Source: Pubdate:2019-09-26 Hits:89
Low frequency component forecast models for the extended range
Yang Qiuming Jiangsu Meteorological Science Institute Research Group of Extended Range Weather Forecast
With the development of many observation methods, including satellite remote sensing, various global data of climatological observation obtained are increasing in recent years. These scientific big data with high data dependency and multiple data property reflect and reflect complicated natural phenomena and relationship[1,2]. By adopting data decomposition, extension and conversion, and other technologies, partial valid data are extracted from the large amount of data; low frequency variability information of middle latitude area of two hemispheres more comprehensive than sampling analysis can be obtained, providing better development foundation for extreme weather event forecast for 10-30 days extended range and forming seamless operational forecast system from weather to climate[3,4]. The major global ISO model closely related to extreme weather event of specific area (for example Yangtze River valley) can be extracted from a large amount of multivariable observation data with long number sequence and high coupling, reducing the system complexity to drive the complicated low frequency variability process and system construction by dynamic data so as to extend the forecast lead time of regional extreme weather process remarkably. Yang Qiuming (2014) established a series of simplified time-varying linear forecast models (such as MLR, ECAR, MLR/PC-CAR), obviously extending the forecast lead time, able to forecast following 40~50 d rainfall and 20~30 d low frequency component variation of the lower reaches of Yangtze River. Among them, low frequency component forecast model (LFCF) based on independent regression equations for MLR forecast model is called LFCF1.0[5]; ECAR forecast model based on complex number autoregressive recursion is called LFCF2.0[6-8]; MLR/PC-CAR forecast model through time coupling of independent regression equations and complex number autoregressive recursion is called LFCF3.0[9]. Based on the big data, these extended range forecast methods are superior to that of small data age, such as ensemble numerical prediction model and classical statistics model, in lead time and stability. It relates to the extended range forecast theoretical research and practice application to improve the forecast accuracy remarkably through the dynamic data update.
Big data forecast completely depends on the source of big data while the big data model needs to confirm through searching the corresponding relationship between mass data and algorithm, specific data calculation without abstract process rather than classical theoretical model and uncertain or random numerical calculation model and classical statistics model. Its features and value can reflect in three dimensions: Extendable capacity, diversified type, dynamic accumulation.
The greatest value of big data lies in that the possibility of finding the principles or incidence relation will be increased through constant data overlay and the forecast accuracy will be remarkably improved with the increase of observation information increment through multiple observation data. Taking full use of meteorological scientific big data resources, to research and find the multiple different time scale (classical) statistical law of these new general atmospheric circulation in different views can get more small data variation signal of representation extended range extreme weather variation.
The consideration of big data in a certain aspect and the analysis on small data can reveal the possible contact and variation between big data and small data. Under the related framework of global teleconnection, the analysis on different ISO models and interannual and interdecadal variations in East Asia and on global sea temperature, snow cover and other external forcing and the time-varying relation of the intensity of ISO variation can get more new principles of relationship and variations between ISO of extreme weather in different areas and different time scales.
On the basis of multiscale and multilayer information, global atmosphere teleconnection (partial physical mechanisms are unknown or cannot be expressed in classical theory or there are mutual effects among different components and the causal relationship is diversified), taking global history and real-time observation big data as the sample, data-driven simplified time-dependent extended range forecast models LFCF1.0(MLR), LFCF2.0 (ECAR) and LFCF3.0 (MLR/PC-CAR) are established and the extended range forecast accuracy and lead time are superior to dynamic models or statistic models in traditional small data age and the advancement cannot be compared by the classical way.
Meteorological big data have such features as great scale, high value, interleaving holographic visibility, not only progress in the amount of greater scale data but also in qualitative leap of data organization and application methods different to the previous. The increase of value should be proportion to that of scale and the real-time application need to be as simple as the operation of small data. This requires very good data organization, clear and obvious in the content and relevance, and natural change properties in big data reflected objectively, able to get the data logic segment for easy analysis and treatment. Different data inevitably include forecast information of different time scale and different critical data are remarkably different in the indicating meaning of extreme weather in different years and specific area.
Therefore, making flexible use of data and analyzing core data of own limitation and its variability are also important. Thinking out of established frame, starting from multiple views and thinking out of the choices of 0 or 1 and the third or more selections could be achieved. Data application is small and beautiful (the objective is simple and specific) rather than big and complete. Weakening non-core data, analyzing data priority for the extreme weather, connection between data and taking full use of data externality can generate huge value much greater than simple sum through data crossing fusion.
Activating data, changing static data into dynamic data and constantly updating critical data to adapt to related new normality of related relationship change, forming a brand-new digital data environment. 30% depends on technology and 70% on data. Analyzing data are major driving force to improve over 10 days weather forecast accuracy.
Knowledge acquired from data are also new data and new knowledge can be acquired from new data. Going forward in this way, dynamic changed knowledge system can be constituted. Data are both resources and results and more useful data will be acquired by using data. Establishing data about data and extracting and purifying data in practice can extend the lead time of extreme weather events forecast remarkably to increase new driving force for innovation and development of climatic prediction methods.
(2019.9.26)
References:
[1] Sue N. Big data: The Harvard computers. Nature, 2008, 455: 36–37.
[2] Overpeck J T, Meehl G A, Bony S, et al. Dealing with data: Climate data challenges in the 21st century. Science, 2011, 331: 700–702.
[3]Hoskins B. The potential for skill across the range of the seamless weather-climate prediction problem: a stimulus for our science. Quarterly Journal of the Royal Meteorological Society,2013,139:573-584.