Prediction model and application of machine learning for supersaturated total dissolved gas generation in high dam...

Supersaturation of total dissolved gas (TDG) caused by high dam discharge is an ecological risk that cannot be ignored in the operation of hydropower stations. The establishment of an efficient and concise TDG generation prediction model is of great significance to the water ecology and water environment protection of hydropower development reaches. The flow conditions and the process of water-gas mass transfer in discharge and energy dissipation are very complicated and difficult to observe in the field, bringing difficulties to the establishment of prediction model and parameter calibration. Increasingly abundant observations make it possible to establish an efficient machine learning prediction model for supersaturated TDG. In this study, extreme learning machine (ELM) and support vector regression (SVR) were used to establish the prediction model. The main influencing factors of supersaturated TDG, obtained by the analysis of the physical process of the generation of supersaturated TDG, were used as the input of the machine learning model. Then, this research took Dagangshan hydropower station and Xiluodu hydropower station as objects, and established machine learning prediction model for supersaturated TDG with several years of observation data in different discharge scenarios. Four models, including ELM, SVR, GA-ELM and GA-SVR, were obtained through genetic algorithm optimization. The relative errors of the simulation results of each model are mostly less than 5%, mean absolute error (MAE) values less than 1.6%, and root mean square error (RMSE) values less than 2.5%. The results showed that these models are highly accurate and time-saving. Based on this, TDG saturation in downstream of Dagangshan hydropower station with different discharge scenarios was simulated by machine learning model, on which the discharge optimization scheme was put forward. The proposed models, as an important supplement to the prediction of supersaturated TDG, enjoy practical significance and engineering value.