Machine Learning in Weather Forecasting
In the past, weather forecasts were entirely based on ground-level observations. Although those got the job done, the accuracy was hit-or-miss at best, especially when it came to precipitation. Just like in any other data-driven field, more data leads to better forecasting: thanks to giant leaps in technological advancement, surface observations are now accompanied by thousands of weather stations, radars, and satellites collecting varied and precise measurements all over the world.
With the enormous volumes of raw data meteorologists collect daily, weather forecasting and machine learning are a match made in heaven. AI models can analyze past forecasts and observations, compare them to weather conditions observed in reality, and help eliminate errors in new forecasts. Meteum combines run-of-the-mill meteorological models with machine learning and crowdsourced user reports to deliver hyper-accurate weather predictions. Without further ado, let’s dive into how we approach machine learning at Meteum to provide some of the best-in-class forecasts and hyperlocal nowcasts.
Traditional Weather Forecasting and Machine Learning
Traditional weather forecasting relies upon a combination of weather observations and numerical models. Meteorologists produce weather forecasts by gathering as much data from earth and space as possible and then processing it through weather prediction models. At Meteum, we use four models developed by third-party providers: the Global Forecast System (GFS) in the United States, the European Centre for Medium-Range Weather Forecasts (ECMWF), Japan Meteorological Agency (JMA), and the Canadian Meteorological Center (CMC). In addition to these, Meteum uses the open-source Weather Research and Forecasting model (WRF). The data from the five models is meticulously processed by Meteum’s algorithms to produce a forecast more accurate than any individual model could ever provide by itself. However, this forecast only includes information on temperature, humidity, winds, and atmospheric pressure.
It takes much more to tackle precipitation. This is where machine learning truly shines, empowering us to provide weather forecasts accurate to the individual building or neighborhood. Our model uses gradient boosting, a machine learning technique for building predictive models, to correct any errors that come from traditional weather forecasting. We employ thousands of parameters to train our neural network, from wind speed and temperature to the height of the planetary boundary layer.
Nowcasting With Deep Learning
Aside from more accurate forecasts, machine learning can improve nowcasting — immediate weather prediction. Nowcasts are typically minute-by-minute precipitation forecasts extending two hours into the future. While nowcasting is technically possible with just radar data, weather models based on machine learning can also consider the data from weather satellites. This comes in handy considering that weather radars are scarce, especially in developing countries.
Deep learning is an AI technology that aims to replicate how humans process things like images. In weather forecasting, prediction models use deep learning primarily to process images from weather satellites, and our model is no exception. The satellite data we receive is combined with radar measurements and numerical model forecasts to generate a precipitation map.
Adding weather satellites to the mix dramatically expands the nowcast’s reach and accuracy: our algorithms can not only detect rain clouds but also predict the direction they’re moving in and estimate the likelihood of rain from them. With machine learning, anyone in the range of a weather satellite can reap the benefits of nowcasting, rather than just those living near a radar station.
It stands to reason that our nowcasts aren’t always perfect, but herein lies the power of machine learning — we can learn from these mistakes. To aid this process, we’ve implemented user reports to refine our nowcasts on the fly. Here’s how it works: our weather app will occasionally ask users to confirm if the weather they experience is correctly reflected in the UI, and we correct the precipitation map if a sufficient number of users report a miscalculation. User reports are considered a ground truth and treated as the most accurate observations; we add them to the data mix we use to further train our models and improve future weather predictions.
In the last few decades, AI and machine learning have revolutionized weather forecasting. We at Meteum are beyond excited for the opportunities the future holds. As machine learning advances and more weather models start integrating it, weather forecasting will become increasingly more accurate. There is also excellent potential for a global expansion of nowcasting, a relatively recent addition to consumer weather forecasting. Only a select few weather services include nowcasting in their forecasts, and, in the past, the tech was limited to people in areas with reliable radar coverage. As Meteum demonstrates, machine learning can be added to weather forecasting to extend nowcasting to places that lack widespread radar coverage.
Our algorithms are built with you in mind: hyperlocal nowcasts help users make day-to-day decisions depending on whether it’s going to rain or snow on a particular day. Detailed and precise weather data is also a massive boost to any business operations, especially in industries highly dependent on meteorological conditions. Learn about the solutions available for your industry and grab a free API key at meteum.ai/b2b to peek into the future of ML-powered weather forecasting!
