Answers are grounded in this paper and cite their sources.
Recent station observations provide a useful basis for air quality index (AQI) forecasting, whereas regional measurements may supply additional context at longer forecast leads. In this study, the forecasts of an independently trained temporal convolutional network (TCN) and a graph-conditioned causal transformer (Graph-CT) are combined. Both use the same input window of 24 h to predict all the hourly leads from 1 to 24 h, and their corresponding outputs are averaged. Air-quality records from 26 stations in the Chengdu–Chongqing economic circle are reconstructed on a common hourly grid, with causal input filling and evaluation restricted to observed targets. Learning rates and training durations are selected on 2021 data, followed by retrospective evaluation on the previously inspected 2022 benchmark. Across five runs, equal fusion achieves a pooled root mean square error (RMSE) of 18.036 ± 0.103 AQI, compared with 19.491 ± 0.063 for TCN, 18.284 ± 0.140 for Graph-CT, 19.413 ± 0.174 for iTransformer, and 19.043 ± 0.455 for Graph WaveNet under the common task. The paired interval includes zero, leaving the advantage over a two-member Graph-CT ensemble unresolved. Validation calibration slightly decreases the RMSE but increases the mean absolute error (MAE), and removing the transformer branch improves the evaluated configuration. Forecast averaging is therefore useful on this benchmark, although the results favour neither equal coefficients as an optimum nor the full architecture over its simpler variants. The evaluation of untouched periods and regions remains necessary for assessing generalization.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: