This work examines different spatial and sociodemographic models for predicting residential fire counts at the census tract level for 118 U.S. fire departments across 25 states. The models give five-year forecasts of residential fire counts for 3392 census tracts which contain over 13 million residents in total. All models described in this paper train on fire incident data from the National Fire Incident Reporting System (NFIRS) over the interval 2006–2011 (inclusive) and are evaluated based on their ability to predict the fire counts that occurred over the interval 2012–2016. Two strictly spatial models are considered- a simple “count” model that serves as a baseline for all other models described in the paper and a model that utilizes kernel density estimation (KDE) with statistically optimized bandwidths. Using data from the American Community Survey (ACS), an examination of the effects of demographic and housing factors on the fire risk is presented. The data suggest that the fire risk per person is generally higher in census tracts with attributes corresponding to socioeconomic disadvantage such as low median incomes and small fractions of residents with college degrees. These trends inform the design of a Bayesian hierarchical Poisson regression model, which is shown to make predictions with a 9% lower root mean squared error (RMSE) relative to the base model. A spatial kernel regression is then conducted on the residuals of this regression, which results in a 15% RMSE improvement relative to the base model. These results are compared to a conditional autoregressive (CAR) model, which incorporates spatial information directly into the hierarchical Poisson regression. Although the RMSE is higher for the CAR model's point estimate forecasts (7% lower than the base model), it allows for the generation of probabilistic forecasts and gives spatially-informed statistical estimates of the effects of the sociodemographic variables. This work highlights the utility of geocoded fire incident and demographic data as well as machine learning techniques that can utilize these datasets to make improved predictions.

这项工作研究了不同的空间和社会人口统计学模型，以预测人口普查区域25个州的118个美国消防部门的住宅火灾计数。这些模型对3392个人口普查区的住宅火灾计数进行了五年预测，该人口总数超过1300万。本文中描述的所有模型均以2006-2011年（含）间隔内国家火灾事件报告系统（NFIRS）的火灾事件数据为训练依据，并基于其预测2012-2016年间发生的火灾计数的能力进行了评估。 。考虑了两个严格的空间模型：一个简单的“计数”模型，该模型用作本文中描述的所有其他模型的基线；以及一个模型，该模型利用具有统计优化带宽的核密度估计（KDE）。使用美国社区调查（ACS）的数据，对人口和住房因素对火灾风险的影响进行了检查。数据表明，人口普查中的人均火灾风险通常较高，其属性对应于社会经济劣势，例如低中位数收入和一小部分具有大学学历的居民。这些趋势为贝叶斯分层Poisson回归模型的设计提供了依据，相对于基本模型，该模型的预测均方根误差（RMSE）降低了9％。然后，对该回归的残差进行空间核回归，相对于基本模型，RMSE改善了15％。将这些结果与条件自回归（CAR）模型进行比较，它直接将空间信息纳入分层Poisson回归中。尽管对于CAR模型的点估计预测，RMSE较高（比基本模型低7％），但它允许生成概率预测，并提供社会人口统计学变量影响的空间信息统计估计。这项工作强调了地理编码火灾事件和人口统计数据的实用性，以及可以利用这些数据集进行改进的预测的机器学习技术。