Prediction of Water Storage Rate for Agricultural Reservoirs Using Recurrent Neural Network and Transformer Time Series Models
Year: 2024
Type: article
Abstract: In Korea, there are 17,100 agricultural reservoirs that play the role of agricultural water supply and flood prevention, but more than 13,700 of them are not equipped with hydrological measurement facilities, making it difficult to operate properly.On the other hand, research on detecting reservoir water volume using satellite imagery and converting it into water yield is actively underway, and if the water yield prediction model can predict the water yield of agricultural reservoirs with high accuracy, it can enable effective operation of measured and unmeasured reservoirs.In this study, we compared and analyzed the ability of recurrent neural network structure and transformer structure time series models to predict water yield for the top 48 agricultural reservoirs by water volume.The 11 years of water yield data from 2013 to 2023 were constructed, and the 2013-2021 data were separated into training data, 2022 data were used as validation data, and 2023 data were used as test data, and the mean absolute error (MAE), root mean square error (RMSE), and correlation coefficient evaluation metrics were used to quantitatively evaluate the water yield prediction ability.Instead of univariate time series forecasting using only historical water storage rate as an input variable, multivariate time series forecasting reflecting meteorological and seasonal effects as input variables was used, The RMSE averages of 48 reservoirs for 20-day forecasts of temporal fusion transformer (TFT), a time series model with a transformer structure, and gated recurrent unit (GRU), a time series model with a recurrent neural network structure, were 7.115% and 17.428%, respectively, showing that the TFT model has a better ability to predict water storage in the medium term than the GRU model.
Source: 대한원격탐사학회지
Authors: Jaeseong Park, Yangwon Lee
Institution: Pukyong National University
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
Open Access status: diamond
Awards: G1081881250, G2202836915
OpenAlex ID: W4406077758