泾河流域降水-径流关系与径流预测研究

Research on precipitation-runoff relationship and runoff prediction in Jinghe River Basin

摘要

【目的】以半干旱区泾河流域为研究对象，分析泾河径流的时空变化特征，并探讨径流对流域降水变化的响应关系。【方法】利用2022—2024年泾河水文站径流和流域降水数据，分析径流时空变化。采用K-Means聚类算法，根据站点降水量对径流的贡献度，将流域划分为4个区，进一步分析不同区域降水-径流滞后关系。最后，基于长短期记忆(LSTM)神经网络，构建1 h、3 h和6 h预见期的泾河径流临近预测模型。【结果】结果显示：(1)泾河日平均径流季节变化显著，汛期是高径流时段，极端峰值集中于7月中下旬，3年均呈唯一极端峰值和多次波动变化。(2)泾河3次典型洪峰的分析发现，当前期土壤水含量相近时，流域内最大小时降水量和降水持续时间是洪峰的关键因子，表现为降水持续时间越长、强度越大和总降水量越大，洪峰历时越长且峰值越高。前期0~28 cm土层土壤水含量小于0.2 m³/m³时，全流域平均滞后时间达30 h,接近大于0.4 m³/m³时的两倍。流域内降水-径流转换存在区域性差异，前期土壤水含量较高时区域1—4平均最优滞后时间分别为11 h、12 h、14 h和28 h,含水量较低时滞后时间分别为23 h、30 h、34 h和35 h。(3)利用LSTM机器学习算法构建了泾河径流临近预测模型。试验了不同影响因子对模型的影响，发现景村站前序流量、雨落坪站前序流量和流域面雨量是关键因子。1 h、3h和6 h预见期最优模型的纳什效率系数分别为0.902、0.774和0.676,均方根误差分别为2.788 m³/s、4.230 m³/s和5.070 m³/s,预测精度随预见期延长而降低。模型对较大洪水和平稳时段的径流模拟效果较好。【结论】揭示的洪峰对流域降水的滞后效应，以及构建的径流预测模型可为泾河流域水资源调控与防洪预报提供技术支撑，也为半干旱区中小流域降水-径流关系研究提供新的思路。

Abstract

[Objective] The Jinghe River Basin in a semi-arid region is selected as the study area to analyze the spatiotemporal variation characteristics of runoff and to investigate the response relationship of runoff to precipitation changes in the river basin.[Methods] Spatiotemporal variations in runoff were analyzed using runoff data from hydrological stations and precipitation data within the river basin from 2022 to 2024. The river basin was divided into four regions using the K-Means clustering algorithm based on the contribution of station precipitation to runoff, and the precipitation-runoff lag relationship in different regions was further analyzed. Finally, short-term runoff prediction models for the Jinghe River with lead times of 1 h, 3 h, and 6 h were developed based on a long short-term memory(LSTM) neural network.[Results] The result showed that:(1) the daily average runoff of the Jinghe River exhibited pronounced seasonal variation, with the flood season being the high-discharge period. Extreme peaks were concentrated in mid-to-late July, with three consecutive years showing a single extreme peak with multiple fluctuations.(2) Analysis of three typical flood peaks in the Jinghe River showed that when antecedent soil moisture content was similar, the maximum hourly precipitation and precipitation duration within the river basin were the key factors influencing the flood peaks. Specifically, longer precipitation duration, higher intensity, and larger total precipitation led to longer flood duration and higher peak discharges. When the antecedent soil moisture content in the 0~28 cm soil layer was below 0.2 m³/m³, the basin-wide average lag time reached 30 h, nearly twice that when the soil moisture content exceeded 0.4 m³/m³. Regional differences were observed in the precipitation-runoff transformation within the river basin. Under higher antecedent soil moisture content conditions, the optimal average lag times for regions 1~4 were 11 h, 12 h, 14 h, and 28 h, respectively. Under lower soil moisture content conditions, the lag times were 23 h, 30 h, 34 h, and 35 h, respectively.(3) A short-term runoff prediction model for the Jinghe River was established using the LSTM machine learning algorithm. The effects of different influencing factors on the model were tested, revealing that antecedent discharge at Jingcun and Yuluoping stations and basin average areal precipitation were identified as key factors. For lead times of 1 h, 3 h, and 6 h, the optimal models achieved Nash-Sutcliffe efficiency coefficients of 0.902, 0.774, and 0.676, respectively, and root mean square errors of 2.788 m³/s, 4.230 m³/s, and 5.070 m³/s, respectively. Prediction accuracy decreased with increasing lead time. The model performed well in simulating runoff during large floods and stable-discharge periods.[Conclusion] The identified lag effect of flood peaks on basin precipitation and the developed runoff prediction models can provide technical support for water resource regulation and flood prediction in the Jinghe River Basin and offer new insights for studying the precipitation-runoff relationships in small and medium-sized river basins in semi-arid regions.

关键词

泾河流域 / 径流变化 / 降水-径流关系 / 滞后性 / 长短期记忆神经网络(LSTM) / 时空变化 / 气候变化 / 半干旱区

Key words

Jinghe River Basin / runoff variation / precipitation-runoff relationship / lag effect / long short-term memory neural network(LSTM) / spatiotemporal variation / climate change / semi-arid region

基金资助

国家自然科学基金气象联合基金重点项目(U2242202)

青海省科技厅基础研究项目(2025-ZJ-741)

陕西省重点研发计划项目(2023-YBSF-191)