Intelligent operation and maintenance method and system for unattended substation
Through the intelligent operation and maintenance method of unmanned substations, the timing feature extraction network and health scoring mechanism are used to solve the problem of lack of early warning and equipment health assessment in traditional operation and maintenance methods, and timely detection and preventive maintenance of equipment abnormal trends. Improve the operating stability and safety of the substation.
Patent Information
- Application Number
- CN202411050506.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-11-29
AI Technical Summary
Traditional substation operation and maintenance methods lack an early warning mechanism and the inability to comprehensively evaluate the health status of the equipment, resulting in the inability to detect abnormalities in time when the equipment parameters deviate from the normal range, affecting operational stability and safety.
Using the intelligent operation and maintenance method of unattended substations, the equipment timing data is collected for preprocessing, the timing feature extraction network is constructed, the timing feature vector is extracted, trend judgment and health score is performed, and multiple times are combined with the equipment's service life and the number of abnormal trend alarms. Level threshold judgment and preventive maintenance.
It realizes timely discovery and early warning of abnormal trends in equipment parameters, comprehensively evaluates the health status of the equipment, improves the operating stability and safety of the substation, and reduces operation and maintenance costs.
Smart Images

Figure CN119048051A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of substation management, and in particular to an intelligent operation and maintenance method and system for an unmanned substation. Background Art
[0002] As a new operation and maintenance model for substations, smart operation and maintenance aims to improve the operating efficiency, safety and reliability of substations, while reducing operation and maintenance costs, and realizing the digital transformation and intelligent upgrading of substation operation and maintenance.
[0003] Traditional substation operation and maintenance methods mainly rely on manual inspections and regular maintenance, or when equipment parameters exceed preset thresholds, alarms are triggered to notify operation and maintenance personnel to handle them. However, the problems with traditional methods are mainly reflected in two aspects: First, there is a lack of early warning mechanism. During the operation of substation equipment, its temperature, current and other parameters may gradually deviate from the normal range, but have not yet reached the traditional alarm threshold. Traditional methods mainly rely on fixed alarm thresholds and lack real-time monitoring and data analysis capabilities for equipment status changes, and cannot provide early warnings; second, traditional methods mainly rely on manual inspections and experience judgments, making it difficult to comprehensively evaluate the health status of equipment in substations and unable to provide a scientific basis for preventive maintenance. Summary of the invention
[0004] In view of this, the present invention aims to provide an unmanned substation intelligent operation and maintenance method and system, aiming to solve the problem that the traditional substation operation and maintenance method cannot provide early warning and is difficult to comprehensively evaluate the health status of the equipment in the substation.
[0005] An unmanned substation intelligent operation and maintenance method comprises the following steps:
[0006] S1: Collect device timing data and preprocess it to obtain preprocessed device timing data;
[0007] S2: Based on the preprocessed device timing data, a timing feature extraction network is constructed to extract the timing feature vector;
[0008] S3: According to the time series feature vector, the trend of the equipment operation status is judged, and an alarm is issued if there is an abnormal trend;
[0009] S4: Output the equipment health score based on the time series feature vector, combined with the equipment’s service life and the number of abnormal trend alarms;
[0010] S5: Perform multi-level threshold judgment based on the equipment health score to obtain the health level, and perform preventive maintenance on the equipment corresponding to the health level;
[0011] Furthermore, the step S1 collects and preprocesses the device timing data to obtain the preprocessed device timing data, including:
[0012] S11: collecting original time series data of the equipment in real time through the sensor network; the equipment includes: transformer, current transformer and voltage transformer; the original time series data is continuous numerical data at each moment, including: transformer oil temperature, transformer oil level, transformer winding temperature, current magnitude, voltage magnitude; the sensor network includes: transformer temperature sensor, transformer oil level sensor, current magnitude sensor and voltage magnitude sensor;
[0013] S12: using an anomaly detection algorithm based on a sliding window to remove outliers from the collected original time series data, and obtaining data after removing outliers;
[0014] S13: annotating missing values of the data after outliers are removed to obtain a missing value annotated sequence, and concatenating the missing value annotated sequence with the data after outliers are removed element by element to obtain a concatenated sequence;
[0015] S14: The spliced sequence is standardized using a Z-Score method to obtain preprocessed device time series data; the preprocessed device time series data includes two parts: a data point sequence and a missing value annotation sequence;
[0016] Furthermore, the step S2 constructs a time series feature extraction network based on the preprocessed device time series data to extract a time series feature vector, including:
[0017] S21: annotate the missing value sequence in the preprocessed device time series data, update the sequence, mark the distance between each data point in the data point sequence and the next nearest missing value, the maximum distance is 10, if the distance is greater than 10, mark the position as -1, and obtain the updated sequence;
[0018] S22: Input the updated sequence into the time series feature extraction network to obtain a time series feature vector, wherein the calculation method of the time series feature extraction network is:
[0019] h t =GRU(concat(X t ,A t ))
[0020] Among them, h t is a time series feature vector, GRU() is a gated recurrent unit operation, concat(x t ,a t ) is the updated sequence, concat() is the concatenation operation, X t is the data point sequence, A tLabel the updated missing value sequence, where t is the time index;
[0021] The present invention proposes a special processing method for missing values in time series data. By calculating the distance between each data point and the next nearest missing value and marking it when the distance is less than or equal to 10, an update sequence containing missing value distance information is generated. This operation retains the position characteristics of the missing values and provides richer input information for step S22, so that the constructed time series feature extraction network can more accurately understand the missing pattern in the data and improve the accuracy of capturing device trends. If the distance between the missing value and a data point is greater than 10, it is considered that the data point is too far away from the missing value, the data correlation is low, and it has no reference value. Therefore, when the distance is greater than or equal to 10, the data point is marked as -1.
[0022] Furthermore, the step S3 determines the trend of the equipment operation status according to the time series feature vector, and issues an alarm if an abnormal trend exists, including:
[0023] S31: Calculate the preliminary prediction value of the device according to the time series feature vector, and the calculation method is:
[0024]
[0025] in, They are the preliminary predicted values of the data points from time t+2 to time t+5;
[0026] S32: Construct a moment importance mechanism, calculate the weight of each moment and update the preliminary prediction value, and obtain the updated preliminary prediction value from t+2 to t+5. The calculation method is:
[0027]
[0028]
[0029]
[0030] Among them, α t+2 ,α t+3 ,α t+4 ,α t+5 are the weights from t+2 to t+5, u t+i is the fusion vector at time t+i, i is the counting index, tanh() is the hyperbolic tangent function, W u is the moment weight parameter matrix, is the initial predicted value of the data point at time t+i, b u is the moment weight bias vector, α t+i is the weight at time t+i, exp() is the exponential operation to remove the natural constant, vα is the moment weight parameter vector, u t+j is the fusion vector at time t+j, j is the counting index; It is the initial forecast value after updating at time t+2;
[0031] Calculated Then, according to the above method Continue to calculate and get the final predicted value of the data points from t+3 to t+5
[0032] When performing time series prediction, traditional methods often assume that all time steps contribute equally to future predictions, or use a fixed weight distribution method; however, in practical applications, different time steps often have different effects on future predictions; the present invention can more accurately capture the information of key time steps by dynamically calculating weights in step S32, thereby improving the accuracy and robustness of predictions;
[0033] S33: Calculate the final prediction value of the device according to the time series feature vector and the updated preliminary prediction value. The calculation method is:
[0034]
[0035] is the final predicted value of the data point at time t+1, and BiLSTM is a bidirectional long short-term memory network operation;
[0036] Calculated Then, continue to calculate in the above way to get the final predicted value of the data points from t+2 to t+5
[0037] Even though the traditional neural network method can utilize bidirectional context information in the training phase, it can only utilize unidirectional context information for prediction in the prediction phase. However, the present invention first uses a gated recurrent unit to preliminarily predict the data values at several future time points based on the time series feature vector, and predicts the context information first. Then, a bidirectional long short-term memory network is used to combine the preliminary prediction value and the time series feature vector, and the bidirectional context information is comprehensively utilized to optimize the prediction of the data value at the next time point, providing more accurate input for subsequent trend judgment. By more accurately predicting the future state of the equipment and performing trend judgment, potential anomalies can be discovered more timely, which helps to reduce the power outage time caused by equipment failure and improve the operation stability and reliability of the substation.
[0038] S34: Extract the future feature vector according to the final prediction value of the device, and the calculation method is:
[0039]
[0040] Among them, h′t is the future feature vector, CNN() is a convolutional neural network operation with a convolution kernel width of 1;
[0041] S35: Combine the future feature vector and the time series feature vector to determine the trend of the equipment operation status. The calculation method is:
[0042] p = softmax(MLP(con.cat(h t ,h′ t )))
[0043] Where p is the probability of device trend category, softmax() is the softmax function, and MLP() is the fully connected neural network operation;
[0044] When p>50%, the device is judged as "having an abnormal trend";
[0045] When p≤50%, the device is judged as "normal trend";
[0046] S36: If the device has an abnormal trend, an abnormal trend alarm is issued.
[0047] Furthermore, the step S4 outputs a device health score based on the time series feature vector, combined with the device's service life and the number of abnormal trend alarms, including:
[0048] S41: Count the service life of the equipment and the number of abnormal trend alarms for each equipment;
[0049] S42: Input the time series feature vector, the equipment service life and the number of abnormal trend alarms into the equipment health score network, and output the equipment health score. The calculation method of the equipment health score network is:
[0050] score = sigmoid(MLP(h t ,T,n))
[0051] Where score is the equipment health score, sigmoid() is the sigmoid function, T is the equipment usage period, and n is the number of abnormal trend alarms.
[0052] In step S42, a health score network is constructed by using a multi-layer perceptron to integrate the recent historical data of the equipment, the service life and the number of abnormal trend alarms, and then the output of the network is mapped to the interval (0,1) through the sigmoid function to more intuitively represent the health score of the equipment; and the equipment health score is used as a quantitative scientific basis for the multi-level threshold judgment in step S5; compared with the traditional method of manually evaluating the health of equipment, the present invention quantitatively represents the health of the equipment, which provides a basis for the preventive maintenance of substation equipment;
[0053] Furthermore, the step S5 performs multi-level threshold judgment according to the equipment health score to obtain a health level, and performs preventive maintenance corresponding to the health level on the equipment, including:
[0054] S51: Set multi-level thresholds to divide the equipment health score into 4 health levels;
[0055] S52: Compare the health score of the device with the set multi-level thresholds to determine the health level of the device;
[0056] S53: Execute preventive maintenance measures corresponding to the health level of the equipment according to the health level;
[0057] Compared with the prior art, the advantages of the present invention are:
[0058] (1) The present invention constructs a time series feature extraction network to capture the potential abnormal trends of the equipment through time series feature vectors, aiming to solve the risks brought about by the equipment parameters gradually deviating from the normal range but not yet reaching the traditional alarm threshold; at the same time, the present invention combines the equipment's service life and the number of abnormal trend alarms to output the equipment health score, thereby comprehensively evaluating the health status of the equipment.
[0059] (2) The present invention proposes a special processing method for missing values in time series data. By calculating the distance between each data point and the next nearest missing value and marking it when the distance is less than or equal to 10, an update sequence containing missing value distance information is generated, which retains the position characteristics of the missing values and enhances the data representation capability.
[0060] (3) The present invention proposes an innovative strategy for judging data trends. First, a gated recurrent unit is used to make a preliminary prediction of the future data of the device to capture the trend of data changes over time. Then, the preliminary predicted future data is combined with the historical data, and the bidirectional context information of the data is fused through a bidirectional long short-term memory network. Compared with traditional neural networks that can only use unidirectional information, the present invention improves the accuracy of trend judgment and provides strong support for timely detection of potential abnormalities of the device. BRIEF DESCRIPTION OF THE DRAWINGS
[0061] Figure 1 A schematic diagram of the process of unattended substation intelligent operation and maintenance provided by the present invention;
[0062] Figure 2 A schematic diagram of the algorithm flow for extracting time series feature vectors provided by the present invention;
[0063] Figure 3 A schematic diagram of the algorithm flow for calculating the final prediction value of a device in combination with context information provided by the present invention. DETAILED DESCRIPTION
[0064] The present invention is further described below in conjunction with the accompanying drawings, but the present invention is not limited in any way. Any changes or substitutions made based on the teachings of the present invention belong to the protection scope of the present invention.
[0065] Embodiment 1: A smart operation and maintenance method for an unattended substation, such as Figure 1 As shown, the following steps are included:
[0066] S1: Collect device timing data and preprocess it to obtain preprocessed device timing data;
[0067] S11: collecting original time series data of the equipment in real time through the sensor network; the equipment includes: transformer, current transformer and voltage transformer; the original time series data is continuous numerical data at each moment, including: transformer oil temperature, transformer oil level, transformer winding temperature, current magnitude, voltage magnitude; the sensor network includes: transformer temperature sensor, transformer oil level sensor, current magnitude sensor and voltage magnitude sensor;
[0068] S12: using an anomaly detection algorithm based on a sliding window to remove outliers from the collected original time series data, and obtaining data after removing outliers;
[0069] S13: annotating missing values of the data after outliers are removed to obtain a missing value annotated sequence, and concatenating the missing value annotated sequence with the data after outliers are removed element by element to obtain a concatenated sequence;
[0070] For example, the original time series voltage data (part) of a transformer is collected in real time through the sensor network. These data contain voltage measurement values at different time points. However, due to sensor failure, communication interruption and other reasons, the data sequence may contain some missing values, which are represented by "NaN":
[0071] [11000,11050,NaN,10980,11100,10960,10950,11020,NaN,10890]
[0072] Step S13 performs missing value annotation on the transformer data point sequence; the annotation rule is: if the data point is not a missing value, the corresponding position in the annotation sequence is marked as 0; if the data point is a missing value, the corresponding position in the annotation sequence is marked as 1; therefore, for the above transformer voltage data sequence, its missing value annotation sequence is:
[0073] [0,0,1,0,0,0,0,0,1,0]
[0074] The length of the annotation sequence is equal to that of the original voltage data sequence;
[0075] S14: The spliced sequence is standardized using a Z-Score method to obtain preprocessed device time series data; the preprocessed device time series data includes two parts: a data point sequence and a missing value annotation sequence;
[0076] S2: Based on the preprocessed device timing data, a timing feature extraction network is constructed to extract the timing feature vector;
[0077] S21: annotate the missing value sequence in the preprocessed device time series data, update the sequence, mark the distance between each data point in the data point sequence and the next nearest missing value, the maximum distance is 10, if the distance is greater than 10, mark the position as -1, and obtain the updated sequence;
[0078] For example, for step S13,
[0079] Voltage data sequence: [11000,11050,NaN,10980,11100,10960,10950,11020,NaN,10890]
[0080] Missing value labeling sequence: [0,0,1,0,0,0,0,0,1,0]
[0081] Next, the missing value annotation sequence is updated to mark the distance between each non-missing data point and the next nearest missing value. According to the rules, starting from the beginning of the sequence, for each non-missing data point (i.e., the position marked as 0), calculate its distance to the next position marked as 1, and mark this distance at the corresponding position. If the distance is greater than 10, it is marked as -1 (indicating that the distance is too far and no special marking is performed). The updated missing value annotation sequence is:
[0082] [2,1,0,5,4,3,2,1,0,-1]
[0083] S22: Input the updated sequence into the time series feature extraction network to obtain a time series feature vector, wherein the calculation method of the time series feature extraction network is:
[0084] h t =GRU(concat(X t ,A t ))
[0085] Among them, h t is a time series feature vector, GRU() is a gated recurrent unit operation, concat(x t ,a t ) is the updated sequence, concat() is the concatenation operation, X t is the data point sequence, At Label the updated missing value sequence, where t is the time index;
[0086] The present invention proposes a special processing method for missing values in time series data. By calculating the distance between each data point and the next nearest missing value and marking it when the distance is less than or equal to 10, an update sequence containing missing value distance information is generated. This operation retains the position characteristics of the missing values and provides richer input information for step S22, so that the constructed time series feature extraction network can more accurately understand the missing pattern in the data and improve the accuracy of capturing device trends. If the distance between the missing value and a data point is greater than 10, it is considered that the data point is too far away from the missing value, the data correlation is low, and it has no reference value. Therefore, when the distance is greater than or equal to 10, the data point is marked as -1.
[0087] S3: According to the time series feature vector, the trend of the equipment operation status is judged, and an alarm is issued if there is an abnormal trend;
[0088] S31: Calculate the preliminary prediction value of the device according to the time series feature vector, and the calculation method is:
[0089]
[0090] in, They are the preliminary predicted values of the data points from time t+2 to time t+5;
[0091] S32: Construct a moment importance mechanism, calculate the weight of each moment and update the preliminary prediction value, and obtain the updated preliminary prediction value from t+2 to t+5. The calculation method is:
[0092]
[0093]
[0094]
[0095] Among them, α t+2 ,α t+3 ,α t+4 ,α t+5 are the weights from t+2 to t+5, u t+i is the fusion vector at time t+i, i is the counting index, tanh() is the hyperbolic tangent function, W u is the moment weight parameter matrix, is the initial predicted value of the data point at time t+i, b u is the moment weight bias vector, α t+i is the weight at time t+i, exp() is the exponential operation to remove the natural constant, vα is the moment weight parameter vector, u t+j is the fusion vector at time t+j, j is the counting index; It is the initial forecast value after updating at time t+2;
[0096] Calculated Then, according to the above method Continue to calculate and get the final predicted value of the data points from t+3 to t+5
[0097] When performing time series prediction, traditional methods often assume that all time steps contribute equally to future predictions, or use a fixed weight distribution method; however, in practical applications, different time steps often have different effects on future predictions; the present invention can more accurately capture the information of key time steps by dynamically calculating weights in step S32, thereby improving the accuracy and robustness of predictions;
[0098] S33: Calculate the final prediction value of the device according to the time series feature vector and the updated preliminary prediction value. The calculation method is:
[0099]
[0100] is the final predicted value of the data point at time t+1, and BiLSTM is a bidirectional long short-term memory network operation;
[0101] Calculated Then, continue to calculate in the above way to get the final predicted value of the data points from t+2 to t+5
[0102] Even though the traditional neural network method can utilize bidirectional context information in the training phase, it can only utilize unidirectional context information for prediction in the prediction phase; however, the present invention uses a gated recurrent unit in step S31 to preliminarily predict the data values at several future time points based on the time series feature vector, and first predicts the context information; then, a bidirectional long short-term memory network is used, combined with the preliminary prediction value and the time series feature vector, and the bidirectional context information is comprehensively utilized to optimize the prediction of the data value at the next time point, providing a more accurate input for subsequent trend judgment; by more accurately predicting the future state of the equipment and performing trend judgment, potential anomalies can be discovered more promptly, which helps to reduce the power outage time caused by equipment failure and improve the operational stability and reliability of the substation;
[0103] S34: Extract the future feature vector according to the final prediction value of the device, and the calculation method is:
[0104]
[0105] Among them, h′ t is the future feature vector, CNN() is the convolutional neural network operation with a convolution kernel width of 1;
[0106] S35: Combine the future feature vector and the time series feature vector to determine the trend of the equipment operation status. The calculation method is:
[0107] p = softmax(MLP(concat(h t ,h′ t )))
[0108] Where p is the probability of device trend category, softmax() is the softmax function, and MLP() is the fully connected neural network operation;
[0109] When p>50%, the device is judged as "having an abnormal trend";
[0110] When p≤50%, the device is judged as "normal trend";
[0111] S36: If the device has an abnormal trend, an abnormal trend alarm is issued.
[0112] For example, the temperature of the transformer is being monitored. The fixed alarm threshold of the traditional monitoring method is set to 85°C; while the real-time temperature monitoring data monitored are: 80°C, 81°C, 82°C, 83°C, 84°C;
[0113] In this example, although the temperature of the device continues to rise, it has not exceeded the set threshold of 85°C, so the traditional method will not trigger an alarm; however, the device trend category probability p calculated based on this temperature sequence is 60%, which means that the device is judged to have an "abnormal trend", which can detect the abnormal trend of the device earlier and issue an abnormal trend alarm, thereby improving the safety of the device;
[0114] In addition, the calculation method proposed in step S3 of the present invention helps to significantly reduce false alarms. For example, under the influence of certain external conditions, the equipment data may fluctuate greatly in a short period of time and then return to normal. At this time, the large fluctuation of the data does not mean that the equipment operation status is abnormal. Traditional methods often issue alarms based on fixed thresholds, which are prone to false alarms due to data fluctuations. In contrast, the present invention uses a neural network model to make a more comprehensive trend judgment on the equipment operation status by integrating time series feature vectors and future feature vectors, and more accurately identifies the real abnormal trend. For short-term data fluctuations caused by external conditions, since no continuous abnormal trend is formed, the system will not trigger an alarm, thereby effectively reducing the false alarm rate.
[0115] For example, due to some reasons, the transformer workload increases temporarily, and the temperature rises to 83°C, 84°C, 85°C briefly, and then quickly drops back to 84°C, 83°C;
[0116] This series of data may trigger an alarm because it exceeds the threshold of 85°C in traditional methods; however, under the calculation mechanism of the present invention, the temperature values at several future time points are first predicted based on the time series feature vector, and then the prediction is optimized using a bidirectional long short-term memory network to obtain a more accurate prediction of the future state. The calculated device trend category probability p is designed to be 45%, which means that the device is judged to have a "normal trend"; although the data fluctuates greatly in a short period of time and once exceeded the alarm threshold of the traditional method, since no continuous abnormal upward trend has been formed, the method of the present invention will not trigger an alarm, thereby avoiding the occurrence of false alarms.
[0117] S4: Output the equipment health score based on the time series feature vector, combined with the equipment’s service life and the number of abnormal trend alarms;
[0118] S41: Count the service life of the equipment and the number of abnormal trend alarms for each equipment;
[0119] S42: Input the time series feature vector, the equipment service life and the number of abnormal trend alarms into the equipment health score network, and output the equipment health score. The calculation method of the equipment health score network is:
[0120] score = sigmoid(MLP(h t ,T,n))
[0121] Where score is the equipment health score, sigmoid() is the sigmoid function, T is the equipment usage period, and n is the number of abnormal trend alarms.
[0122] In step S42, a health score network is constructed by using a multi-layer perceptron to integrate the recent historical data of the equipment, the service life and the number of abnormal trend alarms, and then the output of the network is mapped to the interval (0,1) through the sigmoid function to more intuitively represent the health score of the equipment; and the equipment health score is used as a quantitative scientific basis for the multi-level threshold judgment in step S5; compared with the traditional method of manually evaluating the health of equipment, the present invention quantitatively represents the health of the equipment, which provides a basis for the preventive maintenance of substation equipment;
[0123] For example, the system will calculate a device health score for each device based on the different years of use, the number of abnormal trend alarms, and the time series feature vector of the device, as follows:
[0124] Equipment 1: Transformer A
[0125] Equipment use period: 10 years
[0126] Abnormal trend alarm times: 3 times
[0127] Device health score: 0.85
[0128] Equipment 2: Transformer B
[0129] Equipment use period: 15 years
[0130] Abnormal trend alarm times: 5 times
[0131] Device health score: 0.70
[0132] Equipment 3: Transformer C
[0133] Equipment use period: 8 years
[0134] Abnormal trend alarm times: 1 time
[0135] Device health score: 0.90
[0136] S5: Perform multi-level threshold judgment based on the equipment health score to obtain the health level, and perform preventive maintenance on the equipment corresponding to the health level;
[0137] S51: Set multi-level thresholds to divide the equipment health score into 4 health levels;
[0138] For example, set the following multi-level thresholds to classify the health level of the device:
[0139] Health level A (excellent): Equipment health score ≥ 0.90
[0140] Health level B (good): 0.80 ≤ Equipment health score < 0.90
[0141] Health level C (general): 0.60 ≤ Equipment health score < 0.80
[0142] Health level D (poor): Equipment health score < 0.60
[0143] S52: Compare the health score of the device with the set multi-level thresholds to determine the health level of the device;
[0144] For example, based on the set multi-level thresholds, we can determine the health level of the following devices:
[0145] Equipment 1 (transformer A): the equipment health score is 0.85, which belongs to the health level B (good);
[0146] Equipment 2 (transformer B): the equipment health score is 0.70, which belongs to the health level C (average);
[0147] Equipment 3 (transformer C): the equipment health score is 0.90, which belongs to the health level A (excellent);
[0148] S53: Execute preventive maintenance measures corresponding to the health level of the equipment according to the health level;
[0149] For example:
[0150] When the health level is A (excellent), a comprehensive inspection is carried out every six months, including the appearance of the transformer, the tightness of the connectors, the surface condition of the insulators, the oil level and oil color, etc. At the same time, lubrication maintenance is carried out to lubricate the moving parts such as the transformer's on-load tap changer;
[0151] When the health level is B (good), a comprehensive inspection is carried out every quarter to check for abnormal sounds, vibrations or overheating; at the same time, some worn or aged seals and insulating parts are replaced;
[0152] When the health level is C (normal), a comprehensive inspection is carried out twice a quarter, paying attention to the changes in the oil temperature and oil level of the transformer; at the same time, the weak links of the transformer, such as windings and iron cores, are repaired and reinforced.
[0153] When the health level is D (poor), consider overhauling the transformer or returning it to the factory for repair;
[0154] The present invention is particularly suitable for unattended substations with various types of equipment. The present invention can increase the service life of substation equipment, reduce the failure rate, achieve the maximum maintenance effect under limited resources, and provide data support for preventive maintenance.
[0155] Embodiment 2: The present invention also discloses an unattended substation intelligent operation and maintenance system, comprising:
[0156] Substation data acquisition and preprocessing module: collects equipment time series data and preprocesses it to obtain preprocessed equipment time series data;
[0157] Timing feature extraction module: Based on the preprocessed device timing data, a timing feature extraction network is constructed to extract the timing feature vector;
[0158] Trend judgment module: judge the trend of the equipment according to the time series feature vector, and issue an alarm if there is an abnormal trend;
[0159] Equipment health scoring module: outputs equipment health scores based on the time series feature vector, combined with the equipment’s service life and the number of abnormal trend alarms;
[0160] Health level calculation and preventive maintenance module: Perform multi-level threshold judgment based on the equipment health score to obtain the health level, and perform preventive maintenance on the equipment corresponding to the health level.
[0161] It should be noted that the serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages and disadvantages of the embodiments. And the terms "including", "comprising" or any other variants thereof in this article are intended to cover non-exclusive inclusion, so that a process, device, article or method including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, device, article or method. In the absence of further restrictions, an element defined by the sentence "including a ..." does not exclude the presence of other identical elements in the process, device, article or method including the element.
[0162] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned embodiment methods can be implemented by means of software plus a necessary general hardware platform, and of course by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes a number of instructions for a terminal device (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.
[0163] The above are only preferred embodiments of the present invention, and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.
Claims
1. An unmanned substation intelligent operation and maintenance method, characterized in that: The following steps are involved: S1: collect device timing data and preprocess it to obtain preprocessed device timing data; S2: Based on the preprocessed device timing data, a timing feature extraction network is constructed to extract the timing feature vector; S3: According to the time series feature vector, the trend of the equipment operation status is judged, and an alarm is issued if there is an abnormal trend; S4: Output the equipment health score based on the time series feature vector, combined with the equipment’s service life and the number of abnormal trend alarms; S5: Perform multi-level threshold judgment based on the equipment health score to obtain the health level, and perform preventive maintenance on the equipment corresponding to the health level.
2. The unmanned substation intelligent operation and maintenance method according to claim 1 is characterized in that: The S1 step includes: S11: collecting original time series data of the equipment in real time through the sensor network; the equipment includes: transformer, current transformer and voltage transformer; the original time series data is continuous numerical data at each moment, including: transformer oil temperature, transformer oil level, transformer winding temperature, current magnitude, voltage magnitude; the sensor network includes: transformer temperature sensor, transformer oil level sensor, current magnitude sensor and voltage magnitude sensor; S12: using an anomaly detection algorithm based on a sliding window to remove outliers from the collected original time series data, and obtaining data after removing outliers; S13: annotating missing values of the data after outliers are removed to obtain a missing value annotated sequence, and concatenating the missing value annotated sequence with the data after outliers are removed element by element to obtain a concatenated sequence; S14: The spliced sequence is standardized using a Z-Score method to obtain preprocessed device time series data; the preprocessed device time series data includes two parts: a data point sequence and a missing value annotation sequence.
3. The unmanned substation intelligent operation and maintenance method according to claim 2 is characterized in that: The S2 step includes: S21: annotate the missing value sequence in the preprocessed device time series data, update the sequence, mark the distance between each data point in the data point sequence and the next nearest missing value, the maximum distance is 10, if the distance is greater than 10, mark the position as -1, and obtain the updated sequence; S22: Input the updated sequence into the time series feature extraction network to obtain a time series feature vector, wherein the calculation method of the time series feature extraction network is: h t =GRU(concat(X t ,A t )) Among them, h t is a time series feature vector, GRU() is a gated recurrent unit operation, concat(x t ,a t ) is the updated sequence, concat() is the concatenation operation, X t is the data point sequence, A t Label the updated missing value sequence, where t is the time index.
4. The unmanned substation intelligent operation and maintenance method according to claim 1 is characterized in that: The S3 step includes: First, based on the timing feature vector, the gated recurrent unit is used to predict the preliminary prediction value of the device at time t+2 to t+5. Secondly, based on the timing feature vector and the preliminary prediction value at time t+2 to t+5, the bidirectional long short-term memory network is used to calculate the final prediction value of the device at time t+1. Finally, based on the final prediction value of the device, the future feature vector is extracted, and the trend of the operating status of the device is judged in combination with the timing feature vector.
5. The unmanned substation intelligent operation and maintenance method according to claim 3 is characterized in that: The S3 step includes: S31: Calculate the preliminary prediction value of the device according to the time series feature vector, and the calculation method is: in, They are the preliminary predicted values of the data points from time t+2 to time t+5; S32: Construct a moment importance mechanism, calculate the weight of each moment and update the preliminary prediction value, and obtain the updated preliminary prediction value from t+2 to t+5. The calculation method is: Among them, α t+2 ,α t+3 ,α t+4 ,α t+5 are the weights from t+2 to t+5, u t+i is the fusion vector at time t+i, i is the counting index, tanh() is the hyperbolic tangent function, W u is the moment weight parameter matrix, is the initial predicted value of the data point at time t+i, b u is the moment weight bias vector, α t+i is the weight at time t+i, exp() is the exponential operation to remove the natural constant, v α is the moment weight parameter vector, u t+j is the fusion vector at time t+j, j is the counting index; It is the initial forecast value after updating at time t+2; Calculated Then, according to the above method Continue to calculate and get the final predicted value of the data points from t+3 to t+5 S33: Calculate the final prediction value of the device according to the time series feature vector and the updated preliminary prediction value. The calculation method is: is the final predicted value of the data point at time t+1, and BiLSTM is a bidirectional long short-term memory network operation; Calculated Then, continue to calculate in the above way to get the final predicted value of the data points from t+2 to t+5 S34: Extract the future feature vector according to the final prediction value of the device, and the calculation method is: Among them, h′ t is the future feature vector, CNN() is the convolutional neural network operation with a convolution kernel width of 1; S35: Combine the future feature vector and the time series feature vector to determine the trend of the equipment operation status. The calculation method is: p=softmax(MLP(concat(h t ,h′ t ))) Where p is the probability of device trend category, softmax() is the softmax function, and MLP() is the fully connected neural network operation; When p>50%, the device is judged as "having an abnormal trend"; When p≤50%, the device is judged as "normal trend"; S36: If the device has an abnormal trend, an abnormal trend alarm is issued.
6. The unmanned substation intelligent operation and maintenance method according to claim 5 is characterized in that: The S4 step includes: S41: Count the service life of the equipment and the number of abnormal trend alarms for each equipment; S42: Input the time series feature vector, the equipment service life and the number of abnormal trend alarms into the equipment health score network, and output the equipment health score. The calculation method of the equipment health score network is: score=sigmoid(MLP(h t ,T,n)) Where score is the health score of the device, sigmoid() is the sigmoid function, T is the service life of the device, and n is the number of abnormal trend alarms.
7. The unmanned substation intelligent operation and maintenance method according to claim 6 is characterized in that: The S5 step includes: S51: Set multi-level thresholds to divide the equipment health score into 4 health levels; S52: Compare the health score of the device with the set multi-level thresholds to determine the health level of the device; S53: According to the health level of the equipment, a preventive maintenance measure corresponding to the health level is executed.
8. An unmanned substation intelligent operation and maintenance system, characterized in that: include: Substation data acquisition and preprocessing module: collects equipment time series data and preprocesses it to obtain preprocessed equipment time series data; Timing feature extraction module: Based on the preprocessed device timing data, a timing feature extraction network is constructed to extract the timing feature vector; Trend judgment module: judge the trend of the equipment according to the time series feature vector, and issue an alarm if there is an abnormal trend; Equipment health scoring module: outputs equipment health scores based on the time series feature vector, combined with the equipment’s service life and the number of abnormal trend alarms; Health level calculation and preventive maintenance module: Perform multi-level threshold judgment based on the equipment health score to obtain the health level, and perform preventive maintenance on the equipment corresponding to the health level; To realize an unmanned substation intelligent operation and maintenance method as described in any one of claims 1-7.
Citation Information
Cited By
Transformer substation misoperation mode mining method for large-scale historical data
CN119885095A
Hydroelectric generating set health state monitoring method and system based on machine learning
CN119990887A
Power engineering design power distribution network planning method based on three-dimensional modeling
CN120338324A
A power engineering design power distribution network planning method based on three-dimensional modeling
CN120338324B
State edge monitoring method and device of unattended charging station, electronic equipment and medium
CN121608629A
