基础架构,应用解决方案,云计算,大数据,数据中心,智慧IT,IT架构,人工智能相关的产品白皮书,技术方案,技术架构介绍,适合IT技术专业人士阅读参考
机器能预报多久后的天气?
——机器学习和可预报性
袁行远
彩云天气CEO
Part 1: Problem Description
Part 2: Precipitation Predictability
Part 3: Applications
Part 1: Methodology
/weather40d/101010100.shtml
Problem description
• Given the previous m frames of a video, predict the
next n frames
– It-m+1, … , It-1, It → It+1, It+2, ... , It+n
f ( , )=
predictability measure: decorrelation time
Eulerian persistence
Lagrangian persistence
Previous work
Bellon and Zawadzki (1994): 通过空间平均减少小尺度信息有利于提高大图
预报准确度
Seed and Keenan (2001): 使用Fourier-notch-filter进行空间尺度分解,然
后各个尺度平流
各自的decorrelation time
Zawadzki (1994)将可预报性与大尺度气象参数(比如CAPE,风暴相关的螺旋
度)联系起来
Pierce(2000) 认为风暴相关的螺旋度是指示提前预报时效的有效参数
Carbone (2002)观察到空间尺度1000km,时间持续度1天的持续性降水常有发
生。单站雷达的时空观测限制大,而大陆雷达拼图可以有效延长预报时效。
Data
分辨率:时间15min,空间2km,反射率5dBZ,
空间尺度:2720km*2720km
对流性降水多
Four events:
Variational echo tracking (确定流场)
cost function
The sum of squares of residuals of the conservation equation
其中前面乘的系数为描述数据质量的权重矢量
Smoothness-penalty function
其中前面乘的系数为常量权重
Advection of radar reflectivity (平流方案)
semi-Lagrangian
For each time step, the advection is divided up to N steps of length
Part 2: Predictability of Precipitation
Preliminary results
Preliminary results
Preliminary results
Scale-Dependence of predictability of precipitation
Different threshold of rain (intensity-size relationship)
Scale-Dependence of predictability of precipitation
Decompose radar images according to scales using discrete-cosine-transforms
and low-pass filter in the spectral domain
Part 3: Applications
•基于天气雷达, 结