Summer seasonal precipitation prediction in the middle and lower reaches of the Yangtze River using neural network approaches
- 1APT MOBILE SATCOM LIMITED, Shenzhen, China
Abstract
Accurate prediction of summer precipitation is essential for agricultural planning and flood prevention. However, precipitation is strongly nonlinear. In the middle and lower reaches of the Yangtze River (MLYR), many interacting factors and large interannual variability make conventional forecasting methods unreliable. This study proposes neural network approaches to address this challenge. We select independent predictors from 130 circulation indices of the National Climate Center (NCC) and from the ERA5 monthly dataset, including sea surface temperature (SST), 500 hPa geopotential height, and land surface variables. Predictors are chosen through correlation, partial correlation, and autocorrelation analyses. They are then processed with detrended fluctuation analysis (DFA), standardization, and normalization. Four models are trained: a back propagation (BP) network, a gated recurrent unit (GRU) network, a coupled BP-GRU network, and a particle swarm optimization (PSO)-optimized coupled BP-GRU network. A strict chronological split is used (training: 1980–2000; validation: 2001–2010; testing: 2011–2024). Results are reported as the mean ± standard deviation over ten independent runs. The PSO-optimized coupled network performs best (test MSE = 0.0177 ± 0.0106; spatial correlation r = 0.44). It outperforms the standalone GRU (MSE = 0.0207 ± 0.0003), the coupled BP-GRU (MSE = 0.0357 ± 0.0175), and the standalone BP model (MSE = 0.0447 ± 0.0212). It is also the only network to beat the climatological baseline (MSE = 0.0181). The GRU is the most stable across runs but reproduces spatial patterns poorly. Overall, combining recurrent architectures with physical predictors through evolutionary optimization improves seasonal prediction. However, all models show limited skill (R2 near zero), reflecting the difficulty of this task. The approach offers a reference for flood-season forecasting in the MLYR, with potential use in agriculture and flood prevention.
Related
- A long-term, analysis-ready dataset of rice phenology in middle and lower reaches of the Yangtze River Plain of China from 1981 to 2024.
- Seasonal Dynamics, Botanical Origin, and Yield of Apis mellifera Propolis in the Mexican Altiplano: Influence of Precipitation, Phenology, and Harvesting Methods
- How effective are contrastive learning-based approaches for activity-cliff prediction?
- A Curated Phenotype-Linked HIV-1 Sequence Dataset for Machine Learning Approaches to Tropism Prediction, Disease Progression and Proviral Reservoir Studies.
- Computational prediction of a multi-epitope Human Metapneumovirus vaccine candidate through integrated reverse vaccinology and pan-genomic approaches
BibTeX
@article{ren2026summer,
title = {{Summer seasonal precipitation prediction in the middle and lower reaches of the Yangtze River using neural network approaches}},
author = {Xiran Ren and Lingya Li},
journal = {PLOS One},
year = {2026},
month = nov,
doi = {10.1371/journal.pone.0343910},
url = {https://doi.org/10.1371/journal.pone.0343910},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Summer seasonal precipitation prediction in the middle and lower reaches of the Yangtze River using neural network approaches
AU - Xiran Ren
AU - Lingya Li
PY - 2026
DA - 2026/11/05
JO - PLOS One
DO - 10.1371/journal.pone.0343910
UR - https://doi.org/10.1371/journal.pone.0343910
AB - Accurate prediction of summer precipitation is essential for agricultural planning and flood prevention. However, precipitation is strongly nonlinear. In the middle and lower reaches of the Yangtze River (MLYR), many interacting factors and large interannual variability make conventional forecasting methods unreliable. This study proposes neural network approaches to address this challenge. We select independent predictors from 130 circulation indices of the National Climate Center (NCC) and from the ERA5 monthly dataset, including sea surface temperature (SST), 500 hPa geopotential height, and land surface variables. Predictors are chosen through correlation, partial correlation, and autocorrelation analyses. They are then processed with detrended fluctuation analysis (DFA), standardization, and normalization. Four models are trained: a back propagation (BP) network, a gated recurrent unit (GRU) network, a coupled BP-GRU network, and a particle swarm optimization (PSO)-optimized coupled BP-GRU network. A strict chronological split is used (training: 1980–2000; validation: 2001–2010; testing: 2011–2024). Results are reported as the mean ± standard deviation over ten independent runs. The PSO-optimized coupled network performs best (test MSE = 0.0177 ± 0.0106; spatial correlation r = 0.44). It outperforms the standalone GRU (MSE = 0.0207 ± 0.0003), the coupled BP-GRU (MSE = 0.0357 ± 0.0175), and the standalone BP model (MSE = 0.0447 ± 0.0212). It is also the only network to beat the climatological baseline (MSE = 0.0181). The GRU is the most stable across runs but reproduces spatial patterns poorly. Overall, combining recurrent architectures with physical predictors through evolutionary optimization improves seasonal prediction. However, all models show limited skill (R2 near zero), reflecting the difficulty of this task. The approach offers a reference for flood-season forecasting in the MLYR, with potential use in agriculture and flood prevention.
SN - 1932-6203
ER -