基于机器学习算法的含沙量短临预报模型研究  

Short-Term Forecasting of Suspended Sediment Concentration Based on Machine Learning

作　　者：魏苗 胡新源 周聂 陈娜[2] 易瑞吉 马仲坤 陈华[2] WEI Miao;HU Xin-yuan;ZHOU Nie;CHEN Na;YI Rui-ji;MA Zhong-kun;CHEN Hua(China Energy Xinjiang Jilintai Hydropower Development Co.,Ltd.,Yili 835100,Xinjiang,China;State Key Laboratory of Water Resources Engineering and Management,Wuhan University,Wuhan 430072,Hubei Province,China;China Energy Science and Technology Research Institute Co.,Ltd.,Chengdu 618000,Sichuang Province,China)

机构地区：[1]国家能源集团新疆吉林台水电开发有限公司,新疆伊犁835100 [2]武汉大学水资源工程与调度全国重点实验室,湖北武汉430072 [3]国家能源集团科学技术研究院有限公司,四川成都618000

出　　处：《中国农村水利水电》2024年第9期60-67,共8页China Rural Water and Hydropower

基　　金：国家重点研究与发展计划项目(2022YFC3002701);国家能源集团新疆吉林台水电开发有限公司NLK、TLDSY、SLKT三站泥沙含量自动监测分析决策系统(GDDL-22245)。

摘　　要：入库水流含沙量直接影响水电站机组安全运行,准确预测入库含沙量可为水电站的停机避峰提供决策依据,同时对减少水轮机组泥沙磨损,延长其使用寿命具有重要意义。为开展更为有效和精准的入库含沙量短临预报,研究基于气温、水位、流量等水文气象监测资料,结合长短期记忆模型(LSTM)、支持向量回归模型(SVR)和随机森林模型(RF),探索构建含沙量短临预报模型,并以位于喀什河(KSH)流域的塔勒德萨依(TLDSY)电站库区为研究对象,对所建模型的适用性及可靠性进行验证。研究结果表明:所建SVR模型可有效预测入库水流含沙量的涨落变化趋势,但在含沙量定量预测方面,该模型存在一定的误差,预测结果普遍偏高;RF模型可较为准确的根据过去10h的水文气象信息预测出未来1~3 h的入库水流含沙量,但随着预见期的增加,RF模型稳定性下降明显,可能出现较大的局部误差;LSTM模型同样可较为准确预测出未来1~3h的入库水流含沙量,并且对预见期为4~5h的预测效果更为稳定,其NSE值始终保持在0.6以上,MAE值在0.15kg/m^(3)以下,沙峰预测误差可控制在15%以内。综上所述,研究基于LSTM算法所建的含沙量短临预报模型表现最佳,可基于历史水文气象监测信息,实现更为准确的入库含沙量短临预报,进而为水电站安全高效运行提供更为可靠的数据支撑。The sediment concentration in inflowing water directly affects the safe operation of hydropower stations.Accurate prediction of sed-iment concentration supports decision-making for peak shaving during shutdowns of hydroelectric plants and is crucial for reducing sediment abrasion on turbine units,thereby extending their operational lifespan.To develop a more effective and precise short-term sediment concen-tration forecast,this study employs Long Short-Term Memory(LSTM),Support Vector Regression(SVR),and Random Forest(RF)mod-els to construct a forecast model based on hydro-meteorological monitoring data,including temperature,water level,and flow rate.The ap-plicability and reliability of the models are validated using the reservoir area of the Taledesayi(TLDSY)power station in the Kashi River(KSH)basin as the research subject.The results indicate that the SVR model can effectively predict the fluctuation trend of sediment concen-tration in inflowing water.However,the model exhibits some errors in quantitatively predicting sediment concentration,with generally over-estimated results.The RF model can predict the sediment concentration in inflowing water for the next 1-3 hours fairly accurately based on hydro-meteorological information from the past 10 hours.However,as the forecast period increases,the stability of the RF model decreases significantly,leading to potentially large local errors.The LSTM model can also accurately predict the sediment concentration in inflowing water for the next 1~3 hours,and it demonstrates more stable predictions for a forecast period of 4~5 hours,with NSE value consistently above 0.6,MAE value below 0.15 kg/m^(3),and the peak sediment prediction error within 15%.In conclusion,the short-term sediment concen-tration forecast model constructed based on the LSTM algorithm performs best.It can achieve a more accurate short-term sediment concentra-tion forecast for inflowing water based on historical hydro-meteorological monitoring information,providing quantitative data support for the s

关 键 词：机器学习 LSTM 含沙量预报 短临预报 喀什河 

分 类 号：P338.5[天文地球—水文科学]