Розробка нейронної мережі для прогнозування пасажиропотоків в громадському електротранспорті розумного міста

The research object is a hybrid deep learning model for passenger flow forecasting. These passenger flows constitute complex time series, influenced by a combination of temporal, spatial, and operational factors. The study addresses the fundamental mismatch between stochastic passenger demand and the static supply of transport services. This disparity results in operational inefficiency and a reduced quality of service for passengers. A lack of accurate forecasting tools hinders the optimal daily allocation of rolling stock, thereby limiting the efficiency of transport operators. A hybrid deep learning model was developed and validated to predict daily passenger flows with high accuracy (R² = 0.91). The findings significantly outperform the baseline models and approaches described in scientific sources. This performance is attributed to a sophisticated strategy combining advanced feature engineering. This included the use of cyclic, lagged, and moving average features. This approach was paired with residual modelling, enabling the neural network to capture complex non-linear deviations. Furthermore, robust data preparation methods enhanced the model’s high generalization capabilities. The findings demonstrate that the proposed hybrid approach is an effective tool for operational planning. The results of the neural network work facilitate the optimization of the distribution of rolling stock allocation and improve resource utilization. Consequently, it enhances passenger comfort, contributing to the sustainable development of urban mobility. For practical applications, the model requires reliable historical passenger flow data. It enables operators to mitigate economic losses from underutilized vehicles and prevent overcrowding on high-demand days.