2018Vyčislitelʹnye metody i programmirovanieOpen access

Comparison of data assimilation methods based on the classical, ensemble and local Kalman filter by the example of the advection equation and Lorenz system

Д.А. Ростилов, Maxim Kaurkin, Р.А. Ибраев

Open full text 2 citations

Abstract

Статья посвящена сравнению трех методов усвоения данных наблюденй: фильтр Калмана (Kalman Filter, KF), ансамблевый фильтр Калмана (Ensemble Kalman Filter, EnKF) и локальный фильтр Калмана (Local Kalman Filter, LKF). Выполнены численные эксперименты по усвоению синтетических данных этими методами в двух разных моделях, описываемых системами дифференциальных уравнений. Первая описывается одномерным линейным уравнением адвекции, а вторая - системой Лоренца. Проведено сравнение средних ошибок и времени исполнения этих методов при различных размерах модели, которые согласуются с теоретическим оценками. Показано, что вычислительная сложность ансамблевого и локального фильтров Калмана растет линейно с увеличением размера модели, в то время как у первого метода эта сложность растет со скоростью куба. Рассмотрена эффективность одной из возможных параллельных реализаций локального фильтра Калмана. The paper is devoted to the comparison of three data assimilation methods: the Kalman Filter (Kalman Filter, KF), the ensemble Kalman Filter (EnKF), and the local Kalman Filter (LKF). A number of numerical experiments on data assimilation by these methods are performed on two different models described by systems of differential equations. The first one is a simple one-dimensional linear equation of advection and the second one is the Lorenz system. The mean errors and the execution time of these assimilation methods are compared for different model sizes. The numerical results are consistent with the theoretical estimates. It is shown that the computational complexity of local and ensemble Kalman filters grows linearly with the size of the model, whereas in the classical Kalman Filter this complexity increases according to the cubic law. The efficiency of parallel implementation of the local Kalman filter is considered.

Open-access reader

About this research paper

What this paper is about

Статья посвящена сравнению трех методов усвоения данных наблюденй: фильтр Калмана (Kalman Filter, KF), ансамблевый фильтр Калмана (Ensemble Kalman Filter, EnKF) и локальный фильтр Калмана (Local Kalman Filter, LKF). Выполнены численные эксперименты по усвоению синтетических данных этими методами в двух разных моделях, описываемых системами дифференциальных уравнений. Первая описывается одномерным линейным уравнением адвекции, а вторая - системой Лоренца. Проведено сравнение средних ошибок и времени исполнения этих методов при различных размерах модели, которые согласуются с теоретическим оценками. Показано, что вычислительная сложность ансамблевого и локального фильтров Калмана растет линейно с увеличением размера модели, в то время как у первого метода эта сложность растет со скоростью куба. Рассмотрена эффективность одной из возможных параллельных реализаций локального фильтра Калмана. The paper is devoted to the comparison of three data assimilation methods: the Kalman Filter (Kalman Filter, KF), the ensemble Kalman Filter (EnKF), and the local Kalman Filter (LKF). A number of numerical experiments on data assimilation by these methods are performed on two different models described by systems of differential equations. The first one is a simple one-dimensional linear equation of advection and the second one is the Lorenz system. The mean errors and the execution time of these assimilation methods are compared for different model sizes. The numerical results are consistent with the theoretical estimates. It is shown that the computational complexity of local and ensemble Kalman filters grows linearly with the size of the model, whereas in the classical Kalman Filter this complexity increases according to the cubic law. The efficiency of parallel implementation of the local Kalman filter is considered.

Why it matters

OpenAlex reports 2 citations for this work. Citation counts describe recorded attention and do not establish research quality.

Key contribution

A contribution statement is not available in the OpenAlex record.

Method / approach

Method details are not available in the OpenAlex metadata.

Main findings

Findings are not separately available in the OpenAlex metadata.

Limitations

Limitations are not available in the OpenAlex metadata.

Applications

Application details are not available in the OpenAlex metadata.

Available abstract

Статья посвящена сравнению трех методов усвоения данных наблюденй: фильтр Калмана (Kalman Filter, KF), ансамблевый фильтр Калмана (Ensemble Kalman Filter, EnKF) и локальный фильтр Калмана (Local Kalman Filter, LKF). Выполнены численные эксперименты по усвоению синтетических данных этими методами в двух разных моделях, описываемых системами дифференциальных уравнений. Первая описывается одномерным линейным уравнением адвекции, а вторая - системой Лоренца. Проведено сравнение средних ошибок и времени исполнения этих методов при различных размерах модели, которые согласуются с теоретическим оценками. Показано, что вычислительная сложность ансамблевого и локального фильтров Калмана растет линейно с увеличением размера модели, в то время как у первого метода эта сложность растет со скоростью куба. Рассмотрена эффективность одной из возможных параллельных реализаций локального фильтра Калмана. The paper is devoted to the comparison of three data assimilation methods: the Kalman Filter (Kalman Filter, KF), the ensemble Kalman Filter (EnKF), and the local Kalman Filter (LKF). A number of numerical experiments on data assimilation by these methods are performed on two different models described by systems of differential equations. The first one is a simple one-dimensional linear equation of advection and the second one is the Lorenz system. The mean errors and the execution time of these assimilation methods are compared for different model sizes. The numerical results are consistent with the theoretical estimates. It is shown that the computational complexity of local and ensemble Kalman filters grows linearly with the size of the model, whereas in the classical Kalman Filter this complexity increases according to the cubic law. The efficiency of parallel implementation of the local Kalman filter is considered.

Key concepts: Ensemble Kalman filter, Alpha beta filter, Data assimilation, Kalman filter, Fast Kalman filter, Invariant extended Kalman filter, Extended Kalman filter, Control theory (sociology)

Related papers

Back to paper searchBrowse research topicsOriginal source
Comparison of data assimilation methods based on the classical, ensemble and local Kalman filter by the example of the advection equation and Lorenz system — Research Paper | ScholarLens