Successful mega-update of MultiModel MOS delivers better forecasts
To ensure we remain one of the best weather companies in the world, improving the quality of our forecasts has constantly been at the top of our to-do list. We recently took another huge leap in that area with a large-scale update of our MultiModel MOS. “It was a very exciting process; we do not undertake such a complicated step towards improvement very often,” says Marco Berghege, meteorological services analyst at MeteoGroup. For the past six months, the Meteorological Services Team has been working on the implementation of this complex MOS update. Their goal: to ensure that the automatic forecasts issued by thousands of weather stations will continue to perform excellently in the next few years.
The clear reason: keep finetuning weather models
Every company needs to innovate to stay in the race, and that applies to MeteoGroup as well. We have to deal with the additional factor that the weather model data we purchase gradually gets out of sync with the statistical calculations that we unleash on it.
Marco Berghege: “The weather models that we work with are the ECMWF, EPS, UKMO and NCEP models. In addition to this data, we also have the observations from approximately 20,000 weather stations around the world at our disposal. By comparing weather model information and observations and building statistical equations on that, we eventually achieve better forecasts. That is a post-processing step that we developed in the past, which we call MultiModel MOS. MOS stands for Model Output Statistics, and the term ‘MultiModel’ refers to the four different weather models we use as the basis. But the formulas and algorithms we designed two to three years ago and have applied since then might not be entirely sufficient today. After all, the models we have purchased are continuing to innovate as well. For example, when using a specific weather model mix, we might always add two degrees to a certain area to arrive at the correct maximum temperature. However, over time these models have been adapted to more closely approximate the correct maximum value. If we’re still adding those two degrees as a correction factor, we’ll end up too high.”
In short, it was time for an extensive update. Not a minor adjustment, but a complete overhaul. It was a project that took a considerable amount of calculation, dedication, vigilance and patience.
Processing mindboggling amounts of weather data
The MultiModel MOS is based on a historic observation database covering two years. This database is essential in order to configure a specific MOS forecast for each suitable MOS weather station. For that purpose, each weather station has a unique weighting factor for the purchased weather models, so each weather station gets its own set of equations. A series of equations are applied for each weather element at each expected period of time in order to arrive at a weather forecast which is as reliable as possible. This eventually results in hourly forecasts consisting of approximately 40 parameters. This includes elements such as wind velocity, temperature, dew point, cloud cover, visibility and precipitation. Just consider how many algorithms it takes to generate these forecasts. Every single hour.
Not every MOS weather station is equipped to run a good MultiModel MOS. A station that would be eligible for inclusion in this update should have a historical database which is as complete as possible. Marco: “One single month is not a representative sample; that’s not enough data to base equations on. We ideally prefer to have two complete years.” The data also needs to be very high quality. After all, if you use incorrect input, it will never be possible to come up wityh the right results. In addition to observation data, the update also requires the model data. And again, we are dealing with mindboggling amounts of data; like everything in this project, it’s all about the big numbers. In the first place, all model data needs to be interpolated to the exact location of the weather stations. The weather models work with a grid, for instance, a ten by ten kilometer grid, which is why a weather station is hardly ever at the precise location of that grid point. This means that the model output needs to be recalculated to the observation point, in order to achieve the model expectation for precisely that location. But that’s not all. It is also necessary to recalculate the model output retroactively for all forecast periods for the previous two years. Are you still with us? These are processes that require a huge amount of data processing power. And a massive amount of effort. Marco: “I occasionally had dreams about lists and files. Often, after a weekend off, I had to focus intensely on a flowchart to regain an overview of the entire process we were in. As you have to deal with such an enormous amount of input, this also means that a lot of things can go wrong. On that point, we have also collaborated closely with the Weather Systems Team to thoroughly check and test the equations that were being developed.”
The innovative final step to come to superior results
"Eventually, we reached the finale of this entire endeavor. The enormous pile of historic observation data and the specific model forecasts per station were linked to each other. The moment was at hand. Marco: “At this point in the project, we made the decisions on what the perfect mix should be regarding model weighting factors and additional equations to reach the best possible forecast. Per weather station and per time period. That was the scariest part of the process, since it could all go wrong again at that point. In such a complicated project, things can go wrong at so many stages of the project that you might really have to take four or five steps back in order to move forward again.”
In the end, this final step also made it all the way to the end. The new MultiModel MOS went live in early October. That makes us completely up to date again and has improved the quality of our weather forecasts even more. We do still have to deal with a few hiccups, though. Marco: “Poland is currently experiencing very heavy rainfall – at least, according to the MultiModel MOS it is. That means something has gone wrong with the historical database. Somewhere the information is not being processed correctly or is not stored in the right place. We have quickly rolled back old MOS equations for Poland and will be looking into what went wrong. As I said earlier: sometimes you have to take a few steps back in order to go forward.”
The end of this huge MultiModel MOS update is not a real ending. Just like all the other products, we continue to tinker with it constantly and run constant background checks. That is the only way to remain a successful weather company with a sizable customer base.
The weather models that we work with are the ECMWF, EPS, UKMO and NCEP models. In addition to this data, we also have the observations from approximately 20,000 weather stations around the world at our disposal. By comparing weather model information and observations and building statistical equations on that, we eventually achieve better forecasts.