基于小波-LSTM的流溪河流域短期库水位预测模型  

Short-term Reservoir Water Level Prediction Model of Liuxi River Basin based on Wavelet-LSTM

作　　者：张浩森 龙岩 康爱卿[3] ZHANG Hao-sen;LONG Yan;KANG Ai-qing(School of Water Resources and Electric Power,Hebei University of Engineering,Handan 056038,China;School of Water Resources and Electric Power,Hebei Key Laboratory of Smart Water Conservancy,Handan 056038,China;China Institute of Water Resources and Hydropower Research,Beijing 100038,China)

机构地区：[1]河北工程大学水利水电学院,河北邯郸056038 [2]河北工程大学河北省智慧水利重点实验室,河北邯郸056038 [3]中国水利水电科学研究院,北京100038

出　　处：《海河水利》2023年第11期63-67,76,共6页Haihe Water Resources

摘　　要：考虑到原始监测数据中噪声对库水位预测精度的影响,研究提出了基于小波-LSTM神经网络构建的流溪河流域短期水库水位预测模型。模型以流溪河水库和黄龙带水库2015—2020年的时序数据为研究对象,以小波分解对水位时序数据进行多尺度分析,以2019年8月1日为界划分数据集,以均方根误差RSME、纳什效率系数NSE和平均绝对误差MAE为评价指标,并与LSTM模型在1、6和12 h预见期情境下的预测结果进行了对比分析。结果表明,在1、6和12 h预见期下,小波-LSTM模型和LSTM模型的NSE值均大于0.9,2种模型均取得了不错的预测效果;在相同预见期下,相较于LSTM模型,小波-LSTM模型预测的整体误差和极值误差均更小,模型的整体预测效果更优;对模型输入特征而言,单特征小波-LSTM模型的整体预测误差和极值误差均低于多特征小波-LSTM模型,对水位时序数据的整体预测效果更好,预测中的异常值也相对更少。Considering the influence of noise in the original monitoring data on the prediction accuracy of reservoir water level,a short-term reservoir water level prediction model of Liuxi River Basin based on wavelet-LSTM neural network was proposed.The model takes the time series data of Liuxihe Reservoir and Huanglongdai Reservoir from 2015 to 2020 as the research object,and uses wavelet decomposition to perform multi-scale analysis of water level time series data.The error RSME,NSE and MAE are used as evaluation indicators,and the prediction results of the LSTM model in the 1h,6h and 12h forecast period are compared and analyzed.The research shows that under the forecast period of 1h,6h and 12h,the NSE values of the wavelet-LSTM model and the LSTM model are both greater than 0.9,and both models have achieved good prediction results.Under the same forecast period,compared with the LSTM model,the overall error and extreme value error predicted by the wavelet-LSTM model are smaller,and the overall prediction effect of the model is better.For the input features of the model,the overall prediction error and extreme value error of the single feature wavelet-LSTM model are lower than those of the multi-feature wavelet-LSTM model,and the overall prediction effect of the water level time series data is better,and the outliers in the prediction are relatively less.

关 键 词：水位预测 LSTM循环神经网络 水库水位 小波分解 预见期 

分 类 号：TV697.21[水利工程—水利水电工程] TV124