For the fire department in The Netherlands, the time between notification and arrival at a fire location should be at most eight minutes for most buildings. This time consist of two parts: the time needed to leave the station (turn-out time) and the time needed to get to the location (travel time). VORtech made an analysis of the turn-out times.
First, we used statistics to investigate the extreme values in the data sets and to see if there is a difference in turn-out times during work hours as compared to after-work hours. We used change point analysis to determine which data we can and cannot use to establish characteristic turn-out times.
Data science analyses like these can be used to optimize the internal processes in the fire department, like determining which stations should approach a fire and how many firemen should be available at any time.
Extreme and erroneous values in data sets
A very important issue in this kind of analyses is how to deal with erroneous and extreme values in the data sets. It can be hard to distinguish these two categories. Turn-out times of zero minutes are probably always erroneous. But a long turn-out time can be real, or it could be caused by poor reporting. It proved to be hard to distinguish erroneous and extreme values with statistics alone. It took expert knowledge to select erroneous data points that should be removed from the data set.
On the other hand, the remaining analyses might also have been done with statistical techniques that are insensitive to extreme/erroneous data points. Such statistical techniques use inter quartile distances (distance between first and third quantile) that indicate the width of the distribution. There is not a single “right”method that can always be used. It’s just a matter of finding a reasonable method.
Making consistent data sets
One of the questions that we were asked was wether turn-out times in the evening are similar to those in the weekend. We found that professional firemen clearly had longer turn-out times after work-hours, but the differences between weekends and workdays was negligable. For non-professional firemen, longer turn-out times were only observed in the night. So, for further analysis in the latter case it is more useful to separate daytime data from nighttime data rather than weekends versus workdays.
Use of historical data sets
Another issue that we dealt with was the fact that historical data sets may have been gathered under different circumstances. Hence, taking all historical data together may lead to serious errors. Yet, a longer time series is preferable because it would improve statistics. To see how far back the data is still consistent with today, we used change point analysis. This is a powerful tool that can spot changes in the underlying statistics of data automatically.
In some locations, we clearly observed change points. In those cases, the fire department can be asked whether the change in statistics can be explained. For example, the staffing or the protocols that are used may have changed. But regardless of the cause of a change, the change points indicate how much of the historical data can be used for analysis of today’s situation.
We have completed a succesful data analysis for the fire department. But in the end, the data was limited. When more data becomes available, the analysis can be done more thoroughly. The methodology can then also be extended and improved. Eventually, this could become an operational software tool for monitoring turn-out times. The first steps have been made and already provide valuable insights.