Assessment and pathways of the energy production revolution in the Yellow River Basin, China towards carbon peaking: a machine learning approach

Regional classification

Drawing on the categorization approach proposed by Wang et al.⁴⁶, this study categorizes the nine provinces of the Yellow River Basin into three types: hydro-rich, wind-rich, and PV-rich provinces, based on differences in their renewable electricity generation structures. The specific classification criterion is the share of each province’s hydro and wind power generation in the basin’s total hydro and wind power generation in 2021. The three provinces with the highest respective shares are selected as representative regions (Fig. 2a, b). Except in isolated years, the hydro or wind power generation in these provinces consistently accounts for over 50% of their total renewable energy generation, indicating strong representativeness. Accordingly, Sichuan, Qinghai, and Gansu are classified as hydro-rich provinces; Inner Mongolia, Shanxi, and Shandong as wind-rich provinces; and Ningxia, Henan, and Shaanxi as PV-rich provinces.

Analysis and evaluation of the EIT index

The EIT index comprises 4 first-level indicators and 8 second-level indicators. The target values (a) for each second-level indicator in 2045 are primarily derived from goals set in relevant policy documents such as the Energy Production and Consumption Revolution Strategy (2016–2030) and the Guidelines on Fully and Faithfully Applying the New Development Philosophy to Achieve Carbon Peaking and Carbon Neutrality, estimated via interpolation. The specific indicator weights (W) are shown in Table 3.

As shown in Figure 3, this study categorizes the EIT index into four performance levels: Excellent (>90), Good (80–90), Average (70–80), Fair (50–70), and Poor (<50). From 2010 to 2023, the EIT indices for the entire basin and for the hydro-rich, wind-rich, and PV-rich provinces all exhibit a continuous upward trend. However, by 2023, they remain at either a Good or Fair level, indicating a considerable gap from the Excellent level.

During the 2010–2023 period, the EIT indices for the entire basin and the three province types displayed distinct phased evolutionary characteristics. The period 2010–2011 enters an Initial Stage, where the EIT indices for the entire basin and the three types of major provinces remain largely unchanged. This is influenced by a path-dependency effect stemming from long-term reliance on coal resources, locking firms into high sunk costs. Meanwhile, insufficient marginal returns in the early phase of low-carbon transition lead to continued resource allocation favoring traditional industries, constraining low-carbon investment. Together, these factors result in extremely slow growth in the share of non-fossil energy in production and the share of renewable electricity generation, alongside difficulties in reducing the carbon emission intensity of power production.The period 2012–2016 transitions into a Rapid Breakthrough Stage, characterized by a significant rise in EIT indices across the basin and all provinces. Policy incentives foster a virtuous cycle of ”policy-investment-growth,” with average annual growth in energy industrial investment reaching 28.74%. Initial successes in industrial restructuring, including energy-saving retrofits in high-energy-consuming industries and industrial chain upgrades, contribute to a 4.33% decline in energy intensity.From 2016 to 2019, the trend shifts to a Consolidation and Adjustment Stage, with EIT indices showing fluctuations. On one hand, the subsidy-driven model faces diminishing marginal returns and fiscal pressures. On the other hand, inadequate inter-provincial power coordination and lagging transmission infrastructure lead to high wind and solar curtailment rates in provinces like Gansu, Inner Mongolia, and Qinghai, directly limiting increases in the share of renewable electricity generation.The period 2019–2022 enters a Strategy-Driven Stage, where EIT indices continue to rise overall. The Dual Carbon goals and the strategy for Ecological Protection and High-Quality Development in the Yellow River Basin form a synergistic policy force, propelling the transition forward.The period 2022–2023 moves into a Plateau Stage, with EIT indices stabilizing or even experiencing slight declines. Against the backdrop of the global energy crisis, China’s emphasis on securing coal supply partially slows the phase-out of traditional energy and the cleansing of the energy structure. Simultaneously, tightening constraints from water resources and population indirectly limit further reductions in energy intensity and improvements in energy and processing conversion efficiency, slightly weakening the driving force for low-carbon energy transition.

The EIT indices for hydro-rich provinces consistently remain above the basin’s average, whereas those for PV-rich and wind-rich provinces persistently lag below the average, with PV-rich provinces slightly outperforming wind-rich provinces. The primary reason lies in the fundamental differences in the technical characteristics and system integration capabilities of various renewable energy technologies.Hydropower, characterized by zero-carbon generation and flexible regulation, enhances the EIT index across multiple dimensions by ensuring system stability and improving energy efficiency. In contrast, the output of wind and solar power is highly intermittent and variable. Due to insufficient energy storage facilities and limited system regulation capacity within the basin, their integration still heavily relies on thermal power for peak shaving and backup, resulting in persistently high carbon emission intensity and consequently lower EIT indices (Fig. 3).

Analysis of the energy production revolution effectiveness prediction model

Analysis and processing of influencing factors for the energy production revolution

Considering that different factors exert varying degrees of influence on the effectiveness of the energy production revolution, it is necessary to analyze and select these factors prior to constructing the prediction model. This step aims to enhance model prediction performance, reduce overfitting, simplify the model structure, and ensure the scientific validity of the selected features²⁵.

(1) Variable correlation analysis

Based on the dataset of 15 influencing factors and the EIT index for the nine provinces in the Yellow River Basin, Spearman rank correlation coefficients between each influencing factor and the EIT index were calculated. The specific results are presented in Fig. 4.

As shown in Fig. 4, the Renewable Energy Power Consumption Ratio (PEC), the Number of Patents Granted (PAN), and R&D Expenditure in the Electricity, Heat Production and Supply Industries (RDEH) all show significant positive correlations with the EIT index, with correlation coefficients exceeding 0.7. This indicates that improvements in these factors directly contribute to an increase in the EIT index. Additionally, Natural Gas Production (NGO), the Share of 300MW and Above Units in Thermal Power Installed Capacity (PTP), and Mining Ecological Restoration Funds (FERM) exhibit correlation coefficients between 0.2 and 0.5 with the effectiveness, suggesting a certain degree of positive association. On the other hand, Coal Consumption for Power Generation (CCPE), the Fatality Rate per Million Tons of Coal (CMR), and Raw Coal Production (PVC) show negative correlations with the EIT index, indicating that coal-related factors exert an inhibitory effect on the improvement of the EIT index. The study also finds that correlation coefficients between variables such as Provincial Resident Population (PRP), Urbanization Level (UL), Per Capita GDP (PCG), and R&D Expenditure (RDEH) are generally above 0.4, suggesting potential multicollinearity issues among these variables. Therefore, to enhance model robustness and predictive capability, subsequent analysis will employ PCA for dimensionality reduction to mitigate multicollinearity and optimize the model structure.

(2) PCA processing

The results show a KMO statistic of 0.770, and Bartlett’s test yields a \(\chi ^2\) value of 1189.037 (p < 0.001), indicating that the dataset is appropriate for PCA. Following PCA, 11 principal components were extracted, with a cumulative variance contribution rate reaching 96.25%. This demonstrates that these principal components adequately represent most of the information from the original variables. Consequently, in the subsequent construction of the prediction model, these 11 principal components will be used as new input features, replacing the original 15 influencing factors. This approach simplifies the model structure while reducing the impact of multicollinearity.

Performance evaluation and comparison of EIT index prediction models

(1) Construction and analysis of the EIT index prediction model

After dimensionality reduction via PCA, the dataset is first divided into training and testing sets at a ratio of 8:2 to ensure the independence of the test set. Subsequently, five-fold cross-validation is conducted within the training set for model training and evaluation.

On this basis, three types of prediction models are constructed: first, single models (M1), including SVR, CatBoost, KernelRidge (KRR), ElasticNetCV, and RidgeCV; second, simple averaging ensemble models (M2), where SVR, CatBoost, and KernelRidge are combined in pairs and integrated using equal weights for prediction; third, stacking ensemble models (M3), where SVR, CatBoost, and KernelRidge are used in pairwise combinations as base learners, and ElasticNetCV or RidgeCV is adopted as the meta-learner for second-stage fusion modeling. The hyperparameters of the base learners are selected within reasonable ranges through preliminary experiments and empirical settings to balance model complexity and generalization ability. Considering the limited sample size, large-scale hyperparameter tuning is not performed to avoid overfitting. The parameters of the meta-learner are optimized via cross-validation, thereby improving the overall robustness and predictive performance of the model. The model performance is evaluated using the RMSE, MAE, and \(R^2\) on the test set. The evaluation results for the different models are shown in Figure 5. The results indicate that the stacking ensemble model(M31) constructed with SVR and KernelRidge as base learners and RidgeCV as the meta learner performs best, achieving the RMSE of 3.05, the MAE of 2.58, and the \(R^2\) of 0.95 on the test set.

(2) Robustness analysis based on repeated cross-validation

To further evaluate model robustness, this study performs 100 repetitions of five-fold cross-validation on the full sample for the three types of prediction models mentioned above, following the approach of Kim⁵⁷. For each split, RMSE, MAE, and R\(^2\) are calculated on the validation folds, and the results are reported as mean ± standard deviation (Table 4).The results show that model M31 (SVR + KernelRidge / RidgeCV) achieves the best performance across all metrics, with an RMSE of 2.9957 ± 0.5087, an MAE of 2.1948 ± 0.3891, and an R\(^2\) of 0.9530 ± 0.0188, all outperforming the other models. This indicates that M31 maintains more stable predictive performance under different training/validation splits.

(3) Statistical significance test and confidence interval analysis based on EPA

To evaluate the statistical significance of predictive performance differences between the stacking ensemble models and the benchmark model M31 (SVR + KernelRidge / RidgeCV), this study adopts the Expected Prediction Accuracy (EPA) framework proposed by Akgun et al.⁵⁸. Under a single five-fold cross-validation setting, Out-of-Fold (OOF) predictions for the full sample are obtained, and MSE is used as the loss function. Based on the loss differential series, the EPA test is conducted to assess the statistical significance of differences in predictive performance. A negative EPA statistic indicates that the benchmark model M31 outperforms the competing models.

Step A—cross-sectional dependence test

To examine cross-sectional dependence in the loss differential series, the Pesaran CD test is applied to the OOF residuals of the benchmark model M31. The results show that the average cross-sectional correlation coefficient is close to zero, with a CD statistic of -0.1979 (p > 0.05). The null hypothesis of no cross-sectional dependence cannot be rejected, indicating that prediction errors do not exhibit significant cross-sectional dependence.

Step B—autocorrelation adjustment

Considering the short time dimension and potential temporal dependence in the loss differential series, Newey-West HAC⁵⁹ standard errors are employed for correction to enhance the robustness of statistical inference. Additionally, under small-sample conditions, statistical significance is tested using a t-distribution with degrees of freedom equal to T-1.

Step C—EPA test

Although no significant cross-sectional dependence is detected in Step A, the \(S_{nT}^{(3)}\) statistic is employed to ensure robust overall inference.As shown in Table 5, the EPA statistics of M31 relative to the other models are all negative and statistically significant at the 5% level (p < 0.05). Specifically, M31 remains significant at the 5% level when compared with M23 and M36, while it is highly significant at the 1% level when compared with all other models. These results indicate that M31 significantly outperforms all competing models in terms of predictive accuracy and exhibits strong robustness.

Furthermore, the 95% confidence intervals based on HAC standard errors, following recent developments in robust statistical inference proposed by Salerno et al.⁶⁰, do not cross zero, as shown in Table 6.For example, the confidence interval for M32 is [-15.6563, -6.3327], indicating that the predictive advantage of M31 remains robust after accounting for heteroskedasticity and temporal dependence, and that the differences between models are statistically significant.

(4) Robustness test under alternative dependent variables

To verify the generalization ability of the models across different dependent variables, this study replaces the original EIT index with two sub-indicators from the cleanliness and low-carbon dimensions: (i) the share of renewable electricity generation(in Table 7) ; (ii) the share of non-fossil energy in energy production(in Table 8).This robustness check is commonly used to test model stability under alternative outcome definitions⁶¹.Using the same data partitioning and cross-validation procedures, the models are reconstructed and evaluated on the independent test set using RMSE and R\(^2\). The results show that under both alternative indicators, M31 (SVR + KernelRidge / RidgeCV) consistently achieves the best performance, with the lowest RMSE and the highest R\(^2\). This indicates that its predictive performance is not sensitive to the choice of dependent variable and demonstrates strong generalization ability and robustness.In contrast, the rankings of the other models vary by indicator. For example, M35 performs relatively well in predicting the share of renewable electricity generation but shows a slight decline when predicting the share of non-fossil energy production. Overall, none of the competing models outperforms M31, further confirming its stability and reliability under different dependent variables.

Based on the test set performance, repeated cross-validation results, EPA statistical tests, and alternative dependent variable analysis, the benchmark model M31 consistently demonstrates superior performance in terms of predictive accuracy, stability, and statistical significance.Across different data partitioning strategies, error evaluation frameworks, and dependent variable settings, M31 maintains lower prediction errors and higher explanatory power, with relatively small performance fluctuations. These results indicate that M31 exhibits strong robustness and generalization ability.

Analysis of the carbon emission prediction model

Considering the significant influence of energy intensity on carbon emissions in the Yellow River Basin, this indicator is also included in the initial variable set^26,62. To mitigate multicollinearity and enhance model robustness, Lasso regression is first applied to the training set for feature selection. The optimal penalty parameter is automatically determined via ten-fold cross-validation.Lasso regression employs L1 regularization, which imposes a penalty on the absolute values of the coefficients, shrinking some coefficients to exactly zero and thereby achieving feature selection. In Fig. 6, eliminating features refer to variables whose coefficients are shrunk to zero by the Lasso regression, i.e., features removed from the model. The characteristic coefficient denotes the Lasso regression coefficient of each variable, and its absolute value reflects the relative contribution of that variable to the target variable.Ultimately, 14 significant variables are selected from the initial set of 16 candidate variables.Among them, The Coal Mining Fatality Rate (CMR) and Per Capita GDP (PCG) are eliminated, indicating their relatively weak influence on the effectiveness of the energy production revolution.

The CMR reflects the level of prevention and control in mining subsidence areas, indirectly affecting the efficiency of coal extraction and utilization. For example, in Gansu Province from 2010 to 2018, while the CMR decreased by 77%, carbon emissions increased by 25.8% year-on-year. From 2019 to 2023, as the fatality rate stabilizes, carbon emissions continued to grow at an average annual rate of 2.7%. This demonstrates a clear decoupling and a weakened statistical correlation, leading to its exclusion by the Lasso model. Meanwhile, PCG, which measures regional economic development levels, generally promotes carbon emissions²⁵. In the variable correlation analysis, PCG shows correlations greater than 0.5 with factors such as CCPE, UL, and WCG. Therefore, the information captured by the PCG indicator is already represented in other variables. Consequently, Lasso regression tends to select these other indicators, and PCG is not included in the final model.

Following Lasso-based variable selection, an extended STIRPAT-Ridge regression model is constructed based on the selected features to predict carbon emissions in the Yellow River Basin.The Ridge regression model determines the optimal regularization parameter via five-fold cross-validation, thereby enhancing the model’s generalization ability and stability.As shown in Table 9,the model demonstrates strong performance on the training set, with a MAPE of 2.51%. On the independent test set, the MAPE is 3.19%, indicating excellent fitting performance and good generalization capability. The consistency between training and test results further confirms the stability and reliability of the model, reflecting its high predictive accuracy.Based on the model outputs, carbon emissions in the Yellow River Basin can be quantitatively analyzed under different scenarios, providing a reliable basis for policy evaluation and decision-making.For comparison, the standalone STIRPAT-Ridge model without Lasso pre-selection shows inferior performance across RMSE, MAE, MAPE, and R\(^2\) on both the training and test sets. This result further confirms that, under high-dimensional settings, incorporating Lasso for variable selection effectively improves model fitting performance and predictive capability.

Scenario setting

This study defines the following four future development scenarios. Based on historical data, the actual socio-economic development, the targets established in relevant policy documents, and a comprehensive consideration of national and provincial development plans within the basin, the annual change rates for key parameters are set as shown in Table 10. Due to space limitations, the detailed process of parameter setting is not elaborated here.

1. (1)

   Baseline scenario (A1)

   The Baseline Scenario represents a continuation of the current socio-economic development trajectory. Under this scenario, all influencing factors, such as Coal Consumption for Power Generation and Raw Coal Production, develop according to their existing trends and patterns, guided by the targets set in relevant planning documents.

2. (2)

   Foundational innovation scenario (A2)

   Building upon the Baseline Scenario, the Yellow River Basin accelerates the reduction in the average utilization hours of thermal power generation equipment in plants of 6000 kW and above. It further improves coal combustion efficiency and reduces CCPE. Efforts to optimize patent quality and structure are intensified, enhancing the patenting efficiency of innovation outcomes. These measures aim to more effectively promote the increase in the Renewable Energy Power Consumption Ratio.

3. (3)

   System synergy scenario (A3)

   On the basis of the Foundational Innovation Scenario, this scenario further strengthens the supporting role of human resources in the energy transition. By enhancing inter-provincial infrastructure connectivity within the basin, it accelerates the increase in urbanization levels across the nine provinces. Concurrently, it increases subsidies for talent attraction and strengthens talent team building, for example, by implementing high-level talent introduction programs that provide housing subsidies and research start-up funds for professionals in high-demand fields like energy technology and carbon management.

4. (4)

   Structural synergy scenario (A4)

   Building upon the Foundational Innovation Scenario, this scenario strengthens the process of substituting natural gas for coal, aiming to more rapidly establish a transitional, stable energy bridging pathway¹⁰. On one hand, it further enhances the exploration, development, and technological demonstration of unconventional natural gas resources to boost domestic supply capacity. On the other hand, it vigorously promotes coal-to-gas conversion projects for industrial boilers and gas-fired cogeneration, achieving coordinated electricity and heat supply. Simultaneously, it utilizes gas-fired power for peak shaving to support the large-scale grid integration of renewable energy. This approach promotes an orderly and stable transition of the energy structure through three phases: coal-dominant, gas-electricity synergy, and finally renewable energy-dominant. The process of reducing energy intensity is also accelerated under this scenario.

Forecast results and pathway analysis

Forecast results and analysis for basin-wide province

As shown in Fig. 7, the optimal pathway for the energy production revolution in the Yellow River Basin is the Structural Synergy Scenario (A4). Under this scenario, the EIT index reaches the Excellent level, and carbon peaking is projected to occur during 2028–2030. Figures 7, 8, 9, 10, 11, 12 and 13 adopt a dual-axis design, where the left vertical axis corresponds to the bar charts representing CO\(_2\) emissions, and the right vertical axis corresponds to the line charts representing the EIT index.The basin’s EIT index shows a steady upward trend across all four scenarios. However, only under the A4 pathway does it reach the Excellent level by 2045, with values descending from A4 to A1 as follows: 90.5 (A4), 89.6 (A2), 87.3 (A3), and 86.2 (A1). This indicates that using natural gas as a transitional energy source effectively facilitates the achievement of energy production revolution goals. This finding is consistent with the results of Gerlagh and Smulders⁶³. However, Jinhui Zheng et al.⁶⁴ also point out that the role of natural gas in the energy transition exhibits clear stage-specific characteristics. In regions where resource endowments are dominated by coal, substituting natural gas helps reduce carbon emissions and optimize the energy structure, but its long-term mitigation effects still depend on the development of renewable energy and policy constraints.The projected carbon peaking years across the four scenarios converge, all meeting the target timeline: 2029–2031 for A1 and A2, and 2028–2030 for A3 and A4. Only the Structural Synergy Scenario (A4) results in an Excellent level of EIT index by 2045, therefore,A4 is identified as the optimal pathway. Under this pathway, the basin’s EIT index rises steadily from 66.7 (Fair level) in 2022 to 90.5 (Excellent level) in 2045, and carbon peaking is projected to occur during 2028–2030. This aligns closely with the peak time predicted for the Yellow River Basin by Zhang et al.²⁶ and is consistent with the judgment of Wei et al.⁶⁵ regarding China’s overall carbon peaking timeline.

The fundamental strength of Pathway A4, lies in its ability to effectively control coal production while promoting renewable energy to meet new energy demands, thereby completing the transition from coal dominance to new energy leadership.As a coal-rich region, the Yellow River Basin’s power supply still relies heavily on coal-fired generation. Although the proportion of renewable energy generation continues to increase, achieving a leapfrog replacement in the short term is challenging³⁰. Under pathway A1, due to insufficient maturity in energy storage and grid regulation technologies and limited market competitiveness, the growth in the installed capacity and generation share of renewables is slow. Part of the new demand still depends on coal, suppressing the increase in the non-fossil energy share and the decrease in the carbon emission intensity of power production. This results in insufficient driving force for the basin’s transition, making it difficult for the EIT index to reach an Excellent level. Although A2 pathway improves energy processing and conversion efficiency and reduces the carbon emission intensity of power production by enhancing coal combustion efficiency and accelerating technological progress, the integration of renewable energy systems faces constraints such as grid stability and transmission capacity⁶⁶. Coupled with the long-term investment still required for supporting infrastructure, part of the new demand continues to be met by efficient coal power, keeping the EIT index at a Good level. The pathway A3 enhances electricity supply reliability and reduces energy intensity by strengthening inter-provincial infrastructure connectivity and urbanization. However, this may crowd out investment in renewables, slowing the process of increasing the non-fossil energy share. In contrast, the A4 pathway constructs a “coal-natural gas-renewables” cascade conversion path. On one hand, it leverages the basin’s unconventional natural gas resource advantages to reduce the carbon emission intensity of power production, stabilize the energy self-sufficiency rate, and promote the cleansing of the energy structure. On the other hand, through gas-fired power peak shaving, it effectively alleviates the spatiotemporal mismatch between renewable energy bases in the northwest and load centers in central and eastern China.

Forecast results and analysis for hydro-rich provinces

As shown in Fig. 8, the optimal pathway for the energy production revolution in hydro-rich provinces is the Foundational Innovation Scenario (A2). Under this scenario, the EIT index reaches the Excellent level, and carbon peaking is projected to occur during 2028–2030. The EIT index for hydro-rich provinces shows a trend of initial decline followed by an increase across all four scenarios, ultimately rising to the Excellent level. By 2045, the values are 98.4 (A4, A2), 97.9 (A3), and 96.4 (A1), in descending order. The projected carbon peaking years vary significantly across scenarios, with 2029–2031 under scenario A1, 2028–2030 under scenarios A2 and A4, and 2036-2038 under scenario A3. Since the Foundational Innovation Scenario (A2) yields the best energy production revolution effectiveness by 2045, achieves the earliest carbon peak, and has the lowest peak emissions, it is identified as the optimal pathway. Under this scenario, the EIT index for hydro-rich provinces rises steadily from 85.7 (Good level) in 2022 to 98.4 (Excellent level) in 2045,and carbon peaking is projected to occur during 2028–2030. The above results are consistent with the findings of Li et al.³⁴, which indicate that in hydropower-rich regions, improving energy efficiency and enhancing renewable energy consumption can effectively reduce carbon emission peaks and advance the timing of carbon peaking. This aligns with the lower peak emissions and earlier peaking observed under the A2 scenario in this study.

Across all four scenarios, the EIT index for hydro-rich provinces shows a declining trend between 2022 and 2025. This is likely primarily due to Qinghai, Sichuan, and other areas focusing on building clean energy industrial hubs, which drive a rapid increase in total energy consumption, thereby leading to a temporary decline in the EIT index. After 2025, the EIT index returns to an upward trend. A similar pattern of stage-specific fluctuations is also reported by Sun and Dong⁶⁷, who point out that during periods of rapid energy structure adjustment, energy transition indicators in some regions may exhibit short-term volatility while maintaining a long-term upward trend. This is consistent with the “decline-then-rise” pattern of the EIT index observed in this study.

The pathway A2 demonstrates significant comparative advantages, fundamentally because it employs advanced grid technologies to enhance renewable energy consumption capacity⁶⁸ and effectively utilizes the regulating role of hydropower. This systematically alleviates intermittency issues, thereby accelerating clean energy substitution and reducing carbon emissions. Although the pathway A1 achieves carbon peaking on time, it has the lowest EIT index among the scenarios. This indicates that strengthening policy efforts can effectively improve the energy production revolution effectiveness in hydro-rich provinces. The pathway A3 achieves an Excellent EIT index but has a significantly delayed carbon peaking year. While more advanced grid infrastructure and cross-regional power⁶⁸, intensifying talent recruitment and urbanization construction incurs high human resource costs, coordination costs, and infrastructure expenses. This raises the total cost, increases the unit cost of emission reduction, and delays the peaking process. Compared to the A2 pathway, the A4 pathway has higher peak carbon emissions, mainly because large-scale natural gas infrastructure construction generates direct emissions and may induce additional energy demand, thereby slowing the overall emission reduction process.

Forecast results and analysis for wind-rich province

As shown in Fig. 9, the optimal pathway for the energy production revolution in wind-rich provinces is the Structural Synergy Scenario (A4). Under this scenario, the EIT index reaches the Good level, and carbon peaking is projected to occur during 2027-2029. The EIT index for wind-rich provinces rises rapidly across all four scenarios. Only the Foundational Innovation Scenario (A2) and the Structural Synergy Scenario (A4) reach the Good level, while the others remain at the Average level. By 2045, the values are 83.0 (A4), 80.6 (A2), 76.8 (A3), and 76.4 (A1), in descending order. The results further indicate that the effectiveness of the energy production revolution in wind-rich provinces relies more on system-level coordination, with coordinated pathways outperforming single-technology pathways. This finding is consistent with Yu et al.⁶⁹, who emphasize that system flexibility plays a decisive role in transition efficiency in high wind penetration systems.The projected carbon peaking years are close across scenarios, with 2029-2031 under scenarios A1 and A2, 2028-2030 under scenario A3, and 2027–2029 under scenario A4. The Structural Synergy Scenario (A4) yields the best energy production revolution effectiveness by 2045, has the earliest peaking time, and the lowest peak emissions. Therefore, A4 is identified as the optimal pathway. Under this scenario, the EIT index for wind-rich provinces rises steadily from 56.2 (Fair level) in 2022 to 83.0 (Good level) in 2045, and carbon peaking is projected to occur during 2027-2029. These findings are also in line with Hu et al.⁷⁰, who suggest that northern energy bases can achieve carbon peaking around 2028. However, Han et al.⁷¹ argue that the dependence on long-distance power transmission and the constraints on renewable energy integration may limit the emission reduction benefits of wind power, thereby potentially delaying carbon peaking.

The pathway A2 demonstrates significant comparative advantages, fundamentally because through system integration and orderly substitution, it addresses the intermittency and volatility issues of wind power, maximizing the overall efficiency of the energy system. Under the pathway A1, the EIT index remains at the Average level, indicating that current policy efforts are insufficient to fully unlock the energy revolution potential of wind-rich provinces. The pathway A2 can achieve carbon peaking on time and improve effectiveness through efficiency gains and technological progress. However, it cannot solve the intermittency and volatility of wind power. Continued reliance on coal power leads to slow growth in the renewable energy share. In contrast, the A4 pathway addresses the stability challenge of integrating a high proportion of renewable energy into the grid in wind-rich provinces by leveraging the flexible peak-shaving capability of gas power. This not only ensures energy supply security but also promotes the consumption of wind power, thereby increasing the share of non-fossil energy production. As a result, it boosts the EIT index and facilitates the achievement of emission reduction goals. Moreover, with its pipeline compatibility and operational flexibility, natural gas helps alleviate the institutional challenges associated with wind power integration into the power system, thereby supporting a more manageable and gradual energy transition.

Forecast results and analysis for PV-rich province

As shown in Fig. 10, the optimal pathway for the energy production revolution in PV-rich provinces is the Foundational Innovation Scenario (A2). Under this scenario, the EIT index reaches the Good level, and carbon peaking is projected to occur during 2027–2029. The EIT index for PV-rich provinces shows a continuous upward trend across all four scenarios. By 2045, the values are 90.1 (A4), 89.8 (A2), 87.1 (A3), and 85.7 (A1), in descending order. The projected carbon peaking years differ significantly, with 2034-2036 under scenario A1, 2027–2029 under scenario A2, and 2036-2038 under scenarios A3 and A4. Although the Structural Synergy Scenario (A4) yields the best energy production revolution effectiveness, it fails to achieve peaking on time. Therefore, the Foundational Innovation Scenario (A2) is identified as the optimal pathway. Under this scenario, the EIT index for PV-rich provinces rises steadily from 58.1 (Fair level) in 2022 to 89.8 (Good level) in 2045, and carbon peaking is projected to occur during 2027–2029.This result is consistent with the findings of Liu and Huo⁷² and Wang et al.¹⁶, indicating that in regions with abundant renewable energy resources, technological progress and efficiency improvements can significantly advance the timing of carbon peaking.

The pathway A2 demonstrates clear comparative advantages due to the maximization of marginal benefits from technological innovation and the optimization of cost-effectiveness in the transition pathway. Uncertainty in technology cost is a key factor affecting the total cost of a decarbonized power system, while the continued cost decline of renewables like solar is a general trend^73,74. The pathway A2, through efficiency improvements and breakthroughs in PV technology and energy storage, increases electricity supply reliability and energy processing and conversion efficiency, reduces energy intensity, and steadily increases the share of renewable energy generation.Under the pathway A1, PV-rich provinces not only have the lowest EIT index but also the most delayed carbon peaking time. Although the pathway A4 can enhance electricity supply stability and energy self-sufficiency in the short term, achieving relatively good effectiveness, the expansion of natural gas infrastructure may also create a ”carbon lock-in” effect. This can hinder the future development of renewables and lead to stranded assets⁷⁵. Consequently, the decline in the carbon emission intensity of power production slows, and the carbon peaking time is delayed. The pathway A3 cannot accelerate the improvement of the EIT index or the carbon peaking process, revealing that in regions rich in PV resources, the energy production revolution is unequivocally technology-driven.

Provincial prediction results

The analysis above reveals that the optimal transition pathways for the entire basin and for the hydro-rich, wind-rich, and PV-rich provinces differ due to varying regional resource endowments. Therefore, this study further analyzes the evolutionary trends of each of the nine provinces within the basin under the four development scenarios from 2025 to 2045. The specific results are shown in Fig. 11, where subfigures (a), (b), (c), (d), (e), (f), (g), (h), and (i) represent the EIT index and carbon emission predictions for Ningxia, Henan, Shaanxi, Inner Mongolia, Shanxi, Shandong, Qinghai, Sichuan, and Gansu, respectively.

The effectiveness of the energy production revolution and the carbon emission trajectories among the nine provinces show significant variations, primarily manifested in differences in provincial transition outcomes and the asynchronicity of carbon peaking processes. Firstly, there are marked differences in provincial peaking progress. The four provinces of Henan, Inner Mongolia, Shanxi, and Qinghai are projected to achieve the carbon peak on schedule. Shaanxi, Sichuan, and Gansu are projected to peak later, while Ningxia and Shandong are not projected to achieve the peak within the study horizon. Against this backdrop, the performance of each province under its optimal energy production revolution pathway can be categorized into three tiers.Tier 1 demonstrates Excellent performance and includes Shaanxi, Inner Mongolia, Sichuan, Qinghai, and Gansu, with optimal pathways of A2, A4, A4, A3, and A3, respectively. Tier 2 achieves Average performance and includes Henan and Shanxi, both with an optimal pathway of A4. Tier 3 presents a special case: Ningxia achieves an Excellent EIT level, but its carbon emissions continue to rise; Shandong reaches an Average EIT level, with its carbon emissions also showing a rising trend.

As shown in Fig. (a), Ningxia does not achieve carbon peaking under any of the four original scenarios. As a major energy-exporting province, Ningxia’s EIT index improves significantly across all scenarios by 2045, jumping directly from the Fair to the Excellent level. However, carbon emissions show an increasing trend under all scenarios, consistent with the simulation results of Zhang and Wu⁷⁶. As shown in Fig. 12, considering Ningxia’s specific circumstances, this study introduces policy scenarios with stronger constraints—B2, B3, and B4—based on A2, A3, and A4. These scenarios further enhance the following key indicators: increasing the annual growth rate of the share of 300MW and above units in thermal power installed capacity from the range of -0.1%~1.3% to 0.6%~2.0% to accelerate the phase-out of outdated coal power units; simultaneously increasing the annual growth rate of R&D Expenditure in the electricity, heat production, and supply industries from 5%~8% to 10%~13%. The results indicate that the optimal pathway for Ningxia’s energy production revolution is B4, under which the EIT index reaches the Excellent level and carbon peaking is projected to occur during 2028–2030. This is likely because measures like gas-for-coal substitution under the A4 pathway cannot rapidly change Ningxia’s coal-power-dominated energy export structure, and their emission reduction effects are offset by the expansion of new coal power and coal chemical capacity, leading to continuous carbon growth. In contrast, the B4 pathway significantly increases the proportion of large-capacity thermal power units to reduce coal consumption for power supply and carbon emission intensity. At the same time, it substantially increases R&D investment to promote the application of deep decarbonization technologies like CCUS in the coal power sector, thereby effectively curbing the rebound effect and ultimately driving the achievement of carbon peaking.

As shown in Fig. (f), Shandong’s energy production revolution effectiveness improves to the Average level by 2045 under the A2 and A4 scenarios but remains at the Fair level under the other scenarios. Carbon emissions accelerate upward under all four scenarios, failing to achieve peaking. This is consistent with the prediction of Zhangand Wu⁷⁶. The reason lies in the strong ”carbon lock-in” effect caused by Shandong’s substantial heavy and chemical industrial base. Empirical research by Zhang et al.⁴² further indicates that Shandong is classified as a typical region with a lower level of new-type urbanization but increasing carbon emissions alongside urbanization, reflecting the rigid constraints its high-carbon industrial structure places on low-carbon transition. As shown in Fig. 13, considering Shandong’s specific circumstances, this study introduces policy scenarios with stronger constraints—C2, C3, and C4—based on A2, A3, and A4. These scenarios further enhance the following key indicators: increasing the annual change rate of water use per 10,000 yuan of GDP to the range of -7% to 0%; increasing the annual change rate of energy intensity to the range of -4.5% to -3.8%; and increasing the annual change rate of the number of patents granted to 7% to 13%. Simultaneously, the annual change rate of raw coal production is increased to the range of -9% to 0%, with C4 specifically setting it at -9.5% to 0%. The results indicate that the optimal pathway for Shandong’s energy production revolution is C3, under which the EIT index reaches the Average level and carbon peaking is projected to occur during 2028–2030. This aligns with the research findings of Feng and Yu⁷⁷ under a low-carbon scenario. The C3 pathway, on one hand, simultaneously strengthens the construction of offshore wind power and nuclear energy in response to demand growth, strictly controls coal production, and sets stringent water use targets. This increases the share of non-fossil energy production and the share of renewable electricity generation, driving a rapid decline in the carbon emission intensity of power production. On the other hand, it reinforces energy efficiency constraints and innovation incentives, forcing technological upgrades and structural optimization in heavy industries and continuously improving energy processing and conversion efficiency. These measures further enhance the effectiveness of the energy production revolution, effectively curb the rebound effect, and promote the early achievement of carbon peaking.

As shown in Fig. (b), the optimal pathway for Henan’s energy production revolution is A4, under which the EIT index reaches the Average level, and carbon peaking is projected to be achieved before 2025. The province’s effectiveness improves to the Average level by 2045 under A2 and A4 but remains at the Fair level under other scenarios. Carbon emissions show a declining trend under all four scenarios, consistent with the prediction of Zhang and Wu⁷⁶, primarily benefiting from industrial structure deepening and renewable energy development. As a major industrial and populous province with massive energy consumption, Henan’s high carbon emissions mainly stem from coal consumption. The A4 pathway promotes renewable energy development by substituting for coal, thereby significantly improving energy production revolution effectiveness and leading to a decline in carbon emissions.

As shown in Fig. (c), the optimal pathway for Shaanxi’s energy production revolution is A2, under which the EIT index reaches the Excellent level, but carbon peaking is projected to occur during 2035–2037. The province’s effectiveness improves significantly to the Excellent level by 2045 under all scenarios, with the earliest peaking time under A2. The main reason is that this pathway accelerates improvements in energy processing and conversion efficiency through technological innovation, further reduces energy intensity, and promotes the growth of the renewable electricity generation share, driving a faster increase in the EIT index. Its delayed carbon peaking time is primarily due to the carbon lock in effect potentially induced by its reliance on a transitional energy model.

As shown in Figs. (d) and (e), the optimal pathways for Inner Mongolia and Shanxi are both A4. Their EIT indices reach Excellent and Average levels, with carbon peaking projected to occur during 2029–2031 and 2027–2029, respectively. Among them, Inner Mongolia’s peaking time aligns with the prediction of Feng and Yu⁷⁷, and Shanxi’s result is consistent with the findings of Jia et al.²² and Ren and Xing⁷⁸. Specifically, Inner Mongolia accelerates the substitution of clean energy systems through the synergistic development of ”coal-to-gas” and wind energy; Shanxi breaks the coal-dependent pattern by leveraging the large-scale development of unconventional natural gas.

As shown in Fig. (g), the optimal pathway for Qinghai’s energy production revolution is A3, under which the EIT index reaches the Excellent level and carbon peaking is projected to occur during 2025–2027. The advantage of this pathway lies in its focus on building ultra-high-voltage transmission channels and digital management platforms, effectively resolving the structural contradiction between Qinghai’s abundant clean energy resources and insufficient local consumption capacity. The enhancement of power transmission capacity and dispatching efficiency significantly increases the renewable energy consumption ratio, which in turn stimulates enterprises’ enthusiasm for using green electricity, promotes related investment and technological iteration, and forms a positive ”investment-technology-cost-reinvestment” cycle.

As shown in Fig. (h), the optimal pathway for Sichuan’s energy production revolution is A4, under which the EIT index reaches the Excellent level, but carbon peaking is projected to occur during 2031–2033. Zhang and Wu⁷⁶ also point out that the energy consumption increment brought by Sichuan’s endogenous economic growth is very large, making it unable to achieve carbon peaking on time. The province’s effectiveness improves significantly to the Excellent level by 2045 under all scenarios, with the earliest peaking time under A4. The A4 pathway utilizes the flexible peak-shaving capability of natural gas to effectively support hydropower output and new energy consumption. This improves electricity supply reliability and the share of clean energy while achieving a reduction in the carbon emission intensity of power production, thus promoting the emission reduction process while meeting energy demand.

As shown in Fig. (i), the optimal pathway for Gansu’s energy production revolution is A3, under which the EIT index reaches the Excellent level, but carbon peaking is projected to occur during 2031–2033. Except under A1, the province’s effectiveness improves significantly to the Excellent level by 2045 under other scenarios, with the earliest peaking time under A3. Compared to Qinghai, which is also a hydro-rich province, Gansu’s transition performance lags somewhat. The main reason lies in differences in industrial structure: Gansu has a higher proportion of heavy industry and greater industrial inertia. In 2023, its gross output value of mining and manufacturing was 2.89 times that of Qinghai, indicating that energy-intensive industries still have a significant scale.

A further finding is that there is a significant scale difference between the optimal pathways at the provincial and regional (province-type) levels. The optimal pathways of wind-rich provinces are generally consistent at both the provincial and regional levels, indicating that their resource endowments and industrial structures jointly determine a relatively consistent transition trajectory. Among them, Inner Mongolia and Shanxi can achieve improvements in energy production revolution effectiveness and carbon peaking targets under the A4 pathway, while Shandong requires further intensified efforts. In contrast, the optimal pathways for PV-rich and hydro-rich provinces show significant divergence between the two scales. At the provincial level, Ningxia and Henan have B4 and A4 as their optimal pathways, respectively, while Shaanxi leans toward A2. This may be partly attributed to differences in internal resource structures and development positioning among these three provinces: Shaanxi is an energy production base, rich in coal and gas; Ningxia is also a national energy base, rich in coal resources; while Henan is a major energy-consuming province, relatively scarce in coal resources and highly dependent on external supply. When the scale shifts to the regional level, PV-rich provinces generally tend toward A2. This may be because Shaanxi, as a major contributor to regional carbon emissions, has a strong influence on the overall regional carbon peaking process through its emission reduction performance. The scale difference is most prominent for hydro-rich provinces. The optimal pathways for Qinghai, Sichuan, and Gansu are A3, A4, and A3, respectively, however, the overall optimal pathway unifies to A2.This result indicates that regional optimization leads to different pathway selections compared with provincial-level results under a unified modeling framework. These scenario-based comparative results further indicate that there are significant differences in optimal pathways across provinces in regional energy transition planning. A single provincial optimal pathway cannot be directly generalized to other regions; therefore, policy formulation should fully consider differences in resource endowments and development stages across provinces and implement differentiated “one-province-one-policy” energy transition strategies.

Pathway planning

By comparing the energy production revolution effectiveness and carbon emission prediction results under different scenarios, the pathway A4 is identified as the optimal pathway for maximizing the EIT index in the Yellow River Basin while ensuring the carbon peaking target. Based on this, adhering to the principles of ”categorized policies, integration of near- and long-term goals, system coordination, and phased peaking”⁷⁹, and balancing national energy security strategy with the demands for high-quality development in the basin, this study proposes a transition pathway with the core philosophy of ”structural optimization as the core, technological innovation as the driver, and cross-regional coordination as the safeguard”.This pathway systematically advances the green and low-carbon revolution of the Yellow River Basin’s energy system.

(1) During the current Five-Year Plan period, the EIT index in the Yellow River Basin grows steadily at an average annual rate of 1.86%, while carbon emissions gradually decline after entering a plateau with an average annual growth rate of 1.31%. To address transition challenges and achieve the peaking target on schedule, the main thrust of the energy production revolution during this period is ”increasing renewable energy consumption and accelerating energy structure adjustment.” Specific efforts focus on the following four aspects: First, promote the efficient and clean utilization of coal. Based on existing plans, vigorously promote advanced power generation technologies such as ultra-supercritical units, complete energy-saving retrofits of existing units, and accelerate the phase-out of outdated coal power units to reduce coal consumption for power generation. Actively promote centralized coal utilization and clean technologies, develop cogeneration, and reduce energy intensity. Strictly control new production capacity, further optimize mine layout, and increase the proportion of coal washing and utilization efficiency. Second, strengthen the security of clean energy supply. Vigorously develop clean energy sources such as hydropower, wind power, and solar photovoltaics according to the resource endowments of the three province types. Increase the exploration and development of coalbed methane and shale gas in basins like Ordos and Qinshui to promote the large-scale development of unconventional natural gas. Focus on laying out integrated ”wind-solar-thermal-storage” transmission projects centered on large energy bases such as Ordos, Yulin, and Ningdong, and accelerate the construction of large-scale new energy bases in desert, Gobi, and barren land areas. Simultaneously, accelerate the planning and layout of large-scale energy storage projects and establish a basin-level green power trading market to promote the consumption of new energy. Third, increase investment in key technological innovation.Significantly boost R&D expenditure to tackle key technologies such as energy storage, smart grids, microgrids, and CCUS decarbonization.Fourth, deepen cross-regional coordination mechanisms.Based on existing inter-provincial horizontal ecological protection compensation agreements in the Yellow River Basin, explore linking them with the environmental benefits of cross-provincial green power trading. Make every effort to jointly build major infrastructure such as ultra-high-voltage transmission corridors and supporting peak-shaving power sources.

The main development targets for this period include: increasing the reduction rate of coal consumption for power generation from 0.6% to 0.7%; raising the average annual growth rate of the renewable energy power consumption ratio from 10% to 10.5%; decreasing the average annual growth rate of average utilization hours for power generation equipment in plants \(\ge\) 6000 kW (thermal) from 3% to 2%; increasing the average annual growth rate of the number of patents granted from 6% to 7%; boosting the average annual growth rate of natural gas production from 8% to 11%; lowering the average annual growth rate of raw coal production from 0.3% to 0.1%; and increasing the reduction rate of energy intensity from 3.5% to 3.8%. Achieving these targets ensures the basin can smoothly achieve the carbon peaking goal before 2030.

(2) During the next Five-Year Plan period, the average annual growth rate of the EIT index in the Yellow River Basin slows slightly to 1.77%, and total carbon emissions continue to decline at an average annual rate of 1.77% after peaking. Building on the structural foundation laid and carbon peaking achieved in the previous period, the main direction of the energy production revolution in this stage is ”promoting energy system integration and efficiency leap, consolidating and expanding emission reduction achievements.” Specific efforts focus on the following four aspects:First, establish the systematic dominant position of clean energy. Drive clean energy from being an incremental supplement to becoming the mainstay energy source. Under the premise of ensuring supply security, enable new energy sources like wind and solar to not only meet all incremental electricity demand but also begin orderly replacement of existing coal power. Focus on developing integrated models such as ”PV + storage” and ”wind power + hydrogen production” to enhance their reliability and economic viability.Second, build a smart and coordinated new energy system.Focus on promoting the systematic integration and coordinated optimization of technologies such as energy storage, smart grids, distributed energy, and virtual power plants. Create a number of smart energy demonstration zones with highly coordinated ”source-grid-load-storage” systems, comprehensively improving system operation efficiency and the capacity for efficient renewable energy consumption and cross-regional coordination.Third, accelerate the large-scale application of key decarbonization technologies.Shift the R&D focus from technological breakthroughs to engineering demonstration and industrial promotion. Make every effort to promote the deployment of million-ton-scale CCUS industrial clusters in key sectors, drive significant cost reductions in advanced energy storage technologies, and establish a basin-wide energy big data platform to deeply empower system emission reduction through digitalization.Fourth, deepen cross-regional coordination towards industrial integration and joint standard development.On the basis of existing facilities and mechanisms, advance coordination towards the high end of the value chain. Encourage the joint development of cross-provincial ”zero-carbon industrial parks” based on green power advantages, forming complementary industrial chains. Jointly formulate regionally unified standard systems for green power consumption and carbon emission accounting, and cultivate an integrated regional energy service market.

The main development targets for this period include: increasing the reduction rate of coal consumption for power generation to 0.9%; controlling the average annual growth rate of the renewable energy power consumption ratio above 8.5%; maintaining stability in the average utilization hours for power generation equipment in plants \(\ge\) 6000 kW (thermal); increasing the average annual growth rate of the number of patents granted to 8%; maintaining the average annual growth rate of natural gas production above 10%; controlling the average annual reduction rate of raw coal production above 1.3%; and increasing the reduction rate of energy intensity to 4.0%. Achieving these targets will systematically advance the energy production revolution, ensuring carbon emissions enter a stable downward trajectory after peaking.