基于深度学习的气象预测研究

发布时间：2018-05-25 06:22

本文选题：深度学习 + 精细化预测　；参考：《哈尔滨工业大学》2017年硕士论文

【摘要】：伴随着计算机技术的迅猛发展,深度学习开启了人工智能新时代。以深度学习为代表,伴随其在计算机视觉、语音识别、自然语言处理等领域取得的突破性进展,新技术创新带来的不仅是挑战,同时也给气象预测技术的发展带来了机遇。课题针对气象温度进行时间序列建模,通过分析国内外研究现状及对时间序列预测模型的研究与对比,提出了改进深度学习框架来进行温度时间序列预测的思路。考虑到普通神经网络中出现的天气参数被认为是彼此独立,时序关系一般不被考虑的缺点,在对气象预测模型的构建中,提出了通过滑动时间窗手段改造,让普通神经网络也能学习到历史时序特征。实验表明,在深度前馈网络中加入时序特征的天气预报模型,效果要明显优于不考虑时序的模型。更进一步,针对实验中暴露出的前馈神经网络预报准确率随着预报时间增长快速下降的问题,提出了通过改造循环神经网络(RNN)进行气温预测的方法,并采用专门解决普通循环神经网络长时依赖问题的长短时记忆网络(LONG SHORT-TERM MEMEORY,LSTM)来构建气温预测模型。本文在分析了循环神经网络、RNN-LSTM网络、RNN-GRU网络的基础上,结合气温预测实验模型中出现的过拟合、梯度消失与梯度爆炸等一系列问题,提出使用Re LU激活函数以及加入正则化手段改进等策略,通过优化后的气温预测模型都较以往有更好的收敛效果。在实验中,还包含了对气象数据集的转换、清洗、属性选择、特征提取等工作。在平台应用方面,将实验搬到谷歌最新的深度学习框架TensorFlow-GPU中进行,使用GPU直接参与并行运算,为尝试复杂深度模型实验提供了可能。同时为验证模型的效果,实验不仅有对深度学习框架之间的比较,还加入了与传统ARIMA模型的比较。本文提出深度学习技术在精细化气温预测的应用研究,解决了一系列深度学习技术在气象预测上的具体实现与运用问题,创新了气温预测时序分析方法,拓展了区域化天气预报手段。
[Abstract]:With the rapid development of computer technology, deep learning has opened a new era of artificial intelligence. With the breakthrough in computer vision, speech recognition, natural language processing and so on, the new technology innovation brings not only challenges but also opportunities for the development of meteorological prediction technology. In this paper, time series modeling for meteorological temperature is carried out. Based on the analysis of the current research situation at home and abroad and the research and comparison of time series prediction model, an improved depth learning framework is put forward to predict the temperature time series. Considering the fact that the weather parameters in general neural networks are considered to be independent of each other and the time series relationships are not generally considered, in the construction of meteorological prediction models, a new method is proposed to modify the weather parameters by sliding time windows. So that ordinary neural networks can also learn the characteristics of historical time series. The experimental results show that the effect of the weather prediction model with time series features in the depth feedforward network is obviously better than that without considering the time series. Furthermore, aiming at the problem that the prediction accuracy of feedforward neural network is decreasing rapidly with the increase of forecast time, a method of temperature prediction by modifying the cyclic neural network (RNNN) is put forward. The long term memory network (LONG SHORT-TERM MEORYY LSTM), which is specially used to solve the problem of long time dependence of general circulatory neural networks, is used to construct the temperature prediction model. In this paper, based on the analysis of RNN-LSTM network and RNN-GRU network, a series of problems such as overfitting, gradient vanishing and gradient explosion in the experimental model of temperature prediction are discussed. By using re LU activation function and adding regularization method, the optimized temperature prediction model has better convergence effect than before. In the experiment, the transformation, cleaning, attribute selection and feature extraction of meteorological data sets are also included. In the aspect of platform application, the experiment is carried out in Google's latest depth learning framework (TensorFlow-GPU), and GPU is directly involved in parallel operation, which makes it possible to try out complex depth model experiment. In order to verify the effectiveness of the model, the experiment not only compares the depth learning framework, but also adds a comparison with the traditional ARIMA model. In this paper, the application research of depth learning technology in fine temperature prediction is proposed, which solves a series of problems in the realization and application of depth learning technology in meteorological prediction, and innovates the analysis method of temperature forecasting time series. Regional weather forecast methods have been expanded.
【学位授予单位】：哈尔滨工业大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：P45;TP18

【参考文献】