基于多维度输入的水厂日取水量卷积长短期记忆网络预测  

Convolutional Long and Short-term Memory Network Prediction of Daily Water Intake from Waterworks Based on Multidimensional Input

作　　者：刘怀利[1] 王铭铭[1] 查淳膺 王健[2] 瞿暄 LIU Huai-li;WANG Ming-ming;ZHA Chun-ying;WANG Jian;QU Xuan(Anhui&Huaihe River Institute of Hydraulic Research,Hefei 230088,China;School of Civil and Hydraulic Engineering,Hefei University of Technology,Hefei 230009,China)

机构地区：[1]安徽省(水利部淮河水利委员会)水利科学研究院,安徽合肥230088 [2]合肥工业大学土木与水利工程学院,安徽合肥230009

出　　处：《水电能源科学》2024年第5期72-75,共4页Water Resources and Power

基　　金：安徽省自然科学基金项目(2208085US13);安徽省水利部淮委水利科学研究院科技攻关计划项目(KJGG202001)。

摘　　要：鉴于传统的单维输入卷积长短期记忆网络(CNN-LSTM)的预测精度依赖于历史数据的规律性,构建了基于多维度输入的卷积长短期记忆网络预测模型,其中采用皮尔逊相关分析来识别数据内部特征或外界环境因子相关性,在此基础上构建模型输入,并应用于两个自来水厂日取水量预测。结果表明,相较于传统单维度输入预测模型,构建的基于数据内部特征或外部环境因子的多维度输入CNN-LSTM预测导致A水厂取水量的平均绝对百分比误差分别降低了32%、17%;B水厂的平均绝对百分比误差分别降低了47%、12%,表明基于数据内部特征的多维度输入模型更高。其余评价指标也呈现类似变化;且增大水厂取水量有助于提高模型的预测精度。该模型输入分析方法可为提高预测模型精度提供有效范例。Given that the prediction accuracy of traditional single-dimensional input convolutional neural network-long short term memory(CNN-LSTM)depends on the regularity of historical data,this paper constructs a prediction model based on convolutional long short term memory network with multi-dimensional input,in which Pearson correlation anal-ysis is used to identify the internal characteristics of the data or the correlation of external environmental factors,on the basis of which the model input is constructed and applied to the prediction of daily water intake of two waterworks.Com-pared with the traditional unidimensional input prediction model,the results show that the proposed multidimensional in-put CNN-LSTM prediction leads to a reduction of 32%and 17%of the average absolute percentage error in the daily wa-ter intake of water plant A,and a reduction of 47%and 12%of the average absolute percentage error in water plant B,indicating that the multidimensional input based on the internal features of the data is a higher prediction model.The rest of the evaluation indicators showed similar changes;Increasing the amount of water withdrawn from the water plant helped to improve the prediction accuracy of the model.This model input analysis method can provide an effective exam-ple for improving the accuracy of prediction models.

关 键 词：取水量预测 卷积神经网络 相关性分析 多维度输入 

分 类 号：TV214[水利工程—水文学及水资源]