Retrospective Forecasting Using Provincial Grid Data and a Quota-Based Hybrid Physical-AI Model

Summary

This study proposes a hybrid cost-forecasting model for power grids that combines physical cost quotas with dynamic macroeconomic/technological adjustments and XGBoost-based residual compensation. Achieving 2.34% MAPE, it balances accuracy and interpretability, addressing regulatory transparency needs in tariff setting.

Abstract

The global energy transition and ongoing electricity market reforms require power grid enterprises to balance reliable power supply with increasingly stringent transmission and distribution tariff regulations. Traditional budgeting methods based on historical extrapolation often fail to reflect the physical basis of asset operations, whereas data-driven machine learning models achieve high predictive accuracy but lack the transparency required for regulatory cost verification. To address the trade-off between forecasting accuracy and interpretability, this study proposes a hybrid cost-forecasting model based on cost quotas. The framework uses standardized operating quotas as the physical budgeting baseline and incorporates a dynamic mechanism for quota evolution driven by macroeconomic conditions and technological progress. Extreme Gradient Boosting (XGBoost) is employed to capture nonlinear residuals beyond the quota-based estimates, while SHAP (Shapley Additive exPlanations) is used to interpret the contribution of key cost drivers. The model was evaluated using 16 years of anonymized operational data from a provincial power grid in China. It achieved a mean absolute percentage error (MAPE) of 2.34%, reducing forecasting errors by 61.8%, 46.6%, and 34.1% compared with SARIMAX, standalone XGBoost, and Attention-LSTM models, respectively. The proposed framework integrates engineering cost-quota principles with explainable artificial intelligence, providing both accurate long-term cost forecasts and a transparent decision-support tool for regulatory permitted-cost verification.

Introduction

The global energy transition is reshaping power grid asset management and operational practices through the integration of distributed renewable generation, increasing extreme weather events, and widespread deployment of IoT technologies, all of which demand greater operational flexibility and maintenance resources to ensure system reliability^1,2. At the same time, electricity regulators have strengthened transmission and distribution tariff oversight through frameworks such as the UK's RIIO and the US Federal Energy Regulatory Commission's rate-of-return reviews, emphasizing transparency in cost justification³. China has similarly adopted a "permitted cost plus reasonable return" regulatory mechanism, requiring utilities to demonstrate clear links between engineering activities and financial expenditures⁴. However, utilities often lack transparent quantitative tools that connect physical asset operations with cost forecasts, limiting the effectiveness of regulatory cost verification⁵.

Existing forecasting approaches have important limitations in this environment⁶. Traditional methods, including incremental budgeting and ARIMA-based models, assume relatively stable historical patterns and often perform poorly under structural changes caused by economic fluctuations or extreme weather⁷. Conversely, modern machine learning models such as LSTM and Transformer architectures provide high predictive accuracy for short-term forecasting but lack the engineering interpretability required for regulatory decision-making^8,9. Although recent hybrid forecasting approaches combine statistical and machine learning techniques, they generally overlook the standardized engineering cost quotas that underpin power grid accounting systems¹⁰. To address these limitations, this study proposes a quota-driven hybrid forecasting framework that combines dynamic quota evolution with machine-learning-based residual correction, preserving engineering interpretability while improving forecasting accuracy.

One major research direction stems from incentive regulation in natural-monopoly industries, including the UK's RPI-X mechanism and China's permitted-cost verification framework¹¹. These studies primarily evaluate operational efficiency using Data Envelopment Analysis (DEA) and Stochastic Frontier Analysis (SFA)¹². Previous research has examined long-term relationships between capital expenditure (CAPEX), operating expenditure (OPEX), and cost efficiency to support tariff regulation^13,14. While these approaches provide valuable macroeconomic insights, their outputs are typically expressed as relative efficiency scores rather than monetary forecasts suitable for annual budgeting^15,16. Furthermore, frontier models generally assume relatively stable physical asset structures and operating conditions, limiting their ability to capture abrupt cost changes driven by asset aging, infrastructure replacement, or rapid demand growth^17,18.

A second research direction focuses on statistical and artificial intelligence-based forecasting models. Early studies employed multiple linear regression and autoregressive integrated moving average models with exogenous variables (ARIMAX) for cost prediction¹⁹. More recently, support vector regression, random forests, XGBoost, LSTM, and Transformer models have substantially improved forecasting performance by exploiting nonlinear relationships and high-dimensional feature spaces^20,21,22. These approaches frequently incorporate macroeconomic indicators such as electricity demand, producer price indices (PPI), and climate variables²³. However, power grid operating costs arise from engineering activities, asset deterioration, maintenance scheduling, and management decisions rather than macroeconomic variables alone²⁴. Consequently, purely data-driven models often function as "black boxes," making it difficult to explain forecasted costs during regulatory reviews or permitted-cost verification^25,26.

Engineering cost quota management provides a potential solution to this challenge²⁷. Activity-based costing (ABC) has long relied on standardized operational quotas to estimate labor, material, equipment, and maintenance requirements for routine grid operations²⁸. Major Chinese utilities have developed comprehensive quota databases covering inspection, maintenance, testing, repair, and equipment replacement activities. Nevertheless, these engineering standards are primarily used for project settlement and auditing rather than dynamic long-term forecasting^29,30. Moreover, quota standards are typically revised only every several years, making them insufficiently responsive to changing commodity prices, technological progress, and operational practices³¹. Applying quota systems across millions of distributed assets also poses substantial computational challenges for large-scale forecasting³².

Despite significant advances in statistical forecasting, machine learning, and engineering cost management, no existing framework effectively integrates standardized engineering cost quotas, dynamic macroeconomic adjustment, and explainable machine learning into a unified forecasting model for regulatory cost verification. It was hypothesized that combining dynamically evolving cost quotas with XGBoost-based residual learning can improve long-term forecasting accuracy while preserving the engineering transparency required for regulatory decision-making. To test this hypothesis, a hybrid quota-driven forecasting framework was developed that integrates physical cost modeling, macroeconomic and technological adjustment mechanisms, XGBoost residual compensation, and SHAP-based model interpretation. The proposed approach aims to provide both accurate long-term cost forecasts and transparent evidence to support permitted-cost verification under modern electricity market regulation.

Protocol

This study used anonymized operational and financial data collected from a provincial power grid in eastern China. All data were aggregated and de-identified before analysis, and no personally identifiable or sensitive individual-level information was included. Therefore, ethical approval was not required. Data access and analysis complied with applicable data protection regulations and institutional agreements governing power sector information.

Forecasting framework overview

To integrate physics-based and data-driven forecasting approaches, a hybrid forecasting framework was developed that combines engineering cost quotas with machine learning. Rather than simply combining multiple algorithms, the framework follows the principle that physical models establish the baseline forecast, whereas machine learning compensates for residual errors. This design ensures that the forecasting process is grounded in the physical mechanisms underlying power grid production and operation activities rather than relying solely on historical cost extrapolation.

The framework first establishes a hierarchical mapping between power grid assets, standardized operating activities, and financial cost accounts. Production and operation costs are treated as monetary representations of the resources consumed by physical assets—including substations, transmission lines, distribution feeders, metering devices, and digital inspection equipment—during routine activities such as inspection, maintenance, testing, repair, and replacement. Cost quotas serve as the link between measurable engineering workloads and corresponding financial expenditures.

As illustrated in Figure 1, cost quotas function as standardized accounting units embedded throughout the asset operation and maintenance process rather than as abstract financial allocation rules. Bottom-level asset workloads are converted into standardized operating quotas and subsequently mapped to cost categories, including labor, materials, construction equipment, outsourced services, and emergency supplies. This hierarchical mapping preserves engineering interpretability and regulatory traceability throughout the forecasting process and provides the physical foundation for constructing the static baseline quota model.

Methodological workflow overview

The proposed forecasting framework consists of three sequential stages: (1) construction of a physical baseline using asset-level workload quotas, (2) dynamic evolution of cost quotas through macroeconomic and technological adjustments, and (3) machine-learning-based residual compensation to capture systematic nonlinear effects. As illustrated in Figure 1, the framework establishes a hierarchical mapping from bottom-level assets and standardized operational activities to production and operation cost forecasts. The implementation details of each stage are described in the following subsections.

Physical baseline cost model based on asset-level workloads

The production and operation cost of a power grid, C_total, comprises expenditures associated with multiple business activities, including substation operation, transmission line maintenance, distribution network management, customer service, and supporting systems. In this study, the baseline operating cost is assumed to be determined by the workload generated by each standardized operational activity and its corresponding cost quota.

The static baseline cost is calculated as:

   (1)

where V_i,k,t denotes the workload associated with the i-th asset or operational task within business category k during period t, and Q_i,k represents the corresponding standardized unit cost defined by the engineering cost quota system. Business category includes major operational functions such as substation maintenance, transmission line inspection, distribution network operation, and customer service. The double summation aggregates the costs of all standardized operational activities to estimate the theoretical baseline expenditure required to maintain normal grid operation.

Equation (1) establishes the physical relationship between engineering workloads and financial expenditures by mapping standardized operational activities directly to cost accounts. Unlike purely statistical forecasting models, this formulation provides an interpretable engineering baseline that serves as the foundation for subsequent dynamic quota adjustment and machine-learning-based residual correction. The equation was developed from the operational practices and cost quota system used by provincial power grid enterprises in China. Table 1 summarizes the notation used in Equation (1), including workload (V_i,k,t), standardized unit cost (Q_i,k), the number of operational tasks (N_k), and the business category index (k).

Dynamic evolution mechanism of quotas under external environmental disturbances

Standardized cost quotas (Q_i,k) provide a physically interpretable baseline but do not account for changes in macroeconomic conditions or technological progress. To improve their long-term applicability, a dynamic evolution mechanism was introduced to adjust the baseline quotas in response to both price inflation and technology-driven efficiency gains.

The first adjustment accounts for changes in procurement costs resulting from macroeconomic inflation. Power grid operation and maintenance depend heavily on bulk materials, including copper, aluminum, and silicon steel, whose prices are closely related to fluctuations in the Producer Price Index (PPI). Because  is an index with a base value of 100, it is first converted into a standardized inflation rate:

   (2)

On this basis, the price correction function  is defined as:

   (3)

where  is a lag-weight vector of length L satisfying

The lag structure represents the delayed transmission of macroeconomic inflation to procurement costs within the power grid supply chain. Converting the PPI index into a standardized inflation rate preserves the cumulative effect of price changes while avoiding scaling bias associated with direct use of index values. Equations (2) and (3) are adapted from established macroeconomic inflation adjustment models, with the lag structure calibrated for power-sector procurement cycles^33,34.

Technological progress was incorporated through a cost-reduction factor that reflected improvements in operational efficiency resulting from advances such as unmanned aerial vehicle inspection, intelligent robotics, and digital maintenance technologies. The technology adjustment factor is defined as:

  (4)

In this part, α and β are empirical elasticity coefficients estimated from historical panel data using nonlinear least squares. To ensure that the technological progress factor always represents a reasonable reduction in unit quota cost, the parameter estimation process constrains 0 < Γ(E_tech,t) ≤ 1. It should be noted that this factor primarily reflects the long-term efficiency improvement resulting from the substitution of mature technology. Equation (4) is original to this work, adapting the learning-curve concept from energy technology cost literature^35,36, to grid maintenance operations. Additional costs that may arise during the early stage of digital equipment deployment, such as parallel operation of old and new systems, platform integration, communication testing, and extra maintenance, are not forcibly deducted from the baseline quota; instead, they are identified by the subsequent machine learning residual compensation module:

  (5)

where C_base,t denotes the static baseline cost calculated from bottom-level asset workloads and standardized operating cost quotas; captures the transmission effect of macro price fluctuations on materials, equipment, and external service prices; and Γ(E_tech,t) reflects the efficiency-based reduction in unit operation and maintenance costs after technology maturity. Through the above dynamic evolution mechanism, the quota baseline no longer remains at a static accounting basis but can adaptively adjust with changes in economic environments and technological conditions. Equation (5) is original to this work and represents the novel integration of price and technology corrections into the quota baseline framework.

Systematic nonlinear residual capture under quota constraints

Despite complex evolutionary corrections, the quota model inevitably generates systematic deviations when facing unpredictable disaster-related weather disruptions and sudden policy directives, such as increased customer service complaint-handling costs during temporary tariff-reduction periods. This deviation forms the residual term on both sides of the equation:

R_t=C_actual,t-C_quota,t  (6)

Since conventional physical rules cannot explain this aspect, machine learning can address these limitations. To avoid the curse of dimensionality caused by high-dimensional features, this study uses the XGBoost algorithm based on decision-tree ensembles to model the nonlinear relationship R_t^37,38. A strong-disturbance feature matrix X_t, including meteorological features such as annual extreme freezing days D_ice and macro policy intensity is defined.

For a nonlinear compensator composed of regression trees, the generation logic of the predicted residual  can be expressed as³⁹:

   (7)

where F denotes the space of all possible classification and regression tree structures. To balance fitting accuracy and overfitting prevention, a regularized objective function containing a structural complexity penalty term is constructed and minimized in the m-th iteration:

   (8)

where  is a convex loss function measuring the difference between the true residual and the predicted residual. This paper adopts Huber Loss to enhance the model's robustness to abnormal peak expenditures. The regularization term  is used to constrain the complexity of the tree structure and is defined as:

   (9)

where T_m represents the number of leaf nodes in the m-th tree, w_m represents the corresponding leaf weight vector, and γ and λ denote the leaf-node number penalty coefficient and the weight regularization coefficient, respectively.

The final forecasting equation is:

   (10)

Further expanded as:

   (11)

The formula above mathematically represents the closed-loop structure of the proposed forecasting model. The final production and operation cost demand is not directly generated by the machine learning model; instead, it is obtained by superimposing the nonlinear residual compensation identified by the machine learning module onto the dynamic quota baseline. Among these components, price and technology factors primarily reflect the dynamic evolution of the quota baseline, while factors that are difficult to explicitly characterize through rules, such as climate shocks, policy disturbances, and surges in repair events, are captured by the machine-learning residual compensation module. Equations (10) and (11) are original to this work and synthesize the physical baseline with ML-based residual capture into a unified forecasting framework.

Figure 2 shows that the proposed model's forecasting results exhibit a clear hierarchical generation logic. On the one hand, the baseline quota provides a stable, transparent, and auditable physical foundation for cost demand; on the other hand, price adjustments, technology effects, and external shock residuals enable the model to adapt to dynamic changes under complex environments. Compared with black-box models that directly output predicted values, this decomposition structure can clearly reveal "why costs rise or fall," thereby enhancing the interpretability of model results in budget review and transmission and distribution tariff regulation.

Data sources and collection procedures

Theoretical models must be rigorously validated through empirical data to demonstrate their practical utility. Since the core financial data of the power sector involves sensitive information related to national infrastructure operations, this study extracts high-precision, anonymized monthly accounting data from a typical provincial power grid in eastern China, referred to as E-Grid for convenience, covering 16 consecutive calendar years from 2010 to 2025. This province has experienced a typical economic cycle of shifting from traditional heavy-industry-driven growth to high-end manufacturing, with the compound annual growth rate of grid asset scale reaching 7.4%. The complex evolution of its cost structure, therefore, has potential relevance to other rapidly developing grid systems. The data originate from three primary sources: (1) internal operation and maintenance logs recording asset-level workloads, inspection frequencies, and repair events; (2) financial accounting systems providing monthly cost statements across labor, materials, equipment, and outsourced services; and (3) external environmental databases including meteorological records from the China Meteorological Administration and macroeconomic indicators from the National Bureau of Statistics.

Quality control and missing data treatment

For over 130 initial indicators integrated from multi-source systems, a rigorous quality control procedure was implemented. Missing data points, accounting for less than 3% of the total observations, were handled using linear interpolation for continuous variables with temporal trends, and mode imputation for categorical indicators. Outliers were identified using the interquartile range (IQR) method, with values exceeding 3.0 times the IQR above the third quartile winsorized to the 99th percentile to preserve data integrity and mitigate extreme-value distortion.

Sample size considerations

The dataset comprises 192 monthly observations (January 2010–December 2025), with 156 observations (2010–2022) allocated for training and validation and 36 observations (2023–2025) reserved for out-of-sample testing. While this sample size is relatively modest for deep learning applications, it is appropriate for the XGBoost algorithm, which is specifically designed to perform well with small-to-medium tabular datasets through its regularization and tree-pruning mechanisms. To mitigate potential overfitting risks, (1) strict regularization penalties were employed (γ = 0.1, λ = 1.0), (2) early stopping with a patience of 50 rounds, and (3) conservative tree depth constraints (max depth = 5). These measures collectively ensure model stability and generalization despite the limited sample size.

Data segmentation and multi-source heterogeneous integration

For rigorous testing, the data from January 2010 through December 2022 were assigned to the training-validation interval, containing 156 observations, which is used to train the quota evolution factors and the quota compensation residual network. January 2023–December 2025 is reserved as the held-out out-of-sample test set, containing 36 observations. Why is this period selected as the final testing arena? The reason is that these three years coincided with the acceleration of new-type power system construction, compounded by large-scale extreme El Niño-related high-temperature events and the rapid and uneven growth of distributed renewable generation. The power grid faced unprecedented pressure in material supply chains and repair labor allocation.

Scientific selection and quantitative definition of cost-driving factors are the foundation for ensuring that the machine learning residual network can effectively capture systematic fluctuations. Based on the management logic of standard operation costs in power systems, this study breaks through the single dimension of traditional financial forecasting, which relies only on historical cash flows, and instead reconstructs feature engineering from four core boundaries: physical asset scale, operation and maintenance conditions, macroeconomic evolution, and external climate environment, using original operation logs and external system ledgers. In the actual modeling process, for more than 130 original indicators resulting from multi-source system integration, this study uses Pearson correlation tests to eliminate highly collinear redundant variables, with a threshold of |r| > 0.85. Based on the prior knowledge of senior power grid experts, 42 core input features are ultimately selected to form the feature matrix X_t. To clearly present the underlying data structure and distribution of the input tensor, Table 2 selects 12 representative core features from the four evaluation dimensions above and summarizes their descriptive statistics over the observation period.

To further illustrate the spatial-topological basis of the multi-source feature system, Figure 3 presents an anonymized schematic topology of the provincial power grid under study. The figure overlays voltage-level substations, transmission corridors, distributed renewable clusters, load centers, and representative environmental disturbance zones. The topology helps explain why production and operation costs are jointly affected by asset scale, network structure, emergency repair intensity, and external climate shocks. It also provides a spatial interpretation basis for the residual-driving variables used in the XGBoost compensation module.

Table 2 shows that explanatory variables across different business dimensions exhibit distinctly different statistical forms. Physical asset variables representing endogenous enterprise development dynamics, such as substation capacity and line length, have relatively stable standard deviations and skewness values concentrated between 0.1 and 0.8. Their overall data structure is approximately normally distributed, objectively reflecting the attribute of steady power grid development in the infrastructure construction cycle. In sharp contrast are the meteorological and external environmental disturbance variables at the bottom of the table. For example, the cumulative number of warning-level high-temperature days over the past 90 days and the line trip impact index exhibit extremely strong right-skewedness, with skewness values of 2.15 and 2.45, respectively. This typical heavy-tailed distribution confirms an objective pain point that cannot be ignored in actual power grid operation and maintenance: although extreme weather disasters occur relatively infrequently over the annual timeline, once triggered, they often cause exponential increases in repair labor input and spare-parts consumption. From another perspective, the high non-uniformity and extreme-value skewness in the distribution of these multi-source features reveal the theoretical limitations of traditional linear time-series models, such as ARIMAX, which are based on the assumptions of normality and homoscedasticity when tracing complex power grid costs. This not only further establishes the rationality of introducing a machine learning module beyond the physical accounting baseline but also provides solid statistical support for this paper's selection of the XGBoost tree model, which can efficiently handle sparse feature distributions and nonlinear mappings to approximate cost residuals.

Hyperparameter optimization and evaluation system setup

After determining the feature input space, the setting of model hyperparameters directly affects the fitting performance of the residual approximation network. Since the XGBoost compensation network involves multiple parameters, including tree depth (max depth), learning rate, and regularization penalty terms, and these parameters have nonlinear interactions, conventional grid search not only has high computational complexity but also tends to become trapped in local minima in high-dimensional spaces. Therefore, this study introduces the Tree-structured Parzen Estimator (TPE), a Bayesian optimization method, in the parameter tuning process. The TPE algorithm can dynamically guide subsequent sampling directions using loss-function feedback from previous evaluations. By constructing a posterior kernel density estimate (KDE) of the objective variable, it adaptively narrows the parameter search space, allowing the model to approximate the global optimal hyperparameter configuration without incurring high computational cost. The TPE optimization objective was to minimize validation RMSE over 100 iterations, with early stopping after 50 rounds without improvement.

After completing parameter optimization on the internal validation set, to objectively evaluate the final performance of each model on the out-of-sample test set and to meet regulators' quantitative assessment requirements for cost verification, this study uses the mean absolute percentage error (MAPE) to quantify the relative deviation in the predicted sequence. Meanwhile, to address the practical need to control extreme cost prediction failures in operations, root mean square error (RMSE) is also introduced to impose stronger penalties on larger errors. Finally, the coefficient of determination, R², quantifies the regression's overall explanatory power relative to the true target variance.

The mathematical definitions of the indicators are as follows:

   (12)

    (13)

   (14)

where  represents the actual production and operation cost in the period t,  represents the model-predicted cost,  represents the mean actual cost in the test sample, and N is the number of samples in the test set.

Results

Panoramic dimensionality-reduced comparison of forecasting accuracy

When all Tree-structured Parzen Estimator (TPE)-optimized models were evaluated on the 2023–2025 out-of-sample test set, which included post-pandemic recovery fluctuations and extreme high-temperature events, the performance of mainstream forecasting algorithms could be objectively compared. To provide a rigorous and comprehensive evaluation, four benchmark models representing different methodological approaches were included: Traditional Exponential Smoothing, representing conventional financial forecasting; SARIMAX, representing linear seasonal time-series modeling; standalone XGBoost regression, representing a purely data-driven approach without quota constraints; and Attention-LSTM, an attention-enhanced long short-term memory network widely used for long-sequence forecasting.

As shown in Table 3, the proposed Quota-ML model outperformed all benchmark models in forecasting production and operation costs during the 2023–2025 out-of-sample period. The model achieved an MAPE of 2.34%, reducing forecasting error by 61.8% compared with SARIMAX (6.12%) and by 34.1% relative to the Attention-LSTM model (3.55%). Its RMSE (15.69 million CNY) and MaxAE (23.05 million CNY) were less than half those of the next-best neural network, while an R² value of 0.957 indicated that the model explained more than 95% of the variance in observed costs. These results demonstrate that integrating engineering cost quotas, dynamic quota evolution, and XGBoost-based residual compensation substantially improves both forecasting accuracy and robustness compared with traditional time-series methods and purely data-driven models.

As illustrated in Figure 4, actual production and operation costs exhibited clear seasonal fluctuations and several pronounced peak periods during the test interval. In particular, costs increased sharply during the extreme high-temperature events of July and August 2024. Although the Attention-LSTM model captured the overall seasonal trend, its predictions were relatively smooth and underestimated abrupt cost increases associated with emergency repairs, heavy equipment loading, and increased consumption of emergency supplies. The standalone XGBoost model responded more effectively to local fluctuations but deviated from the observed trajectory during several months because it lacked engineering quota constraints. In contrast, the proposed Quota-ML model closely tracked the observed cost trajectory throughout the test period and accurately reproduced both the summer 2024 cost surge and the secondary summer peak in 2025.

The superior performance of the Quota-ML model demonstrates the benefit of combining engineering cost quotas with machine-learning-based residual correction. The dynamic quota baseline provides a physically interpretable foundation that reflects the evolution of power grid assets and operational workloads, preventing unconstrained learning from limited financial time-series data. The residual compensation network then focuses on nonlinear disturbances that are difficult to represent using engineering quota rules alone, including meteorological events, surges in maintenance work orders, and policy-related changes. Consequently, the proposed model achieved the lowest MAPE and RMSE among all evaluated methods (Table 3) and provided the most accurate tracking of seasonal cost peaks and extreme events (Figure 4), demonstrating its suitability for long-term production and operational cost forecasting.

Ablation validation of core architectural components

In a complex hybrid framework comprising nested submodules, the core academic review concern usually focuses on whether the model suffers from “over-engineering”. Conducting internal ablation experiments by removing core components is necessary to explore the true interrelationships and contribution levels of each module. This study sets two degradation paths for the architecture. First, Structure A, which removes the dynamic evolution mechanism: the quota baseline is forcibly confined to past static physical standards, stripping away the penetration of macro inflation accumulation and technological progress deflation factors in recent years, and only the historical static baseline  is connected to the residual network. Second, Structure B, which removes the nonlinear residual tracking module: the model completely degenerates into an actuarial approach, cutting off the AI-led random fluctuation capture loop and directly taking the pure dynamic quota calculation value , after macro-environment correction, as the final output. The explained variance loss ratio is calculated based on the relative decline in R² between the full model and the degraded model, defined as:

   (15)

Table 4’s ablation study reveals that both components of the proposed framework contribute significantly to forecasting accuracy, but the machine-learning residual compensation plays a more critical role. Removing the dynamic evolution mechanism for price and technology (Variant A) increases MAPE to 4.15% (a 1.81-percentage-point degradation) and reduces R² to 0.837, explaining 12.5% less variance than the full model. In contrast, eliminating ML residual compensation (Variant B) causes a much sharper decline: MAPE rises to 5.62% (+3.28 p.p.), R² drops to 0.686, and the explained variance loss ratio reaches 28.3%. These results demonstrate that while dynamic quota updating improves baseline accuracy, the XGBoost-based residual correction is indispensable for capturing nonlinear cost drivers, together forming a synergistic hybrid architecture.

To facilitate comparison of model performance across the ablation experiments, a normalized radar chart was constructed to visualize five evaluation metrics: MAPE, RMSE, R², MaxAE, and robustness. Error-based metrics (MAPE, RMSE, and MaxAE) were reverse-normalized so that lower errors corresponded to higher scores, whereas positive metrics (R² and robustness) were normalized so that higher values corresponded to higher scores. After normalization, all indicators represent comparable performance scores, with values closer to the outer boundary indicating superior overall performance.

As shown in Figure 5, the complete Quota-ML model achieved consistently high scores across all five performance dimensions, forming the largest and most balanced radar profile. This result indicates that the proposed framework provides an effective balance between relative error control, overall prediction accuracy, explanatory power, suppression of extreme errors, and robustness under varying operating conditions.

In contrast, Variant A, which excluded the dynamic quota evolution mechanism while retaining machine-learning-based residual compensation, exhibited a noticeable reduction in overall performance. This finding suggests that a static quota baseline alone cannot adequately account for structural changes caused by commodity price fluctuations and technological progress. An even greater performance decline was observed for Variant B, in which the residual compensation module was removed, and forecasting relied solely on the dynamic quota baseline. In this case, performance deteriorated substantially, particularly for error-related metrics and robustness.

As summarized in Table 4, the MAPE of Variant B increased from 2.34% for the complete model to 5.62%, representing an increase of 3.28 percentage points, while the R² value decreased by 28.3%. These results indicate that, although the dynamic quota model incorporates macroeconomic inflation and technology-driven efficiency adjustments, it cannot fully capture abrupt cost variations associated with extreme weather, emergency maintenance, policy changes, and abnormal operating conditions.

Overall, the ablation analysis demonstrates that both major components of the proposed framework are essential. The dynamic quota evolution mechanism adapts the engineering baseline to changes in macroeconomic conditions and technological progress, whereas the machine-learning residual compensation module captures nonlinear deviations that cannot be explicitly represented by engineering quota rules. Together, these complementary components form an integrated forecasting framework that combines a physically interpretable quota baseline with data-driven residual learning to achieve accurate and robust long-term cost forecasting.

Interpretability validation of the residual compensation model

Although the forecasting accuracy comparison and ablation experiments demonstrated the importance of the machine-learning residual compensation module, error metrics alone cannot determine whether the captured nonlinear relationships have meaningful physical or operational interpretations. Therefore, SHAP (Shapley Additive exPlanations) was applied to interpret the XGBoost residual model⁴⁰. SHAP has been widely adopted for explaining machine-learning models in power and energy forecasting and related complex systems^41,42. By quantifying the contribution of each input feature to the model predictions, SHAP enables assessment of whether the learned residual patterns are consistent with engineering knowledge.

As shown in Figure 6, external disturbance variables—including the cumulative number of warning-level high-temperature days during the previous 90 days, the severe convection and typhoon-related line outage index, freezing rain and snow duration, and the number of unplanned repair events—exhibited strong positive SHAP values. Samples with high values for these variables were concentrated in the positive SHAP region, indicating that extreme weather and emergency maintenance activities consistently increased actual costs beyond the dynamic quota baseline. These findings demonstrate that the residual compensation module captures meaningful environmental and operational disturbances rather than simply fitting random noise.

Several additional variables, including heavy-load operating duration, the PPI metal price index, the proportion of digital investment, distributed photovoltaic installed capacity, and transmission line length, also showed substantial contributions to the residual predictions. These results indicate that deviations from the quota baseline are jointly influenced by short-term weather events, asset expansion, commodity price transmission, power system transformation, and digitalization. Notably, the proportion of digital investment contributed positively to residual costs in some observations, suggesting that early-stage digital transformation may temporarily increase expenditures due to system integration, platform maintenance, and the parallel operation of legacy and newly deployed systems. Overall, the SHAP analysis validates the business interpretability of the residual compensation module and provides empirical support for understanding the impacts of extreme weather events and digital transformation on production and operation costs.

Statistical significance testing of forecast improvements

To determine whether the forecasting improvements achieved by the proposed framework were statistically significant, formal forecast comparison tests were performed. Following established practice in energy forecasting and econometrics, the Diebold–Mariano (DM) test was used to compare the predictive accuracy of the proposed model with each benchmark. The DM test is well-suited to time-series forecasting because it accounts for autocorrelation in forecast errors without requiring normally distributed residuals.

For each comparison, the null hypothesis assumed that the proposed and benchmark models had equal predictive accuracy, whereas the alternative hypothesis assumed that the proposed model produced lower forecast errors. One-sided DM tests were conducted using squared forecast error as the loss function. To account for heteroscedasticity and autocorrelation in the loss differential series, Newey–West standard errors with automatic lag selection were applied. In addition, the Wilcoxon signed-rank test was performed as a nonparametric alternative that does not rely on distributional assumptions. Together, these complementary tests provide a robust assessment of the statistical significance of the observed forecasting improvements.

As summarized in Table 5, the proposed forecasting framework achieved statistically significant improvements over all benchmark models. The positive DM statistics indicate consistently lower forecast errors than the competing approaches. The largest improvement was observed relative to the ARIMA model (DM = 3.842, p < 0.001), demonstrating the advantage of incorporating engineering cost quotas into the forecasting framework. Compared with ARIMAX, which includes exogenous variables, the proposed model also showed a significant improvement (DM = 3.215, p < 0.001), highlighting the additional value of combining a dynamic quota baseline with machine-learning-based residual compensation.

The proposed model also significantly outperformed the deep learning benchmarks. Forecast improvements were significant relative to the LSTM model (DM = 2.876, p = 0.002) and the Transformer model (DM = 2.543, p = 0.011). Although the improvement over the standalone XGBoost model was smaller (DM = 2.187, p = 0.029), it remained statistically significant. Because both models employ tree-based learning, this result demonstrates that integrating the dynamic quota baseline provides additional predictive value beyond a purely data-driven approach. The Wilcoxon signed-rank test produced significance levels consistent with those of the DM test, providing further evidence that the forecasting improvements of the proposed framework are statistically robust.

DATA AVAILABILITY:

The datasets used in this study have been uploaded to a public repository (DOI: https://doi.org/10.5281/zenodo.21645584).

Figure 1: Mapping relationship among power grid assets, standard operation quotas, and cost accounts. This figure shows that cost quotas are not merely abstract financial allocation rules but standardized accounting units embedded throughout the entire operation and maintenance process for power grid assets. Please click here to view a larger version of this figure.

Figure 3: Anonymized schematic topology of the studied provincial power grid and cost-driving disturbance layers. This figure presents an anonymized schematic topology of the studied provincial power grid. The figure overlays voltage-level substations, transmission corridors, distributed renewable clusters, load centers, and representative environmental disturbance zones. Please click here to view a larger version of this figure.

Figure 4: Prediction trajectories of core models and actual costs during the 2023–2025 out-of-sample test interval. This figure shows that actual production and operating costs during the test period exhibit clear seasonal fluctuations and sudden peaks. During the extreme high-temperature shock in July and August 2024, actual costs rose significantly. Please click here to view a larger version of this figure.

Figure 5: Normalized comparison of forecasting performance across ablation variants. Radar plots compare the normalized performance of the full Quota-ML model and two ablation variants using five evaluation metrics: MAPE score, RMSE score, R² score, MaxAE score, and robustness score. Error-based metrics (MAPE, RMSE, and MaxAE) were reverse-normalized so that higher scores indicate lower prediction errors, whereas R² and robustness were normalized such that higher scores indicate better model performance. The broader radar profile of the full Quota-ML model demonstrates superior overall forecasting performance compared with both ablation variants. Please click here to view a larger version of this figure.

Figure 6: SHAP summary plot showing feature contributions to residual cost prediction. SHAP (Shapley Additive exPlanations) summary plot illustrating the contributions of the most influential variables to the XGBoost residual compensation model. Each point represents one observation, with color indicating the feature value (blue = low; red = high) and horizontal position representing the SHAP value. Higher feature values for variables such as extreme heat days, typhoon-related outage index, ice storm duration, and unplanned repair events are generally associated with more positive SHAP values, indicating increased residual production and operation costs. The plot demonstrates that both environmental disturbances and operational factors contribute substantially to deviations from the dynamic quota baseline. Please click here to view a larger version of this figure.

Table 1: Description of core model variables and parameters. This table presents the key variables in the proposed Quota-ML cost forecasting model, distinguishing between actual and predicted costs; static and dynamic quota baselines; workload and quota parameters; macroeconomic adjustment factors (PPI and technological progress); and residual components. Please click here to download this Table.

Table 2: Classification and descriptive statistics of core input features of the forecasting model. 
This table shows that explanatory variables across different business dimensions exhibit distinct statistical forms. This table summarizes the descriptive statistics of representative physical asset, operational, macroeconomic, and environmental variables used as model inputs. Please click here to download this Table.

Models	MAPE %	RMSE (CNY, Millions)	R² Goodness of Fit	MaxAE (CNY, Millions)
Traditional Exponential Smoothing	8.75	54.22	0.651	125.04
SARIMAX time-series model	6.12	38.54	0.768	84.21
Pure XGBoost regression	4.38	29.16	0.852	51.06
Attention-LSTM neural network	3.55	24.02	0.894	40.53
Proposed Quota-ML Model	2.34	15.69	0.957	23.05

Table 3: Overall performance comparison of different models on the 2023–2025 out-of-sample test set. This table demonstrates that the proposed Quota-ML model significantly outperforms all benchmark models in forecasting power grid production and operation costs over the 2023–2025 out-of-sample period.

Experimental Variant	Removed Core Component	MAPE Degradation Performance	R²	Explained Variance Loss Ratio
Variant A	No price / technology dynamic evolution	4.15% (+1.81 p.p.)	0.837	12.50%
Variant B	No ML residual compensation	5.62% (+3.28 p.p.)	0.686	28.30%
Full Model	Full Proposed Quota-ML Framework	2.34%	0.957	Baseline

Table 4: Ablation experiment results of the quota-machine learning integrated framework. This table presents that both components of the proposed framework contribute significantly to forecasting accuracy, yet the machine learning residual compensation plays a more critical role.

Comparison	Diebold-Mariano Test	Wilcoxon Signed-Rank Test
Proposed Model vs. ARIMA	DM = 3.842*** (p < 0.001)	W = 486.0*** (p < 0.001)
Proposed Model vs. ARIMAX	DM = 3.215*** (p < 0.001)	W = 452.0*** (p < 0.001)
Proposed Model vs. LSTM	DM = 2.876** (p = 0.002)	W = 398.0** (p = 0.003)
Proposed Model vs. Transformer	DM = 2.543* (p = 0.011)	W = 364.0* (p = 0.014)
Proposed Model vs. XGBoost (pure ML)	DM = 2.187* (p = 0.029)	W = 328.0* (p = 0.031)

Table 5: Statistical significance test results for forecast comparisons. The Diebold-Mariano test statistic follows a standard normal distribution under the null hypothesis. Positive DM values indicate superior forecast accuracy of the proposed model. All tests are one-sided with the alternative hypothesis that the proposed model has a lower forecast error than the benchmark. The Wilcoxon signed-rank test statistic W is reported with corresponding p-values. Loss function = squared forecast error.

Discussion

This study demonstrates that integrating industrial cost quotas with machine learning provides an interpretable and accurate framework for long-term power grid cost forecasting. The results show that climate-related variables, particularly extreme high-temperature days and line trip impact indices, are the dominant drivers of cost residuals, highlighting the increasing influence of external disturbances on utility expenditures. These findings suggest that weather-related costs should no longer be treated as incidental operating expenses but instead incorporated into dedicated contingency reserves, with reserve activation linked to forecasted climate risks. The identification of line trip impacts as major cost drivers further underscores the value of predictive maintenance and condition monitoring strategies in reducing both operational disruptions and cost volatility.

The analysis also reveals that digital transformation can increase operational costs during the early stages of implementation due to transitional dual-system operation, rather than immediate efficiency gains. This finding suggests that utilities should evaluate digital investments using life-cycle cost assessment rather than short-term financial performance and should anticipate temporary cost overlaps when deploying smart monitoring, advanced distribution management, or digital substations. The proposed residual compensation framework could also be extended to estimate the point at which digital investments begin generating net cost savings, thereby supporting more effective technology planning and investment decisions.

Beyond utility operations, the proposed framework has important regulatory implications. By quantifying the effects of external drivers on cost residuals, the model provides an objective basis for establishing climate-adjusted contingency reserves and evaluating utility budget requests. Separating baseline quota-driven costs from disturbance-driven residuals also supports more effective performance-based regulation by distinguishing controllable operational efficiency from uncontrollable external shocks. Furthermore, the quota-driven forecasting framework offers greater transparency and auditability than purely statistical or black-box artificial intelligence models, allowing forecasted costs to be directly linked to physical operating conditions and engineering parameters during tariff reviews and regulatory hearings.

Several limitations should be acknowledged. The empirical analysis is based on a single provincial power grid in eastern China, which may limit generalizability to regions with different climatic conditions, regulatory environments, or network structures. Although the 192-month dataset is sufficient for the XGBoost model, it remains relatively small for capturing rare but high-impact events, and forecasting performance depends on the quality and consistency of the underlying cost quota databases. In addition, the dynamic quota mechanism relies on the aggregated producer price index and technology progress factors that may not fully capture regional or component-specific cost variations. Although SHAP analysis improves model interpretability, the reported feature contributions represent predictive associations rather than causal relationships and should therefore be interpreted with appropriate caution.

Future research should focus on integrating unstructured maintenance records through domain-specific large language models, incorporating spatiotemporal graph neural networks to capture disturbance propagation across interconnected grids, extending validation to multiple regions to improve generalizability, and introducing uncertainty quantification to support risk-aware regulatory decision-making. These developments are consistent with recent advances in medium- and long-term energy forecasting and renewable energy integration, which emphasize the importance of combining data-driven intelligence with physically informed modeling frameworks for reliable power system operation^43,44.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
E-Grid anonymized monthly accounting dataset	Provincial power grid in eastern China	N/A	Anonymized monthly production and operation cost dataset (2010–2025) comprising 192 observations, including physical asset, operation and maintenance, macroeconomic, and environmental data. Used for model development (156 observations) and out-of-sample testing (36 observations).
Matplotlib plotting library	Matplotlib development team	3.5.2	Used to generate manuscript figures and graphical outputs, including model-performance plots, topology visualizations, residual diagnostics, radar charts, and SHAP-related visual summaries.
NumPy numerical computing library	NumPy developers	1.22.3	Used for numerical array operations, matrix calculations, and reproducibility control. A fixed random seed of 42 was applied where relevant.
Pandas data manipulation library	pandas development team	1.4.2	Used for data import, integration, restructuring, filtering, and preprocessing, including missing-value imputation, time-series alignment, and preparation of model input variables.
Python programming environment	Python Software Foundation	3.9.13	Programming environment used to implement the complete forecasting workflow, including data preprocessing, feature engineering, model training, hyperparameter optimization, prediction, statistical testing, and performance evaluation.
Scikit-learn machine learning library	scikit-learn contributors	1.0.2	Used for data preprocessing, training-validation splitting, auxiliary model evaluation, and statistical procedures, including Pearson correlation analysis and calculation of selected performance metrics.
SciPy scientific computing library	SciPy contributors	1.9.0	Used for statistical calculations, including Pearson correlation analysis and supporting procedures for outlier treatment and winsorization.
Tree-structured Parzen Estimator implementation	Optuna contributors	3.1.0	Bayesian optimization implementation used to tune XGBoost hyperparameters, including maximum tree depth, learning rate, and regularization parameters. The Tree-structured Parzen Estimator was implemented through the open-source Optuna package.
XGBoost software library	DMLC / XGBoost contributors	1.7.1	Decision-tree-based ensemble learning library used for nonlinear residual compensation. The model implemented a regularized objective function with Huber loss and tree-complexity penalties, as described in Equations 7–9.