Power supply service risk identification system based on big data analysis

Abstract

The invention proposes a power supply service risk identification system based on big data analysis, and relates to the technical field of power supply services, and the system comprises a data collection module which is used for collecting multi-source data related to a power supply service, and the multi-source data comprises work order data, feedback data and public opinion data; the data processing module is used for preprocessing the multi-source data to obtain standardized data; the data fusion module is used for fusing the standardized data to obtain fused data; the risk identification module is used for analyzing the fusion data, identifying the risk type in the power supply service by establishing a power supply service risk assessment index set, and determining the risk level; and the early warning output module is used for generating risk early warning information according to the risk identification result, and the risk early warning information comprises a risk type, a risk level, an influence range and a disposal suggestion. According to the invention, comprehensive identification, accurate evaluation and timely early warning of the power supply service risk can be realized, so that the power supply service risk management level is effectively improved.

Description

Technical Field

[0001] The present invention relates to the technical field of power supply services, and in particular to a power supply service risk identification system based on big data analysis. Background Art

[0002] With the rapid development of the economy and society, power supply has become the basic guarantee for the operation of modern society. As an important function of power companies, power supply services are directly related to the daily lives of thousands of households and the normal operation of enterprises. Power supply services involve multiple links such as power application, fault repair, and complaint handling. The service process is complex and has many influencing factors, which puts forward high requirements for service quality management.

[0003] Identification of power supply service risks is an important means to ensure the quality of power supply services. At present, power supply companies mainly collect work order data, customer feedback and other information, and use manual analysis and simple statistical methods to identify service risks. This risk identification mechanism is of great significance for preventing service quality problems and improving service levels.

[0004] However, the existing power supply service risk identification technology has some shortcomings: first, the data source is single, which makes it difficult to fully reflect the service quality; second, the data processing method is simple and fails to fully tap the value of the data; third, the risk assessment indicator system is not perfect, lacking a multi-dimensional evaluation of service quality; finally, the risk warning mechanism is relatively lagging, making it difficult to timely discover and deal with potential risks. These problems have seriously restricted the effectiveness of power supply service risk management, and it is urgent to explore new risk identification methods. Summary of the invention

[0005] In view of this, the present invention proposes a power supply service risk identification system based on big data analysis. By collecting and analyzing multi-source data such as work order data, feedback data and public opinion data, it can realize comprehensive identification, accurate assessment and timely warning of power supply service risks, thereby effectively improving the level of power supply service risk management and providing decision-making support for improving the quality of power supply services.

[0006] The technical solution of the present invention is implemented as follows: The present invention provides a power supply service risk identification system based on big data analysis, comprising:

[0007] The data collection module is used to collect multi-source data related to power supply services, including work order data, feedback data and public opinion data;

[0008] A data processing module is used to pre-process multi-source data to obtain standardized data;

[0009] A data fusion module is used to fuse the standardized data to obtain fused data;

[0010] The risk identification module is used to analyze the fused data, identify the risk types in the power supply service, and determine the risk level by establishing a power supply service risk assessment indicator set;

[0011] The early warning output module is used to generate risk early warning information based on the risk identification results. The risk early warning information includes: risk type, risk level, impact scope and disposal suggestions.

[0012] On the basis of the above technical solution, preferably, the data processing module includes:

[0013] The data cleaning unit is used to detect and process outliers in multi-source data, identify outlier data points using an outlier detection algorithm based on local outlier factors, and correct outliers using a sliding window median filling method;

[0014] The data standardization unit is used to standardize the cleaned data. It adopts an adaptive standardization method based on data distribution characteristics to convert data of different dimensions into a unified numerical range to obtain standardized data.

[0015] Based on the above technical solution, preferably, in the data standardization unit, the process of the adaptive standardization method is as follows:

[0016] For the data sequence Y, calculate its distribution characteristic parameters α and β:

[0017]

[0018] β=median(I)+σ×skew(I)

[0019] Where Q3 is the upper quartile, Q1 is the lower quartile, m is the sample size, median(Y) is the median, σ is the standard deviation, and skew(Y) is the skewness.

[0020] Standardize the data sequence Y according to the distribution characteristic parameters α and β:

[0021]

[0022] Where ε is a smoothing factor, which is used to prevent the denominator from being zero.

[0023] On the basis of the above technical solution, preferably, in the data fusion module, a data registration algorithm is used to perform spatiotemporal alignment and feature fusion on multi-source data to achieve spatiotemporal consistency alignment of multi-source data. The calculation formula of the data registration algorithm is as follows:

[0024]

[0025]

[0026] Where A(p,t) represents the alignment result of position p at time t; D i (p,t) represents the observation value of the ith data source at location p and time t; d(p,p i ) represents the reference position p of the data source i i The spatial distance between i | represents time t and sampling time t of data source i i The time difference between i represents the weight coefficient of the i-th data source; Q i represents the spatiotemporal resolution coefficient of data source i; Δs i represents the spatial resolution of data source i; Δt i represents the temporal resolution of data source i; a1, a2, and γ are adjustment parameters used to control spatial attenuation, temporal attenuation, and resolution effects.

[0027] Based on the above technical solution, preferably, the risk identification module includes:

[0028] A feature extraction unit, used for extracting service quality features from the fused data;

[0029] A risk assessment unit is used to construct a risk assessment index system and calculate the power supply service risk assessment index value based on service quality characteristics;

[0030] The risk classification unit is used to establish a dynamic threshold model based on historical data statistics, determine the risk level according to the power supply service risk assessment index value and the dynamic threshold model, determine the risk type based on the feature combination pattern, and output the risk identification result.

[0031] Based on the above technical solution, preferably, the execution process of the feature extraction unit is as follows:

[0032] Extract time series features from work order data:

[0033] The work order data is segmented through the sliding time window W(t,Δt), and the window length Δt is adaptively adjusted according to the business characteristics: Δt=max(μ d ·(1+σ d ),Δt min ), where μ d is the average processing time of historical work orders, σ d is the standard deviation, Δt min is the lower limit of the window length;

[0034] Construct an LSTM network, including an attention-enhanced memory unit, control the information flow through the forget gate, input gate, and output gate, and extract the temporal feature f τ =[fτ1 ,f τ2 ,f τ3 ], where f τ1 is the efficiency trend characteristic, through f τ2 is the traffic fluctuation characteristic, f τ3 is the temporal pattern feature;

[0035] Extract semantic features from feedback data:

[0036] Construct a multi-level attention mechanism including word level, sentence level and document level:

[0037] α w =softmax(v w tanh(W w ·h w ))

[0038] α s =softmax(v s tanh(W s ·h s ))

[0039] α d =softmax(v d tanh(W d ·h d ))

[0040] In the formula, α w is the word-level attention weight, with a value range of [0,1], and the sum in each sentence is 1; α s is the sentence-level attention weight, with a value range of [0,1] and a sum of 1 within a document; α d is the document-level attention weight, with a value range of [0,1] and a sum of 1; h w is the word-level hidden layer representation, and its dimension is the word embedding dimension; h s is the sentence-level hidden layer representation, and its dimension is the sentence encoding dimension; h d is the document-level hidden layer representation, and its dimension is the document encoding dimension; W w , W s , W d is the attention weight matrix of each layer; v w 、v s 、v d is the attention vector of each layer;

[0041] By calculating the attention weights of each layer to identify key information, the semantic features f are extracted s =[f s1 ,f s2 ,f s3 ], where f s1 is the topic distribution feature, f s2is the keyword feature, f s3 for emotional characteristics;

[0042] Extract features from public opinion data:

[0043] Construct multimodal features of public opinion texts, including: extracting text features using a multi-level attention mechanism, extracting propagation network features through a graph convolutional network (GCN), analyzing public opinion evolution using a time-decayed attention mechanism, and extracting temporal features;

[0044] Calculate and analyze multimodal features to extract public opinion features f m =[f m1 ,f m2 ,f m3 ], where f m1 is the public opinion influence characteristic, f m2 is the evolution characteristic of public opinion, f m3 is the public opinion clustering feature;

[0045] Interactive feature extraction based on multi-source data:

[0046] Construct a data association graph G(V,E), where the node set V contains work order data, feedback data, and public opinion data, and the edge set E represents the association between entities;

[0047] The weighted summation is performed based on content similarity, temporal correlation and spatial correlation to calculate the correlation strength between nodes and construct a correlation strength matrix;

[0048] Extracting interactive features f based on association strength matrix and data association graph c =[f c1 ,f c2 ,f c3 ], where f c1 is the relationship feature, f c2 is the propagation characteristic, f c3 is the associated feature;

[0049] For the time series feature f τ , semantic features f s 、Public opinion characteristics m , interaction feature f c After normalization, the fusion weight is determined through the attention mechanism to obtain the final quality of service feature F.

[0050] On the basis of the above technical scheme, preferably, the risk assessment index system includes two layers of indicators. The first layer indicators include service response indicators, service quality indicators, user satisfaction indicators and social influence indicators. The second layer indicators are the next level indicators set corresponding to the first layer indicators. The service response indicators include average response time and first response timeliness rate. The service quality indicators include problem solving rate and service standardization. The user satisfaction indicators include complaint rate and bad review rate. The social influence indicators include the spread of public opinion and the degree of negative impact.

[0051] Based on the above technical solution, preferably, the process of calculating the power supply service risk assessment index value based on the service quality characteristics is as follows:

[0052] Construct a feature mapping function to calculate the second-layer index value vector I:

[0053] I = σ(W·F+b)·(1+tanh(F T ·M·F))

[0054] Among them, σ is the sigmoid function; F is the service quality feature; W is the feature mapping matrix; M is the feature interaction matrix; b is the bias vector;

[0055] The power supply service risk assessment index value R is calculated based on the second-layer index value vector I:

[0056] R = ψ(I)·(1+η·‖ΔI‖)

[0057] Where ψ(I) is a nonlinear combination function used to fuse the index values ​​of each dimension, η is a time-varying parameter, ΔI is the change in the index value, and ‖·‖ represents the norm operation.

[0058] Based on the above technical solution, preferably, in the risk classification unit, the dynamic threshold model is established by the following steps:

[0059] Based on historical data, a time series feature sequence is constructed, and the long short-term memory network LSTM is used to extract the time-dependent feature z(t);

[0060] Calculate the adaptive factor λ(t):

[0061]

[0062] In the formula, u1 and u2 are adjustment parameters;

[0063] Calculate dynamic threshold boundaries T1(t) and T2(t), T1(t) <T2(t):

[0064] T1(t)=μ his -λ(t)·σ his

[0065] T2(t) = μ his + λ(t)·σ his

[0066] Wherein, μ his is the mean value of the historical risk assessment index value, and σ his is the standard deviation of the historical risk assessment index value;

[0067] Determine the risk level according to the power supply service risk assessment index value R:

[0068] When R < T1(t), it is determined as low risk;

[0069] When T1(t) ≤ R < T2(t), it is determined as medium risk;

[0070] When R ≥ T2(t), it is determined as high risk.

[0071] On the basis of the above technical solution, preferably, in the risk classification unit, the feature combination mode represents the feature distribution law corresponding to different risk types, including the combination relationship and time series evolution characteristics of each dimension feature in the service quality feature F;

[0072] For each moment t in the historical data, construct a feature pattern matrix M(t):

[0073] M(t) = [F(t), ΔF(t), δF(t)]

[0074] Wherein, F(t) is the historical service quality feature at moment t, ΔF(t) is the change amount of the feature, and δF(t) is the change rate of the feature;

[0075] Based on the feature pattern matrix sequence {M(t)} constructed from historical data, use the density clustering algorithm for analysis, cluster the patterns with similar feature distributions and evolution laws into one category, and each category corresponds to a specific risk type, thus forming a risk feature pattern library;

[0076] Construct a feature pattern matrix M(t0) for the current moment t0, and determine the specific risk type by calculating the similarity between M(t0) and each type of pattern in the risk feature pattern library. The similarity calculation uses a non-linear measurement method considering feature importance.

[0077] The present invention has the following beneficial effects compared with the prior art:

[0078] (1) By constructing a power supply service risk identification system based on big data analysis, the present invention realizes the comprehensive analysis of multi-source data of power supply services and the intelligent identification of risks. Compared with the prior art, it can more comprehensively and accurately discover potential risks in power supply services, and timely give risk warnings and disposal suggestions, effectively improving the intelligent level of power supply service risk management;

[0079] (2) The data processing module of the present invention can effectively identify and correct outliers in multi-source data by introducing an anomaly detection algorithm based on local anomaly factors and a sliding window median filling method to ensure the integrity and accuracy of the data. At the same time, an adaptive standardization method based on data distribution characteristics is used to convert data of different dimensions into a unified numerical interval to ensure the standardization and consistency of data processing results;

[0080] (3) The present invention realizes the spatiotemporal consistency alignment of multi-source data through a data registration algorithm, taking into account the influence of factors such as spatial distance, time difference and data source resolution, solves the spatiotemporal inconsistency problem in multi-source heterogeneous data fusion, and improves the accuracy of data fusion;

[0081] (4) The present invention extracts the time series features of work order data through a sliding time window, extracts the semantic features of feedback data through a multi-level attention mechanism, and extracts the multimodal features of public opinion data through a combination of a graph convolutional network and a time-decayed attention mechanism, thereby being able to extract key feature information from different data sources. At the same time, the interactive feature extraction method based on the data association graph further explores the correlation characteristics between multi-source data, and the service quality features finally generated have higher expressiveness and discrimination;

[0082] (5) The present invention adopts a two-layer risk assessment indicator system and a nonlinear combination assessment method, combined with feature mapping functions and interaction matrices, to achieve a multi-dimensional quantitative assessment of power supply service risks. At the same time, the risk classification module is based on a dynamic threshold model and a feature combination pattern, combined with the time series characteristics of historical data and a density clustering algorithm, which can accurately determine the risk type and risk level, thereby achieving accurate identification and classification of power supply service risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0083] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0084] Figure 1 4 is a system framework diagram of an embodiment of the present invention. DETAILED DESCRIPTION

[0085] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0086] like Figure 1 As shown, the present invention provides a power supply service risk identification system based on big data analysis, including:

[0087] The data collection module is used to collect multi-source data related to power supply services, including work order data, feedback data and public opinion data;

[0088] A data processing module is used to pre-process multi-source data to obtain standardized data;

[0089] A data fusion module is used to fuse the standardized data to obtain fused data;

[0090] The risk identification module is used to analyze the fused data, identify the risk types in the power supply service, and determine the risk level by establishing a power supply service risk assessment indicator set;

[0091] The early warning output module is used to generate risk early warning information based on the risk identification results. The risk early warning information includes: risk type, risk level, impact scope and disposal suggestions.

[0092] Specifically, in one embodiment of the present invention, the data collection module is responsible for collecting multi-source data related to power supply services, including work order data, feedback data and public opinion data. In specific implementation, the data collection module obtains relevant data information in real time by connecting with the power supply service management system.

[0093] Among them, the work order data comes from the business processing system in the power supply service management system. It includes types such as power installation work orders, fault repair work orders, and business consultation work orders. Each work order data record contains fields such as work order number, user information, work order type, work order content, acceptance time, response time, processing process, completion time, and processing results. The data acquisition module calls the data interface of the business processing system to synchronize the latest work order information regularly and track the work order status in real time.

[0094] Feedback data includes user evaluation data and complaint data. User evaluation data comes from the service evaluation system, including service rating, evaluation content, evaluation time and other information; complaint data comes from the complaint handling system, including complaint content, complaint type, complaint time, handling process, handling results and other information. The data collection module collects user evaluation and complaint information in real time through the data interface of the service evaluation system and the complaint handling system.

[0095] Public opinion data comes from external channels such as social media platforms, news websites, and government complaint platforms. It includes news reports, social media discussions, online complaints, and other information related to power supply services. Each piece of public opinion data contains fields such as information content, release time, release platform, dissemination range, and interactive data. The data collection module continuously monitors and collects relevant public opinion information through web crawler technology and third-party data interfaces.

[0096] During the data collection process, the data collection module also needs to ensure the timeliness and integrity of the data. For work order data and feedback data, a quasi-real-time collection method is used, and a fixed collection cycle, such as 5 minutes, is set to synchronize data regularly; for public opinion data, a real-time monitoring method is used, and relevant information is collected immediately once it is found. At the same time, a preliminary integrity check is performed on the collected data to ensure the existence of necessary fields and the correctness of the basic format.

[0097] In addition, the data collection module also needs to uniformly identify and timestamp the collected data. Each piece of data is given a unique identifier and records metadata such as the data collection time and data source.

[0098] Specifically, in one embodiment of the present invention, the data processing module includes:

[0099] The data cleaning unit is used to detect and process outliers in multi-source data, identify outlier data points using an outlier detection algorithm based on local outlier factors, and correct outliers using a sliding window median filling method;

[0100] The data standardization unit is used to standardize the cleaned data. It adopts an adaptive standardization method based on data distribution characteristics to convert data of different dimensions into a unified numerical range to obtain standardized data.

[0101] In this embodiment, the data cleaning unit uses an anomaly detection algorithm based on local anomaly factors to detect and process outliers on multi-source data.

[0102] In specific implementation, first calculate the local density between data points, and identify abnormal data points by comparing the ratio of the local density of a data point to that of its neighborhood points. For work order data, mainly detect outliers in numerical fields such as response time and processing duration; for feedback data, mainly detect outliers in rating data; for public opinion data, mainly detect outliers in statistical indicators such as dissemination volume and interaction volume. After identifying the abnormal data points, use the sliding window median filling method to correct the outliers. The specific calculation process is as follows:

[0103] Calculate the k-distance and k-distance neighborhood:

[0104] k-dist(p) = d(p,o)

[0105] where p is the point to be detected, o is the k-th nearest neighbor of p, d(p,o) is the Euclidean distance between the two points, and k is the preset neighborhood parameter.

[0106] Calculate the reachability distance:

[0107] reach-dist(p,o) = max{k-dist(o), d(p,o)}

[0108] where reach-dist(p,o) is the reachability distance of point p relative to point o, k-dist(o) is the k-distance of point o, and d(p,o) is the Euclidean distance between point p and point o.

[0109] Calculate the local reachability density:

[0110] lrd(p) = 1 / (∑reach-dist(p,o) / |N k (p)|)

[0111] where lrd(p) is the local reachability density of point p, N k (p) is the set of points in the k-distance neighborhood of p, and |N k (p)| is the number of neighborhood points.

[0112] Calculate the local outlier factor:

[0113] LOF(p) = (∑(lrd(o) / lrd(p))) / |N k (p)|

[0114] where LOF(p) is the local outlier factor of point p, lrd(o) is the local reachability density of neighborhood point o, and lrd(p) is the local reachability density of point p. Abnormality determination threshold: when 1 < LOF(p) < 1.5, it is determined as a weak outlier; when LOF(p) ≥ 1.5, it is determined as a strong outlier; when LOF(p) ≈ 1, it is determined as a normal point.

[0115] Sliding window median filling:

[0116] For time series data Y(t), within the time window [tw,t+w]:

[0117] Y′(t)=median({Y(i)|(tw)≤i≤(t+w)})

[0118] Where w is the half-width of the window, Y(i) represents the data value of the time series at time i, {Y(i)|(tw)≤i≤(t+w)} represents the set of all data points in the time window, Y′(t) is the data value after outlier correction, and median() is the function for calculating the median.

[0119] This method sets a fixed-size time window and calculates the median of the data in the window to replace outliers. This method can maintain the timing characteristics of the data and is highly robust to outliers.

[0120] In this embodiment, the process of the adaptive normalization method is as follows:

[0121] For the data sequence Y, calculate its distribution characteristic parameters α and β:

[0122]

[0123] β=median(I)+σ×skew(I)

[0124] Where Q3 is the upper quartile, Q1 is the lower quartile, m is the sample size, median(Y) is the median, σ is the standard deviation, and skew(Y) is the skewness.

[0125] Standardize the data sequence Y according to the distribution characteristic parameters α and β:

[0126]

[0127] Where ε is a smoothing factor, which is used to prevent the denominator from being zero.

[0128] This embodiment adopts an adaptive normalization method based on data distribution characteristics to convert data of different dimensions into a unified numerical range. This method can automatically adjust the normalization parameters according to the distribution characteristics of the data, adapt to the distribution characteristics of different types of data, and ensure the rationality of the normalization results.

[0129] In the specific implementation process, the data processing module also needs to consider the characteristics of different types of data. For time data in work order data, such as response time and processing time, relative time difference is used for standardization; for rating data in feedback data, rating interval mapping is used for standardization; for statistical indicators in public opinion data, logarithmic transformation is used before standardization. Through this classification processing method, it is ensured that different types of data are comparable after standardization.

[0130] Specifically, in one embodiment of the present invention, the data fusion module uses a data registration algorithm to perform spatiotemporal alignment and feature fusion on multi-source data to achieve spatiotemporal consistency alignment of multi-source data. The specific implementation process of the data registration algorithm is as follows:

[0131] First, the algorithm achieves spatiotemporal alignment of multi-source data by calculating the position and time alignment results using the formula A(p,t). The formula is:

[0132]

[0133] Where A(p,t) represents the alignment result of position p at time t; D i (p, t) represents the observation value of the i-th data source at position p and time t, which is the basic input of the original data; d(p, p i ) represents the reference position p of the data source i i The spatial distance between them is used to measure the spatial correlation; |tt i | represents time t and sampling time t of data source i i The time difference between them is used to measure the time correlation; ω i Represents the weight coefficient of the i-th data source, which is used to adjust the importance of different data sources; is the spatial distance attenuation term, which decays exponentially as the spatial distance increases; is the time distance attenuation term, which decays exponentially as the time difference increases; a1 and a2 are adjustment parameters used to control the degree of spatial attenuation and time attenuation.

[0134] This embodiment introduces time dimension and space dimension attenuation terms to achieve time and space alignment of data. For example, a substation failure may affect multiple surrounding communities, and the complaint locations may be distributed in different locations in the affected area. The algorithm uses the spatial attenuation function to reasonably establish the correlation between data at different locations. For example, if a power supply failure occurs in a certain area, the work order system will record it immediately, but related complaints and public opinions may appear in the future. The algorithm uses the time decay function to reasonably associate these data with different times.

[0135] Secondly, the algorithm introduces the spatiotemporal resolution coefficient Qi To adjust the impact of different data sources. The calculation formula of the spatiotemporal resolution coefficient is:

[0136]

[0137] In the formula, Δs i Indicates the spatial resolution of data source i, reflecting the degree of detail of the data in the spatial dimension; Δt i It represents the time resolution of data source i, reflecting the sampling frequency of data in the time dimension; γ is an adjustment parameter used to control the influence of resolution on the final result.

[0138] The spatiotemporal resolution of the data source referred to in this embodiment is an indicator that describes the precision of the data source in the temporal and spatial dimensions. i Indicates the time interval for data collection or update; Δs i Indicates the spatial positioning accuracy of the data.

[0139] The data registration algorithm provided in this embodiment performs weighted fusion of multi-source data by considering spatial distance decay, temporal distance decay and spatiotemporal resolution of data sources. The spatial distance decay term and the temporal distance decay term ensure that data with closer distance and smaller time difference have greater weight. The spatiotemporal resolution coefficient further adjusts the influence of different data sources, so that data sources with higher resolution play a greater role in the fusion process.

[0140] In this way, the algorithm achieves the spatiotemporal consistency alignment of multi-source data, unifying work order data, feedback data, and public opinion data into the same spatiotemporal reference system. This alignment method not only takes into account the spatial location and timestamp of the data, but also the characteristics of the data source (such as resolution), thereby ensuring the accuracy and reliability of the fusion results. The fused data retains the important information of each data source while achieving a unified expression of the data.

[0141] Specifically, in one embodiment of the present invention, the risk identification module includes:

[0142] A feature extraction unit, used for extracting service quality features from the fused data;

[0143] A risk assessment unit is used to construct a risk assessment index system and calculate the power supply service risk assessment index value based on service quality characteristics;

[0144] The risk classification unit is used to establish a dynamic threshold model based on historical data statistics, determine the risk level according to the power supply service risk assessment index value and the dynamic threshold model, determine the risk type based on the feature combination pattern, and output the risk identification result.

[0145] In this embodiment, the feature extraction unit is responsible for extracting service quality features from the fused data. The specific implementation method includes four main parts: time series feature extraction of work order data, semantic feature extraction of feedback data, feature extraction of public opinion data, and interactive feature extraction of multi-source data.

[0146] Extract time series features from work order data:

[0147] The work order data is segmented through the sliding time window W(t,Δt), and the window length Δt is adaptively adjusted according to the business characteristics: Δt=max(μ d ·(1+σ d ),Δt min ), where μ d is the average processing time of historical work orders, σ d is the standard deviation, Δt min is the lower limit of the window length;

[0148] Construct an LSTM network with attention-enhanced memory cells:

[0149] c t =f t ⊙c t-1 +i t ⊙tanh(W c ·[h t-1 ,x t ]+b c )

[0150] h t =o t ⊙tanh(c t )

[0151] a t =softmax(W a tanh(W h ·h t ))

[0152] Among them, c t is the memory unit state at the current moment, and its dimension is the hidden layer size; c t-1 is the memory unit state at the previous moment; h t is the hidden layer output at the current moment, with the same dimension as c t ;h t-1 is the hidden layer output of the previous moment; x t is the input vector at the current moment; f t is the output of the forget gate, with a value range of [0,1], controlling the retention of historical information; i t is the input gate output, with a value range of [0,1], which controls the reception of new information; o tis the output gate output, value range [0,1], control information output; W c is the memory unit weight matrix; b c is the memory unit bias vector; a t is the time step attention weight; W a is the attention weight matrix; W h is the hidden layer transformation matrix.

[0153] The information flow is controlled by the forget gate, input gate and output gate to extract the temporal feature f τ =[f τ1 ,f τ2 ,f τ3 ].

[0154] Among them, f τ1 To obtain the efficiency trend feature, the LSTM network is used to analyze the work order processing time sequence. The calculation method is as follows: first, the work order is segmented according to the time window W(t,Δt), and the average processing time in each window is calculated. The LSTM network is used to extract the change trend of the processing efficiency, and the trend feature vector is output to reflect the overall trend of service efficiency. τ2 To characterize the fluctuation of business volume, we analyze the time distribution characteristics of the number of work orders: we count the number of work orders in each time window, calculate the fluctuation amplitude between adjacent windows, extract the periodic fluctuation pattern, and output the fluctuation feature vector to reflect the change law of business volume. τ3 For the timing pattern features, the timing pattern of work order processing is identified: the work order state transition sequence is extracted, the time characteristics of the processing flow are analyzed, the exception processing pattern is identified, and the pattern feature vector is output to reflect the standardization of the service process.

[0155] Extract semantic features from feedback data:

[0156] Construct a multi-level attention mechanism including word level, sentence level and document level:

[0157] α w =softmax(v w tanh(W w ·h w ))

[0158] α s =softmax(v s tanh(W s ·h s ))

[0159] α d =softmax(v d tanh(W d ·h d ))

[0160] In the formula, α w is the word-level attention weight, with a value range of [0,1], and the sum in each sentence is 1; α s is the sentence-level attention weight, with a value range of [0,1] and a sum of 1 within a document; α d is the document-level attention weight, with a value range of [0,1] and a sum of 1; h w is the word-level hidden layer representation, and its dimension is the word embedding dimension; h s is the sentence-level hidden layer representation, and its dimension is the sentence encoding dimension; h d is the document-level hidden layer representation, and its dimension is the document encoding dimension; W w , W s , W d is the attention weight matrix of each layer; v w 、v s 、v d is the attention vector of each layer;

[0161] By calculating the attention weights of each layer to identify key information, the semantic features f are extracted s =[f s1 ,f s2 ,f s3 ], where f s1 is the topic distribution feature, f s2 is the keyword feature, f s3 is the emotional feature; specifically, f s1 Extraction based on multi-level attention mechanism: word-level attention α w Identify keywords, sentence-level attention α s Extract topic sentence, document level attention α d Determine the topic weight and output the topic distribution vector to reflect the topic composition of the feedback content; s2 Extract through word-level attention mechanism: calculate word importance score, extract high-frequency keywords, analyze word co-occurrence relationship, output keyword feature vector, and reflect the core content of feedback; s3 Sentiment analysis based on multi-layer attention: word-level sentiment polarity recognition, sentence-level sentiment intensity calculation, document-level sentiment tendency judgment, and output of sentiment feature vectors to reflect the user's emotional attitude.

[0162] Extract features from public opinion data:

[0163] Constructing the multimodal features of public opinion texts includes: extracting text features using a multi-level attention mechanism, extracting propagation network features through a graph convolutional network (GCN), analyzing the evolution of public opinion using a time-decayed attention mechanism, and extracting temporal features; specifically, the multi-level attention mechanism reuses the multi-level attention mechanism constructed when extracting semantic features; when extracting propagation network features, a propagation network graph can be constructed, with users participating in the propagation as nodes, the propagation relationships between users as edges, and the connection relationships between nodes to construct an adjacency matrix, thereby forming a propagation network graph, and then using GCN to extract propagation network features; the temporal attention calculation is as follows:

[0164] α t =softmax(λ(t)·v t tanh(W t ·z t ))

[0165]

[0166] Where λ(t) is the time decay function, v t is the temporal attention vector, W t is the time series weight matrix, z t is the temporal hidden state, t0 is the current time, t is the historical time point, T is the time window length, and θ is the attenuation coefficient.

[0167] Calculate and analyze multimodal features to extract public opinion features f m =[f m1 ,f m2 ,f m3 ], where f m1 =[Ic, Ip, Ic·Ip] is the public opinion influence feature, Ic is the content influence, which is calculated by weighted sum of text sentiment intensity, text length normalization value, and content importance score, and Ip is the communication influence, which is calculated by weighted sum of communication scope, communication frequency, and communication depth; f m2 =[Ps, Vs, Ts] is the evolution characteristic of public opinion, Ps = [p1, p2, p3, p4] is the stage characteristic, which represents the characteristic vector of public opinion in the four stages of germination, diffusion, peak and decline, Vs = dN / dt·τ(t) is the speed characteristic, dN / dt is the speed of public opinion propagation, τ(t) is the time attenuation factor, Ts = ∑(a t ·his t ) is the trend characteristic, a t is the temporal attention weight, his t is the historical state vector; f m3=[Gs,Cs,Ds] is the public opinion clustering feature, Gs is the structural feature, which is extracted through the last layer output of GCN, Cs is the clustering feature, and Gs is calculated using the spectral clustering method, Ds = |Ec| / (|Vc|·(|Vc|-1)) is the density feature, |Ec| is the number of subgraph edges, |Vc| is the number of subgraph nodes, where a subgraph refers to a local network structure in the communication network graph, which represents a group of interrelated users and the communication relationship between them under a certain public opinion topic or event.

[0168] Interactive feature extraction based on multi-source data:

[0169] Construct a data association graph G(V,E), where the node set V contains work order data, feedback data, and public opinion data, and the edge set E represents the association between entities;

[0170] The weighted summation is performed based on content similarity, temporal correlation and spatial correlation to calculate the correlation strength between nodes and construct a correlation strength matrix;

[0171] Extracting interactive features f based on association strength matrix and data association graph c =[f c1 ,f c2 ,f c3 ], where f c1 is the relationship feature, f c2 is the propagation characteristic, f c3 is the associated feature;

[0172] In a specific example, the node set V contains three types of data nodes: the work order data node represents the specific information of the work order (such as processing time, status, etc.). The feedback data node represents the content of user feedback (such as evaluation, complaint, etc.). The public opinion data node represents the characteristics of the public opinion text or propagation event. Each node is represented by its feature vector, for example: the work order node is represented by the time series feature f τ , the feedback node uses the semantic feature f s , the public opinion node uses the public opinion feature f m The edge represents the relationship between nodes, and the edge weight is determined by the value in the association strength matrix.

[0173] The element w of the correlation strength matrix W ij Represents node v i and node v j The correlation strength between them is calculated as:

[0174] w ij =α·sim(c i ,c j )+β·exp(-|t i -t j | / τ t)+γ·exp(-d(p i ,p j ) / τ s )

[0175] Where: sim(c i ,c j ) is the node v i and v j Content similarity, |t i -t j | is the time difference of the nodes, d(p i ,p j ) is the spatial distance of the nodes, τ t ,τ s are the time and space attenuation coefficients, and α, β, γ are the weight coefficients.

[0176] The association strength matrix W is a symmetric matrix that represents the association strength between all nodes.

[0177] Based on the data association graph G(V,E) and the association strength matrix W, extract the interaction feature f c =[f c1 ,f c2 ,f c3 ], the specific steps are:

[0178] Relationship feature f c1 Reflects the network structure characteristics between nodes. The extraction method is as follows:

[0179] Calculate the degree centrality of a node:

[0180]

[0181] It represents the node v i The sum of the direct connection strengths with other nodes.

[0182] Calculate the clustering coefficient of a node:

[0183]

[0184] Among them, L i For node v i The number of edges that actually exist between neighbors, k i For node v i Construct the relationship feature f c1 =[degree(v i ),C(v i )].

[0185] Propagation characteristics f c2 Reflects the ability of information to spread in the network. The extraction method is as follows:

[0186] Calculate the propagation efficiency:

[0187]

[0188] Where, d ij For node v i To node v j The shortest path length is N, where N is the total number of nodes in the network. Calculate the propagation range:

[0189]

[0190] In the formula, |V r | is the number of reachable nodes, and |V| is the total number of nodes in the network.

[0191] Calculate the propagation depth:

[0192]

[0193] Where SPL is the shortest path length.

[0194] Construct the propagation feature f c2 =[Pe,Ps,Pd].

[0195] Related feature f c3 Reflect the multi-dimensional correlation between nodes. The extraction method is as follows:

[0196] Calculate content relevance:

[0197]

[0198] Calculate the temporal correlation:

[0199]

[0200] Calculate spatial association:

[0201]

[0202] Construct the associated feature f c3 =[Rc,Rt,Rs].

[0203] For the time series feature f τ , semantic features f s 、Public opinion characteristics m , interaction feature f c After normalization, the fusion weight is determined through the attention mechanism to obtain the final quality of service feature F.

[0204] Specifically, in this embodiment, the risk assessment index system includes two layers of indicators. The first layer of indicators includes service response indicators, service quality indicators, user satisfaction indicators and social influence indicators. The second layer of indicators are the next level indicators corresponding to the first layer of indicators. The service response indicators include average response time and first response timeliness rate. The service quality indicators include problem solving rate and service standardization. User satisfaction indicators include complaint rate and negative review rate. Social influence indicators include the spread of public opinion and the degree of negative impact.

[0205] The process of calculating the power supply service risk assessment index value based on service quality characteristics is as follows:

[0206] Construct a feature mapping function to calculate the second-layer index value vector I:

[0207] I = σ(W·F+b)·(1+tanh(F T ·M·F))

[0208] Among them, σ is the sigmoid function; F is the service quality feature; W is the feature mapping matrix; M is the feature interaction matrix; b is the bias vector; Specifically, the feature mapping matrix W is an m×n matrix, m is the number of second-layer indicators, and n is the service quality feature dimension. The acquisition method is: obtained through historical data training (supervised learning), initialized using expert experience, and then optimized through data iteration. The feature interaction matrix M is an n×n symmetric matrix, which is formed by calculating the mutual information between feature pairs.

[0209] This embodiment realizes nonlinear mapping from feature space to index space, takes into account the interaction between features, ensures the range constraint of index values, and can adapt to situations with different feature combinations.

[0210] The power supply service risk assessment index value R is calculated based on the second-layer index value vector I:

[0211] R = ψ(I)·(1+η·‖ΔI‖)

[0212] Where ψ(I) is a nonlinear combination function used to fuse the index values ​​of each dimension, η is a time-varying parameter, ΔI is the change in the index value, and ‖·‖ represents the norm operation.

[0213] In this embodiment, ψ(I)=∑(w i I i )+∑(w ij I i I j ), w i is the first-order weight, reflecting the importance of a single indicator, w ij is the second-order weight, reflecting the importance of interaction between indicators, I i ,I jIt is the second-layer index value.

[0214] Specifically, in an embodiment of the present invention, the dynamic threshold model is established through the following steps:

[0215] Construct a time series feature sequence based on historical data, and use the long short-term memory network LSTM to extract the time-dependent feature z(t); z(t)=LSTM(F(t),h t-1 ), where F(t) is the quality-of-service feature at time t, and h t-1 is the hidden layer state at the previous moment.

[0216] Calculate the adaptive factor λ(t):

[0217]

[0218] In the formula, u1 and u2 are adjustment parameters; ‖z(t)-z(t - 1)‖ represents the change amplitude of the features at the current moment and the previous moment.

[0219] Calculate the dynamic threshold boundaries T1(t) and T2(t), where T1(t)<T2(t):

[0220] T1(t)=μ his -λ(t)·σ his

[0221] T2(t)=μ his +λ(t)·σ his

[0222] In the formula, μ his is the mean value of the historical risk assessment index values, and σ his is the standard deviation of the historical risk assessment index values;

[0223] Determine the risk level according to the power supply service risk assessment index value R:

[0224] When R<T1(t), it is determined as a low risk;

[0225] When T1(t)≤R<T2(t), it is determined as a medium risk;

[0226] When R≥T2(t), it is determined as a high risk.

[0227] The feature combination pattern represents the feature distribution rules corresponding to different risk types, including the combination relationship and time series evolution characteristics of the features in each dimension of the quality-of-service feature F;

[0228] For each moment t in the historical data, construct a feature pattern matrix M(t):

[0229] M(t)=[F(t),ΔF(t),δF(t)]

[0230] Where F(t) is the historical service quality feature at time t, ΔF(t) = F(t) - F(t-1) is the change amount of the feature, and δF(t) = ΔF(t) / F(t-1) is the change rate of the feature;

[0231] Based on the characteristic pattern matrix sequence {M(t)} constructed based on historical data, a density clustering algorithm is used for analysis, and patterns with similar characteristic distribution and evolution rules are clustered into one category, each category corresponding to a specific risk type, thereby forming a risk characteristic pattern library; in this embodiment, the density clustering algorithm uses DBSCAN, and the similarity between characteristic pattern matrices is calculated during clustering:

[0232]

[0233] In the formula, w k is the feature importance weight, φ k (M i ,M j ) is the similarity measure of the k-th dimension feature. The feature pattern matrix is ​​clustered into several categories according to the similarity to form a risk feature pattern library.

[0234] Construct the characteristic pattern matrix M(t0) for the current time t0, and determine the specific risk type by calculating the similarity between M(t0) and each type of pattern in the risk characteristic pattern library. The similarity calculation adopts a nonlinear measurement method that considers the importance of the feature. The similarity calculation still uses the above similarity calculation formula to calculate S(M(t0),M k ), M k is the kth type of pattern in the risk feature pattern library. According to the similarity S(M(t0),M k ), select the risk type corresponding to the pattern with the highest similarity as the risk type at the current moment.

[0235] Specifically, in one embodiment of the present invention, the early warning output module receives the risk level and risk type, and forms a corresponding impact scope and disposal suggestion.

[0236] The impact range is used to describe the geographical area, user group or business scope that a risk event may affect. The calculation method of the impact range is as follows: Input data association graph and interaction feature f c , using interactive features to propagate features f c2 =[Pe,Ps,Pd], through the propagation range Ps and propagation depth Pd, determine the node set V that the risk may affect r , and combines node attributes (such as geographic location, user type, etc.) to generate the impact range.

[0237] Disposal suggestions are used to guide the response measures for risk events, including priority sorting, resource scheduling and specific operational suggestions. First, pre-define the disposal suggestion template for different risk types. For example: Service quality risk: Response time risk: It is recommended to increase the number of customer service personnel and optimize the response process; Service specification risk: It is recommended to strengthen employee training and improve service specifications. User satisfaction risk: Complaint escalation risk: It is recommended to quickly handle user complaints and provide compensation plans; User loss risk: It is recommended to carry out user return visits and provide preferential policies. Social impact risk: Public opinion dissemination risk: It is recommended to issue an official statement to clarify the facts; Negative impact risk: It is recommended to communicate with the media to control the spread of public opinion. Operation management risk: Resource scheduling risk: It is recommended to optimize resource allocation and increase spare resources; Process execution risk: It is recommended to check the execution of the process and correct problems in time. Adjust the priority of disposal suggestions according to the risk level: High risk: Immediately launch the emergency plan and mobilize more resources; Medium risk: Develop short-term response measures and closely monitor risk changes; Low risk: Regularly check and track, and do a good job of prevention. Determine the resource scheduling plan based on the scope of impact: if the scope of impact is large, it is recommended to mobilize more resources; if the scope of impact is concentrated in a certain area or user group, it is recommended to give priority to solving problems in that area or group.

[0238] Finally, the risk warning information generated by the warning output module includes the following: risk type: such as service quality risk, user satisfaction risk, etc.; risk level: low risk, medium risk or high risk; impact scope: including affected geographical areas, user groups or business scope; disposal suggestions: specific response measures generated based on risk type and level.

[0239] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. The power supply service risk identification system based on big data analysis is characterized by: include: The data collection module is used to collect multi-source data related to power supply services, including work order data, feedback data and public opinion data; A data processing module is used to pre-process multi-source data to obtain standardized data; A data fusion module is used to fuse the standardized data to obtain fused data; The risk identification module is used to analyze the fused data, identify the risk types in the power supply service, and determine the risk level by establishing a power supply service risk assessment indicator set; The early warning output module is used to generate risk early warning information based on the risk identification results. The risk early warning information includes: risk type, risk level, impact scope and disposal suggestions.

2. The power supply service risk identification system based on big data analysis according to claim 1, characterized in that: The data processing module includes: The data cleaning unit is used to detect and process outliers in multi-source data, identify outlier data points using an outlier detection algorithm based on local outlier factors, and correct outliers using a sliding window median filling method; The data standardization unit is used to standardize the cleaned data. It adopts an adaptive standardization method based on data distribution characteristics to convert data of different dimensions into a unified numerical range to obtain standardized data.

3. The power supply service risk identification system based on big data analysis according to claim 2, characterized in that: In the data standardization unit, the process of the adaptive standardization method is as follows: For the data sequence Y, calculate its distribution characteristic parameters α and β: β=median(I)+σ×skew(I) Where Q3 is the upper quartile, Q1 is the lower quartile, m is the sample size, median(Y) is the median, σ is the standard deviation, and skew(Y) is the skewness. Standardize the data sequence Y according to the distribution characteristic parameters α and β: Where ε is a smoothing factor, which is used to prevent the denominator from being zero.

4. The power supply service risk identification system based on big data analysis according to claim 1, characterized in that: In the data fusion module, the data registration algorithm is used to perform spatiotemporal alignment and feature fusion of multi-source data to achieve spatiotemporal consistency alignment of multi-source data. The calculation formula of the data registration algorithm is as follows: Where A(p,t) represents the alignment result of position p at time t; D i (p, t) represents the observation value of the i-th data source at position p and time t; d(p,p i ) represents the reference position p of the data source i i The spatial distance between i | represents time t and sampling time t of data source i i The time difference between i represents the weight coefficient of the i-th data source; Q i represents the spatiotemporal resolution coefficient of data source i; Δs i represents the spatial resolution of data source i; Δt i represents the temporal resolution of data source i; a1, a2, and γ are adjustment parameters used to control spatial attenuation, temporal attenuation, and resolution effects.

5. The power supply service risk identification system based on big data analysis according to claim 1, characterized in that: The risk identification module includes: A feature extraction unit, used for extracting service quality features from the fused data; A risk assessment unit is used to construct a risk assessment index system and calculate the power supply service risk assessment index value based on service quality characteristics; The risk classification unit is used to establish a dynamic threshold model based on historical data statistics, determine the risk level according to the power supply service risk assessment index value and the dynamic threshold model, determine the risk type based on the feature combination pattern, and output the risk identification result.

6. The power supply service risk identification system based on big data analysis according to claim 5, characterized in that: The execution process of the feature extraction unit is as follows: Extract time series features from work order data: The work order data is segmented through the sliding time window W(t,Δt), and the window length Δt is adaptively adjusted according to the business characteristics: Δt=max(μ d ·(1+σ d ),Δt min ), where μ d is the average processing time of historical work orders, σ d is the standard deviation, Δt min is the lower limit of the window length; Construct an LSTM network, including an attention-enhanced memory unit, control the information flow through the forget gate, input gate, and output gate, and extract the temporal feature f τ =[f τ1 ,f τ2 ,f τ3 ], where f τ1 is the efficiency trend characteristic, through f τ2 is the traffic fluctuation characteristic, f τ3 is the temporal pattern feature; Extract semantic features from feedback data: Construct a multi-level attention mechanism including word level, sentence level and document level: α w =softmax(v w ·tanh(W w ·h w )) α s =softmax(v s ·tanh(W s ·h s )) α d =softmax(v d ·tanh(W d ·h d )) In the formula, α w is the word-level attention weight, with a value range of [0,1], and the sum in each sentence is 1; α s is the sentence-level attention weight, with a value range of [0,1] and a sum of 1 within a document; α d is the document-level attention weight, with a value range of [0,1] and a sum of 1; h w is the word-level hidden layer representation, and its dimension is the word embedding dimension; h s is the sentence-level hidden layer representation, and its dimension is the sentence encoding dimension; h d is the document-level hidden layer representation, and its dimension is the document encoding dimension; W w , W s , W d is the attention weight matrix of each layer; v w 、v s 、v d is the attention vector of each layer; By calculating the attention weights of each layer to identify key information, the semantic features f are extracted s =[f s1 ,f s2 ,f s3 ], where f s1 is the topic distribution feature, f s2 is the keyword feature, f s3 for emotional characteristics; Extract features from public opinion data: Construct multimodal features of public opinion texts, including: extracting text features using a multi-level attention mechanism, extracting propagation network features through a graph convolutional network (GCN), analyzing public opinion evolution using a time-decayed attention mechanism, and extracting temporal features; Calculate and analyze multimodal features to extract public opinion features f m =[f m1 ,f m2 ,f m3 ], where f m1 is the public opinion influence characteristic, f m2 is the evolution characteristic of public opinion, f m3 is the public opinion clustering feature; Interactive feature extraction based on multi-source data: Construct a data association graph G(V,E), where the node set V contains work order data, feedback data, and public opinion data, and the edge set E represents the association between entities; Perform weighted summation according to content similarity, time correlation, and spatial correlation, calculate the association strength between nodes, and construct an association strength matrix; Extracting interactive features f based on association strength matrix and data association graph c =[f c1 ,f c2 ,f c3 ], where f c1 is the relationship feature, f c2 is the propagation characteristic, f c3 is the associated feature; For the time series feature f τ , semantic features f s 、Public opinion characteristics m , interaction feature f c After normalization, the fusion weight is determined through the attention mechanism to obtain the final quality of service feature F.

7. The power supply service risk identification system based on big data analysis according to claim 5, characterized in that: The risk assessment index system includes two layers of indicators. The first-layer indicators include service response indicators, service quality indicators, user satisfaction indicators, and social influence indicators; The second-layer indicators are the next-level indicators corresponding to the first-layer indicators. The service response indicators include the average response duration and the first response timeliness rate. The service quality indicators include the problem-solving rate and service standardization. The user satisfaction indicators include the complaint rate and the negative review rate. The social influence indicators include the public opinion diffusion degree and the negative impact degree.

8. The power supply service risk identification system based on big data analysis according to claim 7, characterized in that: The process of calculating the power supply service risk assessment index value based on service quality characteristics is as follows: Construct a feature mapping function to calculate the second-layer index value vector I: I=σ(W·F+b)·(1+tanh(F T ·M·F)) where σ is the sigmoid function; F is the service quality feature; W is the feature mapping matrix; M is the feature interaction matrix; b is the bias vector; Calculate the power supply service risk assessment index value R based on the second-layer index value vector I: R = ψ(I)·(1 + η·‖ΔI‖) In the formula, ψ(I) is a non-linear combination function used to fuse the index values of each dimension. η is a time-varying parameter. ΔI is the change in the index value. ‖·‖ represents the norm operation.

9. The power supply service risk identification system based on big data analysis according to claim 8, characterized in that: In the risk classification unit, the dynamic threshold model is established through the following steps: Construct a time series feature sequence based on historical data, and use a long short-term memory network (LSTM) to extract time-dependent features z(t); Calculate the adaptive factor λ(t): In the formula, u1 and u2 are adjustment parameters; Calculate the dynamic threshold boundaries T1(t) and T2(t), where T1(t) < T2(t): T1(t)=μ his -λ(t)·σ his T2(t)=μ his +λ(t)·σ his In the formula, μ his is the mean of the historical risk assessment index values, σ his is the standard deviation of the historical risk assessment indicator values; Determine the risk level according to the power supply service risk assessment index value R: When R < T1(t), it is determined as a low risk; When T1(t) ≤ R < T2(t), it is determined as a medium risk; When R ≥ T2(t), it is determined as a high risk.

10. The power supply service risk identification system based on big data analysis according to claim 6, characterized in that: In the risk classification unit, the feature combination pattern represents the feature distribution law corresponding to different risk types, including the combination relationship and time series evolution characteristics of each dimension feature in the service quality feature F; For each moment t in the historical data, construct a feature pattern matrix M(t): M(t) = [F(t), ΔF(t), δF(t)] where F(t) is the historical service quality feature at moment t, ΔF(t) is the change in the feature, and δF(t) is the change rate of the feature; Based on the sequence of feature pattern matrices {M(t)} constructed from historical data, use a density clustering algorithm for analysis. Cluster the patterns with similar feature distributions and evolution laws into one category, and each category corresponds to a specific risk type, thus forming a risk feature pattern library; Construct a feature pattern matrix M(t0) for the current moment t0, and determine the specific risk type by calculating the similarity between M(t0) and each type of pattern in the risk feature pattern library. The similarity calculation uses a non-linear measurement method considering feature importance.

Images