Dataset and Replication Archive for "Planning for Leakage-Aware Collaborative Carbon Mitigation: Optimizer-Informed Multi-Agent Reinforcement Learning in Shandong's Dual Metropolitan Areas"
Abstract
This dataset accompanies the manuscript "Planning for Leakage-Aware Collaborative Carbon Mitigation: Optimizer-Informed Multi-Agent Reinforcement Learning in Shandong's Dual Metropolitan Areas" submitted to the Journal of Urban Planning and Development (ASCE). It contains the full numerical inputs, calibrated artefacts, model outputs, and re-executable source listing used to produce every figure, table, and statistical claim reported in the main text and the Supplemental Materials. CoverageSpatial scope: ten cities within Shandong Province forming the inland Jinan metropolitan area and the coastal Jiaodong metropolitan area.Temporal scope: 2006–2024 (190 city–year panel observations).Policy regime: explicitly differentiated provincial allocation of mitigation burdens across the two metropolitan areas.Data Sources (harmonized into the panel)China City Statistical Yearbook and provincial city statistical yearbooks.China Urban Construction Statistical Yearbook.Shandong provincial energy balance and official energy/industrial publications.City-level CO₂ emission inventories harmonized with the China Emission Accounts and Datasets (CEADs) products of Cai et al. (2025) and Shan et al. (2017, 2018a, 2018b, 2020, 2022).Multi-region input–output (MRIO) tables for inter-city embodied-carbon flows from Wang et al. (2025) and Xia et al. (2024, 2025) — cited in the manuscript and used under the respective publishers' terms; not redistributed here.Contents of Dataset.xlsxThe single workbook consolidates the following sheets (column-level field definitions and units are documented in the in-workbook _README sheet): Panel — panel_city_year, panel_macro. A 190-row city–year panel (log GDP relative to 2024, carbon intensity relative to 2024, CO₂-to-BAU gap, same-year mitigation, structural inflow and outflow, net structural exposure, positive reception pressure, previous action as a share of the city budget, cumulative burden share, horizon fraction, and previous implementation effectiveness), together with province-level macro covariates and lagged GDP-growth and carbon-intensity-growth terms. Leakage network — leakage_network, weights_inverse_distance, weights_exponential_decay. The row-normalized inverse-distance baseline matrix and the exponential-decay alternative retained for sensitivity analysis. Optimizer inputs — optimizer_params, city_action_bounds, burden_coefficients. Calibrated parameters, city-specific action bounds, and implementation-burden coefficients. Reinforcement-learning artefacts — seed_summaries (n = 30), city_trajectories (n = 3,300), training_episodes (n = 9,600), ppo_updates (n = 600), common_shock_draws (n = 11,000). PPO-configured multi-agent training outputs under the disclosed hyper-parameters. Sensitivity — sensitivity_records (n = 618), paired_inference, holdout_predictions. Sensitivity-analysis records, paired-inference outputs, and hold-out predictions. Provenance — provenance_hashes. SHA-256 hashes of every sheet for tamper detection and reproducibility checks. Source listing — numpy_scipy_source, dependency_spec, run_config, analytical_outputs, checkpoint_manifest. Re-executable NumPy/SciPy code, dependency pinning, run-time configuration, analytical outputs, and the checkpoint manifest. Software EnvironmentPython 3.13.x with the exact pinned versions listed in the dependency_spec sheet (NumPy, SciPy, Pandas, Matplotlib). All scripts are deterministic given the seeds in seed_summaries. ReusePermitted under the Creative Commons Attribution 4.0 International License (CC BY 4.0). Please cite both this Zenodo record and the associated manuscript (see "Related works" on the record page). Third-party MRIO data must be obtained from the original publishers under their respective terms. Provenance and IntegrityEach sheet carries a SHA-256 hash in provenance_hashes. Re-hashing after any modification will detect silent drift. Files have MD5 checksums archived by Zenodo at deposit time and are re-verified nightly by the repository. ContactCorresponding author: Zhen Peng — School of Architecture and Urban Planning, Qingdao University of Technology, Qingdao, Shandong, China. Email: see the manuscript Acknowledgments.
