Forecasting Weather

Our system takes data of the GFS model of the national US weather service as input. We calculate the correlation between model variables, which represent the atmosphere in a 3 dimensional box, and what happened at a local point on earth's surface in real. This statistical approach is widely called Model Output Statistic (MOS). MOS assigns model data to local data. To make MOS succesful, it is necessary, to create dynamic functions, which downscale large scale meteorological processes to local stamps, the customer wants to know. We do so by using our long term meteorological experience, numerical experiments and statistical methods. We have created various functions of model, geographical and astronomical variables. Thus, our MOS, other than conventianal MOS systems, assigns functions of model data to local data. Most customer requested data are not part of any atmospheric forecast model, which only calculates physical quantities. If only for that reason one needs model plus statistic.

Our MOS system also enables us, to make forecasts for any point on earth's surface. Our weather forecasts are not limited to weather stations. Any point on earth's surface can be characterized by a set of parameters, which represents the static geographical conditions in real and in the atmospheric model. This parameter set, we have calculated for all about 8.000 points on earth's surface, where a weather station is located. For any given point it is now easy, to find the most similar point among these 8.000 parameter sets. There is another important reason, to use model plus statistic instead of just a model. Any weather forecast has to be an estimation of probabilities as the atmosphere is a chaotic system. A good weather forecast gives the probability in which range maximum temperature will lie tomorrow and not a distinct value. With MOS, one can calculate these probabilities. With just one model run realization, it is impossible to calculate any probabilities.

Our MOS keeps learning. We update our MOS equations daily. Although only a forecast of probabilities is a good forecast: When time has gone, the forecast should exactly meet what was observed in real. Even more: A good forecast for the next hours should instantaneously take into account what happened hours before. This is one of our current developments: To put current observations into the MOS equations for the short range forecast. This is also a lack of a pure model forecast. Local observations cannot be taken into account, they were smoothed away, when the model is initialized, to avoid unstable model runs.

How does this adaptation to observational data works at points besides any weather station? The amount of 10.000 weather stations worldwide together with the knowledge about the behaviour of meteorological parameters in dependence of changing geographical conditions enable us to analyse an artificial field of observational data. With this analysed field of current weather conditions, we can adapt the forecast at any point, where a forecast is requested for. Even more: The analysed field leads to better results than trusting the pure observed data. By analysing the observed data, the errors in observational reports can be filtered out. Latest at this point, it gets evident, that a provider of weather forecast also has to be a provider of observational data.