Prediction of Water Quantity and Quality Variables Using Deep Learning Algorithms Adopted in Time Series Domain

Abstract

We developed sequential deep learning models to predict time-series water quantity and quality data. The data set consisting of 10 input variables was prepared from a watershed model Hydrological Simulation Program - FORTRAN (HSPF) fine-tuned to existing conditions in the Nam River Basin on a daily time step for the simulation period 2019-2021. The whole data set was partitioned into training and test sets in the ratio 7 to 3. The predictive accuracy of four deep learning models, created differently in terms of the types of algorithms as well as the number of layers, was tested with respect to mean squared error (MSE). We found that changes in input time steps from 1 to 2 days led to a sharp reduction in prediction errors of all applied models for two target variables during the training and test phases, except for a few cases. In addition, at least 3 important variables were enough to maintain the predictive accuracy of the original deep learning model with 10 variables. The performance of the deep learning model was sensitive to output time steps rather than input time steps. In all test conditions, the MSE values were extremely low, reaching as high as 0.0056. Therefore, sequential deep learning models, regardless of their types and architectures, are most suitable for predictive modeling of time series data compiled on a daily basis or less such as remote sensing data in hydrology and agriculture.