HESTNet: Heterogeneous Ensemble Stacking Network for Top-Down Monthly Urban CO2 Emission Estimation Using Electricity-Centered Multi-Source Data

Existing top-down carbon emission estimation studies often rely on single models or conventional ensemble approaches, which may limit their ability to capture complex nonlinear relationships under small-sample conditions. To address this limitation, this paper proposes a heterogeneous ensemble stacking network (HESTNet) for city-level monthly CO2 emission estimation using electricity-centered multi-source data. A candidate feature set integrating sector-specific electricity consumption, socioeconomic statistics, nighttime light data, and environmental and climatic variables is first constructed, from which 28 core features are selected using autoencoder reconstruction errors. A heterogeneous stacking ensemble comprising CatBoost, TabPFN v2, and TabM is then developed, with five-fold out-of-fold predictions fused by an L2-regularized Ridge meta-learner. Because reliable ground-truth city-level monthly CO2 observations are generally unavailable, the model is trained and evaluated using 360 province-year samples from 30 provincial-level regions in mainland China during 2013–2024, with provincial annual CO2 emissions used as supervised labels. At the supervised province-year evaluation scale, HESTNet achieves an R2 of 0.914, RMSE of 0.0781, MAE of 0.0580, and MAPE of 9.38%. Compared with the conventional stacking baseline, HESTNet yields numerical improvements of 0.028 in R2 and approximately 12.6%, 13.2%, and 14.0% in RMSE, MAE, and MAPE, respectively; however, the paired RMSE difference does not reach statistical significance after Holm–Bonferroni correction (adjusted p = 0.061). The trained model is subsequently applied to city-month-scale predictors to generate relative Emission Proxy Indices (EPIs). Under the constraint of official city-level annual CO2 emissions, the EPIs are normalized into monthly allocation weights to derive model-derived, annual-constrained monthly CO2 estimates. The proposed framework integrates electricity-centered multi-source proxies, heterogeneous ensemble learning, and annual-constrained temporal disaggregation, providing a data-driven approach for characterizing intra-annual variations in city-level CO2 emissions.