A lightweight artificial intelligence algorithm arrangement construction method for power distribution areas

Abstract

The application provides a lightweight artificial intelligence algorithm arrangement construction method for power distribution areas, including the following steps: step S1, analyzing and extracting typical area scenes, summarizing the collaborative management and regional autonomy application requirements of the scenes, and establishing a mapping relationship between the scenes and the application requirements; step S2, for different application requirement scenes, decomposing the artificial intelligence algorithm into tool algorithms that are decoupled and independent of each other for various functions of the scene, and then collating and summarizing the tool algorithms into a tool algorithm library; step S3, according to the scene characteristic set of a specific area, identifying the scene type and outputting an application requirement label, selecting a plurality of required tool algorithms, and making different tool algorithms operate in series to form a final lightweight artificial intelligence algorithm, which is deployed to the edge device of the power distribution area. The application can conveniently realize the artificial intelligence algorithm arrangement technology, reduce the development threshold of power big data application, improve the collaborative management and regional autonomy of the area, and help realize the new power system.

Description

Technical Field

[0001] The present invention relates to the field of smart distribution network technology, and in particular to a lightweight artificial intelligence algorithm arrangement and construction method for distribution network areas. Background Art

[0002] At present, the distribution network is transforming from an uncontrollable traditional distribution network to a partially controllable modern distribution network and a fully controllable future distribution network. The development goal of the distribution network is also gradually evolving from full geographical coverage of user loads, to reliable, efficient, and high-quality power supply services, to a green, low-carbon, intelligent, controllable, and supply-demand interactive service platform. In this transformation process, the application of artificial intelligence algorithms and scenario analysis, management, and resource collaborative control based on edge terminals in the substation face many problems. The access of diverse equipment such as distributed photovoltaics and charging piles has led to complex and changeable operating conditions in the substations, which in turn has led to diverse needs in the substations. Traditional collaborative management and control solutions are not universal and have poor universality.

[0003] With the development of computer technology, artificial intelligence (AI) has been widely applied in electrical engineering control. This model can effectively improve equipment efficiency and significantly reduce errors caused by human error. Furthermore, AI can enhance the utilization of big data in power generation, unlocking hidden insights within the data and providing guidance to power grid companies in pricing and demand-side management.

[0004] However, traditional AI algorithms often have a high barrier to entry, requiring specialized personnel for programming and functional testing, which significantly increases the difficulty of applying AI technology in distribution automation. Furthermore, due to their large number of model parameters and demanding computing power requirements, traditional AI algorithms have certain limitations in their deployment on edge devices within distribution substations. Therefore, there is an urgent need to develop a convenient and implementable AI algorithm orchestration technology for distribution substations. This technology can lower the barrier to entry for developing power big data applications, enhance coordinated substation management and control, and enhance regional autonomy, thereby facilitating the implementation of new power systems. Summary of the Invention

[0005] This invention proposes a lightweight artificial intelligence algorithm orchestration construction method for distribution substations, which can easily implement artificial intelligence algorithm orchestration technology to lower the threshold for power big data application development, improve substation collaborative management and regional self-control capabilities, and help realize new power systems.

[0006] The present invention adopts the following technical solutions.

[0007] A lightweight artificial intelligence algorithm arrangement and construction method for distribution substations includes the following steps:

[0008] Step S1: Combine the equipment records and historical operation data of the distribution area to form a scenario feature set of the area, analyze and extract typical area scenarios, further summarize the collaborative management and control and regional autonomy application requirements of typical scenarios, and establish a mapping relationship between scenarios and application requirements;

[0009] Step S2: For different application demand scenarios, the artificial intelligence algorithm is decomposed into tool algorithms that are decoupled and independent of each other according to their functions, and then organized and summarized into a tool algorithm library;

[0010] Step S3: Based on the scene feature set of a specific substation, identify the scene type and output the application requirement label, select several required tool algorithms, standardize the data format between the tool algorithms, ensure the smooth flow of data between different tool algorithms, and use parameters to set input and output parameters to enable different tool algorithms to operate in series to form the final lightweight artificial intelligence algorithm, which is deployed to the edge equipment of the distribution substation.

[0011] In step S1, the equipment ledger and historical operation data of the distribution station area are used to consider the data such as the area scale, the proportion of new energy access, and the energy storage configuration to form the scene feature set of the area. The scene feature sets of all areas are clustered and analyzed to obtain the typical scene types of the area. The application requirements corresponding to the typical scenes are comprehensively analyzed and a mapping relationship between the two is established. Finally, a scene application requirement identification model is constructed that can output the scene application requirement label of the area by inputting the scene feature set of the area.

[0012] The scenario feature set consists of statistical features and curve features. Specifically, statistical features can be divided into numerical data such as the number of users in the area, the proportion of photovoltaic access, energy storage capacity, total installed wind power capacity, daily average load, peak-valley load difference, and total installed photovoltaic capacity; curve features can be divided into sequence data such as distributed photovoltaic output curve, distribution transformer load curve, and charging and discharging curves of energy storage or charging piles.

[0013] For the i-th substation scene, its statistical characteristics and curve characteristics are:

[0014]

[0015]

[0016] Among them, X i is the statistical feature of the i-th station scene, It is the jth statistical feature among the M statistical features of the i-th station scene. i is the curve feature of the i-th station scene. Assuming that the scene has N curve features and the maximum length of the data sequence is L, then F i is a matrix of the form N×L, Indicates the kth data in the jth curve feature sequence in the i-th station scenario. If there are curve features with inconsistent lengths, the shorter data sequence can be padded with 0s starting from the end;

[0017] Based on the above-mentioned substation scene feature set, the big data mining method of PCA dimensionality reduction algorithm and K-means clustering algorithm is used to realize substation scene classification and obtain typical scenes of all substations; application demand analysis is performed on typical substation scenes; then, the analysis results are used as labels and the substation scene feature set is used as input to establish a substation scene application demand identification model based on a decision tree.

[0018] In step S2, the artificial intelligence algorithm is an artificial intelligence algorithm in a traditional solution for distribution substations; the algorithms corresponding to typical scenario types in distribution substations include an outlier detection algorithm, a missing value filling algorithm, a data normalization algorithm, a household-to-transformer relationship identification main program, a topology identification main program, a voltage monitoring main program, a photovoltaic storage and charging coordinated control optimization, an electric vehicle orderly charging program, a new energy output prediction main program, a distribution load prediction main program, and a functional decoupling tool algorithm for a resident electricity consumption behavior analysis main program;

[0019] In step S2, the artificial intelligence algorithms of traditional solutions are divided into four categories according to the functions of different parts: data preprocessing -> main function implementation -> result visualization -> information interaction; among them, the main function implementation is divided into specific functions of coordinated optimization control of photovoltaic storage and charging, and orderly charging of electric vehicles according to the application needs of substation scenarios; based on the decomposed tool algorithms, a tool algorithm library covering the application needs of typical substation scenarios is constructed; and the tool algorithm library has the functions of re-modification and expansion, which can expand the application breadth of the tool algorithm library according to actual needs.

[0020] The step S3 includes: according to a specific substation scenario, calculating the statistical feature data and curve feature data in the scenario, then inputting the feature set into the substation scenario application demand identification model constructed in step S1, obtaining the application demand label of the current substation scenario to clarify the application demand, and then selecting a suitable tool algorithm from the tool algorithm library, adjusting the data format, and connecting them in series to form an overall lightweight artificial intelligence algorithm, which is deployed to the substation edge device to realize the convenient application of the lightweight artificial intelligence algorithm;

[0021] The arrangement and construction of the lightweight artificial intelligence algorithm is expressed in the following formula:

[0022]

[0023] Where Model is a lightweight artificial intelligence algorithm model used to solve the application needs of the substation scene; P i is the i-th tool algorithm in the tool algorithm library; Ii is the input of tool algorithm P i ; D pre is the set of data preprocessing tool algorithms; T is the set of scene features of the current district, which is composed of curve features F and statistical features X; a i is the flag bit of whether to apply P i tool algorithm, a i = 0 indicates that no tool algorithm P i is used in the overall model; b i is the flag bit of whether to apply D pre data preprocessing algorithm, b i = 0 indicates that the set of scene features T is directly used as the input I i of the tool algorithm, and b i = 1 indicates that the set of scene features T is preprocessed by D pre and then used as the input I i of the tool algorithm P i .

[0024] In step S1, the PCA dimension reduction algorithm is used to reduce the dimension of the set of scene features of all districts, and then the K-means clustering algorithm is used for unsupervised classification to obtain a number of typical district operation scenes. For each typical district scene, the application requirements in the scene are comprehensively analyzed, and then the decision tree algorithm is used to establish the mapping relationship between the set of scene features and the scene application requirements, and a district scene application requirement identification model is constructed.

[0025] In step S1, the specific method of the K-means clustering algorithm for unsupervised classification is as follows:

[0026] Suppose the total sample of the input feature set is S = {x1, x2, x3,..., x m}, where x i represents the feature value of the i-th district scene sample, the number of clusters is selected as k, the cluster center vector is initialized, and k cluster centers are generated, which are {μ1, μ2,..., μ k}. For each district scene sample x i , the class with the closest class center is marked, that is:

[0027] label i = argmin||x i - μ j || Formula Four;

[0028] Each cluster center vector is updated to the mean of all samples belonging to the class according to the following formula, that is:

[0029]

[0030] Repeat the above two steps until the change value of the cluster center is less than a certain threshold to obtain the final substation scene clustering result. Points with similar distances can be regarded as scenes of the same type. Based on the substation clustering result labels, calculate the mean of the scene feature set under each category and use it as the feature set of the typical substation scene.

[0031] In step S1, the PCA dimensionality reduction algorithm is used to reduce the dimensionality of the massive curve feature data of the station scene feature set, that is, the data projection technology is used to reduce the dimensionality of the massive curve feature data. The objective function is expressed as follows:

[0032]

[0033] where X={x1,x2,x3,...,x L}, represents a data sequence X of length L, X T is the transposed vector of X, and w1 is the principal axis direction of the projection.

[0034] When the distribution substation is a high-PV penetration substation, the substation scenario feature set is input into the substation scenario application demand identification model to obtain application demand labels including photovoltaic output forecasting, photovoltaic storage coordinated control, and distribution transformer load forecasting. Then, in step S3, based on the application demand label, the appropriate tool algorithm is selected to construct an overall lightweight artificial intelligence algorithm.

[0035] In step S1, the application requirements of typical scenarios in the substation area are analyzed, and the analysis results are used as labels. A decision tree algorithm is used to construct a mapping relationship between feature sets and application requirements, and a supervised classification model is established to achieve the function of inputting the substation scene feature set and outputting the corresponding application requirements of the scene;

[0036] When the decision tree algorithm is executed, the scene feature set is judged layer by layer from top to bottom until the leaf node at the bottom of the decision tree outputs the application requirements corresponding to the scene feature set.

[0037] In step S2, based on the feedback from practical applications, more fine-grained tool algorithms are added to the tool algorithm library. The added more fine-grained tool algorithms include PCA dimensionality reduction, DBSCAN clustering, and logistic regression.

[0038] The input and output data formats of the tool algorithm are adjustable. When the tool algorithm is an outlier detection tool algorithm, the input data size is set to the size of the station scene feature set after dimensionality reduction, and the output data size is set to the input data size of the subsequent missing value filling tool algorithm.

[0039] The present invention can easily implement artificial intelligence algorithm orchestration technology to lower the threshold for power big data application development, improve the coordinated management and control of substations and regional self-control capabilities, assist in the realization of new power systems, and realize the convenient application of artificial intelligence algorithms in edge equipment of distribution substations. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0041] Attachment Figure 1 It is a schematic flow chart of the construction method of the present invention;

[0042] Attachment Figure 2 It is a schematic diagram of the arrangement and construction principle of the tool algorithm library of the present invention. DETAILED DESCRIPTION

[0043] As shown in the figure, a lightweight artificial intelligence algorithm arrangement and construction method for distribution station area includes the following steps:

[0044] Step S1: Combine the equipment records and historical operation data of the distribution area to form a scenario feature set of the area, analyze and extract typical area scenarios, further summarize the collaborative management and control and regional autonomy application requirements of typical scenarios, and establish a mapping relationship between scenarios and application requirements;

[0045] Step S2: For different application demand scenarios, the artificial intelligence algorithm is decomposed into tool algorithms that are decoupled and independent of each other according to their functions, and then organized and summarized into a tool algorithm library;

[0046] Step S3: Based on the scene feature set of a specific substation, identify the scene type and output the application requirement label, select several required tool algorithms, standardize the data format between the tool algorithms, ensure the smooth flow of data between different tool algorithms, and use parameters to set input and output parameters to enable different tool algorithms to operate in series to form the final lightweight artificial intelligence algorithm, which is deployed to the edge equipment of the distribution substation.

[0047] In step S1, the equipment ledger and historical operation data of the distribution station area are used to consider the data such as the area scale, the proportion of new energy access, and the energy storage configuration to form the scene feature set of the area. The scene feature sets of all areas are clustered and analyzed to obtain the typical scene types of the area. The application requirements corresponding to the typical scenes are comprehensively analyzed and a mapping relationship between the two is established. Finally, a scene application requirement identification model is constructed that can output the scene application requirement label of the area by inputting the scene feature set of the area.

[0048] The scene feature set is composed of statistical features and curve features. Specifically, the statistical features can be divided into numerical data of the number of users in the transformer area, the proportion of photovoltaic access, the energy storage capacity, the total installed capacity of wind power, the daily average load, the peak-valley load difference, and the total installed capacity of photovoltaic; and the curve features can be divided into sequence data such as the distributed photovoltaic output curve, the distribution transformer load curve, and the charge-discharge curve of energy storage or charging piles.

[0049] For the i-th transformer area scene, the statistical features and the curve features are respectively:

[0050]

[0051]

[0052] wherein X i is the statistical feature of the i-th transformer area scene, is the j-th of the M statistical features of the i-th transformer area scene. i is the curve feature of the i-th transformer area scene. Assuming that the scene has N curve features, and the maximum length of the data sequence is L, then F i is a matrix of N x L, represents the k-th data in the j-th curve feature sequence of the i-th transformer area scene. If there are curve features of different lengths, the shorter data sequence can be padded with 0 from the end.

[0053] Based on the above transformer area scene feature set, the PCA dimension reduction algorithm and the big data mining method of the K-means clustering algorithm are used to realize the classification of transformer area scenes and obtain the typical scenes of all transformer areas. The application demand analysis is performed for the typical transformer area scenes. Then, the analysis result is taken as a label, and the transformer area scene feature set is taken as an input to establish a transformer area scene application demand recognition model based on a decision tree.

[0054] In the step S2, the artificial intelligence algorithm is an artificial intelligence algorithm in a traditional solution for a power distribution transformer area. The algorithm corresponding to the typical scene type of the power distribution transformer area includes an outlier detection algorithm, a missing value filling algorithm, a data standardization algorithm, a household transformer relationship identification main program, a topology identification main program, a voltage monitoring main program, a photovoltaic storage charging coordination control optimization, an electric vehicle orderly charging program, a new energy output prediction main program, a power distribution load prediction main program, and a function decoupling tool algorithm of a residential electricity consumption behavior analysis main program.

[0055] In step S2, the artificial intelligence algorithms of traditional solutions are divided into four categories according to the functions of different parts: data preprocessing -> main function implementation -> result visualization -> information interaction; among them, the main function implementation is divided into specific functions of coordinated optimization control of photovoltaic storage and charging, and orderly charging of electric vehicles according to the application needs of substation scenarios; based on the decomposed tool algorithms, a tool algorithm library covering the application needs of typical substation scenarios is constructed; and the tool algorithm library has the functions of re-modification and expansion, which can expand the application breadth of the tool algorithm library according to actual needs.

[0056] The step S3 includes: according to a specific substation scenario, calculating the statistical feature data and curve feature data in the scenario, then inputting the feature set into the substation scenario application demand identification model constructed in step S1, obtaining the application demand label of the current substation scenario to clarify the application demand, and then selecting a suitable tool algorithm from the tool algorithm library, adjusting the data format, and connecting them in series to form an overall lightweight artificial intelligence algorithm, which is deployed to the substation edge device to realize the convenient application of the lightweight artificial intelligence algorithm;

[0057] The arrangement and construction of the lightweight artificial intelligence algorithm is expressed in the following formula:

[0058]

[0059] Where Model is a lightweight artificial intelligence algorithm model used to solve the application needs of the substation scene; P i is the i-th tool algorithm in the tool algorithm library; I i It is the tool algorithm P i Input; D pre is a collection of data preprocessing tool algorithms; T is the scene feature set of the current station area, which consists of curve features F and statistical features X; α i Whether to apply P i The flag of the tool algorithm, α i =0 means that the tool algorithm P is not used in the overall model i β i Whether to apply D pre The flag of the data preprocessing algorithm, β i =0 means directly taking the scene feature set T as the input of the tool algorithm I i , β i =1 means that the scene feature set T is passed through D pre After the data represented is preprocessed, it is used as a tool algorithm P i Input I i .

[0060] In step S1, the PCA dimensionality reduction algorithm is used to reduce the dimensionality of the scene feature sets of all substations, and then the K-means clustering algorithm is used for unsupervised classification to obtain several typical substation operation scenarios. For each typical substation scenario, the application requirements under the scenario are comprehensively analyzed. Then, the decision tree algorithm is used to establish a mapping relationship between the scene feature set and the scene application requirements, and a substation scene application requirement identification model is constructed.

[0061] In step S1, the specific method of performing unsupervised classification using the K-means clustering algorithm is as follows:

[0062] Assume that the total sample of input feature set is S={x1,x2,x3,...,x m}, where x i Represents the characteristic value of the scene sample i in the substation, selects the number of clusters as k, initializes the cluster center vector, and generates k cluster centers, which are {μ1,μ2,...,μ k}, for each stage scene sample x i , mark it as the category closest to the category center, that is:

[0063] label i =argmin||x i -μ j || Formula 4;

[0064] Update each cluster center vector to the mean of all samples belonging to the category as follows:

[0065]

[0066] Repeat the above two steps until the change value of the cluster center is less than a certain threshold to obtain the final substation scene clustering result. Points with similar distances can be regarded as scenes of the same type. Based on the substation clustering result labels, calculate the mean of the scene feature set under each category and use it as the feature set of the typical substation scene.

[0067] In step S1, the PCA dimensionality reduction algorithm is used to reduce the dimensionality of the massive curve feature data of the station scene feature set, that is, the data projection technology is used to reduce the dimensionality of the massive curve feature data. The objective function is expressed as follows:

[0068]

[0069] where X={x1,x2,x3,...,x L}, represents a data sequence X of length L, X T is the transposed vector of X, and w1 is the principal axis direction of the projection.

[0070] When the distribution substation is a high-PV penetration substation, the substation scenario feature set is input into the substation scenario application demand identification model to obtain application demand labels including photovoltaic output forecasting, photovoltaic storage coordinated control, and distribution transformer load forecasting. Then, in step S3, based on the application demand label, the appropriate tool algorithm is selected to construct an overall lightweight artificial intelligence algorithm.

[0071] In step S1, the application requirements of typical scenarios in the substation area are analyzed, and the analysis results are used as labels. A decision tree algorithm is used to construct a mapping relationship between feature sets and application requirements, and a supervised classification model is established to achieve the function of inputting the substation scene feature set and outputting the corresponding application requirements of the scene;

[0072] When the decision tree algorithm is executed, the scene feature set is judged layer by layer from top to bottom until the leaf node at the bottom of the decision tree outputs the application requirements corresponding to the scene feature set.

[0073] In step S2, based on the feedback from practical applications, more fine-grained tool algorithms are added to the tool algorithm library. The added more fine-grained tool algorithms include PCA dimensionality reduction, DBSCAN clustering, and logistic regression.

[0074] The input and output data formats of the tool algorithm are adjustable. When the tool algorithm is an outlier detection tool algorithm, the input data size is set to the size of the station scene feature set after dimensionality reduction, and the output data size is set to the input data size of the subsequent missing value filling tool algorithm.

[0075] Example 1:

[0076] In this example, we first collected equipment inventory data and historical operating data from different distribution substations. From this data, we selected a feature set that represents the substation scenario types. The scenario feature set consists of statistical and curve features. Statistical features primarily include numerical data such as daily maximum load, daily minimum load, daily average load, peak-to-valley load difference, total installed wind power capacity, total installed photovoltaic capacity, and energy storage capacity. Curve features primarily include sequence data such as the substation's daily load curve, photovoltaic output curve, wind power output curve, and the charge and discharge curves of energy storage or charging stations. The scenario feature sets for all substations were dimensionality reduced using the PCA algorithm. Unsupervised classification was then performed using the K-means clustering algorithm to identify several typical substation operation scenarios. For each typical substation scenario, we comprehensively analyzed the application requirements within that scenario. Then, using a decision tree algorithm, we established a mapping between the scenario feature set and the scenario application requirements, thereby constructing a substation scenario application requirement identification model. For example, for a high-PV penetration area, inputting the area's scenario feature set into the identification model yields application requirement tags for photovoltaic output forecasting, photovoltaic-storage coordinated control, and distribution transformer load forecasting. In step S3, staff can quickly and accurately select the appropriate tool algorithm based on these application requirement tags to build an overall lightweight AI algorithm.

[0077] Next, we reviewed the application requirements for all typical scenarios and analyzed the traditional AI algorithms used in each solution. We then decomposed these algorithms into the following functionally decoupled tool algorithms, based on their basic functions: outlier detection algorithm, missing value imputation algorithm, data normalization algorithm, main program for identifying household-to-transformer relationships, main program for topology identification, main program for voltage monitoring, coordinated control optimization for photovoltaic (PV) storage and charging, orderly charging of electric vehicles, main program for forecasting renewable energy output, main program for forecasting distribution load, and main program for analyzing residential electricity consumption behavior. This number of tool algorithms should cover the application requirements of most substation scenarios. Drawing on the concept of extension packages in the Python language, we constructed a tool algorithm library designed to facilitate the implementation of AI algorithms for distribution substations. Furthermore, the tool algorithm library supports further expansion, allowing for the addition of more granular tools such as PCA dimensionality reduction, DBSCAN clustering, and logistic regression based on feedback from future practical applications.

[0078] Next, for a newly added substation scenario, the statistical and curve features of the scenario are obtained from the substation equipment ledger and historical operation data to form a feature set for the substation scenario. This feature set is input into the substation scenario application requirement automatic identification model established in S1 to obtain the application requirement label for the substation scenario, allowing staff to quickly determine the artificial intelligence algorithm that needs to be programmed.

[0079] Finally, the staff selects the appropriate tool algorithm from the tool algorithm library based on the application requirement label, uses custom parameters to set the format of input and output data of each tool algorithm, and then arranges and connects several tool algorithms in series to form a lightweight artificial intelligence algorithm solution that can solve the application needs of the distribution station area, such as Figure 2 As shown; after that, the algorithm is deployed to the edge equipment of the substation area, which can realize the convenient application of lightweight artificial intelligence algorithms in the distribution substation area.

[0080] Example 2:

[0081] In this example, we draw on the concept of extension libraries in Python to organize all tool algorithms into a collection, the tool algorithm library. This library is designed to be extensible later, allowing us to add additional, more fine-grained tool algorithms based on feedback from future practical applications.

[0082] In this example, for a newly added specific substation scenario, the statistical and curve-based feature data for that scenario are calculated based on the substation's ledger data and historical operating data to form a substation scenario dataset. This feature set is then input into the substation scenario application requirement identification model constructed in S1 to obtain the application requirement label for the current substation scenario. After clarifying the application requirements, the staff selects appropriate tool algorithms from the tool algorithm library, adjusts the data format of each tool algorithm by setting parameters, and cascades them into a comprehensive lightweight AI algorithm. For example, for the application requirements of heavy overload prediction and renewable energy output prediction, the staff selects the outlier detection tool algorithm, missing value imputation tool, PCA dimensionality reduction algorithm, renewable energy output prediction, and logistic regression tool algorithms from the tool algorithm library, cascading them to form a lightweight AI algorithm that meets the application requirements of the distribution substation. Finally, the lightweight AI algorithm is deployed to the substation's edge devices, enabling convenient application of AI algorithms at the edge of the distribution substation.

[0083] Example 3:

[0084] In this example, using data from multiple actual substations, we analyzed and collated the following statistical features: total installed photovoltaic capacity, total installed wind power capacity, energy storage capacity, electric vehicle charging pile capacity, daily load factor, daily maximum load, daily minimum load, daily average load, daily electricity consumption, off-peak power coefficient, and percentage of electricity consumption during normal periods. Curve features included the following: photovoltaic output curve, wind power output curve, distribution transformer load curve, energy storage charge and discharge curve, charging pile charge and discharge curve, and distribution transformer three-phase voltage and current curves. This generated a scenario feature set for each substation. This feature set was then reduced using the PCA dimensionality reduction algorithm. Unsupervised cluster analysis was then performed using the K-means clustering algorithm. This yielded nine substation scenario types. After specialized analysis, these nine substation types were categorized as traditional (no renewable energy generation equipment), high-proportion photovoltaic (PV) substation, high-proportion wind power substation, renewable energy power generation (EV) substation, and integrated PV, storage, and charging substation. Labels were then added to match the scenario application requirements. Then, using the reduced-dimensional scene feature set and the manually added substation scene application requirement labels, the problem was transformed into a supervised classification problem, and a scene application requirement identification model suitable for distribution substations was constructed.

[0085] Targeting the application requirements of nine substation scenarios, 19 tool algorithms were implemented. These algorithms are broadly categorized into data mining algorithms, including data dimensionality reduction, missing value detection, missing value filling, and clustering, as well as functional algorithms, such as photovoltaic output forecasting, wind power output forecasting, distribution transformer voltage imbalance monitoring, heavy overload capacity prediction, and photovoltaic-storage-charging coordination optimization. Furthermore, the input and output data formats of these tool algorithms can be customized through parameter settings, ensuring that the different algorithms can operate in tandem and collaboratively. After testing, these tool algorithms were organized into a tool algorithm library.

[0086] An actual substation was selected as the test object, and its substation scenario feature set was input into the scenario application demand identification model. The output application demand labels were heavy overload prediction and new energy output prediction. Based on the application demand labels, the staff selected a total of 7 tool algorithms from the tool algorithm library, and constructed heavy overload prediction application algorithms and new energy output prediction algorithms respectively. Then, according to the lightweight artificial intelligence algorithm arrangement and construction method described in S3, a combination was obtained to obtain an overall solution that meets the application needs of the test substation. The solution construction diagram is shown in the figure below. Figure 2 As shown in the figure, the lightweight artificial intelligence algorithm constructed by inputting the historical load data and renewable energy output data of the test area can output reasonable prediction results. The lightweight artificial intelligence algorithm for distribution area constructed and implemented based on the present invention is effective.

Claims

1. A lightweight artificial intelligence algorithm arrangement and construction method for distribution substations, characterized by: The following steps are included: Step S1: Combine the equipment records and historical operation data of the distribution area to form a scenario feature set of the area, analyze and extract typical area scenarios, summarize the collaborative management and control and regional autonomy application requirements of typical scenarios, and establish a mapping relationship between scenarios and application requirements; Step S2: For different application demand scenarios, the artificial intelligence algorithm is decomposed into tool algorithms that are decoupled and independent of each other according to their functions, and then organized and summarized into a tool algorithm library; Step S3: Based on the scene feature set of the substation, identify the scene type and output the application requirement label, select several required tool algorithms, standardize the data format between the tool algorithms, ensure the smooth flow of data between different tool algorithms, use parameter setting input and output parameters to make different tool algorithms operate in series, form the final lightweight artificial intelligence algorithm, and deploy it to the edge equipment of the distribution station area; In step S1, the equipment ledger and historical operation data of the distribution area are used to consider the area scale, the proportion of new energy access, and the energy storage configuration data to form a scene feature set of the area. The scene feature sets of all areas are clustered and analyzed to obtain the typical scene types of the area. The application requirements corresponding to the typical scenes are comprehensively analyzed and a mapping relationship between the two is established. Finally, a scene application requirement identification model is constructed that can output the scene application requirement label of the area by inputting the scene feature set of the area. The scenario feature set consists of statistical features and curve features. Specifically, the statistical features are divided into numerical data such as the number of users in the area, the proportion of photovoltaic access, energy storage capacity, total installed wind power capacity, daily average load, peak-valley load difference, and total installed photovoltaic capacity; the curve features are divided into distributed photovoltaic output curves, distribution transformer load curves, and charge and discharge curve sequence data of energy storage or charging piles. For the i-th substation scene, its statistical characteristics and curve characteristics are: Among them, X i is the statistical feature of the i-th station scene, is the jth statistical feature among the M statistical features of the i-th station scene; F i is the curve feature of the i-th station scene. Assuming that the scene has N curve features and the maximum length of the data sequence is L, then F i is a matrix of the form N×L, Indicates the kth data in the jth curve feature sequence in the i-th station scenario. If there are curve features with inconsistent lengths, the shorter data sequence will be padded with 0s starting from the end. Based on the above-mentioned substation scene feature set, the big data mining method of PCA dimensionality reduction algorithm and K-means clustering algorithm is used to realize substation scene classification and obtain typical scenes of all substations; application demand analysis is performed on typical substation scenes; then, the analysis results are used as labels and the substation scene feature set is used as input to establish a substation scene application demand identification model based on a decision tree.

2. The method for arranging and constructing a lightweight artificial intelligence algorithm for a distribution substation according to claim 1, characterized in that: In step S2, the artificial intelligence algorithm is an artificial intelligence algorithm in a traditional solution for distribution substations; the algorithms corresponding to typical scenario types in distribution substations include an outlier detection algorithm, a missing value filling algorithm, a data normalization algorithm, a household-to-transformer relationship identification main program, a topology identification main program, a voltage monitoring main program, a photovoltaic storage and charging coordinated control optimization, an electric vehicle orderly charging program, a new energy output prediction main program, a distribution load prediction main program, and a functional decoupling tool algorithm for a resident electricity consumption behavior analysis main program; In step S2, the AI ​​algorithms of traditional solutions are divided into four categories based on their functions: data preprocessing, main function implementation, result visualization, and information interaction. The main function implementation is divided into specific functions such as PV-storage-charging coordinated optimization control and orderly charging of electric vehicles based on the application needs of the substation scenario. Based on the decomposed tool algorithms, a tool algorithm library covering the application needs of typical scenarios in the substation is constructed; in addition, the tool algorithm library has the function of re-modification and expansion, which can expand the application breadth of the tool algorithm library according to actual needs.

3. The method for arranging and constructing a lightweight artificial intelligence algorithm for a distribution substation according to claim 1, characterized in that: The step S3 includes: calculating the statistical feature data and curve feature data in the scenario according to the substation scenario, and then inputting the feature set into the substation scenario application demand identification model constructed in step S1 to obtain the application demand label of the current substation scenario to clarify the application demand, and then selecting a suitable tool algorithm from the tool algorithm library, adjusting the data format, and connecting them in series to form an overall lightweight artificial intelligence algorithm, which is deployed to the substation edge device to realize the convenient application of the lightweight artificial intelligence algorithm; The arrangement and construction of the lightweight artificial intelligence algorithm is expressed in the following formula: Where Model is a lightweight artificial intelligence algorithm model used to solve the application needs of the substation scene; P i is the i-th tool algorithm in the tool algorithm library; I i It is the tool algorithm P i Input; D pre is a collection of data preprocessing tool algorithms; T is the scene feature set of the current station area, which consists of curve features F and statistical features X; α i Whether to apply P i The flag of the tool algorithm, α i =0 means that the tool algorithm P is not used in the overall model i β i Whether to apply D pre The flag of the data preprocessing algorithm, β i =0 means directly taking the scene feature set T as the input of the tool algorithm I i , β i =1 means that the scene feature set T is passed through D pre After the data represented is preprocessed, it is used as a tool algorithm P i Input I i .

4. The method for arranging and constructing a lightweight artificial intelligence algorithm for a distribution substation according to claim 1, characterized in that: In step S1, the PCA dimensionality reduction algorithm is used to reduce the dimensionality of the scene feature sets of all substations, and then the K-means clustering algorithm is used for unsupervised classification to obtain several typical substation operation scenarios. For each typical substation scenario, the application requirements under the scenario are comprehensively analyzed. Then, the decision tree algorithm is used to establish a mapping relationship between the scene feature set and the scene application requirements, and a substation scene application requirement identification model is constructed.

5. The method for arranging and constructing a lightweight artificial intelligence algorithm for a distribution substation according to claim 4, characterized in that: In step S1, the specific method of K-means clustering algorithm for unsupervised classification is as follows: Assume that the total sample of input feature set is S = {x1, x2, x3, ..., x m }, where x i Represents the characteristic value of the scene sample i in the substation, selects the number of clusters as k, initializes the cluster center vector, and generates k cluster centers, which are {μ1,μ2,...,μ k }, for each stage scene sample x i , mark it as the category closest to the category center, that is: label i =argmin||x i -μ j || Formula 4; Each cluster center vector is updated to the mean of all samples belonging to the category as follows: where μ j is the center vector of the j-th cluster, Repeat the above two steps until the change value of the cluster center is less than a certain threshold to obtain the final substation scene clustering result. Points with similar distances can be regarded as scenes of the same type. Based on the substation clustering result labels, calculate the mean of the scene feature set under each category and use it as the feature set of the typical substation scene.

6. The method for arranging and constructing a lightweight artificial intelligence algorithm for a distribution substation according to claim 4, characterized in that: In step S1, the PCA dimensionality reduction algorithm is used to reduce the dimensionality of the massive curve feature data of the station scene feature set, that is, the data projection technology is used to reduce the dimensionality of the massive curve feature data. The objective function is expressed as follows: where X={x1,x2,x3,...,x L }, represents a data sequence X of length L, X T is the transposed vector of X, w1 is the principal axis direction of the projection, and m is the total number of samples in the feature set.

7. The method for arranging and constructing a lightweight artificial intelligence algorithm for a distribution substation according to claim 4, characterized in that: When the distribution substation is a high-PV penetration substation, the substation scenario feature set is input into the substation scenario application demand identification model to obtain application demand labels including photovoltaic output forecasting, photovoltaic storage coordinated control, and distribution transformer load forecasting. Then, in step S3, based on the application demand label, the appropriate tool algorithm is selected to construct an overall lightweight artificial intelligence algorithm.

8. The method for arranging and constructing a lightweight artificial intelligence algorithm for a distribution substation according to claim 1, characterized in that: In step S1, the application requirements of typical scenarios in the substation are analyzed, and the analysis results are used as labels. The decision tree algorithm is used to construct a mapping relationship between feature sets and application requirements, and a supervised classification model is established to achieve the function of inputting the substation scenario feature set and outputting the application requirements corresponding to the scenario; when the decision tree algorithm is executed, the scene feature set is judged layer by layer from top to bottom until the application requirements corresponding to the scene feature set are output at the leaf node at the bottom of the decision tree.

9. The method for arranging and constructing a lightweight artificial intelligence algorithm for a distribution substation according to claim 1, characterized in that: In step S2, based on the feedback from practical applications, more fine-grained tool algorithms are added to the tool algorithm library. The added more fine-grained tool algorithms include PCA dimensionality reduction, DBSCAN clustering, and logistic regression. The input and output data formats of the tool algorithm are adjustable. When the tool algorithm is an outlier detection tool algorithm, the input data size is set to the size of the station scene feature set after dimensionality reduction, and the output data size is set to the input data size of the subsequent missing value filling tool algorithm.

Images