Optimizing Ambulance Allocation in the City of San Diego

Time series analysis of ambulance dispatches in the City of San Diego. Models include Prophet, ARIMA, and LSTM with recurrent neuron network.

Dataset file

raw_dispatch_merged_with_comments.csv

Additional Features Created from the dataset

• New_Zone - Map the incident data before Sep 2015 to the new divisions created on Oct 2015, labeled 0~7 as Zone 1 to Zone 8.
• Zipcode - Geo-decoded addresses from lat-long locations of each incident.
• Mission_Type - 0: Ambulance canceled, 1: Cleared at scene, 2: Transported to destination.

Dataset related files:

• incident_zipcode_newzone.csv - Master_Incident_Number with Zipcode, New_Zone, Mission_Type
• geoloc_coord.csv: Address of the 94,307 distince lat-long points in the dataset with address resolved through geopy by OpenStreetMap API.
• amb_hour.csv: Ambulance unit hours dataset, with scheduled and actual hours claimed, the number of calls are also listed.

Jupyter Notebook files:

• geoloc_zipcode.ipynb: Heatmap visualization of 911 calls by zip codes.
• geoloc_api.ipynb: Read geoloc_coord.csv , send the location through geopy to OpenStreetMap, and write the resolved address back to geoloc_coord.csvas an additional column
• zones_types.ipynb: Read raw_dispatch_merged_with_comments.csv , identify the New_Zone for incidents before September 2015, and identify the Mission_Type``. Write the results into incident_zipcode_newzone.csv.
• forecast_prophet.ipynb: Forecasts of the 24 time series by Prophet.
• forecast_uhu.ipynb: Forecast the unit hour time series in amb_hour.csv using Prophet, ARIMA, and LSTM (RNN) with performance comparison.
• forecast_(XY).ipynb: Forecast on selected series in the 24 time series by districts and mission type in incident_zipcode_newzone.csv using Prophet, ARIMA, and LSTM (RNN) with performance comparison. Here, X is the fire district values from 1-8, Y is the mission type from 1-3 (1: Cancelled, 2: Clear-at-scene, 3: Transported)