Distributed energy storage intelligent scheduling method and system based on transformer area autonomy

Abstract

The invention discloses a distributed energy storage intelligent scheduling method and system based on transformer area autonomy, and the method comprises the steps: firstly obtaining a task instruction that current power enters a target transformer area, and obtaining power load data according to the task instruction; and then, evaluating the load data by using a pre-trained transformer area load evaluation model, and judging whether the load state is normal or not. And if so, determining a main root cause causing the abnormality. And finally, based on load state evaluation and abnormal state analysis results, selecting a proper scheduling strategy from pre-configured energy storage scheduling strategies, and issuing the selected scheduling strategy to the distributed energy storage equipment of the target station area for execution. Through the design, the stability and the scheduling efficiency of the power system are improved, and the method is suitable for intelligent management of a distributed energy system.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a distributed energy storage intelligent scheduling method and system based on substation autonomy. Background Art

[0002] With the rapid development of distributed energy and smart grid technologies, the intelligent dispatching of substation autonomy and distributed energy storage systems has become increasingly important. Traditional power dispatching methods usually rely on centralized control systems, which may not be flexible and efficient enough in distributed energy systems. Summary of the invention

[0003] The object of the present invention is to provide a distributed energy storage intelligent scheduling method and system based on substation autonomy.

[0004] In a first aspect, an embodiment of the present invention provides a distributed energy storage intelligent scheduling method based on substation autonomy, comprising:

[0005] Obtain the task instructions for the current power to enter the target area;

[0006] According to the task instruction, acquiring the power load data of the current power;

[0007] Inputting the current power load data into a pre-trained load assessment model to obtain a corresponding load state assessment result;

[0008] When the load state evaluation result is characterized as an abnormal load state, determining a target abnormal root cause with the highest correlation with the abnormal phenomenon of the abnormal load state, and obtaining an abnormal state analysis result;

[0009] Based on the load state evaluation result and the abnormal state analysis result, selecting a target scheduling strategy from pre-configured energy storage scheduling;

[0010] The target scheduling strategy is sent to the distributed energy storage devices included in the target area for execution.

[0011] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is used to execute the method described in the first aspect.

[0012] Compared with the prior art, the beneficial effects provided by the present invention include: the present invention discloses a distributed energy storage intelligent scheduling method and system based on substation autonomy, including: first, obtaining the task instructions for the current power entering the target substation, and obtaining the power load data according to the task instructions. Then, the load data is evaluated using a pre-trained substation load assessment model to determine whether the load status is normal. If abnormal, the main root cause of the abnormality is determined. Finally, based on the load status evaluation and abnormal status analysis results, a suitable scheduling strategy is selected from the pre-configured energy storage scheduling strategies, and sent to the distributed energy storage equipment in the target substation for execution. Such a design improves the stability and scheduling efficiency of the power system, and is suitable for the intelligent management of distributed energy systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.

[0014] Figure 1 A schematic diagram of the steps of a distributed energy storage intelligent scheduling method based on substation autonomy provided in an embodiment of the present invention;

[0015] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0016] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0017] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.

[0018] In order to solve the technical problems in the aforementioned background technology, Figure 1 The present invention provides a flow chart of a distributed energy storage intelligent scheduling method based on substation autonomy provided in an embodiment of the present invention. The distributed energy storage intelligent scheduling method based on substation autonomy is introduced in detail below.

[0019] Step S201, obtaining a task instruction for the current power to enter the target area;

[0020] Step S202, acquiring the power load data of the current power according to the task instruction;

[0021] Step S203, inputting the current power load data into a pre-trained area load assessment model to obtain a corresponding load state assessment result;

[0022] Step S204, when the load state evaluation result is characterized as an abnormal load state, determining the target abnormal root cause with the highest correlation with the abnormal phenomenon of the abnormal load state, and obtaining an abnormal state analysis result;

[0023] Step S205, selecting a target scheduling strategy from pre-configured energy storage scheduling based on the load state evaluation result and the abnormal state analysis result;

[0024] Step S206: Send the target scheduling strategy to the distributed energy storage devices included in the target substation for execution.

[0025] In an embodiment of the present invention, exemplarily, the server receives a task instruction from the power management system, requiring monitoring and dispatching of the power load of "No. 10 Substation, XX District, XX City". The task instruction includes information such as the specific location of the substation, the monitoring time period, and the main power parameters that need to be paid attention to. According to the received task instruction, the server obtains the real-time power load data of "No. 10 Substation, XX District, XX City" in the past hour through the remote data acquisition system. These data include key parameters such as voltage, current, active power, and reactive power. The server inputs the collected power load data into a pre-trained substation load assessment model. This model is constructed by analyzing and learning historical power load data through a machine learning algorithm. After processing the model, the server obtains a load state assessment result, which shows that the substation is currently in a high load state and has abnormal fluctuations. Since the load state assessment result shows that there is an abnormality, the server further analyzes the data, compares the differences between historical data and current data, and uses anomaly detection algorithms to determine that the main root cause of the abnormality is "a line short circuit or equipment failure somewhere in the substation". According to the high load and abnormal fluctuations in the load state, as well as the identified root causes of the abnormalities, the server selected a strategy of "emergency load reduction and priority power supply of energy storage equipment" from a variety of pre-configured energy storage scheduling strategies. This strategy aims to ensure the stable supply of power in the substation by reducing some non-critical loads and starting distributed energy storage equipment at the same time. The server sends the selected scheduling strategy to all distributed energy storage devices in "Substation No. 10, XX District, XX City". After receiving the instructions, these devices make adjustments according to the strategy requirements, such as starting the discharge of energy storage batteries, cutting off some non-critical circuits, etc., to ensure the smooth operation of power in the substation.

[0026] In an embodiment of the present invention, the substation load assessment model is obtained in the following manner.

[0027] Acquire power load data recorded during the process of transmitting target power to the target substation, and associate the power load data with transformer equipment to obtain sub-power load data corresponding to multiple transformer equipment respectively;

[0028] Based on each of the sub-power load data, determining a load state corresponding to the corresponding transformer device, wherein the load state includes a safe load state and an abnormal load state;

[0029] Taking each of the sub-power load data as a sample instance and the corresponding load state as an instance target value, a sample instance array for training the substation load assessment model is constructed, and based on the sample instance array, a training process is performed on the substation load assessment model;

[0030] The substation load assessment model is used to assess the load status of the substation where the target power is located based on the power load data of the target power.

[0031] In an embodiment of the present invention, illustratively, the server first collects the load data generated when the "XX City Power Grid" transmits electricity to "XX City XX District Various Substations". These data cover the voltage, current and power changes during the entire transmission process from the power station to the user end. Then, the server uses data analysis technology to associate these original power load data with the transformer equipment in each substation. For example, for the three transformer equipment (equipment A, equipment B, equipment C) in "XX City XX District No. 10 Substation", the server calculates their respective sub-power load data, including the average load and peak load of each equipment in different time periods. The server analyzes the sub-power load data of each transformer equipment and determines the load status of each transformer equipment according to the preset safety threshold and abnormal threshold. For example, for equipment A, its average load in the past 24 hours is lower than the safety threshold, so it is judged to be in a "safe load state"; while for equipment B, its load peak at certain time points exceeds the abnormal threshold, so it is judged to be in an "abnormal load state". The server takes the sub-power load data of each transformer as sample input and the corresponding load status (safe or abnormal) as target output, and constructs a large array of sample instances. This array is used to train the load assessment model of the substation. Through machine learning algorithms (such as support vector machines, neural networks, etc.), the server continuously adjusts the parameters of the model so that it can more accurately predict the load status of the transformer equipment based on the input power load data. After the above training process, the server obtains an efficient substation load assessment model. This model can receive the power load data of "Substation No. 10, District XX, XX City" in real time and quickly determine whether the overall load status of the current substation is safe or abnormal. For example, when the model receives a new set of power load data, it automatically analyzes the data and outputs a load status assessment result, such as "The current substation load is high, but still within the safe range" or "Abnormal load fluctuations are detected, and it is recommended to further check the lines and equipment."

[0032] In the embodiment of the present invention, the training process of the substation load assessment model based on the sample instance array can be implemented through the following example.

[0033] The following processing is performed for each of the sub-power load data in the sample instance array:

[0034] Mapping the sub-power load data from data processing domains at multiple levels to obtain conversion domain data corresponding to each data processing domain;

[0035] Performing feature extraction operations on each of the conversion domain data to obtain conversion domain features corresponding to each of the conversion domain data, and performing feature integration processing on each of the conversion domain features to obtain integrated features;

[0036] Using the substation load assessment model, based on the integrated features, the substation load assessment is performed on the target substation to obtain an inferred load state of the target substation;

[0037] In combination with the inferred load state and the corresponding load state, a training process is performed on the substation load assessment model.

[0038] In an embodiment of the present invention, illustratively, after the server obtains the power load data of "No. 10 Substation in XX District, XX City", it starts to process these data. First, the server maps the sub-power load data from multiple data processing domains such as time domain, frequency domain, and wavelet domain. For example, in the time domain, the server arranges the data in chronological order to form time series data; in the frequency domain, the data is converted from the time domain to the frequency domain by Fourier transform to analyze the spectrum characteristics of electricity; in the wavelet domain, the local features of the data are extracted by wavelet transform. These mapping processes help the server to fully understand the characteristics of power load data from multiple angles. Next, the server extracts features from each conversion domain data. In the time domain data, statistical features such as peak value, valley value, and average value of the load are extracted; in the frequency domain data, features such as main frequency and spectrum energy are extracted; in the wavelet domain data, features such as wavelet coefficients and energy distribution are extracted. Then, the server integrates these features from different conversion domains to form a comprehensive and rich integrated feature set, which can more comprehensively reflect the inherent laws and abnormal patterns of power load data. The server inputs the integrated features into the substation load assessment model, and the model infers the load status of "Substation No. 10, XX District, XX City" based on these features. Through complex calculations and learning processes within the model, an inferred load status is output, such as "the current substation is in an abnormal load state and there is a risk of overload." The server compares the load status inferred by the model with the actual observed load status. If a deviation is found between the inferred state and the actual state, the server will adjust the parameters and structure of the model to make it more in line with the actual situation. This process is achieved through the optimization technology in the machine learning algorithm, with the aim of making the model more accurate in future inferences. Through repeated training and adjustment processes like this, the performance of the substation load assessment model will gradually improve, and it will eventually be able to accurately assess the load status of the substation.

[0039] In an embodiment of the present invention, the data processing domain includes a load time series analysis domain, an energy spectrum analysis domain and a power quality fluctuation analysis domain. The data processing domains at multiple levels respectively map the sub-power load data to obtain conversion domain data corresponding to each of the data processing domains, which can be implemented through the following examples.

[0040] From the load time series analysis domain, performing load time series mapping processing on the sub-power load data to obtain load time series conversion domain data corresponding to the load time series analysis domain;

[0041] From the energy spectrum analysis domain, performing energy spectrum mapping processing on the sub-power load data to obtain energy spectrum conversion domain data corresponding to the energy spectrum analysis domain;

[0042] From the power quality fluctuation analysis domain, power quality fluctuation mapping processing is performed on the sub-power load data to obtain power quality fluctuation conversion domain data corresponding to the power quality fluctuation analysis domain.

[0043] In an embodiment of the present invention, illustratively, the server will first process the data from the perspective of the load time series analysis domain. For example, for the power load data of "No. 10 Substation, XX District, XX City" in a certain time period (such as 24 hours a day), the server will arrange these data in time series to form a continuous load time series. This sequence reflects the change of the power load of the substation over time. Through mapping processing, the server obtains the load time series conversion domain data corresponding to the load time series analysis domain, which can intuitively show the fluctuation trend and peak period of the power load. Next, the server will process the same group of sub-power load data from the energy spectrum analysis domain. Through mathematical methods such as Fourier transform, the server converts these time series data into frequency domain data, that is, analyzes the spectrum characteristics of electricity. This conversion process reveals the energy distribution of different frequency components in the power load. For example, the server found that the power load of "No. 10 Substation, XX District, XX City" has obvious energy concentration in a certain frequency band, which can indicate a certain periodic or regular power consumption behavior. Through energy spectrum mapping processing, the server obtains energy spectrum conversion domain data corresponding to the energy spectrum analysis domain. Finally, the server will also process the sub-power load data from the power quality fluctuation analysis domain. This step is mainly to analyze the abnormal fluctuations and disturbances in the power load data to evaluate the stability of the power quality. For example, the server will detect that during a certain period of time in the "No. 10 Substation in XX District, XX City", the power load experienced a short and drastic fluctuation, which may be caused by equipment failure, line problems or external interference. Through the power quality fluctuation mapping processing, these abnormal fluctuations are clearly identified and the power quality fluctuation conversion domain data corresponding to the power quality fluctuation analysis domain is formed. These data are crucial for the subsequent identification and resolution of power quality problems.

[0044] In an embodiment of the present invention, the sub-power load data includes the instantaneous load data corresponding to each sampling time node in the process of transmitting the target power to the corresponding transformer equipment. The sub-power load data is subjected to load time series mapping processing from the load time series analysis domain to obtain the load time series conversion domain data corresponding to the load time series analysis domain. The implementation can be performed through the following example.

[0045] Obtaining the load mean value, root mean square load value, load fluctuation, load dispersion of each instantaneous load data, and peak value data in the instantaneous load data;

[0046] Dividing the load dispersion by the load mean value to obtain the rate of change deviation of the sub-power load data, dividing the peak data by the root mean square load value to obtain the peak-to-average ratio of the sub-power load data, and dividing the peak data by the load mean value to obtain the impact ratio of the sub-power load data;

[0047] Data integration processing is performed on the load mean, the load fluctuation, the change rate deviation, the peak-to-average ratio and the impact ratio to obtain load time series conversion domain data corresponding to the load time series analysis domain.

[0048] In an embodiment of the present invention, exemplarily, the server first obtains the instantaneous load data of a transformer device in "No. 10 Substation, XX District, XX City" at each sampling time node in the past 24 hours. These data reflect the real-time changes of the power load in a short period of time. Next, the server begins to calculate the statistical indicators of these instantaneous load data, including the load mean (i.e., the average value of all instantaneous load data), the root mean square load value (a statistic reflecting the load intensity), the load fluctuation (indicating the severity of the load change), the load dispersion (indicating the dispersion of the load data), and the peak data in the instantaneous load data (i.e., the maximum load value). After obtaining the above statistical indicators, the server further performs data processing. First, the load dispersion is divided by the load mean to calculate the rate of change deviation of the sub-power load data, which reflects the relative fluctuation of the load data. Next, the peak data is divided by the root mean square load value to obtain the peak-to-average ratio of the sub-power load data, which represents the proportional relationship between the maximum load and the average load intensity. Finally, the peak data is divided by the load mean to calculate the impact ratio of the sub-power load data, which reflects the impact degree of the maximum load relative to the average load. The server integrates and processes the calculated load mean, load fluctuation, rate of change deviation, peak-to-average ratio, and impact ratio. These data describe the characteristics of the power load from different perspectives. Through integrated processing, a more comprehensive and richer data representation can be formed, that is, the load time series conversion domain data corresponding to the load time series analysis domain. These data not only reflect the basic statistical characteristics of the power load, but also reveal important information such as the volatility, stability, and peak characteristics of the load, providing strong data support for subsequent load assessment in the substation area.

[0049] In the embodiment of the present invention, the energy spectrum mapping process is performed on the sub-power load data from the energy spectrum analysis domain to obtain the energy spectrum conversion domain data corresponding to the energy spectrum analysis domain, which can be implemented through the following examples.

[0050] Obtaining the spectrum distribution intensity and the power spectrum density mean corresponding to the sub-power load data, and determining the spectrum center frequency of the sub-power load data on the energy spectrum based on the spectrum distribution intensity;

[0051] Determine the spectrum fluctuation of the sub-power load data on the energy spectrum by combining the spectrum center frequency and the spectrum distribution intensity, and obtain the spectrum upper limit frequency in the spectrum distribution intensity;

[0052] Data integration processing is performed on the spectrum distribution intensity, the power spectrum density mean, the spectrum center frequency, the spectrum fluctuation and the spectrum upper limit frequency to obtain energy spectrum conversion domain data corresponding to the energy spectrum analysis domain.

[0053] In an embodiment of the present invention, illustratively, taking "District No. 10, District XX, City" as an example, the server first obtains the sub-power load data of the district within a specific time period, and converts these time series data into frequency domain data through signal processing techniques such as Fourier transform. During the conversion process, the server calculates the amplitude of each frequency component to obtain the spectrum distribution intensity, that is, the energy distribution at each frequency. At the same time, the server also calculates the mean of the power spectrum density, which is a statistic that represents the distribution of the average power of the signal with frequency. Next, the server determines the spectrum center frequency of the sub-power load data on the energy spectrum based on the spectrum distribution intensity. This center frequency can be understood as the frequency point where the energy is most concentrated, which reflects the main periodic components in the power load data. For example, in the power load data of "District No. 10, District XX, City", the server finds that the spectrum center frequency is 50Hz, which indicates that the power load of the district is mainly based on the periodic change of 50Hz. The server also combines the spectrum center frequency and the spectrum distribution intensity to determine the spectrum fluctuation of the sub-power load data on the energy spectrum. This fluctuation reflects the degree of dispersion of the spectrum energy, that is, the distribution of energy at different frequencies. At the same time, the server also obtains the spectrum upper limit frequency from the spectrum distribution intensity, which is the highest frequency part of the energy distribution and is usually related to the highest operating frequency of the equipment in the power system. Finally, the server integrates and processes data such as spectrum distribution intensity, power spectrum density mean, spectrum center frequency, spectrum fluctuation, and spectrum upper limit frequency. These data characterize the characteristics of power load in the frequency domain from different angles. Through integrated processing, the server obtains a data representation that fully reflects the frequency domain characteristics of power load, that is, the energy spectrum conversion domain data corresponding to the energy spectrum analysis domain. These data are of great significance for analyzing the operating status of the power system and identifying potential faults and anomalies.

[0054] In the embodiment of the present invention, performing feature integration processing on each of the conversion domain features to obtain integrated features can be implemented through the following examples.

[0055] Acquire the feature domain attribute corresponding to each of the conversion domain features, and perform weighted average processing on each of the feature domain attributes to obtain a reference domain attribute;

[0056] Compare each of the feature domain attributes with the reference domain attributes to obtain an attribute comparison result corresponding to each of the feature domain attributes;

[0057] When the attribute comparison result indicates that the feature domain attribute is consistent with the reference domain attribute, the corresponding conversion domain feature is determined as a target conversion domain feature corresponding to the conversion domain feature;

[0058] When the attribute comparison result indicates that the feature domain attribute is inconsistent with the reference domain attribute, the feature domain attribute of the corresponding conversion domain feature is updated to the reference domain attribute to obtain a target conversion domain feature corresponding to the conversion domain feature;

[0059] Perform feature integration processing on each of the target conversion domain features to obtain the integrated features.

[0060] In an embodiment of the present invention, exemplarily, the server first obtains conversion domain features from the load time series analysis domain, the energy spectrum analysis domain, and the power quality fluctuation analysis domain. Each conversion domain feature has its corresponding characteristic domain attribute, such as the characteristic domain attribute of the load time series conversion domain data may include the stability and periodicity of the time series; the characteristic domain attribute of the energy spectrum conversion domain data may include the concentration and volatility of the spectrum; the characteristic domain attribute of the power quality fluctuation conversion domain data may include the amplitude and frequency of the fluctuation. Next, the server performs weighted averaging on these characteristic domain attributes to obtain a comprehensive reference domain attribute. This reference domain attribute reflects an average or typical characteristic of all conversion domain features. The server then compares each characteristic domain attribute with the reference domain attribute. For example, if a characteristic domain attribute of a load time series conversion domain data shows a higher stability, while the reference domain attribute shows a slightly lower overall stability, then the comparison result between the two will show inconsistency. When the attribute comparison result shows that the characteristic domain attribute is consistent with the reference domain attribute, the server will retain the conversion domain feature as the target conversion domain feature. If there is inconsistency, the server will update the feature domain attribute of the conversion domain feature to the reference domain attribute to ensure that all conversion domain features are consistent during integration. After completing the above-mentioned attribute comparison and update, the server will perform feature integration processing on all target conversion domain features. This process may include operations such as feature weighting, fusion or dimensionality reduction to obtain a comprehensive integrated feature that can fully reflect the characteristics of power load data. For example, in the power load data analysis of "No. 10 Substation in XX District, XX City", the server may integrate conversion domain features from different analysis domains to obtain an integrated feature that takes into account time series stability and integrates spectrum characteristics and power quality fluctuations. This integrated feature provides a more accurate and comprehensive data basis for subsequent tasks such as substation power load assessment, prediction and fault detection.

[0061] In the embodiment of the present invention, the step of acquiring the power load data recorded during the process of transmitting the target power to the target substation may be implemented through the following examples.

[0062] Obtaining the voltage fluctuation and current change corresponding to each sampling time node during the process of transmitting the target power to the target substation;

[0063] For each of the sampling time nodes, the corresponding normalized value of the voltage fluctuation and the normalized value of the current change are integrated to obtain the instantaneous load data corresponding to the sampling time node;

[0064] Each instantaneous load data is arranged according to the time sequence of the sampling time nodes to obtain the power load data.

[0065] In an embodiment of the present invention, illustratively, taking "No. 10 Substation in XX District, XX City" as an example, the server will monitor and record the voltage and current conditions of the substation when receiving the target power in real time. At each sampling time node (such as every second or every minute), the server will capture the voltage fluctuation and current change data at that moment. These data reflect the actual operating state of the power system at that moment. After obtaining the original voltage and current data, the server will perform standardization processing to eliminate the influence of different dimensions and orders of magnitude on data analysis. For example, the server will convert the voltage and current data into standardized values ​​relative to their average and standard deviation. Next, for each sampling time node, the server will integrate the corresponding standardized values ​​of voltage fluctuations and standardized values ​​of current changes. This integration may include weighted averaging, summation or other mathematical operations to obtain an instantaneous load data that can comprehensively reflect the voltage and current status at that moment. After completing the data integration, the server will arrange each instantaneous load data according to the time series sequence of the sampling time node. In this way, the server obtains a series of instantaneous load data arranged in chronological order, which together constitute the power load data. These power load data not only reflect the operating status of the power system at each moment, but also provide an important data basis for subsequent load analysis, prediction and fault detection. Through the above steps, the server can accurately obtain and process the power load data recorded during the transmission of the target power to the target substation, providing strong data support for the stable operation and optimization of the power system.

[0066] In an embodiment of the present invention, the sub-power load data includes instantaneous load data corresponding to each sampling time node in the process of transmitting the target power to the corresponding transformer device. Based on each of the sub-power load data, the load state corresponding to the corresponding transformer device is determined, which can be implemented through the following examples.

[0067] The following processing is performed for each of the sub-power load data:

[0068] For each of the sampling time nodes corresponding to the sub-power load data, based on the instantaneous load data corresponding to the sampling time node, determine the target load safety level corresponding to the sampling time node;

[0069] When the target load safety level indicates that the area load of the target area through which the target power passes at the sampling time node is within a safe load range, the sampling time node corresponding to the target load safety level is determined as the target sampling time node;

[0070] Based on the number of nodes at the target sampling time node, a load state corresponding to the corresponding transformer device is determined.

[0071] In an embodiment of the present invention, illustratively, taking the transformer equipment of "No. 10 Substation in XX District, XX City" as an example, the server will obtain the sub-power load data of the transformer equipment when receiving the target power. These data include instantaneous load data recorded at each sampling time node (for example, every minute). For each sampling time node, the server will judge the target load safety level of the node based on the instantaneous load data (such as voltage, current, power factor, etc.) and the preset safety threshold. The safety level can be divided into different levels such as "safe", "warning" and "dangerous". After determining the target load safety level of each sampling time node, the server will filter out those sampling time nodes whose safety level is characterized as "safe", that is, at these time points, the substation load through which the target power passes is within the safe load range. These filtered sampling time nodes are called target sampling time nodes. Based on the number of filtered target sampling time nodes, the server will further judge the load status of the corresponding transformer equipment. For example, if the proportion of the number of target sampling time nodes to the total number exceeds a preset value (such as 90%), the server can determine that the transformer equipment is in a "normal load state"; if the proportion is lower (such as less than 60%), it can be determined as a "high load state" or "overload risk state". Specifically, suppose that in the past hour, the server has sampled the transformer equipment of "No. 10 Substation, XX District, XX City" 60 times (once per minute). After analysis, the target load safety level of 54 sampling time nodes is "safe". Based on this data, the server can judge that in the past hour, the transformer equipment has been in a safe load range most of the time, so its load state can be determined as a "normal load state". Through the above steps, the server can monitor and accurately determine the load state of the transformer equipment in real time, thereby providing important guarantees for the stable operation of the power system.

[0072] In the embodiment of the present invention, the determination of the target load safety level corresponding to the sampling time node based on the instantaneous load data corresponding to the sampling time node can be implemented through the following examples.

[0073] Acquire at least one adjacent sampling time node of the sampling time node, and perform difference calculation on the instantaneous load data corresponding to the sampling time node and the instantaneous load data corresponding to each of the adjacent sampling time nodes, so as to obtain the load fluctuation variation corresponding to each of the adjacent sampling time nodes;

[0074] When at least one of the load fluctuation changes exceeds a preset change threshold, determining the target load safety level as a first load safety level;

[0075] When there is no load fluctuation variation exceeding the preset variation threshold, determining the target load safety level as the second load safety level;

[0076] Among them, the first load safety level is used to characterize that the load of the substation through which the target power passes at the sampling time node is within the abnormal load range, and the second load safety level is used to characterize that the load of the substation through which the target power passes at the sampling time node is within the safe load range.

[0077] In an embodiment of the present invention, illustratively, taking "No. 10 Substation, XX District, XX City" as an example, the server obtains the instantaneous load data of the node at a specific sampling time node (such as T1). In order to analyze the volatility of the load, the server will also obtain the instantaneous load data of at least one sampling time node adjacent to T1 (such as T0 one minute before T1 and T2 one minute after T1). The server will then perform a difference operation on the instantaneous load data at time T1 with the instantaneous load data at time T0 and T2 respectively. For example, if the load at time T1 is 100 kilowatts, 90 kilowatts at time T0, and 110 kilowatts at time T2, then the load fluctuation change with T0 is 10 kilowatts, and the load fluctuation change with T2 is also 10 kilowatts. The server will pre-set a threshold value for the load fluctuation change (such as 5 kilowatts) to determine whether the load fluctuates abnormally. In the above example, the load fluctuation changes at two adjacent time nodes are both 10 kilowatts, which exceeds the preset 5 kilowatt threshold. Since there is at least one load fluctuation variation that exceeds the preset variation threshold, the server will determine the target load safety level at time T1 as the first load safety level. This means that at time T1, the power load of "Substation No. 10, District XX, XX City" has experienced abnormal fluctuations, posing a safety hazard. If the load fluctuation variation of all adjacent sampling time nodes does not exceed the preset threshold, the server will determine the target load safety level as the second load safety level, indicating that the power load of the substation fluctuates within a safe range. Through the above steps, the server can monitor the volatility of the power load in real time, promptly detect and warn of potential power grid safety issues, and thus ensure the stable operation of the power system.

[0078] In the embodiment of the present invention, the determination of the load state corresponding to the corresponding transformer device based on the number of nodes at the target sampling time node may be implemented through the following examples.

[0079] Compare the number of nodes at the target sampling time node with the node number threshold to obtain a node number comparison result;

[0080] When the node number comparison result indicates that the node number of the target sampling time node exceeds the node number threshold, determining the load state as the safe load state;

[0081] When the node number comparison result indicates that the node number of the target sampling time node does not exceed the node number threshold, the load state is determined to be the abnormal load state.

[0082] In an embodiment of the present invention, illustratively, the server first sets a node number threshold, which is determined based on historical data and the safe operation requirements of the power grid. For example, assuming that the node number threshold set by the server is 50, it means that among the 100 time nodes sampled continuously, the power load of at least 50 nodes needs to be within the safe range before the overall load state can be considered safe. Subsequently, the server will count the number of target sampling time nodes that are determined to be within the safe load range within a period of time (such as 60 sampling points within an hour). Assume that during this period of time, the power load of 55 sampling time nodes is within the safe range. Next, the server will compare these 55 safe node numbers with the set node number threshold of 50. Since 55 is greater than 50, the node number comparison result shows that the node number of the target sampling time node exceeds the node number threshold. Based on the above comparison results, the server will determine the load state of the transformer equipment as a safe load state. This means that during this period of time, most of the power load borne by the transformer equipment is safe, and there is no continuous overload. On the contrary, if the number of safe nodes counted is less than or equal to the node number threshold, for example, only 45 nodes are within the safe load range, then the server will determine the load status as an abnormal load status. This means that the transformer equipment may have experienced more overload conditions during this period, posing a safety risk and requiring further inspection or adjustment. Through such an analysis process, the server can monitor and accurately evaluate the load status of the transformer equipment in real time, providing important guarantees for the safe and stable operation of the power grid.

[0083] In the embodiment of the present invention, the determination of the load state corresponding to the corresponding transformer device based on the number of nodes at the target sampling time node may be implemented through the following examples.

[0084] Dividing the number of nodes at the target sampling time node by the number of nodes at the sampling time node to obtain the load stability rate of the target substation;

[0085] When the load stability rate exceeds a load stability rate threshold, the load state is determined as the safe load state, and when the load stability rate does not exceed the load stability rate threshold, the load state is determined as the abnormal load state.

[0086] In an embodiment of the present invention, illustratively, taking the transformer equipment of "No. 10 Substation in XX District, XX City" as an example, the server samples the substation 240 times in one day (10 times per hour, a total of 24 hours), and records the power load data of each sampling time node. After analysis, the power load of 216 sampling time nodes is within the safe range and is marked as the target sampling time node. In order to evaluate the load stability of the substation, the server divides the number of these 216 target sampling time nodes by the total number of sampling time nodes 240 to obtain the load stability rate. In this example, the load stability rate is 216 / 240=0.9, that is, 90%. The server will pre-set a load stability rate threshold to determine whether the load state is safe. This threshold is usually determined based on historical data and the safety requirements of the power grid. Assume that in this scenario, the load stability rate threshold is set to 85%. The calculated load stability rate of 90% is compared with the set load stability rate threshold of 85%. Since 90% is greater than 85%, the load stability rate exceeds the load stability rate threshold. Based on the above comparison results, the server will determine the load status of the transformer equipment of "No. 10 Substation in XX District, XX City" as a safe load status. This means that in the past day, the power load of this substation has remained within the safe range most of the time and has a high stability. On the contrary, if the calculated load stability rate is lower than 85%, for example, only 80%, then the server will determine the load status as an abnormal load status. This means that the power load in the substation fluctuates greatly, there are safety hazards, and further inspection or adjustment is required. Through this method, the server can objectively evaluate the load stability of the transformer equipment, promptly discover potential safety problems, and take corresponding measures to ensure the safe and stable operation of the power grid.

[0087] In the embodiment of the present invention, after executing the training process for the substation load assessment model based on the sample instance array, the embodiment of the present invention also provides the following implementation method.

[0088] Obtaining a power dispatching plan for the target power, and searching a load state of a reference substation corresponding to the power dispatching plan from a substation load database to obtain a search result;

[0089] When the search result indicates that the load state of the reference area is not stored in the area load database, obtaining target load data of the reference area in a similar time period, performing a feature extraction operation on the target load data, and obtaining a target load feature;

[0090] Using the trained area load assessment model, based on the target load characteristics, the area load assessment is performed on the reference area to obtain an inferred load state of the reference area;

[0091] The inferred load state of the reference substation is updated into the substation load database to obtain an updated substation load database.

[0092] In an embodiment of the present invention, illustratively, the server first obtains the power dispatch plan of the target power, which details the power distribution and dispatch arrangement in the future. According to this plan, the server needs to search the load state of the corresponding reference area from the area load database. For example, the power dispatch plan may specify "No. 5 Area, XX District, XX City" as the reference area, and the load state in a specific time period (such as 9 am to 11 am) is the information that the server needs to query. After the server performs the search, it is found that the load state information of "No. 5 Area, XX District, XX City" in the specified time period is not stored in the area load database. This means that additional data processing is required to fill this information gap. Due to the lack of directly related data in the database, the server turns to find the target load data of "No. 5 Area, XX District, XX City" in a similar time period (such as 8 am to 10 am). These data include key indicators such as voltage, current, and power factor, which can indirectly reflect the load state of the area. The server performs feature extraction operations on the acquired target load data. For example, by calculating characteristic values ​​such as average voltage, current peak, and power fluctuation, the load characteristics of the area in a similar time period are summarized. These characteristic values ​​will help to accurately infer the load assessment model. Next, the server uses the trained substation load assessment model to infer the load status of "Substation No. 5, District XX, XX City" in the specified time period based on the extracted target load characteristics. The model outputs a numerical value or classification label representing the load status, such as "safe", "warning" or "dangerous". Finally, the server updates the inferred load status information to the substation load database. In this way, in future power dispatching and load assessment processes, the server will be able to directly use these updated data without having to infer again. Through the above process, the server can not only fill in the information gaps in the substation load database in a timely manner, but also ensure the accuracy of power dispatching and load assessment, thereby maintaining the stable operation of the power system.

[0093] In the embodiments of the present invention, the following implementation modes are also provided.

[0094] When the search result indicates that the load status of the reference station is stored in the station load database, and the load status of the reference station indicates that the station load of the reference station is within an abnormal load range, generating a first notification message;

[0095] When the search result indicates that the load status of the reference area is stored in the area load database, and the load status of the reference area indicates that the area load of the reference area is within a safe load range, generating a second notification message;

[0096] The first notification message is used to notify the adjustment of the power dispatching plan on the substation load database, and the second notification message is used to notify that the power load of the reference substation corresponding to the power dispatching plan is normal.

[0097] In an embodiment of the present invention, illustratively, the server first searches the load status of a specific reference substation in the substation load database. For example, for "Substation No. 3, XX District, XX City", the server finds the corresponding load status record in the database. If the search result shows that the load status of "Substation No. 3, XX District, XX City" is in the abnormal load range, such as excessive load or abnormal fluctuation, the server will immediately generate a first notification message. This message may include the specific location of the substation, detailed data of the abnormal load, possible cause analysis, and recommended adjustment measures. The server sends the first notification message to the power dispatching center or relevant management personnel. For example, the message may point out: "The current load of Substation No. 3, XX District, XX City is abnormal. Please adjust the power dispatching plan as soon as possible to reduce the load or conduct equipment inspection." After receiving the notification, the relevant personnel can respond quickly to avoid potential power failures. On the other hand, if the search result shows that the load status of "Substation No. 3, XX District, XX City" is within the safe load range, the server will generate a second notification message. This message is a confirmation of the current power dispatching plan, indicating that the load of the substation is normal and no emergency adjustment is required. The server sends the second notification message to the power dispatching center or management personnel for reference and record. For example, the message can state: "The current load of the No. 3 power distribution area in the XX district of XX city is normal, and the power dispatch plan can continue to be implemented." Such notifications help maintain the transparency of the system and the continuity of management. Through the above process, the server can provide timely feedback based on the actual situation to ensure the flexibility of the power dispatch plan and the safety of the power distribution area load. This real-time monitoring and notification mechanism is crucial to maintaining the stable operation of the power system.

[0098] In the embodiment of the present invention, the aforementioned step S204 may be implemented through the following example execution.

[0099] Identify abnormal phenomena and multiple undetermined abnormal root causes of abnormal load conditions;

[0100] Constructing a sample instance array based on each of the undetermined abnormal root causes and the characteristic parameters corresponding to the abnormal phenomenon;

[0101] Iterative expert model training is performed based on the sample instance array to determine the target abnormal root cause determined through multiple rounds of cycles from each of the pending abnormal root causes; wherein the sample instance of the latter round of expert model training includes characteristic parameters of the abnormal phenomenon and characteristic parameters of the key abnormal root cause determined by the previous round of expert model training, and the characteristic parameters of the key abnormal root cause determined by the previous round of expert model training are different from the characteristic parameter distribution of the characteristic parameters of the abnormal root cause in the sample instance of the previous round of expert model training; the correlation coefficient between the key abnormal root cause and the abnormal phenomenon meets the preset correlation coefficient threshold;

[0102] The correlation and matching relationship between the abnormal phenomenon and the target abnormal root cause is used as the abnormal state analysis result of the abnormal load state.

[0103] In an embodiment of the present invention, illustratively, the server first detects an abnormal load state in a certain substation (for example, "Substation No. 7, XX District, XX City"), which is manifested as a sudden increase in power load, voltage fluctuations and other abnormal phenomena. Then, the server lists multiple pending abnormal root causes that may cause such abnormal phenomena, such as equipment failure, abnormal external power grid power supply, changes in user power consumption behavior, etc. For each pending abnormal root cause and abnormal phenomenon, the server collects relevant characteristic parameters. For example, for the pending abnormal root cause of equipment failure, the characteristic parameters may include equipment temperature, operating time, historical fault records, etc.; for the external power grid power supply abnormality, the characteristic parameters may include grid voltage fluctuations, frequency changes, etc. These characteristic parameters and abnormal phenomena (such as load increase and voltage fluctuations) together constitute a sample instance array. The server uses these sample instance arrays for iterative expert model training. In each round of training, the model learns the relationship between the abnormal phenomenon and the pending abnormal root cause. Importantly, each round of training uses the characteristic parameters of the key abnormal root causes determined by the previous round of training, and the correlation coefficients of these key abnormal root causes and abnormal phenomena meet the preset thresholds. In this way, the model can gradually focus on those abnormal root causes that are more correlated with abnormal phenomena. For example, in the first round of training, the model found that both equipment failure and external power grid power supply abnormality are associated with abnormal load status, but the characteristic parameters of equipment failure are more correlated with abnormal phenomena. In the second round of training, the model will consider more characteristic parameters of equipment failure and further refine the analysis. After multiple rounds of training, the model will determine a target abnormal root cause with the highest correlation with abnormal phenomena. In this example, the model can finally determine that equipment failure is the main cause of the abnormal load state of "No. 7 Substation, XX District, XX City". Finally, the server will use the correlation matching relationship between abnormal phenomena (sudden load increase, voltage fluctuation) and target abnormal root cause (equipment failure) as the abnormal state analysis result of abnormal load status. This result can help power workers quickly locate problems and take corresponding maintenance measures.

[0104] In the embodiment of the present invention, the abnormal phenomenon of the abnormal load state and multiple undetermined abnormal root causes are determined, and the package can be implemented through the following examples.

[0105] Acquire an abnormal analysis indication for an abnormal load state, and extract an abnormal phenomenon of the abnormal load state from the abnormal analysis indication;

[0106] A plurality of pending abnormal root causes of the abnormal load state are determined from a preset hypergraph where the abnormal phenomenon is located.

[0107] In an embodiment of the present invention, illustratively, the server receives an alarm signal about an abnormal load in "Nth Substation in YY District, XX City", which is an abnormal analysis indication. The indication may come from a smart meter, sensor or other monitoring system in the substation. The server analyzes and extracts specific abnormal phenomena from the abnormal analysis indication. For example, the indication may include information such as "The load in Nth Substation in YY District, XX City has continued to rise in the past 30 minutes, and the voltage fluctuation exceeds the normal range". The server identifies these specific manifestations as abnormal phenomena, namely, abnormal load rise and abnormal voltage fluctuation. The server will then consult a preset hypergraph, which is constructed based on historical data and expert knowledge, and contains a complex relationship between various possible abnormal phenomena and potential abnormal root causes. In this hypergraph, nodes may represent different devices, environmental factors or user behaviors, and hyperedges represent the association or causal relationship between these nodes. According to the abnormal phenomena (abnormal load rise and abnormal voltage fluctuation) extracted from the abnormal analysis indication, the server searches for nodes associated with these phenomena in the preset hypergraph. For example, factors that may cause load increase and voltage fluctuation include: internal equipment failure in the substation, abnormal external power supply, sudden changes in user power consumption behavior, etc. The server lists these factors as pending abnormal root causes because they can all be the cause of the current abnormal phenomenon. Through the above steps, the server can quickly identify abnormal phenomena of abnormal load status based on abnormal analysis indications and preset hypergraphs, and determine multiple possible pending abnormal root causes. This provides important clues and directions for subsequent troubleshooting and problem solving.

[0108] In the embodiment of the present invention, a plurality of pending abnormal root causes of the abnormal load state are determined from the preset hypergraph where the abnormal phenomenon is located, which can be implemented through the following examples.

[0109] Determine the substation system to which the abnormal load state belongs, and obtain a preset hypergraph of the substation system; each entity of the preset hypergraph includes an abnormal phenomenon entity where the abnormal phenomenon is located, and potential abnormal root cause entities other than the abnormal phenomenon entity;

[0110] Determine a plurality of key abnormal root cause entities from each of the potential abnormal root cause entities according to the entity relationship corresponding to each entity in the preset hypergraph;

[0111] The abnormal factor represented by the key abnormal root cause entity is determined as the pending abnormal root cause of the abnormal load state.

[0112] In an embodiment of the present invention, illustratively, first, the server needs to determine the specific substation system where the abnormal load state occurs. For example, suppose that the server detects that the "Mth Substation in ZZ District, XX City" has an abnormal load state. Subsequently, the server will obtain a preset hypergraph corresponding to the substation system from its database. This preset hypergraph is a complex relationship network, which contains various possible abnormal phenomenon entities (such as voltage abnormality, excessive load, etc.) and potential abnormal root cause entities associated with it (such as equipment failure, line aging, changes in user electricity consumption behavior, etc.). The server begins to analyze the relationship between the entities in the preset hypergraph. These relationships can be causal relationships, correlation relationships, or conditional relationships. For example, in the hypergraph, the voltage abnormality entity may be connected to multiple potential abnormal root cause entities such as the transformer failure entity and the line poor contact entity. By analyzing these relationships, the server can initially understand which potential abnormal root causes may be related to the current abnormal phenomenon. Next, the server will screen out multiple key abnormal root cause entities from a large number of potential abnormal root cause entities based on the entity relationships in the preset hypergraph, historical data, expert knowledge, etc. These key abnormal root cause entities are causes with a high degree of correlation and a high probability with the current abnormal phenomenon. For example, when analyzing voltage anomalies, the server found that transformer failure and poor line contact were two key abnormal root cause entities because they had caused similar voltage anomalies many times in the past. Finally, the server identified the abnormal factors represented by the screened key abnormal root cause entities as pending abnormal root causes. These pending abnormal root causes will be used as targets for further investigation and verification. In this example, "transformer failure" and "poor line contact" were identified as pending abnormal root causes of the voltage anomaly in "M Substation, ZZ District, XX City". Through the above steps, the server can quickly and accurately determine multiple pending abnormal root causes related to abnormal load conditions based on the preset hypergraph and entity relationship analysis, providing strong support for subsequent problem solving.

[0113] In the embodiment of the present invention, based on each of the undetermined abnormal root causes and the characteristic parameters corresponding to the abnormal phenomenon, a sample instance array is constructed, which can be implemented through the following examples.

[0114] Standardizing the characteristic parameters corresponding to each of the undetermined abnormal root causes and the abnormal phenomena to obtain standardized data;

[0115] Performing feature extraction operations on the standardized data to obtain abnormal factor features that meet the expert model training requirements;

[0116] Construct an array of sample instances containing the characteristics of the abnormal factors.

[0117] In an embodiment of the present invention, illustratively, the server first collects characteristic parameters related to each pending abnormal root cause and abnormal phenomenon. For example, for the pending abnormal root cause of "equipment aging", the relevant characteristic parameters may include equipment operation time, number of maintenance records, equipment temperature, etc.; and for the abnormal phenomenon of "voltage abnormality", the characteristic parameters may include voltage fluctuation range, abnormal duration, etc. Next, the server will standardize these original characteristic parameters to eliminate the dimensional differences and numerical range differences between different characteristic parameters. The standardization method may include minimum-maximum standardization, Z-score standardization, etc. Through standardization, it can be ensured that all characteristic parameters are within the same numerical range, which is convenient for subsequent model training. After the standardization is completed, the server will perform feature extraction operations on the standardized data. The purpose of feature extraction is to extract information useful for model training while reducing the dimension and complexity of the data. For example, for the two characteristic parameters of equipment operation time and number of maintenance records, the server will extract more representative features by calculating their statistics (such as mean, standard deviation, etc.) or applying principal component analysis (PCA) and other methods. In this process, the server will also adjust the feature extraction strategy based on expert knowledge and historical data to ensure that the extracted abnormal factor features can meet the requirements of expert model training. Finally, the server will use the extracted abnormal factor features to construct a sample instance array. Each sample instance contains characteristic parameter values ​​related to a pending abnormal root cause and abnormal phenomenon. For example, a sample instance can contain characteristic parameter values ​​related to equipment aging (such as standardized operating hours, maintenance record statistics, etc.) and characteristic parameter values ​​related to voltage abnormalities (such as standardized voltage fluctuation range, etc.). These sample instance arrays will be used for subsequent expert model training to help the model learn how to identify abnormal root causes and predict abnormal phenomena based on characteristic parameters. Through the above steps, the server can construct a sample instance array that meets the requirements of expert model training based on the characteristic parameters corresponding to each pending abnormal root cause and abnormal phenomenon. This provides an important data basis for subsequent model training and abnormal root cause identification.

[0118] In an embodiment of the present invention, iterative expert model training is performed based on the sample instance array to determine the target abnormal root cause determined after multiple cycles in each of the pending abnormal root causes, which can be implemented through the following examples.

[0119] Performing expert model training on the sample instance array to obtain a target expert model;

[0120] Based on the target expert model, the abnormal phenomenon is analyzed for its root cause, and the correlation coefficient between each of the undetermined abnormal root causes and the abnormal phenomenon is determined respectively;

[0121] From each of the undetermined abnormal root causes, determine a key abnormal root cause whose correlation coefficient meets a preset correlation coefficient threshold;

[0122] Based on the key abnormal root cause, the characteristic parameter distribution of the sample instance is updated, an adjusted sample instance array is determined, and the step of training the sample instance array with an expert model is repeatedly performed until a preset termination state is met to obtain the target abnormal root cause; the adjusted sample instance array includes the key abnormal root cause and the characteristic parameters corresponding to the abnormal phenomenon.

[0123] In the embodiment of the present invention, the server first uses the initial sample instance array to train the expert model. This array contains various characteristic parameters of the pending abnormal root causes and abnormal phenomena. For example, the sample instance array may contain characteristic parameters such as device temperature, grid voltage fluctuation, and changes in user power consumption behavior. Through these data, the server trains an initial expert model, which can be called a "target expert model". Once the target expert model training is completed, the server will use this model to perform abnormal root cause analysis on the abnormal phenomenon. The model will analyze the correlation coefficient between each pending abnormal root cause and the abnormal phenomenon. For example, the model finds that the correlation coefficient between the device temperature being too high and the voltage abnormality is high, while the correlation coefficient between the user's power consumption behavior change and the voltage abnormality is low. The server will then determine from each pending abnormal root cause the key abnormal root cause whose correlation coefficient meets the preset correlation coefficient threshold. In this example, if the correlation coefficient between the device temperature being too high and the voltage abnormality exceeds the preset threshold, then "device temperature being too high" will be determined as a key abnormal root cause. After determining the key abnormal root causes, the server will update the characteristic parameter distribution of the sample instance based on these key abnormal root causes. This means that in the next round of expert model training, the server will pay more attention to the characteristic parameters related to the key abnormal root cause. For example, if "device temperature is too high" is determined as the key abnormal root cause, then in the next round of sample instance array, the characteristic parameters related to device temperature will be given a higher weight. Subsequently, the server will repeat the steps of expert model training using the updated sample instance array. This process will continue to iterate until it meets the preset termination state, such as reaching the preset number of iterations, the model performance improvement is no longer significant, etc. After multiple rounds of iterations, the server will finally determine the target abnormal root cause with the highest correlation with the abnormal phenomenon. In this example, if "device temperature is too high" is always identified as the key abnormal root cause by the model after multiple iterations, and its correlation coefficient with voltage abnormality continues to remain at a high level, then the server will eventually determine "device temperature is too high" as the target abnormal root cause. Through this iterative expert model training method, the server can gradually focus on the key factors that actually cause abnormal phenomena, thereby improving the accuracy and efficiency of abnormal root cause identification.

[0124] In the embodiment of the present invention, expert model training is performed on the sample instance array to obtain a target expert model, which can be implemented through the following examples.

[0125] Determine a feature parameter attribute corresponding to each feature parameter in the sample instance array;

[0126] The core training framework associated with each of the feature parameter attributes is used to perform expert model training on the sample instance array to obtain a target expert model.

[0127] In the embodiment of the present invention, the server first analyzes each feature parameter in the sample instance array to determine their attributes. For example, for the feature parameter "device temperature", its attributes may include "continuous value", "real-time change", etc.; and for the feature parameter "device model", its attributes may include "discrete classification", "static information", etc. According to the attributes of each feature parameter, the server selects the core training architecture associated with it. For example, for continuous numerical features, the server selects a neural network architecture suitable for regression problems, such as a deep neural network (DNN) or a long short-term memory network (LSTM), so as to capture the complex nonlinear relationship between feature parameters. For discrete classification features, the server selects an architecture suitable for classification problems, such as a support vector machine (SVM) or a decision tree. Using the selected core training architecture, the server begins to train the expert model for the sample instance array. Taking device temperature and device model as examples, the server uses the training architecture of continuous numerical features to train a model that can predict the impact of device temperature on power grid load, and uses the training architecture of discrete classification features to train a model that can predict the probability of failure based on the device model. After multiple rounds of training and optimization, the server will eventually obtain one or more target expert models. These models can perform efficient prediction and analysis based on specific feature parameter attributes. For example, for continuously monitored equipment temperature data, the target expert model can predict its impact on the grid load in real time, so as to adjust the scheduling strategy of the energy storage system in a timely manner; for static information such as equipment model, the target expert model can evaluate the failure risk of equipment of different models and provide decision support for equipment maintenance and replacement. Through the above steps, the server can select the appropriate training architecture for expert model training for different types of feature parameters, so as to obtain a more accurate and efficient target expert model. These models play an important role in the autonomous distributed energy storage intelligent scheduling of the substation, and help improve the stability and operation efficiency of the system.

[0128] In the embodiment of the present invention, expert model training is performed on the sample instance array to obtain a target expert model, which can be implemented through the following examples.

[0129] Dividing the sample instance array into a plurality of sub-sample instance arrays; each of the sub-sample instance arrays includes characteristic parameters of the abnormal phenomenon and at least a portion of characteristic parameters of the undetermined abnormal root cause;

[0130] Expert model training is performed on each of the sub-sample instance arrays to obtain a target expert model associated with each of the sub-sample instance arrays.

[0131] In an embodiment of the present invention, illustratively, the server first divides the entire sample instance array into multiple sub-sample instance arrays. The purpose of doing so is to disperse the training pressure and improve the training efficiency, and at the same time, it can also perform more refined model training for different types of abnormal root causes. For example, the server can divide the characteristic parameters related to equipment failure into a sub-sample instance array, and divide the characteristic parameters related to external environmental changes (such as weather, temperature) into another sub-sample instance array. Each sub-sample instance array will contain characteristic parameters of abnormal phenomena and at least a portion of characteristic parameters of undetermined abnormal root causes. Taking equipment failure as an example, the relevant sub-sample instance array may include characteristic parameters such as equipment operating status, fault history records, equipment usage time, and also include characteristic parameters related to abnormal phenomena (such as voltage fluctuations, load anomalies, etc.) that may be caused by equipment failure. The server will perform expert model training on each sub-sample instance array separately. This means that for different types of abnormal root causes, the server will train different target expert models. For example, for an abnormal root cause such as equipment failure, the server will train an expert model that can identify the precursors of equipment failure and predict the impact of the failure; and for an abnormal root cause such as external environmental changes, the server will train another expert model that can analyze the relationship between external environmental factors and the operating status of the power grid. After training, the server will obtain the target expert model associated with each sub-sample instance array. These models are optimized and trained for different types of abnormal root causes, so they have higher accuracy and efficiency when dealing with the corresponding types of abnormalities. For example, when the server detects voltage fluctuations in the power grid, it can use the target expert model related to equipment failure to quickly determine whether the abnormality is caused by equipment failure and take corresponding countermeasures. Through this divide-and-conquer strategy, the server can more effectively use limited computing resources for model training and provide more accurate prediction and analysis capabilities for different types of abnormal root causes.

[0132] In the embodiment of the present invention, the abnormal root cause analysis is performed on the abnormal phenomenon based on the target expert model, and the correlation coefficient between each of the undetermined abnormal root causes and the abnormal phenomenon is determined respectively, which can be implemented through the following examples.

[0133] Based on the target expert model, the abnormal phenomenon is analyzed for its root cause, and the initial correlation coefficient between each of the undetermined abnormal root causes and the abnormal phenomenon is determined respectively;

[0134] For each of the undetermined abnormal root causes, feature masking is performed on the feature parameters of the undetermined abnormal root cause to obtain masked data;

[0135] Determine a model stability evaluation value of the undetermined abnormal root cause for the target expert model based on a characteristic parameter of the abnormal phenomenon and a deviation between an output response of the target expert model to the masked data;

[0136] The correlation coefficient between the undetermined abnormal root cause and the abnormal phenomenon is determined by combining the initial correlation coefficient and the model stability evaluation value.

[0137] In an embodiment of the present invention, illustratively, the server will first use the trained target expert model to perform root cause analysis on the abnormal phenomenon that has occurred. For example, when an abnormal phenomenon of voltage drop occurs in the power grid, the server will preliminarily determine the initial correlation coefficients between each pending abnormal root cause (such as equipment failure, line aging, external interference, etc.) and this abnormal phenomenon through model analysis. These initial correlation coefficients reflect the possibility that each pending abnormal root cause causes the abnormal phenomenon under the model training data. Next, the server will mask the characteristic parameters of each pending abnormal root cause. This is equivalent to "shielding" the influence of the abnormal root cause in the model to observe the performance of the model without this factor. For example, if "equipment failure" is a pending abnormal root cause, the server will mask its related characteristic parameters (such as equipment operating status, historical fault records, etc.) to generate masked data. The server will then input the masked data into the target expert model and observe the output response of the model. By comparing the deviation between the characteristic parameters of the abnormal phenomenon (such as the amplitude and duration of the voltage sag) and the output response of the model to the masked data, the impact of the undetermined abnormal root cause on the model stability can be evaluated. The smaller the deviation value, the smaller the impact of the abnormal root cause on the model output, and vice versa. This deviation value is defined as the model stability evaluation value. Finally, the server will combine the initial correlation coefficient and the model stability evaluation value to comprehensively determine the final correlation coefficient between each undetermined abnormal root cause and the abnormal phenomenon. For example, if the initial correlation coefficient of the abnormal root cause "equipment failure" is high, but after masking its characteristic parameters, the deviation between the model output and the abnormal phenomenon (i.e., the model stability evaluation value) is very small, then this means that although the equipment failure has a certain correlation with the voltage sag, it is not the main cause of this abnormality. Therefore, under comprehensive consideration, the server will adjust the correlation coefficient between "equipment failure" and the voltage sag abnormal phenomenon. Through the above steps, the server can more accurately identify the key factors that cause abnormal phenomena, providing strong support for subsequent troubleshooting and repair.

[0138] In the embodiment of the present invention, the abnormal root cause analysis is performed on the abnormal phenomenon based on the target expert model, and the initial correlation coefficients between each of the pending abnormal root causes and the abnormal phenomenon are determined respectively, which can be implemented through the following examples.

[0139] Taking the characteristic parameters corresponding to each of the undetermined abnormal root causes as the input of the target expert model, taking the characteristic parameters of the abnormal phenomenon as the output of the target expert model, and calculating the expected impact degree of the input of the target expert model through back propagation; the expected impact degree of the input is represented by the expected superimposed impact degree corresponding to each of the undetermined abnormal root causes;

[0140] Based on each of the expected superposition influence degrees, initial correlation coefficients between each of the undetermined abnormal root causes and the abnormal phenomenon are determined respectively; and the initial correlation coefficients match the expected superposition influence degrees of the undetermined abnormal root causes.

[0141] In the embodiment of the present invention, the server uses the characteristic parameters corresponding to each undetermined abnormal root cause as the input of the target expert model, for example, equipment temperature, grid voltage fluctuation, line aging degree, etc. At the same time, the characteristic parameters of the abnormal phenomenon are used as the output of the model, such as voltage anomaly or load anomaly in the grid. In order to determine the degree of influence of the input (undetermined abnormal root cause) on the output (abnormal phenomenon), the server will use the back propagation algorithm. This method is usually used in the training process of neural networks, but here it is used to analyze the expected influence of the input on the output in the trained model. Specifically, the server will start from the output layer of the model and reversely calculate the contribution of each input characteristic parameter to the output, which is called the expected influence of the input. This influence is characterized by the expected superposition influence corresponding to each undetermined abnormal root cause. In the back propagation process, the server will calculate the expected superposition influence of each undetermined abnormal root cause on the output. For example, for the abnormal root cause of "equipment temperature is too high", the server will analyze its influence on the voltage abnormal phenomenon and calculate a specific expected superposition influence value. This value reflects the potential influence of "equipment temperature is too high" on the voltage abnormal phenomenon in the model prediction. Based on the expected superposition impact of each undetermined abnormal root cause, the server will determine their initial correlation coefficients with the abnormal phenomenon. This correlation coefficient matches the expected superposition impact, that is, the greater the impact of the abnormal root cause, the higher its initial correlation coefficient will be. For example, if the expected superposition impact of "equipment temperature is too high" is large, then its initial correlation coefficient with the voltage anomaly will also be high, indicating a strong correlation between the two. Through the above steps, the server can quantify the initial correlation coefficient between each undetermined abnormal root cause and the abnormal phenomenon, providing an important reference for subsequent abnormal root cause analysis and fault handling.

[0142] In the embodiment of the present invention, determining the key abnormal root cause whose correlation coefficient meets the preset correlation coefficient threshold from each of the pending abnormal root causes can be implemented through the following examples.

[0143] determining a preset correlation coefficient threshold associated with a current training cycle;

[0144] The correlation coefficient corresponding to each of the pending abnormal root causes is compared with the preset correlation coefficient threshold, and the key abnormal root cause whose correlation coefficient meets the preset correlation coefficient threshold is determined from each of the pending abnormal root causes.

[0145] In an embodiment of the present invention, exemplarily, at the beginning of each training cycle, the server sets a preset correlation coefficient threshold associated with the current training cycle based on historical data and experience. This threshold is used to determine which pending abnormal root causes have a strong enough correlation with the abnormal phenomenon to be regarded as key abnormal root causes. For example, the server can set the preset correlation coefficient threshold to 0.8, which means that only when the correlation coefficient between the pending abnormal root cause and the abnormal phenomenon reaches or exceeds 0.8, it will be identified as a key abnormal root cause. The server calculates the correlation coefficient between each pending abnormal root cause and the abnormal phenomenon, and compares these correlation coefficients with the preset correlation coefficient threshold. For example, the server calculates that the correlation coefficient between the pending abnormal root cause "equipment aging" and the "grid load abnormality" phenomenon is 0.9, while the correlation coefficient between "external environment temperature change" and the same abnormal phenomenon is 0.7. Based on the comparison results, the server determines from the pending abnormal root causes those key abnormal root causes whose correlation coefficients meet (i.e., are greater than or equal to) the preset correlation coefficient threshold. In the above example, the correlation coefficient of "equipment aging" of 0.9 exceeds the preset correlation coefficient threshold of 0.8, so it will be determined as the key abnormal root cause of "grid load abnormality". The correlation coefficient of "external ambient temperature change" of 0.7 is lower than the threshold, so it will not be considered as a key abnormal root cause. Through the above steps, the server can accurately identify the key abnormal root causes that are most closely related to the abnormal phenomenon from multiple pending abnormal root causes, which is crucial for quickly locating and solving actual problems in grid operation.

[0146] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the above-mentioned distributed energy storage intelligent scheduling method based on substation autonomy. Figure 2 As shown, Figure 2 The block diagram of the computer device 100 provided in the embodiment of the present invention. The computer device 100 includes a memory 111, a processor 112 and a communication unit 113. To achieve data transmission or interaction, the memory 111, the processor 112 and the communication unit 113 are directly or indirectly electrically connected to each other. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.

[0147] For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or to limit the present disclosure to the precise form disclosed. Numerous modifications and variations are possible in accordance with the above teachings. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to suit the intended specific application.

Claims

1. A distributed energy storage intelligent dispatching method based on area autonomy, characterized in that: include: Obtain the task instructions for the current power to enter the target area; According to the task instruction, acquiring the power load data of the current power; Inputting the current power load data into a pre-trained load assessment model to obtain a corresponding load state assessment result; When the load state evaluation result is characterized as an abnormal load state, determining a target abnormal root cause with the highest correlation with the abnormal phenomenon of the abnormal load state, and obtaining an abnormal state analysis result; Based on the load state evaluation result and the abnormal state analysis result, selecting a target scheduling strategy from pre-configured energy storage scheduling; The target scheduling strategy is sent to the distributed energy storage devices included in the target area for execution.

2. The method according to claim 1, characterized in that The load assessment model of the substation area is obtained by the following methods, including: Obtaining power load data recorded during the process of target power being transmitted to the target substation, and associating the power load data with transformers to obtain sub-power load data corresponding to a plurality of transformers; the sub-power load data includes instantaneous load data corresponding to each sampling time node during the process of target power being transmitted to the corresponding transformer; Based on each of the sub-power load data, determining a load state corresponding to the corresponding transformer device, wherein the load state includes a safe load state and an abnormal load state; Taking each of the sub-power load data as a sample instance and the corresponding load state as an instance target value, a sample instance array for training the substation load assessment model is constructed, and the following processing is performed for each of the sub-power load data in the sample instance array: Obtaining the load mean value, root mean square load value, load fluctuation, load dispersion of each instantaneous load data, and peak value data in the instantaneous load data; Dividing the load dispersion by the load mean value to obtain the rate of change deviation of the sub-power load data, dividing the peak data by the root mean square load value to obtain the peak-to-average ratio of the sub-power load data, and dividing the peak data by the load mean value to obtain the impact ratio of the sub-power load data; Performing data integration processing on the load mean, the load fluctuation, the change rate deviation, the peak-to-average ratio and the impact ratio to obtain load time series conversion domain data corresponding to the load time series analysis domain; Obtaining the spectrum distribution intensity and the power spectrum density mean corresponding to the sub-power load data, and determining the spectrum center frequency of the sub-power load data on the energy spectrum based on the spectrum distribution intensity; Determine the spectrum fluctuation of the sub-power load data on the energy spectrum by combining the spectrum center frequency and the spectrum distribution intensity, and obtain the spectrum upper limit frequency in the spectrum distribution intensity; Performing data integration processing on the spectrum distribution intensity, the power spectrum density mean, the spectrum center frequency, the spectrum fluctuation and the spectrum upper limit frequency to obtain energy spectrum conversion domain data corresponding to the energy spectrum analysis domain; From the power quality fluctuation analysis domain, performing power quality fluctuation mapping processing on the sub-power load data to obtain power quality fluctuation conversion domain data corresponding to the power quality fluctuation analysis domain; Performing feature extraction operations on each of the conversion domain data to obtain conversion domain features corresponding to each of the conversion domain data; Acquire the feature domain attribute corresponding to each of the conversion domain features, and perform weighted average processing on each of the feature domain attributes to obtain a reference domain attribute; Compare each of the feature domain attributes with the reference domain attributes to obtain an attribute comparison result corresponding to each of the feature domain attributes; When the attribute comparison result indicates that the feature domain attribute is consistent with the reference domain attribute, the corresponding conversion domain feature is determined as a target conversion domain feature corresponding to the conversion domain feature; When the attribute comparison result indicates that the feature domain attribute is inconsistent with the reference domain attribute, the feature domain attribute of the corresponding conversion domain feature is updated to the reference domain attribute to obtain a target conversion domain feature corresponding to the conversion domain feature; Performing feature integration processing on each of the target conversion domain features to obtain integrated features; Using the substation load assessment model, based on the integrated features, the substation load assessment is performed on the target substation to obtain an inferred load state of the target substation; In combination with the inferred load state and the corresponding load state, executing a training process for the substation load assessment model; The substation load assessment model is used to assess the load status of the substation where the target power is located based on the power load data of the target power.

3. The method according to claim 2, characterized in that The obtaining of the power load data recorded during the process of transmitting the target power to the target area includes: Obtaining the voltage fluctuation and current change corresponding to each sampling time node during the process of transmitting the target power to the target substation; For each of the sampling time nodes, the corresponding normalized value of the voltage fluctuation and the normalized value of the current change are integrated to obtain the instantaneous load data corresponding to the sampling time node; Each instantaneous load data is arranged according to the time sequence of the sampling time nodes to obtain the power load data.

4. The method according to claim 2, characterized in that: The sub-power load data includes instantaneous load data corresponding to each sampling time node in the process of the target power being transmitted to the corresponding transformer device. The determining of the load state corresponding to the corresponding transformer device based on each sub-power load data includes: The following processing is performed for each of the sub-power load data: For each sampling time node corresponding to the sub-power load data, at least one adjacent sampling time node of the sampling time node is obtained, and the instantaneous load data corresponding to the sampling time node is respectively subjected to difference calculation with the instantaneous load data corresponding to each adjacent sampling time node, so as to obtain the load fluctuation change amount corresponding to each adjacent sampling time node; When at least one of the load fluctuation changes exceeds a preset change threshold, determining the target load safety level as a first load safety level; When there is no load fluctuation variation exceeding the preset variation threshold, determining the target load safety level as the second load safety level; The first load safety level is used to indicate that the load of the area through which the target power passes at the sampling time node is within the abnormal load range, and the second load safety level is used to indicate that the load of the area through which the target power passes at the sampling time node is within the safe load range; When the target load safety level indicates that the area load of the target area through which the target power passes at the sampling time node is within a safe load range, the sampling time node corresponding to the target load safety level is determined as the target sampling time node; Based on the number of nodes at the target sampling time node, a load state corresponding to the corresponding transformer device is determined.

5. The method according to claim 4, characterized in that The determining, based on the number of nodes at the target sampling time node, a load state corresponding to the corresponding transformer device includes: Compare the number of nodes at the target sampling time node with the node number threshold to obtain a node number comparison result; When the node number comparison result indicates that the node number of the target sampling time node exceeds the node number threshold, determining the load state as the safe load state; When the node number comparison result indicates that the node number of the target sampling time node does not exceed the node number threshold, determining the load state as the abnormal load state; The determining of the load state corresponding to the corresponding transformer device based on the number of nodes at the target sampling time node further includes: Dividing the number of nodes at the target sampling time node by the number of nodes at the sampling time node to obtain the load stability rate of the target substation; When the load stability rate exceeds a load stability rate threshold, the load state is determined as the safe load state, and when the load stability rate does not exceed the load stability rate threshold, the load state is determined as the abnormal load state.

6. The method according to claim 2, characterized in that After performing a training process on the substation load assessment model in combination with the inferred load state and the corresponding load state, the method further includes: Obtaining a power dispatching plan for the target power, and searching a load state of a reference substation corresponding to the power dispatching plan from a substation load database to obtain a search result; When the search result indicates that the load state of the reference area is not stored in the area load database, obtaining target load data of the reference area in a similar time period, performing a feature extraction operation on the target load data, and obtaining a target load feature; Using the trained area load assessment model, based on the target load characteristics, the area load assessment is performed on the reference area to obtain an inferred load state of the reference area; The inferred load state of the reference substation is updated into the substation load database to obtain an updated substation load database.

7. The method according to claim 6, characterized in that The method further comprises: When the search result indicates that the load status of the reference station is stored in the station load database, and the load status of the reference station indicates that the station load of the reference station is within an abnormal load range, generating a first notification message; When the search result indicates that the load status of the reference area is stored in the area load database, and the load status of the reference area indicates that the area load of the reference area is within a safe load range, generating a second notification message; The first notification message is used to notify the adjustment of the power dispatching plan on the substation load database, and the second notification message is used to notify that the power load of the reference substation corresponding to the power dispatching plan is normal.

8. The method according to claim 1, characterized in that: The determining of the target abnormal root cause with the highest correlation with the abnormal phenomenon of the abnormal load state to obtain the abnormal state analysis result includes: Obtaining an abnormal analysis indication for an abnormal load state, and extracting an abnormal phenomenon of the abnormal load state from the abnormal analysis indication; Determine the substation system to which the abnormal load state belongs, and obtain a preset hypergraph of the substation system; each entity of the preset hypergraph includes an abnormal phenomenon entity where the abnormal phenomenon is located, and potential abnormal root cause entities other than the abnormal phenomenon entity; Determine a plurality of key abnormal root cause entities from each of the potential abnormal root cause entities according to the entity relationship corresponding to each entity in the preset hypergraph; Determine the abnormal factor represented by the key abnormal root cause entity as the pending abnormal root cause of the abnormal load state; Standardizing the characteristic parameters corresponding to each of the undetermined abnormal root causes and the abnormal phenomena to obtain standardized data; Performing feature extraction operations on the standardized data to obtain abnormal factor features that meet the expert model training requirements; Constructing a sample instance array containing the characteristics of the abnormal factors; Performing expert model training on the sample instance array to obtain a target expert model; Taking the characteristic parameters corresponding to each of the undetermined abnormal root causes as the input of the target expert model, taking the characteristic parameters of the abnormal phenomenon as the output of the target expert model, and calculating the expected impact degree of the input of the target expert model through back propagation; the expected impact degree of the input is represented by the expected superimposed impact degree corresponding to each of the undetermined abnormal root causes; Based on each of the expected superposition influence degrees, respectively determine the initial correlation coefficients between each of the undetermined abnormal root causes and the abnormal phenomenon; the initial correlation coefficients match the expected superposition influence degrees of the undetermined abnormal root causes; For each of the undetermined abnormal root causes, feature masking is performed on the feature parameters of the undetermined abnormal root cause to obtain masked data; Determine a model stability evaluation value of the undetermined abnormal root cause for the target expert model based on a characteristic parameter of the abnormal phenomenon and a deviation between an output response of the target expert model to the masked data; Determine the correlation coefficient between the undetermined abnormal root cause and the abnormal phenomenon by combining the initial correlation coefficient and the model stability evaluation value; determining a preset correlation coefficient threshold associated with a current training cycle; Comparing the correlation coefficient corresponding to each of the pending abnormal root causes with the preset correlation coefficient threshold, and determining the key abnormal root cause whose correlation coefficient meets the preset correlation coefficient threshold from each of the pending abnormal root causes; Based on the key abnormal root cause, the characteristic parameter distribution of the sample instance is updated, an adjusted sample instance array is determined, and the step of performing expert model training on the sample instance array is repeatedly performed until a preset termination state is met to obtain a target abnormal root cause; the adjusted sample instance array includes the key abnormal root cause and the characteristic parameters corresponding to the abnormal phenomenon; wherein the sample instance of the latter round of expert model training includes the characteristic parameters of the abnormal phenomenon and the characteristic parameters of the key abnormal root cause determined by the previous round of expert model training, and the characteristic parameters of the key abnormal root cause determined by the previous round of expert model training are different from the characteristic parameter distribution of the characteristic parameters of the abnormal root cause in the sample instance of the previous round of expert model training; the correlation coefficient between the key abnormal root cause and the abnormal phenomenon meets the preset correlation coefficient threshold; The correlation and matching relationship between the abnormal phenomenon and the target abnormal root cause is used as the abnormal state analysis result of the abnormal load state.

9. The method according to claim 8, characterized in that The performing expert model training on the sample instance array to obtain a target expert model includes: Determine a feature parameter attribute corresponding to each feature parameter in the sample instance array; Using a core training framework associated with each of the feature parameter attributes, perform expert model training on the sample instance array to obtain a target expert model; The performing expert model training on the sample instance array to obtain a target expert model also includes: Dividing the sample instance array into a plurality of sub-sample instance arrays; each of the sub-sample instance arrays includes characteristic parameters of the abnormal phenomenon and at least a portion of characteristic parameters of the undetermined abnormal root cause; Expert model training is performed on each of the sub-sample instance arrays to obtain a target expert model corresponding to each of the sub-sample instance arrays.

10. A server system, characterized in that: The method comprises a server, wherein the server is used to execute the method described in any one of claims 1 to 9.

Images