Predicting Total Dissolved Gas Concentration on a Daily Scale Using Kriging Interpolation, Response Surface Method and Artificial Neural Network: Case Study of Columbia River Basin Dams, USA

摘要

Total dissolved gas (TDG) is an important factor for aquatic life and can cause gas bubble trauma in fish if the concentration is higher than 110%. Dissolved gas is entrained in the water over the spillways of dams. Generally, total dissolved gas is simulated and predicted using models based on fluid mechanics, hydrodynamics and mass exchange processes. In the present study, two novel data-driven techniques, namely kriging interpolation method (KIM) and response surface method (RSM), were proposed for predicting total dissolved gas, measured on a daily scale at the upstream and downstream of spillways at four different dams' reservoir sites located in Columbia River, USA. For developing models, we selected several input variables, namely water temperature, barometric pressure, spill from dam and discharge; in addition, total dissolved gas measured as percent of saturation (%) was selected as the predicted variable. Results obtained from the newly proposed models were compared with those obtained with the standard feedforward neural networks (FFNN) model to assess their performances. The proposed models were developed and compared with each other based on several input combinations. Four statistical indexes were utilized to evaluate models' performances: coefficient of correlation (R), Nash-Sutcliffe efficiency (NSE), root-mean-squared error (RMSE) and mean absolute error (MAE). The results obtained clearly show that: (1) the KIM model is better than the RSM and FFNN models at three dams and FFNN is the best for the fourth; (2) the RSM model is ranked in the third place and provided the lowest accuracy; and (3) the highest R and NSE in addition to the lowest RMSE and MAE are obtained when the models include all the four input variables. The R, NSE, RMSE and MAE of the best KIM model among the four dam's reservoirs are 0.973, 0.941, 1.462 and 1.122 while the corresponding values of the best FFNN (RSM) model are 0.962 (0.952), 0.926 (0.906), 1.643 (1.848) and 1.297 (1.426), respectively.