Simulating the call streams to an emergency services call centre

Forecasting inbound call patterns for an emergency services call centre is an important issue with such facilities being an integral part of the community. They operate under stringent parameters and facilitate rapid support to their communities. The literature identifies call centre staff costs as being a significant proportion of the centres operating budget and our research focuses on their efficient allocation. To facilitate this an agent based model of a multi-queue call centre is being developed which will be used to investigate improvements to the business by generating practices that lead to better customer service level metrics at reduced costs through more efficient resource allocation. The NSW Police Assistance Line (NSWPAL) is a multi-site, multi-queue, multi-server emergency services call centre that takes urgent and non-urgent calls from people in the state of New South Wales, Australia. Extensive, and previously unused historical data of call streams for NSWPAL is used as a basis for generating the call queues. In this paper our previous work is extended and focus is on the call stream generation for the model in 15 minute intervals for time-of-day and day-of-week up to a 24 hour period within a day. We identify three methods of determining the number of calls in a 15-minute block; the long-term mean and standard deviation, a cubic equation estimator based on the long-term mean and standard deviation and the Fast Fourier Transform (FFT). The mean and standard deviation of each block for the day of the week is taken from the four years of historical data and used in this paper to show the generation of the call stream within a day of the week. The model is tested by comparing the predicted call stream with actual data. Being an emergency service call centre means that there will be occasion when exogenous events in the community result in unexpected increases in the call rate. By analysing the root mean square (RMS) difference between the simulated time series and the data, such occasions can be identified.