Big Data and Predictive Analytics in Fire Risk Using Weather Data

Abstract

The objective of this article is to study the impact of weather on the damage caused by fire incidents across the United States. The article uses two sets of big data—‐fire incidents data from the National Fire Incident Reporting System (NFIRS) and weather data from the National Oceanic and Atmospheric Administration (NOAA)—to obtain a single comprehensive data set for prediction and analysis of fire risk. In the article, the loss is referred to as “Total Percent Loss,” a metric that is calculated based on the content and property loss incurred by an owner over the total value of content and property. Gradient boosting tree (GBT), a machine learning algorithm, is implemented on the processed data to predict the losses due to fire incidents. An R2 value of 0.933 and mean squared error (MSE) of 124.641 out of 10,000 signify the extent of high predictive accuracy obtained by implementing the GBT model. In addition to this, an excellent predictive performance demonstrated by the GBT model is further validated by a strong fitting between the predicted loss and the actual loss for the test data set, with an R2 value of 0.97. While analyzing the influence of each input variable on the output, it is observed that the state in which a fire incident takes place plays a major role in determining fire risk. This article provides useful insights to fire managers and researchers in the form of a detailed framework of big data and predictive analytics for effective management of fire risk.

A Two‐Stage Data‐Driven Spatiotemporal Analysis to Predict Failure Risk of Urban Sewer Systems Leveraging Machine Learning Algorithms

“…One of the main tasks in the prediction workflow is to evaluate the generalization performance of machine learning algorithms using an appropriate resampling procedure. $$k$$‐fold cross‐validation is a widely used resampling technique that can be used to balance the bias and variance, and estimate the out‐of‐sample predictive performance of machine learning models (Agarwal, Tang, Narayanan & Zhuang, 2020; Alipour, Mukherjee, & Nateghi, 2019; Hastie et al., 2009; James et al., 2013; Jung, 2018; Mukherjee & Nateghi, 2019, 2017; Obringer, Mukherjee, & Nateghi, 2020). This approach involves randomly dividing the set of observations into $$k$$‐folds of approximately equal sizes.…”

“…In XGB, a large number of weak learners are built and combined sequentially to produce a strong learner. The difference between XGB and other gradient boosting algorithms like gradient boosting machines or gradient boosting trees (Agarwal et al., 2020) is due to its regularization formalization to control overfitting. Indeed, the name XGB refers to the goal of taking the computational power to its limits; using OpenMP API (Chapman, Jost, & Van Der Pas, 2008) for parallel processing, XGB is able to utilize the multiple cores that are available on a single machine's CPU for parallel computation (Chen et al., 2018).…”

Machine Learning Algorithms to Distinguish Myocardial Perfusion SPECT Polar Maps

“…Boosting techniques (AB, GB, and XGB) were used to properly combine a series of weak classifiers to obtain a stronger one. A weak classifier occurred when a feature's performance was slightly superior to random guessing (15)(16)(17)(18)(19)(20)(21)(22)(23)(24). Table A in Section Supplementary Data 1 shows parameters used in our ML algorithms.…”

Urban Fire Station Location Planning using Predicted Demand and Service Quality Index

“…In [26], a Random Forest algorithm is used to generate predicted fire risk scores to aide the local fire department in identifying properties that require fire inspections. Such data-driven predictive models are widely used in many prior works [27][28][29][30]. [31] proposes an interesting approach to prediction of fire incidents that is based on a deep neural network.…”