Villegas-Vega et al. (2025) Optimization of LSTM networks through neuroevolution for drought forecasting in Mexico

Identification

• Journal: Theoretical and Applied Climatology
• Year: 2025
• Date: 2025-10-09
• Authors: Ramiro Villegas-Vega, Aldo Márquez-Grajales, Efrén Mezura‐Montes, Fernando Salas-Martínez, Manuel Alejandro Ojeda-Misses, Claudia Romo-Gómez
• DOI: 10.1007/s00704-025-05818-z

Research Groups

• Artificial Intelligence Research Institute, University of Veracruz
• Área Académica de Computación y Electrónica, Instituto de Ciencias Básicas e Ingeniería, Universidad Autónoma del Estado de Hidalgo
• Área Académica de Química, Instituto de Ciencias Básicas e Ingeniería, Universidad Autónoma del Estado de Hidalgo

Short Summary

This study proposes DeepGA-LSTM, a neuroevolution-based method using genetic algorithms to optimize Long Short-Term Memory (LSTM) networks for drought forecasting in Mexico. The DeepGA-LSTM consistently outperformed baseline LSTM and CNN-LSTM models in two Mexican regions (Chihuahua and Zacatecas) using SPEI and SPI indices, demonstrating its effectiveness in finding optimal network architectures.

Objective

• To evaluate the performance of neuroevolution in optimizing the structure and hyperparameters of LSTM networks for drought forecasting, specifically analyzing if the proposed DeepGA-LSTM achieves lower forecasting errors using the Standardized Precipitation-Evapotranspiration Index (SPEI) and the Standardized Precipitation Index (SPI) in two regions of Mexico (Chihuahua and Zacatecas).

Study Configuration

• Spatial Scale: Two Mexican states: Chihuahua and Zacatecas. Zacatecas is further divided into four subregions: Semi-arid, Canyons, High Plain, and Mountains.
• Temporal Scale:
  • Chihuahua: Monthly SPEI-12 and SPEI-24 data from 1901 to 2020, with a one-step forward forecasting scheme.
  • Zacatecas: Monthly SPI data from 1964 to 2020, with a 24-month forecasting horizon.

Methodology and Data

• Models used:
  • DeepGA-LSTM: Long Short-Term Memory (LSTM) networks optimized by a genetic algorithm (neuroevolution).
  • Baseline LSTM: Predefined LSTM architectures from previous studies or empirical tuning.
  • CNN-LSTM: Hybrid model combining Convolutional Neural Network and LSTM layers.
• Data sources:
  • Chihuahua: SPEI monthly data (12 and 24 months) from 1901 to 2020, sourced from the Global SPEI Database of the Consejo Superior de Investigaciones Científicas de España.
  • Zacatecas: 31 precipitation time series from 1964 to 2020, obtained from the Servicio Meteorológico Nacional (SMN) meteorological stations network, used to compute the SPI.

Main Results

• DeepGA-LSTM consistently outperformed baseline LSTM and CNN-LSTM models in both Chihuahua and Zacatecas.
• Chihuahua (SPEI, one-step forward forecasting):
  • For SPEI-12, DeepGA-LSTM achieved RMSE of 0.0644, MAE of 0.0470, R² of 0.8833, and a correlation coefficient (r) of 0.9535.
  • For SPEI-24, DeepGA-LSTM achieved RMSE of 0.0355, MAE of 0.0260, R² of 0.9414, and r of 0.9710.
  • Forecasting performance was notably higher for SPEI-24 due to lower fluctuations and greater smoothness of long-term drought time series.
• Zacatecas (SPI, 24-month forecasting horizon):
  • DeepGA-LSTM's performance across the four subregions ranged from RMSE 0.2951 to 0.4304, MAE 0.2260 to 0.3321, R² 0.1187 to 0.6294, and r 0.3731 to 0.8053.
  • The baseline LSTM model yielded negative R² values in all regions, indicating poor predictive capability compared to a simple mean-based predictor.
  • DeepGA-LSTM showed strong performance in the Semi-arid (R² = 0.6294, r = 0.8053) and High Plain (R² = 0.4672, r = 0.7172) regions.

Contributions

• First study to utilize neuroevolution for optimizing LSTM networks for drought forecasting in Mexico, addressing a gap in the literature for this region.
• Introduction of a novel hybrid encoding for genetic algorithms, combining a block sequence (for network layers, neuron count, and dropout rate) with a real-valued vector (for learning rate, context size, and batch size). This allows for more flexible and comprehensive optimization of LSTM architectures and hyperparameters compared to previous rigid encoding methods.
• Demonstrated the efficacy of neuroevolution in automatically discovering suitable network structures and hyperparameters, reducing reliance on subjective expert knowledge or limited grid search approaches.

Funding

• Secretaría de Ciencia, Humanidades, Tecnología e Innovación (SECIHTI) for a scholarship grant (CVU number 1311162) to the first author.

Citation

@article{VillegasVega2025Optimization,
  author = {Villegas-Vega, Ramiro and Márquez-Grajales, Aldo and Mezura‐Montes, Efrén and Salas-Martínez, Fernando and Ojeda-Misses, Manuel Alejandro and Romo-Gómez, Claudia},
  title = {Optimization of LSTM networks through neuroevolution for drought forecasting in Mexico},
  journal = {Theoretical and Applied Climatology},
  year = {2025},
  doi = {10.1007/s00704-025-05818-z},
  url = {https://doi.org/10.1007/s00704-025-05818-z}
}