Abstract: Previous watershed hydrological and pollutant migration and transformation models are greatly affected in simulation effectiveness by their initial conditions, boundary conditions, numerical resolution, and parameter sensitivity; and the existing deep learning models lack consideration of physical mechanisms in analysis of pollutant flux time series data. Aimed at these two problems, this paper presents a watershed pollutant flux prediction model based on a Long Short-Term Memory (LSTM) neural network. First, we construct a multivariable time series prediction model with the aid of the deep learning framework Keras. To determine its input parameters, we use the meteorological data of the watershed as the driving factor of its pollutant production and collection, and select its previous rainfall as the index of its soil dry-wet degree; then, using these indexes, we analyze the differences in pollutant flux under different rainfall intensities in different months and hydrological periods. Hence, the features–meteorology, soil dry-wet degree, and the level, month and hydrological period of the rainfall–can be added to the model input list. A LSTM-based time simulator is used to identify the inherent characteristics of historical data and the complex nonlinear relationship between the input features. The prediction performance of this LSTM pollutant flux model is evaluated by comparing the simulated watershed daily pollutant fluxes with measurements in-situ and using a comparative analysis with the distributed hydrological and pollutant migration and transformation process model (SWAT model) of the watershed. And the contribution rates of different input features are analyzed. Thus, we have verified the feasibility of the LSTM model and demonstrated a new idea for better predictions of watershed pollutant fluxes.

Key words: watershed pollutant flux prediction, LSTM, time series, non-point source pollution