The achievement by Tomorrow.io’s R&D team leveraged an approach powered by physical models and supercharged with AI/ML allows for vastly improved decision making confidence in advance of weather impact.
Boston, March 15, 2023 – Tomorrow.io, the world’s leading weather intelligence and climate adaptation platform is excited to announce significant advancements in its One Forecast (1F) Model by leveraging operational weather models and machine learning. After extensive tests and validation, Tomorrow.io’s 1F Model’s forecast now provides up to 38% better data for supporting predictive business decisions than publicly available forecasts. The model upgrades will be implemented into Tomorrow.io’s weather intelligence platform during Q2.
Created by Tomorrow.io’s industry leading R&D team, the model incorporates a high-resolution, short-term forecasting system leveraging a unique combination of both machine learning and state-of-the-art numerical weather prediction (NWP) technology. Tomorrow.io’s 1F model rapidly post-processes NWP weather data, reduces inaccuracies and quantifies the uncertainty in the weather forecast, thereby generating probabilistic predictions.
“Our next-generation 1F model is a game-changer for businesses to better understand weather forecasts, make more informed decisions, and take actions to protect critical infrastructure or moving assets in advance of impact,” said Luke Peffers, Chief Weather Officer at Tomorrow.io. “We are pushing the boundaries of what can be reliably predicted through a combination of NWP models and machine learning, resulting in more accurate forecasts and probabilistic outputs that provide a significant advantage for our customers and enabling them to protect their assets and resources.”
Tomorrow.io’s 1F model is a multi-task neural network with a custom loss function that predicts seven key weather variables and counting. In contrast to traditional neural networks that are typically trained to predict one output variable, multi-task neural networks are trained to optimize the performance of multiple output variables at the same time. This is ideal for predicting highly correlated weather variables that are related to the underlying weather phenomena. The model outputs 21 ensemble members, which are used to generate probabilities that can be used to drive better decision making.
“We have demonstrated impressive results for the Contiguous United States (CONUS) model with improved forecast skill across key variables both from a deterministic and probabilistic model verification,” said Tyler McCandless, Director of Data Science at Tomorrow.io. “Our deterministic model results show an improvement over the baseline HRRR model up to 12.5% in overall forecast lead times across CONUS, as quantified by the Root Mean Square Error (RMSE). Beyond the significant improvement in our deterministic model forecast skill, we also demonstrated a tremendous improvement in probabilistic forecasting with up to 38% improvement in our Continuous Ranked Probability Score (CRPS) metric.”
Tomorrow.io’s innovative approach combines high-resolution NWP with deep learning to generate an ensemble forecasting system that is calibrated to the uncertainty in the atmospheric conditions. By training the multi-task neural network using the High-Resolution Rapid Refresh (HRRR) as input, our model achieves high spatial resolution (3 km), high temporal resolution (hourly), and a high refresh rate (updated every hour over CONUS with predictions out to 42 hours). The 1F model also incorporates multiple datasets to capture land use, geography, and other time-invariant locational information. Tomorrow.io’s ML-NWP approach including the use of high-resolution real-time modeling is far superior to alternatives based on coarse resolution reanalyses.
Hear more from Tyler McCandless on this upgrade here:
See Weather Intelligence in Action