Fine-Scale Seasonal Climate Prediction

BACKGROUND

Figure 1. The framework for fine-scale seasonal climate prediction
Figure 1. The framework for fine-scale seasonal climate prediction

Global seasonal climate predictions at about 100-200-km resolution issued by national climate centers provide reliable perspectives of the general circulation conditions about one month to six months in advance. Such forecasts, however, lack the fine-scale details that are critical to regional and local climate-sensitive business and decision-makers. To fill that gap, we are developing a fine-scale seasonal-climate prediction capability through dynamical downscaling. A framework for fine-scale seasonal-climate prediction has been set up. In this framework, the global large-scale seasonal forecasts issued by NCEP’s Climate Forecast System version 2 (CFSv2) are applied to force the Weather Research and Forecasting (WRF) model at fine scales. The version of the WRF model we use has been specially customized and configured for climate purpose.  Both deterministic and ensemble predictions can be performed. Techniques from artificial intelligence such as principal component analysis (PCA) and self-organizing map (SOM) analysis are used to extract the relevant climate information.

We are relying on the CFSv2 operational run outputs for this task. For CFSv2, the real-time analysis is carried out by an atmospheric model at T574 (roughly 27-km grid spacing) in the horizontal and 64 sigma-pressure hybrid levels in the vertical. The real-time forecasts are carried out by the same atmospheric model but at T126 horizontal resolution (roughly 100-km grid spacing) and 64 sigma-pressure hybrid vertical levels. The CFSv2 system includes the interactive Noah land surface model with 4 soil levels, the interactive Modular Ocean Model version 3 (MOM3), and the interactive 3-layer GFDL (Geophysical Fluid Dynamics Laboratory) Sea Ice Simulator sea ice model. A global ocean data assimilation system (GODAS) provides the ocean initial conditions for the CFSv2 analysis and forecasts. The real-time CFSv2 outputs include 9-month forecasts (only 7-month forecasts are available) initialized at 00Z, 06Z, 12Z, and 18Z of each day as shown in Figure 2, as well as a single one-season forecast initialized at 00Z and three 45-day forecasts initialized at 06Z, 12Z and 18Z of each day. We will utilize the 7-month forecasts.

Figure 2. Schematic of the operational daily CFSv2 configuration. Only 7-moth data are available daily for downloading from the NCEP NOMADS webpage.
Figure 2. Schematic of the operational daily CFSv2 configuration. Only 7-moth data are available daily for downloading from the NCEP NOMADS webpage.

Accomplishments in FY2019

In August 2018 we started downloading and archiving the 4-times-daily 7-month CFSv2 forecasts. The downloaded files include the surface files and the flux files initialized at 00Z, 06Z, 12Z, and 18Z each day. The downloaded data files are being used for configuring and tuning the seasonal climate prediction system and for analyzing the climate simulations. We can also examine the climatic conditions out to 7 months for the entire period for which data are available.

We evaluated the CFSv2 real-time forecasts focused on the CONUS domain using the analysis fields as truth. The evaluated variables include 2-m temperature and 10-m winds initially. The forecast lead times ranged from one month to two months, from one season to two seasons. Both spatial distributions and temporal evolution were examined. We have done sensitivity experiments with 4, 8, 12, 16, 20, 24, 28, and 32 ensemble members for determining the optimum number for use in sub-seasonal to seasonal climate forecasts. The ensemble members were constructed by using the CFSv2 forecasts for different cycles. We examined the ensemble mean, bias, standard deviation, and signal-to-noise ratio, and determined on the basis of bias, ensemble spread, and computing resources that 16 ensemble members is optimal for our application. Figure 3 shows the construction of the 16 ensemble members based on the 4-times-daily cycles for 4 days (i.e., 4 x 4 = 16). Figure 4 shows the 16-ensemble member mean, bias, standard deviation, and signal-to-noise ratio of 3-month simulated 2-m temperature.

Figure 3. Schematic of the setup of 16 ensemble members for sub-seasonal to seasonal climate forecasts.
Figure 3. Schematic of the setup of 16 ensemble members for sub-seasonal to seasonal climate forecasts.
Figure 4. Ensemble means (a), bias (b), standard deviation (c) and signal-to-noise ratio (d) of 3-month simulated 2-m temperature for 16 ensemble members.
Figure 4. Ensemble mean (a), bias (b), standard deviation (c), and signal-to-noise ratio (d) of 3-month simulated 2-m temperature for 16 ensemble members.

We have set up the experimental fine-scale seasonal climate forecasting system using WRF driven by the CFSv2 forecasts. The WRF domains were configured with a 30 km -> 10 km -> 3.3 km -> 1.1 km nested hierarchy. We have also downloaded and archived the observations from the NCEP Meteorological Assimilation Data Ingest System (MADIS) for the same time period for conducting point-to-point comparisons of 2-m temperature and relatively humidity, 10-m wind speed and direction, precipitation, surface pressure, and PBLH (planetary boundary layer height). Both low-level winds and PBLH are especially important for pollutant dispersion and transport. PBLH could be computed from the atmospheric soundings.

PLANS for FY2020

We are setting up the fine-scale seasonal forecasting system driven by the latest reanalysis from ECMWF (ERA5) as well as CFSR. The purpose of this work is three-fold: 1) evaluating the performance of the ERA5- and CFSR-driven WRF simulations at fine scales, 2) evaluating the CFSv2-driven WRF simulations using the reanalysis-driven data, and 3) exploring the possibility of increasing the ensemble spread by including ensemble members from different modeling systems. Figure 5 shows the ERA5-driven and CFSR-driven WRF 2-m temperature simulations at 1-km horizontal resolution for January 2019. Clearly, ERA5-driven simulations are colder than the CFSR-driven simulations. We will use the MADIS observations to evaluate the simulations.

Figure 5. ERA5-driven (left) and CFSR-driven (right) WRF 2-m temperature simulations at 1-km resolution valid for January 2019. The model domain centered on Utah.
Figure 5. ERA5-driven (left) and CFSR-driven (right) WRF 2-m temperature simulations at 1-km resolution valid for January 2019. The model domain centered on Utah.

We will extensively evaluate the CFSv2-driven WRF simulations at fine scales (e.g., 3.3 km and 1.1 km horizontal resolutions) using the MADIS observations for varying lead times and for various variables. We will compute bias, correlations, and skill scores. We will conduct the SOM analyses and compare with the reanalysis-driven SOM patterns. We will explore the connection between the climate variability (e.g., ENSO, NAO, monsoon) with local weather centered over Utah.

We will explore the application of bias correction to the CFSv2-driven simulations based on our analysis results. Finally, we will set up the operational daily run of sub-seasonal to seasonal forecasts for Utah.