作與成效評估之挑戰。因此，本模場計畫規劃共兩年研究期程，第一年提出技術需求與獲得研究成果，包括： (1)即時地下水位與基礎水質觀測以強化整治策略規劃； (2)評估含水層水力傳導係數空間分布以增進抽注作業效率； (3)運用光纖高解析監測技術評估垂直向導水特性及水流通量。為強化地質分層與材料分布之空間解析，提高藥劑注入與抽出過程效率。本年度將以第一年之觀 測資料與試驗成果為基礎，導入長短期記憶網路(Long Short-Term Memory, LSTM)結合退火演算法(Simulated Annealing, SA)之深度學習架構，進行污染物濃度時空變 化的預測與溯源分析。LSTM 模型能有效捕捉地下水污染傳輸中具非線性與時序依賴的特性，藉由大量歷史水位、水質與抽灌操作資料之訓練，預測未來污染濃度與趨勢；退火演算法則用於模型參數與初始權重之全域優化，以避免深度學習陷入局部極值，提升預測穩定度與泛化能力。此組合技術可同時進行「污染來源區反演」與「整治策略模擬」，透過比對模擬結果與監測資料差異，不斷迭代修正模型參數，最終建立污染源時空分布的最佳化反演結果。整體技術具備自動化訓練流程與高效參數搜尋能力，可有效提升多重污染(Cu、 Ni、 Pb)傳輸之預測精度，協助識別主要污染來源、評估整治成效，並提供後續決策支援。 ;This project designates the Gi Ding Technology Co., Ltd. industrial site as a pilot-scale test field to validate and optimize soil and groundwater contamination remediation technologies. Site investigations indicate that copper (Cu) and nickel (Ni) contamination in groundwater cover areas of approximately 507 m² and 811.2 m², respectively. In addition, multiple monitoring wells exhibit lead (Pb) concentrations exceeding regulatory standards, demonstrating the presence of multi-metal contamination at the site. Pump-and-treat remediation has been implemented as the primary control measure, and although overall contaminant concentrations show a declining trend, exceedances persist at several localized monitoring points. To enhance remediation efficiency, in-situ soil flushing has been employed as a supplementary technique to promote desorption of metals from the soil matrix followed by extraction and treatment of contaminated groundwater. Pilot-scale testing of this approach has been completed. The site aquifer system consists of interbedded sandy and clayey layers with pronounced heterogeneity. A clay layer at depths of approximately 7–8 m acts as a semi-confining barrier, resulting in uneven contaminant distribution and limiting the effectiveness of extraction and injection operations. Furthermore, groundwater flow directions are strongly influenced by precipitation events and nearby pumping activities, leading to highly variable flow fields and increased uncertainty in remediation operation and performance evaluation. Accordingly, the pilot study is structured over a two-year research period. In the first year, technical needs were identified and key outcomes were achieved, including: (1) implementation of real-time groundwater level and baseline water quality monitoring to strengthen remediation strategy development; (2)assessment of the spatial variability of aquifer hydraulic conductivity to improve the efficiency of pumping and injection operations; and (3) application of high-resolution fiber-optic sensing techniques to evaluate vertical hydraulic connectivity and groundwater fluxes. These efforts aim to refine the spatial resolution of stratigraphic layering and material distribution, thereby enhancing the effectiveness of chemical injection and extraction processes. Building upon the observational data and experimental results obtained in the first year, the second-year research will integrate a deep learning framework combining Long Short-Term Memory (LSTM) networks with a Simulated Annealing (SA) algorithm to predict the spatiotemporal evolution of contaminant concentrations and to conduct source identification analyses. LSTM models are well suited to capture the nonlinear behavior and temporal dependencies inherent in groundwater contaminant transport. Trained on extensive historical datasets of ground water levels, water quality measurements, and pumping–injection operations, the model will forecast future contaminant concentrations and trends. The simulated annealing algorithm will be employed for global optimization of model parameters and initial network weights, mitigating the risk of convergence to local minima and improving prediction robustness and generalization performance. This integrated approach enables simultaneous “contaminant source inversion” and “remediation strategy simulation.” By iteratively minimizing discrepancies between simulated outputs and observed monitoring data, model parameters are progressively refined to derive an optimized reconstruction of the spatiotemporal distribution of contamination sources. Overall, the proposed methodology features an automated training workflow and efficient parameter search capability, substantially enhancing predictive accuracy for multi-metal contaminant transport (Cu, Ni, Pb), supporting identification of primary contamination sources, evaluation of remediation effectiveness, and provision of robust decision support for subsequent site management. |
