A cross-channel temporal correlation transformer for diurnal near-surface air temperature prediction from multivariate observations

Abstract Time series forecasting is fundamental to numerous fields, yet long-range dependencies, multivariate coupling, and nonlinear trend capture remain persistent challenges, especially in fine-grained 24 h surface temperature prediction. To address these issues, this paper presents a novel Cross-Channel Temporal Correlation Transformer (CCT). Its core innovations are threefold: (i) A multi-branch adaptive fusion architecture that, for the first time in time series forecasting, dynamically integrates temporally enhanced, multi-scale trend, and original embedding features through a learnable gated attention mechanism, avoiding the information loss inherent in single-branch designs. (ii) A cross-temporal multi-head attention block augmented with relative positional bias and a parallel local convolutional enhancement branch, which jointly capture global patch relationships and fine-grained local dynamics. (iii) A trend-aware attention module that employs a learnable group of multi-scale convolutions and cross-variable attention, enabling adaptive trend decomposition and interaction beyond fixed-scale moving averages. These components enable the decoupled modeling and adaptive fusion of daily-scale weather evolution and hourly-scale details. Evaluated on a meteorological dataset covering six representative Chinese cities across coastal and inland climates, CCT outperforms 12 mainstream models, achieving an average MSE reduction of 6.2% and MAE reduction of 7.5%. In extreme heat conditions such as Chongqing, the MSE reduction exceeds 9%.