چكيده به لاتين
Abstract
The earth’s atmosphere is a complex system and its short range forecasting is an open and difficult problem. Almost all proposed systems use numerical models for weather forecasting. These models use a set of partially differential equations and apply them to a heterogeneous coordinate system that partitions the under study region into a 3D grid of similar cells. Complexity and precision of models are directly affected by the partially differential equations. Therefore, once there is not enough knowledge about the system, constructing the prediction model becomes impossible or the precision of constructed model is reduces.
In this thesis, a new scalable and data-driven framework is proposed for short-range weather forecasting. This framework has no need for analytical knowledge about the atmosphere and is constructed based on a data history that describes state of the atmosphere in the past. In addition it can control the trade-off between speed and accuracy. In other words, it can forecast slow and accurate or it can run fast and inaccurate.
The proposed framework, includes a global model that consists of a set of small and local models. Each local model is used to forecast a parameter in a specific point. Each local model should map from all potentially effective parameters in the neighborhood of the point to the corresponding parameter. Considering the huge number of potentially effective parameters, each local model has a feature selection module that filters parameters and significantly reduces them. In addition to feature selection module, each local model has a regression ensemble module that maps selected parameters in to target parameters. The output of each local model is a part of the whole state of the system.
In this thesis, proposing a new framework for atmosphere modeling, several improved methods are proposed for feature selection and regression ensemble that have superior efficiency compared to similar reported methods. The proposed framework is applied to the standard NCEP dataset from 1999 to 2010. The implementation results show that the proposed method is not only scalable but also its forecasting precision is comparable to well-known numerical weather forecasting systems such as GFS.
Keywords:
Numerical Weather Prediction, Data Assimilation, Regression Ensemble, Feature Selection.
