Data-Driven Risk Warning of Electricity Sales Companies in the Whole Business Process

Abstract

As China’s power market reforms deepen, the scale of market operations and the number of participants have reached new highs, introducing increasingly complex threats and heightened risk scenarios. Traditional risk early warning systems for electricity sales companies are heavily influenced by subjective factors, incomplete data, and poor real-time performance, which cannot meet the requirements of sustainable development. To achieve efficient, full-chain, and sustainable risk control, this paper proposes a data-driven risk warning method for electricity sales companies, encompassing the entire sales process. Firstly, based on data correlations across the electricity sales process, appropriate data sources for risk warnings are identified. Key elements are then extracted using Principal Component Analysis (PCA), while historical business data is adaptively clustered, with risk warning levels classified using the Adaptive Sparrow Optimization Density Peak Clustering Algorithm (DPC-SSA). Lastly, dynamic risk warnings are generated through the stacking identification model. The effectiveness and practicality of the proposed method are validated through an analysis using real data from a provincial power trading management platform.

1. Introduction

In 2023, the total electricity traded in China’s power market reached 5.7 trillion kilowatt-hours, a 7.9% increase year-on-year. Additionally, the number of registered entities in trading organizations reached 708,000; the number of retail customers meeting the conditions of market-based transactions is gradually increasing, and the electricity sales business of electricity retailers is becoming increasingly complex [1,2,3]. As market-oriented reforms progress [4], diversified competition patterns continue to develop in China’s power market. Transactions in this market now feature diversification, digitalization, and scaling, which introduce uncertainties and heighten associated risks [5].

Electricity sales companies, important main players in the electricity market, have huge differences in their qualifications, credit, and operational capabilities, making them a major source of market risk [6]. Violations and breaches of trust by electricity sales companies affect the order of market transactions and bring operational pressure to the power regulator; on the other hand, they jeopardize the economic interests of other market members and reduce the motivation of market players to participate in market construction [7]. Consequently, an efficient, end-to-end risk warning method is urgently needed for electricity sales companies to identify high-risk segments in business processes and promote healthy, high-quality market development.

Risk control in the electricity market is a measure that cannot be ignored for its sustainable development, and nowadays, research is mainly focused on risk categories and market players. Risk categories include the seasonal risk of transaction prices [8], price fluctuation risk, electricity tariff risk in the electricity market, and the risk of power user loss of power grid companies under the new electricity reform [9].

Electricity sales companies also have wholesale business risks, retail business risks, and internal business risks in their operations [10]. Under the medium- and long-term electricity market, power sales companies also have certain risks of supply stability and cost fluctuations [11,12]. From the perspective of market players, some have arrears of payment and overuse of market power in the electricity market. Comprehensive identification of market power risks of electricity sales companies requires extensive multidimensional time scales and hierarchical regulatory structures [13].

Therefore, to study the risk of power sales companies, it is necessary to consider multiple dimensions and build a comprehensive risk indicator system. According to the actual situation of the market, appropriate indicators should be chosen to build an indicator system; the authors of [14] considered the market share, capital, and employee qualification of power sales companies, and the authors of [15,16] mentioned market force. Compared to foreign electricity markets, the domestic electricity market forces are highly concentrated, the Herfindahl–Hirschman index is higher than the threshold of 1800, and the abuse of market forces is more common [17]. The authors of [18] used an improved support vector machine to automatically identify market power abuses and reduce credit risk in electricity market transactions. Some scholars have also measured the risks faced by electricity retailers in the power sales market from multiple perspectives and constructed a risk evaluation system for electricity retailers [19]. The authors of [20] used the Intuitionistic Mold Lake Hierarchy Analysis Method (IFAHP) based on intuition to identify the credit risk of electricity sales companies and established a credit risk indicator system for electricity sales companies from four dimensions: external basic environment, operational credit risk, financial credit risk, and transactional credit risk. The construction of indicators can be optimized through the combination of Environmental, Social, and Governance (ESG) factors using principal component analysis and grey correlation methods [21].

In terms of risk assessment models, unique models have been developed in various regions of the world to assess the credit risk of power sales companies [22,23,24,25]. The authors of [26] proposed an identification method for assessing the potential for collusion in the electricity market by calculating the Jacobian matrix of profits to identify the presence of collusion in electricity sales companies. The authors of [27] proposed a credit rating method for retailers based on the Bayesian Best Worst Method (BBWM)–Cloud model. In [28], the researchers developed a credit risk evaluation model for electricity sales companies by combining the Intuitionistic Fuzzy Analytic Hierarchy Process (IFAHP) with the cloud model. The authors of [29] developed four credit models to assess credit risk under extreme economic conditions using innovative Conditional Value-at-Risk (CVaR) techniques applied to structural models (Xtreme-S), transition models (Xtreme-T), and quantile regression models (Xtreme-Q). The authors of [20,30,31,32] integrated extenics theory, cloud modeling, and cloud entropy optimization algorithms to evaluate the health of electricity market operations.

In terms of risk early warning, the authors of [33] developed an enhanced backpropagation (BP) neural network model to model the relationship between market power behavior and risk levels, achieving early warning for market power in the electricity spot market. The authors of [34] established a warning indicator system to capture market power risks and applied the Value at Risk (VaR) method based on the Generalized Autoregressive Conditional Heteroskedasticity (GARCH) model to determine thresholds for risk indicators, thus enabling early warnings of market power risk levels. The authors of [35] proposed a method for identifying potentially dangerous behaviors in the electricity market using a cloud model and fuzzy Petri nets to provide real-time early warnings. The authors of [36] used logit regression and artificial neural networks (ANN), the Naïve Bayes algorithm, and the C4.5 decision tree model to predict the overall risk score. Data-driven modeling [37], a common method, has not yet been applied in the early warning of the risk of electricity sales companies.

The above analysis primarily focuses on identifying the security threats posed by power sales companies from the perspective of a certain part of the market in which they participate, but it does not cover the whole process of the business. The main deficiencies are as follows:

• The early warning method based on the indicator system for identifying the risk level of electricity sales companies relies on expert experience in setting risk thresholds, which is influenced by subjective factors. Additionally, the method has a certain degree of lag and cannot be updated in real time to reflect the actual conditions of the electricity sales company.
• The current research on early warning methods primarily analyzes a certain part of the electricity sales company and lacks comprehensive risk warnings throughout the whole transaction process, which is unable to meet the requirements of its sustainable development.

Therefore, to address the aforementioned deficiencies, this paper proposes a data-driven risk warning method for electricity sales companies in the whole business process. The core innovations are expressed in the following aspects:

(1)

Based on the data relevance of the whole business process, this method analyzes the risk factor clusters throughout the whole process. Risk warning data sources for electricity sales companies are identified from the Qualifications Management, Retail Transaction, Wholesale Trading, Settlement, Information Disclosure, and Credit Evaluation processes to improve comprehensiveness.

(2)

The key elements of risk assessment are refined using PCA, and DPC-SSA is applied for adaptive clustering and automatic classification of risk warning levels based on historical business data. The stacking model is employed to implement dynamic risk early warning for electricity sales companies. This method enables the scientific management and control of risks generated by electricity sales companies, reduces losses from potential risks, im-proves the stability and reliability of the electricity market, and fosters a more favorable business environment for market participants. This method ensures that electricity sales companies can participate sustainably in the electricity market.

2. Design of Risk Warning Method for Electricity Sales Companies

The framework of the risk warning method for electricity sales companies, based on complete power sales data, is illustrated in Figure 1. It consists of three modules: data analysis and processing, adaptive clustering and risk level classification, and dynamic risk warning.

2.1. Data Analysis and Processing

First, data correlation analysis is conducted across all aspects of the electricity sales process to mine risk-related data elements and construct a data structure for risk identification. Considering the heterogeneity and volume differences of data sources, data standardization is applied:

$$X^{\prime }={\displaystyle \frac{x-x_{\min }}{x_{\max }-x_{\min }}}$$

(1)

where $X^{\prime }$ represents the standardized indicator data, and $x_{\max }$ and $x_{\min }$ represent the maximum and minimum values of the indicator data, respectively.

Simultaneously, to achieve risk level classification, negative data indicators are processed for alignment:

$$X_{ij}=M_j-x_{ij}$$

(2)

where $X_{ij}$ represents the aligned data, $M_j$ represents the maximum value of the j-th indicator, and $x_{ij}$ denotes the i-th feature value of the j-th sample.

2.2. Adaptive Clustering and Risk Warning Level Classification

High-dimensional data are extracted using PCA, and the DPC-SSA algorithm is applied for adaptive clustering of the reduced-dimensional data. This result, combined with integrated principal component scores, classifies the clusters into risk levels and labels the risk data.

2.3. Dynamic Risk Warning

A stacking model is constructed using automatically labeled historical data as training samples to implement dynamic risk warnings for electricity sales companies.

3. Data Analysis and Processing

3.1. Analysis of Risk Factors in the Whole Process

To identify potential risk factors, it is essential to examine the business processes of electricity sales companies. This paper categorizes the electricity sales business process into six stages: qualification management, retail transactions, wholesale transactions, settlement, information disclosure, and credit evaluation, as shown in Figure 2.

• Qualification Management Stage [38]. In this stage, third-party organizations review an electricity sales company’s eligibility, technical staffing, and performance bond in accordance with the “Regulations on the Management of Electricity Sales Companies”. Requirements for market entry, technical staffing, and performance bond indicate the company’s profitability and risk resilience to a certain extent.
• Retail Transaction Stage. In this stage, the electricity sales company enters agency purchase contracts with power users. Metrics such as the number of power users, user churn rate, and user satisfaction reflect the service quality and competitiveness of the electricity sales company.
• Wholesale Transaction Stage. During wholesale transactions, an electricity sales company’s purchasing behavior reflects whether it leverages market power for improper benefits, impacting transaction stability and security.
• Settlement Stage. Upon transaction completion, the power trading center performs settlement calculations and assesses the company’s deviations. The company’s deviation and profit-loss status indicate its technical strength and operational capability [39].
• Information Disclosure Stage. Beyond regular transactions, the electricity sales company must disclose its operational, financial, and pricing information. The frequency of disclosure errors, omissions, and delays reflects the company’s level of responsibility to a certain extent.
• Credit Evaluation Stage. In this stage, the power trading center evaluates the electricity sales company’s credit status and financial condition, indicating its creditworthiness and financial health.

3.2. Construction of Risk Warning Data Source

After analyzing the risk factors of electricity sales companies, a risk warning data source is constructed to support subsequent risk warnings. As shown in Figure 3, the risk data source is divided into three layers: stage, risk, and data. The stage layer represents the entire business process of the electricity sales company, while the risk layer captures the associated risks within each stage. The data layer quantifies the risk layer and serves as the direct data source for risk warnings.

4. Adaptive Clustering and Risk Level Classification

The DPC-SSA algorithm is used to cluster historical data after dimensionality reduction, combined with the principal component composite score to enable automatic classification and data labeling of risk warning levels for the electricity sales company. The steps for adaptive clustering and risk level classification with the DPC-SSA algorithm are illustrated in Figure 4.

4.1. Principal Component Analysis

PCA is a multivariate statistical technique that combines multiple correlated indicators into a set of uncorrelated composite indicators [40]. PCA is used to identify key risk indicators for power sales, remove redundant information, and improve the stacking model’s performance. The main steps of PCA are as follows:

(1)

Preprocess the dataset X to eliminate the effects of dimensionality.

(2)

Calculate the covariance matrix V:

$$V={\displaystyle \frac{1}{\mathrm{n}}}\overline{X}{\overline{X}}^T$$

(3)

(3)

Perform eigen decomposition on the covariance matrix to compute the eigenvalues and corresponding eigenvectors $w_i$, arrange the eigenvalues in descending order, and obtain the cumulative contribution rate λ.

(4)

Based on the set cumulative contribution rate, select the top m eigenvalues and their corresponding eigenvectors $W_k$, where the eigenvectors $W_k$ form the loading coefficient matrix. The extracted m principal components are denoted as $F$:

$$F=W_k^T\overline{X}=\left[F_1,F_2,\cdots F_m\right]$$

(4)

(5)

Calculate the composite score of principal components. The principal component score matrix is $Y=X\ast F$, and the composite score is calculated as:

$$Z=Y\ast \varphi =y_1{\varphi }_1+y_2{\varphi }_2+\cdots y_m{\varphi }_m$$

(5)

where $\varphi$ is the weight vector after normalizing the contribution rate.

4.2. DPC-SSA Algorithm

The traditional Density Peak Clustering (DPC) algorithm can automatically identify density peaks in the data and cluster data around these peaks [41,42]. It is characterized by a simple principle, few parameters, and no need for iteration, allowing it to identify clusters of various shapes and sizes. However, the truncation distance parameter $d_c$ and selection of cluster centers must be manually chosen, introducing subjectivity [43].

To address this, this paper proposes the DPC-SSA clustering algorithm, which uses the Sparrow Search Algorithm (SSA) to conduct a global search for the truncation distance $d_c$, reducing manual intervention and enhancing clustering accuracy. Additionally, an adaptive clustering center selection strategy determines the number of clusters and cluster centers. The process of the DPC-SSA clustering algorithm is shown in Figure 5.

4.2.1. Truncation Distance Optimization

The SSA is a swarm intelligence optimization algorithm inspired by simulating the foraging and anti-predation behaviors of sparrows [44]. Leveraging the SSA’s robust search optimization capabilities, it optimizes the truncation distance to achieve optimal clustering results.

During optimization, sparrow group discoverers decide whether to continue foraging or flee based on environmental safety. The position update formula for discoverers is as follows:

$$X_{i,j}^{t+1}=\left\{\begin{array}{l}{X}_{i,j}^t\cdot \exp \left({\displaystyle \frac{-i}{\tau \cdot T_{\max }}}\right),R_2<S{T}_{\max }\\ X_{i,j}^t+Q\cdot L,R_2>S{T}_{\max }\end{array}\right.$$

(6)

where t represents the current iteration, $T_{\max }$ represents the maximum number of iterations, $X_{i,j}^t$ is the position of the j-th dimension of the i-th sparrow, $Q$ is a normally distributed random number, $L$ is the unit matrix, $\tau$ is a random number in (0, 1], $R_2$ is the alert value within (0, 1), and $S{T}_{\max }$ represents the safety value in [0.5, 1].

Followers adjust their positions by following the discoverers, and the position update formula for followers is:

$$X_{i,j}^{t+1}=\left\{\begin{array}{l}Q\cdot \exp \left({\displaystyle \frac{X_{worst}^t-X_{i,j}^t}{i^2}}\right),i>{\displaystyle \frac{N}{2}}\\ X_p^{t+1}+\left|X_{i,j}^t-X_p^{t+1}\right|\cdot A^{+}\cdot L,i<{\displaystyle \frac{N}{2}}\end{array}\right.$$

(7)

$$A^{+}=A^T{(A{A}^T)}^{-1}$$

(8)

where $X_{worst}^t$ represents the global worst position, $X_p^{t+1}$ is the best-discovered position, $N$ is the number of sparrows, and $A$ is a matrix consisting of −1 and 1.

Some sparrows act as vigilantes, monitoring the foraging environment for threats. The position update formula for guards is:

$$X_{i,j}^{t+1}=\left\{\begin{array}{l}{X}_{best}^t+\beta \cdot \left|X_{i,j}^t-X_{best}^t\right|,f_i>f_g\\ X_{i,j}^I+K\cdot \left({\displaystyle \frac{\left|X_{i,j}^t-X_{worst}^t\right|}{(f_i-f_w)+\epsilon }}\right),f_i=f_g\end{array}\right.$$

(9)

where $X_{best}^t$ denotes the global best position, $\beta$ is the step-size control parameter, $K$ is a random number, $\epsilon$ is a very small value to prevent division by zero, $f_i$ represents the fitness value of the current sparrow individual, and $f_g$ and $f_w$ represent the fitness values of the best and worst sparrow individuals, respectively.

4.2.2. Adaptive Cluster Center Selection

To avoid the issues of multiple selection or omission in determining the cluster centers and numbers from decision graphs, this paper uses an adaptive cluster center selection strategy to determine the cluster centers and the number of clusters [45]. The following are the relevant definitions:

Definition 1.

The decision value ${\gamma }_i$ is the product of the relative distance $\delta$ and local density $\rho$, calculated as follows:

$${\gamma }_i={\rho }_i{\delta }_i$$

(10)

where $i=1,2,\dots n$, $n$ is the number of sample points in the dataset.

Definition 2.

The values ${\gamma }_i$ are sorted in descending order, and the top $n$ data points are selected. The change rate $k$ indicates the trend of ${\gamma }_i$, and $k$ is expressed as follows:

$$k_i^{i+t}={\displaystyle \frac{{\gamma }_{i+t}-{\gamma }_i}{t}}$$

(11)

Definition 3.

The second-order rate of change $\alpha$ is used to represent the speed of the trend change of ${\gamma }_i$, characterizing how quickly ${\gamma }_i$ decreases:

$$\alpha ={\displaystyle \frac{k_j^u}{k_i^j}}$$

(12)

where $i=1$, $u=j+1$, so $\alpha =k_j^{j+1}/k_1^j$, $k_1^j$ is the rate of change from the first data point to the j-th data point. $k_j^{j+1}$ represents the slope from the j-th data point to the next data point.

Definition 4.

An inflection point is a boundary between cluster centers and other points, where the second-order rate of change reaches its maximum among all data points, representing the critical point of the fastest overall trend change of the decision value.

$$X=\arg \max (\alpha (i,j,u))$$

(13)

The data points corresponding to the j $\gamma$-values before the inflection point are selected as cluster centers, with the number of cluster centers being j. This strategy is based on the fact that the decision value ${\gamma }_i$ of the clustering center and the non-clustering center differ greatly. By observing the ${\gamma }_i$ change trend, the boundary between cluster centers and non-cluster centers is identified when the change rate ${\gamma }_i$ transitions from rapid to gradual.

5. Construction of Risk Warning Model for Electricity Sales Companies

5.1. Stacking Model

The risk warning stacking model [46] constructed in this paper consists of two layers: base learners in the first layer and a meta-learner in the second layer. The original dataset is divided into equally sized sub-training sets, which are fed to the k base learners in the first layer to train and generate predictions. The second layer then uses the k outputs from the first layer as input data to train the meta-learner, which produces the final identification results. To reduce the risk of overfitting in the ensemble model, cross-validation is applied to the data.

(1)

Selection of base learners. When selecting base learners, each base model should have distinct learning characteristics to ensure varied strengths in handling different data aspects. Therefore, this paper employs Support Vector Machines (SVM) [47], Random Forest (RF) [48], K-Nearest Neighbors (KNN) [49], and Extreme Gradient Boosting (XGBoost) [50] as base learners in the risk identification model.

(2)

Selection of meta-learner. To prevent overfitting, the meta-learner should have strong generalization capabilities to effectively refine and optimize the base learners’ predictions. Therefore, Gradient Boosting Decision Trees (GBDT) [51] are selected as the meta-learner for the second layer. The structure of the stacking model is illustrated in Figure 6.

5.2. Performance Metrics of the Stacking Model

The performance of the stacking model is evaluated using the confusion matrix, accuracy (A), recall (R), precision (P), and F1 score. The confusion matrix visually represents the model’s performance across categories, as shown in

Table 1. Confusion matrix.

Identifying Values	True Values
	True Value 1	True Value 2
Identifying values 1	$N_{TN}$	$N_{FN}$
Identifying values 2	$N_{FP}$	$N_{TP}$

.

In Table 1, $N_{TN}$ represents the number of cases where both the identified values and true values are 1, $N_{TP}$ represents cases where both are 2; $N_{FN}$ represents cases where the identified values are 1 but the true values are 2, and $N_{FP}$ represents cases where the identified values are 2 but the true values are 1.

The formulas for calculating accuracy (A), recall (R), precision (P), and F1 score are as follows [52]:

$$A={\displaystyle \frac{N_{TP}+N_{TN}}{N_{TP}+N_{TN}+N_{FP}+N_{FN}}}$$

(14)

$$R={\displaystyle \frac{N_{TP}}{N_{TP}+N_{FN}}}$$

(15)

$$P={\displaystyle \frac{N_{TP}}{N_{TP}+N_{FP}}}$$

(16)

$$F1={\displaystyle \frac{2PR}{P+R}}$$

(17)

6. Simulation Verification and Analysis

As an example, we analyzed whole-process business data on the participation of electricity sales companies in the medium-to-the-term electricity market in a whole year for a province in China, encompassing 2448 samples from 204 companies.

6.1. Extraction of Principal Components of Data

Data samples are aligned and standardized, with PCA employed for dimensionality reduction. The number of principal components is selected based on eigenvalues greater than 1. Figure 7 displays the eigenvalues of the principal components and cumulative contribution rate after dimensionality reduction.

Observably, the eigenvalues level off after the 8th principal component, with a cumulative contribution rate of 85.7% by the 10th component, which contains 85.7% of the data information of the original dataset. Therefore, the first 10 principal components effectively summarize the main information of the 36 risk features in the risk feature library, resulting in a risk feature set with a matrix size of 2448 × 10 for electricity sales companies. This method achieves efficient data dimensionality reduction, retains the key information in the data, and has a low impact on the final results.

The principal component loading matrix is constructed using the feature vectors of the principal components, and the loading coefficients of each index are calculated. It represents both the projection of the original variables onto the principal components and the relationship between the components and the original variables. The matrix heat map of loading coefficients is shown in Figure 8.

The correlation of each principal component with risk can be observed in the figure. The higher the number of heat values of the indicator, the more strongly the indicator is considered to correlate with the principal component. Therefore, the final selection of the composition of each principal component is shown in

Table A1. Composition of principal components.

Principal Component	Explanation Factor	Risks
1	Total asset turnover ratio Return on net assets Debt-to-asset ratio Adequacy of performance bond lines Cash flow ratio Kilowatt-hour profit Power sales revenue Loss rate	Default risk Debt risk Insufficient performance bond risk
2	Performance bond amount Transaction deviation rate Planned power quantity Settlement power quantity	Deviation risk Insufficient performance bond risk
3	Wholesale price Wholesale prices year-on-year Low selling price ratio Settlement price	Market power risk
4	Score of information filling standardization Number of missing disclosures Number of failures	Continued access risk Credit risk Non-standard information disclosure risk
5	Number of arrears Amount of arrears	Non-payment risk
6	Deviation in power quantity Number of technical staff	Deviation risk Continued access risk
7	User satisfaction User attrition rate	User loss risk
8	Total assets	Continued access risk
9	Number of untimely disclosures	Non-standard information disclosure risk
10	Number of disclosure errors	Non-standard information disclosure risk

.

As shown in

Table A2. Principal component score coefficient matrix.

1	2	3	4	5	6	7	8	9	10
0.014	0.028	−0.030	−0.015	−0.024	−0.042	0.037	−0.578	0.013	−0.077
0.009	0.039	−0.008	−0.015	−0.021	−0.082	0.333	0.445	−0.141	−0.301
−0.011	0.022	0.033	−0.003	−0.001	0.004	0.508	−0.005	0.021	−0.065
−0.016	0.006	0.263	−0.004	−0.018	0.021	0.048	0.016	−0.003	0.058
0.145	−0.014	0.038	−0.041	−0.002	−0.106	−0.055	−0.017	−0.036	−0.073
−0.012	0.007	0.271	−0.011	−0.009	0.015	0.057	0.018	0.003	0.058
−0.006	0.003	0.030	−0.005	0.486	−0.006	0.017	0.029	−0.016	0.007
0.004	0.043	−0.013	−0.011	0.035	−0.063	0.383	−0.023	0.023	0.498
0.004	−0.003	0.062	−0.019	0.493	−0.012	−0.022	0.000	−0.026	0.001
−0.027	−0.016	−0.011	0.255	−0.016	0.025	−0.007	0.039	0.023	0.018
0.010	0.012	0.010	−0.262	0.008	0.006	−0.011	−0.005	−0.006	−0.014
−0.010	−0.012	−0.010	0.262	−0.008	−0.006	0.012	0.005	0.006	0.014
−0.140	0.013	0.043	0.020	0.007	0.103	−0.055	0.002	0.005	−0.065
−0.011	0.007	0.267	−0.011	−0.023	0.014	0.055	0.017	0.004	0.057
−0.002	−0.010	−0.011	0.251	−0.003	−0.028	0.016	−0.009	0.002	0.007
−0.223	0.008	0.052	0.373	0.006	0.120	−0.066	−0.009	0.009	−0.092
0.145	−0.017	0.024	−0.041	−0.002	−0.093	−0.064	−0.013	−0.033	−0.076
0.140	−0.014	−0.042	−0.021	−0.007	−0.096	0.055	0.000	−0.004	0.065
0.132	−0.037	0.005	−0.055	0.009	−0.010	−0.031	0.023	0.017	−0.043
−0.025	−0.014	0.260	0.007	0.234	0.057	0.018	0.029	0.006	−0.020
0.091	−0.003	−0.011	−0.057	−0.012	0.119	0.012	0.026	−0.009	−0.032
0.132	−0.028	0.007	−0.063	0.008	−0.008	0.027	0.039	−0.022	−0.050
0.117	−0.003	0.027	−0.036	0.013	−0.104	−0.215	−0.200	0.009	0.016
0.100	−0.023	−0.039	−0.067	0.008	0.042	0.060	0.044	0.073	−0.025
−0.021	−0.017	0.035	0.007	−0.005	0.011	−0.149	0.034	−0.094	0.592
−0.065	−0.021	0.024	0.021	−0.005	0.512	−0.009	0.007	−0.004	0.005
−0.055	−0.036	0.011	0.019	−0.004	0.506	−0.037	0.013	−0.001	−0.028
0.011	0.103	−0.038	0.024	0.002	−0.134	−0.117	−0.081	0.618	−0.292
−0.017	0.157	0.006	0.003	−0.014	−0.190	−0.055	0.297	0.292	0.337
0.009	0.172	−0.046	0.013	0.031	−0.233	−0.238	0.036	−0.645	−0.132
−0.017	0.181	−0.012	0.013	0.009	0.011	−0.148	0.129	0.027	−0.009
−0.016	0.192	−0.013	0.006	−0.008	0.050	−0.151	0.141	0.086	−0.010
−0.052	0.295	0.026	−0.009	−0.012	−0.019	0.261	−0.246	−0.052	0.009
−0.061	0.285	0.033	−0.005	0.001	−0.001	0.306	−0.308	−0.064	−0.027
0.005	0.172	−0.007	−0.023	0.007	−0.084	−0.079	0.137	−0.008	−0.004

, the linear equation for each principal component score is derived by calculating the score coefficient matrix for each component. Finally, the composite score of the principal components for each sample is calculated to evaluate the risk warning level of the electricity sales company. A higher composite score indicates a lower risk level, while a lower score signifies significant risks, requiring timely risk control measures.

6.2. Risk Clustering of Electricity Sales Companies

The 10 principal component variables, obtained via PCA dimensionality reduction, are used for the cluster analysis of the sample data with the DPC-SSA algorithm. For this analysis, the maximum number of sparrows in SSA is set to 30, with a maximum of 300 iterations. The contour coefficient serves as the fitness value, and its trend with iteration count is shown in Figure 9.

The trend graph shows that as iterations increase, the fitness value gradually improves, reaching an optimal level at the 164th iteration. Compared to manual selection, SSA optimization finds the optimal truncation distance based on input data characteristics, achieving better clustering results.

After identifying the optimal truncation distance, the inflection point discriminant graph of data points is shown in Figure 10. Following the proposed clustering center selection strategy, when $i=3$, the second-order rate of change α is maximized, indicating an optimal cluster count of 3. The first three γ corresponding data points are selected as cluster centers.

To validate this selection strategy, contour coefficients for different cluster counts are calculated, with the results shown in Figure 11. The silhouette coefficient is maximized when the number of clusters is 3, consistent with the proposed strategy. The clustering outcome, shown in Figure 12, divides samples into three distinct clusters with clear boundaries, indicating effective clustering.

6.3. Classification of Risk Warning Levels

After clustering the data samples using the DPC-SSA algorithm, electricity sales companies are divided into three categories. To assess the actual risk implications of each cluster, the distribution of PCA principal component scores is shown in Figure 13. The figure shows clear distinctions in the composite score boundaries across the three categories of electricity sales companies. A boxplot (Figure 14) further analyzes the score distribution for each category of samples. Since the data are normalized, a higher composite score corresponds to a lower risk level.

Cluster 1 electricity sales companies generally score between 0.28 and 1.04, higher than the other categories, indicating the highest composite score and lowest risk. Cluster 2 companies have a more dispersed data distribution, ranging from −0.6 to 0.3, indicating a moderate risk level relative to Cluster 3. Cluster 3 companies, with scores from −0.99 to −0.58, generally score lower than the other clusters, indicating higher risk.

Based on clustering results and PCA scores, using −0.58 and 0.28 as the cut-off points for risk levels, this paper classifies electricity sales companies as low-, moderate-, and high-risk, corresponding to Clusters 1, 2, and 3, respectively. Additionally, samples are randomly selected from high-risk electricity sales companies for further analysis, with normalized data results shown in Figure 15. Principal component values for these high-risk companies occasionally fall below 0.1, indicating potentially serious risk issues.

According to the annual reports, samples 2, 4, and 5 exhibit severe debt issues. Their gearing ratios are 90.17%, 92.36%, and 89.42%, with cash flow ratios of 31.93, 34.37, and 23.29, indicating severe debt issues. The risk assessment results show that Principal Component 1 for these samples is low, indicating significant debt risks for these electricity sales companies. Principal Component 4 primarily represents dishonesty risk. As shown in Figure 15, samples 1, 4, and 5 may face significant dishonesty risks. In the credit evaluation of the province’s electricity sales companies, these companies received a credit rating of C and are listed on the joint dishonesty punishment list, validating the assessment results.

The Principal Component 5 values for the five high-risk samples are less than 0.1, indicating potential arrears risk. According to the arrears list, these companies have accumulated 17.49 million yuan in arrears, with average arrears of 3.49 million yuan, reflecting serious non-payment risk. The seventh principal component values for samples 1, 2, 4, and 5 are notably low, indicating potential customer attrition issues. According to customer satisfaction surveys from the power trading center, these companies’ scores are around 50, with a maximum of 56, substantially lower than typical companies. Sample 2’s customer attrition rate reaches 84.7%, indicating dissatisfaction with pricing and service quality.

When comparing the risk evaluation results of this paper with the provincial power trading center’s risk management list, all high-risk samples identified by the model are included in the risk management list. This finding corroborates the reliability of the proposed model.

6.4. Analysis of Risk Warning Results

Clustered sample data are divided into positive and negative classes, with low- and medium-risk electricity sales companies as positive samples and high-risk companies as negative samples. The clustered samples are split, with 70% as the training set and 30% as the test set. The accuracy, recall, precision, and F1 score for the risk prediction model are calculated. Comparing the stacking model of this paper with five single classification models, SVM, RF, KNN, XGBOOST, and GBDT, the recognition results of each model are shown in Figure 16. From the figure, it can be observed that stacking is higher than the other single classification models in all the metrics except precision, and compared with the GBDT model, the accuracy of stacking is improved by 9.8%, recall by 6.0%, precision by 2.2%, and F1 score by 4.1, which is a significant improvement compared with the single classification models.

Additionally,

Table 2. Stacking model confusion matrix.

Real Value	Identifying Value
	Positive Sample	Negative Sample
Positive sample	330	6
Negative sample	2	28

shows the confusion matrix, with 336 positive samples and 30 negative samples. The stacking model identified 28 negative samples (high-risk electricity sales companies), yielding a false detection rate of only 6.67% for negative samples.

The stacking model’s prediction results are analyzed using two principal component indicators, as shown in Figure 17.

Since the warning model focuses on managing high-risk companies, its ability to identify negative samples requires particular attention. As shown in the figure, the first and second principal components of two misclassified samples are very similar to those of normal samples, leading to their misclassification as positive by the stacking model. Deploying the warning model enables additional risk data collection from electricity sales companies, and ongoing model training improves risk prediction accuracy and minimizes errors.

To demonstrate this warning model’s advantages, this paper compares it to multivariate and neural network warning models. As shown in

Table 3. Comparison of risk warning models.

Methods	No Data Labels Required	Real Time	Self-Update
Z-score warning method	√	×	×
Neural networks warning methods	×	√	√
Warning method of this paper	√	√	√

, the proposed method provides accurate risk warnings without requiring risk data labels. It requires minimal prior knowledge and can self-update on the basis of transaction data, achieving real-time risk warnings. This method meets practical requirements better than the other two models.

7. Conclusions

This paper proposes a data-driven risk warning method covering the entire business process of electricity sales companies. The method automatically identifies risks and provides dynamic warnings using regulatory data from the business process. Data analysis from a provincial trading center system indicates that:

(1)

PCA facilitates extracting key factors representing risk features from business data, reducing the amount of data for management and processing.

(2)

The proposed DPC-SSA clustering algorithm addresses cutoff distance issues and manual cluster center determination in DPC, providing effective clustering results for risk data.

(3)

Using principal component scores for adaptive risk level classification reduces the impact of subjective human factors, improving classification accuracy.

(4)

The constructed stacking model, applied to risk warning for electricity sales companies, exhibits high warning accuracy and enables dynamic warning and tracking of high-risk companies, allowing them to participate in the electricity market in a sustainable manner.

(5)

This risk warning method does not require risk data labels, can self-update on the basis of transaction data, and provides real-time, effective warnings for high-risk electricity sales companies.

Future research directions of this study may include enhancing the accuracy of the early warning model, such as investing a larger amount of raw data for training, which can be used to solve the problem of identifying some samples incorrectly. Additionally, a distributed computing framework can be employed to assign training tasks to multiple computing nodes for parallel processing, which would accelerate the training process and reduce operational costs. Furthermore, sustainable risk warning for electricity sales companies in the spot electricity market could be a future research direction.