Spatially-Aware Temperature Forecasting using Graph Attention and Temporal Convolution Networks

Abstract

Accurate local temperature forecasting is difficult
in topographically complex regions such as Nepal, where atmospheric
behavior is governed by strong spatiotemporal dependencies
that conventional statistical and deep learning models fail to
fully capture. In this study, a new Graph Attention Network-
Temporal Convolutional Network (GAT-TCN) model is proposed
for multi-step temperature forecasting over Bagmati Province,
Nepal. A 3×3 grid of ERA5 reanalysis data is encoded as a
spatial graph, where a GAT learns heterogeneous dependencies
among neighboring grid points while a TCN captures temporal
dynamics from a 24-hour historical window using causal, dilated
convolutions. Spatial and temporal representations are combined
through a late-fusion strategy, and 14-day forecasts are generated
autoregressively. Experimental results show that the proposed
late-fusion GAT-TCN model outperforms GAT, TCN, GCNLSTM,
GAT-LSTM, and an early-fusion GAT-TCN variant,
achieving an R2 of 0.944 and MAE of 0.149°C on Day 1,
degrading gracefully to R2 of 0.768 at Day 14. Transfer learning
experiments using a progressive layer-unfreezing strategy further
show that the pretrained model adapts effectively to mountain
and Terai regions, confirming the generalizability of the proposed
spatiotemporal framework.