92nd American Meteorological Society Annual Meeting (January 22-26, 2012)

Wednesday, 25 January 2012
Multi-Model Ensemble Forecasting in the Energy Weather Industry
Hall E (New Orleans Convention Center )
Dr. Daniel B. Kirk-Davidoff, MDA Information Systems, Inc (EarthSat Wx), Gaithersburg, MD; and C. C. Cassidy, R. J. Much, and L. J. Heitkemper

MDA Information Systems has provided weather forecast services to the energy and agricultural sectors for over thirty years. We forecast for hundreds of locations in the United States and worldwide. In the early days, this involved meteorologists looking at output from a range of forecast models, gaining experience about the biases of particular models in particular parts of the world, and using that experience to adjust high and low temperature at each city. Over the years, we have made efforts to, in Ursula LeGuin's words, “domesticate the hunch” , systematizing this accumulation of experience about model bias in the effort to improve forecast skill.

Our original hourly forecasting system integrated a limited set of 3 domestic MOS models and 1 global model. The combination of these various model inputs was simplistic and therefore prone to discontinuities in the flow. As a result, the system was heavily reliant on operational forecasters to smooth the flow at the junction points of various models. Operational forecasters contributed much of their own knowledge about each model and the individual biases to come up with a final forecast. This method was not easily scalable, particularly outside of the United States, labor intensive, and was unable to take advantage of newer model data sources.

Using the literature on multi-model ensemble forecasting (e.g. Krishnamurti et al. 2000) and our own experience we determined that a dynamically weighted ensemble approach would allow us to generate a more accurate first guess forecast. The new hourly forecasting system ingests a total of 10 domestic and international models and uses various statistical techniques to remove station-level biases and weight the models according to their recent skill. The statistical data is continually updated as new model data arrives, providing the most recent forecast information possible. The first guess forecast calibration takes place over an adjustable training period and is designed to throw out anomalous observations that would pollute the statistical analysis.

Compared to the original hourly forecast system, we have seen an average reduction of 0.16 °C mean absolute error (MAE) in temperature forecasts (nearly 10%) when verifying the 72-hour forecast for a set of 24 domestic stations during a 9 month period. The improved first guess forecast allows the operational forecasters to focus their labor more efficiently by concentrating on short-term forecast issues (timing of fronts, precipitation, etc.) that the system may have trouble identifying precisely. We have also designed a variety of tools to monitor the real-time status of the data flow into the system as well as track the verification of historical forecasts. The system can generate reports that compare the forecasts between all available models to examine where the system may be struggling and allows us to make adjustments when necessary. This also allows the operational forecasters to identify stations which may need to be watched and edited more frequently.


Krishnmurti, T.N., C.M. Kishwawal, Z. Zhang, T. LaRow, D. Bachiochi, E. Williford, 2000: Multimodel Ensemble Forecasts for Weather and Seasonal Climate. J. Climate 13:4196-4216.

Supplementary URL: