Since the beginning of time, humans have tried to accurately forecast the weather, using historical data and patterns to infer what might occur in the future. While the most obvious consumer benefit of simply enhancing convenience (e.g., helping to schedule a vacation so it does not coincide with bad weather or ensuring that the Little League game is not scheduled for a rainy day), even slight improvements in predicting the weather could yield significant gains for businesses, government agencies, and other entities that depend on accurate forecasts.
More accurate forecasts could enable farmers to pick the optimal days for planting or harvesting. Train or plane schedules could be modified to account for expected weather interruptions, with costly assets (e.g., moving extra labor or equipment into place to manage weather-related disruptions) accurately accounted for in annual budgets. Businesses that are weather dependent, such as landscaping companies or utility companies that need to do maintenance, can more accurately match labor and resources to expected weather events.
AI can help make more accurate forecasting a reality. Weather forecasting is an ideal use case for AI, because there is a treasure trove of historical and currently available weather data (with more sensors being frequently added), which can feed data into algorithms that literally require both quality and quantity in order to be effective at mating past occurrences with future predictions.
How AI Can Help Improve Weather Prediction
While no one can ever fully predict the future, artificial intelligence (AI) techniques apply reinforcement learning on past predictions and actual outcomes. By comparing predictions with outcomes, the model is able to learn and improve simulation capabilities, forecasting much further into the future and with greater accuracy.
To aid in weather predicting, data is fed into an algorithm that uses deep learning techniques to learn and make predictions based on past data. Deep learning techniques have already been shown to be successful in areas like image and speech recognition and natural language processing (NLP), and it can be applied to the weather and climate field as well. Still, weather patterns are made up of a complex number of data points, making weather prediction a highly data- and compute-intensive exercise. This portends the use of significant compute power, as well as a plethora of storage to retain the data. Furthermore, deep learning algorithms are only as effective as the inputs they are trained on, making data quality and labeling a crucial component of this technique. Nevertheless, the power of AI and deep learning systems can be used to ingest and analyze these multi-factor datasets much more quickly and accurately than ever before.
Results are beginning to emerge. For example, a 2017 article published by the American Meteorological Society (AMS) indicated that the use of modern AI techniques is improving the ability to sift through the overwhelming amounts of weather data to extract accurate insights, and to provide timely guidance for human weather forecasters and decision-makers. The key benefit is that AI can provide more flexible and powerful models capable of identifying complex relationships between a large number of modeled and observed weather features. The paper also noted that AI can be used to directly predict the effects of high-impact weather, including power generated by variable sources, such as solar or wind, energy consumption in a specific area, or airport arrival capacity.
Expanding Sources of Weather Data
Another reason why weather forecasting is a compelling AI use case is the sheer amount of data that is generated every day. More than 1,000 weather-focused satellites currently orbit Earth, which track and send data about cloud patterns, winds, temperature, and weather systems. Hundreds of thousands of terrestrial stations on Earth are constantly gathering real-time data. Perhaps most interestingly, driven by the expansion of the Internet of Things (IoT), new sources of data are coming online every day (e.g., internet-enabled traffic lights, solar panels, weather thermometers, and smart air conditioning (A/C) devices) that can feed even more granular data into a weather-prediction algorithm.
Tractica forecasts that AI technology used in weather forecasting will reach $553.4 million annually in spending on hardware, software, and services by 2025, up from $31.7 million in 2018. Fueling the market are large companies that have invested in weather prediction systems that integrate public and private weather data.
IBM began focusing on using its computer systems to improve forecasts in 1996 and has been refining its project ever since. In 2016, IBM acquired the Weather Company’s properties, including weather.com, Weather Underground, The Weather Company brand, and WSI, its global business-to-business brand.
The purchase provided IBM with access to the Weather Company’s network of sensors and models, providing a massive pipeline of weather data it could feed into IBM’s AI platform Watson to attempt to improve predictions. As of 2016, the Weather Company claimed that its models used more than 100 terabytes of third-party data daily. The result of the combination is IBM Deep Thunder, which can provide customized information for business clients by using hyper-local forecasts, at a 0.2 to 1.2-mile resolution; all information that is useful for transportation companies, utility companies, and even retailers.
Panasonic has also focused on developing a weather forecasting model. In 2013, the company purchased AirDat, which makes TAMDAR, a specialty weather sensor installed on commercial airplanes that account for several thousand flights a day. Data from these sensors is combined with public weather data, resulting in Panasonic Global 4D Weather, a global weather forecasting platform. Panasonic scored significant public relations points in 2017 with Hurricane Irma. The system’s predictions about the storm turned out to be the most accurate of any model, particularly with respect to identifying the storm’s eventual track four to seven days in advance.
But it is not just large companies that are focused on weather forecasting using AI technology. Startups, such as ClimaCell, TempoQuest, and Earth Networks, are using new technology to better predict weather events.
ClimaCell, based in Boston, Massachusetts, is focused on extracting weather data from cellular networks and combining it with historical and weather data from ground-based weather stations, radar sources, and satellites. The company developed software that creates high-definition weather maps based on the ways weather can affect cellular signals, which can then be mapped to specific weather events. The company claims its approach doubles the reliability of radars, with up to 10X the ground resolution.
Meanwhile, Boulder, Colorado-based TempoQuest is focused on ingesting weather data collected by satellites, drones, and radar. It then uses software that runs on extremely powerful NVIDIA graphics processing units (GPUs) than can generate high-resolution weather forecasts in minutes, resulting in a highly accurate and fast forecast that can be used by commodities and weather derivatives traders, energy traders, commercial transportation businesses, and passenger airlines.
Finally, Earth Networks, a Germantown, Maryland-based company, operates 12,000 weather stations in more than 90 countries, along with 1,800 sensors in 50 countries for its Total Lightning Network. The Total Lightning Network can detect in-cloud lightning to generate storm alerts associated with severe weather like tornadoes and hail. The company also incorporates a number of global weather models, combining it with its own proprietary datasets to provide forecast data to a variety of industries, government agencies, and businesses.