Kalman Filter with Adaptive Noise Models for Statistical Post-Processing of Weather Forecasts
2017 (English)Independent thesis Advanced level (degree of Master (Two Years)), 20 credits / 30 HE credits
Student thesis
Abstract [en]
We develop Kalman filter with adaptive noise models for statistical post-processing of 2-metre temperature forecasts for the purpose of reducing the systematic errors that numerical weather prediction models usually suffer. For this, we propose time-varying dynamic linear models for the system noise covariance matrix and the measurement noise covariance matrix, and we study how that affects the mean predictions of the underlying state and the observed data. Five Kalman filter models are introduced, a discrete Kalman filter model with the distinctive feature that the measurement (observation) at time t is the observed forecast error at that time, two Kalman filter with adaptive noise models where the measurement noise covariance matrix is time-varying, a Kalman filter model where the forecasts of the 10-metre wind components are included as explanatory variables, and a Kalman filter with heavy-tailed noise using the Student’s t-distribution under a Bayesian approach. Ten weather stations located in Sweden are selected trying to obtain a heterogeneous sample and six different forecasts issued are filtered with different sets of initial values.
The implementation of these methods has been done in Python and R.
Place, publisher, year, edition, pages
2017. , p. 92
Keywords [en]
Adaptive Kalman Filter, Surface Temperature Forecast, Systematic Errors, Statistical Forecast Correction, Dynamic Linear Models, State-Space Models, Bayesian Forecasting.
National Category
Computer Sciences Probability Theory and Statistics Meteorology and Atmospheric Sciences
Identifiers
URN: urn:nbn:se:liu:diva-136291ISRN: LIU-IDA/STAT-A--17/001-SEOAI: oai:DiVA.org:liu-136291DiVA, id: diva2:1087371
Subject / course
Statistics
Presentation
2017-02-14, Grace Hopper, Linköping University, 10:15 (English)
Supervisors
Examiners
2017-04-072017-04-062025-02-01Bibliographically approved