A Comparison of LSTM and XGBoost for Predicting Firemen Interventions

In several areas of the world such as France, fire brigades are facing a constant increase in the number of their commitments, some of the main reasons are related to the growth and aging of the population and others to global warming. This increase occurs principally in constant human and material resources, due to the financial crisis and the disengagement of the states. Therefore, forecasting the number of future interventions will have a great impact on optimizing the number and the type of on-call firefighters, making it possible to avoid too few firefighters available during peak load or an oversized guard during off-peak periods. These predictions are viable, given firefighters' labor is conditioned by human activity in general, itself correlated to meteorological data, calendars, etc. This article aims to show that machine learning tools are mature enough at present to allow useful predictions considering rare events such as natural disasters. The tools chosen are XGBoost and LSTM, two of the best currently available approaches, in which the basic experts are decision trees and neurons. Thereby, it seemed appropriate to compare them to determine if they can forecast the firefighters' response load and if so, if the results obtained are comparable. The entire process is detailed, from data collection to the predictions. The results obtained prove that such a quality prediction is entirely feasible and could still be improved by other techniques such as hyperparameter optimization.