LSTM於屏東平原地層下陷區地下水位預測模型研究

MODELLING GROUNDWATER LEVEL FOR LAND SUBSIDENCE AREA IN PINGTUNG PLAIN BY LSTM

蔡存孝(Tswn-Syau Tsay) ； 謝勝信(Sheng-Hsin Hsieh)

《農業工程學報》 66卷3期 (2020/09) Pp. 42-56

https://doi.org/10.29974/JTAE.202009_66(3).0005

摘要

本文以資料驅動角度，建立屏東沿海地層下陷嚴重區域，地下水長短期記憶（LSTM）模型預測地下水水位，本文依據屏東沿海林邊溪南北共11個觀測井第2、3、4層含水層日地下水位資料，分別以3個腳本建立LSTM模型。資料經過正規化後分割為訓練資料、驗證資料及測試資料，訓練過程以Early Stopping避免過度擬合的問題，研究發現LSTM模型對於測試資料MSE值發現，腳本3較腳本1及腳本2更能精確預測結果。可能因為林邊溪以南的地下水比林邊溪以北的地下水受到更多不可預知的干擾，大庄（2）、枋寮（2）、德興（2）需要再加入影響地下水位的其他特徵或再建構更優化的LSTM模型，才能降低測試資料MSE值，增加模型的精確度。大潭（2）、崎峰（2）、崎峰（3）、崎峰（4）、新埤（2）、新埤（3）、新埤（4）及德興（3）而言，無論使用單層LSTM模型或3層LSTM模型，30次LSTM模型訓練即可訓練出LSTM模型讓測試資料MSE值在0.001以下，代表實際最高與最低地下水位相差10公尺時，預測地下水位與實際的地下水位會產生32公分的均方差，LSTM模型MSE表現最差的是枋寮（2）及崎峰（4），表現最好的是大潭（2）及新埤（4），整體而言，模型預測值與實際值的相關係數均在0.98以上。

關鍵字

地下水 ； 長短期記憶法 ； 屏東平原

並列摘要

Groundwater level prediction model by Long Short-Term Memory (LSTM) was developed for Pingtung coastal area where land subsidence problems have been for many years. There are 11 LSTM models for 3 scenarios were developed based upon daily groundwater level records for 2^(nd), 3^(rd) and 4^(th) stratified aquifers of 11 monitoring wells on the north and south of Lin-Bian Creek in Pingtung, respectively. Data set of daily groundwater level was normalized and split into training data set, verification data set and testing data set. Early Stopping algorithm was applied during training process to avoid overfitting problem. It was found that LSTM model with scenario 3 obtained more accurate predictions than the other scenarios. Furthermore, it is probably the groundwater on the south of Lin-Bian Creek received unpredicted disturb, other features were required for the 3 models of this area to increase model accuracy. For models on the north of Lin- Bian Creek, 30 times of Early Stopping model training were sufficient for obtaining LSTM model's testing MSE less than 0.001 either by 1-LSTM-Layer model or 3-LSTM-Layer model. As the groundwater level was normalized, MSE less than 0.001 indicated a root mean square error of 32 cm for the difference of maximum and minimum groundwater level of 10 meters. The worst models by LSTM are FangLiao2 and QiFeng4. The best models are DaTang2 and XinPi4 which perform very good prediction ability. Overall, correlation coefficients of predicted and actual groundwater levels are all above 0.98.

並列關鍵字

Groundwater ； LSTM ； Pingtung Plain