NEURAL COMPUTING AND APPLICATIONS, cilt.2022, ss.1-16, 2022 (SCI-Expanded)
Passenger demand forecasting is of great importance to the on-demand ride systems. With the accurate forecasting of
demand, it can be determined from which regions and when the passengers demand a vehicle. In this way, passenger and
vehicle waiting times, fuel costs of vehicles can be reduced. In the literature, various models such as time series, long shortterm memory (LSTM), convolutional neural network (CNN), and hybrid of CNN-LSTM are used for demand forecasting
in on-demand ride service systems. These models forecast demands by considering temporal and spatial data separately or
together. In models that use spatial and spatial–temporal data, generally, the city is divided into zones in the form of a grid.
This partitioning method has some disadvantages, such as misleading the forecasting accuracy by considering regions
without demand and ignoring the geographical conditions. In this study, two new models, ConvLSTM2D-clustering and
CNN-LSTM-clustering are proposed to overcome these disadvantages and make more accurate and robust forecasts. The
proposed models use clustering instead of grid partitioning in dividing the city into zones and take time-of-day, time-ofweek variables into account in forecasting as well as passenger demand. The presented models have been used in the
passenger demand forecasting of Turkcell Technology Company, which provides on-demand ride services for its
employees in Istanbul, Turkey. Experimental results, validated on real-world data provided by Turkcell, show that the
proposed models partition the city more effectively and achieve 14–55% better short- and long-term forecasting performances than the compared models in terms of mean squared error.