Comprehensive evaluation system for electric power economic risk and customer value
The comprehensive evaluation system for power economic risk and customer value, which integrates multi-source data fusion and dynamic weight adaptive adjustment, solves the problems of timeliness and interpretability in risk and value evaluation in existing technologies, and realizes refined decision support for the power system.
Patent Information
- Application Number
- CN202511993683.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-26
- Publication Date
- 2026-02-10
AI Technical Summary
The existing power risk and customer value assessment system is inadequate in terms of timeliness, precision, and compliance, making it difficult to meet the precision decision-making needs of market-oriented operations. Furthermore, the existing system ignores the deep interaction between risk and value, resulting in poor interpretability of the assessment results.
It employs modules for multi-source data acquisition and preprocessing, multi-source data fusion, two-dimensional coupled evaluation modeling, dynamic weight adaptive adjustment, and intelligent decision-making and visualization output. It combines a fractional memory kernel weighted attention mechanism, an interactive attribution mechanism, and the Wasserstein centroid method to conduct a comprehensive evaluation of risk and value.
It enables precise quantification and coupled evaluation of power economic risks and customer value, provides interpretable decision-making basis, supports regional pricing, differentiated services and risk mitigation strategies, and improves the robustness of evaluation results and the reliability of business decisions.
Smart Images
Figure CN121503818A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power system operation technology, specifically to a comprehensive evaluation system for power economic risks and customer value. Background Technology
[0002] With the continued deepening of power system reform, the power market is showing a trend of diversified participation and increasingly complex market-based trading mechanisms. Meanwhile, the volatility and uncertainty brought about by the high proportion of renewable energy grid connection further exacerbate the complexity of power system operation and market management. Against this backdrop, accurately controlling power economic risks (such as electricity price fluctuation risks and supply-demand imbalance risks) and scientifically quantifying customer value have become key prerequisites for power supply companies and electricity sales companies to optimize resource allocation, formulate differentiated operating strategies, and enhance core competitiveness. Power economic risks and customer value are not isolated; they are significantly coupled. Customer electricity consumption behavior and value demands directly affect market supply and demand balance, thereby inducing or mitigating economic risks. Conversely, the evolution of market risks will have a reverse effect on customers' electricity consumption decisions and value judgments. Therefore, a synergistic evaluation of the two is of significant practical importance.
[0003] Currently, most evaluation technologies in the power sector focus on single-dimensional analysis and have not yet formed a mature comprehensive risk and value evaluation system, which is insufficient to meet the refined decision-making needs of market-oriented operations.
[0004] Existing power risk modeling and customer value assessment methods often employ statistical or neural network approaches based on short-term memory assumptions (such as traditional sliding windows, ARIMA, or standard LSTM). These methods fail to adequately characterize the long-term dependence and non-exponential decay behavior of historical events, resulting in poor sensitivity to long-term impacts from policy shocks, renewable energy grid integration, and heavy-tailed events, leading to delayed or missed early warnings. Furthermore, existing systems typically treat economic risk and customer value as independent modules or couple them using only linear concatenation of scalar fractions, ignoring the differences in distribution patterns and deep interaction effects between the two types of information. This results in lost interactive information, poor interpretability, and crude decision-making. In addition, most systems use fixed or empirical indicator weights, making it difficult to adapt in real-time to market supply and demand, policy guidance, or customer group evolution. They also frequently use sensitive customer data directly in fine-grained modeling, introducing privacy compliance and causal spurious correlation risks. This makes weight adjustments susceptible to short-term noise or strategic probing, affecting the robustness of assessment results and the reliability of business decisions.
[0005] The aforementioned shortcomings directly restrict the effectiveness of the evaluation system in terms of timeliness, precision, and compliance, and also make it difficult for enterprises to implement interpretable, goal-oriented differentiated pricing and risk mitigation strategies based on the evaluation results.
[0006] In response, this application proposes a comprehensive evaluation system for power economic risks and customer value to address the aforementioned issues. Summary of the Invention
[0007] The purpose of this invention is to provide a comprehensive evaluation system for power economic risks and customer value, in order to solve the problems raised in the background art.
[0008] To achieve the above objectives, the present invention provides the following technical solution:
[0009] A comprehensive evaluation system for power economic risks and customer value, including:
[0010] The multi-source data acquisition and preprocessing module is used to collect and receive electricity market transaction data, power grid operation data, customer behavior data and macroeconomic data, and to perform timestamp unification, noise reduction, missing value imputation and standardization on the data to obtain a synchronous preprocessed data stream.
[0011] The multi-source data fusion module is used to perform domain-specific feature extraction, privacy processing, and fusion on the synchronous preprocessed data stream to obtain a fused feature set, which serves as the input for subsequent modeling.
[0012] The dual-dimensional coupled evaluation modeling module is used to construct an economic risk sub-model and a customer value sub-model based on the fused feature set, and to perform coupled modeling on the two at the coupling layer to generate a coupled evaluation vector. The coupled evaluation vector represents the economic risk characteristics and customer value characteristics at a given point in time.
[0013] The dynamic weight adaptive adjustment module is used to adaptively adjust the weights of evaluation indicators in real time according to changes in electricity market supply and demand, policy guidance indicators and the evolution of customer group characteristics, and to perform weighted calculation on the coupled evaluation vector based on the weights to obtain the weighted evaluation result.
[0014] The intelligent decision-making and visualization output module is used to generate risk warnings, customer segmentation, and operational and pricing suggestions based on the weighted evaluation results, and output them in the form of visual reports and decision instructions.
[0015] Furthermore, the economic risk sub-model employs a fractional memory kernel weighted attention mechanism to weighted encode the temporal features;
[0016] Specifically, when performing time-series encoding on historical electricity prices, load, and grid connection rates, a decaying memory weight with a Mittag-Leffler function as the kernel is used to characterize long memory and fractional decay behavior. The memory kernel K(t,τ) is expressed as:
[0017]
[0018] In the formula, Eα(⋅) is the Mittag-Leffler function, α∈(0,1) is the fractional memory parameter (preferred range 0.3–0.9), λ>0 is the scale parameter, and t and τ are the current and historical time points, respectively;
[0019] This kernel is used to weight the hidden states of a temporal encoder (such as an LSTM or convolutional temporal layer) according to historical time points to generate a risk temporal representation with stronger time dependence, thereby improving the sensitivity and predictive ability to long-period and non-exponential decaying effects. The structural parameters of the sub-model are preferably: the temporal window length is the most recent 168 hours, the hidden state dimension is 64–256, and the memory kernel parameters can be fine-tuned online with the rolling window through small grid search or Bayesian optimization.
[0020] Furthermore, when generating the customer value vector, the customer value sub-model, in addition to using the subjective / objective composite score obtained by information entropy + TOPSIS, also uses an interaction attribution mechanism to measure the marginal contribution of the interaction between features to the value score, so as to reduce computational complexity.
[0021] The interactive attribution mechanism uses an approximate allocation method with random subset sampling to estimate the marginal contribution of the Shapley style, obtaining attribution results (the number of samplings M is configurable, such as M=200–2000). These attribution results are then added as interactive components to the customer value description vector to reflect the nonlinear interaction between customer characteristics and market / risk characteristics in the coupling layer, thereby driving more refined coupled decision-making. The coupling layer is used to couple the risk score sequence output by the economic risk sub-model with the customer value score output by the customer value sub-model through interactive features (e.g., customer price elasticity, peak-valley shift response, demand-side shift ratio) to obtain a coupled evaluation vector reflecting the interaction between risk and value.
[0022] Furthermore, the customer value sub-model employs the following process when generating customer value scores:
[0023] S21. Calculate standardized R, F, and M evaluation indicators based on customer electricity consumption behavior and transaction records;
[0024] S22. Assign objective weights to the evaluation indicators based on the information entropy method;
[0025] S23. Use TOPSIS to sort the evaluation indicators and the objective weights to obtain the customer's current value score;
[0026] S24. Combine the potential value prediction model based on historical behavior to output the potential value score, realize the parallel characterization of current value and potential value and output it to the coupling layer.
[0027] Furthermore, the coupling layer treats the regional / temporal risk distribution and customer value distribution as probability measures and fuses them based on the Wasserstein barycenter to obtain the coupling measure μ* of the coupling distribution (without directly performing linear concatenation of scalar scores), and derives the coupling evaluation vector from the coupling distribution;
[0028] The formula for calculating the coupling measure μ* is:
[0029]
[0030] In the formula, μk represents the probability measure estimated from the k-th sub-model (e.g., the risk distribution of a certain region or the value distribution of a certain type of customer);
[0031] λk is the corresponding initial fusion weight (which can be initialized by mutual information or business reward-driven methods).
[0032] W2() is the second-order Wasserstein distance; by solving for the above optimal centroid and calculating the quantile, expectation and higher-order moments of μ*, the features are used as components of the coupled evaluation vector to input the next step of weight calculation and decision logic.
[0033] The centroid solution can be obtained by using the Sinkhorn-iterative approximation or the sample-based optimal transport Monte Carlo approximation (the number of samples is usually between 1k and 10k) to balance accuracy and computational efficiency.
[0034] Furthermore, the dynamic weight adaptive adjustment module includes a weight generation submodule and a weight normalization submodule;
[0035] The weight generation submodule calculates the original weight score vector based on real-time market supply and demand indicators, policy sensitivity index, and customer group evolution signals.
[0036] The weight normalization submodule normalizes the original score vector to obtain the final weight vector, and the weight normalization is performed using the softmax normalization method.
[0037] Furthermore, the multi-source data fusion module includes a privacy protection process during the feature extraction process. The privacy protection process includes performing local differentiation processing on sensitive identifiers or using domain-based summary statistics to replace single record transmission, so as to generate a structured privacy-preserving feature stream.
[0038] The fusion process simultaneously performs feature selection and feature transformation to ensure that subsequent modeling retains key discriminative information while meeting privacy compliance requirements.
[0039] Furthermore, the system also includes:
[0040] The feedback and online update module is used to collect the implementation effect and subsequent operation data of the intelligent decision-making and visualization output module, and feed the implementation effect and subsequent operation data back to the multi-source data fusion module and the dual-dimensional coupled evaluation modeling module to complete the online iterative update of the model.
[0041] Furthermore, the intelligent decision-making and visualization output module includes a rules engine submodule and a report / interactive interface submodule. The rules engine triggers graded risk warning thresholds, customer segmentation labeling strategies, and regionalized pricing / service suggestions based on weighted evaluation results, and can generate executable operation and maintenance and marketing decision instructions. The report / interactive interface provides multi-perspective visualization (including time series trends, spatial distribution, and customer group profiles), and supports exporting decision results in a machine-readable format for external systems to execute.
[0042] Compared to existing technologies, the comprehensive evaluation system for power economic risk and customer value provided by this invention treats risk and value as probabilistic measures and couples them using optimal transmission methods such as the Wasserstein barycenter. It then extracts quantiles, tail features, and higher-order moments from the coupled distribution as coupled evaluation vectors. Distribution-level fusion avoids the information loss and neglect of interaction effects caused by directly concatenating scalar scores in traditional methods. This ensures that the evaluation results not only provide a "single-point" risk / value judgment but also offer complete information on the distribution pattern. Combined with interaction attribution based on approximate Shapley, the system can reveal the nonlinear interaction between customer value and market / risk factors and its marginal contribution, thus providing a quantifiable and interpretable basis for regional pricing, differentiated services, target customer retention strategies, and risk mitigation measures. Attached Figure Description
[0043] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.
[0044] Figure 1 This is a schematic diagram of the overall structure provided for an embodiment of the present invention. Detailed Implementation
[0045] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.
[0046] As attached Figure 1 As shown:
[0047] Example 1:
[0048] A comprehensive evaluation system for power economic risks and customer value, comprising:
[0049] The multi-source data acquisition and preprocessing module 101 is used to acquire and receive electricity market transaction data, power grid operation data, customer behavior data and macroeconomic data, and to perform timestamp unification, noise reduction, missing value imputation and standardization on the data to obtain a synchronous preprocessed data stream.
[0050] Specifically, the system accesses the electricity market matching interface, the power grid SCADA / EMS operation bus, smart meter / concentrator reporting and CRM / billing transaction logs, and external macroeconomic and industry APIs in parallel at the distributed acquisition layer. Data enters the streaming bus according to timestamps (it is recommended to use Kafka as a buffer and Flink or Spark Streaming for real-time processing). The access strategy is layered according to sampling resolution—for power grid measurement points that require high-frequency sensing, 1–60 seconds of raw sampling is retained and aggregated in the streaming process at 1-minute or 5-minute windows. Market matching data retains minute-level prices and clearing volumes, and customer-side metering is stored in the database at a default 15-minute or hourly granularity.
[0051] Furthermore, after entering stream processing, timestamp alignment is first performed (alignment benchmark is set to 1 minute, allowing upward / downward aggregation by scenario). Then, short-term noise suppression (e.g., 3-point median filtering or 5-point moving average), missing data handling (pre-push padding, allowing continuous padding for up to 60 minutes, marking missing data and recording quality metadata for values exceeding the threshold), and rolling normalization (rolling window is 30 days by default to suppress seasonality) are performed. Finally, each preprocessed record is written to columnar storage (Parquet) and published in a message queue in a unified serialization format. The output data stream is a "synchronous preprocessed data stream", which includes a unified timestamp, source tag, quality identifier, and standardized numerical fields for subsequent fusion modules to subscribe to.
[0052] The multi-source data fusion module 102 is used to perform domain-specific feature extraction, privacy processing and fusion on the synchronous preprocessed data stream to obtain a fused feature set, which is used as the input for subsequent modeling.
[0053] Specifically, the multi-source data fusion module includes a privacy protection process during feature extraction. The privacy protection process includes performing local differentiation processing on sensitive identifiers or using domain-based summary statistics to replace single record transmission, so as to generate a structured privacy-preserving feature stream.
[0054] The fusion process simultaneously performs feature selection and feature transformation to ensure that subsequent modeling retains key discriminative information while meeting privacy compliance requirements.
[0055] Furthermore, the synchronous preprocessed data stream is read by the streaming / batch hybrid platform (Spark ML / Flink + vector database) and features are extracted according to the domain: the peak-to-valley ratio, short-term volatility and spectral energy of the power grid and load sequences are calculated within a 24-hour sliding window (short-time Fourier or wavelet transform can be used, and the sliding window step size is usually 1 hour).
[0056] The market sequence calculates 24-hour and 7-day rolling volatility and retracement indicators; on the client side, standardized R, F, and M vectors are constructed and the event sequence is mapped into a dense vector through behavioral embedding (the embedding dimension is usually set to 32–128).
[0057] At the privacy protection level, the system uses keyed hashing (e.g., 2^16 buckets) on identifiers and applies local differential privacy noise (Laplace mechanism, ε=0.5–1.0) to the aggregation statistics, while retaining group-level statistics to support modeling; feature selection is achieved by mutual information threshold (threshold can be set to 0.01) and correlation coefficient elimination (eliminating absolute correlation >0.95), and high-dimensional features are compressed to approximately 256–512 dimensions using PCA or autoencoder when necessary; the final output is a "fusion feature set", a dense vector stream indexed by time and domain (containing source, time window identifier, feature quality score), which is available for batch / stream reading by the two-dimensional coupled modeling module.
[0058] The dual-dimensional coupled evaluation modeling module 103 is used to construct an economic risk sub-model and a customer value sub-model based on the fused feature set, and to perform coupled modeling on the two at the coupling layer to generate a coupled evaluation vector, wherein the coupled evaluation vector represents the economic risk characteristics and customer value characteristics at a given point in time.
[0059] Specifically, the economic risk sub-model uses a fractional memory kernel weighted attention mechanism to weight and encode time-series features;
[0060] When performing time-series encoding on historical electricity prices, load and grid connection rate, a decaying memory weight with the Mittag-Leffler function as the core is used to characterize long memory and fractional decay behavior.
[0061] This kernel is used to weight the hidden states of a temporal encoder (such as an LSTM or convolutional temporal layer) according to historical time points to generate a risk temporal representation with stronger time dependence, thereby improving the sensitivity and predictive ability to long-period and non-exponential decay effects. The preferred structural parameters of the sub-model are: a temporal window length of the most recent 168 hours, a hidden state dimension of 64–256, and memory kernel parameters that can be fine-tuned online with the rolling window through small-scale grid search or Bayesian optimization.
[0062] When generating customer value vectors, the customer value sub-model, in addition to using the subjective / objective composite score obtained by information entropy-TOPSIS, also uses an interaction attribution mechanism to measure the marginal contribution of feature interactions to value scoring, thereby reducing computational complexity.
[0063] The interactive attribution mechanism uses an approximate allocation method with random subset sampling to estimate the marginal contribution of the Shapley style, obtaining attribution results (the number of samplings M is configurable; in the embodiments, M=200–2000 is commonly used). These attribution results are then added as interactive components to the customer value description vector to reflect the nonlinear interactive influence of customer characteristics and market / risk characteristics in the coupling layer, thereby driving more refined coupled decisions. The coupling layer is used to couple the risk score sequence output by the economic risk sub-model with the customer value score output by the customer value sub-model through interactive features (e.g., customer price elasticity, peak-valley shift response, demand-side shift ratio) to obtain a coupled evaluation vector reflecting the risk-value interactive influence.
[0064] The customer value sub-model uses the following process to generate customer value scores:
[0065] S21. Calculate standardized R, F, and M evaluation indicators based on customer electricity consumption behavior and transaction records;
[0066] S22. Assign objective weights to the evaluation indicators based on the information entropy method;
[0067] S23. Use TOPSIS to sort the evaluation indicators and the objective weights to obtain the customer's current value score;
[0068] S24. Combine the potential value prediction model based on historical behavior to output the potential value score, realize the parallel characterization of current value and potential value and output it to the coupling layer.
[0069] The coupling layer treats the regional / temporal risk distribution and customer value distribution as probability measures and fuses them based on the Wasserstein barycenter to obtain the coupling measure μ* of the coupling distribution (without directly performing linear concatenation of scalar scores), from which the coupling evaluation vector is derived.
[0070] The centroid solution can be obtained by using the Sinkhorn-iterative approximation or the sample-based optimal transport Monte Carlo approximation (the number of samples is usually between 1k and 10k) to balance accuracy and computational efficiency.
[0071] Furthermore, within the dual-dimensional coupled modeling, the economic risk sub-model uses the most recent week (default 168 hours) or an adaptive window as the input sequence to perform time-series encoding on time-series features such as electricity price, load and grid connection rate. The model adopts a time-series encoder with attention mechanism (such as a two-layer LSTM, with hidden dimensions that can be set to 64–256) and introduces a fractional memory kernel in the attention calculation to characterize long memory and non-exponential decay effects.
[0072] Specifically, the aggregate weights of the hidden states at historical moments use a Mittag-Leffler kernel. The fractional-order parameter α is selected in the range of 0.3–0.9 to match the long-term dependence characteristics of the power system, and the scale λ is used to control the decay rate and can be finely adjusted online with the rolling window.
[0073] This kernel replaces the classic exponential decay in attention weight calculation, enabling remote but information-rich events (such as policy releases or long-term trend changes) to influence the current risk representation in a non-exponential decay manner. The encoded output is a risk score sequence (including confidence intervals and time-sensitive weights) that is separated by region / time. This sequence is output as a structured time-series record for use by the coupling layer.
[0074] When generating current and potential value for each customer, the customer value sub-model first calculates standardized R, F, and M indicators based on current and historical electricity consumption behavior, and objectively assigns indicator weights within the sample using the information entropy method. Then, it ranks and scores the current value using TOPSIS. To characterize the nonlinear interaction effects between features and improve interpretability, the model performs interaction attribution in parallel—using a random subset sampling approximation Shapley method (sampling times M are configurable, typically 200–2000) on a distributed computing environment (Spark + GPU worker) to estimate the marginal contribution of each feature and feature group, and adds these attribution components as interaction features to the customer value description vector. Simultaneously, to predict potential value, an interpretable gradient boosting regression (LightGBM, with the number of trees adjustable between 50–200 and a learning rate of 0.01–0.1) is used to output the potential value score for future periods (e.g., 30 days). The customer value sub-model outputs a "value vector" for each customer, including the current value score, potential value score, important interaction attribution vector, and uncertainty estimate, which flows to the coupling layer by customer ID and timestamp.
[0075] Wasserstein barycenter fusion and coupled evaluation vector extraction in the coupling layer: Instead of simply concatenating scalar scores, the coupling layer treats the risk distribution and customer value distribution on the regional / time scale as probability measures and calculates their Wasserstein barycenters to obtain representative coupling measures. This is formally achieved by solving:
[0076]
[0077] Where μk is the measure given by the sub-model and λk is the initial fusion coefficient, the coupled distribution μ∗ is obtained. The centroid solution is implemented in engineering by sample-level approximation and Sinkhorn entropy regularization iteration (to ensure computational scalability, with the number of samples in the range of 1k–10k) or by the approximation algorithm of the POT (Python Optimal Transport) library. Quantiles, conditional expectation, tail risk measure and higher-order moments are extracted from μ∗ as components of the coupled evaluation vector. At the same time, interactive estimates such as price elasticity and peak-valley shift response are incorporated into this vector. The output of the coupling layer is the "coupled evaluation vector stream", each record of which contains a summary of the coupled distribution, several key statistics and corresponding source weight metadata, and is submitted to the dynamic weight and decision module for weighted evaluation.
[0078] The dynamic weight adaptive adjustment module 104 is used to adaptively adjust the weights of the evaluation indicators in real time according to changes in electricity market supply and demand, policy guidance indicators and the evolution of customer group characteristics, and to perform weighted calculation on the coupled evaluation vector based on the weights to obtain the weighted evaluation result.
[0079] Specifically, the dynamic weight adaptive adjustment module includes a weight generation submodule and a weight normalization submodule;
[0080] The weight generation submodule calculates the original weight score vector based on real-time market supply and demand indicators, policy sensitivity index, and customer group evolution signals.
[0081] The weight normalization submodule normalizes the original score vector to obtain the final weight vector, and the weight normalization is performed using the softmax normalization method.
[0082] Furthermore, the dynamic weighting module uses business objectives (such as risk triggering and customer response / churn within the next 24 hours) as a benchmark to discriminate the weights of each component of the coupled evaluation vector. In engineering implementation, a learnable mutual information approximator (such as MINE based on neural networks to estimate the amount of information between indicators and future objectives in a rolling window) or a multiplicative update mechanism driven by online evaluation rewards is used to generate the original weight scores, and the weights are exponentially smoothed to prevent abrupt changes.
[0083] The weights and coupled evaluation vectors are combined with linear or nonlinear mappings to form a "weighted evaluation result". The rule engine triggers hierarchical warnings, customer segmentation tags, personalized pricing and operation and maintenance actions based on historical percentile thresholds and business strategies, and sends them to external execution systems (such as EMS control interfaces or CRM marketing engines) in JSON / REST or message queue format. At the same time, all decisions and model snapshots, along with implementation effects (warning hits, load deviations, customer response rates and churn indicators), are collected in real time and fed back to the feedback module.
[0084] The feedback module generates sample increments, hyperparameter fine-tuning suggestions, and incremental training tasks (LSTM mini-batch fine-tuning, LightGBM incremental training, etc.) according to the triggering strategy (such as the early warning hit rate being lower than 80% or the customer churn rate rising abnormally). After verification, it automatically applies and records the new model identifier, thus forming a closed-loop data flow chain of "preprocessed data → feature fusion → risk / value vector → coupling distribution and coupling vector → weighted evaluation → decision output → implementation effect → model update".
[0085] The intelligent decision-making and visualization output module 105 is used to generate risk warnings, customer segmentation, operation and pricing suggestions based on the weighted evaluation results, and output them in the form of visual reports and decision instructions.
[0086] Specifically, the intelligent decision-making and visualization output module includes a rules engine submodule and a report / interactive interface submodule. The rules engine triggers graded risk warning thresholds, customer segmentation labeling strategies, and regionalized pricing / service suggestions based on weighted evaluation results, and can generate executable operation and maintenance and marketing decision instructions. The report / interactive interface provides multi-perspective visualization (including time series trends, spatial distribution, and customer group profiles) and supports exporting decision results in a machine-readable format for external systems to execute.
[0087] Specifically, the system also includes a feedback and online update module 106, which is used to collect the implementation effect and subsequent operation data of the intelligent decision-making and visualization output module, and feed the implementation effect and subsequent operation data back to the multi-source data fusion module and the dual-dimensional coupled evaluation modeling module to complete the online iterative update of the model.
[0088] As shown above, this system treats risk and value as probability measures and uses optimal transmission methods such as the Wasserstein barycenter for coupling. It then extracts quantiles, tail features, and higher-order moments from the coupled distribution as coupled evaluation vectors. Distribution-level fusion avoids the information loss and neglect of interaction effects caused by directly concatenating scalar scores in traditional methods. This ensures that the evaluation results not only provide a "single-point" risk / value judgment but also offer complete information on the distributional structure. Combined with interaction attribution based on approximate Shapley, the system can reveal the nonlinear interaction between customer value and market / risk factors and its marginal contribution, thus providing a quantifiable and interpretable basis for decision-making regarding regional pricing, differentiated services, target customer retention strategies, and risk mitigation measures.
[0089] By introducing fractional-order memory kernels (such as Mittag-Leffler-based kernels) into the economic risk sub-model and embedding them into attention calculations, this system can retain and amplify the long-term impact of historical key events in a non-exponential decay form. This allows distant but persistently influential market signals or policy shocks to continuously affect the current risk representation and scoring. This mechanism compensates for the shortcomings of traditional time-series models in terms of long-cycle dependence and poor sensitivity to heavy-tailed events, thereby improving the ability to identify complex, multi-factor driven risks such as abnormal electricity prices, structural load migration, and policy regulation in advance. It enhances the timeliness of early warnings and the policy response window, which is conducive to the early deployment of operation and maintenance and market decisions.
[0090] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.
Claims
1. A comprehensive evaluation system for power economic risks and customer value, characterized in that, include: The multi-source data acquisition and preprocessing module is used to collect and receive electricity market transaction data, power grid operation data, customer behavior data and macroeconomic data, and to perform timestamp unification, noise reduction, missing value imputation and standardization on the data to obtain a synchronous preprocessed data stream. The multi-source data fusion module is used to perform domain-specific feature extraction, privacy processing, and fusion on the synchronous preprocessed data stream to obtain a fused feature set. The dual-dimensional coupled evaluation modeling module is used to construct an economic risk sub-model and a customer value sub-model based on the fused feature set, and to perform coupled modeling on the two at the coupling layer to generate a coupled evaluation vector. The dynamic weight adaptive adjustment module is used to adaptively adjust the weights of evaluation indicators in real time according to changes in electricity market supply and demand, policy guidance indicators and the evolution of customer group characteristics, and to perform weighted calculation on the coupled evaluation vector based on the weights to obtain the weighted evaluation result. The intelligent decision-making and visualization output module is used to generate risk warnings, customer segmentation, and operational and pricing suggestions based on the weighted evaluation results, and output them in the form of visual reports and decision instructions.
2. The comprehensive evaluation system for power economic risk and customer value according to claim 1, characterized in that, The economic risk sub-model uses a fractional memory kernel weighted attention mechanism to weighted encode time-series features; The memory kernel K(t,τ) is represented as: In the formula, Eα(⋅) is the Mittag-Leffler function, α∈(0,1) is the fractional memory parameter, λ>0 is the scale parameter, and t and τ are the current and historical time points, respectively.
3. The comprehensive evaluation system for power economic risk and customer value according to claim 1, characterized in that, When generating customer value vectors, the customer value sub-model, in addition to using the subjective / objective composite score obtained by information entropy-TOPSIS, also uses an interaction attribution mechanism to measure the marginal contribution of feature interactions to value scoring, thereby reducing computational complexity. The interactive attribution mechanism uses an approximate allocation method with random subset sampling to estimate the marginal contribution of the Shapley style, obtain the attribution result, and add the attribution result as an interactive component to the customer value description vector, so as to reflect the nonlinear interactive influence of customer characteristics and market / risk characteristics in the coupling layer, thereby driving more refined coupled decision-making.
4. The comprehensive evaluation system for power economic risk and customer value according to claim 3, characterized in that, The customer value sub-model uses the following process to generate customer value scores: S21. Calculate standardized R, F, and M evaluation indicators based on customer electricity consumption behavior and transaction records; S22. Assign objective weights to the evaluation indicators based on the information entropy method; S23. Use TOPSIS to sort the evaluation indicators and the objective weights to obtain the customer's current value score; S24. Combine the potential value prediction model based on historical behavior to output the potential value score, realize the parallel characterization of current value and potential value and output it to the coupling layer.
5. The comprehensive evaluation system for power economic risk and customer value according to claim 4, characterized in that, The coupling layer treats the regional / temporal risk distribution and customer value distribution as probability measures and fuses them based on the Wasserstein barycenter to obtain the coupling measure μ* of the coupling distribution, from which the coupling evaluation vector is derived; The formula for calculating the coupling measure μ* is: In the formula, μk represents the probability measure estimated by the k-th sub-model; λk is the corresponding initial fusion weight; W2() is the second-order Wasserstein distance; by solving for the above optimal centroid and calculating the quantile, expectation and higher-order moments of μ*, the features are used as components of the coupled evaluation vector to input the next step of weight calculation and decision logic; The centroid solution can be obtained using the Sinkhorn-iterative approximation or the sample-based optimal transport Monte Carlo approximation to balance accuracy and computational efficiency.
6. The comprehensive evaluation system for power economic risk and customer value according to claim 1, characterized in that, The dynamic weight adaptive adjustment module includes a weight generation submodule and a weight normalization submodule; The weight generation submodule calculates the original weight score vector based on real-time market supply and demand indicators, policy sensitivity index, and customer group evolution signals. The weight normalization submodule normalizes the original score vector to obtain the final weight vector, and the weight normalization is performed using the softmax normalization method.
7. The comprehensive evaluation system for power economic risk and customer value according to claim 1, characterized in that, The multi-source data fusion module includes a privacy protection process during feature extraction. This process includes performing local differentiation on sensitive identifiers or using domain-based summary statistics to replace single record transmission, in order to generate a structured, privacy-preserving feature stream.
8. The comprehensive evaluation system for power economic risk and customer value according to claim 1, characterized in that, The intelligent decision-making and visualization output module includes: a rule engine submodule and a report / interactive interface submodule; The rules engine submodule triggers graded risk warning thresholds, customer segmentation labeling strategies, and regionalized pricing / service suggestions based on weighted evaluation results, and can generate executable operation and maintenance and marketing decision instructions; The report / interactive interface submodule provides multi-perspective visualization and supports exporting decision results in a machine-readable format for external systems to execute.
9. The comprehensive evaluation system for power economic risk and customer value according to claim 1, characterized in that, The system also includes: The feedback and online update module is used to collect the implementation effect and subsequent operation data of the intelligent decision-making and visualization output module, and feed the implementation effect and subsequent operation data back to the multi-source data fusion module and the dual-dimensional coupled evaluation modeling module to complete the online iterative update of the model.