Use of one-dimensional CNN for input data size reduction in LSTM for improved computational efficiency and accuracy in hourly rainfall-runoff modeling.

A deep learning architecture, denoted as CNNsLSTM, is proposed for hourly rainfall-runoff modeling in this study. The architecture involves a serial coupling of the one-dimensional convolutional neural network (1D-CNN) and the long short-term memory (LSTM) network. In the proposed framework, multiple layers of the CNN component process long-term hourly meteorological time series data, while the LSTM component handles short-term meteorological time series data and utilizes the extracted features from the 1D-CNN. In order to demonstrate the effectiveness of the proposed approach, it was implemented for hourly rainfall-runoff modeling in the Ishikari River watershed, Japan. A meteorological dataset, including precipitation, air temperature, evapotranspiration, longwave radiation, and shortwave radiation, was utilized as input. The results of the proposed approach (CNNsLSTM) were compared with those of previously proposed deep learning approaches used in hydrologic modeling, such as 1D-CNN, LSTM with only hourly inputs (LSTMwHour), a parallel architecture of 1D-CNN and LSTM (CNNpLSTM), and the LSTM architecture, which uses both daily and hourly input data (LSTMwDpH). Meteorological and runoff datasets were separated into training, validation, and test periods to train the deep learning model without overfitting, and evaluate the model with an independent dataset. The proposed approach clearly improved estimation accuracy compared to previously utilized deep learning approaches in rainfall = runoff modeling. In comparison with the observed flows, the median values of the Nash-Sutcliffe efficiency for the test period were 0.455-0.469 for 1D-CNN, 0.639-0.656 for CNNpLSTM, 0.745 for LSTMwHour, 0.831 for LSTMwDpH, and 0.865-0.873 for the proposed CNNsLSTM. Furthermore, the proposed CNNsLSTM reduced the median root mean square error (RMSE) of 1D-CNN by 50.2%-51.4%, CNNpLSTM by 37.4%-40.8%, LSTMwHour by 27.3%-29.5%, and LSTMwDpH by 10.6%-13.4%. Particularly, the proposed CNNsLSTM improved the estimations for high flows (≧75th percentile) and peak flows (≧95th percentile). The computational speed of LSTMwDpH is the fastest among the five architectures. Although the computation speed of CNNsLSTM is slower than LSTMwDpH's, it is still 6.9-7.9 times faster than that of LSTMwHour. Therefore, the proposed CNNsLSTM would be an effective approach for flood management and hydraulic structure design, mainly under climate change conditions that require estimating hourly river flows using meteorological datasets.

Capabilities of deep learning models on learning physical relationships: Case of rainfall-runoff modeling with LSTM.

This study investigates the relationships which deep learning methods can identify between the input and output data. As a case study, rainfall-runoff modeling in a snow-dominated watershed by means of a long short-term memory (LSTM) network is selected. Daily precipitation and mean air temperature were used as model input to estimate daily flow discharge. After model training and verification, two experimental simulations were conducted with hypothetical inputs instead of observed meteorological data to clarify the response of the trained model to the inputs. The first numerical experiment showed that even without input precipitation, the trained model generated flow discharge, particularly winter low flow and high flow during the snow melting period. The effects of warmer and colder conditions on the flow discharge were also replicated by the trained model without precipitation. Additionally, the model reflected only 17-39% of the total precipitation mass during the snow accumulation period in the total annual flow discharge, revealing a strong lack of water mass conservation. The results of this study indicated that a deep learning method may not properly learn the explicit physical relationships between input and target variables, although they are still capable of maintaining strong goodness-of-fit results.

Hourly-scale coastal sea level modeling in a changing climate using long short-term memory neural network.

In this study, a coastal sea level estimation model was developed at an hourly temporal scale using the long short-term memory (LSTM) network, which is a type of recurrent neural networks. The model incorporates the effects of various phenomena on the coastal sea level such as the gravitational attractions of the sun and the moon, seasonality, storm surges, and changing climate. The relative positions of the moon and the sun from the target location at each hour were utilized to reflect the gravitational attractions of the sun and the moon in the model simulation. The wind speed and direction, mean sea level pressure (MSLP), and air temperature near the target point at each hour were used to consider the effects of storm surges and seasonality of the coastal sea level. In addition to the hourly local variables, the annual global mean air temperature was considered as input to the model to reflect the effect of global warming on the coastal sea level. The model was implemented using several input lengths of the annual global mean air temperature to estimate the coastal sea level at the Osaka gauging station in Japan. Several statistics such as the mean, the Nash-Sutcliffe efficiency, and the root mean square error were used to evaluate model performance. The results show that the proposed model accurately reconstructed the effects of the gravitational attractions of the sun and the moon on the coastal sea levels. The model also considered the effects of fluctuations in the wind speed and MSLP although the coastal sea levels during were underestimated strong winds and low MSLP conditions. Lastly, introducing a longer duration annual global mean air temperature improved model accuracy. Consequently, the best results show 0.720 of the NSE value for the test process.