###
工程科学与技术:2024,56(1):54-64
←前一篇   |   后一篇→
本文二维码信息
码上扫一扫!
基于融合注意力机制LSTM网络的地下水位自适应鲁棒预测
(1.四川大学 电气工程学院,四川 成都 610065;2.成都兴蓉市政设施管理有限公司,四川 成都 610065)
Adaptive Robust Prediction of Groundwater Level Based on Fusion Attention Mechanism LSTM Network
(1.School of Electrical Eng., Sichuan Univ., Chengdu 610065, China;2.Chengdu Xingrong Municipal Facilities Management Co. Ltd., Chengdu 610065, China)
摘要
图/表
参考文献
相似文献
附件
本文已被:浏览 227次   下载 162
投稿时间:2023-04-24    
中文摘要: 地下水水位是旱天污水管网地下水入渗量的重要影响因素,快速精准地预测地下水水位能有效提升旱天污水管网地下水入渗量估算准确度,辅助优化管网病害治理与维护策略。针对目前城市复杂水文预测存在的准确度低、灵敏度低、泛化能力弱等问题,本文提出了一种新的鲁棒自适应水位预测算法。首先,对水文数据进行预处理,解决了数据时间跨度大、噪声多、缺失及异常、非平稳等问题。其次,针对不同输入特征对预测指标的影响,在模型训练阶段提出一种新的空间变量注意机制,可快速识别与水位关联的关键变量,并对输入特征赋予不同的影响权重。然后,针对不同序列长度对预测效果的影响,还设计了自适应时间注意力机制,帮助网络自适应地找出与不同时间序列长度预测指标相关的编码器隐藏状态,以更好地捕捉时间上的依赖关系。在此基础上,以上下文向量作为输入,提出一种融合注意力机制的长短时记忆网络水文预测算法。最后,通过意大利Petrignano水文数据验证了所提算法的有效性,并与GRU、Elman、LSTM、VA–LSTM和S–LSTM等方法进行预测性能比较。结果表明,基于融合注意力机制的LSTM网络在面临大规模、噪点多的复杂数据时有优于其它几种算法的预测效果,表明该算法具有强自适应性和鲁棒性。本文研究结果可以为市政排水策略合理调整、及时控制提供参考。
Abstract:Groundwater level is an important factor affecting groundwater infiltration of sewage pipe network in dry weather. Accurate prediction of groundwater level can effectively improve the accuracy of groundwater infiltration estimation in dry weather, and assist in optimizing pipe network disease control and maintenance strategies. Aiming at the problems of low accuracy, low sensitivity, and weak generalization ability in the current urban complex hydrological prediction, a new robust adaptive water level prediction algorithm was proposed in this paper. First, a prior processing was carried out on the hydrological data, which solved the problems of large time span, high noise, missing and abnormal, and non-stationary data. Secondly, in view of the influence difference of input features on predictive indicators, a new spatial variable attention mechanism was proposed in model training stage, which can quickly identify key variables associated with water levels and assign different influence weights to input features. Furthermore, in view of the influence difference of various sequence lengths on the prediction effect, an adaptive temporal attention mechanism was also designed to adaptively find out the hidden state of the encoder related to the predictors of different sequence lengths, so as to capture time dependencies. On this basis, with the context vector as the input, an LSTM hydrological prediction algorithm integrating attention mechanism was proposed. Finally, the effectiveness of the proposed algorithm was verified by the hydrological data of Petrignano, Italy. The prediction performance was compared with GRU, Elman, LSTM, VA–LSTM and S–LSTM methods. The results showed that the proposed STA–LSTM network based on the fusion attention mechanism has a better prediction effect than other algorithms when faced with complex, large-scale, and noisy data, indicating the strong adaptability and robustness of the algorithm. The research results of the paper provide a reference for the reasonable adjustment and timely control of municipal drainage strategies.
文章编号:202300315     中图分类号:TV211.12    文献标志码:
基金项目:国家重点研发计划项目(2020YFB1709705)
作者简介:第一作者:佃松宜(1972-),男,教授,博士.研究方向:先进控制与人工智能算法;机器人与自动化.E-mail:scudiansy@scu.edu.cn;通信作者:郭斌,E-mail:bguodxl@scu.edu.cn
引用文本:
佃松宜,厉潇滢,杨丹,芮胜阳,郭斌.基于融合注意力机制LSTM网络的地下水位自适应鲁棒预测[J].工程科学与技术,2024,56(1):54-64.
DIAN Songyi,LI Xiaoying,YANG Dan,RUI Shengyang,GUO Bin.Adaptive Robust Prediction of Groundwater Level Based on Fusion Attention Mechanism LSTM Network[J].Advanced Engineering Sciences,2024,56(1):54-64.