Deep insight into daily runoff forecasting based on a CNN-LSTM model

Rainfall-runoff forecasting is expected to play a crucial role in hydrology. In recent years, machine learning models have been found to be effective in runoff simulation, and convolutional neural network (CNN) and long short-term memory (LSTM) in particular have been applied widely in hydrology. However, there are few studies investigating the applicability of the combination of CNN and LSTM (CNN-LSTM) to runoff simulation and the influence of its input parameters on the prediction performance of the model. This paper thus proposes a daily runoff forecasting model based on a CNN-LSTM model and investigates the influence of various input parameters, including the characteristics of input variables, input time step, dataset size, and lead time. The proposed model is then applied in the Feilaixia catchment. Results show that the CNN-LSTM model for runoff forecasting outperforms the LSTM model. Sensitivity analyses suggest that the settings of four input parameters have a strong influence on the prediction performance, and the degree of influence of each parameter differs. The model with runoff and rainfall data inputs yielded the best performance compared to models with other input variables. Increasing excessive input time step will lead to performance degradation and overfitting problem. As for the dataset size, both the length and the stationarity of the time series should be taken into consideration. Current case is 32-year dataset with a segmentation ratio of 0.85:0.15. Lead time is a critical factor in runoff prediction and over 3-day-ahead predictions are of low accuracy. Some discussions are also depicted to translate the recommended values into something interpretable in hydrology. This study enhances the understanding of linkage between hydrological mechanisms and runoff forecasting based on deep learning.