Catastrophic wildfire index for forecasting utility-caused wildfires
A service inputs wind data and utility data corresponding to a utility component into a first machine learning model to determine an outage risk prediction representing a probability that the utility component will have an outage. The service determines a probability of ignition at a vicinity of the utility component, and determines a set of wildfire impact measurements by simulating a wildfire in the vicinity of the utility component. The service inputs the outage risk prediction, the probability of ignition, and the set of wildfire impact measurements into a second machine learning model, and receives as output from the second machine learning model, a catastrophic wildfire risk score corresponding to the utility component. The service outputs a graphical representation on a dashboard representing the catastrophic wildfire risk score.
1 . A method comprising:
inputting wind data and utility data corresponding to a utility component into a first machine learning model, wherein the first machine learning model is trained using training examples indicating at least a representation of wind speed and attributes of a given utility component as paired with a label indicating whether or not an outage was experienced, and wherein as new outages are experienced, the first machine learning model is re-trained using new training examples indicating, for the new outages, at least a representation of new wind speed and new attributes of a utility component that experienced a new outage;
receiving, as output from the first machine learning model, an outage risk prediction representing a probability that the utility component will have an outage;
determining a probability of ignition at a vicinity of the utility component;
determining a set of wildfire impact measurements by simulating a wildfire in the vicinity of the utility component;
inputting the outage risk prediction, the probability of ignition, and the set of wildfire impact measurements into a second machine learning model;
receiving, as output from the second machine learning model, a catastrophic wildfire risk score corresponding to the utility component; and
outputting a graphical representation on a dashboard representing the catastrophic wildfire risk score.
2 . The method of claim 1 , wherein the first machine learning model comprises a Bayesian framework that predicts probability distributions of utility component outages.
3 . The method of claim 1 , wherein determining the probability of ignition at the vicinity of the utility component comprises:
inputting fuel type, fuel dryness, and wind speed at the vicinity of the utility component into a model; and
receiving as output from the model the probability of ignition at the vicinity of the utility component.
4 . The method of claim 1 , wherein simulating the wildfire in the vicinity of the utility component comprises:
running a simulation for a range of time assuming ignition at the utility component and based on the wind data and environmental data in the vicinity of the utility component; and
extracting the set of wildfire impact measurements from an end of the range of time, each wildfire impact measurement of the set indicative of one or more measures of destruction caused by the simulated wildfire.
5 . The method of claim 1 , wherein the set of wildfire impact measurements comprises an initial attack assessment projecting a likelihood of success of containment by firefighting within a threshold amount of time.
6 . The method of claim 5 , wherein the initial attack assessment is determined by:
inputting fire behavior metrics and environmental metrics into a third machine learning model; and
receiving, as output from the third machine learning model, an indication of whether an initial attack would be likely to succeed.
7 . The method of claim 1 , wherein the second machine learning model is trained by:
constructing training examples by:
for each historical instance of a fire caused by a utility outage:
generating training data by determining historical outage risk prediction and historical probability, and adding a historical set of wildfire impacts; and
generate a training example by collating the training data with a label indicating whether the fire is consequential; and
training the second machine learning model to predict whether a given fire is consequential using the training examples.
8 . A non-transitory computer-readable medium comprising memory with instructions encoded thereon that, when executed, cause one or more processors to perform operations, the instructions comprising instructions to:
input wind data and utility data corresponding to a utility component into a first machine learning model, wherein the first machine learning model is trained using training examples indicating at least a representation of wind speed and attributes of a given utility component as paired with a label indicating whether or not an outage was experienced, and wherein as new outages are experienced, the first machine learning model is re-trained using new training examples indicating, for the new outages, at least a representation of new wind speed and new attributes of a utility component that experienced a new outage;
receive, as output from the first machine learning model, an outage risk prediction representing a probability that the utility component will have an outage;
determine a probability of ignition at a vicinity of the utility component;
determine a set of wildfire impact measurements by simulating a wildfire in the vicinity of the utility component;
input the outage risk prediction, the probability of ignition, and the set of wildfire impact measurements into a second machine learning model;
receive, as output from the second machine learning model, a catastrophic wildfire risk score corresponding to the utility component; and
output a graphical representation on a dashboard representing the catastrophic wildfire risk score.
9 . The non-transitory computer-readable medium of claim 8 , wherein the first machine learning model comprises a Bayesian framework that predicts probability distributions of utility component outages.
10 . The non-transitory computer-readable medium of claim 8 , wherein the instructions to determine the probability of ignition at the vicinity of the utility component comprise instructions to:
input fuel type, fuel dryness, and wind speed at the vicinity of the utility component into a model; and
receive as output from the model the probability of ignition at the vicinity of the utility component.
11 . The non-transitory computer-readable medium of claim 8 , wherein the instructions to simulate the wildfire in the vicinity of the utility component comprise instructions to:
run a simulation for a range of time assuming ignition at the utility component and based on the wind data and environmental data in the vicinity of the utility component; and
extract the set of wildfire impact measurements from an end of the range of time, each wildfire impact measurement of the set indicative of one or more measures of destruction caused by the simulated wildfire.
12 . The non-transitory computer-readable medium of claim 8 , wherein the set of wildfire impact measurements comprises an initial attack assessment projecting a likelihood of success of containment by firefighting within a threshold amount of time.
13 . The non-transitory computer-readable medium of claim 12 , wherein the initial attack assessment is determined by:
inputting fire behavior metrics and environmental metrics into a third machine learning model; and
receiving, as output from the third machine learning model, an indication of whether an initial attack would be likely to succeed.
14 . The non-transitory computer-readable medium of claim 8 , wherein the second machine learning model is trained by:
constructing training examples by:
for each historical instance of a fire caused by a utility outage:
generating training data by determining historical outage risk prediction and historical probability, and adding a historical set of wildfire impacts; and
generate a training example by collating the training data with a label indicating whether the fire is consequential; and
training the second machine learning model to predict whether a given fire is consequential using the training examples.
15 . A system comprising:
memory with instructions encoded thereon; and
one or more processors that, when executing the instructions, are caused to perform operations comprising:
inputting wind data and utility data corresponding to a utility component into a first machine learning model, wherein the first machine learning model is trained using training examples indicating at least a representation of wind speed and attributes of a given utility component as paired with a label indicating whether or not an outage was experienced, and wherein as new outages are experienced, the first machine learning model is re-trained using new training examples indicating, for the new outages, at least a representation of new wind speed and new attributes of a utility component that experienced a new outage;
receiving, as output from the first machine learning model, an outage risk prediction representing a probability that the utility component will have an outage;
determining a probability of ignition at a vicinity of the utility component;
determining a set of wildfire impact measurements by simulating a wildfire in the vicinity of the utility component;
inputting the outage risk prediction, the probability of ignition, and the set of wildfire impact measurements into a second machine learning model;
receiving, as output from the second machine learning model, a catastrophic wildfire risk score corresponding to the utility component; and
outputting a graphical representation on a dashboard representing the catastrophic wildfire risk score.
16 . The system of claim 15 , wherein the first machine learning model comprises a Bayesian framework that predicts probability distributions of utility component outages.
17 . The system of claim 15 , wherein determining the probability of ignition at the vicinity of the utility component comprises:
inputting fuel type, fuel dryness, and wind speed at the vicinity of the utility component into a model; and
receiving as output from the model the probability of ignition at the vicinity of the utility component.