A team of scientists from Aalto University and the Finnish Meteorological Institute have developed an open-source AI model for predicting electrical outages caused by storm damage. The model can predict storm location within 15km and classifies the amount of transformer damage with 81% accuracy, allowing power companies to prepare for outages and repair them more quickly.
The work was described in an article published in the European Geosciences Union's (EGU) Natural Hazards and Earth System Sciences (NHESS) journal. The model predicts damage to power transformers from large low-pressure storms up to 10 days in advance, categorizing the results as either no damage, low damage (less than 140 transformers damaged), or high (more than 140). The predictions are based on a support-vector classifier, which achieves 81% precision and 61% recall. Using this model, power companies can prepare materials and repair crews, restoring power to customers more quickly.
Because Finland is a heavily forested country, its overhead power lines are often damaged by falling trees, especially during strong extratropical storms; on average, about 46% of the country's power outages were caused by these storms. Because the power suppliers are required by law to provide their customers with financial compensation for prolonged outages, the companies maintain a large workforce for rapid repair. While several researchers have applied AI techniques to predict power outages from hurricanes, as well as damage to trees (not surprisingly, random forests work quite well for this task), there has been little work specifically on power outages due to extratropical storms.
The Finnish team's model works by first identifying storm objects in weather data; these are defined as polygons where surface wind gusts exceed a certain threshold during a 1-hour time interval; then storm object motion is tracked across these time intervals. To predict the severity of the storm damage, the model extracts several features from each storm object, including wind speed and direction, air temperature and pressure, and cloud cover. The model also incorporates forest data from the storm's current location. These features are then used as input to a classifier. According to the team,
Our method is the first that employs the extratropical storm objects as polygons and combines them with meteorological and non-meteorological features to predict power outages.
Using historical weather and power outage data, the team trained and evaluated several different classifier models, including random forests (RFC), support vector machines (SVC), Gaussian naive Bayes (GNB), Gaussian process (GP), and multi-layer perceptron (MLP). Because most storms in the dataset did not cause any damage, the output classes are imbalanced, which can make the selection of evaluation metrics tricky (for example, accuracy is often a poor choice). The team noted that for their use case, the worst outcome is a prediction of no damage for a storm that does indeed cause damage. Thus, they chose the SVC, which "miss[es] the smallest number of destructive storms, although it confuses the amount of caused damage."
While traditional weather forecasting methods are often based on stochastic methods and physical models, several large tech companies have recently published research on using AI and deep learning for weather prediction. Last year, Microsoft collaborated with scientists from the University of Washington to develop a convolutional neural network (CNN) model to predict "several basic atmospheric variables on a global grid." Earlier this year, a team from Google's DeepMind branch worked with the U.K. Met Office to create a deep-learning based generative model for nowcasting; that is, "high-resolution forecasting of precipitation up to two hours ahead."
The Finnish team's source code for storm tracking and classification are available on GitHub. The weather and forest datasets are also publicly available; however the power outage data is proprietary.
