Przeskocz do nawigacji głównej Przeskocz do wyszukiwania Przeskocz do głównej treści

Graph Neural Network via Dynamic Weights and LSTM with Attention for Traffic Forecasting

  • Silesian University of Technology

Wyniki badań: Wkład do czasopismaArtykuł z konferencjirecenzja

5 Cytowania z bazy Scopus

Abstrakt

In the problem of predicting values, various methods, regression or recurrent techniques are used, which can predict values based on previous ones with great efficiency. In our research, we propose our own algorithm to predict the situation on the streets, which makes it possible to create a system capable of determining the fastest and optimal routes for emergency vehicles in the event of a sudden reported accident, taking into account the occurrence of unexpected traffic jams or congestion on the road. Recently, Graph Neural Network (GNN), which can analyze and process relationships in graphs, have gained popularity. These layers use various aggregation functions to collect information and exchange it between different nodes and their neighbors. In addition, popular networks such as Long Short-Term Memory Network (LSTM), Gated Recurrent Unit model (GRU) or classical Recurrent Neural Network (RNN) are used in the problem of time series forecasting. In our research, we propose an algorithm consisting of two parts: the first part is responsible for spatial processing of data using graph layers and a novel aggregation function using dynamic weights, the second part: temporal processing using LSTM with additional attention. We show that in the problem of traffic volume forecasting, spatial analysis is also important, where streets, intersections can be correlated with each other, and transferring this information to the LSTM model improves and generates better results. In the end, our architecture outperforms typical and popular solutions that do not include attention mechanisms, obtaining MAE = 6.33, RMSE = 3.52, MAPE = 8.99% for the PeMSD7(M) dataset and MAE = 6.64, RMSE = 3.68, MAPE = 9.43% for the PeMSD7(L) dataset.

Język oryginałuangielski
Strony (od–do)119-126
Liczba stron8
CzasopismoProcedia Computer Science
Tom257
Identyfikatory DOI
Status publikacjiOpublikowano - 2025
Wydarzenie16th International Conference on Ambient Systems, Networks and Technologies Networks, ANT 2025 / 8th International Conference on Emerging Data and Industry 4.0, EDI40 2025 - Patras, Grecja
Czas trwania: 22 kwi 202524 kwi 2025

Obszary tematyczne ASJC Scopus

  • Informatyka ogólna

Fingerprint

Zanurz się w tematy badawcze publikacji „Graph Neural Network via Dynamic Weights and LSTM with Attention for Traffic Forecasting”. Razem tworzą niepowtarzalny odcisk palca.

Cytowanie