Spatiotemporal Population Flow Forecasting via Temporal Fusion Graph Convolutional Neural Network: A comparative analysis

In this study, we aim to predict the daily county-level population flow time series extracted from LBS data across the contiguous US. We introduce a popular deep learning framework, Temporal Fusion Graph Convolutional Neural Network (TFGCN), for multivariable and multi-horizon population flow time series forecasting. The main scopes of our study include two parts:

• First, to extensively compare the performance of different cutting-edge TFGCN models in predicting population flow time-series.
• Second, to analyze the effects of adjacency matrices on model performance and determine the best that can describe the spatial structure of the nationwide population flow.

Environment

We use the torch >= 1.10.2 and Python 3 for implementation.

We follow the framework of pytorch_geometric_temporal to prepare data and run the model.

Please execute the following command to get the source code:

git clone https://github.com/SonghuaHu-UMD/PA_Forecast_PGT
cd PA_Forecast_PGT

Data Preparation

Unzip the data.7z and put it into the data folde, and then you can get:

• us_node_values_pre.npy: temporal node signal, i.e. daily population inflow across the US counties from SafeGraph.
• us_node_static_pre.npy: static node feature.
• us_adj_mat_dist_1_pre.npy: adjacency matrix measured by distance.
• us_adj_mat_fb_pre.npy: adjacency matrix measured by Facebook social connection.
• us_adj_mat_gravity_pre.npy: adjacency matrix measured by gravity model.
• us_adj_mat_od_pre.npy: adjacency matrix measured by OD volume,

The folder data_prepare contains the codes for data processing.

Model Training

The script 1.2-Model_multime.py is used for training and evaluating the main model. After getting all the data prepared, run this script and you will get:

• Trained models saved at checkpoint.
• Results saved at Results.

The script 1.3-Model_Best_Result_Output.py is used to transform the best prediction to original scale.

Results

Model performance across different models

Model performance across different adjacency matrices

Top and last three counties' forecasting results