The Application of Machine Learning Methods to Time Series Forecasting: Improving Forecasting Techniques for Smart City Planning in New York City

Abstract

Smart Cities strive to leverage information and communication technologies in order to analyze the growing amount of available data on the ecosystems of modern cities with the aim of making central infrastructure components and services of a city more interconnected and efficient. A key element of this is knowledge about future conditions which can be obtained through forecasting models. The most recent developments seen in the M4 and M5 forecasting competitions illustrate how far Machine Learning methods have evolved. Therefore, this paper compares and contrasts established statistical forecasting models, ensemble methods as well as recurrent neural networks with respect to point forecasts and prediction intervals. Using publicly available high-frequency data from the New York City Open Data platform, we analyze large univariate time series of traffic flows and Emergency Medical Services (EMS) data. We detect multiple levels of seasonality in the high-frequency data considered in this paper, which complicates finding suitable forecasting models. We find that recurrent neural networks produce the most accurate point forecasts and uncertainty measures for traffic data. For the EMS data, our results show that both recurrent neural networks and Exponential Smoothing state space models which can account for complex seasonal patterns perform best with respect to point forecasts, whereas the quality of the prediction intervals of the latter is the highest.