Behavior of Internal Variables of Long and Short-Term Memory Neural Network for Rainfall-Runoff Modeling

Abstract:

Long Short-Term Memory (LSTM) network, which is a kind of deep neural network, has been utilized for rainfall-runoff modeling in some studies. However, the behavior of the internal variables of LSTM for rainfall-runoff modeling has not been investigated well. In this study, we implemented a rainfall-runoff model using LSTM at a snow-dominated watershed to investigate the behavior of its internal variables. Daily precipitation and air temperature date were used as input to model, and daily flow discharge data were utilized as the targeted data. Then, the model parameters (the weights and biases of the LSTM) were calibrated and validated with the input and target datasets. After the calibration and validation, artificial values were given to the model instead of the precipitation and air temperature data in order to investigate the response of the internal variables to the input data in detail. The results indicated that the behavior of the internal variables does not represent a physical process.