Forecasting urban carbon monoxide concentrations using a hybrid ResNet transformer with STL decomposition
Carbon monoxide (CO) concentration in densely populated urban areas poses significant risks to public health, particularly in cities with high traffic density and limited dispersion capacity. Accurate short-term forecasting of CO concentration enables early warning systems and informed policy responses, yet remains challenging due to the non-stationary, multi-scale nature of pollutant time series. This study proposes a hybrid deep learning architecture combining one-dimensional Residual Networks (ResNet-1D) and Transformer encoders, augmented with Seasonal-Trend decomposition using Loess (STL) and a residual prediction strategy for complex temporal data modeling. Using six years of ground-level CO monitoring data (2019–2024) from the Semarang Mijen Air Quality Monitoring Station, Indonesia, four model configurations are evaluated through a systematic ablation study with multi-seed evaluation (5 independent random seeds). The proposed model (M3) employs a residual prediction approach: rather than forecasting raw CO concentrations, the model predicts only the stochastic residual component from STL decomposition and reconstructs the full forecast by combining the predicted residual with deterministic trend and seasonal components. This strategy achieves R 2 = 0.918 ± 0.019, MAE = 204.3 ± 37.5 µg/m 3 , and RMSE = 315.1 ± 36.9 µg/m 3 on held-out 2024 data, representing a gain of 6.9 percentage points in R 2 over the univariate baseline (M1, R 2 = 0.849 ± 0.035). An optimization-informed Asymmetric Huber Loss (α = 3.0, selected through systematic sensitivity analysis) reduces the under-prediction rate to 15.1% across all samples and 16.2% for peak concentrations (> 2000 µg/m 3 ), reflecting a design philosophy that prioritizes safety–critical decision support. Multi-horizon benchmarking shows that M3 maintains R 2 ≈ 0.90–0.92 from T+1 through T+48 (24 h ahead) under per-partition STL reconstruction, and R 2 = 0.902 at T+1 to 0.838 at T+48 under a more conservative causal reconstruction that avoids target-indexed trend and seasonal values (STL is still fitted on the complete partition), while a HistGradientBoosting Regressor (HGBR) baseline degrades from R 2 = 0.955 at T+1 to 0.050 at T+48 under autoregressive rollout. This reduced degradation relative to the baseline models is associated with the combined STL–residual prediction framework within the evaluated configurations, in which the neural network learns only the approximately stationary residual component. These findings suggest that STL-based residual prediction is a more effective intervention than architectural scaling for multi-horizon CO forecasting in this setting.
Authors
- Muhammad Imam Ghozali (ORCID: https://orcid.org/0009-0007-4773-9207)
- Alif Catur Murti (ORCID: https://orcid.org/0000-0003-4653-0616)
- Wibowo Harry Sugiharto (ORCID: https://orcid.org/0000-0002-5940-0805)
- Ali Bardadi (ORCID: https://orcid.org/0000-0001-9907-2466)
- Agus Sugiharto
Institutions
- Diponegoro University (ID)
- Sriwijaya University (ID)
- Muria Kudus University (ID)
Publication Details
- Journal
- Discover Sustainability
- Published
- 2026-11-15
- DOI
- https://doi.org/10.1007/s43621-026-05283-2
- Primary Topic
- Air Quality Monitoring and Forecasting
- Type
- article
- Field-Weighted Citation Impact
- 0.00