Cost-effectiveness uncertainty may bias the decision of coal power transitions in China

China’s coal power sector, despite recent air pollution control measures, still contributed substantially to national emissions in 2017 (35% of CO2, 17% of SO2, 19% of NOx, and 8% of PM2.5). With increasing electrification demand and diminishing returns from end-of-pipe controls, accelerating the transition away from coal is seen as a key pathway to jointly achieve carbon neutrality and clean air. Historical policies since the 11th Five-Year Plan eliminated over 119 GW of small, old, inefficient capacity, but upcoming retirements are challenging because many units were built around 2010 and have similar lifespans. Evolving policy priorities (health protection, climate mitigation, economic considerations) compound decision complexity. Prior analyses often assume perfect adherence to unit phaseout rankings and underexplore implementation disturbances and uncertainties from the siting of new-built capacity. This study develops a unit-level uncertainty assessment framework for 2018–2060 to evaluate multiple phaseout strategies under implementation disturbances and new-built location uncertainty, quantify costs, climate and health impacts, and assess how uncertainty may bias preference-based decisions.

The study builds on plant/unit-level analyses of coal power phaseout aimed at maximizing health co-benefits or minimizing stranded assets, and multicriteria methods for managing trade-offs among environmental and economic objectives. Prior work identified super-polluting units, estimated stranded assets, and assessed health co-benefits and climate impacts of retirement strategies. However, these analyses often strictly follow deterministic phaseout priorities, with limited consideration of policy implementation disturbances or the uncertainty of new-built unit geolocations. This work addresses that gap by modeling uncertainty in phaseout order and new builds, and by integrating health (GEMM), atmospheric response (GEOS-Chem adjoint), and economic valuation (VSL, carbon pricing) within a Monte Carlo framework to evaluate multiple strategies historically and prospectively.

Data and unit-level characterization: The study uses the China coal-fired Power plant Emissions Database (CPED) for unit locations, capacity, start year, coal consumption rate, and emissions (CO2 and pollutants). Unit-level PM2.5-related premature deaths (2018) are estimated using the Global Exposure Mortality Model (GEMM) for chronic exposure-response and the GEOS-Chem adjoint (v35i) to attribute health burdens to unit emissions via sensitivities to SO2, NOx, NH3, BC, OC, and primary PM2.5. Population data from LandScan 2018 and PM2.5 from the TAP dataset are used; cause-specific baseline mortality rates are from GBD 2019. Death intensity (deaths per MW) is defined to capture unit-level heterogeneity in health risk.

Strategies: Five strategies are defined: Historical (operate units for a 40-year lifetime), BAU (phaseout consistent with historical practice inferred via Cox proportional hazards regression using age, capacity, coal consumption rate), Health (prioritize high death intensity), Carbon (prioritize high CO2 intensity), and Age-to-Capacity (prioritize high age-to-capacity ratio to reduce stranded assets). For all non-Historical strategies, all existing capacity is assumed retired by 2050 (average lifetime shortened by about 25.8 years).

Uncertainty and projection framework: Provincial coal power generation demand consistent with carbon peak and neutrality goals is imposed. A unit-level Monte Carlo simulation (n=10,000) is implemented by dividing units into deciles by each strategy’s priority and randomly disturbing 32% of units’ phaseout order (derived from the BAU Cox model C-index). Annually, within each province, capacity is retired following the (possibly disturbed) priority; generation from in-fleet units is computed using installed capacity and capacity factors (assumed to decline 2.5% per year), with complete phaseout of existing units by 2050. Where existing units cannot meet demand, new-built units with higher efficiency and advanced controls are added; their sites are randomly chosen from retired unit locations. Emissions from new-built units are calculated using provincial emission factors and removal efficiencies (ultra-low emission standards). Health burdens from new-built units are estimated consistent with the unit-level method.

CCUS modeling: Provincial supply curves of CCUS retrofits are constructed using harmonized demand and literature-based CCUS penetration pathways, with large-scale commercialization beginning around 2030 and most capacity retrofitted by 2060. A CCUS algorithm allocates installations preferentially to new-built units; if insufficient, existing units with lower phaseout priority are retrofitted. CCUS assumes 90% CO2 capture with a 15% energy penalty, retrofit capital cost 3300–8500 RMB/kW, and a 15-year life extension, with feedback on capacity needs.

Cost-effectiveness calculation: Annual monetized net benefit for each simulated turnover is Benefit_net = decarbonization benefit + health co-benefit − stranded asset cost. Decarbonization benefit uses a carbon price CP=50 RMB/tCO2 and is computed relative to the Historical strategy. Health co-benefit monetizes avoided premature deaths using province-specific VSLs. Stranded asset costs are estimated as the remaining share of overnight capital cost at retirement relative to a 40-year lifetime (retirements beyond 40 years incur zero stranding). Cumulative values (2018–2060) are computed.

Preference analysis: A multi-criteria decision-making (MCDM) approach normalizes cumulative decarbonization benefits, health benefits, and stranded asset costs (min-max normalization). Weighted preferences α (climate), β (health), and γ (cost saving/stranded asset avoidance) are applied, with α+β+γ=3, to compute a normalized net benefit for strategy selection. Preference analysis is conducted for undisturbed and disturbed cases (using high-uncertainty pathways) to assess how implementation uncertainty biases optimal strategy selection. Sensitivity analyses explore lower demand, faster phaseout, higher health risk, and CCUS prioritization, as well as utilization-hour assumptions.

• Emissions trajectory: Coal power CO2 emissions peak at 4.1–4.3 Gt around 2030 and fall to 31–36 Mt by 2060 (about −99%), aided by extensive CCUS; differences among strategies in 2030 are <5% due to shared identification of low-efficiency units and uniform assumptions about new capacity efficiency.
• Strategy decarbonization differences: Carbon strategy reduces CO2 by about 210 Mt more than Historical in 2030, roughly 1.4 times the reduction achieved by Age-to-Capacity.
• Health outcomes: Health strategy avoids substantially more deaths in the near-to-mid term; annual avoided deaths are 108–150% higher than Carbon strategy during 2019–2040. Cumulatively (2018–2060), Health avoids ~563,000 premature deaths (CI 515,900–611,000), while Carbon avoids ~259,200 (CI 213,100–305,000).
• Net benefits over time: All strategies experience negative annual net benefits initially (2019–2021). Health strategy annual net benefit is −20.6 bn RMB (CI −25.1 to −16.5) in 2019, turning positive by 2022 at +12.2 bn RMB (CI 4.9–19.3). BAU reaches a high-probability positive annual net benefit by 2024; Age-to-Capacity shows a similar trend but with earlier adaptation than BAU.
• Implementation disturbance effects: Disturbances can delay the first year of positive annual net benefits by 3 years for Health and Age-to-Capacity, and by 6 years for BAU (from 2021 to 2027). By 2030, 30–55% of the uncertainty in benefits stems from phaseout implementation disturbance. As existing units retire by ~2050, new-built capacity becomes increasingly dominant (≈65% of generation in 2030 to ≈85% in 2040), making siting and build decisions critical to net benefit and its uncertainty.
• Cumulative net benefits (2018–2060): Health yields the highest cumulative monetized net benefits: 587.9 bn RMB (CI 435.7–739.3). Carbon yields the lowest: −378.1 bn RMB (CI −527.7 to −230.6), largely due to low carbon prices; adopting a higher carbon price (e.g., 75.5 USD/tCO2) could make Carbon competitive (becoming positive and comparable to others).
• Cumulative CO2 reductions: Carbon strategy achieves >5 Gt cumulative reductions relative to Historical when new capacity meets 270 gce/kWh. With additional combustion efficiency improvements, cumulative reductions can exceed 7 Gt (Carbon strategy).
• Risk of negative cumulative outcomes: Under BAU or Age-to-Capacity, there remains a non-zero probability of negative cumulative net benefits due to implementation disturbance. Faster phaseout (all existing retired by 2040) increases this risk by narrowing benefit gaps versus Historical.
• Overlap and divergence in priorities: Limited correlations (R²<0.2) among Health, Carbon, and Age-to-Capacity priority rankings. About 20–30% of prioritized retired units are unique to each strategy; 37.3% (828 units) are shared and represent low-hanging fruits to be retired promptly.
• Preference-based selection: Strategy optimality depends on preferences. Cost-saving emphasis (γ>2) favors Age-to-Capacity. Health emphasis (β>2 and γ<0.5) favors Health (rapid retirement of super-polluting units in dense areas). In balanced cases (α≈β≈γ≈1), BAU can be advantageous due to balancing emissions and stranded assets. Implementation uncertainty can cause mis-selection relative to the true preference optimum (e.g., a preference set α=0.6, β=1.8, γ=0.6 might prefer Health or Age-to-Capacity over BAU under disturbance).

The findings show that unit-level heterogeneity profoundly shapes outcomes of coal power transition strategies across climate, health, and economic metrics. Strategic targeting (Health vs Carbon vs Age-to-Capacity vs BAU) leads to materially different cumulative benefits and costs, and implementation disturbances can materially delay benefits and bias preference-based decisions away from optimal choices. Health-focused phaseouts yield large health gains and the highest net benefits under current VSL and carbon price assumptions, but entail higher costs; they are preferable when public health is prioritized. Carbon-focused phaseouts strongly reduce emissions but yield negative net benefits under low carbon prices; higher carbon prices make them competitive. Age-to-Capacity minimizes stranded assets, offering robust positive normalized benefits across many preferences, but may not be optimal in all cases. The BAU-like strategy balances emission reduction and stranded asset avoidance and may remain practical and feasible given historical experience, but is vulnerable to longer delays from implementation disturbance.

Uncertainty plays a critical role: disturbances in phaseout priority and new-build siting contribute significantly to benefit uncertainty (30–55% by 2030) and can cause missed opportunities to select the best strategy for given preferences. As new-built capacity’s share grows, the design of new builds (location, efficiency), maintaining utilization hours of advanced units, and strategic CCUS prioritization become increasingly important. Sensitivity tests suggest giving CCUS priority to new-built units while retrofitting strong-performing existing units and maintaining utilization hours can increase net benefits and widen differences among strategies, making non-BAU strategies more competitive. Policymakers should tailor phaseout to explicit, time-varying preferences, minimize implementation disturbances, and coordinate phaseout with new-build and CCUS planning to optimize benefits and manage risks.

This study develops a unit-level Monte Carlo uncertainty assessment framework to evaluate China’s coal power transition under multiple strategies aligned with distinct policy preferences. By integrating emissions, health impacts, CCUS deployment, stranded asset costs, and preference-based MCDM, it quantifies how implementation disturbances and new-build uncertainty affect cost-effectiveness and strategy selection. Key contributions include demonstrating substantial health and climate benefits from targeted phaseouts, identifying significant risks of delayed and negative net benefits under disturbance, and providing guidance for preference-aligned strategy selection. Policy implications stress minimizing implementation disturbances, strategically planning new builds and CCUS retrofits, and aligning actions with explicit preferences to avoid missed opportunities.

Future work will address limitations by incorporating higher-resolution and updated unit-level data, improved representations of demand and technology evolution, and periodic (every 5 years, aligned with China’s Five-Year Plans) updates to track cost-effectiveness uncertainty and maintain robust, adaptive phaseout strategies.

• Scenario and model resolution: Coal power demand scenarios from IAMs have limited temporal resolution and may not capture feedbacks from changing cost-effectiveness under different strategies.
• Data gaps: The baseline CPED contains some missing data requiring imputation; while it captures key fleet features, residual uncertainties remain.
• Pollution control retrofits: Future end-of-pipe control retrofits are not modeled (most units already meet ultra-low emission standards); further controls may have limited benefits but substantial costs and CO2 implications.
• Health impact modeling: GEMM and GEOS-Chem adjoint-based estimates carry uncertainties related to mortality risk functions, atmospheric processes, meteorology, and demographic changes, potentially underestimating health co-benefits.
• Parameter uncertainty: The study emphasizes cost-effectiveness uncertainty from phaseout decisions and does not propagate full parameter uncertainty distributions (e.g., emission factors).
• Faster phaseout sensitivity: Accelerated retirement (e.g., by 2040) can increase the risk of negative cumulative outcomes by shrinking the advantage over Historical, highlighting sensitivity to phaseout pace.
• New-build siting: Randomized siting of new builds among retired locations captures location uncertainty but may not reflect all siting constraints or grid/system planning considerations.