基于差分Causal LSTM模型的气象图像短时预测研究  

Research on short-time prediction of meteorological images based on differential-Causal LSTM Model

作　　者：张晓晖[1] 白文奇 杨松楠 王晓娟 ZHANG Xiaohui;BAI Wenqi;YANG Songnan;WANG Xiaojuan(Faculty of Automation and Information Engineering,Xi’an University of Technology,Xi’an 710048,China)

机构地区：[1]西安理工大学自动化与信息工程学院,陕西西安710048

出　　处：《西安理工大学学报》2023年第4期529-535,共7页Journal of Xi'an University of Technology

基　　金：陕西省自然科学基础研究计划资助项目(2021JLM-58)。

摘　　要：为解决气象图像序列在短时预测时预测精度低的问题,利用一种具有级联记忆单元的Causal LSTM,将图像梯度差分惩罚因子引入训练过程,来提高预测模型对短时序列动态和突变的建模能力,提出了差分Causal LSTM模型。研究首先通过循环神经网络建立气象图像短时预测模型,然后分析了ConvLSTM模型对气象雷达回波图与卫星云图序列的预测效果,对于ConvLSTM模型预测气象图像存在严重模糊的问题,使用差分Causal LSTM模型进行优化,结果表明改进的模型能够有效改善模糊,提升预测结果的准确性。改进后的差分Causal LSTM模型在HKO-7数据集的测试样本中,关键成功指数(CSI)提高了0.019,在气象云图数据集中提高了0.078,模糊程度有所减弱。Due to the low accuracy of meteorological image sequences’short-term prediction,we propose a differential-Causal LSTM model by using Causal LSTM with cascaded memory units,which is an introduction to the image gradient difference penalty term into the training process to improve the prediction model’s ability to capture the dynamics and abrupt changes of short-time sequences.We first establish the meteorological image short-time prediction model by the recurrent neural network and then analyze the prediction effect by the ConvLSTM model on weather radar echogram and satellite cloud sequences.The results show that the improved model in this paper can effectively reduce the blurring and improve the accuracy of prediction results.The differential-Causal LSTM model improves the critical success index(CSI)by 0.019 in the HKO-7 dataset,CSI also improved by 0.078 in the meteorological cloud image dataset,and the blurring is reduced.

关 键 词：ConvLSTM Causal LSTM 端到端模型 图像梯度差分损失 

分 类 号：P456.1[天文地球—大气科学及气象学]