Hourly air temperature forecasts get the humidity-wind-pressure coupling early: six Chinese cities cut average error 6.2%
Synopsis
Near-surface air temperature forecasts 24 hours ahead must capture the daily cycle, synoptic evolution and hourly fluctuations at once, while common decomposition transformers defer or hide cross-variable interactions. A cross-channel temporal correlation transformer exchanges information among meteorological channels at the embedding stage via 1×1 convolutions, then fuses temporal, trend and raw branches with gated attention and multi-scale trend attention. On six Chinese city datasets it lowers MSE by 6.2% and MAE by 7.5% on average against mainstream models, with MSE reduction above 9% in Chongqing's extreme heat.
Interpretation
A 24-hour near-surface temperature forecast can now couple variables at the input stage, letting temperature, dew point, pressure and wind speed exchange information before the network's later layers. Earlier decomposition and transformer forecasters typically deferred cross-variable interaction to late stages or modeled it implicitly, extracted trends with fixed-scale moving averages or single-kernel convolutions, and fused features by simple addition or concatenation. On meteorological data from six Chinese cities spanning coastal and inland climates, the model beats mainstream baselines on all 12 MSE and MAE metrics; the abstract reports average MSE down 6.2% and MAE down 7.5%, while the conclusion cites 13 mainstream baselines, so the two baseline counts disagree.
The forecast is split into parallel temporal, trend and raw branches whose weights are set dynamically from meteorological context, so daily-cycle detail and weather-scale evolution are learned separately. Single-branch designs lose information, and fixed-scale decomposition struggles to cover hourly fluctuations and weather-process trends at the same time. In ablations, removing gated fusion raises Changsha MSE from 0.5464 to 0.6381 and removing positional encoding raises it to 0.6258; a trend-only branch reaches MSE 0.3869 in Macau versus 0.3431 for the full model.
The trend branch replaces fixed-scale moving averages with a learnable bank of multi-scale convolutions (kernel sizes 3, 5, 7, 9) plus cross-variable attention. Existing methods rely on fixed-scale moving averages or single-kernel causal convolution, covering only one trend scale. In Macau, Chengdu and Fuzhou, variants without multi-scale trend extraction or trend fusion attention mostly stay within 10% MSE increase, yet both consistently trail the full model.
Perspective
Results cover three coastal cities (Macau, Guangzhou, Fuzhou) and three inland cities (Chengdu, Chongqing, Changsha), with a fixed 24-hour forecast horizon and 48-hour inputs of 14 meteorological variables. Data come from the NOAA National Centers for Environmental Information GHCN-Daily public station records, split chronologically into 70% training, 10% validation and 20% test. The model has about 1.8 million trainable parameters and trains in roughly 2 hours on a single RTX 4090 on the Macau dataset. The authors propose future work on physics-informed neural network constraints, graph neural networks for inter-station spatial modeling, and multimodal fusion with satellite remote sensing and numerical weather prediction products.
Conclusions are limited to six Chinese cities and a 24-hour horizon; generalization across regions and seasons is not tested on a larger station set. The authors note extreme peaks remain hard to predict because of their abruptness, and the model mainly preserves the direction of change and avoids over-smoothing. Ablations show the largest degradation when positional encoding or gated fusion is removed, so performance is sensitive to those two designs. The abstract cites 12 mainstream models while the conclusion cites 13 mainstream baselines, and the source does not explain the difference. Physics consistency, joint spatial modeling, extreme-event losses and multimodal fusion are listed as future work and not yet implemented.