Zero-carbon smart park cloud platform data optimization processing method and system

Abstract

The application provides a zero-carbon smart park cloud platform data optimization processing method and system, relates to the technical field of data processing, and comprises the following steps: step 1, a multidimensional state space is constructed according to four high-dimensional state vectors, and the covariance matrix of the state vector of a photovoltaic grid-connected point and the remaining three state vectors is calculated respectively to obtain three independent characteristic subspaces, wherein the four high-dimensional state vectors include the photovoltaic grid-connected point, an energy storage access point, a main load point and a power grid interaction point. The application improves the net load power prediction accuracy, optimizes the rationality of the energy storage charging and discharging instructions, guarantees the dynamic adaptation of the zero-carbon smart park energy system operation state, realizes the accurate processing of energy data and efficient low-carbon scheduling by constructing a multidimensional state space and a multi-level virtual energy guide diagram, combining power time sequence section collaborative reconstruction, integrating an attention mechanism time sequence prediction and joint iterative updating.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a data optimization processing method and system for a zero-carbon smart park cloud platform. Background Technology

[0002] In the operation and management of zero-carbon smart parks, cloud platforms typically need to aggregate and analyze power data from multiple nodes, such as photovoltaic power generation, energy storage systems, major loads, and grid connection points, in order to achieve net load forecasting and optimized energy storage scheduling. Existing technical solutions generally preprocess the collected time-series data through data cleaning, filtering, or simple statistical interpolation methods to eliminate noise or outliers, and then train a time-series prediction model based on the processed historical data. However, such methods may face certain limitations when dealing with data mutations caused by the switching operations of specific equipment within the park.

[0003] For example, when a Static Var Generator (SVG) in a power plant is switched on or off due to voltage regulation needs, it simultaneously affects multiple data sequences within a short period, including the output of the photovoltaic inverter, the power consumption of the main load, and the power at the grid interconnection point. A common practice is to only smooth or replace the data sequence of that node (e.g., the grid interconnection point) when a spike in power data is detected, based on data from the time windows before and after that node. This approach may fail to fully consider how the SVG switching event, as a systemic event, synchronously alters the dynamic relationships between the operating states of multiple nodes, including photovoltaics, loads, and energy storage. Due to the neglect of changes in multi-node coupling relationships under such events, reconstruction based solely on single-point data sometimes fails to accurately reconstruct the true power flow state of the system at the time of the event. The potential discrepancy between this reconstructed data and the actual physical relationship, if directly used for subsequent model training, may affect the reliability of net load range predictions, thus introducing uncertainty risks to energy storage dispatch decisions based on these predictions. Summary of the Invention

[0004] The technical problem to be solved by this invention is to provide a data optimization and processing method and system for a zero-carbon smart park cloud platform, thereby enhancing the stability of park energy operation.

[0005] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0006] Firstly, a data optimization and processing method for a zero-carbon smart park cloud platform, the method comprising:

[0007] Step 1: Based on the four high-dimensional state vectors, construct a multi-dimensional state space, and use the state vector of the photovoltaic grid connection point in the multi-dimensional state space as a reference to calculate the covariance matrix of itself and the remaining three state vectors to obtain three independent feature subspaces. The four high-dimensional state vectors include the photovoltaic grid connection point, the energy storage access point, the main load point, and the grid interaction point.

[0008] Step 2: Calculate the generalized angle between each pair of the three feature subspaces. Divide the multidimensional state space into multiple decision subdomains based on the generalized angle. In each decision subdomain, set the projection of the state vector of three representative points (energy storage access point, main load point, and grid interaction point) onto the decision subdomain. Connect the representative points of each decision subdomain with the reference state vector point to construct a multi-level virtual energy guidance map.

[0009] Step 3: Analyze the path conduction coefficients of each level in the virtual energy steering diagram. When an abnormal change in the conduction coefficient of any level is detected and matches the switching event log, the power time series section of the corresponding event is collaboratively reconstructed to generate reconstructed power data.

[0010] Step 4: Train the time series prediction model with integrated attention mechanism using reconstructed power data and meteorological time series sections to obtain the future net load power prediction range, and verify the coverage of the prediction range with the real-time power time series sections to obtain the adjusted prediction range.

[0011] Step 5: Based on the adjusted prediction interval, determine the energy storage unit charging and discharging commands, and perform joint iterative updates on the collaborative reconfiguration process and time series prediction model according to the command execution feedback.

[0012] Secondly, the zero-carbon smart park cloud platform data optimization and processing system includes:

[0013] The space construction module is used to construct a multi-dimensional state space based on four high-dimensional state vectors, and to calculate the covariance matrix of itself and the remaining three state vectors based on the state vector of the photovoltaic grid connection point in the multi-dimensional state space, so as to obtain three independent feature subspaces. The four high-dimensional state vectors include the photovoltaic grid connection point, the energy storage access point, the main load point, and the grid interaction point.

[0014] The guidance graph construction module is used to calculate the generalized angle between each pair of the three feature subspaces. Based on the generalized angle, the multidimensional state space is divided into multiple decision subdomains. In each decision subdomain, the projection of the state vector of three representative points—energy storage access point, main load point, and grid interaction point—is set in the decision subdomain. The representative points of each decision subdomain are connected to the reference state vector point to construct a multi-level virtual energy guidance graph.

[0015] The anomaly detection module is used to analyze the path conduction coefficients of each level in the virtual energy steering diagram. When an abnormal change in the conduction coefficient of any level is detected and matches the switching event log, the power time series section of the corresponding event is reconstructed collaboratively to generate reconstructed power data.

[0016] The interval verification module is used to train a time series prediction model with an integrated attention mechanism using reconstructed power data and meteorological time series sections to obtain the future net load power prediction interval, and to perform coverage verification between the prediction interval and the real-time power time series sections to obtain the adjusted prediction interval.

[0017] The iterative optimization module is used to determine the charging and discharging commands of the energy storage unit based on the adjusted prediction interval, and to jointly iteratively update the collaborative reconfiguration process and the time series prediction model according to the command execution feedback.

[0018] The above-described solution of the present invention has at least the following beneficial effects:

[0019] By constructing a multi-dimensional state space and independent feature subspaces, the dynamic correlation between photovoltaic grid-connected points and energy storage, loads, and grid interaction points is captured, enabling comprehensive perception of multi-node data coupling changes caused by system-level events such as equipment switching, and avoiding data processing deviations due to neglecting multi-node correlations. Based on generalized angle partitioning of decision subdomains and the construction of multi-level virtual energy guidance maps, the logic of cross-regional energy transmission is clearly outlined, providing accurate topological correlation basis for anomaly detection, improving the accuracy of identifying abnormal changes in transmission coefficients, and ensuring the accuracy of matching anomaly events with switching logs. Power time-series cross-section collaborative reconstruction is carried out for anomaly events, relying on multi-node correlation... The system achieves data repair, restoring the true power flow state of the system at the time of the event and improving data quality. The time-series prediction model integrating the attention mechanism, combined with the reconstructed power data and meteorological time-series cross-section training, can fully explore the time-series correlation and interaction between power and meteorological factors. With the prediction interval coverage verification mechanism, it improves the reliability and adaptability of future net load power prediction. Based on the energy storage scheduling and joint iterative update mechanism of the adjusted prediction interval, a closed-loop system of data processing, prediction, scheduling, and feedback optimization is formed, which improves the rationality of energy storage charging and discharging decisions, enhances the stability and flexibility of park energy operation, and helps to achieve the goal of energy-efficient utilization in zero-carbon smart parks. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the zero-carbon smart park cloud platform data optimization processing method provided in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of a zero-carbon smart park cloud platform data optimization and processing system provided in an embodiment of the present invention. Detailed Implementation

[0022] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0023] like Figure 1 As shown, embodiments of the present invention propose a data optimization and processing method for a zero-carbon smart park cloud platform, the method comprising the following steps:

[0024] Step 1: Based on the four high-dimensional state vectors, construct a multi-dimensional state space, and use the state vector of the photovoltaic grid connection point in the multi-dimensional state space as a reference to calculate the covariance matrix of itself and the remaining three state vectors to obtain three independent feature subspaces. The four high-dimensional state vectors include the photovoltaic grid connection point, the energy storage access point, the main load point, and the grid interaction point.

[0025] Step 2: Calculate the generalized angle between each pair of the three feature subspaces. Divide the multidimensional state space into multiple decision subdomains based on the generalized angle. In each decision subdomain, set the projection of the state vector of three representative points (energy storage access point, main load point, and grid interaction point) onto the decision subdomain. Connect the representative points of each decision subdomain with the reference state vector point to construct a multi-level virtual energy guidance map.

[0026] Step 3: Analyze the path conduction coefficients of each level in the virtual energy steering diagram. When an abnormal change in the conduction coefficient of any level is detected and matches the switching event log, the power time series section of the corresponding event is collaboratively reconstructed to generate reconstructed power data.

[0027] Step 4: Train the time series prediction model with integrated attention mechanism using reconstructed power data and meteorological time series sections to obtain the future net load power prediction range, and verify the coverage of the prediction range with the real-time power time series sections to obtain the adjusted prediction range.

[0028] Step 5: Based on the adjusted prediction interval, determine the energy storage unit charging and discharging commands, and perform joint iterative updates on the collaborative reconfiguration process and time series prediction model according to the command execution feedback.

[0029] In this embodiment of the invention, by constructing a multi-dimensional state space and independent feature subspaces, the dynamic correlation between photovoltaic grid-connected points and energy storage, loads, and grid interaction points is captured. This enables comprehensive perception of multi-node data coupling changes caused by system-level events such as equipment switching, avoiding data processing deviations due to neglecting multi-node correlations. Based on a generalized angle division of the decision subdomain and the construction of a multi-level virtual energy guidance map, the logic of energy transmission across regions is clearly outlined, providing accurate topological correlation for anomaly detection, improving the accuracy of identifying abnormal changes in the transmission coefficient, and ensuring the accuracy of matching anomaly events with switching logs. Furthermore, power time-series cross-section collaborative reconstruction is carried out for anomaly events, relying on multi-node... Data restoration is achieved through point correlation, restoring the true power flow state of the system at the time of the event and improving data quality. The time-series prediction model integrating the attention mechanism, combined with the reconstructed power data and meteorological time-series cross-section training, can fully explore the time-series correlation and interaction between power and meteorological factors. With the prediction interval coverage verification mechanism, the reliability and adaptability of future net load power prediction are improved. Based on the energy storage scheduling and joint iterative update mechanism of the adjusted prediction interval, a closed-loop system of data processing, prediction, scheduling, and feedback optimization is formed, which improves the rationality of energy storage charging and discharging decisions, enhances the stability and flexibility of park energy operation, and helps to achieve the goal of energy-efficient utilization in zero-carbon smart parks.

[0030] In another preferred embodiment of the present invention, the process of determining the four high-dimensional state vectors is as follows:

[0031] Step 001 involves collecting active power time-series data from four topology nodes—the photovoltaic grid connection point, energy storage access point, main load point, and grid interaction point—via the cloud platform. It also collects irradiance and temperature time-series data from meteorological monitoring equipment to form a meteorological time-series profile. Finally, it collects state transition records generated by the static var generator as a switching event log. Specifically, this includes: first, activating the built-in data acquisition module of the zero-carbon smart park cloud platform; and then, according to a preset acquisition frequency, collecting operational data of key topology nodes in the park's energy system, environmental meteorological data, and the status of core power equipment. Systematic and continuous data collection is conducted. For four core topology nodes: the connection point between the photovoltaic array output and the park's low-voltage distribution bus (PV grid connection point); the connection point between the energy storage battery pack and the park's distribution bus via the energy storage converter (energy storage access point); the main incoming line connection point for major power loads such as centralized high-energy-consuming equipment and comprehensive office buildings within the park (main load point); and the boundary connection point between the park's internal power distribution system and the city's public power grid (grid interaction point). Relying on high-precision active power sensors pre-deployed at each node, instantaneous active power data is collected in real-time during each collection cycle. The system collects four independent active power time-series profiles, each corresponding to a node in the topology, by arranging the instantaneous active power values ​​within a continuous collection period in chronological order. Simultaneously, environmental irradiance and temperature data are collected from distributed meteorological monitoring terminals within the park. The instantaneous irradiance and temperature values ​​within the continuous collection period are then arranged chronologically to form irradiance and temperature time-series profiles, respectively. These two time-series profiles of individual meteorological elements are then integrated to form a complete meteorological time-series profile covering key environmental influencing factors. Through the communication interface between the cloud platform and the park's static var generator control system, the system obtains real-time information on the device's operational status transitions, focusing on recording the precise time of the transition, the operating status before the transition, and the operating status after the transition. This information is then archived chronologically to form a switching event log. Finally, the four active power time-series profiles, the complete meteorological time-series profile, and the switching event log are uploaded to the dedicated data storage center of the zero-carbon smart park cloud platform via an encrypted transmission protocol for categorized storage and unified management.

[0032] Step 002: Four power time series sections and meteorological time series sections are time-aligned using sliding time windows of equal length. Within each aligned time window, the mean, variance, and frequency domain energy characteristics of each power time series section are calculated as statistical and spectral features, respectively. The mean irradiance and mean temperature of the corresponding meteorological time series sections within the window are extracted as meteorological features. The power and meteorological features of each topological node within each time window are concatenated and normalized to form the feature vector of the corresponding topological node in the corresponding time window. The feature vectors of each topological node in all time windows are arranged in chronological order to form a high-dimensional state vector representing the operating state of the corresponding topological node, resulting in four high-dimensional states. The vector, specifically, includes: First, combining the real-time operating characteristics of the park's energy system, data acquisition frequency, and response cycle of equipment status changes, a fixed and uniform 5-minute sliding time window length is set. Then, four active power time series sections and the integrated meteorological time series section are used as synchronization alignment objects. Using the set sliding time window as the basic interception unit, the window slides synchronously along the time axis, segmenting all participating time series sections. This ensures that each sliding time window contains active power data from photovoltaic grid-connected points, energy storage access points, main load points, and grid interaction points within the same time period, as well as corresponding irradiance and temperature meteorological data. This achieves precise matching of multi-source heterogeneous time series data in the time dimension. In each complete... Within a time-aligned sliding window, multi-dimensional feature calculations are performed on the extracted data from four active power time-series sections. When calculating power statistical features, the mean of each active power time-series section within the current window is calculated first, followed by the variance. For power frequency-domain energy feature calculation, a Fourier transform operation is first performed on the time-domain active power data within the window, converting the time-domain data to frequency-domain data. The amplitude of each frequency component is then squared, and finally, the squared results of all frequency component amplitudes are summed to obtain the frequency-domain energy feature of the active power time-series section within the current window. Simultaneously, feature extraction is performed on the meteorological time series data within the window, calculating the mean irradiance and mean temperature respectively. These two mean values ​​together constitute the meteorological features of the current window. For each topological node, in a fixed order of mean power, power variance, frequency domain energy features, mean irradiance, and mean temperature, the power statistical features and power frequency domain energy features extracted within the current window are sequentially combined with the meteorological features within the same window to form the initial feature set of the topological node within the current window. Subsequently, min-max normalization is performed on the initial feature set, and the final value obtained is the normalized feature value. All normalized feature values ​​are arranged in order to form the feature vector of the topological node in the corresponding time window.For each topology node, the feature vectors generated within all sliding time windows are sequentially concatenated according to their order on the time axis, forming a multidimensional data sequence with time-series attributes. This data sequence can fully characterize the dynamic changes in the operating state of the corresponding topology node throughout the entire acquisition period. This multidimensional data sequence is the high-dimensional state vector of the topology node. Through the above series of operations, four high-dimensional state vectors are finally obtained for the four topology nodes: photovoltaic grid connection point, energy storage access point, main load point, and grid interaction point.

[0033] This embodiment compensates for the shortcomings of data acquisition that only focuses on a single type of data by collecting power data, meteorological data, and key equipment switching logs from core topology nodes, thus avoiding analytical biases caused by missing data. By using sliding windows of the same time length to align multiple time-series data, it ensures the matching of different types of data in the time dimension and avoids feature correlation distortion caused by data asynchrony. By comprehensively extracting statistical, frequency domain, and meteorological features from the power data, it fully captures the time-domain variation patterns, frequency characteristics, and environmental influencing factors of the topology node's operating status. Compared to methods that only extract single-dimensional features, it can more accurately depict the overall state of the node's operating status. Feature normalization eliminates the dimensional differences between different types of features, improving the effectiveness of feature vectors. Finally, the constructed high-dimensional state vector fully represents the dynamic changes in the node's operating status, improving the reliability of the entire data optimization process from the source.

[0034] In a preferred embodiment of the present invention, step 1 includes:

[0035] Step 100: Using the four high-dimensional state vectors as basic elements, they are synchronously arranged according to their corresponding sampling time points to construct a multi-dimensional state space with time as the first dimension and state vector as the second dimension. In the multi-dimensional state space, the high-dimensional state vector corresponding to the photovoltaic grid connection point is used as the reference state vector. The covariance of the reference state vector and the high-dimensional state vector corresponding to the energy storage access point in the time dimension is calculated to form the first covariance matrix. Specifically, this includes: First, retrieving the high-dimensional state vectors of the four topology nodes (photovoltaic grid connection point, energy storage access point, main load point, and grid interaction point) generated in the previous stage, and using the sampling time points corresponding to each high-dimensional state vector as the basis for calculation. Using a unified time point as a benchmark, four high-dimensional state vectors are synchronously arranged to ensure a one-to-one correspondence between all vectors in the time dimension, eliminating any time misalignment. Based on these synchronously arranged high-dimensional state vectors, a two-dimensional multi-dimensional state space is constructed, with the time dimension as the first dimension and the high-dimensional state vectors of each topology node as the second dimension. This multi-dimensional state space intuitively reflects the change in the operating state of each topology node over time and the temporal correlation between nodes. Within the constructed multi-dimensional state space, the high-dimensional state vector corresponding to the photovoltaic grid-connected point is selected as the benchmark state vector. The relationship between this benchmark state vector and energy storage is then calculated. The covariance of the high-dimensional state vector corresponding to the access point in the time dimension is used to form the first covariance matrix. The specific covariance calculation process is as follows: First, calculate the mean of the reference state vector and the energy storage access point state vector in the time dimension. The mean of the reference state vector is the sum of the eigenvalues ​​corresponding to all time points of the vector divided by the total number of time points. The calculation method of the mean of the energy storage access point state vector is the same as that of the reference state vector. Then, for each time point, calculate the difference between the eigenvalue of the reference state vector at that time point and the mean of the reference state vector, and the difference between the eigenvalue of the energy storage access point state vector at that time point and the mean of the energy storage access point state vector. The difference between the two values ​​is multiplied to obtain the deviation product at that time point. Finally, the deviation products of all time points are added together and divided by the total number of time points to obtain the covariance value corresponding to the combination of individual feature dimensions. The arrangement rule of the first covariance matrix is ​​that the row dimension of the matrix corresponds to the feature dimension of the reference state vector (i.e., the order of power mean, power variance, frequency domain energy feature, irradiance mean, and temperature mean), and the column dimension corresponds to the feature dimension of the energy storage access point state vector (consistent with the feature order of the row dimension). The covariance values ​​corresponding to each pair of feature dimensions obtained above are filled and arranged in the order of row and column dimensions to finally form the first covariance matrix.

[0036] Step 101: Calculate the covariance of the reference state vector and the high-dimensional state vector corresponding to the main load point in the time dimension to form the second covariance matrix. Specifically, this includes: maintaining the structure of the multidimensional state space unchanged, still using the high-dimensional state vector corresponding to the photovoltaic grid-connected point as the reference state vector, calculating the covariance of this reference state vector and the high-dimensional state vector corresponding to the main load point in the time dimension to form the second covariance matrix; the calculation process is completely consistent with the calculation process of the first covariance matrix in Step 100, that is, firstly, solve for the mean of the reference state vector and the main load point state vector in the time dimension, and then calculate... The deviations of the eigenvalues ​​of the two vectors at each time point from their respective means are multiplied together, and the product of the deviations at all time points is accumulated. Finally, the result is divided by the total number of time points to obtain a single covariance value. The arrangement rules of the second covariance matrix are the same as those of the first covariance matrix. The row dimension is the feature dimension of the baseline state vector (in the order of power mean, power variance, frequency domain energy feature, irradiance mean, and temperature mean), and the column dimension is the feature dimension of the main load point state vector (in the same order as the row dimension). The covariance values ​​of all the corresponding features are filled and arranged in row and column order to obtain the second covariance matrix.

[0037] Step 102: Calculate the covariance of the reference state vector and the high-dimensional state vector corresponding to the grid interaction point in the time dimension to form the third covariance matrix. Specifically, this includes: keeping the reference state vector and the multidimensional state space unchanged, calculating the covariance of the photovoltaic grid-connected point reference state vector and the high-dimensional state vector corresponding to the grid interaction point in the time dimension to form the third covariance matrix; the calculation logic is consistent with the first two covariance matrices. First, calculate the time dimension mean of the reference state vector and the grid interaction point state vector, then calculate the product of the deviations of the two vector eigenvalues ​​at each time point, and sum all the deviation products and divide by the total number of time points to obtain a single covariance value; the arrangement rules of the third covariance matrix are also consistent. The row dimension is the feature dimension of the reference state vector (in the order of power mean, power variance, frequency domain energy feature, irradiance mean, and temperature mean), and the column dimension is the feature dimension of the grid interaction point state vector (consistent with the row dimension feature order). Fill and arrange the covariance values ​​corresponding to each pair of feature dimensions in row and column order to finally obtain the third covariance matrix.

[0038] Step 103: Perform matrix decomposition on the first covariance matrix, extracting the eigenvectors corresponding to the preset first k largest eigenvalues ​​to form a first independent feature subspace representing the dynamic correlation between the photovoltaic grid connection point and the energy storage access point. Perform matrix decomposition on the second covariance matrix, extracting the eigenvectors corresponding to the preset first L largest eigenvalues ​​to form a second independent feature subspace representing the dynamic correlation between the photovoltaic grid connection point and the main load point. Perform matrix decomposition on the third covariance matrix, extracting the eigenvectors corresponding to the preset first J largest eigenvalues ​​to form a third independent feature subspace representing the dynamic correlation between the photovoltaic grid connection point and the grid interaction point. The independent feature subspace specifically includes: First, the QR algorithm is used to perform eigenvalue decomposition on the first covariance matrix, and the convergence threshold is set to 1e-6 (that is, when the difference between the eigenvalues ​​calculated in two iterations is less than 1e-6, the iteration is considered to have converged). It should be clarified that the first covariance matrix is ​​a 5×5 square matrix (corresponding to the 5 feature dimensions of the high-dimensional state vector: power mean, power variance, frequency domain energy feature, irradiance mean, and temperature mean). Therefore, the orthogonal matrix Q and the upper triangular matrix R obtained by decomposition are both 5×5 matrices, and their arrangement rules strictly correspond to the dimensional order of the first covariance matrix. The specific matrix transformation solution process is as follows: First, perform orthogonal triangular decomposition on the first covariance matrix, that is, decompose the first covariance matrix into the product of a 5×5 orthogonal matrix Q and a 5×5 upper triangular matrix R. The orthogonal matrix Q is arranged in such a way that both its row and column dimensions correspond to the five feature dimensions of the high-dimensional state vector (in the order of mean power, power variance, frequency domain energy feature, mean irradiance, and mean temperature). Each column of Q is a mutually orthogonal unit vector, and the order of the columns corresponds one-to-one with the order of the feature vectors obtained in the subsequent solution. The upper triangular matrix R is arranged in the same way as Q, with its row and column dimensions also corresponding to the order of the five feature dimensions. All elements below the main diagonal of R (i.e., row index greater than column index) are 0, and all elements above the main diagonal (row index less than or equal to column index) are 0. The elements of the index are the calculated values ​​generated during the decomposition process; then, the upper triangular matrix R and the orthogonal matrix Q are multiplied in the order of R first and Q last to obtain a new 5×5 matrix. This new matrix is ​​used as the input matrix for the next iteration. The operation of multiplying the 5×5 orthogonal matrix Q obtained by orthogonal triangular decomposition with the 5×5 upper triangular matrix R to obtain a new matrix is ​​repeated. After each iteration, a set of eigenvalue approximations are obtained. The iteration continues until the difference between the eigenvalue approximations obtained in two adjacent iterations is less than the set convergence threshold 1e-6. The iteration stops and the final eigenvalues ​​are output. At the same time, based on the orthogonal matrix Q in each iteration, the eigenvectors corresponding to each eigenvalue are obtained. These eigenvectors satisfy the relationship that the product of the first covariance matrix and the vector is equal to the product of the vector and the corresponding eigenvalue.

[0039] After completing the eigenvalue decomposition, all the obtained eigenvalues ​​are sorted in descending order. At the same time, the cumulative contribution rate of the eigenvalues ​​is calculated (calculated by dividing the sum of the first n eigenvalues ​​by the sum of all eigenvalues). When the cumulative contribution rate reaches 90%, the number of eigenvalues ​​corresponding to this point is recorded, and this number is determined as the value of k (i.e., k is the smallest positive integer that makes the cumulative contribution rate of the eigenvalues ​​≥ 90%). The eigenvectors corresponding to these k eigenvalues ​​are retrieved, and these k eigenvectors are combined as basic vectors to construct the first independent feature subspace. This subspace can accurately represent the dynamic relationship between the photovoltaic grid connection point and the energy storage access point. Next, the same eigenvalue decomposition operation is performed on the second covariance matrix: the QR algorithm is used for decomposition, with a convergence threshold set to 1e-6. The second covariance matrix is ​​also a 5×5 square matrix (the dimension order is consistent with the first covariance matrix). Therefore, the orthogonal matrix Q and the upper triangular matrix R obtained by decomposition are also 5×5 matrices, and the arrangement rules are exactly the same as those of Q and R obtained by decomposing the first covariance matrix (rows and columns correspond to the order of the 5 eigendimensional dimensions, the columns of Q are orthogonal unit vectors, and the elements below the main diagonal of R are 0). The specific solution process is also the same: orthogonal triangular decomposition to obtain the 5×5 orthogonal matrix Q and the 5×5 upper triangular matrix R, matrix multiplication to obtain the new matrix, and iteration. Perform decomposition and multiplication operations until the eigenvalues ​​converge and the corresponding eigenvectors are solved. After obtaining all eigenvalues ​​and corresponding eigenvectors of the second covariance matrix, sort the eigenvalues ​​from largest to smallest and calculate the cumulative contribution rate. Take the smallest positive integer corresponding to a cumulative contribution rate ≥ 90% as the value of L. Combine the corresponding L eigenvectors to construct a second independent eigenspace that can characterize the dynamic correlation between the photovoltaic grid connection point and the main load point. Finally, perform eigenvalue decomposition on the third covariance matrix, also using the QR algorithm, with the convergence threshold set to 1e-6. The third covariance matrix is ​​a 5×5 square matrix (the dimension order remains unchanged). The positive eigenvalues ​​obtained from the decomposition are... Both the intersection matrix Q and the upper triangular matrix R are 5×5 matrices, and the arrangement rules are kept consistent (rows and columns correspond to the order of the 5 feature dimensions, Q column vectors are orthogonal units, and R below the main diagonal is 0). Following the same orthogonal triangular decomposition, iterative operation, convergence judgment, and eigenvalue and eigenvector solution process, the eigenvalues ​​and eigenvectors of the third covariance matrix are obtained. After sorting the eigenvalues ​​from largest to smallest, the cumulative contribution rate is calculated, and the smallest positive integer corresponding to the cumulative contribution rate ≥ 90% is taken as the value of J. The corresponding J eigenvectors are retrieved and combined to construct a third independent feature subspace that can characterize the dynamic correlation between the photovoltaic grid connection point and the grid interaction point.

[0040] This embodiment constructs a multi-dimensional state space by synchronously arranging four high-dimensional state vectors at time points, achieving precise matching of the operational status data of multiple topological nodes in the time dimension and avoiding biases in correlation analysis caused by time misalignment. Calculating the covariance matrix between the photovoltaic grid-connected point's state vector and the other three nodes quantifies the linear correlation between the benchmark node and each node, capturing the collaborative patterns of node operational status changes. Compared to single-point data analysis, this approach is more systematic and correlated. Eigenvalue decomposition of the covariance matrix and extraction of eigenvectors corresponding to the first few largest eigenvalues ​​to construct independent feature subspaces effectively filters redundant information and noise interference in the covariance matrix, extracting the core features of inter-node correlations and improving the accuracy and effectiveness of correlation representation. The three independent feature subspaces correspond to the dynamic correlation between the photovoltaic grid-connected point and the other three nodes, propelling the entire data optimization process from data integration to in-depth correlation feature mining.

[0041] In a preferred embodiment of the present invention, step 2 includes:

[0042] Step 200: Calculate the minimum angle between the principal component vector sets of the first and second independent feature subspaces, as the first generalized angle; calculate the minimum angle between the principal component vector sets of the first and third independent feature subspaces, as the second generalized angle; calculate the minimum angle between the principal component vector sets of the second and third independent feature subspaces, as the third generalized angle. Specifically, this includes: First, clarifying that the principal component vector set of each independent feature subspace is the core feature vector set used when constructing that subspace. The principal component vector set of the first independent feature subspace is the k feature vectors extracted in step 103, the principal component vector set of the second independent feature subspace is L feature vectors, and the principal component vector set of the third independent feature subspace is J feature vectors. The core of calculating the generalized angle is to solve for all pairwise... To calculate the minimum angle between vector combinations, we first construct all possible vector combinations, totaling k×L combinations, for the k principal component vectors of the first independent feature subspace and the L principal component vectors of the second independent feature subspace. For each pair of vectors to be calculated, we multiply the elements of the corresponding dimensions of the two vectors and sum all the results. We then square the elements of each dimension of each vector, sum the results, and take the square root of the sum. We calculate the cosine of the angle between the two vectors by dividing the dot product obtained in the first step by the product of the magnitudes of the two vectors. We convert the cosine value obtained by the inverse cosine function into the corresponding angle, with the unit being degrees. After calculating the angle between all k×L vectors, we iterate through all the results and select the angle with the smallest value as the first generalized angle. This angle represents the minimum correlation angle between the photovoltaic grid connection point-energy storage access point correlation subspace and the photovoltaic grid connection point-main load point correlation subspace.

[0043] Using the same calculation logic as the first generalized angle, we first construct all combinations of k principal component vectors of the first independent feature subspace and J principal component vectors of the third independent feature subspace, totaling k×J combinations. Then, we successively calculate the dot product, magnitude, cosine value, and angle for each vector combination. Finally, we select the minimum value from all the calculated angles and take it as the second generalized angle. This angle represents the minimum correlation angle between the photovoltaic grid connection point-energy storage access point association subspace and the photovoltaic grid connection point-grid interaction point association subspace. Maintaining the same calculation logic as the previous two, we first construct all combinations of L principal component vectors of the second independent feature subspace and J principal component vectors of the third independent feature subspace, totaling L×J combinations. Then, we successively complete the calculation process of dot product, magnitude, cosine value, and angle for each combination. Finally, we select the minimum value from all the angle results as the third generalized angle. This angle represents the minimum correlation angle between the photovoltaic grid connection point-main load point association subspace and the photovoltaic grid connection point-grid interaction point association subspace.

[0044] Step 201: Compare the first, second, and third generalized angles with preset angle thresholds. Define the regions in the multidimensional state space that satisfy the condition that the first generalized angle is greater than the angle threshold and both the second and third generalized angles are less than the angle thresholds as first-class decision subdomains; define the regions in the multidimensional state space that satisfy the condition that the second generalized angle is greater than the angle threshold and both the first and third generalized angles are less than the angle thresholds as second-class decision subdomains; define the regions in the multidimensional state space that satisfy the condition that the third generalized angle is greater than the angle threshold and both the first and second generalized angles are less than the angle thresholds as second-class decision subdomains; and define the regions in the multidimensional state space that satisfy the condition that the third generalized angle is greater than the angle threshold and both the first and second generalized angles are less than the angle thresholds. The region of values ​​is divided into the third type of decision subdomain, which specifically includes: First, a preset angle threshold of 60 degrees is set. This threshold is a boundary value derived from an in-depth analysis of the correlation characteristics of the park's energy system. That is, from the quantitative logic of vector correlation, the size of the angle between two vectors directly reflects their correlation strength. The smaller the angle, the higher the projection overlap between the vectors, and the stronger the operational state coordination between the corresponding nodes (i.e., strong correlation); the larger the angle, the lower the projection overlap, and the weaker the coordination (i.e., weak correlation). When the angle is 60 degrees, the corresponding vector cosine value is 0.5, which is the key dividing point for vector projection overlap. When the cosine value is greater than 0.5 (angle less than 60 degrees), the information overlap between vectors exceeds 50%, indicating a strong correlation between nodes; when the cosine value is less than 0.5 (angle greater than 60 degrees), the information overlap is less than 50%, indicating a weak correlation. Considering the actual operating characteristics of the park's energy system, changes in operating conditions such as equipment switching and load fluctuations cause changes in node correlations, and the corresponding generalized angle changes significantly exceed the critical value of 60 degrees. Therefore, setting this angle as a threshold can accurately and stably distinguish between strong and weak correlations between different nodes. Then, step 200 is calculated... The first, second, and third generalized included angles are compared with the 60-degree threshold. Based on the comparison results, the multidimensional state space is divided into targeted regions. The specific division rules are as follows: In the multidimensional state space, regions that satisfy the condition that the first generalized included angle is greater than 60 degrees and the second and third generalized included angles are both less than 60 degrees are selected. These regions are classified as the first type of decision subdomain. The core characteristics of this region are that the correlation between the photovoltaic grid connection point and the energy storage access point, and the correlation between the photovoltaic grid connection point and the grid interaction point are both strong, while the correlation between the photovoltaic grid connection point and the main load point is weak. This corresponds to the energy operating condition in the park where energy storage is the dominant mode of operation.

[0045] In the multidimensional state space, regions satisfying a second generalized angle greater than 60 degrees and both the first and third generalized angles less than 60 degrees are selected and classified as the second type of decision subdomain. The core characteristics of this region are strong correlations between the photovoltaic grid-connected point and the energy storage access point, and strong correlations between the photovoltaic grid-connected point and the main load point, while the correlation between the photovoltaic grid-connected point and the grid interaction point is weak, corresponding to the energy condition dominated by grid interaction in the park. In the multidimensional state space, regions satisfying a third generalized angle greater than 60 degrees and both the first and second generalized angles less than 60 degrees are selected and classified as the third type of decision subdomain. The core characteristics of this region are strong correlations between the photovoltaic grid-connected point and the energy storage access point, and strong correlations between the photovoltaic grid-connected point and the grid interaction point, while the correlation between the photovoltaic grid-connected point and the main load point is weak, corresponding to the energy condition dominated by load consumption in the park.

[0046] Step 202: For each divided decision subdomain, orthogonally project the high-dimensional state vectors of the energy storage access point, main load point, and grid interaction point onto the principal plane of the corresponding decision subdomain to obtain three projected coordinate points. These three projected coordinate points are then set as three representative points within the corresponding decision subdomain. Specifically, after the decision subdomain is divided, since the high-dimensional state vectors of each node contain a lot of dimensional information, directly using them for energy transmission relationship analysis would increase complexity. Therefore, orthogonal projection is needed to reduce the dimensionality of the high-dimensional data. Simultaneously, the core operating state characteristics of each node within the corresponding subdomain are extracted. Finally, representative points are set to provide accurate and concise node anchors for the subsequent construction of energy transmission links. The specific implementation process is as follows: First, the principal plane of each decision subdomain is defined. The core function of the principal plane is to accurately carry the core correlation characteristics of the corresponding subdomain. Its construction logic is strongly correlated with the dominant operating conditions of the subdomain. The principal plane is formed by the strongly correlated nodes within the subdomain corresponding to the core... The associated feature vectors jointly span ("spanning" means constructing a two-dimensional plane through two linearly independent core vectors, which constitute the basis of the plane, and all points in the plane can be represented by a linear combination of these two basis vectors). For example, the dominant operating condition of the first type of decision subdomain is energy storage operation, corresponding to the strongly correlated node pairs of photovoltaic grid connection point and energy storage access point, and photovoltaic grid connection point and grid interaction point. Therefore, its principal plane is jointly spanned by two linearly independent core vectors selected from the principal component vectors of the first independent feature subspace (photovoltaic grid connection point - energy storage access point) and the third independent feature subspace (photovoltaic grid connection point - grid interaction point). Similarly, the principal plane of the second type of decision subdomain (grid interaction dominant) is spanned by the core vectors of the first independent feature subspace and the second independent feature subspace, and the principal plane of the third type of decision subdomain (load consumption dominant) is spanned by the core vectors of the first independent feature subspace and the third independent feature subspace.

[0047] After defining the principal plane, orthogonal projection and representative point setting operations are performed for each decision subdomain. For the high-dimensional state vectors of the energy storage access point, main load point, and grid interaction point, their orthogonal projections on the principal plane of the current decision subdomain are calculated one by one (the core of orthogonal projection is to retain the core information of the vector in the principal plane direction and remove redundant information perpendicular to the principal plane). The specific steps are as follows: First, from the core correlation feature vectors of the subdomain, two linearly independent vectors that can best represent the correlation characteristics of the subdomain are selected as the orthogonal basis vectors of the principal plane (if the selected core vectors...). If the orthogonality is not satisfied, it needs to be converted into an orthogonal vector through Schmidt orthogonalization to ensure the accuracy of subsequent projection calculations. The second step is to decompose the high-dimensional state vector of the target to be projected into the directions of the two orthogonal basis vectors, calculating the projection component in each direction. The calculation logic of the projection component is to accurately lock the effective information of the target vector in the direction of the basis vector. The specific calculation process is as follows: first, calculate the dot product between the target vector and the basis vector (the dot product result reflects the degree of correlation between the two vectors in the same direction); then divide the dot product result by the square of the magnitude of the basis vector (eliminating the basis vector length). The influence of degree on the projection result is considered. Finally, the result is multiplied by the basis vector to obtain the projection component of the target vector in the direction of the basis vector (this component is a vector in the same direction as the basis vector). In the third step, the projection components of the target vector in the directions of the two orthogonal basis vectors are vector-added. The resulting vector is the orthogonal projection vector of the target high-dimensional state vector on the current principal plane. This vector completely retains the core information of the target vector in the direction of the principal plane, while reducing the dimension from high-dimensional to two-dimensional. In the fourth step, a two-dimensional coordinate system is established with the two orthogonal basis vectors of the principal plane as coordinate axes. The orthogonal projection vector obtained above is substituted into the coordinate system and converted into the corresponding two-dimensional coordinate points. These coordinate points are the orthogonal projection coordinate points of the target high-dimensional state vector on the principal plane. The projection coordinate points of the energy storage access point, the main load point, and the grid interaction point obtained above are directly set as three representative points in the current decision subdomain. These three representative points are not arbitrarily selected, but are a condensed representation of the core operating state of each node under the subdomain conditions. Their coordinate information accurately corresponds to the core features of the node's high-dimensional state vector related to the dominant operating conditions of the subdomain.

[0048] Step 203: Within each decision subdomain, starting from the baseline state vector point of the corresponding decision subdomain, connect it to three representative points of the corresponding decision subdomain to form a local virtual energy conduction link within the corresponding decision subdomain. Combine the local virtual energy conduction links of all decision subdomains to form a multi-level virtual energy guidance map that reflects the cross-regional conduction relationship of energy in the multi-dimensional state space. Specifically, this includes: after setting the baseline point (projection point of photovoltaic grid connection point) and representative points of each decision subdomain, local energy conduction links can be constructed based on these nodes. Then, by integrating the links of each subdomain, a multi-level virtual energy guidance map covering all operating conditions is formed, which intuitively and systematically presents the energy conduction path and correlation between nodes under different operating conditions. The specific implementation process is as follows: Within each decision subdomain, the starting and ending points of the link are first determined. The starting point is the reference state vector point of the subdomain. Since the photovoltaic grid connection point is the core reference node of the entire park's energy system, the reference point is the orthogonal projection point of the high-dimensional state vector of the photovoltaic grid connection point on the main plane of the current subdomain. Its calculation method is completely consistent with the calculation method of the orthogonal projection of the target high-dimensional vector in step 202, ensuring that the reference point and the three representative points are in the same main plane and the same coordinate system, thus ensuring the rationality of the link connection. The ending point is the projection coordinate point of the three representative points of the subdomain (energy storage access point, main load point, and grid interaction point). Starting from the reference state vector point, straight line segments are used to connect to the three representative points respectively. Each straight line segment constitutes a local virtual energy transmission link. The direction of the link is clearly defined as from the reference point to the representative point. This direction setting fits the actual energy flow logic of the park. The photovoltaic grid connection point is the core energy node. The electricity generated by it is transmitted to the energy storage access point (energy storage), the main load point (load consumption), and the grid interaction point (grid connection or power purchase). The link itself intuitively represents the path of energy transmission from the photovoltaic grid connection point to the other three nodes under the operating conditions of this subdomain.

[0049] After all local virtual energy transmission links in all decision subdomains are constructed, the entire link is summarized and integrated. Since different decision subdomains correspond to different energy operation conditions in the park (energy storage-dominated, grid interaction-dominated, and load consumption-dominated), the local links of each subdomain reflect the energy transmission patterns under the corresponding operation conditions. During the integration process, the operation condition attributes of each subdomain link are retained, and the correlation mapping between links in different subdomains is established (for example, when the park's operation condition switches from energy storage-dominated to grid interaction-dominated, the corresponding energy transmission link will smoothly transition from the first type of subdomain link to the second type of subdomain link). The resulting multi-level virtual energy guidance map presents a structure that is layered by operation condition and correlated across operation conditions. It can clearly show the energy transmission path within a single operation condition area and intuitively reflect the changing patterns of energy transmission relationships when switching between different operation conditions, fully covering the energy interaction logic under all energy operation conditions in the park.

[0050] This embodiment quantifies the correlation strength between two independent feature subspaces using a generalized angle, achieving an intuitive representation of the dynamic correlation between nodes. Compared to qualitative analysis, it can more accurately distinguish the differences in operating conditions corresponding to different correlation strengths. The decision subdomain division based on the generalized angle and preset threshold can decompose the complex multidimensional state space into clear sub-regions according to the dominant energy operation type, simplifying the analysis difficulty of multi-node correlation under all operating conditions and making it easier to capture the energy transmission patterns under different operating conditions. Orthogonal projection maps high-dimensional state vectors onto the main plane of the decision subdomain, which not only retains the core features of the node operating state but also achieves dimensionality reduction and simplification of high-dimensional data. The set representative points can accurately anchor the core operating state of each node under the corresponding operating conditions. The multi-level virtual energy guidance map intuitively presents the energy transmission path and cross-regional correlation under different operating conditions, clearly depicting the energy interaction logic between the photovoltaic grid connection point and other nodes. It can effectively support the accurate identification of systemic data mutations caused by equipment switching and make up for the deficiency of ignoring the changes in multi-node correlation.

[0051] In a preferred embodiment of the present invention, step 3 includes:

[0052] Step 300: Based on the geometric relationships of each link in the multi-level virtual energy guidance diagram, calculate the first path conduction coefficient of the link connecting the reference state vector point and the energy storage access point representative point, the second path conduction coefficient of the link connecting the reference state vector point and the main load point representative point, and the third path conduction coefficient of the link connecting the reference state vector point and the grid interaction point representative point. Specifically, after completing the construction of the multi-level virtual energy guidance diagram, it is necessary to quantify the energy conduction intensity based on the geometric relationships of each link in the diagram, that is, calculate the path conduction coefficient. The core function of the path conduction coefficient is to characterize the energy conduction from the photovoltaic grid connection point (reference point) to the other three nodes (representative points). The effectiveness of the link is directly related to its geometric characteristics. The specific process is as follows: First, extract the core geometric parameters of each local virtual energy conduction link. For the three links in each decision subdomain (reference point-energy storage access point representative point, reference point-main load point representative point, and reference point-grid interaction point representative point), extract the link segment length and the angle between the link direction and the core feature vector of the subdomain's main plane (i.e., the link direction angle). The calculation process for the segment length is as follows: obtain the coordinate values ​​of the two ends of the link (reference point and corresponding representative point) in the two-dimensional coordinate system of the main plane, and calculate the coordinate difference between the two points in the x-axis direction and the coordinate difference in the y-axis direction. The length of the link segment is obtained by squaring the two coordinate differences, summing them, and then taking the square root of the sum. The link direction angle is calculated by taking one of the orthogonal basis vectors of the principal plane as the positive x-axis and calculating the angle between the link segment and the positive x-axis. Specifically, the tangent value is calculated from the coordinate difference between two points, and then converted to an angle (in degrees) using the arctangent function. Next, the transmission coefficients of the three paths are calculated based on the extracted geometric parameters. The calculation logic is that the shorter the segment length and the smaller the angle between the link direction and the core feature vector, the smaller the energy transmission resistance and the larger the transmission coefficient. The specific calculation process is to take the reciprocal of the link segment length and multiply it by the subdomain principal plane. The characteristic scale coefficient (determined by the average magnitude of the two orthogonal basis vectors of the principal plane, used to eliminate the influence of scale differences in the principal plane of different subdomains on the coefficient) is used to obtain the length influence factor; the cosine value of the link direction angle is taken (the closer the cosine value is to 1, the more aligned the link direction is with the core characteristic vector, and the smoother the energy transmission) to obtain the direction influence factor; the length influence factor and the direction influence factor are multiplied together, and then multiplied by a preset benchmark coefficient (the benchmark coefficient is 0.85, which is determined based on the rated power and transmission efficiency of the park's energy system, combined with the rated power range and conventional power transmission efficiency level of most industrial and commercial park low-voltage power distribution systems, 0.85 can realize the conversion of geometric parameters into quantified values ​​of energy conduction intensity, ensuring that the numerical range of the conduction coefficient matches the actual energy conduction efficiency, thus obtaining the final path conduction coefficient. Following the above process, the first path conduction coefficient connecting the reference state vector point and the energy storage access point, the second path conduction coefficient connecting the reference state vector point and the main load point, and the third path conduction coefficient connecting the reference state vector point and the grid interaction point are calculated respectively. These three conduction coefficients within the same decision subdomain are interconnected and jointly reflect the energy conduction balance under this operating condition.

[0053] Step 301: Continuously monitor the values ​​of the first path conduction coefficient, the second path conduction coefficient, and the third path conduction coefficient; when the instantaneous change of any path conduction coefficient exceeds the preset mutation threshold, an abnormal mutation in the conduction coefficient is determined, and the time of occurrence of the abnormal mutation is recorded. Specifically, to promptly capture systemic data mutations caused by equipment switching, it is necessary to continuously monitor the calculated first, second, and third path conduction coefficients in real time. The specific implementation process is as follows: First, preset the mutation threshold. The mutation threshold is determined based on the maximum fluctuation of the conduction coefficient under normal operating conditions in the historical operating data of the park. Specifically, the standard deviation of the conduction coefficient in the historical data is calculated. Combined with the actual operating characteristics of the park's energy system, the standard deviation of the path conduction coefficient under normal operating conditions is approximately 6.7%. Therefore, three times the standard deviation, i.e., 20%, is used as the mutation threshold (this value can effectively distinguish normal). (To prevent misjudgments regarding fluctuations and anomalous changes); subsequently, a continuous monitoring process is initiated, acquiring the instantaneous values ​​of the conduction coefficients of each path in real time in chronological order. The change range of each instantaneous value compared to the previous value is calculated (the change range is calculated by first calculating the absolute value of the difference between the current value and the previous value, then dividing the absolute value of the difference by the previous value to obtain the relative change range, expressed as a percentage). Finally, the calculated relative change range is compared with a preset 20% mutation threshold. If the relative change range of any path conduction coefficient exceeds 20%, an anomalous mutation in the conduction coefficient is directly determined. Simultaneously, the precise time of the anomalous mutation is recorded (accurate to the same time unit as the original data acquisition cycle), and the topology node corresponding to the abnormal path conduction coefficient is marked (e.g., if the first path conduction coefficient is abnormal, the energy storage access point is marked).

[0054] Step 302: Match the occurrence time of the abnormal mutation with the time stamp recorded in the switching event log. If the match is successful, then based on the value of the path conduction coefficient in the virtual energy steering diagram at the time of occurrence, perform data correction on the original power time series section of the topology node corresponding to the path conduction coefficient of the abnormal mutation at the time of occurrence. After the correction is completed, the reconstructed power data is obtained. Specifically, after determining the abnormal mutation of the conduction coefficient, it is necessary to verify the root cause of the anomaly by combining the switching event log, and accurately correct the original power data of the abnormal node based on the multi-node collaborative association characteristics. Finally, the reconstructed power data is obtained. The specific process is as follows: First, retrieve the static var generator switching event log collected and stored in step 001, and extract the precise time of all device state transitions (accurate to the second, compared with the original data) from the log. The core information, such as the consistent acquisition cycle, the state before and after the conversion, is then precisely matched with the state transition time of the extracted switching event recorded in step 301. If the difference between the time of the abnormal mutation and the state transition time of a certain switching event is within the preset time error allowable range (the time error range does not exceed 1 second, i.e., 1 time the original data acquisition cycle, to ensure that the matching result is not affected by the data acquisition timing deviation), then the matching is considered successful. A successful matching can directly indicate that the abnormal mutation of the conduction coefficient is a systemic data mutation caused by the state transition (switching event) of the static var generator. If no switching event that meets the error requirements is found, the conduction coefficient calculation process, monitoring and judgment logic and the quality of the original data acquisition need to be re-examined to eliminate problems such as calculation errors or data acquisition failures.

[0055] After successful matching, based on the coordinated correlation of energy conduction among topological nodes under normal operating conditions, and using the path conduction coefficients and corresponding node power data without abnormal mutations as a benchmark, the original power time-series cross-sectional data of abnormal nodes are corrected. The specific steps are as follows: First, benchmark data extraction and confirmation. First, the data source is clarified. The two path conduction coefficients without abnormalities come from the path conduction coefficient sequences at various times calculated and stored in real time in step 300 (this sequence is recorded in chronological order, synchronized with the original data acquisition cycle, and the value range is 0.3 to 0.9, which conforms to the normal conduction efficiency range of the park's energy system); the corresponding topological... The original power time-series profile values ​​of the nodes are derived from the instantaneous active power value sequence of each node collected and stored in step 001, such as 0 to 500kW for photovoltaic grid-connected points and 100 to 800kW for main load points. Subsequently, three sets of core data corresponding to the time of anomaly occurrence (denoted as t0) are precisely extracted: first, the instantaneous values ​​of the conduction coefficients of the two paths without anomalies (for example, if the conduction coefficient of the first path is abnormal, the instantaneous value of the conduction coefficient of the second path is extracted as 0.72, and the instantaneous value of the conduction coefficient of the third path is extracted as 0.68); second, the topology nodes corresponding to these two paths without anomalies (such as the main load point corresponding to the second path and the power grid corresponding to the third path). The first step is to construct a linear regression correction model. This involves considering the following parameters: 1) the original power time-series profile values ​​at time t0 (e.g., the original power at the main load point at time t0 is 420kW, and the original power at the grid interaction point at time t0 is 180kW); 2) the abnormal values ​​of the original power time-series profiles of the topology nodes corresponding to the abnormal paths (e.g., the energy storage access point corresponding to the first path is 250kW, an abnormal value caused by switching events); 3) the abnormal values ​​of the original power time-series profiles of the topology nodes corresponding to the abnormal paths (e.g., the original power at the energy storage access point at time t0 is 250kW, an abnormal value caused by switching events); 4) the linear regression correction model is constructed based on the characteristic that the path conduction coefficient and the corresponding node power are linearly correlated under normal operating conditions, using the conduction coefficients of two paths without abnormalities and the corresponding node power. Using numerical values ​​as samples, a linear regression equation is fitted and constructed as the basis for correcting abnormal power values. The specific form of the linear regression equation is y = ax + b, where the dependent variable y is the abnormal node power value to be corrected (unit: kW); the independent variable x is the path conduction coefficient of the abnormal path (unitless, ranging from 0.3 to 0.9); the regression coefficient a is the slope, representing the weight of the influence of the conduction coefficient on the node power; the constant term b is the intercept, representing the baseline power offset when there is no conduction loss; two sets of extracted sample data are used, such as (second path conduction coefficient 0.72, main load point power 420kW) and (third path conduction coefficient 0.72).Substituting 68 (power at the grid interconnection point is 180kW) into the equation, the specific values ​​of the regression coefficient a and the constant term b are calculated using the least squares method to ensure that the model can accurately represent the linear correlation between the conduction coefficient and the node power under normal operating conditions; the third step is to calculate the theoretical normal value of the abnormal path conduction coefficient; the theoretical normal value of the abnormal path conduction coefficient at time t0 is obtained by extrapolating the time change trend of the non-abnormal path conduction coefficient. Specifically, the conduction coefficient values ​​of the two non-abnormal paths are extracted from 10 normal times before and after time t0 (i.e., times when no abnormality or switching event occurs), and the time change trend curves of these two coefficients are fitted respectively (such as linear curves). The trends are defined as Y = k1t + c1 and y = k2t + c2, where Y represents the value of the transmission coefficient, t represents time, k1 and k2 are the slopes of the time-varying trend curves of the transmission coefficients of the two non-abnormal paths, specifically representing the rate of change of the corresponding transmission coefficients over time. A positive slope indicates that the transmission coefficient gradually increases over time, while a negative slope indicates that the transmission coefficient gradually decreases over time. The larger the absolute value of the slope, the faster the rate of change. c1 and c2 are the intercepts of the two trend curves, specifically representing the baseline offset of the corresponding transmission coefficient at the beginning of the fitting time interval, i.e., the baseline value of the transmission coefficient at the beginning of the fitting interval. Based on the two curves... The variation pattern is used to comprehensively deduce the theoretical value of the abnormal path conduction coefficient at time t0 under normal operating conditions (for example, the extrapolated theoretical normal value of the first path conduction coefficient at time t0 is 0.65). This value eliminates the interference of switching events and reflects the true conduction intensity under normal operating conditions. In the fourth step, the extrapolated theoretical normal value of the abnormal path conduction coefficient (e.g., 0.65) is substituted into the linear regression equation constructed in the second step to calculate the corrected power value of the abnormal node at time t0. Subsequently, in the original power time series section of the abnormal node, the abnormal value at time t0 (e.g., 250kW) is replaced with the calculated corrected power value (364kW), completing the process. Precise correction of anomalous data at single moments: For all moments with abnormal abrupt changes in conduction coefficient, the correction operations described in steps two through four above are repeated to ensure that all abnormal power data affected by switching events are corrected. After correction, the corrected power time-series data of each topology node (PV grid connection point, energy storage access point, main load point, grid interaction point) is summarized and spliced ​​with the original power time-series data at moments without anomalies in chronological order to form complete reconstructed power data. This reconstructed power data completely eliminates the systemic data mutation interference caused by static var generator switching events, and fully and accurately reflects the true operating status of the park's energy system.

[0056] In this embodiment, the path conduction coefficient transforms the energy conduction relationship into a quantifiable numerical indicator, enabling precise characterization of the energy interaction intensity under different operating conditions. Compared to qualitative descriptions of energy conduction, this provides a clear quantitative basis for subsequent anomaly monitoring. By determining the mutation threshold based on historical data and combining it with a monitoring and judgment method based on relative change amplitude, it can identify systemic conduction coefficient anomalies caused by switching events, effectively avoiding confusion between normal fluctuations and abnormal mutations, and improving the accuracy and reliability of anomaly identification. By matching the anomaly moment with the switching event log, the root cause of the abnormal mutation is clarified, ensuring the targeted nature of data correction. At the same time, the use of the synergistic correlation between the normal conduction coefficient and node power for correction ensures the rationality of the corrected power data, compensating for the shortcomings of ignoring multi-node correlation, directly discarding abnormal data, or simple interpolation. The final reconstructed power data eliminates the interference of systemic data mutations and completely preserves the true operating status information of the energy system.

[0057] In a preferred embodiment of the present invention, step 4 includes:

[0058] Step 400 involves organizing the reconstructed power data into a success rate training sequence according to time order, and organizing the synchronized meteorological time series sections into a meteorological training sequence. The power training sequence and the meteorological training sequence are then used to train the integrated attention mechanism time series prediction model, resulting in a trained time series prediction model. Specifically, this includes: First, arranging the reconstructed power data obtained in step 302 in chronological order to form a power training sequence. The time step of this sequence is consistent with the original data acquisition cycle. Each time step corresponds to the reconstructed active power value of a topological node. The sequence length covers historical data from the complete operating cycle of the park's energy system, ensuring it includes power variation characteristics under different operating conditions (such as sunny days, cloudy days, peak load, and off-peak load). Second, retrieving the complete meteorological time series section data collected and stored in step 001, synchronizing and aligning it according to the same time axis as the power training sequence, and removing time-misaligned data points to form a meteorological training sequence. Each time step of this sequence contains two core meteorological elements: the average irradiance and the average temperature, which are consistent with the power training sequence. To accurately capture the temporal correlation characteristics of power data and the cross-modal correlation between meteorological and power data, a temporal neural network was selected as the basic framework. This framework integrates self-attention and cross-attention mechanisms to construct a temporal prediction model with dual-branch input and single-branch output. The specific construction process is as follows: Two parallel input branches are set up. The first branch is used to input power training sequences or historical power sequences, with the input dimension consistent with the feature dimension of the power sequence. The second branch is used to input meteorological training sequences or historical meteorological sequences, with the input dimension consistent with the feature dimension of the meteorological sequence, ensuring that both types of data can be accurately input into the model for subsequent processing. Independent feature encoding layers are configured for each of the two input branches. Each encoding layer contains a one-dimensional convolutional layer and a time-step embedding layer. The kernel size of the one-dimensional convolutional layer is set to 3, and the stride is set to 1, used to extract local temporal correlation features of the sequence. The time-step embedding layer converts the time-step information of the sequence into a high-dimensional embedding vector, and then adds it element-wise to the convolutional feature vector to enhance the temporal discriminativeness of the features, ultimately outputting a high-dimensional feature vector sequence.

[0059] The self-attention layer is placed after the power feature encoding layer. Its input is a sequence of power feature vectors. Three learnable mapping matrices (query mapping matrix Ma, key mapping matrix Mb, and value mapping matrix Mc) map the power feature vectors to query vector U, key vector Z, and value vector Y, respectively, capturing the long-term correlation characteristics of the power data. The arrangement of the three mapping matrices is explicitly defined: the number of rows in query mapping matrix Ma is consistent with the dimension of the power feature vectors, and the number of columns is the preset attention feature dimension; the row and column dimensions of key mapping matrices Mb and value mapping matrix Mc are exactly the same as those in query mapping matrix Ma, ensuring uniform dimension after vector mapping and meeting the matching requirements for attention calculation. The cross-attention layer is placed after the self-attention layer. Its input is the power feature vector sequence output from the self-attention layer and the meteorological feature vector sequence output from the meteorological feature encoding layer. Specifically, the power feature vectors are mapped to query vector U through query mapping matrix Ma, and the meteorological feature vectors are mapped to key vector Z and value vector Y through key mapping matrix Mb and value mapping matrix Mc, respectively. This cross-modal mapping is used to mine the correlation characteristics between meteorological and power data. The three mapping matrices here are arranged as follows: the number of rows in the query mapping matrix Ma is consistent with the dimension of the power feature vector, and the number of columns is the dimension of the attention feature; the number of rows in the key mapping matrix Mb and the value mapping matrix Mc are consistent with the dimension of the meteorological feature vector, and the number of columns is the same as the number of columns in the query mapping matrix Ma, ensuring the dimensionality matching after cross-modal vector mapping.

[0060] The fusion layer is positioned after the attention mechanism layer. It receives the power feature sequence optimized by the attention layer and the meteorological feature sequence optimized by the cross-attention layer. It performs feature fusion by adding elements one by one and outputs the fused feature, achieving deep integration of the two types of feature information. The output layer is set as a fully connected network structure. The input is the fused feature. The output layer adopts a parallel prediction structure, which can simultaneously output the net load power prediction values ​​for multiple consecutive time points in the future, as well as the probability distribution parameters (mean and variance) corresponding to each prediction value. After the model is built, the constructed power training sequence is used as the first branch of the model input features, and the meteorological training sequence is used as the second branch of the input features. At the same time, the training objective of the model is set to output the net load power values ​​for each consecutive time point in the next 24 hours. This duration covers the complete cycle of daily scheduling of the park's energy system and can meet the actual scheduling needs.

[0061] During training, mean squared error is used as the loss function, followed by iterative parameter updates using a mini-batch gradient descent algorithm. The algorithm's specific execution process is as follows: First, the training set is divided into multiple fixed-size batches, each containing 64 samples to ensure stability and efficiency during training. Then, the data is input into the model batch by batch, and the loss function value for each batch is calculated via forward propagation. Next, based on the loss function value, the gradient of all learnable parameters in the model is calculated using the chain rule, with the gradient direction representing the direction in which the parameters increase the loss function. Finally, along the opposite direction of the gradient, all learnable parameters, including the query mapping matrix Ma, key mapping matrix Mb, value mapping matrix Mc, and convolutional kernel weights, are updated at a preset learning rate of 0.005. This iterative process continues until the loss function value converges to a preset threshold of 0.001. At this point, the model has fully learned the relevant correlation patterns, training stops, and the trained time-series prediction model is obtained.

[0062] Step 401a: After training, the time-series prediction model encodes the input power history sequence and meteorological history sequence using time-series features, obtaining power feature sequences and meteorological feature sequences respectively. Then, using the attention mechanism in the trained time-series prediction model, it calculates the self-attention weights of the feature vectors at each time step in the power feature sequence and the cross-attention weights between the power feature sequence and the meteorological feature sequence. Specifically, when the trained time-series prediction model is put into use, it first encodes the input historical data using time-series features, and then calculates the relevant weights through the attention mechanism. The specific process is as follows: First, obtain the historical input data for the period to be predicted, including the power history sequence (reconstructed power data of consecutive time steps before the time to be predicted) and the meteorological history sequence (meteorological data synchronized with the power history sequence). Then, perform feature encoding on the two types of sequences respectively. Input the power history sequence into the power feature encoding layer of the model. First, perform convolution operations through a one-dimensional convolutional layer with a kernel size of 3 and a stride of 1. Slide through the power history sequence to extract each time step. The local correlation features of the irradiance and temperature data at two adjacent time steps are used to obtain local convolutional feature vectors. Then, through a time step embedding layer, the positional information of each time step is converted into an embedding vector with the same dimension as the convolutional feature vector. The embedding vector is then added element-wise to the convolutional feature vector to obtain a high-dimensional feature vector containing time and positional information. Finally, all high-dimensional feature vectors from all time steps are arranged sequentially to form a power feature sequence. Each element of this sequence is the power feature vector for the corresponding time step, containing core information such as the power amplitude and trend at that moment. The historical meteorological sequence is input into the meteorological feature encoding layer of the model, using the same network structure and parameter settings as the power feature encoding layer. Specifically, a one-dimensional convolutional layer with a kernel size of 3 and a stride of 1 is used to extract local meteorological correlation features, which are then incorporated into the time step embedding layer to integrate time and positional information. Finally, the irradiance and temperature data are converted into a high-dimensional vector sequence, i.e., a meteorological feature sequence. Each element of this sequence is the meteorological feature vector for the corresponding time step, containing potential correlation information about the impact of meteorological conditions on power.

[0063] After completing the temporal feature encoding, the self-attention weights and cross-attention weights are calculated respectively. The specific process is as follows: For the power feature sequence, the model calculates the correlation weight between any two time step feature vectors in the sequence through the self-attention mechanism, thereby highlighting historical power features that have an important impact on prediction. Specifically, the first step is to linearly map each feature vector yj in the power feature sequence through the learnable query mapping matrix Ma, key mapping matrix Mb, and value mapping matrix Mc to obtain the corresponding query vector. Key vector Value vector The query vectors, key vectors, and value vectors of all time steps are arranged sequentially to form the query vector sequence U, key vector sequence Z, and value vector sequence Y. The second step is to calculate the dot product of the transposes of the query vector sequence U and the key vector sequence Z to obtain the attention score matrix D. Each element Dmn in this matrix represents the degree of influence of the power feature at time step m on the power feature at time step n. The third step, to avoid excessively large dot product results due to high feature vector dimensions, divides each element of the attention score matrix D by a scaling factor, where the scaling factor is the square root of the key vector dimension fh, to obtain the scaled attention score matrix Dt = D / The fourth step is to perform a softmax function operation on the scaled attention score matrix Dt, converting each element in the matrix into a value between 0 and 1, and the sum of the elements in each row is 1. The resulting matrix is ​​the self-attention weight matrix, and the element Nm in the matrix is ​​the self-attention weight of the feature vector of the m-th time step to the feature vector of the n-th time step, which represents the degree of influence of the power features of different time steps on the current prediction time.

[0064] For power feature sequences and meteorological feature sequences, the model calculates the cross-modal association weights between the two sequences through a cross-attention mechanism, thereby exploring the impact of meteorological features on power prediction. Specifically, the process is as follows: First, the power feature vector output from the self-attention layer is used as the source of the query vector, mapped to a query vector sequence U through a query mapping matrix Ma; each feature vector in the meteorological feature sequence is used as the source of the key vector and value vector, mapped to a key vector sequence Z through a key mapping matrix Mb, and a value vector sequence Y through a value mapping matrix Mc; Second, the dot product of the transpose matrices of the power query vector sequence U and the meteorological key vector sequence Z is calculated to obtain the cross-modal attention score matrix. The elements in this matrix represent the degree of correlation between the power characteristics and meteorological characteristics at a certain time step; the third step is to divide each element in the cross-modal attention score matrix Dc by the square root of the key vector dimension. The scaling operation is completed to avoid the problem of excessively large dot product results. The fourth step is to perform normalization processing, that is, to perform a softmax function operation on the scaled cross-modal attention score matrix Dc to obtain a cross-attention weight matrix. Each element in the matrix has a value range from 0 to 1, and the sum of the elements in each row is 1. This element is the cross-attention weight of the power feature and meteorological feature at the corresponding time step. The magnitude of the weight value directly represents the degree of influence of the meteorological feature at the corresponding time step on the power prediction result. The larger the weight value, the closer the correlation between the meteorological feature and the power change at that time step is, the higher the importance will be given in the subsequent feature fusion process, and the more significantly it can affect the generation of the net load power prediction value. The smaller the weight value, the weaker the correlation between the meteorological feature and the power change at that time step is, the lower the degree of influence on the power prediction result is, and the smaller the contribution ratio in the feature fusion will be.

[0065] Step 401b involves weighted fusion of the power feature sequence and the meteorological feature sequence based on self-attention weights and cross-attention weights to form a fused feature. Specifically, after calculating the attention weights, the power feature sequence and the meteorological feature sequence are weighted and optimized based on the self-attention weights and cross-attention weights obtained in step 401a. The two optimized feature sequences are then fused to form a fused feature that incorporates both types of core information. The specific implementation process is as follows: a matrix multiplication operation is performed between the self-attention weight matrix and the value vector sequence Y of the power feature sequence. That is, the value vector at each time step is multiplied by the self-attention weight corresponding to that time step. Then, all weighted values ​​are multiplied... The vectors are summed to obtain the optimized power feature sequence, which highlights the historical time step features that contribute significantly to power prediction. The cross-attention weight matrix is ​​then multiplied by the value vector sequence Y of the meteorological feature sequence, i.e., the meteorological value vector at each time step is multiplied by the corresponding cross-attention weight. All weighted meteorological value vectors are then summed to obtain the optimized meteorological feature sequence, which highlights meteorological features that significantly influence power prediction. Finally, the optimized power feature sequence and the optimized meteorological feature sequence are added element-wise to obtain a fusion feature that includes both power temporal correlation characteristics and meteorological cross-modal correlation characteristics.

[0066] Step 401c: After training, the time-series prediction model generates net load power prediction values ​​for each consecutive future time point and probability distribution parameters representing the uncertainty of each prediction value in parallel through the output layer based on the fused features. Specifically, this includes: inputting the fused features obtained in step 401b into the output layer of the trained time-series prediction model, completing power prediction through the parallel prediction structure of the output layer, and simultaneously generating probability distribution parameters representing prediction uncertainty. The specific process is as follows: inputting the fused features into the fully connected output layer of the model, which contains the same number of neurons as the number of time points in the next 24 hours. Each neuron corresponds to the calculation of the net load power prediction value for one time point. The calculation process is as follows: first, performing matrix multiplication between the fused features and the weight matrix of the fully connected output layer. The value range of this weight matrix is ​​[-0.1, 0.1]. This matrix is ​​used to complete the linear transformation of the fused features, obtaining the linearly transformed feature vector; then, a bias term is added to this feature vector, with the value of the bias term set to 0.0. 1. Complete basic linear calculations; finally, perform nonlinear mapping on the calculation results using the ReLU activation function to output the net load power prediction value corresponding to each time point; since the output layer adopts a parallel prediction structure, the prediction values ​​of all time points can be calculated synchronously without iterative generation point by point. The number of prediction values ​​is consistent with the preset 24-hour future prediction duration. Each time point corresponds to one prediction value, realizing parallel prediction of power at multiple time steps and meeting the prediction needs of daily energy dispatch in the park; to accurately characterize the uncertainty of each prediction value, the model generates the corresponding probability distribution parameters simultaneously while outputting the power prediction value. The normal distribution is selected as the probability distribution model for the prediction value. The generated probability distribution parameters include the mean and variance corresponding to each prediction value. The mean is the net load power prediction value calculated in the above steps, and the variance is generated by calculating the independent parameter branch of the fully connected output layer. The magnitude of the variance characterizes the dispersion of the prediction value. The larger the variance, the higher the uncertainty of the prediction result.

[0067] Step 402: Calculate the upper and lower confidence bounds of the predicted values ​​for each future time point based on the probability distribution parameters. The upper and lower confidence bounds form the future net load power prediction interval. Specifically, based on the probability distribution parameters obtained in step 401c, use the normal distribution confidence interval calculation method to calculate the upper and lower confidence bounds of the predicted values ​​for each future time point, thereby forming a complete net load power prediction interval covering the future prediction period. The specific process is as follows: Using the normal distribution confidence interval calculation method, a confidence level of 95% is set. According to the statistical characteristics of the normal distribution, the two-sided quantile corresponding to this confidence level is 1.96. As a fixed constant, it represents that 95% of the sample data will fall within the range of the mean ± 1.96 times the standard deviation. For the net load power forecast value at each future time point, the upper confidence bound and the lower confidence bound are calculated. Specifically, the upper confidence bound = net load power forecast value + 1.96 × the square root of the corresponding variance; the lower confidence bound = net load power forecast value - 1.96 × the square root of the corresponding variance. The upper confidence bound and the lower confidence bound of each future time point are paired to form the net load power forecast interval for that time point. The forecast intervals of all time points are arranged in chronological order to obtain a complete net load power forecast interval covering the forecast period of the next 24 hours.

[0068] Step 403: Collect real-time power time-series profiles of actual power generated within the time period covered by the future net load power prediction interval, count the number of data points in the real-time power time-series profiles that fall into the corresponding prediction interval, and calculate the actual coverage ratio. Specifically, this includes: collecting real-time power time-series profile data of actual power generated by the park's energy system within the 24-hour time period covered by the future net load power prediction interval, ensuring that the collection period and time range of the real-time data are completely consistent with the prediction interval, and avoiding the impact of data misalignment on the accuracy of the coverage ratio calculation; for each future time point, compare the real-time power value at that time with the corresponding prediction interval. If the real-time power value is between the upper confidence bound and the lower confidence bound, it is determined that the data point falls into the prediction interval and is marked as successfully covered; if the real-time power value exceeds the upper confidence bound or is lower than the lower confidence bound, it is determined as a failure to cover; and count the coverage ratio. The calculation process for the actual coverage ratio, including the number of successfully covered data points and the total number of data points, is as follows: Actual coverage ratio = Number of successfully covered data points ÷ Total number of data points × 100%. Combined with the expected coverage ratio corresponding to the 95% confidence level preset in step 402, the reliability of the prediction interval is intuitively judged by comparing the actual coverage ratio with the expected coverage ratio. If the actual coverage ratio is close to or reaches 95%, it indicates that the uncertainty characterization of the prediction interval matches the fluctuation characteristics of the real power data, effectively covering the real power value, and the prediction interval has high reliability. If the actual coverage ratio is significantly higher than 95%, it indicates that the prediction interval is set too wide; although the coverage capability is strong, prediction accuracy is sacrificed, and there is room for reliability optimization. If the actual coverage ratio is consistently lower than 95%, it indicates that the uncertainty characterization of the prediction interval is insufficient, failing to effectively cover the real power data, and the prediction interval has poor reliability.

[0069] Step 404: Compare the actual coverage ratio with the preset expected coverage ratio. If the actual coverage ratio is consistently lower than the expected coverage ratio, backpropagate the probability distribution parameters output by the time-series prediction model based on historical prediction error data within a set time period to correct the probability distribution parameters, output new probability distribution parameters, and generate new upper and lower confidence bounds to form an adjusted prediction interval. Specifically, this includes: setting the preset expected coverage ratio to 95%, which is consistent with the confidence level set in step 402; comparing the actual coverage ratio calculated in step 403 with the expected coverage ratio; if the actual coverage ratio is consistently higher than or equal to the expected coverage ratio, it indicates that the reliability of the prediction interval meets the requirements and no adjustment is needed; if the actual coverage ratio is consistently lower than the expected coverage ratio, it indicates that the uncertainty representation of the prediction interval is insufficient and cannot effectively cover real-time power data, requiring correction of the model parameters; when the actual coverage ratio... When the coverage ratio remains below the expected level, historical prediction error data from the past 7 days is extracted (historical prediction error is the difference between the real-time power value and the corresponding predicted value). With the goal of improving the coverage ratio of the prediction interval to the expected level, the error loss is calculated (the loss function is the square of the difference between the actual coverage ratio and the expected coverage ratio). The error loss is then passed from the output layer to the probability distribution parameter generation layer of the model through the backpropagation algorithm. The weight coefficients in the variance calculation process are iteratively adjusted (the initial value range of these weight coefficients is [0.05, 0.15]) to appropriately increase the generated variance value. The upper and lower confidence bounds of the predicted value at each future time point are recalculated using the corrected probability distribution parameters to generate the adjusted prediction interval. The width of the adjusted prediction interval is appropriately increased, which can effectively improve the coverage capability of real-time power data and ensure that the reliability of the prediction results meets the energy dispatching needs of the park.

[0070] This embodiment trains a time-series prediction model based on reconstructed power data and synchronous meteorological data, eliminating data mutation interference caused by equipment switching, ensuring the authenticity and integrity of the training data, enabling the model to learn the real laws of energy system operation, and improving the basic accuracy of power prediction. The feature processing method integrating self-attention and cross-attention mechanisms captures both the temporal correlation characteristics of the power data itself and explores the influence of meteorological conditions on power changes, achieving deep fusion of multi-source data. Compared with traditional prediction models with single data input, it has higher prediction accuracy and stronger generalization ability. The synchronous generation of net load power prediction values ​​and probability distribution parameters constructs a prediction interval that effectively characterizes the uncertainty of prediction results, providing a more comprehensive decision-making basis for park energy dispatch and solving the problem of unquantifiable risk in prediction. By statistically analyzing the actual coverage ratio of the prediction interval in real time and performing backpropagation correction based on historical prediction errors, the prediction interval is dynamically adjusted to ensure that its reliability always meets the expected requirements.

[0071] In a preferred embodiment of the present invention, step 5 includes:

[0072] Step 500a: Extract the upper and lower confidence bounds of each time point in the adjusted prediction interval as deterministic boundary constraints for the net load power at the corresponding time. With the objective function of minimizing the sum of the total electricity purchase cost and equipment loss cost of the park in the future scheduling cycle, construct a robust optimization scheduling model. Specifically, this includes: first, extracting deterministic boundary constraints by extracting the upper and lower confidence bounds of each time point in the adjusted prediction interval generated in step 404; setting the upper confidence bound of each time point as the upper limit constraint for the net load power at that time, and the lower confidence bound... The numerical value is set as the lower limit constraint of the net load power at that moment, forming a set of deterministic boundary constraints for the net load power at each time point in the future scheduling cycle. Based on the extraction of deterministic boundary constraints, the objective function of the robust optimization scheduling model is constructed, with the core objective of minimizing the sum of the total electricity purchase cost and equipment loss cost of the park in the future scheduling cycle. The specific calculation process is as follows: First, the total electricity purchase cost is calculated. For each time point t (t=1, 2, ..., T, where T is the total number of time points in the scheduling cycle) within the future scheduling cycle, the electricity purchased by the park from the grid at that moment is calculated. Electricity purchase price per unit of the grid during the corresponding period The total electricity purchase cost for the future dispatch cycle is obtained by multiplying the costs by the total cost of electricity purchased at all points in time and then summing them up. The calculation formula is as follows: For core scheduling equipment such as energy storage units, based on the correspondence between their charging and discharging power and equipment loss coefficient, the equipment loss power at each time point t within the future scheduling cycle is calculated. Unit loss cost The product of these costs, plus the sum of the equipment wear and tear costs at all points in time, yields the equipment wear and tear cost for the future scheduling cycle. The calculation formula is: Finally, the objective functions are integrated by summing the total electricity purchase cost and equipment loss cost to obtain the objective function of the robust optimization scheduling model. The output value of this function is the total cost that needs to be minimized in the future scheduling cycle. The calculation formula is as follows:

[0073] .

[0074] Step 500b involves using the rated power, charge / discharge efficiency, current state of charge (SOC), and allowable SOC range of the energy storage unit as operational constraints for the robust optimization scheduling model. The robust optimization scheduling model is then solved to obtain a set of energy storage unit charge / discharge power values ​​that satisfy all the aforementioned deterministic boundary constraints and operational constraints. Specifically, this includes: firstly, setting operational constraints by using the core operational parameters of the energy storage unit as constraints for the robust optimization scheduling model. These constraints include four types: firstly, rated power constraints, setting the upper limit of the energy storage unit's charging power to 500kW and the upper limit of its discharging power to 500kW to ensure that the real-time charge / discharge power of the energy storage unit does not exceed the rated power range during scheduling; secondly, charge / discharge efficiency constraints, based on the inherent charge / discharge efficiency parameters of the energy storage unit, setting the conversion relationship between input power and actual stored power during charging as actual stored power = product of charging input power and charging efficiency, with a charging efficiency value of 0.95; and setting the conversion relationship between stored power and output power during discharging as discharge output power = stored power released... The system employs four types of constraints: 1) the product of discharge power and discharge efficiency, with the discharge efficiency set at 0.95 to ensure the rationality of the energy conversion process; 2) the current state of charge (SOC) constraint, using the real-time SOC values ​​of the energy storage units as initial constraints to define the initial state for model solving; and 3) the allowable SOC range constraint, setting an upper limit threshold of 0.95 and a lower limit threshold of 0.05 for the SOC of the energy storage units to ensure that the SOC of the energy storage units remains within a safe operating range during scheduling. Based on these four types of operational constraints, combined with the deterministic boundary constraints constructed in step 500a, the robust optimization scheduling model is solved and the set of charging and discharging power values ​​is obtained. The two types of constraints are imported into the robust optimization scheduling model, and a linear programming algorithm is used to solve the objective function. During the solution process, the feasible solution space that satisfies all constraints needs to be traversed, and the optimal solution that minimizes the objective function value is selected. This optimal solution is the charging and discharging power value of the energy storage units at each time point in the future scheduling cycle. The charging and discharging power values ​​at all time points are integrated to form the set of charging and discharging power values ​​of the energy storage units.

[0075] Step 500c: Arrange the set of energy storage unit charging and discharging power values ​​in chronological order to generate a sequence of charging and discharging power commands for the energy storage units in the future scheduling cycle. Specifically, this includes: arranging the set of energy storage unit charging and discharging power values ​​obtained in step 500b in chronological order according to the scheduling cycle of the next 24 hours. Each time point corresponds to a charging and discharging power command value. Positive values ​​represent that the energy storage unit is in a charging state, negative values ​​represent that the energy storage unit is in a discharging state, and zero values ​​represent that the energy storage unit is in a standby state. Finally, a sequence of energy storage unit charging and discharging power commands covering the entire scheduling cycle of the next 24 hours is generated.

[0076] Step 501: After executing the charging and discharging power command sequence, collect the actual power data of the four topology nodes, calculate the actual net load power curve, and calculate the difference between the actual net load power curve and the adjusted predicted interval median sequence to form a power deviation sequence. Specifically, this includes: first, collecting actual power data and calculating the net load power curve; executing the charging and discharging power command sequence generated in step 500c; and after the scheduling cycle ends, collecting the actual power data of the four topology nodes of the park's energy system. The four topology nodes are the photovoltaic grid connection point, the energy storage access point, the main load point, and the grid interaction point. Based on the power coupling relationship of the four topology nodes, the actual net load power value at each time point is calculated by algebraically summing the node power. Specifically, the actual net load power value = main load point power. - The power at the photovoltaic grid connection point, the power at the energy storage access point, and the power at the grid interaction point are combined. The actual net load power values ​​at all time points are then arranged in chronological order to form the actual net load power curve. After obtaining the actual net load power curve, the power deviation sequence is generated. The median value of each time point in the adjusted prediction interval is calculated by taking the arithmetic mean of the upper and lower confidence bounds of the corresponding time point. The median values ​​of all time points are arranged in chronological order to form the median sequence of the adjusted prediction interval. The difference between the value of each time point in the actual net load power curve and the value of the corresponding time point in the median sequence of the adjusted prediction interval is calculated. The difference result is the actual net load power value minus the median value of the prediction interval. The difference results of all time points are arranged in chronological order to form the power deviation sequence.

[0077] Step 502: Integrate the power deviation sequence, corresponding meteorological data, switching event logs, and abnormal mutation information of the conduction coefficient to form a closed-loop feedback dataset. Based on the statistical distribution characteristics of the power deviation in the closed-loop feedback dataset, identify the time periods where the deviation exceeds a preset deviation threshold and extract the associated abnormal mutation records of the conduction coefficient within those time periods. According to the mutation records, adjust the mutation threshold used to determine the abnormal mutation of the conduction coefficient. Use the historical power data, meteorological data, and corresponding actual net load power data in the closed-loop feedback dataset as new training samples and input them into the time series prediction model to perform training and update the network parameters of the time series prediction model. By periodically performing mutation threshold parameter adjustment and time series prediction model training, complete the joint iterative update of the collaborative reconstruction process and the time series prediction model. Specifically, this includes: First, constructing the closed-loop feedback dataset by integrating the power deviation sequence obtained in step 501, meteorological data within the corresponding scheduling period, switching event logs of the park's energy system equipment, and abnormal mutation information of the system's conduction coefficient. The four types of data are precisely aligned according to the time axis to ensure that the power deviation value at each time point matches the corresponding data. Meteorological conditions, equipment operating status, and conduction coefficient status are used to construct a closed-loop feedback dataset. After completing the construction of the closed-loop feedback dataset, the extraction of abnormal deviation periods and conduction coefficient mutation records is carried out. Based on the statistical distribution characteristics of the power deviation sequence in the closed-loop feedback dataset, the mean and standard deviation of the power deviation are calculated. A preset deviation threshold of 2 × standard deviation is set. Periods in the power deviation sequence where the absolute value exceeds the preset deviation threshold are identified as abnormal deviation periods. The corresponding conduction coefficient mutation records associated with these abnormal deviation periods are extracted to clarify the correspondence between abnormal deviations and conduction coefficient mutations. Based on the extracted conduction coefficient mutation records, the conduction coefficient mutation threshold parameter is adjusted. The initial value of the conduction coefficient mutation threshold is set to 0.02. The accuracy of the current mutation threshold in identifying abnormal conduction coefficient states is analyzed. If conduction coefficient mutations within abnormal deviation periods are not effectively identified, the mutation threshold is gradually reduced in a fixed step size of 0.005. After each adjustment, the accuracy of identifying abnormal conduction coefficient mutations is re-verified until conduction coefficient mutations within abnormal deviation periods can be completely identified. If too many misidentified conduction coefficient mutations occur within normal periods, the threshold is reduced to 0.005.The mutation threshold is gradually increased with a fixed step size of 005. After each adjustment, the accuracy of identifying abnormal mutations in the conduction coefficient is re-verified until the number of false identification records within normal periods is reduced to an acceptable range. The accuracy of judging abnormal mutations in the conduction coefficient is optimized through parameter adjustment. Simultaneously with the adjustment of the conduction coefficient mutation threshold parameters, the parameters of the time-series prediction model are updated. Historical power data, meteorological data, and corresponding actual net load power data from the closed-loop feedback dataset are used as new training samples and integrated with the original training set to form an updated training set. The updated training set is input into the time-series prediction model, and the model training is re-executed according to the training process set in step 400. All network parameters, such as the model's mapping matrix, convolutional kernel weights, fully connected layer weights, and bias terms, are iteratively updated to optimize the model's prediction accuracy. Finally, a joint iterative update is performed. Following a preset 24-hour cycle, the operations of adjusting the conduction coefficient mutation threshold parameters and training the time-series prediction model are repeated. Through the continuous input of closed-loop feedback data and the dynamic updating of model parameters, the joint iterative update of the collaborative reconstruction process and the time-series prediction model is completed, ensuring that both are always adapted to changes in the operating status of the park's energy system.

[0078] This embodiment constructs a robust optimization scheduling model based on deterministic boundary constraints of net load power, and performs optimization solutions in conjunction with the full-dimensional operational constraints of energy storage units. The generated charging and discharging power command sequence can minimize the total electricity purchase cost and equipment loss cost of the park while meeting the reliable range of power fluctuations, thus improving the economy and stability of the park's energy dispatch. By calculating the difference between the actual net load power curve and the median sequence of the prediction interval, the generated power deviation sequence can accurately reflect the prediction accuracy of the time series prediction model. The constructed closed-loop feedback dataset integrates multi-source operating data. Through the identification of abnormal deviation periods and the extraction of transmission coefficient mutation records, abnormal correlation factors in the system operation process can be located. Based on this, the adjustment of the transmission coefficient mutation threshold can improve the accuracy of abnormal state judgment. At the same time, the iterative update of the time series prediction model with new training samples can continuously optimize the model's prediction performance and enhance the model's adaptability to changes in the operating state of the park's energy system. The joint iterative update of the collaborative reconstruction process and the time series prediction model forms a closed-loop optimization mechanism, breaking down information silos in each link, improving the synergy between power prediction and dispatch optimization of the park's energy system, and ensuring the long-term reliability and efficiency of the system.

[0079] like Figure 2 As shown, embodiments of the present invention also provide a zero-carbon smart park cloud platform data optimization and processing system, including:

[0080] The space construction module is used to construct a multi-dimensional state space based on four high-dimensional state vectors, and to calculate the covariance matrix of itself and the remaining three state vectors based on the state vector of the photovoltaic grid connection point in the multi-dimensional state space, so as to obtain three independent feature subspaces. The four high-dimensional state vectors include the photovoltaic grid connection point, the energy storage access point, the main load point, and the grid interaction point.

[0081] The guidance graph construction module is used to calculate the generalized angle between each pair of the three feature subspaces. Based on the generalized angle, the multidimensional state space is divided into multiple decision subdomains. In each decision subdomain, the projection of the state vector of three representative points—energy storage access point, main load point, and grid interaction point—is set in the decision subdomain. The representative points of each decision subdomain are connected to the reference state vector point to construct a multi-level virtual energy guidance graph.

[0082] The anomaly detection module is used to analyze the path conduction coefficients of each level in the virtual energy steering diagram. When an abnormal change in the conduction coefficient of any level is detected and matches the switching event log, the power time series section of the corresponding event is reconstructed collaboratively to generate reconstructed power data.

[0083] The interval verification module is used to train a time series prediction model with an integrated attention mechanism using reconstructed power data and meteorological time series sections to obtain the future net load power prediction interval, and to perform coverage verification between the prediction interval and the real-time power time series sections to obtain the adjusted prediction interval.

[0084] The iterative optimization module is used to determine the charging and discharging commands of the energy storage unit based on the adjusted prediction interval, and to jointly iteratively update the collaborative reconfiguration process and the time series prediction model according to the command execution feedback.

[0085] It should be noted that this system is a system corresponding to the above method. All implementation methods in the above method embodiments are applicable to this embodiment and can achieve the same technical effect.

[0086] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A data optimization and processing method for a zero-carbon smart park cloud platform, characterized in that, The method includes: Step 1: Based on the four high-dimensional state vectors, construct a multi-dimensional state space, and use the state vector of the photovoltaic grid connection point in the multi-dimensional state space as a reference to calculate the covariance matrix of itself and the remaining three state vectors to obtain three independent feature subspaces. The four high-dimensional state vectors include the photovoltaic grid connection point, the energy storage access point, the main load point, and the grid interaction point. Step 2: Calculate the generalized angle between each pair of the three feature subspaces. Divide the multidimensional state space into multiple decision subdomains based on the generalized angle. In each decision subdomain, set the projection of the state vector of three representative points (energy storage access point, main load point, and grid interaction point) onto the decision subdomain. Connect the representative points of each decision subdomain with the reference state vector point to construct a multi-level virtual energy guidance map. Step 3: Based on the geometric relationships of each link in the multi-level virtual energy steering diagram, calculate the first path conduction coefficient of the link connecting the reference state vector point and the energy storage access point representative point, the second path conduction coefficient of the link connecting the reference state vector point and the main load point representative point, and the third path conduction coefficient of the link connecting the reference state vector point and the grid interaction point representative point; continuously monitor the values ​​of the first path conduction coefficient, the second path conduction coefficient, and the third path conduction coefficient; when the instantaneous change amplitude of any path conduction coefficient exceeds the preset mutation threshold, it is determined that an abnormal mutation of the conduction coefficient has occurred, and the time of occurrence of the abnormal mutation is recorded; match the time of occurrence of the abnormal mutation with the timestamp recorded in the switching event log; if the match is successful, based on the values ​​of the path conduction coefficients that have not experienced abnormal mutations in the virtual energy steering diagram at the time of occurrence, perform data correction on the original power time series profile of the topology node corresponding to the path conduction coefficient that experienced abnormal mutations at the time of occurrence; after the correction is completed, summarize to obtain the reconstructed power data; Step 4: Train the time series prediction model with integrated attention mechanism using reconstructed power data and meteorological time series sections to obtain the future net load power prediction range, and verify the coverage of the prediction range with the real-time power time series sections to obtain the adjusted prediction range. Step 5: Based on the adjusted prediction interval, determine the energy storage unit charging and discharging commands, and perform joint iterative updates on the collaborative reconfiguration process and time series prediction model according to the command execution feedback.

2. The data optimization and processing method for the zero-carbon smart park cloud platform according to claim 1, characterized in that, The process of determining the four high-dimensional state vectors is as follows: The active power time series profiles of four topological nodes—the photovoltaic grid connection point, the energy storage access point, the main load point, and the grid interaction point—are collected through the cloud platform. The irradiance time series profiles and temperature time series profiles provided by the meteorological monitoring equipment are also collected to form the meteorological time series profiles. The state transition records generated by the static var generator are collected as the switching event logs. Four power time series sections and meteorological time series sections are time-aligned using sliding time windows of the same length. Within each aligned time window, the mean, variance, and frequency domain energy characteristics of each power time series section are calculated as statistical and spectral characteristics of each power time series section. The mean irradiance and mean temperature of the meteorological time series section within the corresponding window are extracted as meteorological characteristics. The power and meteorological features of each topological node in each time window are concatenated and normalized to form the feature vector of the corresponding topological node in the corresponding time window. The feature vectors of each topological node in all time windows are arranged in chronological order to form a high-dimensional state vector representing the operating state of the corresponding topological node, resulting in four high-dimensional state vectors.

3. The data optimization and processing method for the zero-carbon smart park cloud platform according to claim 2, characterized in that, Step 1 includes: Four high-dimensional state vectors are used as basic elements and synchronously arranged according to their respective sampling time points to construct a multi-dimensional state space with time as the first dimension and state vector as the second dimension. In the multi-dimensional state space, the high-dimensional state vector corresponding to the photovoltaic grid connection point is used as the reference state vector. The covariance of the reference state vector and the high-dimensional state vector corresponding to the energy storage access point in the time dimension is calculated to form the first covariance matrix. Calculate the covariance of the baseline state vector and the high-dimensional state vector corresponding to the master load point in the time dimension to form the second covariance matrix; Calculate the covariance of the reference state vector and the high-dimensional state vector corresponding to the power grid interaction point in the time dimension to form a third covariance matrix; Matrix decomposition is performed on the first covariance matrix to extract the eigenvectors corresponding to the preset first k largest eigenvalues, forming a first independent feature subspace representing the dynamic correlation between the photovoltaic grid connection point and the energy storage access point. Matrix decomposition is performed on the second covariance matrix to extract the eigenvectors corresponding to the preset first L largest eigenvalues, forming a second independent feature subspace representing the dynamic correlation between the photovoltaic grid connection point and the main load point. Matrix decomposition is performed on the third covariance matrix to extract the eigenvectors corresponding to the preset first J largest eigenvalues, forming a third independent feature subspace representing the dynamic correlation between the photovoltaic grid connection point and the grid interaction point. Here, k, L, and J are the smallest positive integers that make the cumulative contribution rate of the eigenvalues ​​≥ 90%.

4. The data optimization and processing method for the zero-carbon smart park cloud platform according to claim 3, characterized in that, Step 2 includes: Calculate the minimum angle between the principal component vector sets of the first independent feature subspace and the second independent feature subspace, and use it as the first generalized angle; calculate the minimum angle between the principal component vector sets of the first independent feature subspace and the third independent feature subspace, and use it as the second generalized angle; calculate the minimum angle between the principal component vector sets of the second independent feature subspace and the third independent feature subspace, and use it as the third generalized angle. The first, second, and third generalized included angles are compared with preset angle thresholds. Regions in the multidimensional state space that satisfy the condition that the first generalized included angle is greater than the angle threshold and both the second and third generalized included angles are less than the angle threshold are classified as first-class decision subdomains; regions in the multidimensional state space that satisfy the condition that the second generalized included angle is greater than the angle threshold and both the first and third generalized included angles are less than the angle threshold are classified as second-class decision subdomains; and regions in the multidimensional state space that satisfy the condition that the third generalized included angle is greater than the angle threshold and both the first and second generalized included angles are less than the angle threshold are classified as third-class decision subdomains. For each decision subdomain, the high-dimensional state vectors of the energy storage access point, main load point and grid interaction point are orthogonally projected onto the main plane of the corresponding decision subdomain to obtain three projected coordinate points, and the three projected coordinate points are set as three representative points in the corresponding decision subdomain. Within each decision subdomain, starting from the baseline state vector point of the corresponding decision subdomain, three representative points of the corresponding decision subdomain are connected to form a local virtual energy transmission link within the corresponding decision subdomain. The local virtual energy transmission links of all decision subdomains are combined to form a multi-level virtual energy guidance map that reflects the cross-regional transmission relationship of energy in the multi-dimensional state space.

5. The data optimization and processing method for the zero-carbon smart park cloud platform according to claim 4, characterized in that, Step 4 includes: The reconstructed power data is organized into a success rate training sequence in chronological order, and the synchronized meteorological time series sections are organized into a meteorological training sequence. The power training sequence and the meteorological training sequence are used to train the time series prediction model with integrated attention mechanism to obtain the trained time series prediction model. After training, the time series prediction model outputs the net load power prediction values ​​for consecutive future time points and the probability distribution parameters corresponding to each prediction value based on the input power history sequence and meteorological history sequence. The upper and lower confidence bounds of the predicted values ​​for each future time point are calculated based on the probability distribution parameters, and the upper and lower confidence bounds form the future net load power prediction range. Collect real-time power time series profiles actually generated within the time period covered by the future net load power prediction interval, count the number of data points in the real-time power time series profiles that fall into the prediction interval at the corresponding time, and calculate the actual coverage ratio. The actual coverage ratio is compared with the preset expected coverage ratio. If the actual coverage ratio is consistently lower than the expected coverage ratio, the probability distribution parameters output by the time series prediction model are backpropagated and corrected based on the historical prediction error data within the set time period. New probability distribution parameters are output, and new upper and lower confidence bounds are generated to form the adjusted prediction interval.

6. The data optimization and processing method for the zero-carbon smart park cloud platform according to claim 5, characterized in that, After training, the time-series forecasting model outputs predicted net load power values ​​for consecutive future time points and probability distribution parameters for each predicted value, based on the input historical power and meteorological sequences. The post-trained time series prediction model encodes the input power history sequence and meteorological history sequence into time series features to obtain power feature sequences and meteorological feature sequences. Then, it uses the attention mechanism in the post-trained time series prediction model to calculate the self-attention weights of the feature vectors at each time step in the power feature sequence and the cross-attention weights between the power feature sequence and the meteorological feature sequence. The power feature sequence and the meteorological feature sequence are weighted and fused based on self-attention weight and cross-attention weight to form a fused feature; After training, the time-series prediction model generates, in parallel, predicted net load power values ​​for each consecutive future time point and probability distribution parameters characterizing the uncertainty of each predicted value through the output layer based on the fusion features.

7. The data optimization and processing method for a zero-carbon smart park cloud platform according to claim 6, characterized in that, Step 5 includes: Based on the adjusted prediction interval, a robust optimization scheduling algorithm is used to solve the problem and obtain the charging and discharging power command sequence of the energy storage unit in the future scheduling cycle. After executing the charging and discharging power command sequence, the actual power data of the four topology nodes are collected, the actual net load power curve is calculated, and the difference between the actual net load power curve and the adjusted prediction interval median sequence is calculated to form a power deviation sequence. The power deviation sequence, corresponding meteorological data, switching event logs, and abnormal abrupt changes in the conduction coefficient are integrated to form a closed-loop feedback dataset. Based on the statistical distribution characteristics of the power deviation in the closed-loop feedback dataset, time periods where the deviation exceeds a preset deviation threshold are identified, and the associated abnormal abrupt changes in the conduction coefficient within these time periods are extracted. According to the abrupt change records, the parameters of the abrupt change threshold used to determine the abnormal abrupt changes in the conduction coefficient are adjusted. The historical power data, meteorological data, and corresponding actual net load power data in the closed-loop feedback dataset are used as new training samples and input into the time series prediction model for training to update the network parameters of the time series prediction model. By periodically performing the adjustment of the abrupt change threshold parameters and the training of the time series prediction model, the joint iterative update of the collaborative reconstruction process and the time series prediction model is completed.

8. The data optimization and processing method for the zero-carbon smart park cloud platform according to claim 7, characterized in that, Based on the adjusted prediction interval, a robust optimization scheduling algorithm is used to solve the problem, obtaining the charging and discharging power command sequence of the energy storage units within the future scheduling period, including: The values ​​of the upper and lower confidence bounds of each time point in the adjusted prediction interval are extracted as deterministic boundary constraints of the net load power at the corresponding time. A robust optimization scheduling model is constructed with the objective function of minimizing the sum of the total electricity purchase cost and equipment loss cost of the park in the future scheduling cycle. The rated power, charge / discharge efficiency, current state of charge, and allowable range of state of charge of the energy storage unit are used as the operating constraints of the robust optimization scheduling model. The robust optimization scheduling model is solved to obtain a set of energy storage unit charge / discharge power values ​​that satisfy all the aforementioned deterministic boundary constraints and operating constraints. Arrange the set of energy storage unit charging and discharging power values ​​in chronological order to generate a sequence of charging and discharging power commands for energy storage units in future scheduling cycles.

9. A zero-carbon smart park cloud platform data optimization and processing system, wherein the system implements the method as described in any one of claims 1 to 8, characterized in that, include: The space construction module is used to construct a multi-dimensional state space based on four high-dimensional state vectors, and to calculate the covariance matrix of itself and the remaining three state vectors based on the state vector of the photovoltaic grid connection point in the multi-dimensional state space, so as to obtain three independent feature subspaces. The four high-dimensional state vectors include the photovoltaic grid connection point, the energy storage access point, the main load point, and the grid interaction point. The guidance graph construction module is used to calculate the generalized angle between each pair of the three feature subspaces. Based on the generalized angle, the multidimensional state space is divided into multiple decision subdomains. In each decision subdomain, the projection of the state vector of three representative points—energy storage access point, main load point, and grid interaction point—is set in the decision subdomain. The representative points of each decision subdomain are connected to the reference state vector point to construct a multi-level virtual energy guidance graph. The anomaly detection module is used to analyze the path conduction coefficients of each level in the virtual energy steering diagram. When an abnormal change in the conduction coefficient of any level is detected and matches the switching event log, the power time series section of the corresponding event is reconstructed collaboratively to generate reconstructed power data. The interval verification module is used to train a time series prediction model with an integrated attention mechanism using reconstructed power data and meteorological time series sections to obtain the future net load power prediction interval, and to perform coverage verification between the prediction interval and the real-time power time series sections to obtain the adjusted prediction interval. The iterative optimization module is used to determine the charging and discharging commands of the energy storage unit based on the adjusted prediction interval, and to jointly iteratively update the collaborative reconfiguration process and the time series prediction model according to the command execution feedback.

Images