MSPA-CSE: Statistical Methods for Precipitation Nowcasting and Verification

本项目旨在开发一种基于贝叶斯分层建模框架内相关大气动力学的多分辨率统计表示的强降水短期预报新方法。主要输入数据将是来自气象雷达的反射数据，降水演变的动力学将被表示为一个低维随机过程。其中一组研究问题包括隐藏动力学与降水场之间映射的性质，决定隐藏动力学演变及其与大气物理关系的传播因子，以及映射和传播因子函数对气象变量的参数依赖性。将测试隐藏低维过程与多尺度降水场之间的各种以气象参数为条件的线性和非线性图。传播器将由准平稳一阶马尔可夫过程近似，其中有限的时间依赖性表示为对气象状态的参数依赖性。该模型将应用于在伊利诺伊州、俄亥俄州和印第安纳州的THOR观测项目中收集的雷达和其他气象数据，并对其进行测试。该方法将与现有预测工具对该地区的短期预测进行比较。在项目的后期，该方法将扩展到使用顺序蒙特卡罗技术的顺序预测算法的开发。第二组研究问题涉及有效核查技术的发展。将探讨在数据和预报中自动识别风暴单体的方法，并对它们的形状、位置和强度进行逐单体比较。最后，该项目将包括对风暴细胞预测的假设检验技术的研究。所开发的方法可以改进洪水预报和对这些预报的不确定性进行定量估计，从而有利于应急服务。

This project seeks to develop a new approach to short-term forecasts of heavy precipitation based on a multi-resolution statistical representation of the relevant atmospheric dynamics incorporated within a Bayesian hierarchical modeling framework. The primary input data will be reflection data from weather radars and the dynamics of the evolution of the precipitation will be represented as a low-dimensional stochastic process. One set of research questions includes the nature of the map between the hidden dynamics and the precipitation field, of the propagator determining the evolution of the hidden dynamics and its relation to atmospheric physics, and of the parametric dependence of the map and propagator functions on meteorological variables. A variety of linear and nonlinear maps, conditioned on meteorological parameters, between the hidden low-dimensional process and the multi-scale precipitation field will be tested. The propagator will be approximated by a quasi-stationary first-order Markov process where the limited time dependence is represented as a parametric dependence on the meteorological regime. The model will be applied to and tested with radar and other meteorological data collected during the THOR observing project over Illinois, Ohio and Indiana. The method will be compared with short-term forecasts for this region obtained from existing prognostic tools. Later in the project, the approach will be extended to the development of a sequential prediction algorithm using sequential Monte Carlo techniques. A second set of research questions relate to the development of effective verification techniques. Methods for automatically identifying storm cells in data and forecasts and making cell-by-cell comparisons of their shape, location and intensity will be explored. Finally the project will include research into hypothesis testing techniques for storm cell predictions. The methodology developed could lead to improvements in flood forecasting and a quantitative estimate of the uncertainty in these forecasts with consequent benefit to emergency services.