Fire Alarm Volume Forecasting with ML Models

Forecasting the number of expected fire alarms in different areas for future dates can help authorities to proactively provision resources to manage them. This can be modeled as a time series forecasting problem, and Machine Learning (ML) techniques can be used to analyze historical patterns to produce forecasts. However, choosing a ML model that offers the best forecasting on the given data can sometimes be challenging. Furthermore, accurate predictions are usually unattainable using history alone, especially with the presence of foreign influences. This work investigates the effect of fire alarm device installations on the likely volume of fire alarms generated in a specific geography. The study involves data on the daily reported fire alarm numbers and daily new fire alarm device installation over a period of 15 months in five areas. Multiple ML models are trained and tested on the provided data, and the best-performing model is selected for each area. The results show that the best-performing model can be different for different areas and for different feature sets. Moreover, incorporating additional features related to the installed fire alarm devices boosts the accuracy of fire alarm prediction.