Data-driven power system operating mode identification method, system and medium

Abstract

The application discloses a power system operation mode identification method and system based on data driving, and a medium, comprising: obtaining time series production operation data in a preset historical time period of a power system; inputting the time series production operation data into an improved near neighbor propagation clustering model for clustering to obtain an operation mode clustering result; selecting key influence factors representing load sequence characteristics as time series load characteristics; performing prediction according to the time series load characteristics to obtain a prediction result; calculating a daily load total amount, a daily load rate and a daily load peak-valley difference corresponding to a target date according to the prediction result and taking the daily load total amount, the daily load rate and the daily load peak-valley difference as daily load characteristics; and establishing a future power grid operation mode identification model based on a mapping relationship between the daily load characteristics and typical operation mode labels to perform future typical operation mode identification. The application can well improve the deficiency of formulating typical operation modes based on experience and accurately analyze and identify typical operation modes of a power grid.

Description

Technical Field

[0001] This invention relates to the field of power system dispatching and operation control technology, specifically to a data-driven method, system, and medium for identifying power system operation modes. Background Technology

[0002] With the rapid development and construction of new power systems, and the large-scale grid connection of new energy sources, pumped storage, and energy storage equipment, the power grid's operation modes have become more flexible and varied. The characteristics of power grid operation and dispatch control have become more complex. Power systems are gradually exhibiting diversified operation modes, bidirectional power flow, and more complex stability mechanisms. Uncertainty and flexibility are increasingly dominating the power grid's operation modes. Traditional model- and experience-based power system analysis methods are increasingly unable to cope with these complex and changing operation modes, significantly impacting the operation, planning, protection, and stability analysis of power systems.

[0003] The operating mode of a power system reflects its operational status at a specific time scale (e.g., year, month, day) and time granularity (e.g., hour, minute, second). It is primarily determined by factors such as line power flow, generator output, load level, and grid topology. Annual operation analysis of the power system is often the most crucial, typically involving several typical operating modes. These modes form the basis for calculating the power system's operational safety margin, transient stability, reliability assessment, and planning decisions. Because the operating mode of a power system is mainly influenced by load variations or seasonal output, it exhibits relatively fixed patterns. Therefore, typical operating modes are often selected based on seasonal variations and load levels.

[0004] Power system operation mode identification is a typical big data analysis problem. Considering the variability of load and power flow, traditional methods are limited by expert experience and historical patterns. Determining the number and modes of power system operation modes using traditional methods greatly reduces the reliability of the results and cannot accurately analyze and identify typical power grid operation modes. In addition, the modeling complexity of mechanistic model-driven methods further restricts the application of refined power system operation mode identification. Summary of the Invention

[0005] The technical problem to be solved by this invention is that traditional methods for identifying power system operation modes are limited by expert experience and historical patterns. Determining the number and patterns of power system operation modes through traditional methods will greatly reduce the reliability of the results and will not be able to accurately analyze and identify typical power grid operation modes.

[0006] The purpose of this invention is to provide a data-driven method, system, and medium for identifying power system operation modes. This invention considers time-series production and operation data, combining data-driven and deep learning-based frameworks for analyzing and identifying typical power system operation modes. Compared with existing technologies, this approach, with data-driven principles as its core concept and hourly load sequences as its analytical basis, proposes an improved AP time-series clustering algorithm (i.e., adaptive dynamic sequence clustering method) for classifying and analyzing typical power system operation modes in different scenarios. This invention effectively overcomes the shortcomings of relying on experience to determine typical operation modes, accurately analyzing and identifying typical power grid operation modes, and helping researchers obtain more reasonable typical operation modes.

[0007] This invention is achieved through the following technical solution:

[0008] In a first aspect, the present invention provides a data-driven method for identifying the operating mode of a power system, the method comprising:

[0009] Acquire time-series production and operation data of the power system within a preset historical time period, and divide the time-series production and operation data into hourly time series data using data slicing;

[0010] The time-series production operation data is input into the improved nearest neighbor propagation clustering (AP clustering) model for clustering to obtain the operation mode clustering results and serve as the labels for typical operation modes. The improved nearest neighbor propagation clustering (AP clustering) model is an adaptive threshold nearest neighbor propagation clustering AP (Affinity Propagation) clustering model that uses dynamic time regularization distance as the evaluation index.

[0011] Key influencing factors that can be predicted or artificially defined to characterize the load sequence characteristics are selected. These key influencing factors include total daily load, daily load factor (the ratio of the average daily load to the maximum daily load), and daily load peak-to-valley difference. These key influencing factors are then used as time-series load characteristics.

[0012] Based on the time-series load characteristics, a day-ahead load level prediction model is used to make predictions and obtain prediction results.

[0013] Based on the forecast results, calculate the total daily load, daily load rate, and daily load peak-valley difference for the target date and use them as daily load characteristics.

[0014] Based on the mapping relationship between daily load characteristics and typical operation mode labels, a future power grid operation mode identification model is established; and the future power grid operation mode identification model is used to identify future typical operation modes.

[0015] This invention introduces time-series load data and employs an improved nearest-neighbor propagation clustering (AP clustering) model to obtain typical and special operating modes. It visualizes the characteristics of power system operating modes based on variance dispersion, intra-cluster correlation, and seasonal consistency indices. Key features of daily load levels are extracted, and a day-ahead load forecasting model is constructed using a TCN network to calculate key feature indices for predicted daily load levels. Finally, based on the spatial mapping relationship between key features of daily load levels and corresponding operating modes, a future typical operating mode identification model is established using a TCN network to obtain the identification results of future typical operating modes. This invention effectively overcomes the shortcomings of formulating typical operating modes based on experience, accurately analyzes and identifies typical power grid operating modes, and helps researchers obtain more reasonable typical operating modes.

[0016] Furthermore, this invention takes into account that the basic idea of ​​Affinity Propagation (AP) clustering is to treat all sample points as network nodes, and select suitable cluster centers and complete clustering by continuously transmitting information between different nodes.

[0017] First, let's introduce some key parameters in nearest neighbor propagation clustering (AP clustering): s i,k Representative sample point X i With X k The similarity between them is used to determine the similarity of sample X. k Is it suitable as a sample X? i Cluster centers are typically represented by negative Euclidean distances.

[0018] s i,k =-||X i -X k || 2

[0019]

[0020] s k,k =median(S n×n )

[0021] In the formula, S n×n Indicates by s i,k The resulting n×n dimensional similarity matrix; s k,k This is called the reference degree, and in this invention, it is set as the similarity matrix S. n×n The median of all values ​​in the similarity matrix S is objectively defined as the median of the similarity matrix S. n×n Points with the same coordinates on both the horizontal and vertical axes represent the probability that point k is a cluster center. Before the iteration begins, it is assumed that all sample points have an equal probability of becoming cluster centers. k,k The size of the cluster will affect the final number of clusters.

[0022] r i,k and a i,k This indicates that from sample point X respectively i With X k Determine the angle of X k As X i The suitability of cluster centers:

[0023]

[0024]

[0025] Among them, s i,k For point X i With X k The similarity between them, a i,k This represents the degree of belonging of all points except point k to point i; s i,j r represents the attraction of points other than point k to point i; k,k To represent the similarity value of point k as a cluster center, r i′,k This represents the similarity value of point k as the cluster center of all points except point i.

[0026] r i,k and a i,k The convergence rate and stability are adjusted by the damping coefficient γ, which typically takes a value between 0 and 1. i,k and a i,k The update method is as follows:

[0027]

[0028] Where the superscripts t+1 and t represent the iteration number, Let k in the t-th iteration be the similarity value of the cluster centers of all points except i. Let point k in the (t+1)th iteration be the similarity value of the cluster centers of all points except point i. To represent the degree of belonging of all points except point k to point i in round t, Let represent the degree of belonging of all points except point k to point i in the (t+1)th round.

[0029] Most studies set the damping coefficient γ to a constant value, but a fixed damping coefficient γ is obviously insufficient to guarantee high efficiency and stability when dealing with different targets.

[0030] Therefore, this invention improves the AP clustering algorithm by using an adaptive threshold AP (Affinity Propagation) clustering model with dynamic time-normalized distance as the evaluation metric as the improved nearest neighbor propagation clustering (AP clustering) model.

[0031] The improved nearest neighbor propagation clustering (AP clustering) model replaces the Euclidean distance representation in the AP clustering algorithm with dynamic time warping distance. Dynamic time warping finds an optimal curved path P by adjusting the relationship between corresponding data at different times in two sequences, and the k-th element p of the optimal curved path P is... k =(i,j) k The optimal path satisfies the conditions of boundary, monotonicity, and continuity to produce the optimal distance value; where i is the x-coordinate of the k-th element and j is the y-coordinate of the k-th element.

[0032] Furthermore, the formula for calculating the dynamic time-warped distance is:

[0033]

[0034]

[0035] Where, d DTW (x,y) represents the dynamic time warping distance, K is the sequence length of the time-series production operation data, and dist(p) is the distance between the two points. k () represents the Euclidean distance between sample points of sequence x and sequence y; Let x be the i-th node. Let y be the j-th node of sequence y.

[0036] Furthermore, the time-series production operation data is input into the improved nearest neighbor propagation clustering (AP clustering) model for clustering. The specific steps are as follows:

[0037] Step 41: Normalize the time-series production operation dataset X = {X1, X2, ..., Xn} and select the dynamic time warping distance (i.e., d). DTW Instead of Euclidean distance, construct a similarity matrix S. n×n And let the dynamic time-warped distance (i.e., d) be... DTW The negative of the calculated result equals the performance s i,k ;

[0038] Step 42: Calculate the membership degree and attraction degree of the nodes in the time-series production operation data, i.e., r and a; the initial state is r = 0, a = 0;

[0039] Step 43: Given the degree of affiliation and the degree of attraction, update the performance s of each point in the similarity matrix. i,k Simultaneously, the affiliation and attraction are updated based on the updated performance; interference during the iteration process is suppressed, and an initial threshold γ0 and optimization step size e are set.

[0040] Step 44: Sum the attraction and affiliation information for each node, make a decision, and determine whether the iteration should stop. If so, update the initial threshold γ0 using the threshold update method. The updated threshold is γ0.l If not, continue iteratively calculating the attraction and belonging information of each node and execute the decision.

[0041] Step 45, update γ l And repeat step 44 until all γ values ​​meet the conditions. l Clustering results were obtained for all γ groups, and all γ groups were clustered together. l The corresponding clustering results are stored in a similarity matrix, and the iterative threshold γ corresponding to the highest similarity value is selected. l As the optimal threshold γ best ;

[0042] Step 46, query the optimal threshold γ best The corresponding clustering results are used as the clustering results of the running mode and output.

[0043] Furthermore, the formula for the threshold update method is:

[0044] γ l =γ0-e(l-1)

[0045] In the formula, l is the iteration number; e is the optimization step size.

[0046] Furthermore, the method also includes:

[0047] Three clustering evaluation indicators are introduced to visualize the clustering results of operation modes, showcasing the characteristics of power system operation modes; and the effectiveness of the improved nearest neighbor propagation clustering (AP clustering) model is verified. The three clustering evaluation indicators include variance dispersion, intra-cluster correlation, and seasonal consistency.

[0048] Among them, variance dispersion is a stability index of the improved nearest neighbor propagation clustering (AP clustering) model; intra-cluster correlation is a credibility index reflecting the clustering results of operation mode; seasonal consistency index is an index reflecting the impact of season on the operation mode of the power system. The seasonal consistency index refers to the ratio of the number of days with unchanged daily operation mode to the total number of days in a certain season.

[0049] Specifically, the specific expressions for each phase index are as follows:

[0050] (1) Variance Dispersion

[0051] This invention selects the Explained Variance Regression Score (EVS) to evaluate the variance dispersion of the model. This index can reflect the robustness of the model more intuitively than variance, and its value ranges from [0,1]. The closer it is to 1, the better the aggregation of sample points.

[0052]

[0053] In the formula, Var represents the variance calculation method, y represents the cluster center, and y′ represents other sample sequences.

[0054] (2) Intraclass correlation

[0055] This invention designs an intra-cluster correlation index based on the Spearman correlation coefficient to reflect the relevance and reliability of clustering results. By calculating the Spearman correlation between each sequence in each sample set and its corresponding cluster center, and then traversing all clusters, the distribution of the correlation matrix obtained reflects the clustering rationality of the proposed method.

[0056]

[0057] In the formula, γ s R represents the Spearman correlation coefficient, where n represents the number of variables, and R0 represents the number of variables. i and S i Let R and S represent the rank of the i-th sample point, respectively, and let S and R represent the mean of the variables.

[0058] (3) Seasonal consistency

[0059] This invention constructs a seasonal consistency (SC) index, which is defined as the ratio of the number of days with an unchanged daily operating pattern to the total number of days in a certain season.

[0060]

[0061] In the formula, Ms is the total number of days in the s-th season, and I(p j s) represents the total number of samples in the s-th season whose daily operating mode does not match the corresponding season. It is easy to see that the higher the seasonal consistency, the stronger the consistency between the power system's operating mode and the season.

[0062] Furthermore, the clustering results of operational modes include representative operational patterns and special patterns that cannot be classified;

[0063] When the clustering result of the operation mode is a special pattern that cannot be classified, the Spearman similarity distance between the special pattern that cannot be classified and the sample center of each class is calculated, and the Spearman similarity distance is integrated into the class with the lowest similarity.

[0064] Furthermore, both the day-ahead load level prediction model and the future power grid operation mode identification model adopt time-domain convolutional networks as the basic algorithm. The time-domain convolutional network uses dilated convolution operation to realize the interaction between neurons; the mathematical expression of the dilated convolution operation is shown in the following formula:

[0065]

[0066] In the formula, F(t) represents the dilated convolution operation, and X... t Let f(v) represent the sequence data, f(v) represent the filter function, u be the length of the input sequence data, and v be the value of the v-th element in the input sequence data.

[0067] Regression prediction is achieved using a day-ahead load level prediction model. Specifically, based on the concept of time series prediction, the Spearman correlation between loads at specific times a certain number of days in advance is analyzed. For samples of different typical operating modes, appropriate historical loads are selected as input features of the day-ahead load level prediction model. The model is then input into a time-domain convolutional network and reasonable parameters are set for regression prediction.

[0068] Furthermore, the future power grid operation mode identification model designs the output layer of the time-domain convolutional network as a classification layer to achieve identification and classification; the specific steps are as follows:

[0069] Based on the obtained operation mode classification results (i.e. operation mode clustering results) and load level prediction results, the daily average load, daily load rate and daily load peak-valley difference are selected as input features of the future power grid operation mode identification model, and the corresponding operation mode is selected as the label.

[0070] Train a temporal convolutional network to obtain the mapping relationship between daily load features and typical operating mode labels.

[0071] Secondly, the present invention also provides a data-driven power system operation mode identification system, which uses the aforementioned data-driven power system operation mode identification method; the system includes:

[0072] The acquisition unit is used to acquire time-series production and operation data of the power system within a preset historical time period;

[0073] An improved clustering unit is used to input time-series production operation data into an improved nearest neighbor propagation clustering (AP clustering) model for clustering, obtain operation mode clustering results and use them as typical operation mode labels; the nearest neighbor propagation clustering (AP clustering) model is an adaptive threshold AP clustering model that uses dynamic time warping distance as an evaluation index;

[0074] The feature extraction unit is used to select key influencing factors that can be predicted or set manually to characterize the load sequence. The key influencing factors include the total daily load, the daily load rate, and the daily load peak-valley difference; and the key influencing factors are used as time-series load features.

[0075] The forecasting unit is used to make forecasts based on time-series load characteristics using a day-ahead load level forecasting model, and obtain forecast results.

[0076] The daily load characteristic calculation unit is used to calculate the total daily load, daily load rate, and daily load peak-valley difference for the target date based on the prediction results and use them as daily load characteristics.

[0077] The identification model construction and identification unit is used to establish a future power grid operation mode identification model based on the mapping relationship between daily load characteristics and typical operation mode labels; and to use the future power grid operation mode identification model to identify future typical operation modes.

[0078] Furthermore, the improved nearest neighbor propagation clustering model replaces the Euclidean distance representation in the nearest neighbor propagation clustering algorithm with dynamic time warping distance. Dynamic time warping finds an optimal curved path P by adjusting the relationship between corresponding data at different times in two sequences, and the k-th element p of the optimal curved path P is... k =(i,j) k The optimal path satisfies the conditions of boundary, monotonicity, and continuity to produce the optimal distance value; where i is the x-coordinate of the k-th element and j is the y-coordinate of the k-th element.

[0079] Furthermore, the formula for calculating the dynamic time-warped distance is:

[0080]

[0081]

[0082] Where, d DTW (x,y) represents the dynamic time warping distance, K is the sequence length of the time-series production operation data, and dist(p) is the distance between the two points. k () represents the Euclidean distance between sample points of sequence x and sequence y; Let x be the i-th node. Let y be the j-th node of sequence y.

[0083] Furthermore, the system also includes:

[0084] The visualization and validation unit is used to introduce three clustering evaluation indicators to visualize the clustering results of operation modes, show the characteristics of power system operation modes, and validate the effectiveness of the improved nearest neighbor propagation clustering model. The three clustering evaluation indicators include variance dispersion, intra-cluster correlation, and seasonal consistency.

[0085] Among them, variance dispersion is a stability index of the improved nearest neighbor propagation clustering model; intra-class correlation is a credibility index of the clustering results of operation mode; seasonal consistency index is an index of the impact of season on the operation mode of the power system, which refers to the ratio of the number of days with unchanged daily operation mode to the total number of days in a certain season.

[0086] Thirdly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described data-driven power system operation mode identification method.

[0087] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0088] 1. This invention is based on a data-driven method, system, and medium for identifying power system operation modes. This invention considers time-series production and operation data, and combines a data-driven and deep learning-based framework for analyzing and identifying typical power system operation modes. Compared with existing technologies, this solution takes data-driven as its core concept and hourly load sequences as its analysis basis, and proposes an adaptive dynamic sequence clustering method (i.e., an improved nearest neighbor propagation clustering model) for classifying and analyzing typical power system operation modes in different scenarios.

[0089] 2. This invention relates to a data-driven method, system, and medium for identifying power system operation modes. The proposed data-driven and deep learning-based future operation mode identification model possesses advantages such as simple modeling, accurate results, and high efficiency. It can obtain more representative modes in power system planning, providing assurance for more frequent operation mode switching in power system operation. Furthermore, compared with existing technologies, this invention effectively addresses the shortcomings of traditional methods that are limited by expert experience and historical patterns, further enhancing the application value of refined power system operation mode identification in practical engineering cases. Attached Figure Description

[0090] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0091] Figure 1 A general framework diagram for developing a method for identifying and analyzing the operation mode of a power system.

[0092] Figure 2 This is a flowchart of the steps in the power system operation mode identification and analysis method of the present invention.

[0093] Figure 3 This is a distance difference map. Figure 3 (a) is the Euclidean distance difference diagram of the present invention; Figure 3 (b) is a distance difference diagram of DTW (Dynamic Time Warping) of the present invention.

[0094] Figure 4 This is a flowchart illustrating the adaptive γ-clustering execution process of the present invention.

[0095] Figure 5 This is a diagram of the TCN network structure framework of the present invention.

[0096] Figure 6 This is a diagram showing the typical operation mode partitioning results of Example 1 in Embodiment 1 of the present invention. Figure 6 (a) is a diagram showing the winter low-load mode as a typical operating mode division result of Example 1 of the present invention; Figure 6 (b) is a diagram showing the summer low-load mode as a typical operating mode division result of Example 1 of the present invention; Figure 6 (c) is a diagram showing the summer high-load mode as a typical operating mode division result of Example 1 of the present invention; Figure 6 (d) is the result diagram of the typical operation mode division in Example 1 of this invention, specifically the winter high load mode.

[0097] Figure 7 This is a time-series result diagram of the daily operation mode of Example 1 of Embodiment 1 of the present invention.

[0098] Figure 8 This illustrates the correlation between historical loads in a typical scenario of Example 2 of Embodiment 1 of the present invention. Figure 8 (a) shows the correlation between historical loads in typical winter conditions in Example 2 of this invention; Figure 8 (b) shows the correlation between historical loads in the summer typical mode of Example 2 of this invention.

[0099] Figure 9 The analysis results of the load level prediction index for the operating mode in Example 2 of Embodiment 1 of this invention are as follows: Figure 9 (a) Analysis results of load level prediction index for winter high-load operation mode in Example 2 of this invention; Figure 9 (b) Analysis results of load level prediction index for winter low-load operation mode in Example 2 of this invention; Figure 9 (c) Analysis results of load level prediction index for summer high-load operation mode in Example 2 of this invention; Figure 9 (d) Analysis results of load level prediction index for summer high load operation mode in Example 2 of this invention.

[0100] Figure 10 This is a typical test of the target's future operating mode identification in Example 2 of the present invention.

[0101] Figure 11 This is a block diagram of the data-driven power system operation mode identification system of the present invention. Detailed Implementation

[0102] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0103] Traditional methods for identifying power system operation modes are limited by expert experience and historical patterns. Determining the number and patterns of power system operation modes using traditional methods will greatly reduce the reliability of the results and will not be able to accurately analyze and identify typical power grid operation modes.

[0104] Therefore, to address the above problems, this invention designs a data-driven method and system for identifying power system operation modes. This invention considers time-series production and operation data, combining a data-driven and deep learning-based framework for analyzing and identifying typical power system operation modes. Compared with existing technologies, this approach uses data-driven principles as its core concept and hourly load sequences as its analytical basis, proposing an improved AP time-series clustering algorithm (i.e., an adaptive dynamic sequence clustering method) for classifying and analyzing typical power system operation modes in different scenarios. This invention effectively overcomes the shortcomings of relying on experience to determine typical operation modes, accurately analyzing and identifying typical power grid operation modes, and helping researchers obtain more reasonable typical operation modes.

[0105] Specifically, this invention approaches the issue from both qualitative and quantitative perspectives, designing reasonable frameworks for the identification and analysis of power system operation modes, such as... Figure 1 As shown, Figure 1 The first part is a power system operation mode analysis method based on 8760-hour time-series load to obtain typical operation modes; Figure 1 The second half focuses on future load level forecasting and future operation mode identification, aiming to provide a comprehensive view of the power grid and enhance its flexibility, openness, interactivity, economy, and sharing. First, using an 8760-hour load level as the target, an improved AP (Affinity Propagation) time-series clustering algorithm is employed to display the spatiotemporal distribution of power system operation modes. Second, using variance dispersion, intra-class correlation, and seasonal consistency as indicators, the performance differences between the proposed model, empirically selected algorithms, and classic clustering algorithms are compared. Third, based on the obtained operation mode identification results, predictable or artificially set scalars such as daily total load, daily load factor, and daily load peak-valley difference are selected as features to design a reasonable day-ahead load forecasting model for accurate prediction of load levels over a certain period. Finally, deep learning is used to establish an identification model for future power grid operation modes, and this is combined with an evaluation index system for visualization, providing an intuitive description.

[0106] Example 1

[0107] like Figure 2 As shown, this invention relates to a data-driven power system operation mode identification method. The flowchart of the steps of this invention, based on 8760 hours of time-series production and operation data and a data-driven power system operation mode identification and analysis method, is as follows: Figure 2 The specific steps are as follows:

[0108] Step 1: Obtain time-series production and operation data of the power system within a preset historical time period, specifically extracting historical 8760-hour load data as the core feature for analyzing the operation mode on different dates.

[0109] Step 2: Based on 8760 hours of time-series production operation data, set the dynamic time warping distance as the distance evaluation index and generate the distance matrix D in the original AP clustering algorithm.

[0110] The AP clustering algorithm treats all sample points as network nodes. By continuously transmitting information between different nodes, it selects suitable cluster centers and completes the clustering process. Wherein, s i,k Representative sample point X i With X k The similarity between them is used to determine X k Is it suitable as X? i The cluster centers are represented by dynamic distance normalization instead of Euclidean distance.

[0111] s i,k =-||X i -X k || 2 (1)

[0112] Dynamic time warping seeks an optimal curved path P between data points in two sequences, and the elements p of the optimal curved path P are... k =(i,j) k To ensure that the optimal path produces the optimal distance value, conditions such as boundary conditions, monotonicity, and continuity must be met.

[0113]

[0114]

[0115] Figure 3 (a) and Figure 3 (b) illustrates the differences between the two distance measurement methods. Dynamic time warping measures asynchronous similarity based on the optimal matching path, enabling asynchronous data with similar patterns to match each other, thereby ensuring that the algorithm has stronger robustness and adaptability.

[0116] r i,k and a i,k This indicates that from sample point X respectively i With X k Determine the angle of X k As X i The suitability of cluster centers:

[0117]

[0118]

[0119] r i,k and a i,k Both convergence speed and stability are adjusted by the damping coefficient γ, which typically takes a value between 0 and 1. i,k and a i,k The update method is as follows:

[0120]

[0121] In the formula, the superscripts t+1 and t represent the number of iterations.

[0122] This invention designs an adaptive γ based on the grid search approach. Its calculation process is as follows: Figure 4 As shown. Figure 4 In this context, e is set to 0.1, i represents the number of executions, and N... i Represents γ i The number of clusters obtained.

[0123] Step 3: Calculate the average similarity value between each class of samples and its cluster center at the initial threshold γ0. Combine this with a grid search method to obtain the optimal threshold γ for the improved nearest neighbor propagation clustering (AP clustering) model. best .

[0124] An adaptive threshold for an improved nearest neighbor propagation clustering (AP clustering) model, combining grid search method, is designed. The specific steps are as follows:

[0125] Step 31: Set the initial threshold γ0 and the optimization step size e for the improved nearest neighbor propagation clustering (AP clustering) model;

[0126] Step 32, update the initial threshold γ0 of the improved nearest neighbor propagation clustering (AP clustering) model using formula (6);

[0127] γ l =γ0-e(l-1) (6)

[0128] Where, γ l The threshold is the value after iteration, l is the iteration number, and e is the optimization step size;

[0129] Step 33: Determine whether the number of generated clusters is less than a preset value (e.g., 10) or the threshold γ after iteration. l Is it less than 0? If yes, proceed to step 34; otherwise, increment the iteration by 1 and return to step 32. The preset value is determined based on the actual needs of the power system.

[0130] Step 34: Calculate the threshold γ based on the Spearman index of intra-class correlation. lThe average similarity value between each sample class and its cluster center;

[0131] Step 35: Store all average similarity values ​​in the similarity matrix, and determine the threshold γ again. l Is it less than 0? If yes, proceed to step 36; otherwise, increment the iteration by 1 and return to step 32.

[0132] Step 36, the iterative threshold γ corresponding to the highest average similarity value. l As the optimal threshold γ best .

[0133] Step 37, output the optimal threshold γ best End the program.

[0134] Step 4: Use the obtained 8760 hours of time-series production operation data as input to improve the AP clustering model to obtain representative operation patterns and special patterns that cannot be classified.

[0135] Specifically, the improved nearest neighbor propagation clustering (AP clustering) model replaces the Euclidean distance representation in the AP clustering algorithm with dynamic time warping distance. Dynamic time warping finds an optimal curved path P by adjusting the relationship between corresponding data at different times in two sequences, and the k-th element p of the optimal curved path P is... k =(i,j) k The optimal path satisfies the conditions of boundary, monotonicity, and continuity to produce the optimal distance value; where i is the x-coordinate of the k-th element and j is the y-coordinate of the k-th element.

[0136] The specific steps of clustering in step 4 are as follows:

[0137] Step 41: Normalize the time-series production operation dataset X = {X1, X2, ..., Xn} and select the dynamic time warping distance (i.e., d). DTW Instead of Euclidean distance, construct a similarity matrix S. n×n And let the dynamic time-warped distance (i.e., d) be... DTW The negative of the calculated result equals the performance s i,k ;

[0138] Step 42: Calculate the membership degree and attraction degree of the nodes in the time-series production operation data, i.e., r and a; the initial state is r = 0, a = 0;

[0139] Step 43: Given the degree of affiliation and the degree of attraction, update the performance s of each point in the similarity matrix. i,k Simultaneously, the affiliation and attraction are updated based on the updated performance; interference during the iteration process is suppressed, and an initial threshold γ0 and optimization step size e are set.

[0140] Step 44: Sum the attraction and affiliation information for each node, make a decision, and determine whether the iteration should stop. If so, update the initial threshold γ0 using the threshold update method. The updated threshold is γ0. l If not, continue iteratively calculating the attraction and belonging information of each node and execute the decision.

[0141] Step 45, update γ l And repeat step 44 until all γ values ​​meet the conditions. l Clustering results were obtained for all γ groups, and all γ groups were clustered together. l The corresponding clustering results are stored in a similarity matrix, and the iterative threshold γ corresponding to the highest similarity value is selected. l As the optimal threshold γ best ;

[0142] Step 46, query the optimal threshold γ best The corresponding clustering results are used as the clustering results of the running mode and output.

[0143] Step 5: Calculate the Spearman similarity distance between the special patterns that cannot be classified and the sample centers of each class, and integrate them into the class with the lowest similarity.

[0144] Step 6: Introduce indicators such as variance dispersion, intra-class correlation, and seasonal consistency to visualize the clustering results of operation modes and display the characteristics of power system operation modes.

[0145] (1) Variance Dispersion

[0146] This invention selects the Explained Variance Regression Score (EVS) to evaluate the variance dispersion of the model. This index can reflect the robustness of the model more intuitively than variance, and its value ranges from [0,1]. The closer it is to 1, the better the aggregation of sample points.

[0147]

[0148] In the formula, Var represents the variance calculation method, y represents the cluster center, and y′ represents other sample sequences.

[0149] (2) Intraclass correlation

[0150] This invention designs an intra-cluster correlation index based on the Spearman correlation coefficient to reflect the relevance and reliability of clustering results. By calculating the Spearman correlation between each sequence in each sample set and its corresponding cluster center, and then traversing all clusters, the distribution of the correlation matrix obtained reflects the rationality of the proposed clustering method.

[0151]

[0152] In the formula, γ s Represents the Spearman correlation coefficient, where n represents the number of variables, and R0 represents the value of R0. i and S i Let R and S represent the rank of the i-th sample point, respectively, and let S and R represent the mean of the variables.

[0153] (3) Seasonal consistency

[0154] This invention constructs a seasonal consistency (SC) index, which is defined as the ratio of the number of days with an unchanged daily operating pattern to the total number of days in a certain season.

[0155]

[0156] In the formula, Ms is the total number of days in the s-th season. This represents the total number of samples where the daily operating mode in the s-th season does not match the corresponding season. It is easy to see that the higher the seasonal consistency, the stronger the consistency between the power system's operating mode and the season.

[0157] Step 7: Select predictable or artificially set factors such as total daily load, daily load factor (ratio of average to maximum daily load), and daily load peak-valley difference as input features for the subsequent future power grid operation mode identification model.

[0158] Step 8: Construct a day-ahead load level prediction model based on data-driven and deep learning methods.

[0159] The current load level prediction model is based on the Temporal Convolutional Network (TCN) algorithm, with the TCN output layer designed as a regression layer to achieve regression prediction. The TCN network is a deep neural network model that combines dilated causal convolution (DCC) and residual connections (RC). The TCN network architecture is as follows: Figure 5 As shown.

[0160] Figure 5 The diagram is divided into two parts. The left side shows the complete structure of a TCN, which is composed of multiple stacked TCN residual blocks. Each TCN residual block has an important parameter pair (k, d), representing the kernel size and dilation coefficient, respectively. The right side shows the internal structure of the TCN residual block.

[0161] TCN uses dilated convolution to achieve interaction between neurons. The mathematical expression of dilated convolution is as follows:

[0162]

[0163] In the formula, F(t) represents the dilated convolution operation, and X... t Let f(v) represent the sequence data, f(v) represent the filter function, u be the length of the input sequence data, and v be the value of the v-th element in the input sequence data.

[0164] Based on the concept of time series forecasting, this study analyzes the Spearman correlation between loads at corresponding times n days in advance, and selects appropriate historical loads as model input features for samples from different typical scenarios. These loads are then input into the TCN network with appropriate parameters for regression prediction.

[0165] Step 9: Based on the prediction results obtained from the load forecasting model in Step 8, calculate the total daily load, daily load rate, and daily load peak-valley difference for the target date and use them as daily load characteristics.

[0166] Step 10: Based on deep learning research, establish a future power grid operation mode identification model by mapping the existing daily load characteristics and typical operation mode labels.

[0167] The future power grid operation mode identification model will still be based on the TCN algorithm, with the TCN output layer designed as a classification layer to achieve the identification and classification purpose. The specific steps are as follows:

[0168] Step 101: Based on the obtained operation mode classification results (i.e. operation mode clustering results) and load level prediction results, select the daily average load, daily load rate and daily load peak-valley difference as input features of the future power grid operation mode identification model, and select the corresponding operation mode as the label;

[0169] Step 102: Train a temporal convolutional network to obtain the mapping relationship between daily load features and typical operating mode labels.

[0170] Step 11: Output the future power grid operation mode identification results, and end.

[0171] Finally, root mean square error, mean absolute percentage error, and accuracy are introduced as evaluation indicators for prediction results and identification efficiency. A case study is conducted using actual annual load data from a region in southwest my country to verify the effectiveness of the model.

[0172] Root mean square error I RMSE Mean absolute percentage error I MAPE These three metrics—accuracy (A), accuracy, and precision (A)—provide feedback on the overall performance of the regression and identification models from different perspectives. Their specific expressions are as follows:

[0173]

[0174]

[0175] A = (T) P +T N ) / n (13)

[0176] In the formula, n represents the number of sample points, y i and y i ' and ' represent the i-th true value and the predicted result, respectively. T P and T N These represent the correct results identified as true and the correct results identified as false, respectively. The sum of these two is the number of correctly identified samples.

[0177] The specific implementation is as follows:

[0178] (1) Introduction to Example 1

[0179] To verify the effectiveness of the proposed model in analyzing and identifying power system operation modes, Example 1 uses annual load data actually collected from a region in Southwest my country as a test example. Based on past load levels and socio-economic development trends, the maximum load in this region will reach 16.8 million kilowatts, and the maximum and minimum calculated loads for 220kV substations are 16.5 million kilowatts and 5.6 million kilowatts, respectively. To ensure the rationality of the simulation, missing and outlier values ​​in the actual load data were supplemented and corrected.

[0180] (2) Analysis of the results of Example 1

[0181] This invention preprocesses 8760 hours of historically collected production load data, employs an improved nearest neighbor propagation clustering (AP clustering) model for cluster analysis, and obtains the classification results of typical operating modes through visualization analysis, such as... Figure 6 (a)- Figure 6 As shown in (d), and the temporal changes in the daily operating mode, as follows: Figure 7 As shown. Figure 6 (a)- Figure 6 In (d), each subgraph represents a mode of operation, and the dark thick lines represent the cluster centers of each type, i.e., the typical operating mode. Figure 7 In the diagram, each dot represents a daily operating mode, different colors represent the operating patterns obtained from clustering, and pentagrams represent the cluster centers for each operating mode.

[0182] Depend on Figure 6 (a)- Figure 6 (d) shows that even without pre-setting the number of clusters, the optimal clustering result obtained using the improved nearest neighbor propagation clustering (AP clustering) model is consistent with the four typical operating modes of "high load in summer, low load in summer, high load in winter, and low load in winter" in actual practice. Figure 7 As shown, Figure 7 The category numbers correspond to four typical operating modes. Power system operating modes exhibit a certain degree of discreteness and are still primarily influenced by seasonal load and the peak and off-peak characteristics of hydropower output. Daily operating mode variations within the same season are relatively stable, but some extreme cases require special handling. Therefore, blindly adopting experience-based typical operating mode analysis methods to determine operating mode boundaries is unreasonable.

[0183] Example 1 uses three metrics—variance discreteness, intra-class correlation, and seasonal consistency—to measure the performance differences between the proposed model, the traditional AP algorithm, the K-means algorithm, and the traditional empirical partitioning method, thereby verifying the effectiveness and reliability of the proposed model. See Table 1 for details.

[0184] Table 1 Comparison of Indicators from Different Methods in Case Studies

[0185]

[0186] As shown in Table 1, excluding the seasonal consistency index, the other two indices were analyzed class by class. A higher EVS value (Equal Values) of 1 indicates lower dispersion for that class of samples. s The closer the index is to 1, the higher the similarity between the sample type and the corresponding typical operating mode. Combining these two indicators, the operating mode classification results obtained by the method proposed in this invention have lower variance dispersion and higher intra-class similarity compared to other data-driven methods, verifying that the proposed method has stronger stability and rationality than classical clustering algorithms. Furthermore, since empirical classification mainly relies on seasonal factors, the closer the SC index is to 1, the more the classification results take into account the influence of seasonal factors, and the more similar they are to seasonal loads and actual hydropower output patterns. However, the typical operating mode classification results obtained based on empirical methods, which only consider seasonal factors, have relatively poor intra-class correlation and variance dispersion indices, further proving that considering only a single characteristic to formulate a mode classification strategy is not advisable, and the results are relatively less complete. This further reflects the superiority of the proposed model and the comprehensiveness of the results.

[0187] (3) Introduction to Example 2

[0188] Example 2 again uses annual load data collected from a location in southwestern my country as the test example. Based on the obtained operational mode classification results, five groups of samples are randomly selected from each typical mode as test targets. The remaining samples are divided into training and validation sets in an 8:2 ratio to obtain the mapping relationship between input features and labels.

[0189] (4) Analysis of the results of Example 2

[0190] Considering that the data samples are all daily load sequences in 30-minute timeframes, directly designing loss functions and establishing prediction models based on features such as total daily load and peak / valley values ​​would consume a significant amount of effort in the data preprocessing and model design stages. Therefore, this invention divides this module into two parts for separate execution: First, a short-term prediction model is established for the 30-minute daily load sequences, combining data-driven and deep learning to ensure the accuracy and stability of the prediction results as much as possible. Second, a series of required target features are extracted from the prediction results.

[0191] The historical samples collected in Example 2 were divided into categories. Five groups of samples were randomly selected from each typical category as test targets. Samples prior to the test targets could be used as training sets. Cross-validation was used to train the model to reflect the predictive performance of the proposed model in different periods. Example 2 compared the proposed model with traditional CNN, LSTM, and GRU models to verify the reliability of the proposed model's prediction results.

[0192] Based on the concept of time series forecasting, this study analyzes the Spearman correlation between loads at corresponding times n days in advance and selects appropriate historical loads as model input features for samples of different typical methods. Figure 8 (a) and Figure 8 (b) shows the correlation between historical loads at corresponding times n days in advance for typical winter and summer seasons, where n ranges from 1 to 84.

[0193] like Figure 8 (a) and Figure 8 As shown in (b), historical loads under both typical modes exhibit significant periodic fluctuations on a weekly basis, showing high correlation on dates that are multiples of 7 days prior to the test target. Furthermore, the correlation between historical loads and the test target decreases under both typical modes. Due to differences in electricity consumption behavior caused by meteorological characteristics, the correlation of the winter typical mode decreases significantly with increasing time scale. Based on the legend, historical loads 1 day, 7 days, and 14 days in advance are selected as input features for the local prediction model. The correlation trend of the summer typical mode is relatively stable; therefore, historical loads 7 days, 14 days, and 21 days in advance are selected as inputs for the local prediction model.

[0194] Figure 9 (a)- Figure 9(d) and Table 2 show the results of three predictive indicators that effectively reflect daily load levels: daily average load, daily load rate, and daily load peak-to-valley difference, for four typical testing methods. To enhance the readability of the legend, the predictive indicator results for different targets were normalized, with the upper and lower limits of normalization set at 0.8 and 0.2, respectively. Table 2 contains the accurate results for each indicator. Furthermore, to minimize the randomness of the deep learning model, all results are the average of 10 model training iterations.

[0195] Table 2 compares the prediction results of different prediction methods for different typical scenarios.

[0196]

[0197]

[0198] Reference Figure 9 (a)- Figure 9 (d) As shown in Table 2, there are a total of 4 sets of calculation examples and 12 sets of predicted targets. The model proposed in this invention has the lowest prediction error in 10 of these sets, and the I of the proposed model is... RMSE and I MAPE Compared to the second-best performing models in each example, the proposed model achieved an average improvement of 14.31% and 19.25%, respectively. As detailed in Table 2, even in the examples where the performance index was not the lowest, the proposed model's I... RMSE The average difference between the index and the corresponding method of the optimal index in these examples can also be controlled within 13.5%. Meanwhile, I MAPE The average difference of the indicators relative to I RMSE The difference will be even smaller, and can be kept below 8%. Therefore, the proposed model demonstrates good predictive performance for different time and space scenarios, and is capable of undertaking the task of predicting future load levels in the short to medium term.

[0199] Based on the obtained operational mode classification results and load level prediction results, the daily average load, daily load factor, and daily load peak-valley difference were selected as the model input features to train the TCN network and obtain the mapping relationship between features and labels. A total of three random validations were conducted to evaluate the rationality of the proposed model in identifying future power system operational modes. Figure 10 The simulation results of one of the verifications are shown in Table 3, which presents the index of the identification results.

[0200] Table 3 Analysis of Future Mode Identification Assessment Indicators like Figure 10As shown in Table 3, using the predicted load level as input to the identification model, the three rounds of simulations in Example 2 all achieved 100% prediction of the operating mode corresponding to the test target, simultaneously balancing identification accuracy and computational efficiency. As long as the acquired features such as future load are relatively accurate, the TCN-based power system identification model can accurately identify the future power system operating mode. Furthermore, using the predicted daily load data as input, the typical pattern segmentation steps in Part 1 were repeated to determine the pattern affiliation of 20 samples across four typical modes. Identification accuracy and computation time were still used to compare the performance of the deep learning-based model and the unsupervised clustering-based model. Table 4 lists the specific comparison results of the two methods.

[0201] Table 4 Performance analysis of deep learning-based models and unsupervised clustering-based models Based on the identification results in Tables 3 and 4, the clustering-based identification method, due to its unsupervised nature, has higher computational efficiency than the deep learning method. However, the identification accuracy of the clustering method is 25% lower than the proposed method. This further demonstrates that combining deep learning methods with supervised learning can yield a better solution. In conclusion, the deep learning-based power system identification model of this invention can better achieve accurate identification of future operating modes. Compared with some classic algorithms, TCN has better identification performance.

[0202] This invention proposes a model framework for power system operation mode analysis and identification. Within this framework, daily load levels are first used as the basis for analyzing power system operation modes across multiple time scales. A data-driven method, including preprocessing and time-series clustering, is proposed for power system operation mode analysis and pattern recognition. Several indicators are designed to visualize the operation modes. Secondly, a day-ahead load level prediction model and a future power system operation mode identification model are established based on deep learning. Finally, taking a power system in Southwest China as an example, the analysis and identification effects of power system operation modes are analyzed. The results validate the effectiveness of the proposed framework and method. The proposed method helps researchers obtain more reasonable typical operation modes and formulate more accurate power system operation rules in conjunction with future load levels.

[0203] (1) The classification model of typical operation modes of power systems constructed based on the sequence clustering algorithm can obtain more reasonable typical modes compared with the empirical classification mode. In addition, the improved nearest neighbor propagation clustering (AP clustering) model has the advantage of not needing to pre-determine the number of clusters, and can obtain more representative modes in power system planning, providing a guarantee for more frequent operation mode switching in power system operation.

[0204] (2) The data-driven and deep learning-based future operation mode identification model has the advantages of simple modeling, accurate results, and high efficiency. Furthermore, considering the introduction of future load level forecasting, this model can provide certain reference value for identifying future power system operation modes in both engineering and theoretical fields.

[0205] Example 2

[0206] like Figure 11 As shown, the difference between this embodiment and Embodiment 1 is that this embodiment provides a data-driven power system operation mode identification system, which uses the data-driven power system operation mode identification method of Embodiment 1; this system corresponds one-to-one with the data-driven power system operation mode identification method of Embodiment 1; the system includes:

[0207] The acquisition unit is used to acquire time-series production and operation data of the power system within a preset historical time period;

[0208] An improved clustering unit is used to input time-series production operation data into an improved nearest neighbor propagation clustering (AP clustering) model for clustering, obtain operation mode clustering results and use them as typical operation mode labels; the improved nearest neighbor propagation clustering (AP clustering) model is an adaptive threshold AP clustering model that uses dynamic time warping distance as an evaluation index;

[0209] The feature extraction unit is used to select key influencing factors that can be predicted or set manually to characterize the load sequence. The key influencing factors include the total daily load, the daily load rate, and the daily load peak-valley difference; and the key influencing factors are used as time-series load features.

[0210] The forecasting unit is used to make forecasts based on time-series load characteristics using a day-ahead load level forecasting model, and obtain forecast results.

[0211] The daily load characteristic calculation unit is used to calculate the total daily load, daily load rate, and daily load peak-valley difference for the target date based on the prediction results and use them as daily load characteristics.

[0212] The identification model construction and identification unit is used to establish a future power grid operation mode identification model based on the mapping relationship between daily load characteristics and typical operation mode labels; and to use the future power grid operation mode identification model to identify future typical operation modes.

[0213] As a further implementation, the improved nearest-neighbor propagation clustering model replaces the Euclidean distance representation in the nearest-neighbor propagation clustering algorithm with dynamic time warping distance. Dynamic time warping finds an optimal curved path P by adjusting the relationship between corresponding data at different times in two sequences, and the k-th element p of the optimal curved path P is... k =(i,j) kThe optimal path satisfies the conditions of boundary, monotonicity, and continuity to produce the optimal distance value; where i is the x-coordinate of the k-th element and j is the y-coordinate of the k-th element.

[0214] As a further implementation, the formula for calculating the dynamic time-warped distance is:

[0215] Where, d DTW (x,y) represents the dynamic time warping distance, K is the sequence length of the time-series production operation data, and dist(p) is the distance between the two points. k () represents the Euclidean distance between sample points of sequence x and sequence y; Let x be the i-th node. Let y be the j-th node of sequence y.

[0216] As a further implementation, the system also includes:

[0217] The visualization and validation unit is used to introduce three clustering evaluation indicators to visualize the clustering results of operation modes, show the characteristics of power system operation modes, and validate the effectiveness of the improved nearest neighbor propagation clustering model. The three clustering evaluation indicators include variance dispersion, intra-cluster correlation, and seasonal consistency.

[0218] Among them, variance dispersion is a stability index of the improved nearest neighbor propagation clustering model; intra-class correlation is a credibility index of the clustering results of operation mode; seasonal consistency index is an index of the impact of season on the operation mode of the power system, which refers to the ratio of the number of days with unchanged daily operation mode to the total number of days in a certain season.

[0219] The execution process of each unit can be carried out according to the steps of the data-driven power system operation mode identification method in Example 1, and will not be described in detail in this example.

[0220] Meanwhile, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-mentioned data-driven power system operation mode identification method.

[0221] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0222] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0223] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0224] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0225] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A data-driven method for identifying the operating mode of a power system, characterized in that, The method includes: Acquire time-series production and operation data of the power system within a preset historical time period; The time-series production operation data is input into the improved nearest neighbor propagation clustering model for clustering to obtain the operation mode clustering results and serve as typical operation mode labels; the improved nearest neighbor propagation clustering model is an adaptive threshold nearest neighbor propagation clustering model that uses dynamic time regularization distance as the evaluation index. Key influencing factors are selected, including total daily load, daily load rate, and daily load peak-valley difference; and these key influencing factors are used as time-series load characteristics. Based on the aforementioned time-series load characteristics, a day-ahead load level prediction model is used to make predictions, and the prediction results are obtained. Based on the prediction results, the total daily load, daily load rate, and daily load peak-valley difference corresponding to the target date are calculated and used as daily load characteristics. Based on the mapping relationship between the daily load characteristics and typical operating mode labels, a future power grid operating mode identification model is established; and the future power grid operating mode identification model is used to identify future typical operating modes. The future power grid operation mode identification model adopts a temporal convolutional network as the basic algorithm, and the temporal convolutional network uses dilated convolution operation to realize the interaction between neurons; The future power grid operation mode identification model is designed by using the output layer of the time-domain convolutional network as a classification layer to achieve identification and classification. The improved nearest neighbor propagation clustering model replaces the Euclidean distance representation in the nearest neighbor propagation clustering algorithm with dynamic time warping distance. Dynamic time warping finds an optimal curved path P by adjusting the relationship between corresponding data at different times in two sequences, and the k-th element p of the optimal curved path P is... k = (i, j) k The optimal path satisfies the conditions of boundary, monotonicity, and continuity to produce the optimal distance value; where i is the x-coordinate of the k-th element and j is the y-coordinate of the k-th element. The formula for calculating the dynamic time warp distance is: ; ; in, For dynamic time-normalized distance, The sequence length of the time-series production operation data. Let x be the Euclidean distance between sample points of sequence x and sequence y; Let x be the i-th node. Let j be the j-th node of sequence y; The daytime load level prediction model uses a temporal convolutional network as its basic algorithm, and designs the TCN output layer as a regression layer; the temporal convolutional network uses dilated convolution operations to realize the interaction between neurons; The future power grid operation mode identification model includes: based on the obtained operation mode classification results and load level prediction results, selecting the daily average load, daily load rate and daily load peak-valley difference as input features of the future power grid operation mode identification model, selecting the corresponding operation mode as the label; and training a time-domain convolutional network to obtain the mapping relationship between daily load features and typical operation mode labels.

2. The data-driven power system operation mode identification method according to claim 1, characterized in that, The time-series production operation data is input into an improved nearest neighbor propagation clustering model for clustering. The specific steps are as follows: The time-series production operation dataset is normalized, and dynamic time warping distance is selected instead of Euclidean distance to establish a similarity matrix; And set the negative of the calculated result of the dynamic time-warped distance to equal the performance; Calculate the affiliation and attraction of nodes in the time-series production operation dataset; Given the degree of belonging and the degree of attraction, update the performance of each point in the similarity matrix; at the same time, update the degree of belonging and the degree of attraction based on the updated performance. To suppress interference during the iteration process, an initial threshold γ0 and an optimization step size e are set. For each node, sum the attraction and belonging information, make a decision, and determine whether the iteration should stop. If so, update the initial threshold γ0 using the threshold update method. The updated threshold is γ0. l If not, continue iteratively calculating the attraction and belonging information of each node and execute the decision. Update γ l The decision-making and judgment steps are repeated until all γ values ​​meet the conditions. l Clustering results were obtained for all γ groups, and all γ groups were clustered together. l The corresponding clustering results are stored in a similarity matrix, and the iterative threshold corresponding to the highest similarity value is selected. As the optimal threshold γ best ; Query the optimal threshold γ best The corresponding clustering results are used as the clustering results of the running mode and output.

3. The data-driven power system operation mode identification method according to claim 2, characterized in that, The formula for the threshold update method is: ； In the formula, l is the iteration number; e is the optimization step size.

4. The data-driven power system operation mode identification method according to claim 1, characterized in that, The method also includes: Three clustering evaluation indicators are introduced to visualize the clustering results of the operation modes, showcasing the characteristics of the power system operation modes; and to verify the effectiveness of the improved nearest neighbor propagation clustering model; the three clustering evaluation indicators include variance dispersion, intra-cluster correlation, and seasonal consistency. Among them, the variance dispersion is a stability index of the improved nearest neighbor propagation clustering model; the intra-class correlation is a reliability index reflecting the clustering results of the operation mode; the seasonal consistency index is an index reflecting the impact of season on the operation mode of the power system, and the seasonal consistency index refers to the ratio of the number of days with unchanged daily operation mode to the total number of days in a certain season.

5. The data-driven power system operation mode identification method according to claim 1, characterized in that, The clustering results of the operating modes include representative operating modes and special modes that cannot be classified. When the clustering result of the operation mode is a special pattern that cannot be classified, the Spearman similarity distance between the special pattern that cannot be classified and the sample centers of each class is calculated, and the Spearman similarity distance is integrated into the class with the lowest similarity.

6. A data-driven power system operation mode identification system, characterized in that, The system includes: The acquisition unit is used to acquire time-series production and operation data of the power system within a preset historical time period; An improved clustering unit is used to input the time-series production operation data into an improved nearest neighbor propagation clustering model for clustering, obtain the operation mode clustering results and use them as typical operation mode labels; the improved nearest neighbor propagation clustering model is an adaptive threshold nearest neighbor propagation clustering model that uses dynamic time regularization distance as an evaluation index. A feature extraction unit is used to select key influencing factors, including total daily load, daily load rate, and daily load peak-valley difference; and to use these key influencing factors as time-series load features. The prediction unit is used to make predictions based on the time-series load characteristics using a day-ahead load level prediction model, and obtain prediction results. The daily load characteristic calculation unit is used to calculate the total daily load, daily load rate, and daily load peak-valley difference corresponding to the target date based on the prediction results and use them as daily load characteristics. The identification model construction and identification unit is used to establish a future power grid operation mode identification model based on the mapping relationship between the daily load characteristics and typical operation mode labels; and to use the future power grid operation mode identification model to identify future typical operation modes. The future power grid operation mode identification model adopts a temporal convolutional network as the basic algorithm, and the temporal convolutional network uses dilated convolution operation to realize the interaction between neurons; The future power grid operation mode identification model is designed by using the output layer of the time-domain convolutional network as a classification layer to achieve identification and classification. The improved nearest neighbor propagation clustering model replaces the Euclidean distance representation in the nearest neighbor propagation clustering algorithm with dynamic time warping distance. Dynamic time warping finds an optimal curved path P by adjusting the relationship between corresponding data at different times in two sequences, and the k-th element p of the optimal curved path P is... k = (i, j) k The optimal path satisfies the conditions of boundary, monotonicity, and continuity to produce the optimal distance value; where i is the x-coordinate of the k-th element and j is the y-coordinate of the k-th element. The formula for calculating the dynamic time warp distance is: ; ; in, For dynamic time-normalized distance, The sequence length of the time-series production operation data. Let x be the Euclidean distance between sample points of sequence x and sequence y; Let x be the i-th node. Let j be the j-th node of sequence y; The daytime load level prediction model uses a temporal convolutional network as its basic algorithm, and designs the TCN output layer as a regression layer; the temporal convolutional network uses dilated convolution operations to realize the interaction between neurons; The future power grid operation mode identification model includes: based on the obtained operation mode classification results and load level prediction results, selecting the daily average load, daily load rate and daily load peak-valley difference as input features of the future power grid operation mode identification model, selecting the corresponding operation mode as the label; and training a time-domain convolutional network to obtain the mapping relationship between daily load features and typical operation mode labels.

7. The data-driven power system operation mode identification system according to claim 6, characterized in that, The system also includes: The visualization and verification unit is used to introduce three clustering evaluation indicators to visualize the clustering results of the operation mode, show the characteristics of the power system operation mode, and verify the effectiveness of the improved nearest neighbor propagation clustering model. The three clustering evaluation indicators include variance dispersion, intra-cluster correlation and seasonal consistency indicators. Among them, the variance dispersion is a stability index of the improved nearest neighbor propagation clustering model; the intra-class correlation is a reliability index reflecting the clustering results of the operation mode; the seasonal consistency index is an index reflecting the impact of season on the operation mode of the power system, and the seasonal consistency index refers to the ratio of the number of days with unchanged daily operation mode to the total number of days in a certain season.

8. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the data-driven power system operation mode identification method as described in any one of claims 1 to 5.

Images