Scholarlink

Abstract <p>The paper presents the results of modeling average daily runoff for Russia’s small and medium rivers using the long short-term memory (LSTM) neural network architecture. Meteorological data from the ERA5-Land (temperature) and MSWEP (precipitation) datasets were used as input data. The models were trained by comparing simulation results with observational data on river runoff at 1100 gauging stations. The model was implemented in three versions: training on all time series simultaneously with the inclusion of physiographic catchment attributes obtained from the HydroATLAS dataset; training on all time series without these attributes; training the model separately for each catchment. The GR4J and HBV conceptual hydrological models were also calibrated for all the runoff observation sites. The results were assessed by the median value of the Nash–Sutcliffe model efficiency coefficient (NSE) for each station during the test observation period. The LSTM architecture incorporating physiographic attributes and trained on the entire sample demonstrated the best result: the median value of NSE for 1100 stations was equal to 0.64. For the GR4J and HBV models, the NSE values were 0.44 and 0.35, respectively. For LSTM without using physiographic attributes, NSE is 0.43, and the corresponding value for LSTM trained separately for each object is 0.28. Thus, it is demonstrated that a model can be developed based on the LSTM architecture that takes into account both hydrometeorological data and physiographic factors.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Modeling Average Daily Runoff Time Series of Russia’s Small and Medium Rivers Using Recurrent Neural Networks

  • 期刊论文

  • 发表日期：

    2026年3月1日

查看全文
  • D. V. Abramov,
  • L. S. Kurochkina,
  • V. M. Moreido

摘要

Abstract

The paper presents the results of modeling average daily runoff for Russia’s small and medium rivers using the long short-term memory (LSTM) neural network architecture. Meteorological data from the ERA5-Land (temperature) and MSWEP (precipitation) datasets were used as input data. The models were trained by comparing simulation results with observational data on river runoff at 1100 gauging stations. The model was implemented in three versions: training on all time series simultaneously with the inclusion of physiographic catchment attributes obtained from the HydroATLAS dataset; training on all time series without these attributes; training the model separately for each catchment. The GR4J and HBV conceptual hydrological models were also calibrated for all the runoff observation sites. The results were assessed by the median value of the Nash–Sutcliffe model efficiency coefficient (NSE) for each station during the test observation period. The LSTM architecture incorporating physiographic attributes and trained on the entire sample demonstrated the best result: the median value of NSE for 1100 stations was equal to 0.64. For the GR4J and HBV models, the NSE values were 0.44 and 0.35, respectively. For LSTM without using physiographic attributes, NSE is 0.43, and the corresponding value for LSTM trained separately for each object is 0.28. Thus, it is demonstrated that a model can be developed based on the LSTM architecture that takes into account both hydrometeorological data and physiographic factors.