A power distribution network power failure early warning method based on multi-source data fusion

Abstract

The present application relates to the technical field of power distribution network monitoring, and more particularly to a power distribution network outage early warning method based on multi-source data fusion, comprising: collecting historical operation data of the power distribution network; preprocessing the collected historical operation data; constructing a time series sample set based on the preprocessed data and labeling the corresponding outage label; constructing a double-path outage prediction model based on a long short-term memory network, wherein the double-path outage prediction model includes a basic risk path and a fault increment path, the basic risk path is based on numerical features for time series modeling, and the fault increment path is based on comprehensive coupling quantitative scoring for time series modeling; collecting real-time operation data of the power distribution network, inputting the preprocessed data into the double-path outage prediction model, outputting outage probability prediction values and triggering early warning. The present application realizes the deep fusion of multi-source heterogeneous data of the power distribution network and the accurate prediction of outage risk by constructing a coupling quantitative processing method of maintenance record text data and a double-path outage prediction model.

Description

Technical Field

[0001] This invention relates to the field of power distribution network monitoring technology, and in particular to a power outage early warning method for power distribution networks based on multi-source data fusion. Background Technology

[0002] With the continuous expansion of power distribution networks and the increasing demands for power supply reliability from users, the impact of power outages on social production and daily life is becoming increasingly significant. Power outage early warning systems, as a crucial means of improving power supply reliability, have become one of the core technologies for intelligent operation and maintenance of power systems. During the operation of power distribution networks, massive amounts of multi-source heterogeneous data are generated, including numerical monitoring data such as load, voltage, and current, as well as textual data such as equipment maintenance records and fault descriptions. How to effectively integrate this multi-modal data to achieve accurate power outage risk prediction is a pressing technical challenge that needs to be addressed.

[0003] Existing power outage early warning methods for distribution networks primarily rely on the analysis of numerical monitoring data, using statistical analysis or machine learning models to assess equipment operating status. However, these methods are insufficient in utilizing unstructured textual data such as equipment maintenance records. While some studies have attempted to integrate textual features, they typically employ late-stage fusion strategies, simply concatenating textual and numerical feature vectors before inputting them into the prediction model. This approach ignores the inherent correlation between textual and numerical data, limiting the accuracy of early warnings. Furthermore, existing methods often treat all features uniformly during time-series modeling, failing to effectively distinguish between the risks of natural evolution of equipment operating status and the sudden risks caused by fault events, thus affecting the model's ability to characterize outage mechanisms.

[0004] Chinese patent CN116304928A discloses a method for predicting power supply equipment faults. This method collects operating condition data of the power supply equipment, extracts features from the operating condition data to obtain feature data, and inputs the feature data into a pre-set fault prediction model to obtain the fault prediction result. This method constructs decision tree models for different equipment types and uses historical operating data to train the model to predict the future operating status of the equipment. However, this method mainly relies on numerical operating parameters for fault prediction and does not involve the quantitative processing and fusion of maintenance record text data. Furthermore, the decision tree model it uses is a static classification model, lacking the ability to model the temporal evolution of equipment operating states, and thus failing to capture the dynamic changes in power outage risk over time. Summary of the Invention

[0005] In view of this, the present invention provides a power outage early warning method for distribution networks based on multi-source data fusion, in order to solve the technical problems in the prior art of insufficient utilization of unstructured text data and insufficient fusion depth of multi-source heterogeneous data in power outage early warning, which leads to limited accuracy of power outage prediction.

[0006] The technical solution of this invention is implemented as follows: This invention provides a power outage early warning method for distribution networks based on multi-source data fusion, comprising: S1. Collect historical operating data of the power distribution network, including basic equipment information, operation monitoring data, and maintenance record text data; S2. Preprocess the collected historical operation data, including cleaning and normalizing the numerical data to obtain numerical features, and performing coupled quantization on the maintenance record text data to obtain a comprehensive coupled quantization score. S3. Construct a time series sample set based on the preprocessed data and label it with the corresponding power outage tags; S4. Construct a dual-path power outage prediction model based on long short-term memory network. The dual-path power outage prediction model includes a basic risk path and a fault incremental path. The basic risk path is modeled in time series based on preprocessed numerical features, and the fault incremental path is modeled in time series based on comprehensive coupled quantitative scoring. S5. Collect real-time operation data of the power distribution network, input it into the dual-path power outage prediction model after preprocessing, output the predicted power outage probability value and trigger an early warning.

[0007] Based on the above technical solutions, preferably, in step S1, the basic equipment information includes equipment number, equipment type, commissioning date, and rated lifespan, and the equipment aging degree is calculated based on the equipment commissioning date and rated lifespan; the operation monitoring data includes power load and historical power outage count; the maintenance record text data includes maintenance date, fault description, fault cause, handling measures, and expected review time; the equipment aging degree is calculated based on the equipment commissioning date and rated lifespan.

[0008] Based on the above technical solutions, preferably, step S2, which involves performing coupled quantitative processing on the maintenance record text data to obtain a comprehensive coupled quantitative score, specifically includes: S21. Perform keyword matching on the fault description text in the maintenance record to identify the fault type, and assign a basic sensitivity coefficient to each fault type based on the degree of impact of the fault type on the power outage event. S22. Extract the fault propagation chain by identifying causal keywords in the maintenance records, count the number of nodes in the fault propagation chain, and calculate the propagation weight coefficient based on the length of the fault propagation chain. S23. Construct a numerical data coupling function, which is calculated based on power load, equipment aging level, and historical power outage frequency. S24. Calculate the time decay factor based on the time interval between the maintenance record date and the current predicted date, and calculate the review reminder factor based on the proximity of the current time to the expected review time; S25. Based on the basic sensitivity coefficient, propagation weight coefficient, numerical data coupling function, time decay factor, and review reminder factor, calculate the comprehensive coupling quantitative score for each maintenance record.

[0009] Based on the above technical solution, preferably, in step S24, the formula for calculating the time decay factor is as follows: ; The formula for calculating the review reminder factor is: ; in, The time decay factor, The time interval between the maintenance record date and the current predicted date. The attenuation coefficient is... To verify the reminder factor, To estimate the review time, For the current time, To verify the influence intensity coefficient, This is the time window width parameter.

[0010] Based on the above technical solutions, preferably, step S3 includes: At each time step, a comprehensive coupled quantitative score is extracted, which includes power load, equipment aging level, number of historical power outages, and all maintenance records corresponding to that time step, to construct a time-series feature vector sequence. Set early warning time windows based on business needs, and label samples according to whether a power outage event will occur within the future early warning time window; divide all samples into training set, validation set, and test set in chronological order. Based on the above technical solutions, preferably, step S4 includes: S41. Decouple the comprehensive coupled quantitative scoring and extract pure fault semantic features; S42. Set up a basic risk path and a fault incremental path in the long short-term memory network. The basic risk path is input with preprocessed numerical features to learn the basic probability of power outage caused by the natural evolution of the equipment's operating status. The fault incremental path is input with pure fault semantic features to learn the incremental contribution of fault records to the power outage risk. S43. Design an adaptive modulation fusion mechanism to dynamically reconstruct the coupling relationship between basic risk and fault incremental risk through modulation coefficients, output the power outage probability prediction value, and obtain a dual-path power outage prediction model.

[0011] Based on the above technical solutions, preferably, in step S41, the decoupling formula is: ; in, For the pure fault semantic features after decoupling, A comprehensive coupled quantitative score for time step t. This represents the electrical load value at time step t. This represents the degree of equipment aging at time step t. This represents the historical number of power outages at time step t. For numerical data coupling functions, The time decay factor, This is a reminder factor for review.

[0012] Based on the above technical solutions, preferably, in step S42, the hidden state update process of the basic risk path is as follows: ; in, The normalized power load is the amount of electricity generated at time step t. The equipment aging degree is normalized to time step t. The number of historical power outages after normalization at time step t. Let be the basic risk hidden state vector at time step t. Long Short-Term Memory (LSTM) network units representing basic risk paths, This is the hidden state from the previous time step; The hidden state update process for the fault incremental path is as follows: ; in, Let be the fault increment hidden state vector at time step t. Long Short-Term Memory (LSTM) network units representing fault incremental paths, This is the hidden state of the previous time step.

[0013] Based on the above technical solutions, preferably, the adaptive modulation fusion mechanism in step S43 is as follows: the hidden states of the two paths are fused using a residual connection method, and the fusion formula is: ; in This represents the hidden state of the overall risks after integration. The hidden state of the basic risk path, This represents the hidden state of the fault incremental path. The modulation coefficient is the modulation coefficient at time step t; the modulation coefficient The input is a concatenated vector obtained through a fully connected network, consisting of the hidden state of the basic risk path, the hidden state of the fault incremental path, the power load, and the degree of equipment aging. The fused comprehensive risk hidden state is then mapped to a predicted power outage probability value through a fully connected layer and an activation function.

[0014] Based on the above technical solution, preferably, step S4 further includes training the dual-path power outage prediction model using a time-series sample set. The training process uses a combined loss function to optimize the model parameters. The combined loss function includes the main task loss, path independence constraint loss, and modulation coefficient regularization loss. The main task loss is constructed based on the cross-entropy loss function to optimize the accuracy of power outage probability prediction. The path independence constraint loss is constructed based on feature correlation measurement to ensure the independence between the basic risk path and the pure fault semantic features. The modulation coefficient regularization loss is constructed based on physical constraints to constrain the rationality of the modulation coefficient.

[0015] The present invention has the following advantages over the prior art: (1) This invention achieves deep fusion of multi-source heterogeneous data in the distribution network and accurate prediction of power outage risks by constructing a coupled quantification processing method for maintenance record text data and a dual-path power outage prediction model. This method transforms unstructured maintenance record text into quantifiable risk scores and incorporates numerical features of equipment operating status into the quantification process, overcoming the problem of insufficient utilization of text data in existing technologies. By modeling the natural evolution risk of equipment operating status and the incremental risk of fault events through a dual-path long short-term memory network, and using an adaptive modulation fusion mechanism to dynamically reconstruct the coupling relationship between the two types of risks, this invention solves the problem that existing technologies fail to effectively distinguish the evolution patterns of different types of risks, improves the accuracy and reliability of power outage early warning, and provides an effective technical means for intelligent operation and maintenance of the distribution network.

[0016] (2) The maintenance record text coupling quantitative processing method proposed in this invention transforms the semantic information in the maintenance record into a comprehensive coupling quantitative score by identifying the fault type, extracting the fault propagation chain, constructing a numerical data coupling function, calculating the time decay factor and the review reminder factor. This method considers multiple factors such as fault severity, chain effect, equipment operating status, time decay, and review risk, and realizes early fusion of text features and numerical features in the quantification stage. Compared with simple text vectorization or late feature splicing methods, it can more fully mine the power outage risk information contained in the maintenance record and improve the fusion depth and feature expression capability of multimodal data.

[0017] (3) The feature decoupling and dual-path modeling strategy adopted in this invention extracts pure fault semantic features by decoupling the comprehensive coupling quantitative scoring, so that the basic risk path and the fault incremental path can model different types of power outage risks based on independent feature spaces. The basic risk path captures the gradual risk evolution law of equipment caused by factors such as load fluctuation and aging accumulation, while the fault incremental path captures the sudden impact and time-series cumulative effect of fault events. The independent modeling of the two paths avoids the confusion and redundancy of feature information, making the model have a clearer physical interpretability of the formation mechanism of power outage risk, which is convenient for operation and maintenance personnel to understand the early warning results and formulate targeted disposal measures. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The flowchart below shows the power outage early warning method for distribution networks based on multi-source data fusion according to the present invention. Figure 2 This is a schematic diagram of the text coupling and quantization processing flow of the maintenance record of the present invention; Figure 3 This is a schematic diagram of the dual-path power outage prediction model architecture of the present invention; Figure 4 This is a detailed structural diagram of the adaptive modulation fusion mechanism of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] like Figure 1 As shown, this invention provides a power outage early warning method for distribution networks based on multi-source data fusion, comprising: S1. Collect historical operating data of the power distribution network, including basic equipment information, operation monitoring data, and maintenance record text data; S2. Preprocess the collected historical operation data, including cleaning and normalizing the numerical data to obtain numerical features, and performing coupled quantization on the maintenance record text data to obtain a comprehensive coupled quantization score. S3. Construct a time series sample set based on the preprocessed data and label it with the corresponding power outage tags; S4. Construct a dual-path power outage prediction model based on long short-term memory network. The dual-path power outage prediction model includes a basic risk path and a fault incremental path. The basic risk path is modeled in time series based on preprocessed numerical features, and the fault incremental path is modeled in time series based on comprehensive coupled quantitative scoring. S5. Collect real-time operation data of the power distribution network, input it into the dual-path power outage prediction model after preprocessing, output the predicted power outage probability value and trigger an early warning.

[0022] In one embodiment of the present invention, step S1 specifically includes: collecting multi-source heterogeneous data from business systems such as the power distribution network dispatch management system, production management system, and equipment management system. Basic equipment information includes static attributes such as equipment number, equipment type, commissioning date, and rated lifespan. Equipment types include major power distribution network equipment such as transformers, switchgear, circuit breakers, cables, and overhead lines. Operation monitoring data includes dynamic operating parameters such as power load and historical outage counts. Power load data is typically collected hourly or minutely, while historical outage counts are accumulated daily or weekly. Maintenance record text data includes unstructured information such as maintenance date, fault description, fault cause, handling measures, and expected review time. This text data typically comes from business systems such as equipment defect records, fault repair records, and periodic inspection records.

[0023] The calculation of equipment aging is performed as follows: First, the equipment's service life (in years) is calculated based on the difference between the current date and the equipment's commissioning date. Then, the service life is compared to the rated life of the equipment type to obtain a preliminary aging ratio. This ratio is then normalized; a ratio between 0 and 1 is directly used as the equipment's aging level. When the ratio exceeds 1, it is set to 1, indicating that the equipment has exceeded its rated lifespan and entered a high-risk operation phase. This indicator reflects the increasing trend of power outage risk due to physical aging and insulation degradation of the equipment.

[0024] Historical power outage count Statistics, setting the time window length Within this time window, the cumulative number of power outages for each device is counted. The length of the time window is adjusted according to the business scenario and device type. For devices with frequent failures, such as overhead lines, the time window can be set to 90 days or 120 days to capture recent power outage frequency; for devices with low failure rates, such as main transformers, the time window can be set to 180 days or 365 days to obtain stable statistical characteristics.

[0025] In one embodiment of the present invention, step S2, the preprocessing of numerical data, includes: performing data quality checks on numerical data such as power load and historical power outage frequency, identifying and processing outliers, missing values, and duplicate values. Outlier detection employs statistical methods, such as the three-standard-deviation rule or box-ring method. Figure 4 The quantile method removes data points that are clearly outside the reasonable range. For missing values, different imputation strategies are selected based on the missing percentage and business characteristics. When the missing percentage is low, forward imputation is used, that is, the missing position is filled with the nearest valid value; when the missing percentage is moderate and the data changes slowly, linear interpolation is used; features with an excessively high missing percentage may be considered for complete removal or assigned values ​​using business rules.

[0026] Normalization is performed using the minimum-maximum normalization method, mapping features of different dimensions and numerical ranges to a unified interval, typically [0, 1]. Preprocessing of maintenance record text data includes text cleaning and standardization. First, the text is segmented using a Chinese word segmentation tool to divide continuous text into word sequences, then stop words, punctuation marks, and other semantically meaningless components are removed. For equipment terminology and abbreviations, a terminology dictionary is established for standardized mapping; for example, different expressions such as "switch trip," "switch open," and "switch action" are unified into the standard term "switch trip." For numerical and unit expressions, different formats such as "10 kV," "10 kV," and "10 kV" are unified into a standard format.

[0027] like Figure 2 As shown, step S2, which involves performing coupled quantization processing on the maintenance record text data to obtain a comprehensive coupled quantization score, specifically includes: S21. Perform keyword matching on the fault description text in the maintenance records to identify fault types, and assign a basic sensitivity coefficient to each fault type based on its impact on the power outage event. A fault type knowledge base is established, containing keyword sets for various fault types and corresponding basic sensitivity coefficients. Fault types are categorized into three levels based on severity: high-risk faults, medium-risk faults, and minor faults. High-risk faults include those that directly cause power outages, such as line short circuits, main transformer tripping, busbar faults, and line grounding; their basic sensitivity coefficient ranges from 1.5 to 2.0. Medium-risk faults include those that may indirectly affect power supply reliability, such as switch failure, secondary circuit abnormalities, and localized overheating; their basic sensitivity coefficient ranges from 1.0 to 1.5. Minor faults include those that do not temporarily affect power supply but require attention, such as localized overheating, minor discharge, and signs of insulation aging; their basic sensitivity coefficient ranges from 0.5 to 1.0. Keyword matching employs regular expressions or dictionary matching methods. When a characteristic keyword of a certain fault type appears in the maintenance record text, the record is classified as the corresponding fault type and assigned the corresponding basic sensitivity coefficient.

[0028] S22. Extract the fault propagation chain by identifying causal keywords in the maintenance record, count the number of nodes in the fault propagation chain, and calculate the propagation weight coefficient based on the length of the fault propagation chain. The fault propagation chain reflects the chain reaction process of a fault event. For example, the description "transformer oil temperature rise leads to overheat protection action, triggering transformer tripping and causing power loss in downstream lines" contains multiple causal relationship nodes. Causal keywords include conjunctions indicating causal relationships such as "due to," "leading to," "causing," "furthermore," "causing," and "ultimately." By identifying these keywords, the maintenance record text is parsed into a causal chain form, with each node representing a fault phenomenon or consequence. The number of nodes in the propagation chain is recorded as... The more nodes there are, the wider the scope of the fault's impact and the stronger the chain reaction.

[0029] Propagation weight coefficient The calculation formula is: ; in, For propagation weighting coefficients, This is the gain coefficient for a single node in the propagation chain, reflecting the incremental contribution of each additional propagation node to the risk score. A typical value is 0.2 to 0.5. This represents the number of nodes in the fault propagation chain. The baseline value of this formula is 1, when... This indicates that the repair record lacks a clear causal description and is a simple, single fault. This indicates no chain reaction. When When it increases, The linear increase reflects the amplified risk effect brought about by the expansion of the fault propagation range.

[0030] S23. Construct a numerical data coupling function based on power load, equipment aging level, and historical outage frequency. Since the risk of power outage varies under different operating conditions for the same fault type, equipment operating under high load, severely aged equipment, and with a history of frequent outages has a higher probability of causing a power outage. The coupling function uses a linear weighted form. ; in, For numerical data coupling function values, The power load at the current time step. Historical average load is typically calculated as the average load over the past year or six months. Load deviation reflects the degree to which the current load deviates from the average level. The value is a normalized value for the degree of equipment aging, ranging from 0 to 1, and is calculated in step S1. For time window The number of historical power outages within the area To calculate the length of the statistical time window, the unit is days. The power outage frequency reflects the number of power outages per unit of time. , , These are weighting coefficients for load, aging, and number of power outages, used to adjust the contribution of each factor to the coupling function. The weighting coefficients can be adjusted based on historical data statistical analysis or expert experience.

[0031] The baseline value of this function is 1. When the equipment is at its average load level, with no significant aging and no historical power outages, all deviations are 0. When the load exceeds the average value, the aging process is severe, and power outages are frequent, the weighted values ​​of each item are positive. A value greater than 1 indicates that the current equipment status amplifies the risk of a power outage due to a fault. Conversely, when the load is below average, the equipment is relatively new, and there is no record of power outages, <1 indicates that the risk of a power outage due to a fault is relatively low in the current state.

[0032] S24. Calculate the time decay factor based on the time interval between the maintenance record date and the current predicted date, and calculate the review reminder factor based on the proximity of the current time to the expected review time.

[0033] The impact of a failure event on equipment risk gradually diminishes over time. Equipment that has just experienced a failure may have unresolved hidden dangers or potential related defects, while the risk of equipment that has been faulty for a longer period of time tends to stabilize after repair and verification. The formula for calculating the time decay factor is as follows: ; in, is the time decay factor for time step t, ranging from 0 to 1, and e is the base of the natural logarithm. This is the attenuation coefficient, which controls the rate of attenuation and can be adjusted according to different fault types and equipment characteristics. The time interval, in days, between the maintenance record date and the current predicted date corresponding to time step t. This exponential decay function... A value of 1 indicates that the impact on the repair records for that day is greatest, and so on. As the value increases, the attenuation factor gradually decreases. As the value approaches infinity, the decay factor approaches 0.

[0034] The review reminder factor is calculated based on the proximity of the current time to the expected review time. After equipment repair, a scheduled review time is usually set to verify the repair effectiveness and identify potential problems. When the current time approaches the review time, the equipment may be in a period of risk accumulation, potentially leading to power outages due to incomplete repairs or the emergence of new defects. The formula for calculating the review reminder factor is... ; in, This is the review reminder factor for time step t, with a value greater than or equal to 1. To verify the influence intensity coefficient, the peak increase of the verification factor was controlled. The estimated review time is in days and is determined by the estimated review date recorded in the maintenance record. This refers to the current time corresponding to time step t, in days. This represents the difference in days between the current time and the review time. The time window width parameter controls the width of the Gaussian function. For Gaussian functions, in The Gaussian term reaches a peak value of 1 when the current time is far from the review time, and then decays rapidly. The base value of this formula is 1; as the current time moves far from the review time, the Gaussian term approaches 0. A value close to 1 indicates that the review factor has no effect. As the current time approaches the review time, the Gaussian term approaches 1. Reaching the maximum value This reflects the cumulative effect of potential risks that may exist in the equipment before the review.

[0035] S25. Based on the basic sensitivity coefficient, propagation weight coefficient, numerical data coupling function, time decay factor, and review reminder factor, calculate the comprehensive coupling quantitative score for each maintenance record.

[0036] The formula for calculating the comprehensive coupled quantitative score is as follows: ; in, A comprehensive coupled quantitative score for time step t. The basic sensitivity coefficient for fault types. This is the fault propagation weighting coefficient. For numerical data coupling functions, The time decay factor, This serves as a review and reminder factor. The formula employs multiplicative multifactor coupling to reflect the synergistic effect of various factors on power outage risk. A larger value for a particular factor significantly improves the overall risk score. This quantification method transforms unstructured maintenance record text into quantifiable risk scores. Simultaneously, it deeply integrates numerical features of equipment operating status during the quantification process, achieving early fusion of multimodal data and providing information-rich input features for subsequent time-series modeling.

[0037] In one embodiment of the present invention, step S3 specifically includes: extracting the power load at each time step t. Equipment aging degree Historical power outages This includes the comprehensive coupled quantitative score of all maintenance records corresponding to that time step. For cases where multiple maintenance records exist for the same device at the same time step, the comprehensive coupled quantitative scores of all maintenance records are summed to obtain the total text score for that time step. If there is no maintenance record for a certain time step, then... This represents the incremental risk of no fault events at that time step. Construct a sequence of time-series feature vectors. Where T is the time series length, i.e. the number of historical time steps contained in each sample, which is set according to business needs and data characteristics.

[0038] Based on business needs, a warning time window is set. For each time series sample, after constructing the time series feature vector sequence, it is queried whether a power outage event will occur for that device within the future warning time window. If a power outage occurs, the sample is labeled as 1; otherwise, it is labeled as 0. The length of the warning time window is set according to the operation and maintenance response capability and business scenario. A time window that is too short will result in insufficient time for post-warning response, while a time window that is too long will reduce the urgency and accuracy of the warning. Typical values ​​are 7 days, 15 days, or 30 days. Considering that power outage events are low-frequency events, the number of power outage samples in actual data is far less than the number of non-power outage samples, leading to sample imbalance. When constructing the training set, oversampling or undersampling strategies can be used to adjust the ratio of positive to negative samples. Oversampling increases the number of power outage samples through duplication or synthesis, while undersampling reduces the number of non-power outage samples through random sampling, so that the ratio of positive to negative samples reaches a relatively balanced state, typically 1:3 to 1:5.

[0039] All samples are divided into training, validation, and test sets in chronological order, with a typical ratio of 7:1.5:1.5. The time series data must be divided strictly according to chronological order to ensure that the test set's time range is significantly later than the training set, avoiding data leakage that could lead to artificially inflated model performance. The training set is used for model parameter optimization, the validation set for hyperparameter tuning and early stopping detection, and the test set for final model performance evaluation.

[0040] In one embodiment of the present invention, such as Figure 3 and 4 As shown, step S4 specifically includes: S41. Decouple the comprehensive coupled quantitative score and extract pure fault semantic features. The comprehensive coupled quantitative score constructed in step S2 integrates textual and numerical features. During the time-series modeling stage, these two types of features need to be separated so that the dual-path model can independently model the basic risk and the incremental fault risk. The decoupling formula is as follows: ; in, For the pure fault semantic features after decoupling, A comprehensive coupled quantitative score for time step t. This represents the electrical load value at time step t. This represents the degree of equipment aging at time step t. This represents the historical number of power outages at time step t. For numerical data coupling functions, The time decay factor, This is a review reminder factor. The division operation is the inverse operation of the multiplication coupling in step S25. By eliminating the modulation effect of numerical factors, time factors, and review factors on text features, it makes... Only retain fault type sensitivity and propagation chain weight Information, namely By decoupling the operations, it is ensured that the basic risk path learns the power outage probability of natural equipment evolution only from numerical features, and the fault increment path learns the incremental impact of fault events only from pure fault semantic features, thus maintaining feature independence between the two paths.

[0041] S42. In the Long Short-Term Memory (LSTM) network, a basic risk path and a fault incremental path are set. The basic risk path is input with normalized preprocessed numerical features. The basic probability of power outage caused by the natural evolution of the operating status of learning equipment is used to learn the incremental contribution of fault records to power outage risk by inputting pure fault semantic features of the fault incremental path.

[0042] The basic risk path is modeled temporally using a Long Short-Term Memory network, and the hidden state update process is as follows: ; in, This is the underlying risk hidden state vector at time step t, typically with dimensions set to 64 or 128. The long short-term memory network unit representing the basic risk path includes internal structures such as input gates, forget gates, output gates, and cell states. The concatenated input feature vector contains the normalized electricity load. Normalized equipment aging degree and normalized historical power outages , This represents the hidden state vector from the previous time step. Long Short-Term Memory (LSTM) networks effectively capture long-term dependencies in time-series data through gating mechanisms. The input gate determines how much of the current input information is remembered, the forget gate determines how much historical information is retained, and the output gate determines how much information is output from the current hidden state. Basic risk path modeling describes the temporal evolution of power outage risks caused by factors such as load fluctuations, accumulated equipment aging, and historical outage frequency, under conditions of no fault records or minimal fault impact. This evolution is gradual and continuous.

[0043] The fault incremental path also uses a long short-term memory network for time-series modeling, and the hidden state update process is as follows: ; in, This is the fault increment hidden state vector at time step t, typically with dimensions set to 64 or 128. The long short-term memory network unit representing the fault incremental path has the same structure as the LSTM unit of the basic risk path, but with independent parameters. The input is pure fault semantic features. This is the hidden state vector from the previous time step. The fault incremental path model models the additional incremental impact of fault records on the power outage probability. Since maintenance records are discrete events, in most time steps... This indicates no fault record, only the time step when the fault occurred. Take non-zero values. Capture the suddenness of fault events and the cumulative and decaying effects of fault impacts over time through fault incremental paths.

[0044] S43. Design an adaptive modulation fusion mechanism to dynamically reconstruct the coupling relationship between basic risk and incremental fault risk through modulation coefficients, outputting the predicted power outage probability value to obtain a dual-path power outage prediction model. Specifically, the hidden states of the two paths are fused using a residual connection method, and the fusion formula is as follows: ; in, Let be the integrated risk hidden state vector after fusion at time step t, with dimensions and and same, The hidden state vector of the basic risk path, This represents the hidden state vector of the fault incremental path. Let be the modulation coefficient at time step t, and be a scalar value. The contribution of the incremental fault risk is dynamically adjusted through the modulation coefficient.

[0045] Modulation coefficient The data is obtained through a fully connected neural network, with the input being a concatenated vector of the hidden states of basic risk paths, incremental fault paths, power load, and equipment aging. The calculation formula is as follows: ; in, The modulation coefficient, For activation functions, the Sigmoid function is typically used to limit the output to between 0 and 1, or a scaled Sigmoid function is used, such as... The output is mapped to a range of 0 to 3, allowing modulation coefficients greater than 1 to amplify the effects of faults. Here is the weight matrix of the fully connected layer, with dimension 1. Where d is the dimension of the hidden state vector. For scalar bias, This represents a vector concatenation operation, which hides the underlying risk state vector. Fault Increment Hidden State Vector Normalized power load scalar and normalized equipment aging scale The modulated coefficients are concatenated into a long vector. This concatenation enables the modulation coefficients to sense the timing status of the two paths, as well as the current load and aging conditions, thus achieving state-aware dynamic modulation.

[0046] The fused comprehensive risk hidden state is mapped to a predicted power outage probability value through a fully connected layer and an activation function. ; in, The probability of a power outage predicted at time step t. This is the weight vector of the fully connected layer at the output layer. This is the output layer bias scalar. This is the hidden state vector of the integrated risk after fusion.

[0047] The aforementioned dual-path decoupling modeling and dynamic modulation fusion mechanism, compared to traditional single-path models or simple multi-feature splicing models, can more accurately characterize the formation mechanism of power outage risks. Traditional single-path models input all features uniformly, failing to distinguish the evolution patterns of different types of risks. While simple splicing models fuse multiple features, they do not consider the coupling relationships between features. This invention ensures that the two paths independently model different types of risks through feature decoupling, and achieves adaptive fusion of the two types of risks through learnable modulation coefficients. This maintains feature independence while reconstructing dynamic coupling relationships, improving the model's physical interpretability and prediction accuracy.

[0048] In one embodiment of the present invention, step S4 further includes training the dual-path power outage prediction model using a time-series sample set. The training process uses a combined loss function to optimize the model parameters. The combined loss function includes the main task loss, path independence constraint loss, and modulation coefficient regularization loss.

[0049] The main task loss is constructed based on the cross-entropy loss function to optimize the accuracy of power outage probability prediction. Cross-entropy loss measures the difference between the predicted probability distribution and the true label distribution; minimizing this loss makes the model's output power outage probability closer to the actual situation. To address the sample imbalance problem caused by the sparsity of power outage samples, a weighted cross-entropy loss is used, assigning higher weights to power outage class samples, typically with a weight ratio of 5:1 to 10:1. The main task loss function is: ; in, is the loss value for the main task, N is the total number of training samples, and i is the sample index. Let i be the true label of the i-th sample. Let be the predicted power outage probability for the i-th sample. These are the weighting coefficients for the power outage category samples. These are the weighting coefficients for samples that are not affected by power outages.

[0050] The path independence constraint loss is constructed based on feature correlation metrics to ensure the independence between the underlying risk path and the pure fault semantic features. By calculating the Pearson correlation coefficient between the hidden state of the underlying risk path and the decoupled pure fault semantic features, its absolute value is constrained to be close to zero. The loss function is: ; in, This represents the loss value due to path independence constraints. This represents the Pearson correlation coefficient function, which measures the degree of linear correlation between two variables. The hidden state sequence of the basic risk path from time step 1 to T. This is a numerical sequence of pure fault semantic features from time step 1 to time step T, where T is the time series length. This represents the absolute value operation. By minimizing this loss, the model is forced to maintain a low correlation between the base path and the fault features during training, ensuring the effective implementation of feature decoupling.

[0051] The modulation coefficient regularization loss is constructed based on physical constraints to ensure the rationality of the modulation coefficients, including range constraints and temporal smoothness constraints. The regularization loss function is: ; in, Here, t represents the modulation coefficient regularization loss value, and t is the time step index. The modulation coefficient at time step t This is the upper bound of the modulation coefficient. This is the lower bound of the modulation coefficient. This represents taking the maximum value between x and 0, when Exceeding A penalty is incurred when the range is exceeded. These are the weighting coefficients for the range constraints. These are the weighting coefficients for the smoothness constraint. It is the square of the difference between the modulation coefficients in adjacent time steps, which measures the temporal variation of the modulation coefficients.

[0052] The total loss function is a weighted sum of the three losses: in, This is the total loss value. Loss to the main task For path independence constraint loss, For modulation coefficient regularization loss, These are the weighting coefficients for the path independence constraint loss. This represents the weighting coefficient for the regularization loss. The combined loss function is used to perform end-to-end optimization of the model parameters via backpropagation, employing an adaptive learning rate optimizer such as Adam. A typical learning rate is 0.001, and the batch size is set based on the sample size, typically 32 or 64. During training, the main task loss and warning accuracy on the validation set are monitored. If the validation set performance no longer improves after several consecutive epochs, an early stopping strategy is employed to prevent overfitting. The early stopping patience value is typically set to 10 to 20 training epochs.

[0053] In one embodiment of the present invention, step S5 specifically includes: collecting real-time operating data of the power distribution network, including the current power load, equipment aging status, historical outage counts, and the latest maintenance record text data. The data collection frequency is set according to business needs; for important equipment, it can be set to hourly or higher frequencies, and for general equipment, it can be set to daily frequencies. The real-time data is preprocessed according to the method in step S2. Numerical data is standardized using normalization parameters determined during the training phase to ensure that the numerical range is consistent with the training data. The maintenance record text data is cleaned, segmented, and keyword matched. The comprehensive coupled quantization score is calculated according to the coupled quantization process in step S2. The preprocessed data is used to construct a time-series feature vector sequence according to the method in step S3. The feature data of the most recent T time steps are extracted to form an input sequence, which is then input into the trained dual-path outage prediction model. The model forward propagates to calculate and output the predicted outage probability value. .

[0054] Set early warning threshold When the predicted probability of power outage The system triggers power outage warnings, with warning thresholds calibrated based on business needs and historical data statistics, typically ranging from 0.6 to 0.8. Upon triggering a warning, the system sends warning information to maintenance personnel, including equipment number, equipment type, predicted outage probability, warning time, analysis of key influencing factors, and recommended measures. The analysis of key influencing factors, by examining the contribution of basic risk paths and incremental fault paths, as well as the magnitude of the modulation coefficient, determines whether the power outage risk primarily stems from natural equipment aging or the impact of fault events, providing a basis for maintenance decision-making.

[0055] Different risk levels of warning thresholds can be set according to business needs to achieve tiered warnings. For example, a high-risk warning can be set to... The corresponding emergency measures include immediate repair or load transfer; the medium-risk warning is set as follows: Correspondingly, strengthen inspections or arrange planned maintenance; low-risk areas are designated as... For routine monitoring, no special handling is required. A tiered early warning mechanism enables the rational allocation of operational resources and improves early warning response efficiency.

[0056] Through the above technical solution, this invention achieves deep fusion of maintenance record text data and numerical monitoring data, constructing a dual-path power outage early warning model with physical interpretability. Compared with existing technologies, this invention adopts an innovative coupling-decoupling-recoupling strategy in multi-source data fusion. In the coupling quantization stage, it fully integrates textual and numerical features; in the time-series modeling stage, it decouples features to independently model different types of risks; and in the fusion output stage, it reconstructs the dynamic coupling relationship through learnable modulation coefficients. This approach fully utilizes the complementary information of multimodal data while maintaining the model's interpretability and generalization ability, enabling accurate prediction of power outage events in the distribution network and providing technical support for intelligent operation and maintenance of power systems.

[0057] 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 within the protection scope of the present invention.

Claims

1. A power outage early warning method for distribution networks based on multi-source data fusion, characterized in that, include: S1. Collect historical operating data of the power distribution network, including basic equipment information, operation monitoring data, and maintenance record text data; S2. Preprocess the collected historical operation data, including cleaning and normalizing the numerical data to obtain numerical features, and performing coupled quantization on the maintenance record text data to obtain a comprehensive coupled quantization score. S3. Construct a time series sample set based on the preprocessed data and label it with the corresponding power outage tags; S4. Construct a dual-path power outage prediction model based on long short-term memory network. The dual-path power outage prediction model includes a basic risk path and a fault incremental path. The basic risk path is modeled in time series based on preprocessed numerical features, and the fault incremental path is modeled in time series based on comprehensive coupled quantitative scoring. S5. Collect real-time operation data of the power distribution network, input it into the dual-path power outage prediction model after preprocessing, output the predicted power outage probability value and trigger an early warning.

2. The power outage early warning method for distribution networks based on multi-source data fusion as described in claim 1, characterized in that, In step S1, the basic equipment information includes equipment number, equipment type, commissioning date, and rated lifespan, and the equipment aging degree is calculated based on the commissioning date and rated lifespan; the operation monitoring data includes power load and historical power outages; and the maintenance record text data includes maintenance date, fault description, fault cause, handling measures, and estimated review time.

3. The power outage early warning method for distribution networks based on multi-source data fusion as described in claim 2, characterized in that, Step S2 involves performing coupled quantization processing on the maintenance record text data to obtain a comprehensive coupled quantization score, specifically including: S21. Perform keyword matching on the fault description text in the maintenance record to identify the fault type, and assign a basic sensitivity coefficient to each fault type based on the degree of impact of the fault type on the power outage event. S22. Extract the fault propagation chain by identifying causal keywords in the maintenance records, count the number of nodes in the fault propagation chain, and calculate the propagation weight coefficient based on the length of the fault propagation chain. S23. Construct a numerical data coupling function, which is calculated based on power load, equipment aging level, and historical power outage frequency. S24. Calculate the time decay factor based on the time interval between the maintenance record date and the current predicted date, and calculate the review reminder factor based on the proximity of the current time to the expected review time; S25. Based on the basic sensitivity coefficient, propagation weight coefficient, numerical data coupling function, time decay factor, and review reminder factor, calculate the comprehensive coupling quantitative score for each maintenance record.

4. The power outage early warning method for distribution networks based on multi-source data fusion as described in claim 3, characterized in that, In step S24, the formula for calculating the time decay factor is: ； The formula for calculating the review reminder factor is: ； in, The time decay factor, The time interval between the maintenance record date and the current predicted date. The attenuation coefficient is... To verify the reminder factor, To estimate the review time, For the current time, To verify the influence intensity coefficient, This is the time window width parameter.

5. A power outage early warning method for distribution networks based on multi-source data fusion as described in claim 2, characterized in that, Step S3 includes: At each time step, a comprehensive coupled quantitative score is extracted, which includes power load, equipment aging level, number of historical power outages, and all maintenance records corresponding to that time step, to construct a time-series feature vector sequence. Set early warning time windows according to business needs, and label samples based on whether a power outage event will occur within the future early warning time window; divide all samples into training set, validation set and test set in chronological order.

6. The power outage early warning method for distribution networks based on multi-source data fusion as described in claim 1, characterized in that, Step S4 includes: S41. Decouple the comprehensive coupled quantitative scoring and extract pure fault semantic features; S42. Set up a basic risk path and a fault incremental path in the long short-term memory network. The basic risk path is input with preprocessed numerical features to learn the basic probability of power outage caused by the natural evolution of the equipment's operating status. The fault incremental path is input with pure fault semantic features to learn the incremental contribution of fault records to the power outage risk. S43. Design an adaptive modulation fusion mechanism to dynamically reconstruct the coupling relationship between basic risk and fault incremental risk through modulation coefficients, output the power outage probability prediction value, and obtain a dual-path power outage prediction model.

7. The power outage early warning method for distribution networks based on multi-source data fusion as described in claim 6, characterized in that, In step S41, the decoupling formula is: ； in, For the pure fault semantic features after decoupling, A comprehensive coupled quantitative score for time step t. This represents the electrical load value at time step t. This represents the degree of equipment aging at time step t. This represents the historical number of power outages at time step t. For numerical data coupling functions, The time decay factor, This is a reminder factor for review.

8. The power outage early warning method for distribution networks based on multi-source data fusion as described in claim 6, characterized in that, In step S42, the hidden state update process of the basic risk path is as follows: ； in, The normalized power load is the amount of electricity generated at time step t. The equipment aging degree is normalized to time step t. The number of historical power outages after normalization at time step t. Let be the basic risk hidden state vector at time step t. Long Short-Term Memory (LSTM) network units representing basic risk paths, This is the hidden state from the previous time step; The hidden state update process for the fault incremental path is as follows: ； in, Let be the fault increment hidden state vector at time step t. Long Short-Term Memory (LSTM) network units representing fault incremental paths, This is the hidden state of the previous time step.

9. A power outage early warning method for distribution networks based on multi-source data fusion as described in claim 6, characterized in that, The adaptive modulation fusion mechanism in step S43 specifically involves fusing the hidden states of the two paths using a residual connection approach. The fusion formula is as follows: ； in This represents the hidden state of the overall risks after integration. The hidden state of the basic risk path, This represents the hidden state of the fault incremental path. The modulation coefficient is the modulation coefficient at time step t; the modulation coefficient The input is a concatenated vector obtained through a fully connected network, consisting of the hidden state of the basic risk path, the hidden state of the fault incremental path, the power load, and the degree of equipment aging. The fused comprehensive risk hidden state is then mapped to a predicted power outage probability value through a fully connected layer and an activation function.

10. A power outage early warning method for distribution networks based on multi-source data fusion as described in claim 6, characterized in that, Step S4 also includes training the dual-path power outage prediction model using a time-series sample set. The training process uses a combined loss function to optimize the model parameters. The combined loss function includes the main task loss, path independence constraint loss, and modulation coefficient regularization loss. The main task loss is constructed based on the cross-entropy loss function to optimize the accuracy of power outage probability prediction. The path independence constraint loss is constructed based on the feature correlation metric to ensure the independence between the basic risk path and the pure fault semantic features. The modulation coefficient regularization loss is constructed based on physical constraints to constrain the rationality of the modulation coefficient.

Images