【作者】 杨彬；

【导师】 艾波； 高松；

【作者基本信息】 山东科技大学 ， 测绘工程（专业学位）， 2023， 硕士

【摘要】 风暴潮由风暴增水和天文潮共同组成,现阶段对于天文潮的推演和预测已经非常准确,然而风暴增水由于影响因素的多样化给预测的准确性带来困难。当前风暴增水预报以数值模式结果为主,但其成因受非线性多变量的影响,各种风暴潮分量之间相互作用影响导致其峰值和时间难以准确预测,尤其在极端天气等影响下更具有一定的局限性,需要预报人员凭借专业知识和经验进行人工订正。本文基于深度学习方法进行风暴增水的预测,利用站点时间序列和空间场时间序列数据的学习训练,挖掘更为精确的预测因素之间关系,依托现有深度学习预测方法,得出更为精确的预测模型。本文的主要研究内容如下:(1)依托长短期记忆神经网络构建基于多模型平差接力的长时序风暴增水预测模型,实现对站点风暴增水的时序预测。首先,利用站点前时序的风速、风向、气压和风暴增水数据作为模型输入,通过多个子模型接力预测风暴增水时间序列,通过找出子模型预测结果的误差规律,减小随子模型迭代造成的误差累积,并结合均方根误差和相关系数筛选出不同角度划分模型、不同平差方法模型和不同前进时间模型中的最优模型。最后,将模型应用于不同区域位置及台风影响下的风暴增水预测,通过均方根误差、相关系数进行指标评价验证,并分析预测值与真实值曲线走势,结果表明基于多模型平差接力的长时序风暴增水预测模型具有较好的预测效果,有实际应用的可行性。(2)依托卷积长短时记忆神经网络构建基于多尺度海域分割预测方法的风暴增水预测模型,实现空间场风暴增水的预测。卷积长短时记忆神经网络能够充分利用卷积神经网络突出的图像特征提取能力,更加有效的处理时空数据。首先,利用有限体积海岸海洋模型(FVCOM)预报增水数据作为模型的训练输入,通过均方根误差和相关系数检验得出最优训练模型,然后,在此模型的基础上逐区域检验误差,超过设定误差阈值的区域重新提取该区域数据进行训练,若该子区域仍无法达到误差要求可进一步划分区域训练,该过程可通过双线性插值使模型训练更多数据找出数据时空规律。最后,组合训练过程中的子模型得到基于多尺度海域分割预测方法的FConvLSTM模型。(3)通过风暴增水再分析数据和台风数据对基于多尺度海域分割预测方法得到的模型进行验证与精度对比分析。首先,利用FVCOM预测数据训练模型流程得出基于再分析数据训练的多尺度海域分割预测模型,通过对整体海域预测模型和多尺度海域分割预测模型预测结果的均方根误差和相关系数,以及差值场的对比,得出多尺度海域分割预测方法的适用性。然后,进一步通过台风时期的预测结果验证此方法在极端天气时的适应性。最后,分析比较通过再分析数据和FVCOM预测数据训练得出的FConvLSTM模型在台风时期的预测精度,发现再分析数据相比预报数据得出的模型更适合台风时期风暴增水的预测。更多还原

【Abstract】 Storm surge is composed of storm surge and astronomical tide.At present,the deduction and prediction of astronomical tide are very accurate.However,due to the diversification of influencing factors,it is difficult to predict storm surge in accurately.At present,the forecast of storm surge is mainly based on numerical model results,but its causes are affected by nonlinear multivariate.The interaction between various storm surge components makes it difficult to accurately predict its peak value and time.Especially it has certain limitations under the influence of extreme weather,and requires forecasters to make manual corrections with professional knowledge and experience.In this paper,the prediction of storm surge is carried out based on the deep learning method.The station time series and spatial field time series data are used to learn and train,and the more accurate relationship between predictors is found.Based on the existing deep learning prediction method,a more accurate prediction model is obtained.The main research contents of this paper are as follows:(1)Based on the long short-term memory neural network,a prediction model of long-term storm surge based on multi-model gradual adjustment is constructed to realize the time series prediction of storm surge at the site.Firstly,the wind speed,wind direction,air pressure and storm surge data in the previous period of the station are used as model input.and the storm surge time series is predicted by multiple sub-models.By finding out the error law of the sub-model prediction results,the error accumulation caused by the sub-model iteration is reduced,and the optimal models in different angle division models,different adjustment method models and different forward time models are selected by combining the root mean square error and correlation coefficient.Finally,the model is applied to the prediction of storm surge in different regions and under the influence of typhoon.The root mean square error and correlation coefficient are used to evaluate and verify the indicator,and the relationship between the predicted value and the real value curve trend is analyzed.The results show that the prediction model of long-term storm surge based on multi-model gradual adjustment has a good prediction effect and has the feasibility of practical application.(2)Based on the convolutional long short-term memory neural network,a storm surge prediction model based on multi-scale sea area segmentation prediction method is constructed to realize the prediction of storm surge in spatial field.The convolutional long short-term memory neural network can make full use of the outstanding image feature extraction ability of Convolutional Neural Network to process spatio-temporal data more effectively.Firstly,the FVCOM forecast storm surge data is used as the training input of the model,and the optimal training model is obtained by root mean square error and correlation coefficient test.Then,on the basis of this model,the error is tested region by region.If the error of the region exceeds the set error threshold,the data of the region is re-extracted for training.If the sub-region still fails to meet the error requirements,the region can be further divided for training.This process can make the model train more data to find out the spatial and temporal law of the data through bilinear interpolation.Finally,the FConvLSTM model based on multi-scale sea area segmentation prediction method is obtained by combining the sub-models in the training process.(3)Through the storm surge of reanalysis data and typhoon data,the model based on the multi-scale sea area segmentation prediction method is verified and the accuracy is compared and analyzed.Firstly,the model of multi-scale sea area segmentation prediction based on reanalysis data training is obtained through the FVCOM prediction data training model process.By comparing the root mean square error and correlation coefficient of the prediction results of the whole sea area prediction model and the multi-scale sea area segmentation prediction model,as well as the difference field,the applicability of the multi-scale sea area segmentation prediction method is obtained.Then,the adaptability of this method in extreme weather is further verified by the prediction results of typhoon period.Finally,the prediction accuracy of FConvLSTM model trained by reanalysis data and FVCOM prediction data during typhoon period is analyzed and compared.It is found that the model trained by reanalysis data is more suitable for the prediction of storm surge during typhoon period than the the model trained by prediction data.更多还原