Power grid business performance evaluation method and system

Abstract

The invention discloses a power grid business performance evaluation method and system, and the method comprises the steps: carrying out the filtering and formatting processing of a historical business data flow of a target power grid, and obtaining a dynamic business data set; constructing an initial performance evaluation model, and training the initial performance evaluation model based on the dynamic service data set to obtain a first performance evaluation model; extracting time sequence features and business change features in the dynamic business data set, and generating dynamic business change trend data; matching the dynamic business change trend data with a business rule base to obtain the change amplitude of the business index; if the variation amplitude exceeds a threshold range, calculating a weight coefficient of the business index, performing weighting processing on the dynamic business data set, and training the first performance evaluation model based on the obtained weighted dynamic business data set to obtain a second performance evaluation model; and generating a performance evaluation result according to the second performance evaluation model. According to the method, the timeliness of a performance management system is improved.

Description

Technical Field

[0001] The present invention relates to the field of information processing technology, and in particular to a power grid service performance evaluation method and system. Background Art

[0002] Performance management of power grid companies is a crucial research area in the digital transformation of the energy industry. It is directly related to operational efficiency, resource allocation optimization, and the scientific nature of decision support. With the development of smart grid technology, power grid business data is characterized by high-frequency collection, multi-dimensional correlation, and dynamic changes.

[0003] Existing methods for power grid performance evaluation mostly rely on artificially defined fixed indicator systems, which are difficult to adapt to the rapid changes in power grid business scenarios and the real-time data requirements, resulting in performance strategy adjustments that are difficult to reflect the actual operating status.

[0004] It can be seen that how to improve the existing power grid business performance evaluation method and enhance the timeliness of the performance management system has become a technical problem that needs to be solved urgently by those skilled in the art. Summary of the Invention

[0005] The present invention provides a power grid business performance evaluation method and system to solve the technical problem of improving the existing power grid business performance evaluation method, so as to achieve the effect of improving the timeliness of the performance management system.

[0006] In order to solve the above technical problems, an embodiment of the present invention provides a power grid service performance evaluation method, comprising:

[0007] The acquired historical business data stream of the target power grid is filtered and formatted in sequence to obtain a structured dynamic business data set;

[0008] Constructing an initial performance evaluation model based on a random forest classification algorithm, and training the initial performance evaluation model based on the dynamic business data set to obtain a first performance evaluation model;

[0009] Extracting time series features and business change features from the dynamic business data set, and generating dynamic business change trend data based on the time series features and the business change features;

[0010] Matching the dynamic business change trend data with a pre-built business rule library to obtain the change range of each business indicator in the dynamic business change trend data;

[0011] If the variation exceeds a preset threshold range, the weight coefficients of the business indicators are calculated based on a linear regression algorithm, and the dynamic business data set is weighted according to the weight coefficients to obtain a weighted dynamic business data set;

[0012] Incrementally training the first performance evaluation model based on the weighted dynamic business data set to obtain an updated second performance evaluation model;

[0013] The acquired real-time business data stream of the target power grid is input into the second performance evaluation model to generate a performance evaluation result of the real-time business data stream.

[0014] As one preferred solution, the extracting of time series features and business change features from the dynamic business data set includes:

[0015] Based on the STL algorithm, the continuous monitoring data in the dynamic business data set is decomposed at multiple scales to obtain continuous parameter decomposition characteristics including long-term trend components, seasonal components and residual components;

[0016] Extracting parameter fluctuation characteristics within a preset time period before and after a discrete business event in the dynamic business data set by a sliding window method;

[0017] The continuous parameter decomposition feature and the parameter fluctuation feature are time-aligned to obtain a time series feature.

[0018] As one preferred solution, the extracting of time series features and business change features from the dynamic business data set includes:

[0019] Performing frequency domain analysis on the time series characteristics according to a Fourier transform algorithm, and determining significant frequency components obtained according to the analysis as periodic variation characteristics;

[0020] Perform abnormal monitoring on the continuous monitoring data based on the CUSUM algorithm, and determine the parameter range and change rate before and after the extracted abnormal mutation point as the abnormal mutation feature;

[0021] The periodic change feature and the abnormal mutation feature are normalized to generate a multi-dimensional business change feature vector.

[0022] As one preferred solution, generating dynamic business change trend data based on the time series characteristics and the business change characteristics includes:

[0023] Inputting the time series characteristics and the business change characteristics into a pre-built ETS model to predict the business parameter change trend in a preset time period, and obtaining short-term trend data including trend values and confidence intervals;

[0024] Calculate the feature similarity between the business scenario in the forecast time period and the historical typical scenario using a dynamic time warping algorithm to obtain a typical pattern label for the business scenario;

[0025] The short-term trend data is fused with the typical pattern label to generate the dynamic business change trend data.

[0026] As one preferred solution, the incremental training of the first performance evaluation model based on the weighted dynamic business data set to obtain an updated second performance evaluation model includes:

[0027] Dividing the weighted dynamic service data set into a training data set and a validation data set based on a preset division ratio;

[0028] Incrementally training the first performance evaluation model based on the training data set, wherein, during the training process, decision tree parameters in a random forest classification algorithm are adjusted to adapt to changes in the importance of business indicators in the weighted dynamic business data set;

[0029] Using the validation set to perform real-time evaluation on the first performance evaluation model during training, the evaluation includes determining the performance of the model by calculating an evaluation indicator;

[0030] When the evaluation index reaches a preset performance threshold, the training is stopped to obtain an updated second performance evaluation model.

[0031] Another embodiment of the present invention provides a power grid service performance evaluation system, including:

[0032] The pre-processing module is used to filter and format the acquired historical business data stream of the target power grid in sequence to obtain a structured dynamic business data set;

[0033] A first training module is used to construct an initial performance evaluation model based on a random forest classification algorithm, and train the initial performance evaluation model based on the dynamic business data set to obtain a first performance evaluation model;

[0034] An extraction module, configured to extract time series features and business change features from the dynamic business data set, and generate dynamic business change trend data based on the time series features and the business change features;

[0035] A matching module, configured to match the dynamic business change trend data with a pre-built business rule library to obtain a change range of each business indicator in the dynamic business change trend data;

[0036] a calculation module configured to calculate a weight coefficient of each of the business indicators based on a linear regression algorithm if the change range exceeds a preset threshold range, and perform weighted processing on the dynamic business data set according to the weight coefficient to obtain a weighted dynamic business data set;

[0037] A second training module, configured to perform incremental training on the first performance evaluation model based on the weighted dynamic business data set to obtain an updated second performance evaluation model;

[0038] A generation module is used to input the acquired real-time business data flow of the target power grid into the second performance evaluation model to generate a performance evaluation result of the real-time business data flow.

[0039] As one preferred solution, the extraction module is specifically used to:

[0040] Based on the STL algorithm, the continuous monitoring data in the dynamic business data set is decomposed at multiple scales to obtain continuous parameter decomposition characteristics including long-term trend components, seasonal components and residual components;

[0041] Extracting parameter fluctuation characteristics within a preset time period before and after a discrete business event in the dynamic business data set by a sliding window method;

[0042] The continuous parameter decomposition feature and the parameter fluctuation feature are time-aligned to obtain a time series feature.

[0043] As one preferred solution, the extraction module is further configured to:

[0044] Performing frequency domain analysis on the time series characteristics according to a Fourier transform algorithm, and determining significant frequency components obtained according to the analysis as periodic variation characteristics;

[0045] Perform abnormal monitoring on the continuous monitoring data based on the CUSUM algorithm, and determine the parameter range and change rate before and after the extracted abnormal mutation point as the abnormal mutation feature;

[0046] The periodic change feature and the abnormal mutation feature are normalized to generate a multi-dimensional business change feature vector.

[0047] As one preferred solution, the extraction module is further configured to:

[0048] Inputting the time series characteristics and the business change characteristics into a pre-built ETS model to predict the business parameter change trend in a preset time period, and obtaining short-term trend data including trend values and confidence intervals;

[0049] Calculate the feature similarity between the business scenario in the forecast time period and the historical typical scenario using a dynamic time warping algorithm to obtain a typical pattern label for the business scenario;

[0050] The short-term trend data is fused with the typical pattern label to generate the dynamic business change trend data.

[0051] As one preferred solution, the second training module is specifically used to:

[0052] Dividing the weighted dynamic service data set into a training data set and a validation data set based on a preset division ratio;

[0053] Incrementally training the first performance evaluation model based on the training data set, wherein, during the training process, decision tree parameters in a random forest classification algorithm are adjusted to adapt to changes in the importance of business indicators in the weighted dynamic business data set;

[0054] Using the validation set to perform real-time evaluation on the first performance evaluation model during training, the evaluation includes determining the performance of the model by calculating an evaluation indicator;

[0055] When the evaluation index reaches a preset performance threshold, the training is stopped to obtain an updated second performance evaluation model.

[0056] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0057] 1) This invention uses a linear regression algorithm to dynamically calculate weight coefficients based on the magnitude of indicator changes, and performs weighted processing on dynamic business data sets, enabling the model to automatically focus on current abnormal or key business indicators (such as sudden load increases and abnormal fluctuations in equipment parameters), solving the problem of rigid static model weights. Combined with an incremental training mechanism, new data features are incorporated in real time while retaining historical model knowledge, ensuring that the evaluation model is dynamically optimized as the power grid business status changes, significantly improving the response accuracy to real-time abnormal scenarios.

[0058] 2) This invention achieves a transition from single-value indicator evaluation to a multi-dimensional analysis approach combining "trend prediction and pattern recognition" by extracting time series characteristics (long-term trends, seasonal fluctuations) and business change characteristics (abnormal mutations, cyclical patterns) to generate dynamic business change trend data. This technical feature enables the evaluation results to not only reflect the real-time status of the indicator but also integrate historical trends and business scenario evolution patterns (such as peak load patterns and precursors to equipment failures). This effectively captures potential risks in complex business scenarios and addresses the issues of delayed evaluation and missed assessments caused by traditional methods' one-sided reliance on real-time values. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 is a flow chart of a method for evaluating power grid service performance in one embodiment of the present invention;

[0060] Figure 2 Schematic diagram of a power grid service performance evaluation system in one embodiment of the present invention. DETAILED DESCRIPTION

[0061] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. The purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts shall fall within the scope of protection of the present invention.

[0062] In the description of this application, the terms "first," "second," "third," etc. are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first," "second," "third," etc. may explicitly or implicitly include one or more of the features. In the description of this application, unless otherwise specified, "plurality" means two or more.

[0063] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or an indirect connection through an intermediate medium, or it can be a communication between the two components. The terms "vertical", "horizontal", "left", "right", "up", "down" and similar expressions used herein are for illustrative purposes only, and do not indicate or imply that the device or component referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items. For those of ordinary skill in the art, the specific meanings of the above terms in this application can be understood according to specific circumstances.

[0064] In the description of this application, it should be noted that, unless otherwise defined, all technical and scientific terms used in this application have the same meanings as those commonly understood by those skilled in the art. The terms used in this specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. Those skilled in the art will understand the specific meanings of the above terms in this application in specific circumstances.

[0065] An embodiment of the present invention provides a method for evaluating the performance of a power grid service. For details, see Figure 1 , Figure 1 The figure shows a flow chart of a method for evaluating the performance of a power grid service in one embodiment of the present invention, which includes steps S1-S7:

[0066] S1: Filter and format the acquired historical business data stream of the target power grid in sequence to obtain a structured dynamic business data set;

[0067] Specifically, the historical business data stream of the target power grid is filtered and formatted in sequence. The filtering process includes eliminating data segments with excessive missing values and abnormal data exceeding three times the rated value (such as voltage and current surge and drop data) according to preset data quality rules (such as data integrity ≥ 95%, outlier threshold) to ensure the reliability of the input data; the formatting process includes converting unstructured data (such as JSON, CSV) into a unified structured format through a streaming computing framework (such as Apache Flink), including timestamp, device ID, monitoring parameters (such as voltage, current, load rate), and supplementing equipment ledger metadata (such as equipment model, rated parameters, and region), ultimately forming a dynamic business data set containing 10+ dimensional information.

[0068] For example, sensors deployed at key grid nodes, such as voltage transformers and current transformers, collect grid operating data every second, including parameters such as voltage, current, and frequency, at a sampling frequency of 1,000 times per second. This raw data is transmitted to a streaming computing platform via an IoT gateway device using the MQTT protocol at a rate of 100 messages per second. The streaming computing framework uses Apache Flink, setting filtering rules at the data inlet and using regular expressions to remove invalid data with incorrect formatting or out-of-range data (e.g., voltage below 220V or above 380V), retaining over 95% of valid data. The filtered data is formatted using predefined schema mapping rules, converting the raw string data into a structured JSON format containing fields such as timestamp, device ID, voltage, and current values. The formatted data is aggregated by minute using a time window function, and statistical metrics such as average voltage and maximum current for each device are calculated to form a structured, dynamic business data set. This data is transmitted to downstream analysis systems via a Kafka message queue at a rate of 50 messages per second, providing basic data support for real-time monitoring and early warning. During the entire processing process, the system delay is controlled within 500 milliseconds to ensure the real-time and accuracy of the data.

[0069] In one embodiment of the present invention, for structured dynamic business data sets, distributed parallel computing technology is applied to efficiently clean and integrate the data. If the data missing rate exceeds a preset threshold, historical data is interpolated and supplemented to obtain a complete high-quality data set.

[0070] S2: constructing an initial performance evaluation model based on a random forest classification algorithm, and training the initial performance evaluation model based on the dynamic business data set to obtain a first performance evaluation model;

[0071] Specifically, the structured dynamic business data set is divided into training set and validation set in a ratio of 7:3, and stratified sampling is used to ensure that the training set covers multiple business scenario samples such as normal, warning, and fault, to avoid category imbalance affecting the model generalization ability.

[0072] The random forest classifier was initialized in a machine learning framework such as Scikit-learn. The number of decision trees was set to 50, the maximum depth was set to 15, and the minimum number of sample splits was set to 10. The Gini coefficient was used as the splitting criterion. The parameters were optimized through 10-fold cross-validation to ensure that the classification accuracy of the validation set was ≥90% and the F1 score was ≥0.85. This resulted in the first performance evaluation model with multi-dimensional feature processing capabilities.

[0073] S3: extracting time series features and business change features from the dynamic business data set, and generating dynamic business change trend data based on the time series features and the business change features;

[0074] Preferably, in one embodiment of the present invention, extracting the time series features and business change features from the dynamic business data set includes:

[0075] Based on the STL algorithm, the continuous monitoring data in the dynamic business data set is decomposed at multiple scales to obtain continuous parameter decomposition characteristics including long-term trend components, seasonal components and residual components;

[0076] Extracting parameter fluctuation characteristics within a preset time period before and after a discrete business event in the dynamic business data set by a sliding window method;

[0077] The continuous parameter decomposition feature and the parameter fluctuation feature are time-aligned to obtain a time series feature.

[0078] Among them, each continuous parameter time series (such as the minute-level voltage data of a bus) is decomposed into three core components through the STL algorithm: long-term trend component (Trend): reflects the changing trend of the parameter over a long time range (such as the annual trend of seasonal load growth, and the slow decline trend of insulation resistance caused by aging of power grid equipment); seasonal component (Seasonal): captures the fluctuation characteristics with periodic laws (such as the periodic changes in the morning and evening peaks of daily load, and the weekday / weekend differences in weekly load); residual component (Residual): retains random fluctuations or abnormal mutation information that cannot be explained by trend and seasonal components (such as the instantaneous voltage drop caused by sudden short-circuit faults).

[0079] During the decomposition process, a sliding window is set (the window size is usually the period length, such as 24 hours for daily period decomposition) to ensure the accuracy of feature separation at different time scales. The trend, seasonal, and residual components obtained from the decomposition are aligned with the timestamps of the original time series to generate a continuous parameter decomposition feature containing these three components. For example, for voltage data, a feature vector with three dimensions is generated: "voltage trend value," "voltage seasonal fluctuation value," and "voltage residual fluctuation value."

[0080] For discrete business events (non-continuous events such as equipment startup and shutdown, planned maintenance, and fault tripping), a sliding window approach is used to extract parameter fluctuation characteristics before and after the event. The time point of the discrete event (denoted as t0) is identified through event logs (such as equipment status change records in the SCADA system) or parameter mutation detection (such as a sudden load increase exceeding twice the historical mean). With t0 as the center, sliding windows of T minutes are defined before and after (for example, T = 30 minutes, extracting data from 30 minutes before to 30 minutes after the event), ensuring coverage of the entire time range of the event's impact.

[0081] For continuously monitored parameters within the window (such as current and device temperature), the following fluctuation characteristics are calculated:

[0082] Extreme value characteristics: the maximum and minimum values of the parameters within the window and their difference from the steady-state value before the event (such as the deviation between the peak current and the rated current when the equipment starts);

[0083] Change rate: The slope of the parameter within the window (such as the voltage drop per minute), reflecting the speed of parameter change caused by the event;

[0084] Fluctuation range: The standard deviation or range of the parameter within the window, which measures the degree of parameter instability caused by the event.

[0085] Generate an event-driven feature vector containing event types (such as "transformer closing" and "line tripping") and fluctuating feature values.

[0086] The continuous parameter decomposition features and discrete event fluctuation features are integrated in the time dimension to form a complete time series feature:

[0087] The timestamps of the two types of feature data are aligned to the same time granularity (such as minutes), ensuring that the continuous decomposition features (such as the minute-level trend value output by STL) and the event fluctuation features (such as the rate of change per minute before and after the event) correspond one-to-one on the time axis.

[0088] The continuous parameter decomposition characteristics (trend, season, residual) of each time point are horizontally spliced with the associated discrete event fluctuation characteristics (such as extreme values and rates when events occur) in chronological order to form a multidimensional time series feature matrix that includes trend influence, seasonal patterns, and event fluctuations.

[0089] Preferably, in one embodiment of the present invention, extracting the time series features and business change features from the dynamic business data set includes:

[0090] Performing frequency domain analysis on the time series characteristics according to a Fourier transform algorithm, and determining significant frequency components obtained according to the analysis as periodic variation characteristics;

[0091] Perform abnormal monitoring on the continuous monitoring data based on the CUSUM algorithm, and determine the parameter range and change rate before and after the extracted abnormal mutation point as the abnormal mutation feature;

[0092] The periodic change feature and the abnormal mutation feature are normalized to generate a multi-dimensional business change feature vector.

[0093] Perform discrete Fourier transform (DFT) on time series features (such as minute-level data of parameters such as load rate and voltage) to convert the time domain signal into power spectral density (PSD) distribution in the frequency domain and identify the energy proportion of different frequency components.

[0094] Specifically, a frequency screening threshold is set (such as energy proportion > 5%) to extract frequency components with significant periodicity, typically including a 24-hour cycle (daily load fluctuations), a 7-day cycle (weekly load differences), etc., corresponding to periodic patterns such as "weekdays-weekends" and "peak-valley periods" in power grid business.

[0095] For each significant frequency component, its amplitude and phase in the time domain are calculated to generate periodic variation characteristics. For example, the amplitude of a 24-hour periodic component reflects the magnitude of daily load fluctuations, while the phase reflects the timing of peak hours. Example output: Features such as "daily load cycle amplitude" and "weekly load cycle phase" are extracted to form a dimensional vector describing periodic variation.

[0096] For each time series of continuously monitored parameters (such as current and device temperature), set the mean shift threshold (such as ±2 times the historical standard deviation) and the cumulative sum increment (such as k=0.5), and calculate the cumulative sum statistics in real time.

[0097] When the cumulative sum statistic exceeds the preset control limit (such as H = 4 times the standard deviation), the current time point is marked as an abnormal mutation point, indicating that the parameter has a mutation that significantly deviates from the historical trend (such as a sudden temperature rise before equipment failure or a sudden current increase during a short circuit fault).

[0098] Taking the mutation point as the center, extract parameter data for N minutes before and after (e.g., N = 10 minutes) and calculate the following features:

[0099] Parameter range: the difference between the maximum and minimum values within the mutation interval (reflecting the magnitude of the mutation, such as a sudden drop in voltage from 10kV to 8kV, the range is 2kV);

[0100] Change rate: the ratio of the parameter change within the mutation interval to the time (e.g., if the current increases by 500A within 10 minutes, the rate is 50A / minute);

[0101] Mutation direction: Marks whether the parameter is an upward or downward mutation (e.g., a sudden increase in load rate is a positive mutation, and a sudden drop in voltage is a negative mutation). Generates an abnormal mutation feature vector containing the mutation amplitude, rate, and direction.

[0102] The periodic change features and abnormal mutation features are integrated into a unified feature space. The periodic features (such as amplitude and phase) and abnormal mutation features (range and rate) are normalized separately, and all features are scaled to the range of 0-1 to eliminate the impact of dimensional differences on model training.

[0103] Preferably, in one embodiment of the present invention, generating dynamic business change trend data based on the time series characteristics and the business change characteristics includes:

[0104] Inputting the time series characteristics and the business change characteristics into a pre-built ETS model to predict the business parameter change trend in a preset time period, and obtaining short-term trend data including trend values and confidence intervals;

[0105] Calculate the feature similarity between the business scenario in the forecast time period and the historical typical scenario using a dynamic time warping algorithm to obtain a typical pattern label for the business scenario;

[0106] The short-term trend data is fused with the typical pattern label to generate the dynamic business change trend data.

[0107] Based on the seasonal fluctuation characteristics of the time series, select an additive ETS model (suitable for scenarios where the seasonal amplitude is stable) or a multiplicative model (suitable for scenarios where the seasonal amplitude increases with the trend). For example, if the seasonal fluctuation amplitude of load data increases due to high temperatures in the summer, a multiplicative model can be used.

[0108] We use Maximum Likelihood Estimation (MLE) to fit historical data, minimize the root mean square error (RMSE) between the predicted and actual values, and iteratively optimize smoothing parameters (trend smoothing coefficient, slope smoothing coefficient, and seasonal smoothing coefficient). We also introduce business rule constraints: If an equipment maintenance event is detected (identified through discrete event features), the residual term weight is temporarily increased to enhance the impact of abnormal data on the forecast.

[0109] Perform minute-by-minute forecasts of service parameters (such as voltage and load factor) for a preset time period (e.g., one hour) in the future, generating a trend value sequence (e.g., predicted voltage value 10.2 kV). Based on the historical residual distribution, a bootstrap method is used to calculate a 95% confidence interval (e.g., [10.1 kV, 10.3 kV]) to quantify the uncertainty of the forecast results.

[0110] Cluster analysis is performed on historical data to identify several typical business scenarios (such as "load peak mode", "equipment startup mode", and "fault recovery mode"). Each scenario contains a corresponding feature template (for example, the feature template of the load peak mode is "the daily load curve exceeds the rated value by 110% during the 18:00-22:00 period").

[0111] The feature vector of the forecast period (including time series decomposition features and business change features) is converted into a dynamic time series matrix. The time warping distance between the feature vector and the historical typical scenario template is calculated using the DTW algorithm. The scenario with the smallest distance is the matching typical pattern. For example, if the load curve of the current forecast period has the smallest distance to the feature template of the historical "summer high temperature load pattern", it is marked as "summer high temperature load pattern" and accompanied by a matching confidence level (e.g., 92%).

[0112] Trend values are modified to adapt to specific scenarios based on typical pattern labels. For example, if a "summer high temperature load pattern" is matched, a 5%-10% correction factor is added to the load rate trend value to reflect the historical average load premium under that pattern. The pattern matching confidence level is combined with the ETS model prediction confidence interval to generate a composite confidence level (e.g., taking the minimum or weighted average of the two), improving the reliability of trend data.

[0113] The integrated dynamic business change trend data includes the following core fields:

[0114] Trend information: parameter name (such as "main transformer load rate"), predicted trend value, and confidence interval;

[0115] Mode information: typical mode labels (such as “winter cold wave load mode”), matching confidence;

[0116] Time information: prediction start / end time, timestamp granularity (such as minutes).

[0117] S4: matching the dynamic business change trend data with a pre-built business rule library to obtain the change range of each business indicator in the dynamic business change trend data;

[0118] For continuous indicators (voltage, current), calculate the deviation rate from the rated value (such as voltage ±5%, current ±10%); for comprehensive indicators (equipment failure rate, load balancing rate), calculate the hourly change rate (such as failure rate increase ≥2% / hour).

[0119] Compare the real-time deviation rate / change rate with the preset threshold of the rule base to generate a data list containing the indicator name, current value, threshold range, and degree of deviation, which serves as the core basis for judging indicator abnormalities.

[0120] For example, in dynamic business trend analysis, historical business data is first fitted using a time series model. For example, using the ARIMA algorithm with parameters p = 2, d = 1, and q = 1, the sales data for the past 12 months is modeled to predict sales trends for the next three months. Based on a pre-established business rule library, the normal range of sales fluctuation is set to ±5%. If the forecast results show that the sales fluctuation in a particular month reaches 8%, the model parameter adjustment mechanism is triggered.

[0121] S5: If the change range exceeds a preset threshold range, the weight coefficients of the business indicators are calculated based on a linear regression algorithm, and the dynamic business data set is weighted according to the weight coefficients to obtain a weighted dynamic business data set;

[0122] Specifically, we screen historical data segments of indicators with excessive changes in the past three months (such as periods with a sudden 15% load increase), extract abnormal indicator values (such as real-time load rate), related indicator values (such as current and equipment temperature in the corresponding period), and manually labeled performance levels (levels 1-5) to form a special training set that includes abnormal scenarios.

[0123] Taking performance level as the dependent variable and standardized abnormal indicators and related indicators as independent variables, the significant features (p value < 0.05) were screened through stepwise regression method, and the multicollinearity was suppressed by ridge regression regularization processing (α = 0.1) to obtain a stable business indicator weight coefficient vector.

[0124] Convert the weight coefficient to sample weight (for example, the sample weight of high-weight indicators is 1.5-2.0, and the low-weight indicator remains at 1.0). Use tools such as Pandas to add weight labels to the dynamic business data set to generate a weighted dynamic business data set, ensuring that the positive and negative sample weight ratio is ≤3:1 to meet the requirements of incremental training.

[0125] S6: performing incremental training on the first performance evaluation model based on the weighted dynamic business data set to obtain an updated second performance evaluation model;

[0126] Preferably, in one embodiment of the present invention, the incremental training of the first performance evaluation model based on the weighted dynamic business data set to obtain an updated second performance evaluation model includes:

[0127] Dividing the weighted dynamic service data set into a training data set and a validation data set based on a preset division ratio;

[0128] Incrementally training the first performance evaluation model based on the training data set, wherein, during the training process, decision tree parameters in a random forest classification algorithm are adjusted to adapt to changes in the importance of business indicators in the weighted dynamic business data set;

[0129] Using the validation set to perform real-time evaluation on the first performance evaluation model during training, the evaluation includes determining the performance of the model by calculating an evaluation indicator;

[0130] When the evaluation index reaches a preset performance threshold, the training is stopped to obtain an updated second performance evaluation model.

[0131] Specifically, the decision tree structure and initial parameters of the first performance evaluation model are retained, and only the feature importance input interface is updated to avoid computational loss from repeated training.

[0132] The weighted data set is input in batches, and the splitting threshold of the decision tree is updated batch by batch through incremental training. The selection probability of the splitting nodes corresponding to high-weight indicators is increased by 30% (for example, load-related indicators are triggered first when the decision tree splits). After 5-10 iterations, the F1 score of the validation set is increased by ≥5%, generating a second performance evaluation model that adapts to the current business status.

[0133] S7: Inputting the acquired real-time business data flow of the target power grid into the second performance evaluation model to generate a performance evaluation result of the real-time business data flow.

[0134] Specifically, sensor data is acquired in real time through the Kafka message queue, and converted into a format isomorphic to the weighted dynamic business data set (minute-level timestamp, standardized indicator value) through Flink preprocessing. Dynamic normalization is performed using the mean / standard deviation within the current 1-hour sliding window to avoid evaluation bias caused by outdated historical statistical features.

[0135] A 5-minute sliding window aggregation is performed on high-frequency data at the second level, combining batch inference and single-sample inference to output evaluation results including real-time performance level, confidence level, and contribution of key indicators. Results with a confidence level of less than 70% trigger a secondary verification mechanism to ensure the accuracy of abnormal scenario identification.

[0136] Another embodiment of the present invention provides a power grid service performance evaluation system. For details, see Figure 2 , Figure 2 The figure shows a schematic diagram of a power grid service performance evaluation system according to one embodiment of the present invention, which includes:

[0137] The pre-processing module 11 is used to filter and format the acquired historical business data stream of the target power grid in sequence to obtain a structured dynamic business data set;

[0138] A first training module 12 is configured to construct an initial performance evaluation model based on a random forest classification algorithm, and train the initial performance evaluation model based on the dynamic business data set to obtain a first performance evaluation model;

[0139] An extraction module 13 is configured to extract time series features and business change features from the dynamic business data set, and generate dynamic business change trend data based on the time series features and the business change features;

[0140] A matching module 14 is configured to match the dynamic business change trend data with a pre-built business rule library to obtain a change range of each business indicator in the dynamic business change trend data;

[0141] a calculation module 15 configured to calculate weight coefficients of the business indicators based on a linear regression algorithm if the change range exceeds a preset threshold range, and perform weighted processing on the dynamic business data set according to the weight coefficients to obtain a weighted dynamic business data set;

[0142] A second training module 16 is configured to perform incremental training on the first performance evaluation model based on the weighted dynamic business data set to obtain an updated second performance evaluation model;

[0143] The generating module 17 is configured to input the acquired real-time business data stream of the target power grid into the second performance evaluation model to generate a performance evaluation result of the real-time business data stream.

[0144] Preferably, in one embodiment of the present invention, the extraction module is specifically used to:

[0145] Based on the STL algorithm, the continuous monitoring data in the dynamic business data set is decomposed at multiple scales to obtain continuous parameter decomposition characteristics including long-term trend components, seasonal components and residual components;

[0146] Extracting parameter fluctuation characteristics within a preset time period before and after a discrete business event in the dynamic business data set by a sliding window method;

[0147] The continuous parameter decomposition feature and the parameter fluctuation feature are time-aligned to obtain a time series feature.

[0148] Preferably, in one embodiment of the present invention, the extraction module is further used to:

[0149] Performing frequency domain analysis on the time series characteristics according to a Fourier transform algorithm, and determining significant frequency components obtained according to the analysis as periodic variation characteristics;

[0150] Perform abnormal monitoring on the continuous monitoring data based on the CUSUM algorithm, and determine the parameter range and change rate before and after the extracted abnormal mutation point as the abnormal mutation feature;

[0151] The periodic change feature and the abnormal mutation feature are normalized to generate a multi-dimensional business change feature vector.

[0152] Preferably, in one embodiment of the present invention, the extraction module is further used to:

[0153] Inputting the time series characteristics and the business change characteristics into a pre-built ETS model to predict the business parameter change trend in a preset time period, and obtaining short-term trend data including trend values and confidence intervals;

[0154] Calculate the feature similarity between the business scenario in the forecast time period and the historical typical scenario using a dynamic time warping algorithm to obtain a typical pattern label for the business scenario;

[0155] The short-term trend data is fused with the typical pattern label to generate the dynamic business change trend data.

[0156] Preferably, in one embodiment of the present invention, the second training module is specifically used to:

[0157] Dividing the weighted dynamic service data set into a training data set and a validation data set based on a preset division ratio;

[0158] Incrementally training the first performance evaluation model based on the training data set, wherein, during the training process, decision tree parameters in a random forest classification algorithm are adjusted to adapt to changes in the importance of business indicators in the weighted dynamic business data set;

[0159] Using the validation set to perform real-time evaluation on the first performance evaluation model during training, the evaluation includes determining the performance of the model by calculating an evaluation indicator;

[0160] When the evaluation index reaches a preset performance threshold, the training is stopped to obtain an updated second performance evaluation model.

[0161] Compared with the prior art, the embodiments of the present invention have the following advantages:

[0162] 1) This invention uses a linear regression algorithm to dynamically calculate weight coefficients based on the magnitude of indicator changes, and performs weighted processing on dynamic business data sets, enabling the model to automatically focus on current abnormal or key business indicators (such as sudden load increases and abnormal fluctuations in equipment parameters), solving the problem of rigid static model weights. Combined with an incremental training mechanism, new data features are incorporated in real time while retaining historical model knowledge, ensuring that the evaluation model is dynamically optimized as the power grid business status changes, significantly improving the response accuracy to real-time abnormal scenarios.

[0163] 2) This invention achieves a transition from single-value indicator evaluation to a multi-dimensional analysis approach combining "trend prediction and pattern recognition" by extracting time series characteristics (long-term trends, seasonal fluctuations) and business change characteristics (abnormal mutations, cyclical patterns) to generate dynamic business change trend data. This technical feature enables the evaluation results to not only reflect the real-time status of the indicator but also integrate historical trends and business scenario evolution patterns (such as peak load patterns and precursors to equipment failures). This effectively captures potential risks in complex business scenarios and addresses the issues of delayed evaluation and missed assessments caused by traditional methods' one-sided reliance on real-time values.

[0164] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

Claims

1. A method for evaluating power grid service performance, characterized in that: include: The acquired historical business data stream of the target power grid is filtered and formatted in sequence to obtain a structured dynamic business data set; Constructing an initial performance evaluation model based on a random forest classification algorithm, and training the initial performance evaluation model based on the dynamic business data set to obtain a first performance evaluation model; Extracting time series features and business change features from the dynamic business data set, and generating dynamic business change trend data based on the time series features and the business change features; Matching the dynamic business change trend data with a pre-built business rule library to obtain the change range of each business indicator in the dynamic business change trend data; If the variation exceeds a preset threshold range, the weight coefficients of the business indicators are calculated based on a linear regression algorithm, and the dynamic business data set is weighted according to the weight coefficients to obtain a weighted dynamic business data set; Incrementally training the first performance evaluation model based on the weighted dynamic business data set to obtain an updated second performance evaluation model; The acquired real-time business data stream of the target power grid is input into the second performance evaluation model to generate a performance evaluation result of the real-time business data stream.

2. The power grid service performance evaluation method according to claim 1, wherein: The extracting of time series features and business change features from the dynamic business data set includes: Based on the STL algorithm, the continuous monitoring data in the dynamic business data set is decomposed at multiple scales to obtain continuous parameter decomposition characteristics including long-term trend components, seasonal components and residual components; Extracting parameter fluctuation characteristics within a preset time period before and after a discrete business event in the dynamic business data set by a sliding window method; The continuous parameter decomposition feature and the parameter fluctuation feature are time-aligned to obtain a time series feature.

3. The power grid service performance evaluation method according to claim 2, wherein: The extracting of time series features and business change features from the dynamic business data set includes: Performing frequency domain analysis on the time series characteristics according to a Fourier transform algorithm, and determining significant frequency components obtained according to the analysis as periodic variation characteristics; Perform abnormal monitoring on the continuous monitoring data based on the CUSUM algorithm, and determine the parameter range and change rate before and after the extracted abnormal mutation point as the abnormal mutation feature; The periodic change feature and the abnormal mutation feature are normalized to generate a multi-dimensional business change feature vector.

4. The power grid service performance evaluation method according to claim 1, wherein: The generating of dynamic business change trend data based on the time series characteristics and the business change characteristics includes: Inputting the time series characteristics and the business change characteristics into a pre-built ETS model to predict the business parameter change trend in a preset time period, and obtaining short-term trend data including trend values and confidence intervals; Calculate the feature similarity between the business scenario in the forecast time period and the historical typical scenario using a dynamic time warping algorithm to obtain a typical pattern label for the business scenario; The short-term trend data is fused with the typical pattern label to generate the dynamic business change trend data.

5. The power grid service performance evaluation method according to claim 1, wherein: The incrementally training the first performance evaluation model based on the weighted dynamic business data set to obtain an updated second performance evaluation model includes: Dividing the weighted dynamic service data set into a training data set and a validation data set based on a preset division ratio; Incrementally training the first performance evaluation model based on the training data set, wherein, during the training process, decision tree parameters in a random forest classification algorithm are adjusted to adapt to changes in the importance of business indicators in the weighted dynamic business data set; Using the validation set to perform real-time evaluation on the first performance evaluation model during training, the evaluation includes determining the performance of the model by calculating an evaluation indicator; When the evaluation index reaches a preset performance threshold, the training is stopped to obtain an updated second performance evaluation model.

6. A power grid business performance evaluation system, characterized in that: include: The pre-processing module is used to filter and format the acquired historical business data stream of the target power grid in sequence to obtain a structured dynamic business data set; A first training module is used to construct an initial performance evaluation model based on a random forest classification algorithm, and train the initial performance evaluation model based on the dynamic business data set to obtain a first performance evaluation model; An extraction module, configured to extract time series features and business change features from the dynamic business data set, and generate dynamic business change trend data based on the time series features and the business change features; A matching module, configured to match the dynamic business change trend data with a pre-built business rule library to obtain a change range of each business indicator in the dynamic business change trend data; a calculation module configured to calculate a weight coefficient of each of the business indicators based on a linear regression algorithm if the change range exceeds a preset threshold range, and perform weighted processing on the dynamic business data set according to the weight coefficient to obtain a weighted dynamic business data set; A second training module, configured to perform incremental training on the first performance evaluation model based on the weighted dynamic business data set to obtain an updated second performance evaluation model; A generation module is used to input the acquired real-time business data flow of the target power grid into the second performance evaluation model to generate a performance evaluation result of the real-time business data flow.

7. The power grid service performance evaluation method according to claim 6, characterized in that: The extraction module is specifically used to: Based on the STL algorithm, the continuous monitoring data in the dynamic business data set is decomposed at multiple scales to obtain continuous parameter decomposition characteristics including long-term trend components, seasonal components and residual components; Extracting parameter fluctuation characteristics within a preset time period before and after a discrete business event in the dynamic business data set by a sliding window method; The continuous parameter decomposition feature and the parameter fluctuation feature are time-aligned to obtain a time series feature.

8. The power grid service performance evaluation method according to claim 7, characterized in that: The extraction module is further used to: Performing frequency domain analysis on the time series characteristics according to a Fourier transform algorithm, and determining significant frequency components obtained according to the analysis as periodic variation characteristics; Perform abnormal monitoring on the continuous monitoring data based on the CUSUM algorithm, and determine the parameter range and change rate before and after the extracted abnormal mutation point as the abnormal mutation feature; The periodic change feature and the abnormal mutation feature are normalized to generate a multi-dimensional business change feature vector.

9. The power grid service performance evaluation method according to claim 6, wherein: The extraction module is further used to: Inputting the time series characteristics and the business change characteristics into a pre-built ETS model to predict the business parameter change trend in a preset time period, and obtaining short-term trend data including trend values and confidence intervals; Calculate the feature similarity between the business scenario in the forecast time period and the historical typical scenario using a dynamic time warping algorithm to obtain a typical pattern label for the business scenario; The short-term trend data is fused with the typical pattern label to generate the dynamic business change trend data.

10. The power grid service performance evaluation method according to claim 6, wherein: The second training module is specifically used to: Dividing the weighted dynamic service data set into a training data set and a validation data set based on a preset division ratio; Incrementally training the first performance evaluation model based on the training data set, wherein, during the training process, decision tree parameters in a random forest classification algorithm are adjusted to adapt to changes in the importance of business indicators in the weighted dynamic business data set; Using the validation set to perform real-time evaluation on the first performance evaluation model during training, the evaluation includes determining the performance of the model by calculating an evaluation indicator; When the evaluation index reaches a preset performance threshold, the training is stopped to obtain an updated second performance evaluation model.

Images