This is a project which is currently making use of HPC facilities at Newcastle University. It is active.
For further information about this project, please contact:
Scalable Spatiotemporal Graph Learning for Urban Sensor Networks
This project focuses on developing advanced deep learning models to analyse and forecast complex spatiotemporal patterns across large-scale urban sensor networks. By modeling the city as a dynamic graph, we aim to capture non-Euclidean dependencies between traffic flow, weather conditions, and air quality using Graph Neural Networks (GNNs).
Specifically, the research involves training Spatiotemporal Graph Attention Networks (ST-GATs) and Masked Autoencoders (MAE) on high-frequency data streams. The goal is to create robust, self-supervised representations of urban dynamics that can improve forecasting accuracy even in the presence of missing or noisy sensor data. The project utilises a containerised high-performance computing pipeline to handle the significant computational requirements of training large-scale graph models on historical datasets.
This project utilises a containerised Python 3.11 deep learning stack designed for the GPU-L cluster (NVIDIA H100s). The core software framework is PyTorch with PyTorch Geometric, utilising NVIDIA CUDA (v12.x) for GPU-accelerated training.
The workflow is encapsulated in Docker to ensure reproducibility and is configured to perform high-throughput I/O operations directly on the SCRATCH and NOBACKUP fast storage volumes to avoid latency bottlenecks.
The processing pipeline involves:
Data Engineering: Construction of large-scale sensor graphs using NetworkX, Pandas/Polars, and GeoPandas.
Deep Learning: Training complex spatiotemporal architectures, specifically Graph Attention Networks (GATs) and Masked Autoencoders (MAE)."