Pairwise Convolutional Relationship Modeling with Learned and Fixed Weighted Aggregation for Multi-Horizon Temperature Forecasting

In weather forecasting, secondary variables such as humidity, wind speed, pressure, solar ra-
diation, and cloud amount contain distinct relationships with and predictive information about a
primary target such as temperature. Rather than immediately combining all variables into a sin-
gle global representation, this study proposes first learning the temporal relationship between
temperature and each secondary meteorological variable independently. Each temperature–
secondary-variable pair is processed by a one-dimensional convolutional network over a fixed
historical window and produces an intermediate temperature forecast. The individual pairwise
forecasts are then combined through a weighted sum, allowing the final prediction to incorpo-
rate information learned from every target-specific relationship. Inspired by relational multi-
variate forecasting methods such as MTGNN(Multivariate Time Series Graph Neural Network),
the proposed architecture replaces global all-to-all interaction modeling with target-conditioned
pairwise relationship extraction. Separate direct models are trained for forecast horizons from
one to seven days, and both learned and fixed aggregation weights are examined to distinguish
the contribution of the pairwise models from that of the final combination mechanism