| 摘要: | 地下水作為人類生活及經濟發展中至關重要的存在，亦是穩定的水資源來源之一，因此面臨水資源短缺時，如何妥善利用地下水，成為一個非常重要的課題。而過往文獻大多以月份作為時間尺度，且通常以該地區過往歷史水位資料作為建模過程中的唯一輸入因子，如此可見目前對於以小尺度資料分析角度出發，且利用多項因子探討並預測地下水水位的研究仍相當缺乏。
故本研究以高雄大寮地區為例，除了降雨、潮汐、溫度及濕度的歷史小時觀測資料外，更加入過往文獻中，難以估計卻深刻地影響著地下水位變動的抽水資料。以及結合小波進行深度學習模型建立：藉由小波特徵萃取方式得到各項因子在時間域下的變動特徵，以及影響地下水位的延遲時間，進而透過遞歸神經網路（recurrent neural networks，RNN）、長短期記憶模型（long short-term memory，LSTM）及門控循環單元神經網絡 (Gate Recurrent Unit, GRU)等深度學習方法，歸納並預測出多項變動因子在不同的時間延遲下對於地下水位的影響，最後以均方根誤差(root mean square error, RMSE)、決定係數(coefficient of determination,R2)等評估係數評估模型是否可信，並在最終 LSTM 模型的結果中得到 RMSE : 0.97、MAE：0.76、MSE：0.95、R2：0.5；表示本研究能以事先瞭解並預測大寮地區可能產生的不同水位變動情形，為地下水位預測提供了一種可行而準確的方法，這將可作為智慧地下水資源管理與風險評估的一個重要參考，達到地下水資源永續利用的目標。;Groundwater, as a vital existence in human life and economic development, is also one of the stable sources of water resources. Therefore, how to properly utilize groundwater becomes a very important issue when faced with water shortages. However, most of the previous literature uses monthly data as the time scale, and usually uses the historical water level data of the area as the only input factor in the modeling process without considering pumping information and rainfall. This shows that the current studies of small-scale data which is based on the use of multiple factors with hydrological mechanisms to explore and predict the groundwater level is still quite lacking.
Therefore, this study proposed a novel framework combining wavelet analysis and deep learning models called wavelet-deep learning models and taking the Daliao area of Kaohsiung as an example. From the historical hourly observation data during 2017/08/23-2020/01/30, including groundwater level, smart pumping measurement, tidal, and meteorological data. After abstracting important features of each factor with groundwater level by wavelet transform, using deep learning algorithms such as recurrent neural networks (RNN), long short-term memory (LSTM) model and gate recurrent unit (GRU) to summarize and predict the impact of multiple variable factors on the groundwater level under different time lags. The results of hourly prediction show that the performance of the LSTM model, GRU model and RNN model are both reliable in which values of the root mean square error (RMSE) were obtained 0.97, 1.04 and 1.01, respectively.
This study provides a feasible and accurate approach for groundwater level prediction by understanding and predicting different water level changes that may occur in the Daliao area in advance. As a result, the study will be an important reference for groundwater resources management and risk assessment, and achieve the goal of sustainable use of groundwater resources. |
