A true synergy between research, forecast technology, and decision support
provides unique and innovative forecast solutions
CFAN's weather and climate predictions, spanning time scales from days to decades, are based upon a sophisticated statistical/dynamical system that utilizes forecasts from multiple weather and climate modeling centers. CFAN adds considerable value to the weather and climate model outputs by increasing their information content through statistical adjustments, interpretation of the forecasts, and characterizing uncertainty of the forecasts.
The engine that drives CFAN’s prediction systems includes the following elements:
Selection of the best forecast models through comprehensive forecast model evaluation
Making the best models better through extensive statistical adjustments based upon model historical forecasts and recent model performance.
Analysis of large-scale weather regimes: their predictability and relationships to surface weather and extreme weather events.
Advanced forecast modules for business-relevant variables for the energy and agricultural sectors, plus severe weather events such as hurricanes and floods
Downscaling of forecasts for target locations, using topographic and land surface information
Advanced ensemble clustering techniques to extract information from the forecast ensembles
Objective confidence assessment for each forecast based on analysis of recent forecast performance relative to observations and historical predictability analyses.
Why ensemble forecasts?
Predictions of weather and climate variability are issued with lead times of hours, days, months, years and decades. However, every single forecast is to some extent uncertain. There are three main sources of uncertainty in weather and climate prediction:
Initial conditions uncertainty: weather observations are neither perfect nor complete. Our incomplete knowledge of the exact state of the atmosphere (the initial conditions) leads to uncertain forecasts right from the beginning.
Model simplifications: computational constraints and inadequate understanding inevitably lead to modelers to approximate the exact equations for weather and climate.
Nonlinearities and chaotic behavior of the atmosphere leads to fundamental uncertainties in prediction that varies with the state of the atmosphere.
Ensemble prediction methods have been developed to tackle the problem of these unavoidable uncertainties. Instead of making a single forecast of the most likely future weather, a set (or ensemble) of forecasts is produced. In ensemble prediction, a model is run multiple times with slightly varying initial conditions and model parameters. Ensembles can be created using multiple models, or multiple versions of the same model. Ensemble weather simulations, run operationally for the past two decades, aim to describe the potential realizations of the future system state and thus provide a path to estimate the probabilities of particular events. This set of forecasts aims to give an indication of the range of possible future states of the atmosphere.
In the figure below, a single deterministic forecast (indicated by the gold dot) can be rather far from the actual observed outcome (indicated by the purple dot). The objective of the probability forecast is to bound the observed outcome in a probability space (indicated by the darker blue region). The forecast is useful if the probability space is much smaller than climatology (indicated by the light blue region). The ensemble spread (size of the blue region) is related to the uncertainty of the forecast. One of the challenges is that model forecasts are biased (indicated by the green region). A forecast calibration scheme effectively moves the green region closer to the dark blue region, to realize the full predictability of the ensemble forecast. The biggest forecast challenge is to identify extreme events that lie outside of the recent climatology, such as indicated by the red dot.
What are the advantages of probability forecasts?
To support decision making, the predicted distribution of forecast outcomes needs to be interpreted statistically.
Probabilistic information from ensemble forecasts can help users make better decisions by accounting for uncertainty in the forecasts and assessing the probability of exceeding a threshold of vulnerability. Probability distributions of outcomes derived from an ensemble forecast support a utility theory approach for decision making, which takes into account the complexity of decision problems with identified options and multiple sources of information, risk, benefits, and uncertainty.
Probabilistic forecasts are contrasted with deterministic forecasts. A deterministic forecast is single-valued forecast, as represented by the orange dot in the above figure. A deterministic forecast may be created by using only one forecast realization, or by taking the average of the ensemble members to determine a single value. Most weather forecast providers provide a single forecast outcome (a deterministic forecast.) Some users prefer deterministic forecasters. Therefore, some of our forecasts are reported as a single, ensemble mean value, such as our 15 day city temperature forecast tables:
At the other extreme are decision makers who want to see the forecasts of each of the ensemble members, such as potential hurricane tracks.
The image above shows the forecasted hurricane tracks from the 51 ensemble members from the ECMWF Ensemble Forecast System. For extended-range predictions of hurricane tracks, this ensemble size is too small to accurately characterize the true uncertainty of the forecast. To better assess the track forecast uncertainty, The figure below shows track location probabilities for the same forecast, based on 1530 synthetic tracks that were generated using Monte Carlo techniques with the current forecast and historical errors from model hindcasts.
Uncertainty in the surface temperature forecast is illustrated graphically from the calibrated probability distribution of the multi-model forecast ensemble, see below for the probability forecast for surface temperature at Boston:
An additional way of portraying the distribution of forecast outcomes is to calculate the probability of exceedence of a particular threshold. The figure below illustrates the probability of cold waves for selected cities using thresholds provided by a client in the energy sector
An additional example of threshold exceedance probabilities is illustrated for precipitation.