 |
|
 |
|
|
First NOAA Workshop on Artificial Intelligence Using neural nets to improve week 3-4 precipitation and 2 meter temperature forecast |
|
|
April 25, 2019 The 1st NOAA Workshop on Leveraging AI in the Exploitation of Satellite Earth Observations & Numerical Weather Prediction was held in National Center for Weather and Climate Prediction, College Park, MD on April 23-25, gathering more than 400 scientists, program managers, and leaders from the public, academic and commercial sectors to exchange ideas, share lessons learned, and to discuss both the future potential and limitations of AI. Being selected by the workshop organizing committee, Dr. Yun Fan of Climate Prediction Center gave a presentation on the work of using artificial neural networks (NN) to improve the Climate Forecast System (CFS) week 3-4 precipitation and 2 meter temperature (T2m) forecast in the Post-Forecast and Extreme Weather session. Taking the advantage of flexibility of NN architectures that can handle big dataset and account for nonlinear impact, pattern relationship and co-variability among predictors and predictands, his preliminary results showed great potential of NN to improve week 3-4 forecast skill over CFS and multiple linear regression (MLR). As presented in Fig. 1, some CFS forecasts and corresponding observations are out of phase, which are reversed by NN, but not by MLR, prompting the value of NN. More information about the motivation, the NN basic as well as additional variables and cases explored are detailed in his presentation, which is available at https://www.star.nesdis.noaa.gov/star/documents/meetings/2019AI/Thursday/ |
|
|
Fig. 1 Time series (Jan 2017 - Oct 2018) of spatial anomaly correlation between observation and week 3-4 T2m forecast by NN, MLR, and CFS indicated by red solid, green dotted and black dotted line, respectively. Each four-panel inset shows the anomaly of forecast by CFS (lower left), MLR (upper right) and NN (lower right) in comparison with that of observation (upper left) at a specific time. |
|
North American Multi-Model Ensemble Teleconference Assessment of variances and NAD-NorCal precipitation relationship in NMME forecast April 4, 2019 Drs. Huug van den Dool and Emily Becker of Climate Prediction Center (CPC) gave a joint presentation on the topics of 1) EOF patterns in 200 hPa height in NMME models and observation, and 2) NMME representation of stationary wave pattern and the “North American Dipole index” in the April NMME teleconference. |
|
|
Principal Scientist Emeritus Huug van den Dool studied the time series of “all” model members (many realizations) versus that of observation (a single realization) to explain how using the ensemble mean as input can result in a simpler time series and lead to forecast overconfidence. His EOF analysis of 200 hPa height revealed that the 1st mode was AO-like for both observation and all models, while the explained variance was 30% in observation and a large range from 20% to 40% among “all” model members. It was also demonstrated that nearly 50% (ranging from 37% to 58%) of the variance, when using ensemble means, was explained by the 1st EOF mode (one degree of freedom). The forecast based on the ensemble mean could be too simple when a lot of the “noise” variance being eliminated. Dr. van den Dool also showed the trends appear mostly in EOF2, if not EOF1, with spatially uniform maps and mainly up or down time series for all models calculated from ensemble means. CPC NMME Lead Dr. Becker explored the relationship between northern California (NorCal) precipitation and “North American Dipole” (NAD), a peak/trough of stationary wave pattern revealed in an early work by Wang et al. (2014). Her study demonstrated the relationship between DJF NAD index and NorCal precipitation was fairly strong (r=-0.55, Fig. 1 top panel) in observation and even stronger (r=-0.88) in NMME ensemble mean forecast but weaker (r=-0.16, Fig. 1 bottom panel) between forecasted NAD index and observed NorCal precipitation, though the former has some relationship with observed NAD index (r=0.36). |
Fig. 1 Scatter plot of observed NorCal precipitation with observed (top) and forecasted (bottom) NAD index, respectively. |
|
Reference Wang, S.-Y., L. Hipps, R. R. Gillies, and J.-H. Yoon, 2014: Probable causes of the abnormal ridge accompanying the 2013-14 California drought: ENSO precursor and anthropogenic warming footprint. Geophys. Res. Lett., doi: 10.1002/2014GL059748 |
|
