基于非平稳贝叶斯网络模型的月径流预报模拟研究

Simulation Study On Monthly Runoff Forecasting Based on Nonstationary Bayesian Network Model

精准实时的月径流预报是实现流域水资源优化调度和防汛抗旱的基本保证。受到气候变化和人类活动的双重影响,流域径流过程的趋势型非平稳特征较为显著,打破了传统月径流预报方法的平稳性假设前提。为此,聚焦于非平稳条件下水文变量间动态可变的相依性特征,构建一种考虑非平稳性的动态图形建模和贝叶斯网络结合(NGMBNs)模拟方法,并将其运用到新疆地区喀什流域的月径流预报中。具体实施过程如下:首先通过合理的数据转换方法使研究变量序列满足正态分步假设,极大提升了贝叶斯网络模型的预报精度;其次,通过动态调整分步预报阶段的贝叶斯网络结构和参数,实现对非平稳条件下的月径流预报模拟。基于5种模型性能指标(归一化均方根误差,Kling-Gupta效率系数,纳什效率系数,一致性指数,决定系数)的计算,NGM-BNs的预报精度明显优于4种常见的数据驱动模型(包括非平稳和平稳条件下支持向量回归(SVR)模型和自适应模糊神经网络(ANFIS)模型)。由于非平稳贝叶斯网络模型将变量间的时变特性纳入到模型构建过程中,其在捕捉极值流量方面展现了更强的模拟预报能力。本研究的开展可为气候变化背景下流域防洪减灾提供可靠的技术支撑和理论保证。

Accurate and real-time monthly runoff forecasting is essential for optimizing water resource management and flood and drought mitigation in river basins.Influenced by both climate change and human activities,the non-stationary characteristics of basin runoff processes are pronounced,challenging the stationary assumptions of traditional forecasting methods.This paper focuses on the dynamically variable dependencies among hydrological variables under non-stationary conditions,developing a novel simulation method that integrates Non-stationary Graphical Modeling and Bayesian Networks(NGM-BNs).This method is applied to monthly runoff forecasting in the Kashgar River Basin in Xinjiang.The process is as follows:initially,the research variable sequence is transformed through reliable data conversion methods to meet the normal distribution assumption,significantly enhancing the predictive accuracy of the Bayesian network model.Subsequently,the Bayesian network structure and parameters are dynamically adjusted during the forecasting phase to realize monthly runoff forecast simulation under non-stationary conditions.Based on five performance metrics—normalized root mean square error,Kling-Gupta efficiency coefficient,Nash efficiency coefficient,index of agreement,and coefficient of determination—NGM-BNs demonstrated superior forecasting accuracy compared to four common data-driven models,including non-stationary and stationary conditions of Support Vector Regression(SVR)and Adaptive Neuro-Fuzzy Inference System(ANFIS)models.The non-stationary Bayesian network model′s ability to incorporate time-varying characteristics of variables enhances its capability to capture extreme flow events.This study can provide reliable technical support and theoretical assurance for flood and disaster reduction in river basins under a changing climate.