Most weather services around the world prepare their forecasts based on a single global model (either US or European) and then apply a local model such as WRF to the data. No matter how accurate the local model, its limits are determined by the accuracy of the global model. We have chosen a different approach to eliminate most of the potential errors.
ML combination of global models
First, we use our vast database of historical weather and all past forecasts of available global weather models to create a single output that minimises the errors and biases of individual models. This is achieved using machine learning algorithms that evaluate past errors in different meteorological situations and locations. As a result, we obtain a global forecast of the most probable weather scenario based on all the models available. Such output is computed 4 times each day.
Our detailed numerical weather prediction model
Using blended output, we make use of Meteosource’s regional models which increase resolution and take account of local conditions and geography. A meteorologist on duty can also adjust some of the inputs at this stage based on his or her experience.
AI model adjustments for cities
For places with measurements from weather stations - usually towns/cities and more populous areas - we run another level of AI adjustments to consider special weather patterns for a given location and weather situation.
Nowcasting in real-time
Weather is constantly changing and even the best model outputs might need to be adjusted before the next run of global models is available. Using all the available measurements (from stations, radar, etc.) for a given location, we update our forecasts in real-time by adjusting our outputs based on the latest information. Therefore, we can bring you the most up-to-date weather forecast available- hour-by-hour and minute-by-minute.
