A multi-modal geospatial–temporal LSTM based deep learning framework for predictive modeling of urban mobility patterns

Abstract

Urban mobility prediction is crucial for optimizing resource allocation, managing transportation systems, and planning urban development. We propose a novel framework, GeoTemporal LSTM (GT-LSTM), designed to address the intricate spatiotemporal dynamics of urban environments. GT-LSTM integrates temporal dependencies with geographic information through a multi-modal approach that combines attention mechanisms and Recurrent Neural Networks (RNNs). This method allows the model to focus on relevant spatial features while capturing sequential relationships in time-series data. The approach uses attention mechanisms to dynamically weight geographic features and LSTM layers to model temporal patterns, resulting in enhanced predictive accuracy. Evaluations using a real-world multi-modal urban transportation dataset demonstrate the performance of GT-LSTM, with significant reductions of 15% in Mean Absolute Percentage Error (MAPE) and 20% in Root Mean Square Error (RMSE) compared to traditional methods. The model also shows substantial improvements over traditional techniques, including Convolutional LSTM and Graph Convolutional Networks. The effectiveness of GT-LSTM in capturing both spatial and temporal dynamics highlights its potential for real-time urban mobility prediction and provides valuable insights for urban planners, policymakers, and transportation authorities to improve decision-making and system efficiency.