基于CNN-LSTM-AM模型的沿黄九省(区)用水量预测

Water consumption prediction in nine provinces (regions) along Yellow River based on CNN-LSTM-AM model

摘要

【目的】针对用水量预测中自然-社会二元特征未系统整合、时空异质性建模不足的局限性，基于深度学习方法构建沿黄九省(区)用水量时空协同预测框架。【方法】选取29个影响用水量的特征因子构建数据初集；通过随机森林算法进行特征因子重要性排序并排除冗余特征；综合考虑各深度学习算法的特点与适用场景，构建了基于卷积神经网络、长短期记忆网络和注意力机制的混合预测模型，并与其他基线模型进行对比；针对极端误差问题，设计双注意力协同工作机制对模型进行优化。【结果】研究区范围内CNN-LSTM-AM模型的模拟结果优于其他模型，MAE、MAPE和RMSE分别降低了7.7%~40.6%、22.6%~44.1%、0.7%~32.1%,整体性能优越。引入双重注意力协同工作机制后，模型能在保持整体精度波动较小的情况下降低极端误差。模型具有较好的泛化能力，能较为精确地预测研究区未来用水量。【结论】在研究区内，当前模型表现出良好的适用性与预测精度，可为用水量时空协同预测提供新的技术手段。未来研究应结合任务需求，考虑模型适配性、复杂性和稳定性之间的平衡，通过多维度分析来构建完整的预测体系。

Abstract

[Objective] To address the limitations of unsystematic integration of natural and social dual characteristics and insufficient modeling of spatiotemporal heterogeneity in water consumption prediction, a spatiotemporal collaborative prediction framework for water consumption in the nine provinces and regions along the Yellow River is constructed based on deep learning method. [Methods] A preliminary dataset was constructed using 29 characteristic factors influencing water consumption. The importance of these factors was ranked using the random forest algorithm, and redundant features were eliminated. Considering the characteristics and applicable scenarios of different deep learning algorithms, a hybrid prediction model based on convolutional neural network(CNN), long short-term memory(LSTM) network, and attention mechanism(AM) was established and compared with other baseline models. To address the problem of extreme errors, a dual-attention collaborative mechanism was designed to optimize the model. [Results] In the study area, the CNN-LSTM-AM model achieved better simulation result than other models, with mean absolute error(MAE), mean absolute percentage error(MAPE), and root mean square error(RMSE) reduced by 7.7%~40.6%, 22.6%~44.1%, and 0.7%~32.1%, respectively, indicating superior overall performance. After introducing the dual-attention collaborative mechanism, extreme errors were reduced while maintaining small fluctuations in overall accuracy. The model demonstrated good generalization ability and was able to predict future water consumption in the study area with high accuracy. [Conclusion] In the study area, the current model shows good applicability and prediction accuracy, providing a new technical approach for spatiotemporal collaborative prediction of water consumption. Future research should consider the balance among model adaptability, complexity, and stability based on task requirements, and construct a comprehensive prediction system through multi-dimensional analysis.

关键词

深度学习 / 用水量预测 / 卷积神经网络 / 长短期记忆网络 / 注意力机制 / 自然-社会二元特征 / 时空异质性 / 水资源规划

Key words

deep learning / water consumption prediction / convolutional neural network / long short-term memory network / attention mechanism / naturac-social binary features / spatio temporal heterogeneity / water tesources planning

基金资助

国家重点研发计划(2023YFF1304202)

流域水循环模拟与调控国家重点实验室项目(SKL2024YJZD02)

云南省重点研发计划项目(202303AC100020)