Prediction of dam discharge or water level in an urban drainage basin is necessary for response-time reduction and stable operation of the facilities. Recently, many time-series prediction models using deep learning have been actively studied and developed. However, most studies are limited to simple applications of deep learning models, and proposals for new algorithms that improve accuracy and usability are insufficient. Therefore, it is necessary to improve these to utilize deep learning based hydrological predictions in the actual field. In this dissertation, a revised training algorithm and pre-processing technique that reflect hydrological characteristics are proposed to increase the accuracy of the water level and discharge prediction model. In addition, an algorithm that automatically optimizes the structure and parameters of the learning model is presented to increase usability and accuracy. The optimized learning model was applied to a dam catchment and urban drainage basin, and the prediction accuracy was analyzed. The error of dam inflow prediction was reduced 17.8 %, and the error of water level of the sewer pipe was reduced 8.9 %. These results will contribute in the stable operation of dams and underground drainage facilities through the accurate inflow and rapid water level prediction.