Apollo Gerolymbos, Head of Data Analytics at London Fire says “Many people see us as an emergency response service, but we would prefer that people stay safe by never having an emergency in the first place. We spend more of our time on prevention work than firefighting so it is important that we target prevention at those people and places where fire is most likely to happen. Last year we visited over 83,000 homes to give free home fire safety advice, with over 68,000 of these visits targeted at people who are statistically considered ‘high risk’”.
LFB commissioned Tudor Thomas, a data scientist through the ASI Data Science fellowship for a six-week project with two phases;
to build an “address catalogue” with details of every building in London
analyse this data to predict fires.
Apollo adds: “Firefighters collect lots of datapoints at the incidents they attend. Coupled with machine generated data and external datasets we find ourselves to be an information rich organisation. It is our responsibility as analysts to find value and insights in this data to contribute to the safety of Londoners and visitors to the capital.”
To find all addresses within London, LFB used the national address gazetteer which they already use for call handling and mobilising and linked it with extra information:
Energy Performance Certificate (EPC) data published by MHCLG
LFB fire history data.
LFBs fire history data already contains the UPRN making it easy to link. The EPC data contains a text address for each record and GeoPlaced matched this data to the UPRN, which resulted in a complete address catalogue with a rich set of attributes to analyse.
Tudor’s 1st step was to examine past fires in one London borough. For example, whether property tenancy could change the fire risk of a building. Statistical methods can test these assumptions and add rigour to the approach.
Unfortunately, only weak relationships could be seen, and no single attribute was a strong enough indicator of fire. By combining characteristics in a predictive model, linked by the UPRN, Tudor could build stronger indicators of risk.
Tudor’s machine learning algorithm also quantified how useful each attribute is once the analysis is complete. The EPC data was found to be very useful, particularly as it gives a strong predictor when other datasets only give weak predictors.
The next step was to re-train the model on historical data for the whole city to find other properties at risk. Tudor also removed all properties which have suffered from a fire in the past. The remaining high-risk households could not have been highlighted by historical data alone.
Apollo said “This analysis gives us the potential to target inspections on a household level and to ensure our fire stations are as prepared as possible for the most likely future demand on our service. Without machine learning techniques we cannot make the best use of the intelligence available to us to target risk”.
Tudor said “The UPRN was the perfect way to link data from multiple sources to the address catalogue. The predictive power of a machine learning model is only as good as the data we put in. To that end, EPC data was important in predicting where London’s next fire might occur.”
The project has been instructive in revealing the potential of machine learning with location data.