Explainable Machine Learning for Extreme Weather Event Prediction Using Long-Term Environmental Measurements
This study presents an explainable multi-tower machine learning framework using 15-minute meteorological data from six Oak Ridge Reservation towers to forecast five types of extreme weather events, demonstrating that gradient boosting models outperform deep learning and CNN architectures while identifying key tower-based and physics-based features that drive prediction accuracy across varying time horizons.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Weather is not just a backdrop for daily life; for the massive research facilities and critical infrastructure nestled in the rugged hills of East Tennessee, it is a constant operational variable. When a sudden gust of wind threatens to topple a crane, or a rapid drop in temperature risks freezing sensitive equipment, the difference between safety and disaster often comes down to minutes. For decades, scientists have relied on weather models that look at the atmosphere as a broad, sweeping canvas, useful for predicting rain across a state but often too coarse to catch the sharp, localized shifts that happen in complex terrain like valleys and ridges. To manage the Oak Ridge Reservation, a site home to national security labs and research centers, operators need to know what is happening right outside their doors, not just what is happening a hundred miles away. The challenge lies in the sheer complexity of the local landscape, where cold air pools in low spots and winds accelerate over hilltops, creating unique microclimates that standard forecasts miss.
A team of researchers has now developed a new way to predict these dangerous, localized weather events by teaching computers to listen to a network of six weather towers scattered across the reservation. Instead of relying on a single model or a single viewpoint, they built a system that fuses data from all six towers simultaneously, allowing the computer to see the weather as a connected, three-dimensional puzzle rather than isolated points. By analyzing years of high-resolution measurements, the team trained machine learning algorithms to recognize the specific signatures of five distinct extreme events: dangerous heat and humidity, biting wind chill, freezing temperatures, sudden high winds, and periods of stagnant air. Their work demonstrates that simpler, well-tuned computer models can outperform more complex, deep-learning systems for this specific task, providing facility managers with reliable warnings from thirty minutes up to six hours in advance.
The researchers began by gathering a massive dataset from six instrumented towers, each perched at different elevations and facing different slopes, from flat valley floors to steep, forested ridges. These towers recorded temperature, humidity, wind speed, and pressure every fifteen minutes over several years. The team first used this data to define exactly what constituted a dangerous event. They identified five specific scenarios that matter most for site safety: a combination of high heat and moisture that creates a hazardous environment, a wind chill condition where cold air and wind combine to lower the felt temperature, a drop in temperature below freezing, a sudden spike in wind speed, and a period of very low wind that causes air to stagnate. To make the computer models effective, the researchers did not just feed them raw numbers. They engineered the data to include "lag" features, which tell the model what the weather was doing in the recent past, and "rolling" statistics, which smooth out the data to show trends, dips, and spikes over time. They also added a layer of physics-based context, such as atmospheric stability and pressure gradients, which are derived from meteorological expertise to help the model understand the larger forces driving the local weather.
To test their approach, the team ran three hundred different experiments, comparing six different types of computer architectures. These ranged from gradient boosting models, which are powerful decision-making tools that build a prediction by combining many simple rules, to deep learning and convolutional neural networks, which are often used for complex pattern recognition in images and speech. The goal was to see which type of model could best predict the five extreme events at different time horizons, from thirty minutes out to six hours. The results were clear and surprising to those who might assume that the most complex models would always win. The gradient boosting models consistently outperformed the deep learning and neural network approaches across the board. For predicting temperature-related events like wind chill and freezing conditions, the best models achieved an accuracy that was nearly perfect, maintaining high reliability even when looking six hours into the future. This is because temperature changes slowly, driven by the thermal inertia of the ground and the air, giving the models a long "memory" to work with.
Wind events, however, told a different story. Predicting sudden high winds proved to be the most difficult challenge, with accuracy dropping significantly as the prediction window extended beyond thirty minutes. This is because wind gusts are driven by chaotic, turbulent air movements that change in seconds, making them much harder to forecast than temperature. The study found that while the complex deep learning models struggled to capture these rapid, turbulent shifts, the simpler gradient boosting models handled them better, though they still faced limits. A key finding was that the location of the weather data mattered immensely. For wind events, the towers on the ridges provided the most critical information, as they were the first to feel the acceleration of air over the hills. For temperature events, the towers in the valleys were most important, as they captured the pooling of cold air. By combining data from all towers, the models learned to see the full picture, using the ridge towers to predict wind and the valley towers to predict cold, creating a unified view that no single tower could provide.
The researchers also tested whether adding those extra physics-based variables, like atmospheric stability and pressure gradients, helped the models. They found that these extra details were only useful for specific situations. For wind events, especially when looking further into the future, the physics-based data provided a noticeable boost in accuracy, helping the model see the larger forces at play once the immediate, chaotic turbulence had faded. However, for temperature events, these extra variables offered little to no benefit. The local temperature readings from the towers were already so strong and clear that adding more data just introduced noise. This suggests that for some types of weather, the dense network of local sensors is sufficient, while for others, understanding the broader atmospheric context is essential.
The practical implications of this work are significant for the safety and operations of critical facilities. The study confirms that a single, unified model can effectively monitor a complex landscape, replacing the need for separate, isolated forecasts for each location. For facility managers, this means they can receive reliable alerts for freezing temperatures up to six hours in advance, allowing time to protect equipment. They can get warnings for dangerous wind chill conditions an hour or two ahead to stop outdoor work, and receive thirty-minute alerts for sudden high winds to secure cranes and heavy machinery. The system also identifies periods of stagnant air, which is crucial for facilities handling hazardous materials that rely on ventilation. By proving that simpler, faster models can outperform complex deep learning systems for this specific task, the researchers have provided a blueprint for building efficient, explainable warning systems. The models do not just predict; they explain their reasoning, showing exactly which tower and which weather variable triggered the alert, giving operators the confidence to act on the information. This approach turns a vast network of sensors into a single, intelligent observer, capable of seeing the subtle shifts in the atmosphere that precede danger.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.