Abstract: The data of the surface carbon dioxide partial pressure ［p(CO2)］ and corresponding physicochemical properties in water of a karst deep-water reservoir in Yunnan Guizhou Plateau—Wanfeng Lake Reservoir from 2016 to 2017 was collected to analyze the correlation between the CO2 flux and physicochemical properties in water. The convolutional neural network and long-short-term memory neural network hybrid model (CNN-LSTM model) was established to predict the CO2 flux in reservoir, which was based on the collected sample data. Research shows that the CO2 flux in summer in Wanfeng Lake Reservoir only has a significant correlation with pH and oxidation-redox potential (ORP), while in winter it is significantly related to water temperature (T), pH, alkalinity (ALK), total dissolved solid (TDS) and conductivity (Cond). In a complete year, all physicochemical properties in water are important factors influencing CO2 flux. Finally, 80% of the training data are used to train the CNN-LSTM model and 20% of the test data to test the absolute mean error (MAE), root mean square error (RMSE) and correlation (R2) of the model. The CNN model, LSTM model and DNN model were established to compare with the CNN-LSTM model, the R2 between the predicted value and the measured value of the four models are higher than 0.90, and the MAE and RMSE of the proposed CNN-LSTM model were 2.64 and 3.85 mmol/(m2·d), which was lower than those of the other three models. The CNN-LSTM model can perform more effectively in feature extraction and data prediction.

Key words: environmental engineering; CO2 flux; deep learning; CNN-LSTM neural network model

摘要： 整合了2016年—2017年云贵高原岩溶深水水库——万峰湖水库表层CO2分压［p(CO2)］及对应的水质指标，计算了水-气界面CO2通量并分析其与水质的线性相关性，最终在收集的样本数据下建立了水库CO2通量预测的卷积神经网络与长短时记忆神经网络混合模型(CNN-LSTM模型)。研究表明：万峰湖水库夏季的CO2通量仅与pH和氧化还原电位(ORP)有显著的相关性，而冬季的CO2通量与水温(T)、pH、碱度(ALK)、总溶解固体物质浓度(TDS)和电导率(Cond)均有显著的相关性，在一个完整的水文年内，6个水质指标均为CO2通量的重要影响因素。使用80%训练集数据训练CNN-LSTM模型，20%测试集数据测试模型的绝对均值误差(MAE)，均方根误差(RMSE)和相关性(R2)，并且建立CNN神经网络模型、LSTM神经网络模型和全连接神经网络模型(DNN)与之对比。4种模型预测值与实测值的相关性(R2)均高于0.90， CNN-LSTM模型的MAE与RMSE分别为2.64、3.85 mmol/(m2·d)，均低于另外3种神经网络模型，CNN-LSTM模型能在样本数量较小的情况下取得最好的CO2通量预测效果。

关键词: 环境工程；CO2通量；深度学习；CNN-LSTM神经网络模型