A method and system for identifying abnormal photovoltaic data in a substation

Abstract

The present invention relates to a method and system for identifying photovoltaic data anomalies in a substation, which belongs to the field of photovoltaic power generation technology and solves the problem of inaccurate identification and attribution of anomalies in existing photovoltaic data. It includes: collecting and preprocessing photovoltaic power generation data and meteorological data in the substation, obtaining data at each time point in the time period to be analyzed; transferring the data at each time point into the substation power generation prediction model to obtain the power generation prediction value at each time point; combining the power generation prediction value at each time point with the actual power generation value at the same time point in the photovoltaic power generation data in the substation to construct data point pairs at each time point; using a clustering algorithm to obtain abnormal point pairs from the data point pairs at each time point; obtaining corresponding substation business data according to the time point of the abnormal point pair, and obtaining the rule category corresponding to the substation business data according to the preset photovoltaic business abnormality rules as the abnormal cause of the corresponding abnormal point pair. Efficient and accurate abnormality identification and attribution are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a method and system for identifying photovoltaic data anomalies in a substation. Background Art

[0002] The identification of abnormal photovoltaic power generation data is crucial to ensuring the safe and economical operation of photovoltaic power stations. It is of great significance to improving the efficiency of photovoltaic power generation, reducing operation and maintenance costs, and ensuring the safe and stable operation of power stations.

[0003] The identification of abnormal data in photovoltaic power generation includes obtaining abnormal data and analyzing the causes of the abnormality. Some existing technologies only focus on identifying abnormal data and issuing early warnings, which require manual analysis; some compare the actual power generation curve with the standard power generation curve. If there is an inconsistency, it is considered that the power generation equipment has failed, and then the operating status of the photovoltaic power generation equipment is monitored and diagnosed; some are judged manually according to set abnormal judgment rules.

[0004] Existing anomaly recognition methods have difficulty in identifying complex and diverse data anomalies, and the causes of anomalies are relatively simple, resulting in low anomaly recognition efficiency and inaccurate recognition results and anomaly attribution. Summary of the invention

[0005] In view of the above analysis, the embodiments of the present invention aim to provide a method and system for identifying photovoltaic data anomalies in a substation, so as to solve the problems of inaccurate identification and attribution of anomalies in existing photovoltaic data.

[0006] On the one hand, an embodiment of the present invention provides a method for identifying abnormal photovoltaic data in a substation, comprising the following steps:

[0007] Collect and pre-process the photovoltaic power generation data and meteorological data of the substation area to obtain the data at each time point in the period to be analyzed;

[0008] The data at each time point is transferred into the power generation prediction model of the substation area to obtain the power generation prediction value at each time point;

[0009] The predicted power generation value at each time point is compared with the actual power generation value at the same time point in the photovoltaic power generation data of the substation area to construct the data point pairs at each time point; the clustering algorithm is used to obtain the abnormal point pairs from the data point pairs at each time point;

[0010] The corresponding area business data is obtained according to the time point of the abnormal point pair, and the rule category corresponding to the area business data is obtained according to the preset photovoltaic business abnormal rule as the abnormal cause of the corresponding abnormal point pair.

[0011] Based on the further improvement of the above method, the photovoltaic power generation data and meteorological data of the substation area are preprocessed to obtain the data at each time point in the period to be analyzed, including:

[0012] During the period to be analyzed, the actual power generation value in the photovoltaic power generation data of the substation is collected at each time point as the reference time point, and the data of total solar radiation intensity, net solar radiation intensity, hourly temperature and hourly relative humidity at each reference time point are supplemented by linear interpolation method;

[0013] The actual power generation value at each benchmark time point is taken as the real data, and the total solar radiation intensity, net solar radiation intensity, hourly temperature and hourly relative humidity at each benchmark time point are taken as characteristic data to obtain the data at each time point in the period to be analyzed.

[0014] Based on the further improvement of the above method, the power generation prediction model of the substation area is obtained by optimizing the parameters of the multilayer perceptron using the improved fruit fly optimization algorithm to obtain the initial optimal parameters and then training the multilayer perceptron again.

[0015] Based on the further improvement of the above method, the position of the fruit fly in the improved fruit fly optimization algorithm is a row vector composed of the weight matrix and the bias vector in the multi-layer perceptron, and the fitness value of the improved fruit fly optimization algorithm is obtained by taking the position of the fruit fly as the parameter of the multi-layer perceptron to predict the training samples, and calculating the sum of the root mean square error and the mean absolute error based on the prediction results.

[0016] Based on the further improvement of the above method, in the improved fruit fly optimization algorithm, the fruit fly with the smallest fitness value is taken as the leader in each iteration. When the difference in fitness value between other fruit flies and the leader is greater than or equal to a threshold, the position is updated according to the current position of the fruit fly and the current moving distance of the fruit fly; otherwise, the weighted average position of all fruit flies whose fitness value difference with the leader is less than the threshold is calculated, and the position is updated based on the weighted average position and the current moving distance of the fruit fly; the current moving distance of the fruit fly is obtained by calculating the product of the position difference between the fruit fly and the leader and the update step size of the current iteration.

[0017] Based on the further improvement of the above method, a clustering algorithm is used to obtain abnormal point pairs from the data point pairs at each time point, including: using the DBSCAN algorithm to cluster the data point pairs at each time point to obtain multiple categories, calculating the average absolute error between the predicted power generation value and the actual power generation value of each category, taking the one with the smallest average absolute error as the normal category, and the other categories as abnormal categories; the data point pairs at each time point in the abnormal category are abnormal point pairs.

[0018] Based on the further improvement of the above method, the photovoltaic business abnormality rule is a judgment condition set according to the three rule categories of meter reading abnormality, user behavior abnormality and power generation equipment failure.

[0019] Based on the further improvement of the above method, the judgment condition for meter reading abnormality is that the current actual power generation is 0 but the power generation of last month is not 0; the user behavior abnormality is subdivided into over-capacity, power generation exceeding contract capacity and night power generation, among which the judgment condition for over-capacity is that the actual power generation exceeds 30% of the regional theoretical value; the judgment condition for the actual power generation exceeding the contract capacity is that the instantaneous active value in the preset daytime period is greater than the contract capacity; the judgment condition for night power generation is that the current and actual power generation in the preset nighttime period are both greater than the minimum threshold; the judgment condition for power generation equipment failure is that the actual power generation is 0 for three consecutive days.

[0020] Based on the further improvement of the above method, the power generation prediction model of the substation area is obtained by freezing the parameters of the power generation prediction model trained in the non-current substation area except the last hidden layer and the output layer, and retraining it according to the training samples of the current substation area.

[0021] On the other hand, an embodiment of the present invention provides a system for identifying abnormal photovoltaic data in a substation, including:

[0022] The data preprocessing module is used to collect and preprocess the photovoltaic power generation data and meteorological data of the substation area, and obtain the data at each time point in the period to be analyzed;

[0023] The power generation prediction module is used to input the data at each time point into the power generation prediction model of the substation area to obtain the power generation prediction value at each time point;

[0024] The abnormal data identification module is used to construct data point pairs at each time point by comparing the predicted power generation value at each time point with the actual power generation value at the same time point in the photovoltaic power generation data of the substation area; and to obtain abnormal point pairs from the data point pairs at each time point by using a clustering algorithm;

[0025] The abnormal data attribution module is used to obtain the corresponding area business data according to the time point of the abnormal point pair, and obtain the rule category corresponding to the area business data according to the preset photovoltaic business abnormal rules as the abnormal cause of the corresponding abnormal point pair.

[0026] Compared with the prior art, the present invention can achieve at least one of the following beneficial effects:

[0027] 1. Extract data features from photovoltaic data and meteorological data, and use a combination of machine learning judgment and business rule recognition judgment to increase the comprehensiveness of recognition and have a wide range of applicability; first screen out abnormal categories through clustering and then use abnormal rules to attribute the abnormal categories, which improves the efficiency of abnormal recognition and attribution.

[0028] 2. Combined with the application scenarios, the initialization position, position update and fitness value of the fruit fly in the fruit fly optimization algorithm are improved, and the parameters of the multi-layer perceptron are optimized. This speeds up the convergence speed and makes it easier to obtain the global optimal parameters, thereby improving the prediction accuracy of the power generation.

[0029] 3. The predicted value and actual value of power generation are used as two dimensions of data point pairs for cluster analysis. The running speed is fast and the data collection cost is low. The accuracy and precision of anomaly detection are high, which enhances the ability to identify anomalies in photovoltaic power generation data.

[0030] In the present invention, the above-mentioned technical solutions can also be combined with each other to achieve more preferred combination solutions. Other features and advantages of the present invention will be described in the subsequent description, and some advantages can become obvious from the description, or can be understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained through the contents particularly pointed out in the description and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The accompanying drawings are only used for the purpose of illustrating specific embodiments and are not to be considered as limiting the present invention. In the entire drawings, the same reference symbols represent the same components;

[0032] Figure 1 This is a flow chart of a photovoltaic data anomaly identification method in Embodiment 1 of the present invention;

[0033] Figure 2 This is a schematic diagram of the structure of a photovoltaic data anomaly identification system in Example 2 of the present invention. DETAILED DESCRIPTION

[0034] The preferred embodiments of the present invention are described in detail below in conjunction with the accompanying drawings, wherein the accompanying drawings constitute a part of this application and are used together with the embodiments of the present invention to illustrate the principles of the present invention, but are not used to limit the scope of the present invention.

[0035] Example 1

[0036] A specific embodiment of the present invention discloses a method for identifying photovoltaic data anomalies, such as Figure 1 As shown, the following steps are included:

[0037] S1. Collect and pre-process the photovoltaic power generation data and meteorological data of the substation area to obtain the data at each time point in the period to be analyzed.

[0038] It should be noted that each photovoltaic power station in each substation is equipped with a data collection system. The collected photovoltaic power generation data include but are not limited to: actual power generation value, total solar radiation intensity and net solar radiation intensity; the collected meteorological data include but are not limited to: hourly temperature, hourly dew point temperature, hourly relative humidity, hourly rainfall, hourly wind direction, hourly wind speed and hourly air pressure. The collected data are removed from obviously abnormal data, such as data with empty values ​​or too large values.

[0039] Furthermore, since the total solar radiation intensity, net solar radiation intensity and meteorological data are generally collected at hourly intervals, and the actual power generation value is collected at 15-minute intervals, the total solar radiation intensity, net solar radiation intensity and meteorological data are preprocessed by using the linear interpolation method, with the collection time points of the actual power generation values ​​as the reference time points.

[0040] Considering that a lot of data is initially collected, and some data are not highly correlated with actual power generation, in order to improve the accuracy of subsequent power generation prediction, this embodiment performs correlation analysis on the collected historical data and extracts strongly correlated data as feature data during implementation.

[0041] Specifically, after the historical data are supplemented with the data at each time point according to the above method, the normality test is performed on the data, such as the Shapiro-Wilk test method Shapiro-Wilk, the Pearson correlation analysis method is used for the data that conforms to the normality, and the Spearman correlation analysis method is used for the data that does not conform to the normality. The correlation between each data and the actual power generation value is calculated to obtain strongly correlated data. The strongly correlated data obtained are the total solar radiation intensity, the net solar radiation intensity, the hourly temperature and the hourly relative humidity, among which the hourly relative humidity is negatively correlated, and the other three data are positively correlated.

[0042] Through correlation analysis, the total solar radiation intensity, net solar radiation intensity, hourly temperature and hourly relative humidity are taken as the characteristic data at each time point, and the actual power generation value at each time point is taken as the real data. After normalization of the characteristic data and the real data, the data at each time point in the period to be analyzed are obtained.

[0043] Preferably, the data is normalized using the Min-max normalization method.

[0044] S2. The data at each time point is transferred to the power generation prediction model of the substation to obtain the power generation prediction value at each time point.

[0045] It should be noted that since the characteristic data and the target data in the data at each time point are not in a simple linear relationship, the present embodiment trains the multi-layer perceptron model to obtain the power generation prediction model for the substation area. Moreover, in order to overcome the shortcomings of the multi-layer perceptron model, such as overfitting, slow convergence speed, and easy to fall into local extreme values, the present embodiment uses the fruit fly optimization algorithm to optimize the parameters of the multi-layer perceptron to obtain the initial optimal parameters and then trains the multi-layer perceptron again to obtain the model.

[0046] It should be noted that the multilayer perceptron includes: input layer, hidden layer and output layer in sequence, the activation function uses the ReLU function, and the loss function uses the root mean square error. The multilayer perceptron model is constructed by setting the number of hidden layers and the number of neurons in each layer. The parameters of the multilayer perceptron include: weights and biases, where the weights are the parameters connecting the input layer neurons, the neurons between the hidden layers, and the neurons from the hidden layer to the output layer. Each connection has a weight; the bias is to provide a translation for the activation function of the neuron. In addition to the neurons in the input layer, all other neurons have a bias. In other words, there is a weight matrix and a bias vector between each pair of layers in the multilayer perceptron.

[0047] Furthermore, the Fruit Fly Optimization Algorithm (FOA) is a heuristic optimization algorithm that simulates the foraging behavior of fruit flies and is used to solve optimization problems. During the training process, the parameters of the multilayer perceptron are optimized using the improved fruit fly optimization algorithm to obtain the initial optimal parameters, including:

[0048] ① Initialize the fruit fly population and the position of each fruit fly. The position of the fruit fly is a row vector composed of the weight matrix and bias vector in the multi-layer perceptron.

[0049] It should be noted that the fruit flies in the fruit fly population are represented by , represents the total number of fruit flies, The position of the fruit fly is , The total number of items representing the position, that is, the total number of parameters of the multilayer perceptron.

[0050] Generate initial values ​​for each weight and bias at each fruit fly's position. The weight values ​​belonging to the same weight matrix are calculated based on the number of neurons in the previous layer and the number of neurons in the next layer connected by the weight matrix. The value range is calculated according to the following formula, and the initial value is generated using uniform distribution to ensure that the initial value of the weight is not too large or too small, which helps the propagation of the gradient in the network:

[0051] ,

[0052] in, and denote the lower and upper limits of the uniform distribution, respectively. and Represents the number of neurons in the previous layer and the number of neurons in the next layer connected by the weight matrix.

[0053] The initial value of each bias is a random number following a uniform distribution U[0,1).

[0054] ② Iterate and execute the following steps: use the position of each fruit fly as the parameter of the multi-layer perceptron to predict the training samples, calculate the sum of the root mean square error and the mean absolute error according to the prediction results to obtain each fitness value, and select the fruit fly with the smallest fitness value as the leader; update the position of the non-leader fruit fly according to the position of the leader and the difference in fitness with the leader, and then enter the next iteration until the stopping condition is reached.

[0055] It should be noted that the training samples are the total solar radiation intensity, net solar radiation intensity, hourly temperature and hourly relative humidity at each time point in the historical period obtained after preprocessing the historically collected photovoltaic power generation data and meteorological data; the actual power generation value at each time point in the historical period is used as the real data for comparison with the prediction results.

[0056] When calculating the fitness value, the sum of the root mean square error and the mean absolute error is calculated. The root mean square error focuses on larger errors and is more sensitive to outliers, while the mean absolute error treats all errors equally. The two are combined to reduce the excessive impact of a single outlier on the overall fitness and provide a more comprehensive error assessment.

[0057] Furthermore, in order to improve efficiency, the training samples are divided into multiple small batches in each iteration process, and the training samples of the small batches are passed into the multi-layer perceptron for forward propagation in turn, the sum of the root mean square error and the mean absolute error is calculated, and then the multiple small batches are averaged to obtain the fitness value.

[0058] In this embodiment, the smaller the fitness value, the higher the fitness. Therefore, the fruit fly with the smallest fitness value is the current optimal solution, that is, the leader. When updating the position of the non-leader fruit fly, it is necessary to encourage other fruit flies to move closer to the leader, but also to avoid converging to the local optimal solution too early and missing the global optimal solution. Therefore, according to the difference in fitness values ​​and the comparison results of the threshold, the position update is divided into two cases:

[0059] When the difference between the fitness values ​​of other fruit flies and the leader is greater than or equal to the threshold When , the position is updated according to the current position of the fruit fly and the current moving distance of the fruit fly; otherwise, the difference between the calculated fitness value and the leader is less than the threshold The weighted average position of all fruit flies is calculated, and the position is updated based on the weighted average position and the current movement distance of the fruit fly.

[0060] Specifically, the weighted average position is calculated according to the following formula:

[0061] ,

[0062] in, Indicates that the current Iterations, Indicates The weighted average position of the iteration The value of the item, represents the number of flies whose fitness value difference with the leader is less than a threshold plus the number of the leader; Indicates In the iteration The position of the fruit fly The value of the item; Indicates In the iteration The weights of the fruit flies are evenly distributed according to the actual situation, or they are ranked according to the difference in fitness values ​​between the fruit flies and the leader. The smaller the ranking, the greater the weight.

[0063] The current movement distance of the fruit fly is obtained by calculating the product of the position difference between the fruit fly and the leader and the update step size of the current iteration, where the update step size of the current iteration is the initial step size. It is calculated based on the current number of iterations according to the following formula, which decreases as the number of iterations increases:

[0064] ,

[0065] in, Indicates The update step size of the iteration, Indicates the maximum number of iterations.

[0066] Finally, the position of the non-leader fly is updated according to the following formula:

[0067] ,

[0068] in, Indicates The leader in the iteration The position of The value of the item, and Respectively represent The position of the fruit fly The value of the item before and after the update, ; Indicates The leader in the iteration The fitness value of Indicates In the iteration The fitness value of a fruit fly.

[0069] It should be noted that if the updated position of the fruit fly exceeds the preset search range, the position of the fruit fly is corrected by clipping to ensure that it is within the search range.

[0070] The updated position of the fruit fly is used in the next iteration. The stopping condition of the iteration is that the maximum number of iterations is reached or the fitness value no longer decreases. If the stopping condition is reached, step ③ is executed; otherwise, the number of iterations is increased by 1 and step ② is continued. If a fruit fly with a smaller fitness value is found in the new iteration process, then this fruit fly will become the new leader.

[0071] ③ Output the latest leader’s position to obtain the initial optimal parameters.

[0072] It should be noted that in order to further improve the performance of the multilayer perceptron, after the initial optimal parameters are obtained using the fruit fly optimization algorithm, the multilayer perceptron is inserted into the multilayer perceptron, and the multilayer perceptron is trained again using the training samples. The hyperparameters such as the learning rate and batch size are adjusted, and the parameters of the multilayer perceptron are updated again in the back propagation. After the training, the trained multilayer perceptron is obtained as the power generation prediction model of the substation.

[0073] It should be noted that if a power generation prediction model for a substation is obtained by training according to the above method, the power generation prediction models of other substations can be quickly obtained through transfer learning of model pre-training and fine-tuning.

[0074] Specifically, considering that the first few layers of the power generation prediction model are the extraction of common features, the parameters of the power generation prediction model trained for the non-current substation are frozen except for the last hidden layer and the output layer, and then re-trained according to the training samples of the current substation, the power generation prediction model of the current substation can be obtained.

[0075] During implementation, the data at each time point in the time period to be analyzed obtained in step S1 is transmitted as feature data to the power generation prediction model of the substation area to obtain the power generation prediction value at each time point.

[0076] S3. The predicted power generation value at each time point is compared with the actual power generation value at the same time point in the photovoltaic power generation data of the substation to construct data point pairs at each time point; and the abnormal point pairs are obtained from the data point pairs at each time point using a clustering algorithm.

[0077] It should be noted that constructing data point pairs at each time point is to establish a one-to-one correspondence between the predicted power generation value and the actual power generation value, and obtain continuous and full-time power samples of the photovoltaic power generation equipment in the area during the period to be analyzed.

[0078] The DBSCAN clustering algorithm is used to perform cluster analysis on the data point pairs at each time point to distinguish the normal point pairs from the abnormal point pairs. DBSCAN (Density-Based Spatial Clustering of Applications with Noise) is a density-based clustering algorithm that divides areas with sufficiently high density in space into the same category and regards those data that do not belong to any category as noise points. Each category is assigned a unique label, and the noise points are marked with a specific label.

[0079] When using the DBSCAN clustering algorithm, the k-distance scree plot is used to determine the search neighborhood radius, and the minimum number of points for a cluster is set according to 2 times the feature dimension. After clustering, the evaluation indicators are calculated based on the clustering results, such as the average distance within the cluster and the maximum distance within the cluster. If the preset accuracy is reached, the clustering operation is completed. Otherwise, the search neighborhood radius and / or the minimum number of points are adjusted and clustering is performed again until the preset accuracy is reached. In this embodiment, the search neighborhood radius is set to 0.2; the minimum number of points for a cluster is 4.

[0080] After completing the clustering operation, the clustering results are analyzed, and the data of each category is extracted using the assigned labels; the average absolute error between the predicted power generation value and the actual power generation value of each category is calculated, and the category with the smallest average absolute error is taken as the normal category, and the other categories are abnormal categories; the data point pairs at each time point in the abnormal category are abnormal point pairs.

[0081] Compared with the prior art, this embodiment uses power generation as an indicator for clustering, has a fast running speed and low data collection cost, has high accuracy and precision in anomaly detection, and enhances the ability to identify anomalies in photovoltaic power generation data.

[0082] S4. Obtain corresponding area business data according to the time point of the abnormal point pair, and obtain the rule category corresponding to the area business data according to the preset photovoltaic business abnormal rule as the abnormal cause of the corresponding abnormal point pair.

[0083] It should be noted that the substation business data includes: the contract capacity in the marketing system, the actual power generation, instantaneous active power value, current, theoretical power generation in the region and other data in the data collection system of the photovoltaic power station.

[0084] The photovoltaic business anomaly rule is a judgment condition set according to the three rule categories of meter reading anomaly, user behavior anomaly, and power generation equipment failure. By default, the corresponding area business data is obtained according to the time point of the abnormal point pair, which is the area business data within the time point of the abnormal point pair and the previous time point. For example, the time period to be analyzed is February 3, 2024, the interval between each time point is 15 minutes, and the time point of the abnormal point pair obtained is 8:30 on February 3, 2024, then the area business data between 8:15 and 8:30 on February 3, 2024 is obtained.

[0085] The rule category corresponding to the substation business data is obtained according to the preset photovoltaic business exception rules, that is, whether the substation business data matches the judgment condition of any photovoltaic business exception rule is identified. If it matches, the rule category corresponding to the matched photovoltaic business exception rule is used as the exception cause of the exception point pair.

[0086] Specifically, a meter reading anomaly indicates that the electric energy meter is temporarily removed or not working, and the judgment condition is that the current actual power generation is 0 but the actual power generation last month is not 0. When making a judgment, the actual power generation in the business data within the corresponding time period is obtained and identified based on the time point of the abnormal point pair, whether it is 0, and whether the actual power generation in the same time period last month is 0. If it matches the judgment condition, the abnormal cause of the abnormal point pair is a meter reading anomaly.

[0087] Abnormal user behavior is further subdivided into over-capacity, power generation exceeding contract capacity and night power generation. Among them, the judgment condition for over-capacity is that the actual power generation exceeds 30% of the theoretical power generation of the region; the judgment condition for power generation exceeding the contract capacity is that the instantaneous active value in the preset daytime period is greater than the contract capacity; the judgment condition for night power generation is that the current and actual power generation in the preset nighttime period are both greater than the minimum threshold; exemplarily, the preset daytime period is 9:00-15:00, and the preset nighttime period is 22:00-4:00; the minimum threshold is 0.1. When making a judgment, the actual power generation in the business data within the corresponding time period is obtained and identified based on the time point of the abnormal point pair, whether it exceeds 30% of the normal power generation in the area. If it exceeds, it matches the judgment condition of over-installed capacity, and the abnormal cause of the abnormal point pair is over-installed capacity; if it does not exceed, identify whether the corresponding time period meets the preset daytime period or the preset nighttime period, and continue to identify whether the instantaneous active value, contract capacity, or current and actual power generation in the business data match the judgment condition of the time period. If they match, the abnormal cause of the abnormal point pair is that the power generation exceeds the contract capacity, or it is night power generation.

[0088] Power generation equipment failure indicates that the power generation equipment or inverter is abnormal, and the judgment condition is that the actual power generation is 0 for three consecutive days. When making a judgment, the actual power generation in the business data in the corresponding time period is obtained and identified according to the time point of the abnormal point pair. If it is 0, the actual power generation of the previous two days is also 0. If both are 0, the abnormal cause of the abnormal point pair is a power generation equipment failure.

[0089] According to the above method, the substation business data corresponding to each abnormal point pair is matched and identified with each photovoltaic business abnormal rule in turn, and the rule category corresponding to the substation business data is obtained as the abnormal cause of the corresponding abnormal point pair.

[0090] Preferably, different levels of warning signals are issued according to the degree of deviation between the business data and the indicators in the judgment conditions; the date, time point, cluster category, exception rule, exception category and related data are recorded and output as an exception report to facilitate subsequent analysis and processing.

[0091] It should be noted that the abnormal categories obtained by clustering in step S3 are continuously attributed to rule categories by the abnormal business rules in step S4. Some abnormal categories have always corresponded to the same rule categories in multiple identifications and can be directly attributed, thereby reducing the number of judgments of business abnormality rules and improving the efficiency of judging the causes of abnormalities.

[0092] Preferably, a retrospective analysis method is used to identify and attribute historical abnormal data using the above steps, effectively discover potential problems, timely adjust and optimize algorithm parameters, and improve the reaction speed and processing capabilities of the method of this embodiment when facing new scenarios. The analysis and verification of a large amount of historical abnormal data further enriches and improves the abnormal rules of photovoltaic business, enhances adaptability and robustness, and improves the efficiency of abnormal identification and attribution.

[0093] Compared with the prior art, the photovoltaic data anomaly identification method provided in this embodiment extracts data features from photovoltaic data and meteorological data, and uses a combination of machine learning judgment and business rule recognition judgment to increase the comprehensiveness of recognition and has a wider applicability; first, the abnormal categories are screened out through clustering and then the abnormal categories are attributed using abnormal rules, thereby improving the efficiency of abnormal identification and attribution; the fruit fly initialization position, position update and fitness value in the fruit fly optimization algorithm are improved in combination with the application scenario, and the parameters of the multi-layer perceptron are optimized, which makes it easier to obtain the global optimal parameters while accelerating the convergence speed, thereby improving the prediction accuracy of the power generation; the predicted value and actual value of the power generation are used as two dimensions of the data point pair for cluster analysis, which has a fast running speed and low data collection cost, and the accuracy and precision of anomaly detection are high, thereby enhancing the ability to identify anomalies in photovoltaic power generation data.

[0094] Example 2

[0095] Another embodiment of the present invention discloses a system for identifying abnormal photovoltaic data in a substation, thereby implementing the method for identifying abnormal photovoltaic data in a substation in Example 1. The specific implementation of each module refers to the corresponding description in Example 1. Figure 2 As shown, the system includes:

[0096] The data preprocessing module 101 is used to collect and preprocess the photovoltaic power generation data and meteorological data of the substation area, and obtain the data at each time point in the time period to be analyzed;

[0097] The power generation prediction module 102 is used to input the data at each time point into the power generation prediction model of the substation area to obtain the power generation prediction value at each time point;

[0098] The abnormal data identification module 103 is used to construct data point pairs at each time point by comparing the predicted power generation value at each time point with the actual power generation value at the same time point in the photovoltaic power generation data of the substation area; and to obtain abnormal point pairs from the data point pairs at each time point by using a clustering algorithm;

[0099] The abnormal data attribution module 104 is used to obtain the corresponding area business data according to the time point of the abnormal point pair, and obtain the rule category corresponding to the area business data according to the preset photovoltaic business abnormal rule as the abnormal cause of the corresponding abnormal point pair.

[0100] Since the present embodiment and the aforementioned method for identifying abnormal photovoltaic data in a substation can be mutually referenced, the description here is repeated, so it will not be repeated here. Since the principle of the present system embodiment is the same as that of the aforementioned method embodiment, the present system embodiment also has the corresponding technical effects of the aforementioned method embodiment.

[0101] Those skilled in the art will appreciate that all or part of the processes of the above-mentioned embodiments can be implemented by instructing related hardware through a computer program, and the program can be stored in a computer-readable storage medium, wherein the computer-readable storage medium is a disk, an optical disk, a read-only storage memory, or a random access memory, etc.

[0102] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A method for identifying abnormal photovoltaic data in a substation, characterized in that: The following steps are involved: Collect and pre-process the photovoltaic power generation data and meteorological data of the substation area to obtain the data at each time point in the period to be analyzed; The data at each time point is transmitted to the power generation prediction model of the substation to obtain the power generation prediction value at each time point; The predicted power generation value at each time point is combined with the actual power generation value at the same time point in the photovoltaic power generation data of the station area to construct a data point pair at each time point; Obtaining abnormal point pairs from the data point pairs at each time point by using a clustering algorithm, including: clustering the data point pairs at each time point by using a DBSCAN algorithm to obtain multiple categories, calculating the average absolute error between the predicted power generation value and the actual power generation value of each category, taking the category with the smallest average absolute error as the normal category, and the other categories as abnormal categories; the data point pairs at each time point in the abnormal category are abnormal point pairs; Acquire corresponding area business data according to the time point of the abnormal point pair, and acquire the rule category corresponding to the area business data according to the preset photovoltaic business abnormal rule as the abnormal cause of the corresponding abnormal point pair; The power generation prediction model of the substation area is obtained by optimizing the parameters of the multi-layer perceptron using the improved fruit fly optimization algorithm to obtain the initial optimal parameters and then training the multi-layer perceptron again; the position of the fruit fly in the improved fruit fly optimization algorithm is a row vector composed of the weight matrix and the bias vector in the multi-layer perceptron, and the weight values ​​belonging to the same weight matrix are calculated according to the number of neurons in the previous layer and the number of neurons in the next layer connected by the weight matrix, and the initial value is generated using uniform distribution after the value range is calculated according to the following formula: ， in, and denote the lower and upper limits of the uniform distribution, respectively. and Indicates the number of neurons in the previous layer and the number of neurons in the next layer connected by the weight matrix; the initial value of each bias is a random number that follows the uniform distribution U[0,1); The fitness value of the improved fruit fly optimization algorithm is obtained by taking the position of the fruit fly as the parameter of the multi-layer perceptron to predict the training sample, and calculating the sum of the root mean square error and the mean absolute error according to the prediction result; in each iteration of the improved fruit fly optimization algorithm, the fruit fly with the smallest fitness value is taken as the leader, and when the difference between the fitness value of other fruit flies and the leader is greater than or equal to a threshold, the position is updated according to the current position of the fruit fly and the current moving distance of the fruit fly; otherwise, the weighted average position of all fruit flies whose fitness value difference with the leader is less than the threshold is calculated, and the position is updated based on the weighted average position and the current moving distance of the fruit fly; the current moving distance of the fruit fly is obtained by calculating the product of the position difference between the fruit fly and the leader and the update step size of the current iteration.

2. The method for identifying abnormal photovoltaic data in a substation according to claim 1, characterized in that: The pre-processing of photovoltaic power generation data and meteorological data in the substation area to obtain data at each time point in the time period to be analyzed includes: During the period to be analyzed, the actual power generation value in the photovoltaic power generation data of the substation is collected at each time point as the reference time point, and the data of total solar radiation intensity, net solar radiation intensity, hourly temperature and hourly relative humidity at each reference time point are supplemented by linear interpolation method; The actual power generation value at each benchmark time point is taken as the real data, and the total solar radiation intensity, net solar radiation intensity, hourly temperature and hourly relative humidity at each benchmark time point are taken as characteristic data to obtain the data at each time point in the period to be analyzed.

3. The method for identifying abnormal photovoltaic data in a substation according to claim 1, characterized in that: The photovoltaic business abnormality rule is a judgment condition set according to three rule categories: meter reading abnormality, user behavior abnormality and power generation equipment failure.

4. The method for identifying abnormal photovoltaic data in a substation according to claim 3, characterized in that: The judgment condition for the meter reading abnormality is that the current actual power generation is 0 but the power generation last month is not 0; the user behavior abnormality is subdivided into over-capacity, power generation exceeding the contract capacity and night power generation, among which the judgment condition for over-capacity is that the actual power generation exceeds 30% of the regional theoretical value; the judgment condition for the actual power generation exceeding the contract capacity is that the instantaneous active value in the preset daytime period is greater than the contract capacity; the judgment condition for night power generation is that the current and actual power generation in the preset nighttime period are both greater than the minimum threshold; the judgment condition for the power generation equipment failure is that the actual power generation is 0 for 3 consecutive days.

5. The method for identifying abnormal photovoltaic data in a substation according to claim 2, characterized in that: The power generation prediction model for the substation area is obtained by freezing the parameters of the power generation prediction model trained in the non-current substation area except the last hidden layer and the output layer, and then retraining according to the training samples of the current substation area.

6. A photovoltaic data anomaly identification system in a substation, characterized in that: include: The data preprocessing module is used to collect and preprocess the photovoltaic power generation data and meteorological data of the substation area, and obtain the data at each time point in the period to be analyzed; A power generation prediction module is used to transfer the data at each time point into the power generation prediction model of the substation to obtain the power generation prediction value at each time point; An abnormal data identification module is used to construct data point pairs at each time point by comparing the predicted power generation value at each time point with the actual power generation value at the same time point in the photovoltaic power generation data of the substation area; Obtaining abnormal point pairs from the data point pairs at each time point by using a clustering algorithm, including: clustering the data point pairs at each time point by using a DBSCAN algorithm to obtain multiple categories, calculating the average absolute error between the predicted power generation value and the actual power generation value of each category, taking the category with the smallest average absolute error as the normal category, and the other categories as abnormal categories; the data point pairs at each time point in the abnormal category are abnormal point pairs; An abnormal data attribution module is used to obtain the corresponding area business data according to the time point of the abnormal point pair, and obtain the rule category corresponding to the area business data according to the preset photovoltaic business abnormal rule as the abnormal cause of the corresponding abnormal point pair; The power generation prediction model of the substation in the power generation prediction module is obtained by optimizing the parameters of the multi-layer perceptron using the improved fruit fly optimization algorithm to obtain the initial optimal parameters and then training the multi-layer perceptron again; the position of the fruit fly in the improved fruit fly optimization algorithm is a row vector composed of a weight matrix and a bias vector in the multi-layer perceptron, and the weight values ​​belonging to the same weight matrix are calculated according to the number of neurons in the previous layer and the number of neurons in the next layer connected by the weight matrix, and the initial value is generated using uniform distribution after the value range is calculated according to the following formula: ， in, and denote the lower and upper limits of the uniform distribution, respectively. and Indicates the number of neurons in the previous layer and the number of neurons in the next layer connected by the weight matrix; the initial value of each bias is a random number that follows the uniform distribution U[0,1); The fitness value of the improved fruit fly optimization algorithm is obtained by taking the position of the fruit fly as the parameter of the multi-layer perceptron to predict the training sample, and calculating the sum of the root mean square error and the mean absolute error according to the prediction result; in each iteration of the improved fruit fly optimization algorithm, the fruit fly with the smallest fitness value is taken as the leader, and when the difference between the fitness value of other fruit flies and the leader is greater than or equal to a threshold, the position is updated according to the current position of the fruit fly and the current moving distance of the fruit fly; otherwise, the weighted average position of all fruit flies whose fitness value difference with the leader is less than the threshold is calculated, and the position is updated based on the weighted average position and the current moving distance of the fruit fly; the current moving distance of the fruit fly is obtained by calculating the product of the position difference between the fruit fly and the leader and the update step size of the current iteration.

Images