APPROXIMATION OF TANK MODEL BY NEURAL NETWORK TO IMPROVE PREDICTION PERFORMANCE OF DAM INFLOW

Abstract

AI models, especially deep learning models, which are increasingly being used for dam inflow forecasting, are black boxes, and the question remains as to whether they can completely replace conventional models. In addition, considering the possibility that the frequency of heavy rainfall will increase in the future, it is inevitable to ask whether AI models can handle large water outflows that have not been experienced during training. In this study, a conventional tank model was used to generate pseudo data on the relationship between rainfall and dam inflow, and the neural network models were trained and their performances were evaluated. The results showed that as the training period increased, the prediction error decreased and the variability in performance tended to decrease. Prediction performance deteriorated when the outflow was more than twice the maximum dam inflow during training, and the best results were obtained when all data were used for training instead of limiting the training data by the dam inflow.