Welcome to the NOAA Institutional Repository | Accurate and fast neural network emulations of long and short wave radiation for the NCEP Global Forecast System model - :6951 | Office of Oceanic and Atmospheric Research (OAR) | National Weather Service (NWS)
Stacks Logo
Advanced Search
Select up to three search categories and corresponding keywords using the fields to the right. Refer to the Help section for more detailed instructions.
 
 
Help
Clear All Simple Search
Advanced Search
Accurate and fast neural network emulations of long and short wave radiation for the NCEP Global Forecast System model
  • Published Date:
    2012
Filetype[PDF - 4.52 MB]


Details:
  • Corporate Authors:
    National Centers for Environmental Prediction (U.S.)
  • Series:
    Office note (National Centers for Environmental Prediction (U.S.)) ; 471
  • Document Type:
  • Description:
    In this study we used the neural network (NN) emulation approach applied earlier to NCAR Community Atmospheric Model (CAM) and NCEP Climate Forecast System (CFS) radiation to develop NN emulations of the full NCEP Global Forecast System (GFS) model radiation. NN emulations have been developed and tested for the original RRTMG long-wave radiation (LWR) and RRTMG short wave radiation (SWR) parameterizations, which together comprise the full model radiation for the GFS model. The results presented in this note show that the developed NN radiation is very accurate. Also, it is 20 (LWR) and 100 (SWR) times faster than the original LWR and SWR parameterizations. The NN radiation was tested in several parallel 8-day forecasts. During first four days of integration no significant differences between control and NN runs can be observed. The differences observed after four days of integration are small and, after four days of integration, the NN run often demonstrates slightly better performance (higher anomaly correlation, lower bias and RMS errors) than the control run. Comparisons with an additional GFS run using CFS NN radiation demonstrate robustness of the developed NN radiation with respect to changes in the model environment. This is a very important practical result, which shows that the NN radiation does not require frequent updates and may work in the model for many years without retraining. The high speed of NN radiation calculations can be used to: (1) significantly speedup (15-18%) the model integration; (2) calculate radiation more frequently than ones per hour (actually it can be calculated at each integration time step); (3) emulate more advanced and time consuming radiation parameterization (e.g., the newest RRTMMcICA). Also, the NN radiation helps to achieve a significantly better load balance. This study is the first, initial step in evaluating NN radiation in GFS. Further steps will include: (1) more comprehensive tests in a longer series of 10-day forecasts and (2) evaluation of the NN radiation in parallel runs with more frequent radiation calculations.

  • Supporting Files:
    No Additional Files