A power distribution network state prediction method and system based on bias attribution and data augmentation

Abstract

This invention discloses a distribution network state prediction method based on deviation attribution and data augmentation, comprising: acquiring historical operating data of the distribution network; using a pre-constructed distribution network state prediction model, identifying deviations in the historical operating data based on an error index to obtain predicted deviation data; performing deviation attribution analysis based on the standardized prediction residuals to obtain deviation attribution results; wherein, the prediction residuals are the difference between the actual observed values ​​and the predicted deviation data values; based on the deviation attribution results, selecting a target strategy from a preset dimension augmentation strategy library to augment the dimensions of the real-time operating data of the distribution network, and evaluating the performance of the distribution network state prediction model after dimension augmentation; dynamically adjusting the parameters of the distribution network state prediction model based on the deviation attribution results and the performance evaluation results to achieve adaptive optimization of the distribution network state prediction model, and using the optimized prediction model to predict the distribution network state in real time.

Description

Technical Field

[0001] This invention relates to a method and system for predicting the state of a distribution network, and more particularly to a method and system for predicting the state of a distribution network based on deviation attribution and data supplementation. Background Technology

[0002] Power system operational status prediction is a crucial link in ensuring the safe, stable, and economical operation of the power grid. With the large-scale integration of flexible resources such as distributed power sources (e.g., photovoltaics), electric vehicles, and user-side energy storage into medium- and low-voltage distribution networks, the operating characteristics of traditional power systems have changed significantly, exhibiting features such as dispersed nodes, large fluctuations in operational status, and strong coupling between load and source characteristics. This presents unprecedented challenges to operational status prediction. Accurate and reliable operational status prediction is fundamental to achieving refined management, optimized scheduling, and improved operational efficiency of the distribution network.

[0003] However, existing machine learning models commonly suffer from prediction bias in predicting the operating status of power systems. This bias is defined as the systematic difference between the model's predicted average and the actual operating value. The root cause often stems from structural defects at the data level, including but not limited to inappropriate feature selection, outliers and noise in the training data, insufficient sample representativeness, and inconsistencies between the model structure and its inherent assumptions and the actual power grid operating mechanism. Particularly in the complex and variable operating environment of medium- and low-voltage distribution networks, data quality is limited by sampling granularity, communication bandwidth, and the operational capabilities of edge nodes, making it more prone to quality issues such as outliers, missing items, and asynchronous data. Furthermore, if the training samples cannot comprehensively cover various complex operating scenarios, structural "blind spots" will appear in the model, leading to severe biases under critical operating conditions (such as load peaks, sudden changes in photovoltaic output, and islanded grid operation). These biases directly result in increased prediction errors, control failures, and unreasonable resource scheduling, thereby seriously affecting the operational safety, dispatch efficiency, and economy of the power system.

[0004] Currently, while various methods exist to mitigate prediction errors in machine learning models, few offer integrated solutions that can systematically and accurately identify the causes of model biases in power system operation state predictions and, based on these bias attributions, selectively supplement relevant data dimensions. Existing methods either focus on addressing single data defects or lack systematic verification of discrepancies between model assumptions and actual operating mechanisms, making it difficult to comprehensively and dynamically adapt to the complex characteristics of highly dynamic, heterogeneous, and locally uncertain distribution networks. Therefore, constructing a system and method capable of deeply attributing prediction biases and dynamically and purposefully supplementing data dimensions based on the attribution results, thereby continuously improving the model's prediction performance and robustness in complex power system environments, has become a pressing technical challenge in this field. Summary of the Invention

[0005] Purpose of the invention: The purpose of this invention is to provide a method and system for predicting the state of distribution networks based on bias attribution and data supplementation, aiming to improve the prediction accuracy, stability and generalization ability of the model in the complex operating environment of medium and low voltage distribution networks, and to solve the problem of systematic prediction bias caused by data structure defects and model structure assumption mismatch.

[0006] Technical Solution: In a first aspect, this embodiment provides a method for predicting the state of a distribution network based on deviation attribution and data supplementation, including:

[0007] Obtain historical operating data of the power distribution network;

[0008] Using a pre-built distribution network status prediction model, deviations are identified in the historical operating data based on error indices to obtain predicted deviation data;

[0009] Bias attribution analysis is performed based on the standardized prediction residuals to obtain the bias attribution results; wherein, the prediction residuals are the difference between the actual observed values ​​and the predicted bias data;

[0010] Based on the deviation attribution results, target strategies are selected from the preset dimension supplementation strategy library to supplement the dimensions of the real-time operation data of the distribution network, and the performance of the distribution network state prediction model after dimension supplementation is evaluated.

[0011] Based on the deviation attribution results and performance evaluation results, the parameters of the distribution network state prediction model are dynamically adjusted to achieve adaptive optimization of the distribution network state prediction model. The optimized prediction model is then used to make real-time predictions of the distribution network state.

[0012] Furthermore, the method also includes cleaning, time synchronization, and standardization preprocessing operations on the operating data, wherein the data includes load, voltage, current, meteorological, and distributed power output data.

[0013] Furthermore, the error indicators include root mean square error, mean absolute error, and mean absolute percentage error.

[0014] Furthermore, the bias attribution analysis based on the standardized prediction residuals specifically includes:

[0015] Time dimension identification and analysis: The mean and volatility index of time series residuals are calculated by using a sliding window to construct a residual variation trend function;

[0016] Spatial dimension identification and analysis: Error intensity maps are drawn by constructing node residual statistics and combining them with GIS spatial indexes;

[0017] Control behavior identification analysis: This involves introducing control behavior variables and testing the statistical significance of differences in the residual mean under control event conditions;

[0018] Boundary state response identification and analysis: By defining a set of boundary state scenarios and comparing the prediction performance of the prediction model under boundary states and normal states, it is determined whether the response of the prediction model under boundary states is ineffective.

[0019] The deviation attribution result is used to determine the source of the deviation in the prediction model, including one or more of the following: time dimension, spatial dimension, control behavior, and boundary state.

[0020] Furthermore, the selection of target strategies for dimensional supplementation includes:

[0021] Missing mechanism completion dimension: including introducing cloud cover rate, constructing weather variability rate to express the impact of sudden weather changes, and introducing human heat load index as load response variable;

[0022] State condition expression dimension: including introducing holiday tags to determine whether it is a holiday, introducing public event calendar variables, and introducing topological level identifiers;

[0023] Multidimensional refinement dimensions: including breaking down the load into: The weather status coding is extended to one-hot coding;

[0024] Through RMSE, MAPE, Evaluation metrics, SHAP value ranking, and error heatmaps are used to assess the performance of the distribution network condition prediction model after dimensional augmentation.

[0025] Secondly, this embodiment provides a distribution network state prediction system based on deviation attribution and data supplementation, including:

[0026] Data acquisition module: used to acquire historical operating data of the power distribution network;

[0027] Data screening module: used to identify deviations in the historical operating data based on error indicators using a pre-built distribution network status prediction model to obtain predicted deviation data;

[0028] Deviation Attribution Analysis Module: Used to perform deviation attribution analysis based on the standardized prediction residuals to obtain deviation attribution results; wherein, the prediction residuals are the difference between the actual observed values ​​and the predicted deviation data;

[0029] Dimension augmentation module: Based on the deviation attribution results, it selects a target strategy from a preset dimension augmentation strategy library to augment the dimensions of the real-time operation data of the distribution network, and evaluates the performance of the distribution network state prediction model after dimension augmentation.

[0030] Real-time prediction module: Based on deviation attribution results and performance evaluation results, it dynamically adjusts the parameters of the distribution network state prediction model to achieve adaptive optimization of the distribution network state prediction model, and makes real-time predictions of the distribution network state through the optimized prediction model.

[0031] Furthermore, the data acquisition module also includes preprocessing operations such as cleaning, time synchronization, and standardization of the operating data, including load, voltage, current, meteorological data, and distributed power output data.

[0032] Furthermore, the error indicators include root mean square error, mean absolute error, and mean absolute percentage error.

[0033] Furthermore, the bias attribution analysis based on the standardized prediction residuals specifically includes:

[0034] Time dimension identification and analysis: The mean and volatility index of time series residuals are calculated by using a sliding window to construct a residual variation trend function;

[0035] Spatial dimension identification and analysis: Error intensity maps are drawn by constructing node residual statistics and combining them with GIS spatial indexes;

[0036] Control behavior identification analysis: This involves introducing control behavior variables and testing the statistical significance of differences in the residual mean under control event conditions;

[0037] Boundary state response identification and analysis: By defining a set of boundary state scenarios and comparing the prediction performance of the prediction model under boundary states and normal states, it is determined whether the response of the prediction model under boundary states is ineffective.

[0038] The deviation attribution result is used to determine the source of the deviation in the prediction model, including one or more of the following: time dimension, spatial dimension, control behavior, and boundary state.

[0039] Furthermore, the selection of target strategies for dimensional supplementation includes:

[0040] Missing mechanism completion dimension: including introducing cloud cover rate, constructing weather variability rate to express the impact of sudden weather changes, and introducing human heat load index as load response variable;

[0041] State condition expression dimension: including introducing holiday tags to determine whether it is a holiday, introducing public event calendar variables, and introducing topological level identifiers;

[0042] Multidimensional refinement dimensions: including breaking down the load into: The weather status coding is extended to one-hot coding;

[0043] Through RMSE, MAPE, Evaluation metrics, SHAP value ranking, and error heatmaps are used to assess the performance of the distribution network condition prediction model after dimensional augmentation.

[0044] Thirdly, this embodiment provides a computer device including one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of the aforementioned distribution network state prediction method based on deviation attribution and data supplementation.

[0045] Fourthly, this embodiment provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned distribution network state prediction method based on deviation attribution and data supplementation.

[0046] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages:

[0047] (1) This invention uses multi-dimensional residual attribution analysis to accurately extract deviation patterns and construct a set of deviation scenarios, overcoming the shortcomings of traditional methods in complex distribution network environments, such as incomplete deviation attribution and lack of systematic verification. Secondly, based on the deviation attribution results, a targeted data dimension supplementation strategy is implemented to effectively compensate for the prediction deviation caused by data structure defects and model assumption mismatch, thereby enhancing the model's ability to capture key operating mechanisms and its generalization performance.

[0048] (2) This invention achieves dynamic model selection, parameter adjustment and multi-model collaboration by using multi-dimensional quantitative evaluation of residual mean reduction ratio, error space heat map, SHAP value change and business indicators, and combined with feedback-driven structural adaptive mechanism, thereby improving the adaptability, stability and interpretability of the model in the highly dynamic and multi-source heterogeneous distribution network operation environment.

[0049] (3) This invention can be widely applied to multiple scenarios such as load forecasting, distributed power output modeling, voltage over-limit risk identification, and energy storage control strategy optimization. It can significantly improve the accuracy of power system operation status prediction, dispatch hit rate and boundary condition stability, and provide reliable technical support for the safe, economical and efficient operation of the distribution network. It has outstanding practical value and promotion prospects. Attached Figure Description

[0050] Figure 1 This is a flowchart of the prediction bias attribution identification and data dimension supplementation iterative optimization process of the present invention. Detailed Implementation

[0051] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0052] Example 1

[0053] like Figure 1 As shown, this embodiment provides a distribution network state prediction method based on deviation attribution and data supplementation, including:

[0054] (1) Obtain historical operating data of the power distribution network;

[0055] It connects to power system SCADA systems, smart meters, distributed power (such as photovoltaic and wind power) monitoring systems, and meteorological monitoring systems through communication interfaces to collect various operational data in real time or at regular intervals. This includes, but is not limited to, load data, distributed power output data, voltage, current, power flow, meteorological data (temperature, humidity, irradiance, wind speed, etc.), and related historical operational data. The data is preprocessed, including cleaning (such as filling missing values ​​and removing duplicate values), data synchronization, standardization, or normalization, to form historical operational data that can be used later.

[0056] (2) Using a pre-built distribution network status prediction model, the historical operating data is deviated based on the error index to obtain the predicted deviation data;

[0057] Based on the historical operational data, evaluation metrics for the prediction model are calculated, and a subset of data with poor prediction performance is identified. Commonly used evaluation metrics include:

[0058] Root Mean Square Error (RMSE):

[0059]

[0060] Mean Absolute Error (MAE):

[0061]

[0062] Mean Absolute Percentage Error (MAPE):

[0063]

[0064] in, For the sample size, For the true value, These are predicted values. By calculating the above indicators on data subsets with different feature dimensions (such as different load levels, different photovoltaic output ranges, and different weather types) or time periods (such as peak, off-peak, and holidays), it is possible to accurately identify specific scenarios with significant prediction deviations and label these data as deviation datasets.

[0065] Based on the aforementioned bias dataset, potential sources of model bias are diagnosed. The specific implementation of the bias diagnosis is as follows:

[0066] (21) Feature Validity Evaluation: Feature validity evaluation refers to the systematic analysis of each feature in the dataset to determine its predictive contribution and actual influence on the target variable (such as the prediction target, classification label, etc.). In power system operation status prediction, identifying features with high predictive ability, low redundancy, and strong stability can significantly improve the generalization performance and interpretability of the model, and avoid biases introduced by irrelevant or redundant features. This embodiment mainly adopts three key evaluation methods to adapt to different scenarios:

[0067] (211) The Pearson correlation coefficient is used to measure the strength and direction of the linear relationship between two continuous variables. The Pearson correlation coefficients between each input feature (e.g., temperature, photovoltaic irradiance) and the target variable (e.g., load, distributed power output) are calculated. The formula is as follows:

[0068]

[0069] in, for and covariance, and They are respectively and The standard deviation of the correlation coefficient. The range of the correlation coefficient is... The closer the absolute value is to 1, the stronger the linear relationship and the higher the feature effectiveness.

[0070] (212) The Gini Index is used to measure the improvement in data purity of features when constructing decision trees or similar tree models. It is mainly used for classification tasks, but its variants are also applicable to regression trees. Its calculation formula is:

[0071]

[0072] in, Features In the The probability of a feature in a class (or discretized group). The smaller the Gini index, the stronger the feature. In classification tasks, the higher the purity of the segmentation, the stronger the ability to distinguish the target variable.

[0073] (213) Permutation importance: Train the model and calculate the baseline error; randomly perturb the value of each feature (shuffle the order), and calculate the error after perturbation again; the feature importance is equal to the difference between the baseline error and the error after perturbation. The greater the difference, the more important the feature.

[0074] The above evaluation method can identify effective features that contribute significantly to the prediction of power system operating status, and mark or initially remove features with low contribution or high redundancy, thereby optimizing the feature set and improving the accuracy, stability and interpretability of the model.

[0075] (22) Outlier and Noise Detection: Outliers are data points in the dataset that significantly deviate from other observations, while noise is random error introduced into the data due to measurement errors, sensor malfunctions, transmission interference, etc. During the acquisition of power system load, output, and other operational data, sensor malfunctions or communication anomalies may introduce outliers and noise. If left untreated, these will seriously affect the accuracy and stability of the prediction model. This embodiment performs outlier detection on the values ​​of each feature:

[0076] (221) Z-Score method based on statistical data: Calculate the multiple of the standard deviation of each data point from the mean. For data points... The Z-Score calculation formula is as follows:

[0077]

[0078] in, For data points, The mean of the dataset. Let be the standard deviation of the dataset. If... If the value exceeds the preset threshold, the system will consider it an outlier.

[0079] (222) The decision tree structure is constructed by randomly selecting features and randomly splitting data using the isolated forest algorithm. Outliers are usually more likely to be isolated in the tree due to their sparsity, i.e., the path length is shorter.

[0080] (223) Noise filtering of time series data: For random noise that is difficult to completely remove from time series data, this system can use signal processing methods to filter it, such as moving average, median filtering, Kalman filtering or wavelet denoising, to smooth the data and improve data quality.

[0081] The above detection and filtering methods can effectively identify and process outliers and noise in power system operation data, ensuring the cleanliness and reliability of input data and laying the foundation for subsequent predictive modeling.

[0082] (23) Sample representativeness assessment: This refers to the process of verifying whether the training sample data can accurately reflect the overall characteristics of the actual operation of the power system. In power system operation status prediction, due to the complex and variable data acquisition environment and the strong randomness of distributed resource output, if the training sample is not representative, it may lead to a decrease in the generalization performance of the model for unseen scenarios in practical applications, thereby introducing systematic bias. In this embodiment, the Kullback-Leibler (KL) divergence analysis method is used to compare the distribution of key features of the training sample with the distribution of actual operating data (such as recent real-time operating data, specific extreme operating condition data, etc.) to quantify the representativeness of the sample.

[0083] KL divergence is used to measure the difference between two probability distributions. In assessing sample representativeness, KL divergence can be used to compare the distributions of training samples. Distribution of actual operating data The differences between these distributions (in key characteristic dimensions such as load, distributed output, voltage, etc.) are estimated using methods such as histograms or kernel density estimation (KDE), and the KL divergence is calculated.

[0084]

[0085] in, The probability distribution of training samples along a specific feature dimension. This represents the probability distribution of the actual operating data on this dimension. For the set of all possible values. If A large value indicates that the training samples are not representative enough.

[0086] (24) Model Hypothesis Verification: This embodiment will rigorously verify the basic assumptions upon which the constructed prediction model relies. The degree to which these assumptions are satisfied directly affects the validity of the model estimation results and the reliability of the predictions. This includes:

[0087] (241) Verification of the linear relationship hypothesis: This involves testing whether a linear relationship exists between the independent and dependent variables. This is done by plotting scatter plots of the key independent and dependent variables on a computer system, or by creating a component-plus-residual plot (also known as a partial regression plot), to visually observe whether the data points exhibit a linear trend or whether there is an obvious non-linear pattern. If a non-linear curve trend appears in the plot, it indicates that the linear hypothesis is not valid.

[0088] (242) Homoscedasticity Assumption Verification: Test whether the variance of the model's error term is consistent across all observations (i.e., homoscedasticity). First, calculate the model's residuals. It is defined as the actual observed value. Compared with model predictions Difference:

[0089]

[0090] Next, a scatter plot of the residuals versus the fitted values ​​(or the residuals versus the independent variables) will be plotted. If the residuals are uniformly distributed without a clear pattern or fan-shaped trend, it indicates that homoscedasticity is satisfied. For a more rigorous statistical test, the Breusch-Pagan test will also be performed. This test uses an auxiliary regression model, with the squared residuals as the dependent variable and the original model's independent variables as the independent variables, to determine heteroscedasticity. The formula for its test statistic, BP, is as follows:

[0091]

[0092] in, For sample size, The coefficient of determination for the auxiliary regression model is given. If the p-value corresponding to the BP statistic is less than the preset significance level, the homoscedasticity hypothesis is rejected, indicating the existence of heteroscedasticity.

[0093] (243) Independence Hypothesis Verification: This involves testing whether the error terms in the model are independent of each other. For time series data commonly used in power system forecasting, it is necessary to detect the presence of autocorrelation. The autocorrelation function (ACF) and partial autocorrelation function (PACF) plots of the residuals will be calculated and plotted to identify serial correlation in the residuals. Furthermore, the Durbin-Watson test will be performed. First-order autocorrelation is detected by comparing the differences between adjacent residuals; its statistic... The calculation formula is:

[0094]

[0095] in, Let be the residual of the t-th observation. The Durbin-Watson statistic typically ranges from 0 to 4, with values ​​close to 2 indicating no first-order autocorrelation, less than 2 indicating positive correlation, and greater than 2 indicating negative correlation. The independence of the residuals will be determined based on the calculated d-value and the corresponding critical value.

[0096] The above verification steps allow for a comprehensive assessment of whether the model's assumptions hold true. If any assumption is found to be false, it indicates that the model may have structural biases or defects. This will serve as an important component of the bias diagnosis report, guiding subsequent strategies for bias attribution and data dimension supplementation.

[0097] (3) Perform bias attribution analysis based on the standardized prediction residuals to obtain bias attribution results; wherein, the prediction residuals are the difference between the actual observed values ​​and the prediction bias data;

[0098] The implementation of residual standardization processing introduces a standardized residual form, which facilitates the comparative analysis of residuals between different transformer areas and time periods.

[0099] There are actual observation values. Compared with model predictions The residual is then defined as:

[0100]

[0101] To facilitate the comparison of residuals across different transformer areas and time periods, a standardized residual format is introduced:

[0102]

[0103] in, Represents the true value The standard deviation in the sample set, For standardized residuals.

[0104] The deviation attribution analysis includes:

[0105] The implementation of time pattern recognition calculates the mean and volatility of time series residuals using a sliding window to identify the model's inadequacy in adaptability under abrupt changes.

[0106] By calculating the mean and volatility indices of the time series residuals using a sliding window, a residual variation trend function is constructed:

[0107]

[0108]

[0109] in, The width of the time window. The window start time.

[0110] Significant peak or trough intervals correspond to abrupt changes in operating conditions, indicating that the model is not adaptable enough to critical operating conditions such as peak loads and photovoltaic ramp-up.

[0111] The implementation of spatial pattern recognition and error geographic mapping involves constructing node residual statistics and combining them with GIS spatial indexes to draw error intensity maps, thereby identifying weakly robust areas of the model in spatial distribution.

[0112] Construct residual statistics for each node (or transformer area):

[0113]

[0114] By combining GIS spatial indexes, error intensity maps were drawn. ,in:

[0115]

[0116] The map reflects the "weakly robust regions" of the model in terms of spatial distribution.

[0117] The implementation of operational behavior association pattern recognition involves introducing control behavior variables and examining the statistical significance differences in the residual mean under control event conditions to identify the influence of control behaviors not captured by the model and guide feature addition.

[0118] Introduce control behavior variables (such as energy storage start-stop signals) Construct the mean residuals under control event conditions:

[0119]

[0120] If a statistically significant difference exists (verified by t-test or KS test), it indicates that the model has not captured the impact of control behavior on the state, and control features should be added.

[0121] The implementation of boundary state response capability identification is to define a set of boundary state scenarios and compare the difference in prediction performance of the model in boundary state and normal state in order to determine whether the model's response in boundary state is ineffective.

[0122] Define boundary state scenario set (e.g., high temperature, high load, low voltage):

[0123]

[0124] Comparison models in Differences in prediction performance compared to normal conditions:

[0125]

[0126] like (Tolerance threshold) indicates that the model fails to respond in the boundary state.

[0127] The deviation attribution result is used to determine the source of the deviation in the prediction model, including one or more of the following: time dimension, spatial dimension, control behavior, and boundary state.

[0128] Based on the identified bias attribution results, targeted model enhancement strategies are constructed, including introducing high-resolution meteorological indices, constructing slope features, introducing explicit control variables, and supplementing historical samples or simulation data.

[0129] Based on the identified bias attribution results, a model enhancement strategy is constructed:

[0130] If the deviation is concentrated in "hot and humid weather", it is recommended to introduce a high-resolution meteorological index;

[0131] If the deviation is concentrated in the "photovoltaic ramp-up phase", it is recommended to construct a slope feature. ;

[0132] If the deviation is concentrated in "user-controlled behavior", then explicit control variables need to be introduced. ;

[0133] If the residuals in a specific area remain high for a long period of time, it is necessary to supplement the area with historical samples or simulation data.

[0134] (4) Based on the deviation attribution results, select the target strategy from the preset dimension supplementation strategy library to supplement the dimensions of the real-time operation data of the distribution network, and evaluate the performance of the distribution network state prediction model after dimension supplementation.

[0135] The performance of the distribution network state prediction model after dimensionality augmentation is evaluated using the residual mean reduction ratio and error hotspot attenuation maps. After each feature dimension augmentation, this embodiment verifies whether the model's performance is improved under the original bias scenario. Commonly used evaluation metrics include:

[0136] Residual mean reduction ratio:

[0137]

[0138] in, The mean of the residuals before feature addition. The mean of the residuals after feature addition.

[0139] Hotspot area error decay map: plotted before and after supplementation and It is used to visualize the convergence of the error distribution.

[0140] The correlation data dimension supplementation module in this application receives the bias attribution results and, based on the findings of the bias attribution analysis, implements a targeted data dimension supplementation strategy to enhance the model's ability to identify key mechanisms of system operation and improve its generalization performance and robustness under complex working conditions.

[0141] The implementation method of residual structure attribution identification identifies performance degradation regions in specific scenarios based on the residual distribution of the prediction model and constructs a set of deviation scenarios. Based on the residual distribution of the prediction model, this embodiment identifies performance degradation regions of the model in specific scenarios and constructs a "set of deviation scenarios". For example, photovoltaic power output fluctuates drastically in cloudy weather; the aggregated load of electric vehicles deviates periodically within the day; and there are concentrated false alarms in the boundary samples during voltage over-limit prediction.

[0142] The residual scenario can be constructed using a multidimensional residual heatmap, combined with the residual mean index:

[0143]

[0144] The implementation method for dimension augmentation path identification determines and implements four types of dimension augmentation strategies based on deviation pattern characteristics: missing mechanism completion, state condition expression, multi-dimensional refinement, and interactive combination construction. Based on deviation pattern characteristics, this embodiment determines the applicable dimension augmentation types, categorizing them into four types:

[0145] (a) Missing mechanism completion dimension: used to enhance unmodeled but strongly correlated driving factors with error.

[0146] Features introduced or enhanced in this embodiment include, but are not limited to: Example 1: Introducing cloud cover rate Enhance the dynamic representation of photovoltaic shading; Example 2: Construct weather variability:

[0147]

[0148] in Let be the temperature at time t, used to express the impact of sudden weather changes. Example 3: Introducing the Human Heat Load Index (THI) as a load response variable:

[0149]

[0150] In the formula Temperature in Celsius This refers to relative humidity.

[0151] (b) State-conditional dimension: context variables used to provide modeling conditions.

[0152] Features introduced or enhanced in this application's embodiments include, but are not limited to: introducing holiday tags. Indicates whether it is a holiday; introduces public event calendar variables, such as major sporting events and electricity price jump dates; introduces topology hierarchy identifiers. Such as the relative positions of busbars, feeders, and transformers.

[0153] (c) Multidimensional refinement dimension: used to split coarse-grained features and refine them to distinguishable subclasses.

[0154] Features introduced or enhanced in the embodiments of this application include, but are not limited to: user load Split into:

[0155]

[0156] Weather status coding has been extended to one-hot coding:

[0157]

[0158] Automatically generate labels by clustering transformer area features using tools such as KMeans / DBSCAN. .

[0159] (d) Interactive Combinatorial Stereotype Dimension: Used to construct combinations of variables to enhance the expression of nonlinear structures.

[0160] Features introduced or enhanced in this embodiment include, but are not limited to: photovoltaic power output efficiency index.

[0161]

[0162] in, For output power, Irradiance, For component area. Real-time adjustment of elasticity factor:

[0163]

[0164] in Indicates adjustable load capacity. Fast response indicators:

[0165]

[0166] This indicates a trigger signal for a sudden surge in demand caused by rapid changes in load combined with temperature fluctuations.

[0167] The evaluation of the performance of the distribution network state prediction model after dimensional augmentation specifically includes:

[0168] The effectiveness of feature dimension augmentation is verified and dimension selection is performed by comparing prediction model performance metrics, SHAP value ranking, and error heatmaps. This application's embodiments verify the effect of feature augmentation by comparing model performance before and after training.

[0169] Commonly used evaluation metrics include: RMSE, MAPE, Evaluation metrics were used; the importance of new features was ranked using SHAP values; and the improvement in response in deviation areas was visualized and compared using error heatmaps.

[0170] This embodiment comprehensively verifies the effect of feature addition by introducing before-and-after performance comparison analysis, error structure re-clustering analysis, feature importance change analysis, and business-side indicator evaluation. It quantifies the effectiveness of the added features in improving model prediction accuracy, suppressing error hotspots, changing feature contribution, and enhancing business usability. The accuracy indicators of the model before and after adding features are quantitatively compared on the training and validation sets.

[0171] Commonly used evaluation metrics include: Root Mean Square Error (RMSE):

[0172]

[0173] Mean Absolute Percentage Error (MAPE):

[0174]

[0175] Coefficient of determination ( ):

[0176]

[0177] The above metrics are used to measure performance changes after feature expansion.

[0178] The implementation of error structure re-clustering analysis aims to evaluate whether the supplementary features effectively suppress error hotspot regions by performing spatial or temporal pattern clustering on the model residuals. In this application, embodiments perform spatial or temporal pattern clustering on the model residuals and compare whether "error hotspot regions" are suppressed before and after supplementation.

[0179] In this application, the embodiments use algorithms such as K-Means and DBSCAN to perform subset clustering on the residual matrix.

[0180] The implementation method of introducing feature importance transition analysis aims to rank and compare the contribution of features before and after the addition using SHAP values ​​to assess whether the added features have stable explanatory power. This application embodiment uses SHAP (SHapley Additive exPlanations) to rank and compare the contribution of all features before and after the addition.

[0181] If the supplementary feature significantly improves the SHAP value and enters the Top-k, it indicates that it has stable explanatory power.

[0182]

[0183] in Representation of features marginal contribution This represents the model output under the feature subset S.

[0184] The implementation method for evaluating business-side indicators aims to demonstrate that the model improves business applicability in key scenarios by introducing key business indicators such as scheduling hit rate, load reconfiguration accuracy, and boundary condition prediction stability. This application's embodiments introduce key business indicators such as scheduling hit rate, load reconfiguration accuracy, and boundary condition prediction stability.

[0185] (5) Based on the deviation attribution results and performance evaluation results, dynamically adjust the parameters of the distribution network state prediction model to achieve adaptive optimization of the distribution network state prediction model, and make real-time predictions of the distribution network state through the optimized prediction model.

[0186] By matching and mapping scene labels, combined with sliding window residual monitoring and error threshold setting, the model structure switching and parameter fine-tuning are dynamically triggered. It supports a multi-model collaborative mechanism including linear regression, tree model, deep neural network, and graph neural network to adaptively optimize the distribution network status prediction model. The optimized prediction model is used to make real-time predictions of the distribution network status.

[0187] Example 2

[0188] This embodiment also provides a distribution network state prediction system based on deviation attribution and data supplementation, including:

[0189] Data acquisition module: used to acquire historical operating data of the power distribution network;

[0190] Data screening module: used to identify deviations in the historical operating data based on error indicators using a pre-built distribution network status prediction model to obtain predicted deviation data;

[0191] Deviation Attribution Analysis Module: Used to perform deviation attribution analysis based on the standardized prediction residuals to obtain deviation attribution results; wherein, the prediction residuals are the difference between the actual observed values ​​and the predicted deviation data;

[0192] Dimension augmentation module: Based on the deviation attribution results, it selects a target strategy from a preset dimension augmentation strategy library to augment the dimensions of the real-time operation data of the distribution network, and evaluates the performance of the distribution network state prediction model after dimension augmentation.

[0193] Real-time prediction module: Based on deviation attribution results and performance evaluation results, it dynamically adjusts the parameters of the distribution network state prediction model to achieve adaptive optimization of the distribution network state prediction model, and makes real-time predictions of the distribution network state through the optimized prediction model.

[0194] Example 3

[0195] This embodiment also provides a computer device, including one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of the aforementioned distribution network state prediction method based on deviation attribution and data supplementation.

[0196] Example 4

[0197] This embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned distribution network state prediction method based on deviation attribution and data supplementation.

Claims

1. A method for predicting the state of a distribution network based on deviation attribution and data supplementation, characterized in that, include: Obtain historical operating data of the power distribution network; Using a pre-built distribution network status prediction model, deviations are identified in the historical operating data based on error indices to obtain predicted deviation data; Bias attribution analysis is performed based on the standardized prediction residuals to obtain the bias attribution results; wherein, the prediction residuals are the difference between the actual observed values ​​and the predicted bias data; Based on the deviation attribution results, target strategies are selected from the preset dimension supplementation strategy library to supplement the dimensions of the real-time operation data of the distribution network, and the performance of the distribution network state prediction model after dimension supplementation is evaluated. Based on the deviation attribution results and performance evaluation results, the parameters of the distribution network state prediction model are dynamically adjusted to achieve adaptive optimization of the distribution network state prediction model. The optimized prediction model is then used to make real-time predictions of the distribution network state.

2. The distribution network state prediction method based on deviation attribution and data supplementation according to claim 1, characterized in that, The method also includes cleaning, time synchronization, and standardization preprocessing of the operating data, which includes load, voltage, current, meteorological, and distributed power output data.

3. The distribution network state prediction method based on deviation attribution and data supplementation according to claim 1, characterized in that, The error metrics include root mean square error, mean absolute error, and mean absolute percentage error.

4. The distribution network state prediction method based on deviation attribution and data supplementation according to claim 1, characterized in that, The bias attribution analysis based on the standardized prediction residuals specifically includes: Time dimension identification and analysis: The mean and volatility index of time series residuals are calculated by using a sliding window to construct a residual variation trend function; Spatial dimension identification and analysis: Error intensity maps are drawn by constructing node residual statistics and combining them with GIS spatial indexes; Control behavior identification analysis: This involves introducing control behavior variables and testing the statistical significance of differences in the residual mean under control event conditions; Boundary state response identification and analysis: By defining a set of boundary state scenarios and comparing the prediction performance of the prediction model under boundary states and normal states, it is determined whether the response of the prediction model under boundary states is ineffective. The deviation attribution result is used to determine the source of the deviation in the prediction model, including one or more of the following: time dimension, spatial dimension, control behavior, and boundary state.

5. The distribution network state prediction method based on deviation attribution and data supplementation according to claim 1, characterized in that, The target selection strategy for dimensional supplementation includes: Missing mechanism completion dimension: including introducing cloud cover rate, constructing weather variability rate to express the impact of sudden weather changes, and introducing human heat load index as load response variable; State condition expression dimension: including introducing holiday tags to determine whether it is a holiday, introducing public event calendar variables, and introducing topological level identifiers; Multidimensional refinement dimensions: including breaking down the load into: The weather status coding is extended to one-hot coding; Through RMSE, MAPE, Evaluation metrics, SHAP value ranking, and error heatmaps are used to assess the performance of the distribution network condition prediction model after dimensional augmentation.

6. A distribution network state prediction system based on deviation attribution and data supplementation, characterized in that, include: Data acquisition module: used to acquire historical operating data of the power distribution network; Data screening module: used to identify deviations in the historical operating data based on error indicators using a pre-built distribution network status prediction model to obtain predicted deviation data; Deviation Attribution Analysis Module: Used to perform deviation attribution analysis based on the standardized prediction residuals to obtain deviation attribution results; wherein, the prediction residuals are the difference between the actual observed values ​​and the predicted deviation data; Dimension augmentation module: Based on the deviation attribution results, it selects a target strategy from a preset dimension augmentation strategy library to augment the dimensions of the real-time operation data of the distribution network, and evaluates the performance of the distribution network state prediction model after dimension augmentation. Real-time prediction module: Based on deviation attribution results and performance evaluation results, it dynamically adjusts the parameters of the distribution network state prediction model to achieve adaptive optimization of the distribution network state prediction model, and makes real-time predictions of the distribution network state through the optimized prediction model.

7. A distribution network state prediction system based on deviation attribution and data supplementation according to claim 6, characterized in that, The data acquisition module also includes preprocessing operations such as cleaning, time synchronization, and standardization of the operating data, which includes load, voltage, current, meteorological data, and distributed power output data.

8. A distribution network state prediction system based on deviation attribution and data supplementation according to claim 6, characterized in that, The error metrics include root mean square error, mean absolute error, and mean absolute percentage error.

9. A distribution network state prediction system based on deviation attribution and data supplementation according to claim 6, characterized in that, The bias attribution analysis based on the standardized prediction residuals specifically includes: Time dimension identification and analysis: The mean and volatility index of time series residuals are calculated by using a sliding window to construct a residual variation trend function; Spatial dimension identification and analysis: Error intensity maps are drawn by constructing node residual statistics and combining them with GIS spatial indexes; Control behavior identification analysis: This involves introducing control behavior variables and testing the statistical significance of differences in the residual mean under control event conditions; Boundary state response identification and analysis: By defining a set of boundary state scenarios and comparing the prediction performance of the prediction model under boundary states and normal states, it is determined whether the response of the prediction model under boundary states is ineffective. The deviation attribution result is used to determine the source of the deviation in the prediction model, including one or more of the following: time dimension, spatial dimension, control behavior, and boundary state.

10. A distribution network state prediction method based on deviation attribution and data supplementation according to claim 6, characterized in that, The target selection strategy for dimensional supplementation includes: Missing mechanism completion dimension: including introducing cloud cover rate, constructing weather variability rate to express the impact of sudden weather changes, and introducing human heat load index as load response variable; State condition expression dimension: including introducing holiday tags to determine whether it is a holiday, introducing public event calendar variables, and introducing topological level identifiers; Multidimensional refinement dimensions: including breaking down the load into: The weather status coding is extended to one-hot coding; Through RMSE, MAPE, Evaluation metrics, SHAP value ranking, and error heatmaps are used to assess the performance of the distribution network condition prediction model after dimensional augmentation.

11. A computer device, characterized in that, It includes one or more processors, a memory, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the programs, when executed by the processors, implement the steps of a distribution network state prediction method based on deviation attribution and data supplementation as described in any one of claims 1-5.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the distribution network state prediction method based on deviation attribution and data supplementation as described in any one of claims 1-5.

Images