| 引用本文: | 李润童,邢立文,崔宁博,等.基于群智能算法优化LSTM模型的参考作物蒸散量模拟研究[J].灌溉排水学报,2025,():-. |
| Li Runtong,Xing Liwen,Cui Ningbo,et al.基于群智能算法优化LSTM模型的参考作物蒸散量模拟研究[J].灌溉排水学报,2025,():-. |
|
| 摘要: |
| 摘 要：【目的】利用有限气象数据，提高西北干旱地区逐日参考作物蒸散量（ET0）模拟精度。【方法】将西北干旱地区划分为4个亚气候区（温带大陆性干旱区、温带大陆性高温干旱区、高原大陆性半干旱区和温带季风半干旱区），选取8个代表性气象站点1961—2019年逐日气象数据作为输入参数，以FAO-56 Penman-Monteith模型计算的ET0作为标准值。采用遗传算法（GA）和粒子群优化算法（PSO）对长短期记忆网络（LSTM）超参数进行优化，构建了LSTM、GA-LSTM和PSO-LSTM共3种深度学习模型。针对西北干旱地区气象数据缺乏的情况，设计了3种输入组合方案（温度-辐射型、温度型、温度-湿度型），仅利用气温、日照时数和相对湿度等基础气象要素，共建立9种模型组合。选取Priestley-Taylor、Hargreaves-Samani和Romanenko三种经验统计模型的ET0模拟值作为对照，评估深度学习模型的ET0模拟精度。【结果】PSO-LSTM模型模拟精度最高，决定系数（R2）、Nash-Sutcliffe系数（NSE）、均方根误差（RMSE）、相对均方根误差（RRMSE）、平均绝对误差（MAE）和综合性指标（GPI）分别为0.831~0.923、0.801~0.922、0.476~0.866 mm/d、0.190~0.382、0.299~0.627 mm/d和0.208~0.598，其中，温度-辐射型PSO-LSTM1模型在四个气候分区的ET0模拟精度最高，R2为0.893~0.923；智能优化算法显著提升了LSTM模型性能，其中PSO算法优于GA算法。【结论】基于温度-辐射型输入策略的PSO-LSTM模型模拟西北干旱地区ET0精度最高，为西北干旱地区及类似气候区域ET0准确模拟提供了有效方法。 |
| 关键词: 长短期记忆网络；智能优化算法；参考作物蒸散量；西北干旱地区，模型模拟 |
| DOI：10.13522/j.cnki.ggps.2025284 |
| 分类号:S161.4 |
| 基金项目:国家重点研发计划项目（2022YFD1900805）；四川省科技计划项目（2024YFHZ0217，2024ZHCG0101，2024YFHZ0200）；四川省基本科研业务费项目（2024JDKY0029-05）；中央高校基本科研业务费项目（20822041J4119）；国家自然科学基金项目（52309057）；新平柑橘产业科技创新示范县创建（202304BT090019） |
|
| A reference crop evapotranspiration simulation model based on LSTM networks optimized by swarm intelligence algorithms |
|
Li Runtong1, Xing Liwen1, Cui Ningbo1, Jiang Shouzheng1, Wang Zhihui1, Zhu Guoyu1, Liu Jincheng2, He Qingyan3
|
|
1.Sichuan University;2.Daying County Bureau of Agriculture and Rural Affairs;3.Sichuan Academy of Agricultural Machinery Sciences
|
| Abstract: |
| Abstract:【Objective】To achieve high-precision simulation of daily reference crop evapotranspiration (ET0) in Northwest China"s arid regions using limited meteorological data.【Method】The arid region of Northwest China was divided into four climatic sub-zones: temperate continental arid zone, temperate continental high-temperature arid zone, plateau continental semi-arid zone, and temperate monsoon semi-arid zone. Daily meteorological data from eight representative stations (1961–2019) were used as input parameters, with ET0 calculated by the FAO-56 Penman-Monteith model as the reference value. Three deep learning models—LSTM, GA-LSTM, and PSO-LSTM—were developed by optimizing Long Short-Term Memory (LSTM) network hyperparameters using Genetic Algorithm (GA) and Particle Swarm Optimization (PSO). To address the scarcity of meteorological data in Northwest China"s arid regions, three input combination schemes were designed: temperature-radiation type, temperature type, and temperature-humidity type, utilizing only basic meteorological elements (temperature, sunshine hours, and relative humidity), resulting in nine model combinations. Three empirical statistical models (Priestley-Taylor, Hargreaves-Samani, and Romanenko) were selected as benchmarks to evaluate the ET0 simulation accuracy of the deep learning models.【Result】The PSO-LSTM model achieved the highest simulation accuracy, with coefficient of determination (R2), Nash-Sutcliffe efficiency coefficient (NSE), root mean square error (RMSE), relative root mean square error (RRMSE), mean absolute error (MAE), and general performance indicator (GPI) ranging from 0.831~0.923, 0.801~0.922, 0.476~0.866 mm/d, 0.190~0.382, 0.299~0.627 mm/d, and 0.208~0.598, respectively. Among these, the temperature-radiation-based PSO-LSTM1 model demonstrated the highest ET0 simulation accuracy across all four climatic zones, with R2 ranging from 0.893~0.923. Intelligent optimization algorithms significantly enhanced LSTM model performance, with PSO outperforming GA.【Conclusion】The PSO-LSTM model based on the temperature-radiation input strategy exhibited optimal performance for ET0 simulation in Northwest China"s arid regions, providing an effective method for accurate ET0 simulation in this region and similar climatic areas. |
| Key words: neural networks; genetic algorithm; particle swarm optimization; ET0 simulation; arid Northwest China |
