Data-Driven Housing Damage and Repair Cost Prediction Framework Based on The 2010 Kraljevo Earthquake Data
Authorized Users Only
Conference object (Published version)
MetadataShow full item record
This paper presents an earthquake damage and repair cost prediction framework for individual residential buildings and portfolios of residential buildings in a municipal area in a region where the seismological networks are sparse and the structural engineering data on the existing residential building stock is poor. The proposed data driven framework is based on the damage and reconstruction data from an actual earthquake, in this case, the M5.4 November 3, 2010 Kraljevo, Serbia, earthquake. It belongs to a more general class of hybrid building portfolio vulnerability models. The earthquake in the model is defined by its magnitude and epicenter location. The geographical distribution of the intensity of the earthquake at the location of the buildings is modeled using the 2013 Akkar-Sandikkaya-Bommer ground motion prediction model suitable for seismically active crustal regions in Europe, with the peak ground acceleration as the intensity measure. The data on the soil type distributi...on was collected form the municipality building department sources. The residential building stock was classified into six types by identifying typical architecture layouts, structural systems and elements. The residential building damage was surveyed after the 2010 Kraljevo earthquake by local engineers using a locally-developed survey form. The form contained the information about the individual damage, classified into four categories ranging from slight damage to collapse, varying amount of building-specific details, and addresses from which geographic locations of the buildings were derived. A random forest machine-learning algorithm was used to derive a predictive model for residential building portfolio seismic damage and repair cost using a portion of the 2010 Kraljevo data as the learning dataset. The model outputs both the individual building fragility and the aggregate portfolio-level vulnerability data. The calculation of the expected repair cost for each building type was done using an expert-defined matrix that specifies average repair costs for each building type and damage category. The model is verified on a separate test portion of the 2010 Kraljevo dataset, yielding a satisfactory relative error when comparing total predicted to total actual repair costs. The model is limited to regions with similar seismicity and similar building stock. However, there many regions in the Balkans that fit this constraint. The proposed framework is, however, more general. It can be applied to other regions with different seismicity and building stock using the data from a recent earthquake as its learning input dataset and an expert-defined repair cost matrix for analyzing the repair cost scenarios.
Keywords:earthquake / damage state / repair cost / machine learning
Source:16th World Conference on Earthquake Engineering (16WCEE), 2017
- Resilient Kraljevo: Management of the Post-Earthquake Community Reconstruction Effort