New energy station abnormal data detection method and system, computer equipment and medium

Abstract

The invention provides a new energy station abnormal data detection method and system, computer equipment and a medium, and belongs to the technical field of power system operation state monitoring and intelligent operation and maintenance, and the method comprises the steps: obtaining environment meteorological data, equipment operation state data and historical power output data; performing time alignment on the environmental meteorological data, the equipment operation state data and the historical power output data to obtain time sequence data; extracting a time pattern feature from the time series data; modeling a time relation and a sensor relation for the time mode characteristics based on a graph attention mechanism to obtain time sequence characteristics and sensor relation characteristics; fusion features are obtained through fusion; reconstructing according to the fusion features to obtain reconstructed data; and determining an abnormal data detection result of the to-be-evaluated new energy station according to the reconstruction data and reconstruction errors of the corresponding equipment operation state data and environment meteorological data. According to the method, the linkage abnormal mode between the sensors is captured, the limitation of single-point modeling is avoided, and the abnormal data detection accuracy is improved.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of power system operation state monitoring and intelligent operation and maintenance, and particularly relates to a new energy station abnormal data detection method and system, a computer device and a medium. BACKGROUND

[0002] With large-scale access of wind farms, photovoltaic power stations and other new energy stations to the power grid, the renewable energy output in the power system presents significant randomness and volatility. In order to ensure the safe and stable operation of the power grid and improve the level of new energy consumption, it is necessary to carry out fine operation monitoring on new energy stations, and it is necessary to timely identify and locate abnormal data and abnormal operation states in a large number of monitoring quantities of new energy stations.

[0003] In view of the above-mentioned needs, the existing technology often uses outlier detection methods based on statistical characteristics and clustering, isolation forest, autoencoder and other machine learning algorithms to detect abnormal data through single-point or a small number of feature modeling. However, this abnormal data detection ignores the spatial correlation and linkage abnormality between multiple devices and multiple measuring points in the new energy station, and when the scale of the station is expanded and the number of sensors is increased, the existing detection method is prone to high false positive rate, missing complex correlation abnormalities and other problems, and the accuracy of abnormal data detection is low. SUMMARY

[0004] In order to solve the problem of low accuracy of existing new energy station abnormal data detection, the present application provides a new energy station abnormal data detection method, system, computer device and medium.

[0005] In order to achieve the above-mentioned purpose, the present application provides the following technical solutions: A new energy station abnormal data detection method, comprising: Obtaining environmental meteorological data, device operation state data and historical power output data of a new energy station to be evaluated; time aligning the environmental meteorological data, device operation state data and historical power output data to obtain time series data; Extracting time-dependent relationships from the time series data to obtain time pattern features; modeling time relationships and sensor relationships of the time pattern features through the corresponding graph structure of the new energy station to be evaluated based on a graph attention mechanism to obtain time sequence features and sensor relationship features; and fusing the time pattern features, time sequence features and sensor relationship features to obtain fused features; Reconstructing the fused features to obtain reconstructed data; and determining an abnormal data detection result of the new energy station to be evaluated according to the reconstructed data and the reconstruction errors of the corresponding device operation state data and environmental meteorological data.

[0006] Optionally, the application provides a new energy station abnormal data detection method, wherein a graph attention mechanism detection prediction integrated model is used to process time series data to obtain an abnormal data detection result; the graph attention mechanism detection prediction integrated model comprises a time sequence feature extraction module, a graph attention feature fusion module, a sequence pattern capturing module and a reconstruction branch module connected in sequence; the graph attention feature fusion module comprises a feature-oriented graph attention network layer and a time-oriented graph attention network layer connected in parallel; The time sequence feature extraction module is used to extract time pattern features from the time series data; The feature-oriented graph attention network layer is used to extract sensor relationship features from the time pattern features; The time-oriented graph attention network layer is used to extract time sequence features from the time pattern features; The sequence pattern capturing module is used to fuse the time pattern features, the time sequence features and the sensor relationship features to obtain fused features; The reconstruction branch module is used to determine the abnormal data detection result according to the fused features.

[0007] Optionally, the application provides a new energy station abnormal data detection method, which further comprises: The environmental meteorological data, the equipment operation state data and the historical power output data are aligned based on a preset time scale to obtain time-aligned multi-source data; When the time-aligned multi-source data has missing values, the missing values are filled by interpolation or previous value filling to obtain multi-source data with filled missing values; Outliers are removed from the multi-source data with filled missing values to obtain multi-source data with removed outliers; The multi-source data with removed outliers is denoised and normalized to obtain time series data.

[0008] Optionally, based on the prior relationship between the equipment and the sensor measurement points in the new energy station to be evaluated, a plurality of graph structures are constructed to represent the correlation between the environmental meteorological data, the equipment operation state data and the historical power output data; the graph structure comprises a physical topology graph, a geographical adjacency graph and a power correlation graph; the application provides a new energy station abnormal data detection method, which further comprises: The equipment or sensor measurement points in the new energy station to be evaluated are taken as nodes, and the electrical connection relationship and the equipment structure connection relationship of the equipment or sensor measurement points are taken as edges to construct a physical topology graph; The equipment or sensor measurement points in the new energy station to be evaluated are taken as nodes, and the geographical positions of the equipment or sensor measurement points are taken as edges to construct a geographical adjacency graph; Take the equipment or sensor measuring point in the new energy station to be evaluated as a node, calculate the correlation coefficient according to the historical power output data corresponding to the equipment or sensor measuring point, when the correlation coefficient is greater than a preset correlation threshold, build an edge between the nodes corresponding to the correlation coefficient, and obtain a power correlation graph.

[0009] Optionally, the new energy station abnormal data detection method provided by the application further comprises: Determine the physical connection relationship weight between the equipment or sensor measuring points according to the physical topology graph; Determine the environmental influence correlation weight between the equipment or sensor measuring points according to the geographical adjacency graph; Determine the operation correlation relationship weight between the equipment or sensor measuring points according to the power correlation graph; Determine the comprehensive attention weight matrix according to the physical connection relationship weight, the environmental influence correlation weight and the operation correlation relationship weight; Based on the comprehensive attention weight matrix, the sensor relationship features are extracted from the time mode features through the graph attention mechanism.

[0010] Optionally, the new energy station abnormal data detection method provided by the application further comprises: Calculate the reconstruction data of the equipment or sensor measuring point and the corresponding equipment operation state data or environmental meteorological data to obtain the reconstruction error of the equipment or sensor measuring point; According to the equipment or sensor measuring point reconstruction error, the abnormal index of the new energy station to be evaluated is aggregated, and the abnormal data detection result of the new energy station to be evaluated is determined according to the abnormal index.

[0011] Optionally, the abnormal data detection result includes an outlier anomaly, and the graph attention mechanism detection prediction integrated model further comprises a prediction branch, and the new energy station abnormal data detection method provided by the application further comprises: When the abnormal data detection result is an outlier anomaly, the data of the similar nodes of the equipment or sensor measuring point corresponding to the outlier anomaly is estimated and filled to obtain the data estimate value of the equipment or sensor measuring point, wherein the similar nodes are determined based on the physical topology graph, the geographical adjacency graph and the power correlation graph; Replace the abnormal data corresponding to the outlier anomaly with the data estimate value; The replaced fusion features are extracted from the replaced environmental meteorological data, equipment operation state data and historical power output data through the sequentially connected time sequence feature extraction module, graph attention feature fusion module and sequence mode capture module prediction branch; The power prediction result of the new energy station to be evaluated is determined by the prediction branch from the fusion features.

[0012] Optionally, the abnormal data detection result further includes systematic deviation, and the new energy station abnormal data detection method provided by the application further includes: When the abnormal data detection result is the systematic deviation, the physical connection relationship weight, the environmental influence correlation weight and the operation related relationship weight of the device or the sensor measurement point corresponding to the systematic deviation are adjusted, and the weight-adjusted fusion features are extracted from the environmental meteorological data, the device operation state data and the historical power output data according to the adjusted physical connection relationship weight, the environmental influence correlation weight and the operation related relationship weight. The power prediction result of the new energy station to be evaluated is determined by the weight-adjusted fusion features through the prediction branch.

[0013] Optionally, the new energy station abnormal data detection method provided by the application further includes: An abnormal report is generated from the abnormal data detection result by the large language model pre-trained, wherein the abnormal report includes the abnormal type of the abnormal data detection result, the corresponding device or sensor measurement point information, and the recommended measures, and the recommended measures are determined based on the semantic matching between the abnormal type, the corresponding device or sensor measurement point information and the pre-constructed abnormal data knowledge base.

[0014] The application further provides a new energy station abnormal data detection system, comprising: A data preprocessing module is configured to obtain environmental meteorological data, device operation state data and historical power output data of a new energy station to be evaluated, and perform time alignment on the environmental meteorological data, the device operation state data and the historical power output data to obtain time series data. A feature fusion module is configured to extract time-dependent relationships from the time series data to obtain time pattern features, model time relationships and sensor relationships based on a graph attention mechanism through a graph structure corresponding to the new energy station to be evaluated, and obtain time sequence features and sensor relationship features, and fuse the time pattern features, the time sequence features and the sensor relationship features to obtain fusion features. An abnormal data detection module is configured to reconstruct the fusion features to obtain reconstructed data, and determine an abnormal data detection result of the new energy station to be evaluated according to the reconstructed data and the reconstruction errors of the corresponding device operation state data and environmental meteorological data.

[0015] The application further provides a computer device comprising a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any one of the new energy station abnormal data detection method.

[0016] The application further provides a computer-readable storage medium, and the computer program is stored on the storage medium and can execute the steps of any one of the new energy station abnormal data detection methods when loaded by a processor.

[0017] The new energy station abnormal data detection method has the following beneficial effects: The new energy station abnormal data detection method can capture the linkage abnormal mode among sensors by dynamically learning the physical topology, geographical proximity and power correlation relationship among devices based on the graph attention mechanism, thereby avoiding the limitation of single-point modeling; by fusing the time mode feature, time sequence feature and sensor relationship feature, the scheme enhances the comprehensiveness of feature representation, so that the reconstruction error can more accurately reflect the abnormal deviation, significantly reduces the false positive rate and the missed detection rate, and is especially suitable for identifying complex correlation abnormalities in large-scale stations. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present application and the design scheme thereof, the drawings required by the present embodiments will be briefly introduced as follows. The drawings in the following description are only part of the embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0019] Figure 1 A new energy station abnormal data detection method provided for the embodiments of the present application is shown in the figure. Figure 2 A flowchart example of the abnormal data detection and power prediction method provided for the embodiments of the present application. DETAILED DESCRIPTION

[0020] In order to make those skilled in the art better understand the technical scheme of the present application and can be implemented, the present application will be described in detail below in combination with the drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical scheme of the present application, and cannot be used to limit the protection scope of the present application.

[0021] New energy power stations not only have the need for abnormal data detection, but also often need high-precision short-term prediction of the power of the power station. Existing new energy power station power prediction methods mainly include unit power curve models based on physical mechanisms, time series models based on statistical regression, and deep learning-based prediction models widely used in recent years, such as long short-term memory networks and convolutional neural networks. These methods usually model the overall power of the power station or the power of a single unit as an independent time series. Some works introduce external factors such as wind speed and irradiance as auxiliary features, but most do not fully utilize the physical topology relationship, geographical spatial relationship, and power correlation between internal devices in the power station, and have limited ability to model the correlation of multi-source data, making it difficult to fully characterize the complex spatio-temporal coupling characteristics of new energy power stations. In addition, power prediction and abnormal detection are usually implemented by two independent modules in engineering practice, and there is a lack of coordination between models, resulting in redundant overall architecture and high maintenance cost.

[0022] With the development of neural networks, graph convolutional networks and graph attention networks have been gradually applied to power system load forecasting and topology analysis. These graph neural networks can model the correlation between nodes to some extent using network structure information. However, existing works mostly focus on a single task, such as power prediction or power flow estimation, and there is still a lack of solutions that unify the modeling of physical topology, geographical proximity, and power correlation in the context of new energy power stations, and combine them with time series networks to form an integrated framework for prediction and abnormal detection. At the same time, existing abnormal detection models usually only output numerical abnormal scores or alarm identifiers, and lack the ability to provide semantic explanations that are closely related to the operation and maintenance process.

[0023] Furthermore, in the context of operation and maintenance, the existing abnormal handling process of new energy power stations still relies mainly on manual analysis. Operation and maintenance personnel need to comprehensively review historical curves, alarm records, and device information, and manually write abnormal event reports, which is time-consuming and the quality is significantly affected by individual experience. Abnormal knowledge is scattered in the form of documents or individual experience, and there is a lack of systematic abnormal knowledge base and closed-loop utilization model optimization mechanism, which is not conducive to accumulating experience and continuously improving detection and prediction performance in the long-term operation process. Although large-scale pre-trained language models have shown strong capabilities in natural language generation and knowledge abstraction, existing technologies are still in the early stages of integrating language models for automatic generation of abnormal reports and deep integration with abnormal detection models.

[0024] In summary, there is still a lack of a new energy power station operation monitoring method in the existing technology that can simultaneously utilize multi-source monitoring data and multi-relation graph structures, balance high-precision power prediction and abnormal data intelligent detection, and automatically generate abnormal reports using language models and achieve continuous optimization through an abnormal knowledge base in a unified model framework.

[0025] The application aims to overcome the problems in the prior art that new energy station power prediction and abnormality detection are dispersedly implemented, the internal multi-relation structure of the station is not fully utilized, the semantic expression of abnormal alarm information is insufficient, and it is difficult to form a knowledge closed loop, and proposes a new energy station abnormal data detection and power prediction method based on a graph attention mechanism, which realizes the correlation modeling of multi-source monitoring data, high-precision power prediction and intelligent identification of abnormalities under a unified model framework, and improves the operation and maintenance decision support capability by combining a large language model and an abnormal knowledge base.

[0026] Embodiment 1 The application provides a new energy station abnormal data detection method, specifically as shown in the figure, comprising the following steps: Figure 1 The application provides a new energy station abnormal data detection method, specifically as shown in the figure, comprising the following steps: Step 11, obtaining environmental meteorological data, equipment operating state data and historical power output data of the new energy station to be evaluated; time aligning the environmental meteorological data, equipment operating state data and historical power output data to obtain time series data.

[0027] The time series data can be determined by the following steps: Step 111, aligning the environmental meteorological data, equipment operating state data and historical power output data based on a preset time scale to obtain multi-source data after time alignment.

[0028] Step 112, when the multi-source data after time alignment has missing values, filling in estimated values by interpolation or previous value filling to obtain multi-source data after filling in missing values.

[0029] Step 113, removing outliers from the multi-source data after filling in missing values to obtain multi-source data after removing abnormal values.

[0030] Step 114, denoising and normalizing the multi-source data after removing abnormal values to obtain time series data.

[0031] Specifically, first, the full amount of core operating data of the new energy station is obtained, including but not limited to environmental meteorological data, equipment operating state data and historical power output data; the obtained data is preprocessed, including aligning the time scales of different data, filling in missing data, smoothing noise and normalization processing.

[0032] For example, the multi-source data is aligned according to a unified time scale, the data is resampled and interpolated at a selected time step, each time point has corresponding multi-dimensional feature values, for missing data, interpolation, previous value filling or estimation method based on related variables are used to fill in, if the missing duration is long and cannot be reliably filled in, a missing mark is added to the data of the corresponding time period to avoid the influence of false data on model training, outliers obviously beyond the physical possible range of the device are removed to reduce noise interference, high-frequency measurement noise is smoothed to reduce random fluctuations, and normalization or standardization processing is performed on feature data of different dimensions to improve the quality and consistency of input data.

[0033] Step 12, taking the devices or sensor measurement points in the new energy station to be evaluated as nodes, taking the electrical connection relationship and device structure connection relationship of the devices or sensor measurement points as edges, a physical topology graph is constructed; taking the devices or sensor measurement points in the new energy station to be evaluated as nodes, taking the geographical position of the devices or sensor measurement points as edges, a geographical adjacency graph is constructed; taking the devices or sensor measurement points in the new energy station to be evaluated as nodes, according to the historical power output data corresponding to the devices or sensor measurement points, a correlation coefficient is calculated, when the correlation coefficient is greater than a preset correlation threshold, an edge is constructed between the nodes corresponding to the correlation coefficient, and a power correlation graph is obtained.

[0034] Specifically, based on the prior relationship of the devices and sensor measurement points in the new energy station, multiple graph structures are constructed to represent the correlation relationship between variables, including a physical topology graph, a geographical adjacency graph and a power correlation graph. In the physical topology graph, the nodes represent the devices or sensor measurement points in the station, if two nodes have a direct correlation in the actual electrical connection relationship or device structure, an edge is connected between the nodes to represent the physical connection relationship between the devices or measurement points; in the geographical adjacency graph, the nodes represent the devices or sensors in the station, if two nodes are adjacent in geographical position, an edge is connected between the nodes to represent the relationship that adjacent devices are affected by similar environmental conditions; in the power correlation graph, the correlation relationship between the devices or measurement points is determined based on historical operation data, the nodes represent the devices or measurement points in the station, the correlation coefficient between the nodes is calculated by selecting an index representing the operating state of the device, if the correlation coefficient of two nodes in the historical data exceeds a preset threshold, an edge is connected between the nodes to represent that they have strong correlation.

[0035] Step 13, the time-dependent relationship is extracted from the time series data to obtain time pattern features; based on the graph attention mechanism, the time pattern features are modeled for time relationship and sensor relationship through the graph structure corresponding to the new energy station to be evaluated, time sequence features and sensor relationship features are obtained; the time pattern features, time sequence features and sensor relationship features are fused to obtain fused features.

[0036] The abnormality detection and power prediction integrated model adopts a graph attention mechanism in a graph neural network to fuse the correlation characteristics of multi-sensor data and simultaneously performs prediction of future power and identification of abnormal data. The abnormality detection and power prediction integrated model processes time series data through the graph attention mechanism to obtain an abnormal data detection result; the graph attention mechanism detection and prediction integrated model includes sequentially connected time series feature extraction modules, a graph attention feature fusion module, a sequence pattern capturing module, and a reconstruction branch module, and the graph attention feature fusion module includes a feature-oriented graph attention network layer and a time-oriented graph attention network layer in parallel; the time series feature extraction module extracts time pattern features from the time series data; the feature-oriented graph attention network layer extracts sensor relationship features from the time pattern features; the time-oriented graph attention network layer extracts time series features from the time pattern features; the sequence pattern capturing module fuses the time pattern features, the time series features, and the sensor relationship features to obtain fused features; and the reconstruction branch module determines the abnormal data detection result based on the fused features.

[0037] Specifically, the graph attention mechanism detection and prediction integrated model extracts time pattern features using a sequence network such as a gated recurrent unit (GRU) and captures complex dependency relationships between different sensor features and between different time points in combination with a graph attention network; the graph attention mechanism detection and prediction integrated model includes two parallel graph attention modules: a feature-oriented graph attention module for modeling relationships between sensor features and a time-oriented graph attention module for modeling dependency relationships in the time dimension; the feature-oriented graph attention module uses graph structure information to establish edges between nodes, aggregates multiple relationships such as physical topology, geographical proximity, and power correlation, and realizes dynamic learning of the correlation between different sensors; the time-oriented graph attention module takes each time point within a time window as a node of a complete graph to capture dependency patterns between time series at different times; the outputs of the two graph attention modules are fused with the original time features and input to a subsequent sequence model layer to further extract long-term time series features; the detection and prediction integrated model realizes self-supervised learning by jointly training a prediction sub-model and a reconstruction sub-model: the prediction sub-model is used to perform single-step or multi-step prediction of the power value at a future time, and the reconstruction sub-model is used to reconstruct the input time series to capture global data distribution features.

[0038] The step 13 includes: Step 131, determining a physical connection relationship weight between devices or sensor measurement points according to a physical topology graph.

[0039] Step 132, determining an environmental influence correlation weight between devices or sensor measurement points according to a geographical adjacency graph.

[0040] Step 133, determining the operation correlation weight between the devices or sensor points according to the power correlation graph.

[0041] Step 134, determining the comprehensive attention weight matrix according to the physical connection weight, the environment influence correlation weight and the operation correlation weight.

[0042] Step 135, extracting the sensor relationship features from the time mode features by the graph attention mechanism based on the comprehensive attention weight matrix.

[0043] Step 14, reconstructing the reconstructed data according to the fused features; determining the abnormal data detection result of the new energy station to be evaluated according to the reconstruction error of the reconstructed data and the corresponding device operation state data and environment meteorological data.

[0044] Among them, step 14 includes: Step 141, calculating the reconstruction error of the device or sensor point by the reconstructed data and the corresponding device operation state data or environment meteorological data.

[0045] Step 142, aggregating the device or sensor point reconstruction error to obtain the abnormal index of the new energy station to be evaluated, and determining the abnormal data detection result of the new energy station to be evaluated according to the abnormal index.

[0046] Specifically, when obtaining the real-time environment data and operation data of the new energy station to be evaluated, the model outputs the abnormal detection result of each monitoring variable at the current time based on the graph attention mechanism. And when detecting the abnormality, record the time of abnormality, the associated sensor or device and its value deviation. First, calculate the reconstruction error of each monitoring variable at the last time in the current time window, wherein the reconstruction error is the difference between the actual observation value and the model reconstruction value; determine the abnormal score of each monitoring variable according to the reconstruction error of each monitoring variable, and aggregate the abnormal scores of each monitoring variable to obtain the overall abnormal index. Compare the abnormal index with the threshold value to determine whether there is an abnormality.

[0047] In addition to identifying abnormal data, the model provided by the application can also classify abnormal data, that is, after determining the occurrence of an anomaly, the source of the anomaly is located by analyzing the contribution of each monitoring variable to the score of the anomaly indicator, and the attention weight of the graph attention network model is used to assist in identifying the type of anomaly: when the correlation weight of a sensor relative to other sensors changes significantly, the anomaly is determined to belong to a sensor failure or a correlation system anomaly in combination with prior knowledge to refine the anomaly category. For example, abnormal data includes physical limit anomaly, outlier or noise, systematic bias, and data missing; wherein the physical limit anomaly is directly determined by a preset threshold, the outlier or noise and the systematic bias are determined by the anomaly score output by the model, and the data missing is identified by detecting the missing marker or comparing the deviation of the model reconstruction value from the normal value; and the method provided by the application can also detect sensor freezing anomalies and correlation anomalies.

[0048] In addition, a dynamic threshold strategy is adopted for anomaly determination to adapt to the changes in data volatility under different working conditions and seasonal conditions: based on the statistical situation of the anomaly score in a recent period of time, the anomaly determination threshold at the current time is adaptively determined, and the threshold can be set for different monitoring variables, thereby improving the adaptability and robustness of anomaly detection.

[0049] Step 15, when the abnormal data detection result is an outlier anomaly, the data of the equipment or sensor measurement point near the node corresponding to the outlier anomaly is estimated and filled to obtain the data estimate value of the equipment or sensor measurement point, wherein the near node is determined based on a physical topology graph, a geographical adjacency graph and a power correlation graph; the abnormal data corresponding to the outlier anomaly is replaced by the data estimate value; the replaced environmental meteorological data, equipment operating state data and historical power output data are extracted by the replaced environmental meteorological data, equipment operating state data and historical power output data through the sequentially connected time series feature extraction module, the graph attention feature fusion module and the sequence pattern capturing module prediction branch; the power prediction result of the new energy station to be evaluated is determined by the fusion feature through the prediction branch.

[0050] Step 16, when the abnormal data detection result is a systematic bias, the physical connection relationship weight, the environmental influence correlation weight and the operation correlation weight of the equipment or sensor measurement point corresponding to the systematic bias are adjusted, and the weight-adjusted fusion feature is extracted from the environmental meteorological data, the equipment operating state data and the historical power output data by the adjusted physical connection relationship weight, the environmental influence correlation weight and the operation correlation weight; the power prediction result of the new energy station to be evaluated is determined by the weight-adjusted fusion feature through the prediction branch.

[0051] Specifically, after the abnormal data is detected, the real-time data collected can also be corrected based on the abnormal data, for example, the weight of the systematic error caused by the batch sensor error in data processing is reduced, the data of the similar points is used to estimate and fill the individual discrete points, and the power prediction is performed through the graph attention mechanism to detect the prediction branch or the prediction sub-model in the prediction and prediction integrated model.

[0052] Moreover, the prediction branch or the prediction sub-model and the reconstruction branch or the reconstruction sub-model are trained and optimized in a joint training manner. For example, after the two branches generate the predicted values and the reconstructed values based on the training data, the prediction error between the predicted values and the actual values and the reconstruction error between the reconstructed values and the actual values are used as the loss or deviation of the model as a whole, and then the joint loss is reduced through hyperparameter optimization.

[0053] Step 17, generating an abnormal report from the abnormal data detection result by a pre-trained large language model, wherein the abnormal report includes the abnormal type of the abnormal data detection result, the corresponding device or sensor measurement point information, and the suggested measures, which are determined based on the semantic matching of the abnormal type, the corresponding device or sensor measurement point information, and the pre-constructed abnormal data knowledge base.

[0054] Specifically, for the detected abnormal data or event, a pre-trained large language model is called to generate an abnormal report; the abnormal related information is input into the large language model as a prompt, and a prompt word customized for the new energy station is used to guide the generation of a natural language abnormal report in a format meeting the requirements of the station.

[0055] The abnormal report generated based on the large language model includes the time and duration of the abnormal occurrence, the specific device or sensor measurement point name involved in the abnormality, the detailed index of the abnormal data, the determined abnormal type and severity, and the suggested measures for the operation and maintenance of the new energy station; the report gives a description and analysis of the abnormal event in a natural language form, thereby providing comprehensive background information; the large language model organizes the above-mentioned content into a text meeting the management specifications of the station according to a pre-designed template and prompt; the report content can be continuously optimized according to the historical experience in the knowledge base, so that it gradually has a simple inference ability for the abnormal reasons and a targeted disposal suggestion.

[0056] Moreover, the abnormal data detected this time and its related information are stored in the abnormal data knowledge base; the knowledge base records the detailed conditions of the previous abnormalities, including the time period of the abnormal occurrence, the devices and sensors involved, the abnormal type, the data feature description, the cause analysis, and the handling measures and results, etc., and is used for subsequent optimization training of the integrated model.

[0057] Furthermore, the anomaly data knowledge base is used to improve model performance and refine anomaly handling processes over the long term. When new anomaly cases are entered into the knowledge base, the integrated detection and prediction model is updated using transfer learning or incremental training to adapt the model parameters to the new anomaly patterns and reduce the false negative rate when similar anomalies recur. At the same time, the knowledge base supports statistical analysis of anomaly patterns to assist maintenance personnel in developing preventive measures. Through the interpretability analysis results of typical cases in the knowledge base and the model attention weights, this method can analyze the root causes of anomalies and their impact scope, thereby improving the intelligence level of new energy power plant operation monitoring.

[0058] In addition, before calling the large language model to generate an anomaly report, a prompt message is designed, including contextual information such as the name of the detected anomaly event, the equipment or sensor identifier, the time of the anomaly, relevant data indicators, model detection results, and the preliminary anomaly type as part of the input prompt. Combined with the requirements for the report content and format, the large language model is guided to generate an anomaly report that meets the operation and maintenance needs of new energy power stations in accordance with the predetermined style and requirements.

[0059] After the large language model generates an anomaly report, the report content is verified and reviewed. If any information in the report is found to be inconsistent with the actual situation or any inappropriate suggestions are found, experienced engineers will correct it and store the optimized report instance in the knowledge base for reinforcement learning training of the large language model or adjustment of the prompting strategy, so that the anomaly reports generated later are more in line with the actual operation and maintenance needs.

[0060] In summary, the anomaly detection method for new energy power plants provided by this invention shares the same graph neural network structure for both the prediction model and the anomaly detection model, enabling collaborative training. The prediction task helps the model better understand normal data patterns and reduces the false alarm rate; anomaly detection feedback can make the model more sensitive to anomalies, improving its ability to predict and respond to emergencies. Furthermore, by leveraging physical topology, geographical proximity, and power correlation graphs, this method can more accurately distinguish between local anomalies and global changes. When a device's data is abnormal, its anomaly is more easily captured and located by the model because the nodes connected to it affect the attention weights; conversely, if the global change is caused by environmental factors, the model will identify it as a normal trend rather than a fault based on geographical correlation.

[0061] Furthermore, the method of this invention can accommodate various modes such as point anomalies, interval anomalies, and gradual deviations. In particular, by utilizing the two perspectives of prediction branch and reconstruction branch, both instantaneous impact anomalies caused by large prediction errors and mode-changing anomalies caused by large reconstruction errors can be detected, reducing missed detections. The use of dynamic thresholds improves adaptability to different operating states.

[0062] Furthermore, the automated report generation through large-scale models significantly saves manpower and shortens response time. The analysis in the reports can also cover multiple hypothetical causes and recommendations, preventing omissions due to insufficient human experience. At the same time, the establishment of a knowledge base ensures the inheritance and sharing of experience; new failures do not appear out of thin air without any trace, and both models and personnel can learn from history and continuously improve.

[0063] Example 2 Based on Example 1, such as Figure 2 As shown, this invention also provides a specific example of a method for detecting abnormal data and predicting power at renewable energy power plants. Its core is to simultaneously achieve high-precision prediction of power plant power and timely detection and identification of abnormal data within a single framework. Renewable energy power plants can be various renewable energy power plants such as wind power plants or photovoltaic power plants; the method is not limited to any particular type of power plant. By introducing the attention mechanism of graph neural networks and combining various relationships between equipment and measuring points within the power plant, this method constructs an integrated detection and prediction model, enabling the fusion analysis of multi-source data. In actual operation, this model can learn the characteristic patterns of normal power plant operation based on historical data, predicting future power while monitoring real-time data, and quickly detecting any abnormal deviations. Furthermore, this method integrates a large-scale pre-trained language model to generate detailed anomaly reports, helping maintenance personnel understand abnormal situations and take measures; simultaneously, it establishes an anomaly data knowledge base to accumulate experience and continuously improve the model's detection and prediction capabilities. The following will describe the method of this invention in detail with specific implementation steps.

[0064] First, core operational data from renewable energy power plants are acquired as model inputs. This core operational data includes data on key external factors affecting renewable energy power generation, such as environmental meteorological data reflecting the environment in which the power plant is located (e.g., wind speed, wind direction, temperature, atmospheric pressure for wind farms, and solar irradiance, ambient temperature, and wind speed for photovoltaic power plants); operational data reflecting equipment performance and power output capabilities, such as equipment operational data reflecting the working status of each power generation unit or related equipment within the power plant (e.g., rotor speed, generator temperature, grid frequency, voltage, current, power factor, blade pitch angle, and yaw angle for wind turbines, and component voltage, current, inverter output power, and battery temperature for photovoltaic arrays); and operational data crucial for identifying anomalies, such as the active power output time series for a single unit or the entire power plant, and historical data on related reactive power and power quality indicators.

[0065] After collecting the aforementioned data, preprocessing of this data from different sources is necessary to ensure the quality and consistency of the input data. Specifically, since different data acquisition devices may have asynchronous timestamps or different sampling frequencies, the multi-source data needs to be aligned according to a unified time scale. A unified time step can be selected to resample and interpolate the data, ensuring that each time point has corresponding multi-dimensional feature values. Then, for missing data caused by data gaps or communication interruptions in the acquisition system, methods such as interpolation, previous value imputation, or estimation based on other relevant variables are used to fill in the gaps. If the missing time is long and cannot be reliably filled, this time period can be specially marked in the model to avoid erroneous data affecting model training. Before training the model, simple rules or statistical methods can be applied to remove obviously erroneous outliers to reduce noise interference. Data points exceeding the physical limits of the equipment should be marked for later focus. Furthermore, to improve the model's learning effect on normal patterns, high-frequency noise can be smoothed to reduce the interference of random noise on prediction and anomaly detection. In addition, features of different dimensions can be normalized or standardized.

[0066] After the multi-source running data has undergone the above preprocessing, a multivariate time series data tensor for model input can be obtained. Let the preprocessed time series data tensor be denoted as... ,in Indicates a time index. This represents a feature index, such as a sensor measurement point index. The tensor of the preprocessed time-series data is also included. It can be organized into dimensions It is in matrix form, with each row corresponding to a timestamp and each column corresponding to a feature. It considers the total length of the time series. It is the number of features.

[0067] Subsequently, considering the rich interrelationships among various devices and sensors within the new energy power station, this invention incorporates multiple graph structure information into the model to fully utilize these relationships and enhance the model's ability to detect anomalies and predict power. These include a physical topology map reflecting the electrical wiring and topological connections of devices within the station, a geographical adjacency map reflecting the spatial proximity of devices or substations, and a power correlation map constructed based on the correlation relationships of device power or key parameters obtained from historical data statistics.

[0068] In a physical topology diagram, each node represents a device or measurement point in the power station. If two nodes have a direct connection in the actual circuit or equipment structure, they are connected by an edge in the diagram. For example, the feeder relationship of each wind turbine in a wind farm to the power station substation, and the connection relationship of modules-combiner boxes-inverters-step-up transformers in a photovoltaic power station can all be represented in a physical topology diagram.

[0069] In a geographic adjacency graph, nodes still represent equipment or sensors within a site. If two devices are relatively close in space, an edge is connected in the graph. Since adjacent devices are often affected by similar environmental conditions such as wind and sunlight, for example, wind turbines that are very close together in a wind farm may have correlated wind speed and power changes; adjacent modules in a photovoltaic field are subjected to similar sunlight and temperature conditions. Geographic adjacency graphs can help models learn cooperative change patterns among spatially adjacent devices.

[0070] In a power correlation diagram, nodes represent various devices or measurement points. Key indicators such as power output, power factor, and temperature, which represent the operating status of the devices, are selected to calculate the correlation coefficient between nodes. If two nodes have a high correlation coefficient over a long historical period, such as exceeding a certain correlation threshold, connecting them in the diagram indicates a strong correlation between them.

[0071] Specifically, the three types of graphs mentioned above can be formally represented as adjacency matrices. , and Regarding the total number of nodes Each adjacency matrix is Dimension. Among its elements Represents nodes in the physical topology and If there is a direct connection, then the value is 0; It can be given by the distance decay function, for example, the closer the distance, the higher the weight, or use 0 or 1 to indicate whether they are adjacent; Then it can be processed according to the historical correlation coefficient threshold.

[0072] The integrated detection and prediction model provided by this invention uses a graph attention network as its core to achieve joint analysis of multi-sensor, multi-time-series data. The model input is a preprocessed multi-dimensional time-series data window. ,Include These features are in recent The model outputs a power prediction for the next time step and a judgment on whether there are any anomalies at the current time step through a series of neural network layers.

[0073] In terms of overall architecture, the model mainly includes the following modules: temporal feature extraction module, graph attention feature fusion module, sequence pattern capture module, prediction branch and reconstruction branch.

[0074] The temporal feature extraction module first performs a temporal feature extraction operation on the input multivariate time series to smooth noise and extract local temporal patterns. This step can be accomplished through a one-dimensional convolutional layer, convolving the time series of each feature to extract higher-order features; or through a shallow recurrent neural network to process each feature to obtain a preliminary temporal dependency representation. In this embodiment, a one-dimensional convolution is used, which is equivalent to... Perform column-wise convolution filtering to output a feature matrix of the same size. ,in Representative characteristics At any moment The initial higher-order representation.

[0075] Furthermore, the model introduces two graph attention network layers in parallel to fuse inter-feature and inter-temporal relationship information. The feature-oriented graph attention network layer treats each feature within the input time window as a node in the graph, constructing a feature graph based on the comprehensive adjacency information obtained in the "Station Relationship Graph Construction" step. In the simplest implementation, the three relationship graphs can be merged to represent whether a certain association exists between any two feature nodes. Alternatively, multiple relationships can be retained and processed separately using multi-head attention. In the feature graph, nodes... The initial representation can be taken as the entire sequence of the feature in the current window. Alternatively, one could simply take the representation of that feature at the current moment. Then, combined with its past trends, the graph attention network layer will use an adjacency matrix-guided attention mechanism to calculate the relational weights between each feature node, outputting a representation vector for each feature node that incorporates information from its neighbors. Let the feature graph have... If there are 10 nodes, the output of this layer can be represented as a matrix. The same dimension as the input but each element Features are already included Information after interaction with relevant features. The time-oriented graph attention network layer treats each time point within the input window as a node in a graph, connecting them in a complete graph manner. That is, in the time graph, there are... Time nodes Any two different time points are considered neighbors. Each time point The initial features can be a vector composed of all features at that time. , indicating the time of the station The state. Therefore, the feature vector dimension of each node in the time graph is... The graph attention network layer learns attention weights between time points on this fully connected time graph, that is, it identifies which time points have similar system states or similar patterns. The output of the time graph attention network layer is shown below. A new representation for each time node, incorporating neighbor information for each time node. The output can be represented as... It has the same dimension but includes time-related information.

[0076] The extracted features are then processed by the sequence pattern capture module or the GRU layer. and information that integrates relational information and Merging is performed to capture longer periods or complex patterns in the time series. Specifically, the three data points can be concatenated or cascaded along the feature dimensions to obtain a fused representation. For example, combining along the feature dimension to form A 3D matrix. Then, As input, it is fed into a Gated Recurrent Unit (GRU) layer. A GRU is a type of recurrent neural network layer that efficiently learns long-term dependencies through its update and reset gate mechanisms. Here, the GRU layer reads the fused feature vector step-by-step, where each time step has a dimension... By accumulating previously remembered information through hidden states, the GRU layer captures sequence patterns. The hidden state dimension of the GRU layer can be chosen as appropriate. The model learns on its own. After the loop, the GRU layer at the last moment Output a hidden vector It encapsulates the dynamic changes of various features within a window period. This hidden vector will serve as the basis for subsequent predictions and anomaly detection.

[0077] Then, the prediction branch is derived from the output of the GRU layer. The prediction network is used to infer the power value at future times. The prediction task can be single-step or multi-step; here, we take single-step prediction as an example. The prediction network can consist of several fully connected layers. For example, a three-layer fully connected network can be used, with the first two layers using ReLU activation and the last layer having a linear output. The power output, which is of particular interest, is contained in one or more feature components. When only the total power is of interest, a single value can be output at the end of the network. The goal of a prediction network during training is to minimize the error between the predicted and true values, for example, by using mean squared error (MSE) loss. Optimization is performed to enable the model to learn to accurately predict the future. By employing multivariate inputs, the predictive model can provide more accurate power predictions based on the current state of multiple variables.

[0078] In parallel, from the output of the GRU layer An access reconstruction network is used to reconstruct the input time series. The reconstruction network is the decoding part of an autoencoder, recovering the original input from the encoded hidden vectors, thus approximating the distribution of the input data. In implementation, it will... By upsampling or loop unrolling, the reconstructed length is... The multidimensional sequence. The training objective of the reconstruction branch is to minimize the error between the reconstructed sequence and the original input sequence.

[0079] The model's total loss function can be designed as a weighted sum of prediction and reconstruction losses. During training, the model learns an internal representation that accurately predicts and effectively encodes normal behavior by simultaneously minimizing both prediction and reconstruction errors. During inference, for each new time window input, the model produces two types of outputs: one is the predicted value of future power from the prediction branch, and the other is the reconstruction of the current window data from the reconstruction branch.

[0080] It is important to emphasize that, in order to ensure the accuracy of power prediction, abnormal data detection can be performed based on the output of the above model, and the data can be corrected based on the results of abnormal data detection, so as to clarify a more suitable power prediction result. For example, if the actual observed data differs significantly from the normal behavior predicted or reconstructed by the model, it is judged as abnormal and corrected based on the predicted value.

[0081] Specifically, the first step is anomaly score calculation. This is done for the last moment of the current time window. Obtain the next time-step data predicted by the model and the data obtained from model reconstruction. The actual observed value is Define each feature The prediction error is The reconstruction error is These two values ​​reflect the degree of deviation between the current value and the model's "expectation" in that dimension. They can then be combined to obtain anomaly scores for a single dimension or the overall score. To obtain an overall anomaly index, the scores from each dimension can be further aggregated; the aggregation method depends on the application requirements.

[0082] Secondly, to adapt to changes in data volatility under different operating conditions and seasons, this invention preferably employs dynamic thresholds to determine anomalies. Specifically, at each time point, the current threshold is adaptively determined based on the statistical analysis of anomaly scores over a recent period. Furthermore, thresholds can be set separately for the error of each feature.

[0083] Dangdang Exceeding the current threshold When, then it is determined that in Anomalies can occur at any time. At this point, the source of the anomaly can be further located by identifying which features contribute the most to the error. This invention utilizes the attention weights of a graph attention network model to assist in classifying anomaly types. Combined with prior knowledge, anomalies are categorized into the following types: physical limit anomalies where data exceeds physical or safe operational thresholds. For example, model predictions typically do not produce values ​​that violate physical laws; therefore, once a physical limit anomaly occurs... and The values ​​of all outliers will be extremely high, and the outlier score will increase significantly. Occasional single-point outliers, random outliers, or noise, for example, may be partially smoothed out by the autoencoder. Therefore, these outliers, because they do not conform to historical patterns, will cause a large bias in the model's predictions, but the reconstruction error may be slightly smaller and can be mitigated by a high single-point outlier. Identify systematic biases caused by a persistent trend of deviation from the normal data. Since the anomalies do not manifest as sudden points but rather as a slow deviation, they can be identified by reconstructing errors or comparing with similar devices to discover changes in the relationship between the sensor and others, such as changes in correlation attention weights, or the device being in an unreasonable state for a long period of time. Also identify missing or invalid data, such as communication anomalies causing all data to be empty or stagnant for a period of time. Missing data can be identified in the preprocessing stage and skipped in the model prediction or given a special high-weight penalty.

[0084] Furthermore, the method of this invention can also cover equipment malfunction anomalies, sensor freezing anomalies, etc. Different patterns can be defined in the knowledge base for further subdivision. However, regardless of the type of anomaly, the criterion for judgment is "the actual situation does not match the model's expectations," the only difference being the pattern and scope of the discrepancy. Once an anomaly is determined to have occurred, anomaly-related information will be recorded.

[0085] Finally, when an anomaly is detected, traditional methods typically require manual querying of various data to compile a report. This invention, however, introduces a large-scale language model to automatically generate professional anomaly reports. This process includes three stages: prompt design, model generation, and result verification.

[0086] In the prompt design phase, to ensure that the large language model outputs reports that meet the operation and maintenance needs of new energy power plants, the input prompt information must first be designed. Prompts generally include two parts: contextual information and task requirements. Contextual information refers to key information about the anomaly provided by the detection module, such as power plant and equipment information, time information, data indicators, model results, preliminary anomaly type judgment, and other relevant information. Task requirements, in natural language, describe the desired content and format for the large model's output, limiting the response style and scope.

[0087] In the model generation phase, after receiving prompts, the large model uses the language and knowledge learned during training to write a report. Because the large model possesses contextual understanding and knowledge generalization capabilities, it can not only organize the provided data but also potentially infer the causes by combining its stored power system knowledge. For example, "Anomaly Overview: At 14:35 on December 1, 2025, wind turbine No. 5 in wind farm A experienced a sudden power drop anomaly, with output power decreasing from approximately 500kW to 0kW, lasting for about 5 minutes. Possible Cause Analysis: According to on-site SCADA data, the wind speed was approximately 10m / s when the anomaly occurred, while other wind turbines continued to generate power normally. This indicates that wind turbine No. 5 itself is more likely to have experienced an abnormal shutdown. Common causes may include: ① The wind turbine triggered protection due to a fault (such as a gearbox or generator failure); ② Communication interruption of the wind turbine led to abnormal power readings." 0 (but other measurements such as speed and status signals should also be abnormal during the same period); ③ The grid-side protection action caused the wind turbine to trip. Considering that the vibration and temperature data of the wind turbine were not significantly abnormal in the 5 minutes prior to the anomaly, the possibility of a fault is low; it is more likely a communication interruption or a grid-related issue. Recommended measures: Please have on-site maintenance personnel check the status of wind turbine No. 5 as soon as possible. First, verify whether the communication between the SCADA system and wind turbine No. 5 is normal. If the communication module indicator light is abnormal, restart the communication module. If communication is normal, check whether the wind turbine's local controller has any fault alarms or tripping records, and perform corresponding repairs according to the fault category. Before reconnecting to the grid, confirm that all components of the wind turbine are operating normally. At the same time, notify the dispatch center to temporarily shut down the wind turbine, with an estimated repair time of…”. In this report, the large model provides a comprehensive description combining specific data and general power knowledge. It covers everything from event descriptions such as time, equipment, and phenomena, listing multiple possibilities and combining data trend analysis to specific inspection steps and communication requirements.

[0088] During the result verification phase, since the model may generate information that is inconsistent with the actual situation based on its training knowledge, experienced engineers need to quickly review and confirm it when necessary. If the measures suggested by the model are inappropriate, they need to be corrected in a timely manner. To this end, this invention can accumulate excellent report examples in a knowledge base and use reinforcement learning to adjust the model, making its output more in line with actual needs.

[0089] Furthermore, in the method of this invention, the anomaly data knowledge base plays a crucial role in storing experience and providing feedback for optimization. Specifically, the establishment and utilization of the knowledge base includes the following aspects:

[0090] First, there's the knowledge base data structure. The knowledge base can be stored in the form of a database or documents, with each record corresponding to an anomaly event. Key fields include: Event ID, Occurrence Time, Site, Equipment, Anomaly Description, Data Snapshot, Model Detection Results, Anomaly Cause Classification or Manual Confirmation Result, Handling Measures, Result Feedback, and Related Attention Weight Analysis. The "Data Snapshot" can store key sensor and power curves before and after the anomaly, for future review and analysis. The "Model Detection Results" include the model's output anomaly score and major anomaly variables. The "Manual Confirmation Result" records the actual cause entered by maintenance personnel after investigation, such as determining whether it's a sensor malfunction or an external cause. The "Handling Measures" record the actions taken, such as replacing equipment or repairing communications, and whether these actions were effective. The "Attention Weight Analysis" is optional information, recording how certain attention coefficients of the graph attention network model changed during the event to support the understanding of related anomalies. For example, if the correlation attention between a wind turbine and surrounding wind turbines significantly decreases during an anomaly, it may indicate an independent fault in that turbine.

[0091] Secondly, there's the accumulation of knowledge base information. As time progresses, each time an anomaly occurs and is resolved, relevant information should be entered into the knowledge base. For example, standardized entry templates can be developed to ensure the completeness and consistency of information. The classification labels for "four types of anomalies" can be used as a field to facilitate the statistics of the frequency of similar events. This embodiment pays particular attention to deviations caused by equipment aging. Such anomalies are often difficult to detect initially but gradually become apparent; therefore, the knowledge base is invaluable for recording long-term changes.

[0092] Finally, there's model optimization based on the knowledge base. Since the primary purpose of the knowledge base is to feed back into the model, newly added anomalous data can be used periodically to retrain or fine-tune the detection and prediction model. Specific methods include supervised fine-tuning, semi-supervised enhancement, transfer learning to correct biases, knowledge fusion with large models, and auxiliary decision-making and analysis.

[0093] Example 3 This invention also provides a new energy power station abnormal data detection system, comprising: The data preprocessing module is used to acquire environmental meteorological data, equipment operating status data, and historical power output data of the new energy power station to be evaluated; and to perform time alignment on the environmental meteorological data, equipment operating status data, and historical power output data to obtain time series data. The feature fusion module is used to extract time dependencies from time series data to obtain time pattern features; based on the graph attention mechanism, the time pattern features are modeled with time relationships and sensor relationships respectively through the graph structure corresponding to the new energy power station to be evaluated, to obtain time series features and sensor relationship features; the time pattern features, time series features and sensor relationship features are fused to obtain fused features; The abnormal data detection module is used to reconstruct data based on fusion features; and to determine the abnormal data detection results of the new energy power station to be evaluated based on the reconstruction error of the reconstructed data and the corresponding equipment operation status data and environmental meteorological data.

[0094] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the steps in an embodiment of a method for detecting abnormal data at a new energy power station. Specific implementation methods can be found in the method embodiments, and will not be repeated here.

[0095] Furthermore, the present invention also provides a non-transitory computer-readable storage medium containing instructions, on which a computer program is stored. For example, a memory containing instructions, which can be executed by a processor of a computer device to complete the above-described method. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc. When the computer program is executed by the processor, it can implement the steps in an embodiment of a method for detecting abnormal data at a new energy power station. Specific implementation methods can be found in the method embodiments, which will not be repeated here.

[0096] Those skilled in the art will understand that embodiments of the present invention can provide methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0097] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxesFigure 1 A device that provides the functions specified in one or more boxes.

[0098] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0099] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0100] It should be noted that the specific embodiments described above enable those skilled in the art to more fully understand the present invention, but do not limit the present invention in any way. Therefore, although the present invention has been described in detail in this specification and embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention; and all technical solutions and improvements that do not depart from the spirit and scope of the present invention are covered within the protection scope of the present invention patent. No reference numerals in the claims should be construed as limiting the scope of the claims. Any simple variations or equivalent substitutions of technical solutions that can be readily obtained by those skilled in the art within the scope of the technology disclosed in the present invention are within the protection scope of the present invention.

Claims

1. A method for detecting abnormal data at new energy power stations, characterized in that, include: Acquire environmental meteorological data, equipment operating status data, and historical power output data of the new energy power plants to be evaluated; Time-aligned data is obtained by aligning the environmental meteorological data, equipment operating status data, and historical power output data. Temporal dependencies are extracted from the time series data to obtain temporal pattern features. Based on the graph attention mechanism, the temporal and sensor relationships are modeled on the time pattern features through the graph structure corresponding to the new energy power station to be evaluated, to obtain time series features and sensor relationship features. The time pattern features, time series features, and sensor relationship features are fused to obtain the fused features; The reconstructed data is obtained based on the fusion features; the abnormal data detection results of the new energy power station to be evaluated are determined based on the reconstruction error of the reconstructed data and the corresponding equipment operation status data and environmental meteorological data.

2. The method for detecting abnormal data at new energy power stations according to claim 1, characterized in that, The time series data is processed by a graph attention mechanism detection and prediction integrated model to obtain abnormal data detection results. The graph attention mechanism detection and prediction integrated model includes a time series feature extraction module, a graph attention feature fusion module, a sequence pattern capture module, and a reconstruction branch module connected in sequence. The graph attention feature fusion module includes a parallel feature-oriented graph attention network layer and a time-oriented graph attention network layer. The time-series feature extraction module extracts time-pattern features from the time-series data. Sensor relationship features are extracted from the temporal pattern features through the feature-oriented graph attention network layer; Temporal features are extracted from the temporal pattern features through the time-oriented graph attention network layer; The sequence pattern capture module fuses the time pattern features, time series features, and sensor relationship features to obtain the fused features; The reconstruction branch module determines the abnormal data detection result based on the fusion features.

3. The method for detecting abnormal data at new energy power stations according to claim 1, characterized in that, Time-aligned data from the environmental meteorological data, equipment operating status data, and historical power output data yields time-series data including: Based on a preset time scale, the environmental meteorological data, equipment operating status data and historical power output data are aligned to obtain time-aligned multi-source data. When missing values ​​appear in the time-aligned multi-source data, the estimated values ​​are filled by interpolation or previous values ​​to obtain the multi-source data after missing value filling. Outliers are removed from the multi-source data after missing value imputation to obtain multi-source data after outlier removal; The time series data is obtained by denoising and normalizing the multi-source data after outlier removal.

4. The method for detecting abnormal data at new energy power stations according to claim 2, characterized in that, Based on the prior relationships between the equipment and sensor points within the new energy power station to be evaluated, multiple graph structures are constructed to characterize the correlation between the environmental meteorological data, equipment operating status data, and historical power output data. The graph structures include a physical topology graph, a geographical adjacency graph, and a power correlation graph. Before modeling the temporal relationships and sensor relationships of the time pattern features using the graph structures corresponding to the new energy power station to be evaluated, the following steps are also included: The physical topology graph is constructed by taking the equipment or sensor measurement points in the new energy power station to be evaluated as nodes and the electrical connection relationship and equipment structural connection relationship of the equipment or sensor measurement points as edges. Using the equipment or sensor measurement points within the new energy power station to be evaluated as nodes and the geographical locations of the equipment or sensor measurement points as edges, construct the geographical adjacency graph. Using the equipment or sensor measurement points in the new energy power station to be evaluated as nodes, a correlation coefficient is calculated based on the historical power output data corresponding to the equipment or sensor measurement points. When the correlation coefficient is greater than a preset correlation threshold, edges are constructed between the nodes corresponding to the correlation coefficient to obtain the power correlation graph.

5. The method for detecting abnormal data at new energy power stations according to claim 4, characterized in that, Extracting sensor relationship features from the temporal pattern features through the feature-oriented graph attention network layer includes: The weights of the physical connection relationships between device or sensor measurement points are determined based on the physical topology diagram. The environmental impact correlation weights between device or sensor measurement points are determined based on the geographical adjacency map. The operational correlation weights between device or sensor measurement points are determined based on the power correlation diagram. The comprehensive attention weight matrix is ​​determined based on the weights of physical connectivity, environmental impact, and operational relevance. Based on the comprehensive attention weight matrix, the sensor relationship features are extracted from the temporal pattern features using a graph attention mechanism.

6. The method for detecting abnormal data at new energy power stations according to claim 5, characterized in that, The abnormal data detection results of the new energy power station to be evaluated are determined based on the reconstruction errors of the reconstructed data and the corresponding equipment operating status data and environmental meteorological data, including: Calculate the reconstructed data of the device or sensor measurement point and the corresponding device operating status data or environmental meteorological data to obtain the reconstructed error of the device or sensor measurement point. The abnormal indicators of the new energy power station to be evaluated are obtained by aggregating the reconstruction error of the measurement points of the equipment or sensor, and the abnormal data detection results of the new energy power station to be evaluated are determined by the abnormal indicators.

7. The method for detecting abnormal data at new energy power stations according to claim 6, characterized in that, The abnormal data detection results include outlier anomalies. The graph attention mechanism detection and prediction integrated model also includes a prediction branch. After determining the abnormal data detection results of the new energy power station to be evaluated based on the anomaly indicators, it also includes: When the abnormal data detection result is an outlier, the estimated value of the device or sensor measurement point is obtained by filling the estimated value with the data of the nearby nodes of the device or sensor measurement point corresponding to the outlier. The nearby nodes are determined based on the physical topology map, the geographical adjacency map and the power correlation map. Replace the outlier data corresponding to the outlier with the estimated data value; The prediction branch, which is sequentially connected by the time-series feature extraction module, graph attention feature fusion module, and sequence pattern capture module, extracts the replaced fused features from the replaced environmental meteorological data, equipment operating status data, and historical power output data. The power prediction result of the new energy power station to be evaluated is determined by the fusion feature through the prediction branch.

8. The method for detecting abnormal data at a new energy power station according to claim 7, characterized in that, The abnormal data detection results also include systematic biases. After determining the abnormal data detection results of the new energy power station to be evaluated based on the abnormal indicators, the results also include: When the abnormal data detection result is a systematic deviation, the weights of the physical connection relationship, environmental impact association weight, and operation-related relationship weight of the equipment or sensor measurement point corresponding to the systematic deviation are adjusted. The adjusted fusion features are then extracted from the environmental meteorological data, equipment operating status data, and historical power output data using the adjusted weights of the physical connection relationship, environmental impact association weight, and operation-related relationship weight. The power prediction result of the new energy power station to be evaluated is determined by the fusion feature adjusted by the weights of the prediction branch.

9. The method for detecting abnormal data at a new energy power station according to claim 8, characterized in that, After determining the abnormal data detection results of the new energy power station to be evaluated based on the abnormal indicators, the method further includes: An anomaly report is generated from the anomaly detection results using a pre-trained large-scale language model. The anomaly report includes the anomaly type of the anomaly detection results, the corresponding device or sensor measurement point information, and suggested measures. The suggested measures are determined based on semantic matching between the anomaly type, the corresponding device or sensor measurement point information, and a pre-built anomaly data knowledge base.

10. A new energy power station abnormal data detection system, characterized in that, include: The data preprocessing module is used to acquire environmental meteorological data, equipment operating status data, and historical power output data of the new energy power stations to be evaluated; Time-aligned data is obtained by aligning the environmental meteorological data, equipment operating status data, and historical power output data. The feature fusion module is used to extract time dependencies from the time series data to obtain time pattern features; based on the graph attention mechanism, the time pattern features are modeled with time relationships and sensor relationships respectively through the graph structure corresponding to the new energy power station to be evaluated, to obtain time series features and sensor relationship features. The time pattern features, time series features, and sensor relationship features are fused to obtain the fused features; An abnormal data detection module is used to reconstruct data based on the fusion features; and to determine the abnormal data detection result of the new energy power station to be evaluated based on the reconstruction error of the reconstructed data and the corresponding equipment operating status data and environmental meteorological data.

11. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the abnormal data detection method for new energy power stations as described in any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to execute the steps of the abnormal data detection method for new energy power stations as described in any one of claims 1 to 9.

Images