Journal of Geo-information Science >
A Physics-Constrained GRU Neural Network for River Water Quality Prediction
Received date: 2022-05-21
Revised date: 2022-07-09
Online published: 2023-03-25
Supported by
National Key Research and Development Program of China(2021YFB3900903)
The Natural Science Foundation of Hunan Province, China(2021JJ20058)
The Natural Science Foundation of Hunan Province, China(2020JJ4695)
Water conservancy science and technology project of Guizhou, China(KT202110)
River water quality prediction plays a key role in water environment management and water pollution prevention. In recent years, artificial neural networks have been widely used in river water quality prediction. However, the lack of physical mechanism becomes a major limitation of artificial neural networks. As a result, the prediction results obtained by artificial neural networks are usually difficult to interpret and are unstable. To overcome this limitation, this study proposed a physics-constrained Gated Recurrent Unit model (PHY_GRU) for predicting dissolved oxygen in rivers. Specifically, two kinds of physical rules were identified, i.e., the rule of monotonicity and the rule of gradation. These physical rules were modeled as the loss functions of Gated Recurrent Unit (GRU). The water quality dataset in Atlanta, USA was used to evaluate the performance of the PHY_GRU. The experimental results show that: ① Compared with autoregressive integrated moving average, multilayer perceptron, and GRU model, the prediction accuracy and stability of PHY_GRU were the highest ( the root mean square error of prediction was reduced by 94.8%, 62.9% and 37.2%, respectively); ② the PHY_GRU performed best when different physical rules were considered simultaneously; ③ When PHY_GRU used 30% of the training samples of GRU, it could outperform GRU in terms of prediction accuracy and stability. The proposed PHY_GRU effectively incorporates physical mechanism into artificial neural networks and may be helpful for water management.
LI Yuzhao , LIU Qiliang , DENG Min , XU Rui , WANG Maoyang , YANG Hainan . A Physics-Constrained GRU Neural Network for River Water Quality Prediction[J]. Journal of Geo-information Science, 2023 , 25(1) : 102 -114 . DOI: 10.12082/dqxxkx.2023.220331
表1 不同模型溶解氧含量预测结果对比Tab. 1 Comparison of the prediction results of dissolved oxygen concentration obtained by different models |
预测模型 | RMSE | MAE | MAPE/% |
---|---|---|---|
ARIMA | 1.156 | 0.962 | 10.010 |
MLP | 0.159 | 0.115 | 1.207 |
GRU | 0.094 | 0.073 | 0.762 |
PHY_GRU | 0.059 | 0.037 | 0.387 |
图6 4个站点上不同预测模型溶解氧含量预测值与实测值的对比Fig. 6 Fitting comparison of the predicted values of dissolved oxygen concentration obtained by different models at four randomly selected sites |
[1] |
|
[2] |
王紫薇, 蔡红艳, 陈慕琳. 基于大数据的化工企业时空变化特征及其对水域的潜在影响分析方法[J]. 地球信息科学学报, 2022, 24(4):673-683.
[
|
[3] |
|
[4] |
|
[5] |
吴在栋, 林广发, 张明锋, 等. 突发河流污染事件应急资源调度动态规划模型研究[J]. 地球信息科学学报, 2018, 20(6):799-806.
[
|
[6] |
樊敏, 顾兆林. 非机理性水质模型研究综述[J]. 环境科学与管理, 2009, 34(9):63-67.
[
|
[7] |
|
[8] |
陈昌仁, 吕玲玲, 王纪杰, 等. 基于S-P概化模型的滁河干流六合段纳污能力分析[J]. 水生态学杂志, 2011, 32(2):82-85.
[
|
[9] |
|
[10] |
|
[11] |
|
[12] |
Tiyasha, Tung T M, Yaseen Z M. A survey on river water quality modelling using artificial intelligence models: 2000-2020[J]. Journal of Hydrology, 2020, 585:124670. DOI:10.1016/j.jhydrol.2020.124670
|
[13] |
|
[14] |
|
[15] |
|
[16] |
|
[17] |
|
[18] |
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
[30] |
|
[31] |
|
[32] |
|
[33] |
|
[34] |
|
[35] |
|
[36] |
|
[37] |
|
/
〈 |
|
〉 |