Scholarlink

<p>Drought is a slowly evolving climatic phenomenon that significantly affects environmental conditions and human activities, particularly agriculture and water resource management. Accurate predicting drought conditions is crucial to mitigate its impacts through timely interventions and policy formulation. In recent years, artificial intelligence, especially deep learning, has been increasingly applied to drought prediction due to its ability to model complex, nonlinear temporal patterns. However, many approaches rely on predefined architectures and hyperparameters, limiting adaptability and performance. This study proposes DeepGA-LSTM, a neuroevolution-based method that uses genetic algorithms to optimize the architecture and hyperparameters of Long Short-Term Memory (LSTM) networks. The goal is to assess its effectiveness in forecasting drought conditions using the Standardized Precipitation-Evapotranspiration Index (SPEI) and the Standardized Precipitation Index (SPI) in two Mexican regions: Chihuahua and Zacatecas. In Chihuahua, a one-step forward forecasting scheme was applied. The DeepGA-LSTM achieved RMSE values of 0.0644 and 0.0355, MAE of 0.0470 and 0.0260, <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\varvec{R}^{\varvec{2}}\)</EquationSource> </InlineEquation> of 0.8833 and 0.9414, and <InlineEquation ID="IEq2"> <EquationSource Format="TEX">\(\varvec{r}\)</EquationSource> </InlineEquation> of 0.9535 and 0.9710 for the SPEI-12 and SPEI-24 indices, respectively. In Zacatecas, a 24-month forecasting horizon was implemented across four subregions using an encoder-decoder design. The model’s performance ranged from 0.2951 to 0.4304 in RMSE, 0.2260 to 0.3321 in MAE, 0.1187 to 0.6294 in <InlineEquation ID="IEq3"> <EquationSource Format="TEX">\(\varvec{R}^{\varvec{2}}\)</EquationSource> </InlineEquation>, and 0.3731 to 0.8053 in <InlineEquation ID="IEq4"> <EquationSource Format="TEX">\(\varvec{r}\)</EquationSource> </InlineEquation>. Compared with LSTM and Convolutional Neural Network–LSTM (CNN-LSTM) baselines, DeepGA-LSTM consistently outperformed them in both regions, demonstrating its potential as an effective tool for drought forecasting.</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

Optimization of LSTM networks through neuroevolution for drought forecasting in Mexico

  • 期刊论文

  • 发表日期：

    2025年10月9日

查看全文
  • Ramiro Villegas-Vega,
  • Aldo Márquez-Grajales,
  • Efrén Mezura-Montes,
  • Fernando Salas-Martínez,
  • Manuel Alejandro Ojeda-Misses,
  • Claudia Romo-Gómez

摘要

Drought is a slowly evolving climatic phenomenon that significantly affects environmental conditions and human activities, particularly agriculture and water resource management. Accurate predicting drought conditions is crucial to mitigate its impacts through timely interventions and policy formulation. In recent years, artificial intelligence, especially deep learning, has been increasingly applied to drought prediction due to its ability to model complex, nonlinear temporal patterns. However, many approaches rely on predefined architectures and hyperparameters, limiting adaptability and performance. This study proposes DeepGA-LSTM, a neuroevolution-based method that uses genetic algorithms to optimize the architecture and hyperparameters of Long Short-Term Memory (LSTM) networks. The goal is to assess its effectiveness in forecasting drought conditions using the Standardized Precipitation-Evapotranspiration Index (SPEI) and the Standardized Precipitation Index (SPI) in two Mexican regions: Chihuahua and Zacatecas. In Chihuahua, a one-step forward forecasting scheme was applied. The DeepGA-LSTM achieved RMSE values of 0.0644 and 0.0355, MAE of 0.0470 and 0.0260, \(\varvec{R}^{\varvec{2}}\) of 0.8833 and 0.9414, and \(\varvec{r}\) of 0.9535 and 0.9710 for the SPEI-12 and SPEI-24 indices, respectively. In Zacatecas, a 24-month forecasting horizon was implemented across four subregions using an encoder-decoder design. The model’s performance ranged from 0.2951 to 0.4304 in RMSE, 0.2260 to 0.3321 in MAE, 0.1187 to 0.6294 in \(\varvec{R}^{\varvec{2}}\) , and 0.3731 to 0.8053 in \(\varvec{r}\) . Compared with LSTM and Convolutional Neural Network–LSTM (CNN-LSTM) baselines, DeepGA-LSTM consistently outperformed them in both regions, demonstrating its potential as an effective tool for drought forecasting.