Functional Data Analysis for Prediction of Employee Presence

The problem of space occupancy and the prediction of occupancy schedules has rarely been addressed, due to the highly stochastic behavior of occupants and the insufficient quality of data sets. The COVID-19 pandemic has challenged traditional models based on the physical presence of employees in the workplace. This crisis has accentuated the evolution of new trends, such as teleworking, and highlighted the inherent diversity of work environments, as well as the multitude of factors that influence employee deplacements. The ability to anticipate when people will come is essential for effective management of these spaces, particularly in terms of resource planning, safety, and comfort. As a result, predicting when employees will get to work is a complex challenge. Accurate presence prediction tools can optimize resource utilization and support strategic decisions in corporate real estate. This paper proposes a robust approach combining functional data analysis with the integration of exogenous variables adapted to time series (RobFTS). The purpose of this approach is to anticipate the future presence of employees in their workplace. The model is developed using lags of functional time series, as well as external variables such as weather data (temperature, precipitation) and meeting rooms reservations. In addition to standard error metrics, we apply a comprehensive model selection process using AIC, BIC, AICc, HQIC and other information criteria adapted for non-parametric models. This method is then applied to a real database, gathering daily observations of employee presence at a corporate site in Paris. Finally, we contrast this method with other models widely used in time series analysis, Seasonal Autoregressive Integrated Moving Average with Exogenous Regressors (SARIMAX), Long Short Term Memory (LSTM), Adaboost, and Deep Learning for functional data analysis (DFMM). It is found that the mean absolute percentage error (MAPE) is about 4.99% for RobFTS, compared to 17.30%, 12.92%, 11% and 6.50% for SARIMAX, LSTM, Adaboost and DFMM, respectively. The data on which we applied the different models is made available on GitHub<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><i>a</i></sup>, along with the various scripts of the models.