An analysis of ensemble models for the water surface evaporation simulation in the Three Gorges Reservoir

The current study aims to investigate the applicability of ensemble modeling in improving the simulation of water surface evaporation (EW) in the Three Gorges Reservoir. To achieve this objective, a sensitivity analysis is performed to determine the most influential model inputs. Various models are employed for the simulation of EW, including empirical models such as Stelling (STE), Thornthwaite Holzman (T-H), and Ryan Harleman (R-H); statistical models including multiple-linear regression (MLR), Ridge regression (Ridge), and Lasso regression (Lasso); and different Artificial Intelligence (AI) techniques such as Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and General Regression Neural Network (GRNN). The performance of these models is evaluated. Additionally, three ensemble methods, namely simple averaging, weighted averaging, and neural ensemble, are utilized in strategy 1 for empirical models, strategy 2 for statistical models, strategy 3 for AI models, and strategy 4 for multi-class mixed models, with the aim of improving the simulation performance. The results indicate that the dominant parameters are water surface temperature, water surface area, relative humidity, temperature difference of vapor, and wind speed. For the single model, empirical and statistical models can yield valuable results, while most AI models suffer from overfitting issues. Among the ensemble models, the neural ensemble method outperforms the simple averaging and weighted averaging methods. The multi-class mixed ensemble model exhibits the highest simulation accuracy, with NSE values of 0.95 and 0.86 in the training and validation phases, respectively. Compared to the best single model, the ensemble approaches proposed in this study improve the performance of single models in the validation phase by up to 11.63%, 8.21%, 6.88%, and 6.96% for strategies 1 similar to 4, respectively. Furthermore, the results demonstrate that the multi-class mixed ensemble modeling approach is preferable over empirical, statistical, and AI ensemble modeling.