A moving forest model to predict the building-level progression of ordinance-mandated seismic retrofits

Abstract

An algorithm is formulated to dynamically predict the progression of policy-mandated building-specific seismic retrofits over time. Elements of the widely utilized moving forecast model is integrated with the random forest machine learning algorithm. An evolving duration-based window structure is adopted, which is partitioned into training and forecasting phases, each of which is further subdivided into smaller increments of time. Inside the training window, the time since the passage of the ordinance is combined with building, socioeconomic and demographic variables to create a feature matrix. The outcome vector comprises the state of progression of each building (as defined by the ordinance and/or modeler) in the affected inventory at pre-defined time instants. The data subset in the training window is used to construct the random forest classification model that predicts the retrofit progression states within the forecasting part of the window. This process is repeated while incrementally increasing the fraction of the ordinance time horizon that is encapsulated within each subsequent window. The “moving forest” algorithm is implemented on the wood frame building inventory that is under the purview of the Los Angeles Soft-Story Ordinance. For this specific application, the number of units in the building, the percentage of owner-occupied units (at the census-block level) and the time since the passage of the Ordinance, are found to be the most important features. The classification accuracy ranged from approximately 53% during the first six months to 95% in the final year of the considered four-year time-horizon. The proposed model can be used by departments of building and safety to manage the flow of permit and retrofit certification applications associated with an ordinance or for dynamic assessments that aim to model the risk reduction in the affected inventory over time.