【作者】 刘辉；

【导师】 田红旗；

【作者基本信息】 中南大学 ， 载运工具运用工程， 2011， 博士

【摘要】 强风是危及铁路运输安全的主要气象灾害之一。我国有多条横跨恶劣强风区域的铁路线路,包括青藏、兰新等。强风线路沿线风速引发的气动横向力和气动升力是造成列车吹翻事故的根本原因。开展铁路沿线关键区域强风风速实时预测研究,是铁路运营部门在恶劣强风环境下防范事故、进行科学决策和安全行车指挥调度的有效手段,也是研建高水平强风预警指挥系统的核心关键技术之一。为获得不同步长的高精度铁路沿线风速短期预报值,本文引入小波分析、遗传算法、神经网络和自适应卡尔曼滤波等现代智能优化理论,结合时间序列分析理论,开展高精度风速预测智能优化研究。经过多年研究,提出了铁路沿线非平稳风速信号智能预测新算法：(1)对实测非平稳风速信号建立时间序列模型,完成超前多步预测计算。针对所建模型存在的精度不高问题,引入滚动修正优化改进思路,提出了滚动时间序列分析法。并将该方法与小波分析法混合建模,提出了小波分析-滚动时间序列分析法(Wavelet Rolling Time Series Method,简称WRTSM)。预测实例表明：WTSM明显提高了时间序列分析法的预测精度,改善了模型预测延时现象。WTSM兼具小波分析法信号细分与时间序列分析法建模简单的综合算法性能,获得了高精度的大步长预测结果。该研究成果已刊登于国际SCI刊物《Renewable Energy》。(2)对实测非平稳风速信号建立BP (Back Propagation)神经网络模型,完成超前多步预测计算。针对BP预测网络初始权值确定的随意性与主观性,以及网络学习时间过长、预测精度不高等不足,引入遗传算法和时间序列分析理论的潘迪特-吴贤明建模方案,提出了遗传-神经网络优化模型和神经网络结构时间序列确定方法。并将该优化方法与小波分析法混合建模,提出了小波分析-遗传算法-神经网络法(Wavelet Genetic BP Method,简称WGBM)。预测实例表明：WGBM明显改善了上述神经网络的不足,兼有小波分析法信号细分、遗传算法信号全局搜索和神经网络法信号非线性映射等能力,获得了高精度小步长预测结果。该研究成果将刊登于国际SCI刊物《Information Science》。(3)为了获得非平稳风速信号超前单步超高精度预测,运用小波分析法和自适应卡尔曼滤波法混合建模,提出了小波分析-卡尔曼滤波法(Wavelet-Kalman Method,简称WKM)。预测实例表明：WKM吸收了小波分析法信号细分和自适应卡尔曼滤波法实时追踪的综合算法特征,获得了超前单步超高精度预测结果,该研究成果已刊登于《中国电机工程学报》和《中南大学学报(英文版)》。(4)为进一步考核优化算法性能,本文将其推广应用于我国典型强风高海拔铁路-青藏铁路工程。通过对沿线唐古拉山、风火山、五道梁等重点监控区域风速信号进行建模与预测,结果表明所提出的几种风速预测优化改进算法是正确和可行的。此外,由于非平稳随机信号预测机理相似,所提出的算法同样可以推广到风电场风速、机器人路径、机械振动等领域,具有学术与工程双重意义。更多还原

【Abstract】 Strong-wind is one of major meteorological disasters which affect the safety of railway transportation.China has several railways which pass the strong-wind environment. The aerodynamic transverse forces and lift forces caused by strong-wind are the original reason for train derailment under strong-wind condition. Doing research on real-time forecasting for wind speed from certain important zones along railway is not only an effective way to avoid accidents and provide scientific guidance for railway departments, but also is the key technology to establish a strong-wind warning system.To attain the high-precision different steps ahead wind speed predictions along railways, several intelligent optimization theories including wavelet analysis, genetic algorithms, neural networks and kalman filtering are introduced in this study, and optimizations are done. After years of research, the method of identification and prediction algorithms for unsteady wind speed signal has been presented, which include the main contents as follows.(1) Some time series models have been established for the real non-stationary wind speed signals to get the multi-step ahead predictions. Aimed at the un-satisfactory prediction accuracy of those time series models, a new optimization method named rolling time series method (RTSM for short) has been proposed based on modifying the calculation process of time series method (TSM for short). In addition, another optimization method named wavelet-rolling time series method (WTSM for short) has also been presented based on RTSM and wavelet analysis method (WAM for short). From the prediction cases, it can be found that: WTSM significantly improved the wind speed prediction accuracy of traditional time series models, and solve the prediction-delay phenomenon of time series models. WTSM both has the excellent algorithm performances from WAM and TSM. The detailed study is published in the international journal "Renewable Energy".(2) Some neural networks models have been buloit for the real non-stationary wind speed signals to get the multi-step ahead predictions. But the traditional BP neural networks has some dis-advantages, such as it is difficult to determine initial weights and thresholds, network training process is too long and prediction accuracy is not high enough. Aimed at those issues, genetic algorithm and Pandit-Wu modeling algorithm of time series method is introduced, and genetic-neural networks method (GNM for short) and neural networks structure determination method (NNSDM for short) has been proposed. In addition, a new optimization method named wavelet-genetic-BP method (WGBM for short) has been presented based on GNM and WAM. From the prediction cases, it can be found that:WGBM has all the excellent algorithm performances from WAM, GNM and BP, which can be used in the conditions of prediction for multi-center signals. The detailed study will be published in the international journal "Information Science ".(3) In order to obtain super-high precision single-step ahead forecasting for non-stationary wind speed series, a new prediction method named wavelet-kalman method (WKM for short) has been proposed based on WAM and kalman filtering method (KFM for short). From the prediction cases, it can be found that:WTSM has both the excellent algorithm performances from WAM and KFM, which can attain the ultra-precision single-step ahead forecasting results. The detailed study is published in "Journal of Chinese Society for Electrical Engineering" and "Journal of Central South University of Technology (English Edition)".To further check the performance, the proposed algorithms have been applied to Chinese typical high-altitude strong-wind railway-the Qinghai-Tibet Railway. In this study, modeling and forecasting work for wind speed series from the key regions has been done. And the results show that the optimization algorithms are correct and feasible. In addition, the prediction mechanisms of non-stationary random signals are similar, so those new proposed algorithms can be extended to forecasting fields, such as wind speed from wind farms, robot path tracking, mechanical vibration signal processing, etc. In a word, this study is both academic and practical.更多还原