基于神经网络委员会机器的南中国海台风路径预报模型研究

发布时间：2018-05-13 00:18

本文选题：贝叶斯神经网络 + 混合密度网络　；参考：《华东师范大学》2017年硕士论文

【摘要】：台风是一种最具破坏性和毁灭性的极端自然灾害。中国位于西北太平洋西岸,是全球台风最活跃的地区,其中以南中国海海域及其沿岸地区为最。由于台风灾害的不可抗性,提前预测台风的转移路径为潜在受灾地区赢得充分的预警时间做好防灾准备工作是保护人民生命财产安全最重要和最有效的途径。随着人工神经网络技术的兴起和发展,基于神经网络技术的台风路径预报方法已经成为台风路径统计学预报技术中一个重要的研究分支。相比于传统的动力学数值预报技术和统计—动力学预报技术,基于人工神经网络的台风路径预报技术具有非线性拟合能力强,运算资源需求相对较低,对数据的适应性较强和鲁棒性较好等特性。近年来基于多模型集成的台风路径集成预报技术得到了广泛应用,本文的工作在于将人工神经网络及委员会机器集成理论应用于南中国海的台风路径预报模型的建立,尝试获得具备较好的稳定性和泛化能力且能够满足南中国海24小时短期台风路径预报业务需求的集成预报模型。首先,本文提出了一个基于贝叶斯神经网络并且使用Bagging委员会机器集成的台风路径预报模型。贝叶斯神经网络可以在较好地学习台风路径历史数据的同时控制网络模型的复杂度,而基于Bagging委员会机器的集成预报技术避免了传统集成预报技术中难以确定分量模型权重系数的困难。基于上述工作的成果,为了进一步增强预报模型对多模式台风路径的预报能力,本文接着提出了基于混合密度网络并且分别使用Averaging,Bagging委员会机器集成的台风路径预报模型。同时在模型的训练策略上采用了数据集"分块策略"对模型进行递进式的训练和测试,强化预报模型对数据集的学习并且综合观察预报结果的稳定性和泛化性能。在南中国海台风路径数据集上的多组实验表明,本文提出的集成预报模型在预报结果稳定性和泛化能力上有较好的表现,能够满足南中国海24小时短期台风路径预报业务的需求。
[Abstract]:Typhoon is one of the most destructive and devastating extreme natural disasters. Located on the west coast of the Northwest Pacific Ocean, China is the most active area of typhoons in the world, with the South China Sea and its coastal areas being the most active. Due to the irresistibility of typhoon disaster, it is the most important and effective way to protect the safety of people's life and property to predict the transition path of typhoon in advance to win sufficient early warning time for the potential disaster area to prepare for disaster prevention. With the rise and development of artificial neural network (Ann) technology, the forecasting method of typhoon track based on neural network technology has become an important research branch of typhoon track statistical prediction technology. Compared with the traditional dynamic numerical prediction technology and the statistical and dynamic prediction technology, the forecast technique of typhoon track based on artificial neural network has strong nonlinear fitting ability, and the demand for computing resources is relatively low. Better adaptability and robustness to the data. In recent years, the integrated forecasting technology of typhoon track based on multi-model integration has been widely used. The work of this paper is to apply artificial neural network and committee machine integration theory to the establishment of typhoon track prediction model in the South China Sea. An integrated forecasting model which has good stability and generalization ability and can meet the operational requirements of short-term typhoon track prediction in the South China Sea is obtained. Firstly, this paper presents a typhoon track prediction model based on Bayesian neural network and integrated with Bagging committee. Bayesian neural network can control the complexity of network model while learning typhoon track history data well. The integrated prediction technology based on the Bagging Committee machine avoids the difficulty of determining the weight coefficient of the component model in the traditional integrated forecasting technology. Based on the above work, in order to further enhance the forecasting ability of the forecasting model for the multi-model typhoon track, this paper proposes a typhoon track prediction model based on the mixed density network and using the Averagingsbagging Committee machine integration respectively. At the same time, the data set "block strategy" is used to train and test the model in order to enhance the learning of the prediction model and to observe the stability and generalization of the prediction results. A number of experiments on the South China Sea typhoon track data set show that the integrated forecasting model presented in this paper has good stability and generalization ability in forecasting results. Can meet the South China Sea 24-hour short-term typhoon track forecast business demand.
【学位授予单位】：华东师范大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P732.4;TP183

【相似文献】