|
| 引用本文: | 杨玲,杨慧霞,王云云,姚云波,王顺.基于机器学习的总溶解气体预测模型构建与评估.湖泊科学,2025,37(2):508-516. DOI:10.18307/2025.0225 |
| Yang Ling,Yang Huixia,Wang Yunyun,Yao Yunbo,Wang Shun.Construction and evaluation of total dissolved gas prediction model based on machine learning. J. Lake Sci.2025,37(2):508-516. DOI:10.18307/2025.0225 |
|
| |
|
|
| 本文已被：浏览 805次 下载 273次 |
码上扫一扫！ |
|
|
| 基于机器学习的总溶解气体预测模型构建与评估 |
|
杨玲，杨慧霞，王云云，姚云波，王顺
|
|
贵州大学土木工程学院，贵阳 550025
|
|
| 摘要: |
| 随着梯级水库的持续开发，大坝泄流产生的过饱和总溶解气体在河道内难以消散，这将导致鱼类患气泡病甚至死亡，因此开发大坝下游总溶解气体预测模型对保护生物多样性具有重要意义。本文收集美国哥伦比亚河上3个大坝监测站点的日监测数据（水温、气压、流量、大坝溢流以及总溶解气体饱和度），在此基础上利用BP神经网络、随机森林以及提升树3种机器学习算法预测总溶解气体饱和度，并对3种算法的预测性能进行评价和对比。研究发现，随着具有显著性相关性的输入变量个数增加，各模型预测性能呈上升趋势，且不同模型受输入因子的影响不同。在最佳输入变量方案下，随机森林（平均绝对误差(MAE) =1.273%,均方根误差(RMSE)=1.775%, R2=0.952）和提升树（MAE=1.268%, RMSE=1.622%, R2=0.960）的预测性能最佳。在模型验证阶段，提升树模型可以将预测值与实测值的平均相对误差控制在2.4%以内。本研究所构建的模型能够快速准确地预测大坝泄流期间下游河道内总溶解气体饱和度，有助于提前评估过饱和风险，及时调整排放调度方案，对于局部鱼类保护区提前采取防护措施，从而减少总溶解气体对水生生态的影响。研究结果可为深入开展基于机器学习的总溶解气体预测模型提供一定的参考价值。 |
| 关键词: 总溶解气体过饱和 机器学习 预测 梯级水库 |
| DOI：10.18307/2025.0225 |
| 分类号: |
| 基金项目:贵州省科技计划项目（QKHJ-[2019]1117）资助 |
|
| Construction and evaluation of total dissolved gas prediction model based on machine learning |
|
Yang Ling，Yang Huixia，Wang Yunyun，Yao Yunbo，Wang Shun
|
|
College of Civil Engineering，Guizhou University，Guiyang 550025 ，P.R.China
|
| Abstract: |
| The ongoing advancement of cascade reservoirs has resulted in the formation of supersaturated total dissolved gases, which are challenging to disperse in riverine environments. This phenomenon has the potential to precipitate the onset of gas bubble disease and, in extreme cases, may even result in fish mortality. Therefore, the development of a predictive model for total dissolved gases downstream of dams is important for biodiversity conservation. This paper collected data from three dam monitoring stations on the Columbia River in the United States, comprising measurements of water temperature, barometric pressure, flow, dam overflow, and total dissolved gas saturation. These data were used to train three machine learning algorithms, namely, BP neural networks, random forests, and boosting trees, which were then employed to predict total dissolved gas saturation. The performance of the three algorithms was evaluated and compared. It found that as the number of significantly correlated input variables increases, the predictive performance of each model showed an upward trend, and different models were affected differently by input factors. Random forest (MAE=1.273%, RMSE=1.775%, R2=0.952) and boosting tree (MAE=1.268%, RMSE=1.622%, R2=0.960) had the best prediction performance under the optimal input variable scheme. In the model validation phase, boosting tree and random forest showed higher accuracy, with average relative errors of 2.3% and 2.6% between their predicted and measured values. In the model validation phase, the boosting tree model can control the average relative error between the predicted and measured values within 2.4%. The model constructed in this study can rapidly and accurately predict total dissolved gas (TDG) saturation in the downstream channel during dam releases. This enables the risk of TDG saturation to be assessed in advance, thereby facilitating the timely adjustment of the discharge scheduling scheme. Protective measures are taken in advance for localized fish sanctuaries, thus reducing the impact of TDG on aquatic ecology. The results of the study can provide some reference value for in-depth machine learning-based total dissolved gas prediction modeling. |
| Key words: Total dissolved gas supersaturation machine learning prediction cascade reservoirs |
|
|
|
|
