Humans have been forecasting weather since at least 300 BC. Over time, the development of weather instruments and models has improved data collection and forecasting. The advent of computers - including ENIAC (pictured above) - paved the way for Numerical Weather Prediction (NWP) models, which leverage the underlying physics to dramatically improve weather forecasting. We're now entering a new age of weather forecasting, where artificial intelligence (AI) will push the boundaries of what we can forecast.
The Limits of Numerical Weather Prediction
(1) The physics of the problem is fully known. However, this is often not true for complex problems like weather and climate. For example, it is difficult to codify the physics of long-term weather phenomena in a model.
(2) The physics of the problem can be simulated well. What happens when simulations become costly? As physics models become more complex, simulating NWPs becomes computationally prohibitive.
Given these two downsides, it's worth asking: Are there other ways to approach the weather forecasting problem?
Enter Physics + AI
In the past four decades, there has been an explosion of data from sources such as satellite observations, sensors, and NWP models. We've also witnessed incredible advances in AI models and hardware.
Given these advances, we revisited the problem of forecasting at Vāyuh.ai. Leveraging on our years of experience in AI and Climate Science at UC Berkeley and Berkeley Lab, approached the task of sub-seasonal forecasting for temperature in the Continental United States (CONUS). Subseasonal forecasting requires making predictions 3 to 12 weeks in advance, which is known to be an extremely challenging forecasting timescale.
We found that by combining ground truth observations, physics-based forecasts, and AI models, we can improve on existing forecasts across locations, seasons, and temperature regimes. Our models leverage the best of both worlds: they're physics-informed, but greatly benefit from the pattern recognition brought by AI.
Physics-informed machine learning results in models that respect symmetry, conservation laws, and capture dynamics (Karniadakis et al., 2021, Nature Reviews Physics).
However, forecasts themselves sometimes don't move the needle for all stakeholders. In many cases, what matters most is anomalous behaviour: how different is today's temperature relative to historical patterns? At Vāyuh.ai, we have taken our state-of-the-art forecast and sprinkled some good old Bayesian methods to quantify uncertainty in anomalous event detection. Specifically, our models capture the probability of anomalies occurring, allowing you to make informed decisions based on our forecast.
We can assess the quality of our model's probabilistic predictions with receiver operator characteristic (ROC) curves (see figure below). ROC curves are well-suited to answering the question of how well can we trust the probabilistic prediction of an event. The further above the dashed gray line, the better a model's predictions. We used ROC curves to compare our model to CFSv2 (one of the most commonly used NWPs). Our model outperforms CFSv2, demonstrating how physics and AI lead to better forecasting.
Receiver Operating Characteristic (ROC) curves are great ways to evaluate a model's performance on events. A perfect model would be a right angle curve. Dashed gray line denotes chance, the blue curve corresponds to CFSv2, and the orange curve corresponds to Vāyuh.ai. Vāyuh.ai's forecast considerably outperforms CFSv2. This plot was generated for all of CONUS on held-out test set.
Our forecasts are currently being used by a top 5 bank and other energy firms to help improve their understanding of weather events. Our clients see anomalous weather events weeks before their peers, helping them gain a competitive advantage.
At Vāyuh.ai, we're invested in developing the most accurate forecasting product, tailored to your needs. How can improved temperature forecasts help your business? Would you like to know how your produce may be affected ahead of time? Would you like to understand the fluctuations in the power or energy markets?
Drop us a note on LinkedIn or reach out at firstname.lastname@example.org. and we'd be happy to share more.