Vehicle traffic forecasting models based on LSTM in application development

The paper investigates the problem of traffic prediction. The issue is important for residents of modern cities and their groups while planning transportations in a city in particular and their own activities in general, taking into account that increased traffic leads to the increase in air pollution. The objective of the research was to study the influence of different input features of an LSTM model and the length of input time periods on the results of traffic prediction in a city. The proposed solutions make it possible to overcome the limitations concerning requirements for a whole system of observation stations and for a large set of historical data. These requirements are common for models proposed in the researches of other authors. In the paper the problem of traffic prediction, which is a matter of a time series forecasting, is considered for the horizon of the next 6 hours with the integration of the created models into applied programs, for which the appropriate data collection and processing procedures are defined. Experimental investigation was conducted using data collected at 59 observation stations and its results were evaluated based on a number of indicators (mean absolute, mean square and root mean square error) of accuracy and on the coefficient of determination for informativeness. The proposed prediction models are built based on long short-term memory. Experimental investigation confirmed the improvement in terms of accuracy and informativeness of the created models with a longer input time interval (24 hours instead of 6) and with additional input features based on traffic data of observation stations, which were selected from the whole set of stations using ensembles of decision trees based on the Random Forest method.