Recently, developing countries have steadily been pushing for the construction of stream-oriented smart cities, breaking away from the existing old-town-centered development in the past. Due to the accelerating effects of climate change along with such urbanization, it is imperative for urban rivers to establish a flood warning system that can predict the amount of high flow rates of accuracy in engineering, compared to using the existing Computational Fluid Dynamics (CFD) models for disaster prevention. In this study, in the case of streams where missing data existed or only small observations were obtained, the variation in flow rates could be predicted with only the appropriate deep learning models, using only limited time series flow data. In addition, the selected deep learning model allowed the minimum number of input learning data to be determined. In this study, the time series flow rates were predicted by applying the deep learning models to the Han River, which is a highly urbanized stream that flows through the capital of Korea, Seoul and has a large seasonal variation in the flow rate. The deep learning models used are Convolution Neural Network (CNN), Simple Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Bidirectional LSTM (Bi-LSTM) and Gated Recurrent Unit (GRU). Sequence lengths for time series runoff data were determined first to assess the accuracy and applicability of the deep learning models. By analyzing the forecast results of the outflow data of the Han River, sequence length for 14 days was appropriate in terms of the predicted accuracy of the model. In addition, the GRU model is effective for deep learning models that use time series data of the region with large fluctuations in flow rates, such as the Han River. Furthermore, through this study, it was possible to propose the minimum number of training data that could provide flood warning system with an effective flood forecasting system although the number of input data such as flow rates secured in new towns developed around rivers was insufficient.

Since predicting rapidly fluctuating water levels is very important in water resource engineering, Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) were used to evaluate water-level-prediction accuracy at Hangang Bridge Station in Han River, South Korea, where seasonal fluctuations were large and rapidly changing water levels were observed. The hydrological data input to each model were collected from the Water Resources Management Information System (WAMIS) at the Hangang Bridge Station, and the meteorological data were provided by the Seoul Observatory of the Meteorological Administration. For high-accuracy high-water-level prediction, the correlation between water level and collected hydrological and meteorological data was analyzed and input into the models to determine the priority of the data to be trained. Multivariate input data were created by combining daily flow rate (DFR), daily vapor pressure (DVP), daily dew-point temperature (DDPT), and 1-hour-max precipitation (1HP) data, which are highly correlated with the water level. It was possible to predict improved high water levels through the training of multivariate input data of LSTM and GRU. In the prediction of water-level data with rapid temporal fluctuations in the Hangang Bridge Station, the accuracy of GRU’s predicted water-level data was much better in most multivariate training than that of LSTM. When multivariate training data with a large correlation with the water level were used by the GRU, the prediction results with higher accuracy (R2=0.7480−0.8318; NSE= 0.7524−0.7965; MRPE= 0.0807−0.0895) were obtained than those of water-level prediction results by univariate training.

The design flow considering nonstationarity is estimated to determine the design flood related to hydraulic structure quantitatively based on the design process for stream restoration in the Mokgamcheon watershed proposed by Lee et al. (2011). The purpose of this research is to suggest new ways that the design flood was calculated considering nonstationarity at the Mokgamcheon watershed. Storm-unit hydrograph method to calculate design flood and direct frequency analysis were applied and nonstationarity was considered for the frequency analysis through extRemes toolkit developed at NCAR (National Center for Atmospheric Research). Although the method of direct flood frequency analysis due to dealing with flowrates directly has a more reliable than strom-unit hydrograph method, as a result, the method of direct flood frequency analysis underestimated the design flood than strom-unit hydrograph method due to the characteristics of the flow data. Therefore, the flood of storm-unit hydrograph method (100 years frequency) was determined as the design flood in the Mokgamcheon watershed.

The methods for improving the accuracy of water level prediction were proposed in this study by selecting the Gated Recurrent Unit (GRU) model, which is effective for multivariate learning at the Paldang Bridge station in Han River, South Korea, where the water level fluctuates seasonally. The hydrological data (i.e., water level and flow rate) for Paldang Bridge station were entered into the GRU model; the data were provided by the Water Resources Management Information System (WAMIS), and the meteorological data for Seoul Meteorological Observatory and Yangpyeong Meteorological Observatory were provided through the Korea Meteorological Administration. Correlation analysis was used to select the training data for hydrological and meteorological data. Important input data affecting the daily water level (DWL) were daily flow rate (DFR), daily vapor pressure (DVP), daily dew point temperature (DDPT), and 1 h max precipitation (1HP), and were used as the multivariate learning data for water level prediction. However, the DWL prediction accuracy did not improve even if the meteorological data from a single meteorological observatory far from the DWL prediction point were used as the multivariate learning data. Therefore, in this study, methods for improving the predictive accuracy of DWL through multivariate learning that effectively utilize meteorological data from each meteorological observatory were presented. First, it was a method of arithmetically averaging meteorological data for two meteorological observatories and using it as the multivariate learning data for the GRU model. Second, a method was proposed to use the meteorological data of the two meteorological observatories as multivariate learning data by weighted averaging the distances from each meteorological observatory to the water level prediction point. Therefore, in this study, improved water level prediction results were obtained even if data with some correlation between meteorological data provided by two meteorological observatories located far from the water level prediction point were used.