Presentation is loading. Please wait.

Presentation is loading. Please wait.

Donat Perler, MeteoSwiss Joint work with Oliver Marchand, MeteoSwiss

Similar presentations


Presentation on theme: "Donat Perler, MeteoSwiss Joint work with Oliver Marchand, MeteoSwiss"— Presentation transcript:

1 Donat Perler, MeteoSwiss Joint work with Oliver Marchand, MeteoSwiss
Automatic weather interpretation using modern classification algorithms Donat Perler, MeteoSwiss Joint work with Oliver Marchand, MeteoSwiss

2 Overview: The aLMo Model
Model features: Non-hydrostatic Prognostic variables (u, v, w, T, p´, qv, qc, qi) aLMo7 (today) aLMo2 (planned) 2 times per day, +72h 8 times per day, +18h 358 x 325 x 45 grid points 480 x 350 x 45 grid points 7km horizontal grid scale 2km horizontal grid scale COSMO Meeting, Date, Zürich

3 Result: Output of aLMo and its interpretation
2000 plots per simulation COSMO Meeting, Date, Zürich

4 Operational Weather Forecasting
COSMO Meeting, Date, Zürich ?Thunderstorm risk?  ?Thunderstorm alert?

5 Automatic Weather Interpretation at DWD
Pseudo Code for DWD Expert System: […] if ((STABILITY_INDEX < -6) AND (P_CLOUD_BAS-P_CLOUD_TOP > 400hPa) AND (T_CLOUD_TOP < -45C) AND (PERCIP_CONV > 2.0mm/h)) then report(`Thunderstorm´) end COSMO Meeting, Date, Zürich

6 Output of the System used at DWD
Thunder- storm COSMO Meeting, Date, Zürich

7 Main Drawbacks of DWD`s Method
Rules are manually chosen Parameters/threshold values set by hand No measure for the certainty of classification given Questions How can the parameters be tuned more exact? Do all rules have the same importance? Are there other important rules? How does one find these rules? COSMO Meeting, Date, Zürich

8 Machine Learning use! Training set Actual data Neural Networks
Boosting SVM train! COSMO Meeting, Date, Zürich Output Classification Problem

9 Select appropriate model features
Steps in this work Select appropriate model features Principal Component Analysis Use DWD experience from Expert System Compare and evaluate different supervised machine learning algorithms Boosting Support Vector Machines Bayesian- and Neural Networks Introduce the best algorithm to the aLMo for operational use COSMO Meeting, Date, Zürich

10 Thank you for your attention! We very much welcome discussions.
COSMO Meeting, Date, Zürich


Download ppt "Donat Perler, MeteoSwiss Joint work with Oliver Marchand, MeteoSwiss"

Similar presentations


Ads by Google