基于多任务长短时卷积计算网络的降雨预测  

Rainfall prediction based on multi-task long short-term convolutional calculation network

作　　者：王军[1,2] 陈百艳 程勇 WANG Jun;CHEN Bai-yan;CHENG Yong(School of Computer and Software,Nanjing University of Information Science and Technology,Nanjing 210044,China;Technology Industry Department,Nanjing University of Information Science and Technology,Nanjing 210044,China)

机构地区：[1]南京信息工程大学计算机与软件学院,江苏南京210044 [2]南京信息工程大学科技产业处,江苏南京210044

出　　处：《计算机工程与设计》2022年第9期2686-2693,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(41875184、41975183);江苏省“六大人才高峰”创新团队基金项目(TD-XYDXX-004);赛尔网络下一代互联网技术创新基金项目(NGII20170610、NGII20171204);江苏省农业气象重点实验室开放基金项目(KYQ1309)。

摘　　要：为提高降雨预测的准确率,解决现有深度学习降雨预测模型缺乏对多站点气象数据时空关系建模能力的问题,提出基于多任务学习和改进的长短时卷积计算网络降雨预测模型(MTL-LSTC)。在长短时记忆网络内部的输入到状态和状态到状态的转换过程加入卷积运算结构,对气象序列数据进行编码,结合多任务学习方法提取出站点间隐藏的交互信息,建立站点间相关性模型,实现基于多站点气象数据的降雨预测。模拟实验结果表明,MTL-LSTC模型预测结果准确率更高且模型更高效,多站点气象数据的利用率也得到较大提升。To improve the accuracy of rainfall prediction and solve the problem that the existing deep learning rainfall prediction model lacks the ability to model the spatial-temporal relationship of multi-site meteorological data,a rainfall prediction model based on multi-task learning and improved long short-term convolutional calculation network(MTL-LSTC)was proposed.The convolution structure was added to the process of input to state and state to state transition in the long short-term memory network to encode the meteorological sequence data.Combined with multi-task learning method,the hidden interaction information between sites was extracted,and the correlation model between sites was established.The rainfall prediction based on multi-site meteorological data was realized.Experimental results show that the accuracy of MTL-LSTC model is higher and the model is more efficient,and the utilization of multi-site meteorological data is greatly improved.

关 键 词：降雨预测 长短时记忆网络 卷积计算结构 长短时卷积计算网络 多任务学习 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]