PSO-LSTM-TMPH模型在水库调蓄流域径流模拟中的应用

Application of PSO-LSTM-TMPH Model in Runoff Simulation of Reservoir Regulation and Storage Basins

水库的建设和运行使得水库下游流量发生了显著改变,为了在月径流模拟中充分考虑水库的调蓄作用,选择东江流域作为典型流域,使用粒子群算法优化的长短期神经网络(Particle Swarm Optimization-Long Short Term Memory,PSO-LSTM)对东江流域的3个多功能水库枫树坝、新丰江、白盆珠水库的出流进行了模拟,并与传统的水库模型Level Pool Scheme(LPS)进行对比;使用三参数月度水文模型(Three-Parameter Monthly Hydrological Model Based on the Proportionality Hypothesis,TMPH)进行水库入流和区间来水模拟,与上述两种水库出流模型结合分别形成PSO-LSTM-TMPH和LPS-TMPH对东江流域重要站点龙川、河源、岭下、博罗进行预见期为一个月的径流模拟。结果表明:①PSO-LSTM在三大水库的模拟中效果均好于LPS,尤其在新丰江水库出流模拟中,在验证期的纳什效率系数(Nash-Sutcliffe Efficiency,NSE)、均方根误差(Root Mean Square Error,RMSE)分别为0.59、55.59 m^(3)/s,相比LPS提高了0.22,降低了17.01 m^(3)/s,说明该模型可以很好地捕捉多年调节水库复杂的水库出流规则;②PSO-LSTM-TMPH模拟龙川站、河源站、岭下站、博罗站径流的NSE为0.87,0.86,0.91,0.93,相比LPS-TMPH,NSE提高了0.09、0.21、0.07、0.03;③在测试期内,PSO-LSTM-TMPH水库出流模拟效果仍然较好,相比训练期、验证期模拟效果差异小,说明模型的模型泛化能力较强。研究建立的PSOLSTM-TMPH混合模型可以结合深度学习和物理模型各自的优势,适用于人类活动干扰下的径流模拟,可为优化水资源利用、实施干旱调度等提供技术支撑。

The construction and operation of reservoirs have led to significant alterations in downstream flow.In order to fully account for the reservoir regulation effects in monthly runoff simulations,the Dongjiang basin was selected as a representative watershed.Using Particle Swarm Optimization-Long Short-Term Memory(PSO-LSTM)algorithm,the outflows of three multi-functional reservoirs—Fengshubao,Xinfengjiang,and Baipenzhu—in the Dongjiang basin were simulated and a comparison was made with the traditional reservoir model,Level Pool Scheme(LPS).Additionally,Three-Parameter Monthly Hydrological Model Based on the Proportionality Hypothesis(TMPH)was employed for simulating reservoir inflow and interval low-flow,forming the combined models PSO-LSTM-TMPH and LPS-TMPH.These models were applied for one-month lead time runoff simulations at key sites in the Dongjiang basin,including Longchuan,Heyuan,Lingxia,and Boluo.The results indicate that:①PSO-LSTM outperformed LPS in simulating the three major reservoirs,particularly in the outflow simulation of Xinfengjiang Reservoir.During the validation period,the Nash-Sutcliffe Efficiency(NSE)and Root Mean Square Error(RMSE)were 0.59 and 55.59 m^(3)/s,respectively.Compared to LPS,this represented an increase of 0.22 and a decrease of 17.01 m^(3)/s,demonstrating the model′s ability to effectively capture the complex regulations of reservoir outflows over multiple years.②For the PSO-LSTMTMPH model,the NSE for runoff simulation at Longchuan,Heyuan,Lingxia,and Boluo stations was 0.87,0.86,0.91,and 0.93,respectively.Compared to LPS-TMPH,the NSE increased by 0.09,0.21,0.07,and 0.03,indicating improved accuracy.③During the testing period,the PSO-LSTM-TMPH reservoir outflow simulation remained effective,with smaller differences compared to the training and validation periods,suggestting strong model generalization capabilities.The hybrid model,PSO-LSTM-TMPH,established in this study,leverages the strengths of both deep learning and physical models.It is suitable for runoff simulation under human activity disturbances and provides technical support for optimizing water resource utilization and implementing drought scheduling.