Real-time deviation analysis method and system for grid-connected operation data of new energy station

Abstract

The invention provides a real-time deviation analysis method and system for grid-connected operation data of a new energy station, and relates to the technical field of data deviation analysis, and the method comprises the steps: collecting the real-time grid-connected operation data of the station, and carrying out the adaptive decomposition based on a data fluctuation trend, and obtaining a plurality of groups of original feature sequences; screening features through a recursive support degree model and information entropy, and generating a reconstructed feature matrix by using a manifold embedded network; the double-branch verification network analyzes the features from the time sequence and the frequency domain, and generates time sequence feature vectors and frequency domain feature vectors; the topological continuous coherence analyzer extracts the topological relation between the features; the dynamic coupler constructs a multi-dimensional feature space, a deviation source is tracked, analyzed and positioned through a feature propagation link, and a propagation path is established; constructing a fractional order differential operator based on a propagation path, depicting a long-range dependency relationship among parameters, generating a dependency intensity matrix, performing causality test, constructing a dynamic causality graph, quantifying contribution degrees of propagation nodes, analyzing a deviation evolution mechanism, and outputting a deviation analysis result.

Description

Technical Field

[0001] The present invention relates to the technical field of data deviation analysis, and in particular to a real-time deviation analysis method and system for grid-connected operation data of a new energy station. Background Art

[0002] Renewable energy generation, such as solar and wind power generation, poses challenges to the stable operation of the power grid due to its inherent intermittent and volatile nature. The output power of grid-connected renewable energy stations is easily affected by weather conditions and other environmental factors, resulting in fluctuations in power system frequency, voltage, and power. Accurate analysis of these fluctuations and deviations is essential to maintaining grid stability, optimizing power dispatch, and ensuring power supply.

[0003] Traditional deviation analysis methods are usually based on static models and simplified assumptions, which are difficult to capture the dynamic characteristics and complex behaviors of renewable energy generation. They often rely on manual experience and expert knowledge, are inefficient and difficult to adapt to large-scale renewable energy grid connection scenarios. In addition, traditional analysis methods are difficult to effectively identify the root causes of deviations and their propagation paths in the power system, thus limiting the effective control and prevention of deviations.

[0004] In summary, the existing technology mainly has the following defects and deficiencies: Traditional analysis methods are usually based on stationarity and linearity assumptions, and cannot effectively deal with the nonlinear and non-stationary characteristics brought about by fluctuations in the output of renewable energy power generation. Existing methods can usually only identify the appearance of deviations, but it is difficult to deeply analyze the root causes and propagation mechanisms of deviations, resulting in difficulty in effectively controlling and preventing the occurrence of deviations. Traditional analysis methods lack analysis of the propagation path of deviations in the power system, making it difficult to evaluate the impact of deviations on different nodes, thereby limiting the effective control and prevention of deviations. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for real-time deviation analysis of grid-connected operation data of a new energy station, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A real-time deviation analysis method for grid-connected operation data of a new energy station is provided, comprising:

[0008] Collect real-time grid-connected operation data of new energy stations, input the real-time grid-connected operation data into an adaptive decomposition unit, dynamically adjust the decomposition window width based on the data fluctuation trend, perform time domain decomposition on the real-time grid-connected operation data, and obtain multiple groups of original feature sequences; construct a recursive support model, dynamically filter feature dimensions of the multiple groups of original feature sequences through information entropy, and obtain filtered feature subsets; input the filtered feature subsets into a manifold embedding network, construct a high-dimensional feature space based on Riemannian geometry, and generate a reconstructed feature matrix;

[0009] The reconstructed feature matrix is ​​input into a dual-branch verification network, and the dual-branch verification network analyzes the features from the time series branch and the frequency domain branch respectively; the time series branch identifies the periodic change features and the mutation features based on the data correlation of adjacent time periods, and generates a time series feature vector; the frequency domain branch identifies the distortion features based on the energy density distribution, and generates a frequency domain feature vector; a topological persistence coherence analyzer is constructed, and the topological relationship between features is extracted by calculating the Betti number and the persistence diagram, and a feature connection structure is established; the time series feature vector, the frequency domain feature vector and the feature connection structure are input into a dynamic coupler, and the feature structure is kept invariant based on the symplectic geometry principle to construct a multidimensional feature space;

[0010] In the multi-dimensional feature space, the source of the deviation is located and a propagation path is established by tracking and analyzing the feature propagation link. A fractional-order differential operator is constructed based on the propagation path. The long-range dependency relationship between parameters on the propagation path is characterized by non-integer-order derivatives. A dependency intensity matrix is ​​generated. A causal test is performed on the dependency intensity matrix. The causal link from the source of the deviation to each propagation node is identified, a dynamic causal graph is constructed, the contribution of each propagation node is quantified according to the dynamic causal graph, the evolution mechanism of the deviation on the propagation path is analyzed, and the deviation analysis result is output.

[0011] In an optional embodiment,

[0012] Collect the real-time grid-connected operation data of the new energy station, input the real-time grid-connected operation data into the adaptive decomposition unit, dynamically adjust the decomposition window width based on the data fluctuation trend, perform time domain decomposition on the real-time grid-connected operation data, and obtain multiple groups of original feature sequences including:

[0013] Input the real-time grid-connected operation data into an orthogonal wavelet analysis unit, perform continuous wavelet transform using orthogonal wavelet basis functions, obtain a time-domain-frequency domain energy distribution diagram, calculate the energy integral value of each frequency band in the time-domain-frequency domain energy distribution diagram, sort the energy integral values ​​by size, generate a frequency band energy sequence, calculate the energy accumulation ratio based on the frequency band energy sequence, extract the frequency band information whose energy accumulation ratio meets a preset threshold, and construct a target frequency band sequence; map the target frequency band sequence to the time domain, establish a time domain mapping matrix, and calculate the gradient change value of the time domain mapping matrix;

[0014] Performing Hilbert transform on the real-time grid-connected operation data to obtain an analytical signal, calculating an instantaneous frequency value based on a phase function of the analytical signal, setting the instantaneous frequency value as a center frequency, determining a bandwidth range according to the gradient change value, constructing a corresponding bandpass filter, filtering the real-time grid-connected operation data through the bandpass filter, and obtaining a filtered signal sequence;

[0015] Processing the filtered signal sequence using a variational mode decomposition method, setting the decomposition dimension to the length of the target frequency band sequence, and obtaining a mode decomposition sequence;

[0016] Calculating the energy value of each modal component in the modal decomposition sequence respectively to obtain a modal energy sequence, and dividing each modal energy value in the modal energy sequence by the sum of the modal energy sequence to obtain a modal energy proportion sequence;

[0017] Selecting the maximum modal component in the modal energy proportion sequence, extracting the period of the maximum modal component as a reference period, and updating the decomposition window width with the reference period;

[0018] Based on the updated decomposition window width, the real-time grid-connected operation data is segmented to obtain multiple groups of original feature sequences.

[0019] In an optional embodiment,

[0020] A recursive support model is constructed to dynamically filter feature dimensions of the multiple groups of original feature sequences through information entropy, and the filtered feature subsets are obtained, including:

[0021] Input the multiple groups of original feature sequences into a deep belief network, perform layer-by-layer feature transformation through a multi-layer restricted Boltzmann machine, obtain a feature hidden layer representation sequence, calculate the quantum information entropy of the feature hidden layer representation sequence, and construct a feature uncertainty matrix; construct an attention weight network, calculate the correlation scores between feature dimensions based on the feature uncertainty matrix, and generate a feature importance sequence;

[0022] Input the feature importance sequence into a pre-built recursive neural network, extract the feature temporal dependency, calculate the migration entropy value between feature dimensions, and build a conditional entropy matrix based on the migration entropy value;

[0023] Inputting the feature importance sequence and the conditional entropy matrix into a Monte Carlo tree search module, and iteratively optimizing the feature combination scheme using a policy gradient method;

[0024] Calculate the entropy increase rate of the feature combination scheme, determine the feature redundancy threshold based on the entropy increase rate, and eliminate the feature dimensions whose information gain rate is lower than the feature redundancy threshold; calculate the recursive support of each remaining feature dimension based on the feature importance sequence and the conditional entropy matrix, arrange the recursive support in descending order, and accumulate the recursive support values ​​in sequence to obtain the support accumulation value, select the feature dimensions whose support accumulation value reaches the preset ratio threshold, and construct the filtered feature subset.

[0025] In an optional embodiment,

[0026] The filtered feature subset is input into the manifold embedding network, a high-dimensional feature space is constructed based on Riemannian geometry, and a reconstructed feature matrix is ​​generated, including:

[0027] Inputting the filtered feature subset into a feature mapping layer, processing the real component of the filtered feature subset using a hyperbolic tangent function, processing the imaginary component of the filtered feature subset using a cosine function, and generating a complex-valued feature map;

[0028] The complex-valued feature map is projected to the manifold tangent space through an exponential mapping function, and the minimum path integral between feature points is calculated to obtain the geodesic distance; based on the geodesic distance, a local metric tensor is constructed, and the second-order covariant derivative of the local metric tensor is calculated through coordinate transformation to obtain the manifold curvature tensor; based on the manifold curvature tensor, a manifold curvature constraint is constructed through a cross-sectional curvature criterion, and the manifold curvature constraint is input into the metric learning layer;

[0029] In the metric learning layer, a covariant derivative operator is constructed using an affine connection, and a parallel shift equation group is calculated through a third-order connection coefficient, and a characteristic manifold transmission matrix is ​​obtained by solving the equation. A variational functional is established according to the characteristic manifold transmission matrix, and a geodesic flow equation is obtained by minimizing the variational functional. The geodesic flow equation is solved to obtain the characteristic manifold evolution trajectory.

[0030] Performing wavelet decomposition on the evolution trajectory of the characteristic manifold, calculating the covering dimension at each scale, obtaining a fractal feature sequence, calculating the cumulative dimension of the fractal feature sequence at different scales, determining the dimension of the target feature space, and constructing a high-dimensional feature space of corresponding dimension;

[0031] The fractal feature sequence is transformed into the high-dimensional feature space by isometric mapping that maintains the geodesic distance, so as to generate a reconstructed feature matrix.

[0032] In an optional embodiment,

[0033] The reconstructed feature matrix is ​​input into a dual-branch verification network, and the dual-branch verification network analyzes the features from the time series branch and the frequency domain branch respectively; the time series branch identifies the periodic change features and the mutation features based on the data correlation of adjacent time periods, and generates a time series feature vector; the frequency domain branch identifies the distortion features based on the energy density distribution, and generates the frequency domain feature vector, including:

[0034] Inputting the reconstructed feature matrix into a recursive attention unit, the recursive attention unit reads the feature data stream through an input gate, filters the historical feature information through a forget gate, and generates preliminary temporal features through an output gate;

[0035] Based on a preset time window length, a corresponding sliding time window is set for the preliminary time series feature, two adjacent sliding time windows are traversed in sequence, the feature sequence mutual correlation coefficients are calculated, the feature sequence mutual correlation coefficients are formed into a feature correlation matrix, the feature correlation matrix is ​​subjected to singular value decomposition to obtain a left singular matrix, a singular value matrix and a right singular matrix, and the column vector of the left singular matrix corresponding to the maximum singular value in the singular value matrix is ​​selected as a periodic variation feature;

[0036] Performing kernel density estimation on the preliminary time series features to obtain a probability density function of a feature sequence, calculating a deviation between the probability density function and a preset reference density function, and identifying a feature mutation point through an adaptive threshold judgment unit;

[0037] Through the feature fusion module, the periodic change feature and the feature mutation point are weightedly combined to generate a time series feature vector;

[0038] Performing multi-scale wavelet decomposition on the reconstructed feature matrix, calculating wavelet coefficients using orthogonal wavelet basis functions, and taking the sum of the squares of the wavelet coefficients in different frequency bands as the frequency band energy density;

[0039] Calculating a query matrix, a key matrix and a value matrix of the frequency band energy density, obtaining a frequency band attention weight by matrix multiplication, weightedly combining the frequency band attention weight with the frequency band energy density, and identifying a frequency band distortion feature based on an energy density threshold;

[0040] A state transfer equation is constructed, and the position information of the frequency band distortion feature in the time domain and the frequency domain is input into the state transfer equation. The optimal state path is solved by a preset dynamic programming algorithm to generate a frequency domain feature vector.

[0041] In an optional embodiment,

[0042] Constructing a topological persistence coherence analyzer, extracting the topological relationship between features and establishing a feature connection structure by calculating the Betti number and the persistence graph; inputting the time series feature vector, the frequency domain feature vector and the feature connection structure into a dynamic coupler, maintaining the invariance of the feature structure based on the principle of symplectic geometry, and constructing a multidimensional feature space, including:

[0043] Mapping the time series feature vector and the frequency domain feature vector to the feature space, and calculating the distance value between the feature points in the feature space based on the Euclidean distance calculation formula;

[0044] A first distance threshold is set to start constructing a first simplicial complex: when the distance value between the feature points is less than the first distance threshold, a first connecting edge is established between the corresponding feature points; when the three feature points all have the first connecting edge, a first two-dimensional simplicial complex is constructed; when any three feature points among the four feature points all constitute the first two-dimensional simplicial complex, a first three-dimensional simplicial complex is constructed;

[0045] Based on a preset distance increment step, the first distance threshold is increased to obtain a second distance threshold, a second simplicial complex is constructed, and iteration is repeated until a preset maximum distance threshold is reached to obtain a simplicial complex sequence;

[0046] Calculate the zero-order Betti number of the simplicial complex sequence to obtain the number of connected components, calculate the first-order Betti number of the simplicial complex sequence to obtain the number of ring structures, and calculate the second-order Betti number of the simplicial complex sequence to obtain the number of cavity structures; construct a continuous homology group based on the number of connected components, the number of ring structures, and the number of cavity structures, and extract the generation time and extinction time of the topological structure in the continuous homology group; calculate the time interval between the generation time and the extinction time to obtain the topological feature life cycle, set a life cycle threshold, and screen out stable topological features that are greater than the life cycle threshold;

[0047] A symmetric adjacency matrix is ​​constructed based on the stable topological features, and the connection strength between the feature points is used as the matrix element of the symmetric adjacency matrix; a first tensor product operation is performed on the symmetric adjacency matrix and the time series feature vector to obtain a time series feature connection structure, and a second tensor product operation is performed on the symmetric adjacency matrix and the frequency domain feature vector to obtain a frequency domain feature connection structure;

[0048] The time series feature connection structure is used as a generalized coordinate, and the frequency domain feature connection structure is used as a generalized momentum to construct a symplectic geometric system; based on the generalized coordinates, the potential energy is determined, based on the generalized momentum, the kinetic energy is determined, the potential energy and the kinetic energy are added together to construct the energy function of the symplectic geometric system, the canonical equation group is determined, the canonical equation group is solved to obtain the dynamic evolution trajectory of the time series feature connection structure and the frequency domain feature connection structure, a multidimensional feature space is constructed, and the dynamic evolution trajectory is mapped into the multidimensional feature space to realize the dynamic coupling of the feature structures.

[0049] In an optional embodiment,

[0050] In the multidimensional feature space, the source of the deviation is located by tracking and analyzing the feature propagation link, a propagation path is established, a fractional-order differential operator is constructed based on the propagation path, the long-range dependency relationship between parameters on the propagation path is characterized by non-integer-order derivatives, a dependency intensity matrix is ​​generated, a causal test is performed on the dependency intensity matrix, the causal link from the source of the deviation to each propagation node is identified, a dynamic causal graph is constructed, the contribution of each propagation node is quantified according to the dynamic causal graph, the evolution mechanism of the deviation on the propagation path is analyzed, and the output deviation analysis results include:

[0051] In the multidimensional feature space, a differential sequence of feature values ​​at adjacent moments is calculated to establish a feature propagation link, wherein the feature propagation link includes nodes and edges, wherein the nodes represent features, the edges represent propagation relationships between features, and the weights of the edges are determined based on the temporal correlation of the differential sequence;

[0052] In the feature propagation link, an abnormal point that deviates from the expected value is identified as a deviation abnormal point through a threshold analysis method, and the earliest appearing deviation abnormal point is determined as the deviation source along the edge with the largest weight in the feature propagation link; starting from the deviation source, the edge with the largest weight in the feature propagation link is traced forward until the deviation abnormal point, and the path from the deviation source to the deviation abnormal point is determined as the propagation path;

[0053] A fractional-order differential operator is constructed based on the propagation path, the order of the fractional-order differential operator is between zero and one, the characteristic values ​​of adjacent nodes on the propagation path are substituted into the fractional-order differential operator, the prediction error is minimized by an iterative optimization method, the optimal order of the fractional-order differential operator is determined, and the long-range dependency relationship between the parameters on the propagation path is obtained;

[0054] Based on the long-range dependency, the square integral of the error between the output value of the fractional differential operator and the actual eigenvalue on the propagation path is calculated to generate the dependency strength between adjacent nodes on the propagation path and organize it into a dependency strength matrix;

[0055] An autoregressive model containing historical information is constructed for the dependency strength matrix, and the significance of the causal relationship between the nodes in the dependency strength matrix is ​​determined by the F test method, so as to identify the causal link from the deviation source to each node on the propagation path;

[0056] Constructing a dynamic causal graph, taking all nodes on the propagation path as a node set of the dynamic causal graph, and taking the causal links as an edge set of the dynamic causal graph;

[0057] Based on the dynamic causal graph, the path probability from the deviation source through each node on the propagation path to the target node is calculated, the path probability is normalized to obtain a normalized path probability, the normalized path probability is multiplied by the corresponding dependency strength in the dependency strength matrix and the sum is calculated to obtain the contribution of each node on the propagation path;

[0058] According to the contribution of each node on the propagation path and the time correlation relationship, the evolution mechanism of the deviation on the propagation path is analyzed, and the deviation analysis result is output based on the evolution mechanism.

[0059] According to a second aspect of the embodiments of the present invention,

[0060] A real-time deviation analysis system for grid-connected operation data of new energy stations is provided, including:

[0061] The first unit is used to collect real-time grid-connected operation data of new energy stations, input the real-time grid-connected operation data into an adaptive decomposition unit, dynamically adjust the decomposition window width based on the data fluctuation trend, and perform time domain decomposition on the real-time grid-connected operation data to obtain multiple groups of original feature sequences; construct a recursive support model, dynamically filter feature dimensions of the multiple groups of original feature sequences through information entropy, and obtain filtered feature subsets; input the filtered feature subsets into a manifold embedding network, construct a high-dimensional feature space based on Riemannian geometry, and generate a reconstructed feature matrix;

[0062] The second unit is used to input the reconstructed feature matrix into a dual-branch verification network, and the dual-branch verification network analyzes the features from the time series branch and the frequency domain branch respectively; the time series branch identifies the periodic change features and the mutation features based on the data correlation of adjacent time periods, and generates a time series feature vector; the frequency domain branch identifies the distortion features based on the energy density distribution, and generates a frequency domain feature vector; a topological persistence coherence analyzer is constructed, and the topological relationship between features is extracted by calculating the Betti number and the persistence diagram, and a feature connection structure is established; the time series feature vector, the frequency domain feature vector and the feature connection structure are input into a dynamic coupler, and the feature structure is kept invariant based on the symplectic geometry principle to construct a multidimensional feature space;

[0063] The third unit is used to track and analyze the feature propagation link in the multi-dimensional feature space, locate the source of the deviation, establish a propagation path, construct a fractional-order differential operator based on the propagation path, characterize the long-range dependency between parameters on the propagation path through non-integer-order derivatives, generate a dependency intensity matrix, perform a causal test on the dependency intensity matrix, identify the causal link from the source of the deviation to each propagation node, construct a dynamic causal graph, quantify the contribution of each propagation node according to the dynamic causal graph, analyze the evolution mechanism of the deviation on the propagation path, and output the deviation analysis result.

[0064] According to a third aspect of the embodiments of the present invention,

[0065] An electronic device is provided, comprising:

[0066] processor;

[0067] a memory for storing processor-executable instructions;

[0068] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0069] A fourth aspect of the embodiments of the present invention is:

[0070] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0071] In the embodiment of the present invention, through the techniques such as adaptive decomposition, recursive support model and manifold embedding network, the key features in the grid-connected operation data of new energy stations can be effectively extracted and reconstructed to avoid information loss, thereby improving the accuracy of deviation analysis. The dual-branch verification network analyzes the features from the perspectives of time series and frequency domain, and combines the topological continuous coherence analyzer to extract the topological relationship between the features, further improving the comprehensiveness and reliability of the analysis; using the feature propagation link tracking analysis technology, the source of the deviation can be accurately located and the propagation path can be constructed. Through fractional differential operators and causality tests, the causal links from the source of the deviation to each propagation node can be identified, a dynamic causal graph can be constructed, and the contribution of each propagation node to the deviation can be quantified, thereby clearly showing the evolution mechanism of the deviation; using adaptive decomposition technology, the decomposition window width can be dynamically adjusted according to the data fluctuation trend, which enhances the adaptability of the method to different working conditions. At the same time, a multidimensional feature space is constructed based on the principle of symplectic geometry to maintain the invariance of the feature structure, further improving the robustness and stability of the method. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 A schematic flow chart of a method for real-time deviation analysis of grid-connected operation data of a new energy station according to an embodiment of the present invention;

[0073] Figure 2 It is a structural schematic diagram of a real-time deviation analysis system for grid-connected operation data of a new energy station according to an embodiment of the present invention. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0075] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0076] Figure 1 Schematic diagram of the process of real-time deviation analysis method of grid-connected operation data of new energy stations according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0077] S101. Collect real-time grid-connected operation data of new energy stations, input the real-time grid-connected operation data into an adaptive decomposition unit, dynamically adjust the decomposition window width based on the data fluctuation trend, perform time domain decomposition on the real-time grid-connected operation data, and obtain multiple groups of original feature sequences; construct a recursive support model, dynamically filter feature dimensions of the multiple groups of original feature sequences through information entropy, and obtain filtered feature subsets; input the filtered feature subsets into a manifold embedding network, construct a high-dimensional feature space based on Riemannian geometry, and generate a reconstructed feature matrix;

[0078] In this embodiment, the decomposition window width is dynamically adjusted by the adaptive decomposition unit, so that the fluctuation trend of the real-time grid-connected operation data of the new energy station can be captured more accurately, and the accuracy of feature extraction can be improved; the recursive support model and information entropy are used to dynamically screen the feature dimensions, reduce the interference of redundant features, retain key information, and improve the quality and effectiveness of the feature set; through the manifold embedding network and Riemann geometry method, a high-dimensional feature space is constructed, and a reconstructed feature matrix is ​​generated, which can more fully express the nonlinear relationship between complex features; the reconstructed feature matrix retains the important information of the original data, and at the same time, through noise reduction and optimization, the availability of the data is improved, which lays a more robust foundation for subsequent analysis and model application (such as prediction and control).

[0079] S102. Input the reconstructed feature matrix into a dual-branch verification network, and the dual-branch verification network analyzes the features from the time series branch and the frequency domain branch respectively; the time series branch identifies the periodic change features and the mutation features based on the data correlation of adjacent time periods, and generates a time series feature vector; the frequency domain branch identifies the distortion features based on the energy density distribution, and generates a frequency domain feature vector; construct a topological persistence coherence analyzer, extract the topological relationship between features by calculating the Betti number and the persistence diagram, and establish a feature connection structure; input the time series feature vector, the frequency domain feature vector and the feature connection structure into a dynamic coupler, maintain the invariance of the feature structure based on the principle of symplectic geometry, and construct a multi-dimensional feature space;

[0080] In this embodiment, the dual-branch verification network combines timing branch and frequency domain branch analysis to simultaneously capture periodic changes, mutation features (timing branches) and energy distortion features (frequency domain branches), thereby achieving comprehensive extraction and precise identification of multiple features. The topological persistence coherence analyzer calculates the Betti number and persistence graph to extract the topological relationship between features, thereby identifying the global and local connection characteristics of the data and providing structural information for complex system modeling. The dynamic coupler is based on the principle of symplectic geometry to organically integrate the extracted timing features, frequency domain features and topological features while maintaining the invariance of the feature structure to ensure the physical meaning and stability of the multi-dimensional feature space. The timing, frequency domain and topological features are integrated to construct a multi-dimensional feature space, which can more fully express the multi-level information of complex data and provide more accurate and efficient feature representation for subsequent modeling, prediction and analysis.

[0081] S103. In the multi-dimensional feature space, the source of the deviation is located through feature propagation link tracking and analysis, a propagation path is established, a fractional-order differential operator is constructed based on the propagation path, the long-range dependency relationship between parameters on the propagation path is characterized by non-integer-order derivatives, a dependency intensity matrix is ​​generated, a causal test is performed on the dependency intensity matrix, the causal link from the source of the deviation to each propagation node is identified, a dynamic causal graph is constructed, the contribution of each propagation node is quantified according to the dynamic causal graph, the evolution mechanism of the deviation on the propagation path is analyzed, and the deviation analysis result is output.

[0082] In this embodiment, through the tracking analysis of feature propagation links and causal tests, the starting position of the deviation and its impact range can be accurately identified, providing a scientific basis for the source diagnosis of the deviation problem; by using fractional differential operators and non-integer derivatives, the long-range dependency relationship between parameters on the propagation path is deeply characterized, the complex dynamic interaction between features is quantified, and the depth and meticulousness of the analysis are improved; by generating a dependency intensity matrix and identifying causal links, a dynamic causal graph is constructed to reveal the propagation mechanism of the deviation in the system and clarify the causal relationship between nodes; the dynamic causal graph provides a means to quantify the contribution of propagation nodes, and can intuitively evaluate the impact of each node on the propagation of the deviation, providing guidance for optimizing system design or intervention measures; the analysis of the comprehensive propagation path and causal link reveals the evolution mechanism of the deviation in the system, and supports in-depth research on the law of deviation propagation in complex systems; finally, the deviation analysis results are formed to provide an intuitive and quantitative reference for system performance optimization, deviation correction and decision support.

[0083] In an optional implementation, real-time grid-connected operation data of a new energy station is collected, the real-time grid-connected operation data is input into an adaptive decomposition unit, the decomposition window width is dynamically adjusted based on the data fluctuation trend, and the real-time grid-connected operation data is decomposed in the time domain to obtain multiple groups of original feature sequences including:

[0084] Input the real-time grid-connected operation data into an orthogonal wavelet analysis unit, perform continuous wavelet transform using orthogonal wavelet basis functions, obtain a time-domain-frequency domain energy distribution diagram, calculate the energy integral value of each frequency band in the time-domain-frequency domain energy distribution diagram, sort the energy integral values ​​by size, generate a frequency band energy sequence, calculate the energy accumulation ratio based on the frequency band energy sequence, extract the frequency band information whose energy accumulation ratio meets a preset threshold, and construct a target frequency band sequence; map the target frequency band sequence to the time domain, establish a time domain mapping matrix, and calculate the gradient change value of the time domain mapping matrix;

[0085] Performing Hilbert transform on the real-time grid-connected operation data to obtain an analytical signal, calculating an instantaneous frequency value based on a phase function of the analytical signal, setting the instantaneous frequency value as a center frequency, determining a bandwidth range according to the gradient change value, constructing a corresponding bandpass filter, filtering the real-time grid-connected operation data through the bandpass filter, and obtaining a filtered signal sequence;

[0086] Processing the filtered signal sequence using a variational mode decomposition method, setting the decomposition dimension to the length of the target frequency band sequence, and obtaining a mode decomposition sequence;

[0087] Calculating the energy value of each modal component in the modal decomposition sequence respectively to obtain a modal energy sequence, and dividing each modal energy value in the modal energy sequence by the sum of the modal energy sequence to obtain a modal energy proportion sequence;

[0088] Selecting the maximum modal component in the modal energy proportion sequence, extracting the period of the maximum modal component as a reference period, and updating the decomposition window width with the reference period;

[0089] Based on the updated decomposition window width, the real-time grid-connected operation data is segmented to obtain multiple groups of original feature sequences.

[0090] First, the collected real-time grid-connected operation data is preprocessed, such as removing outliers and noise. Some common data preprocessing methods can be used, such as moving average filtering, wavelet denoising, etc. Assuming that there are some spike pulses caused by sensor failure in the collected photovoltaic power output power data, the moving average filtering method can be used to remove them. For example, taking the average value of 5 consecutive sampling points as the new data point can effectively smooth the data and remove the spike pulses.

[0091] Next, the preprocessed real-time grid-connected operation data is input into the adaptive decomposition unit. The core function of this unit is to dynamically adjust the decomposition window width based on the data fluctuation trend. For example, when the data fluctuates violently, the decomposition window width should be smaller in order to capture more subtle changes; when the data fluctuates gently, the decomposition window width can be larger to reduce the amount of calculation. Specifically, statistics such as the variance or standard deviation of the data can be calculated, and the decomposition window width can be dynamically adjusted based on these statistics. For example, a threshold is set, and when the variance of the data exceeds the threshold, the decomposition window width is reduced; when the variance of the data is lower than the threshold, the decomposition window width is increased. Assuming that the initial decomposition window width is 1 hour, when the data variance exceeds the preset threshold, the decomposition window width is reduced to 0.5 hours; when the data variance is lower than the preset threshold, the decomposition window width is increased to 2 hours. In this way, the decomposition window width can be adaptively adjusted according to the fluctuation trend of the data, and the real-time grid-connected operation data can be decomposed in the time domain to obtain multiple groups of original feature sequences.

[0092] Then, the real-time grid-connected operation data is input into the orthogonal wavelet analysis unit. Orthogonal wavelet basis functions are used, preferably Daubechies wavelet is applied to perform continuous wavelet transform, and a time-domain-frequency domain energy distribution diagram is obtained. For example, a continuous wavelet transform is performed on a section of photovoltaic power generation output power data, and a two-dimensional image can be obtained, in which the horizontal axis represents time, the vertical axis represents frequency, and the brightness of the image represents the magnitude of energy. The energy integral value of each band in the time-domain-frequency domain energy distribution diagram is calculated, and the energy integral value is sorted by size to generate a band energy sequence. Assuming that the energy integral values ​​of the five bands are calculated to be 10, 5, 20, 15, and 1 respectively, the generated band energy sequence is 20, 15, 10, 5, and 1. The energy accumulation ratio is calculated based on the band energy sequence, for example, the proportion of the energy accumulation values ​​of the first two bands to the total energy. The band information whose energy accumulation ratio meets the preset threshold (for example, 80%) is extracted to construct a target band sequence. Assuming that the energy accumulation ratio of the first three bands has exceeded 80%, the target band sequence contains the first three bands. The target frequency band sequence is mapped to the time domain, a time domain mapping matrix is ​​established, and the gradient change value of the time domain mapping matrix is ​​calculated.

[0093] After that, the real-time grid-connected operation data is Hilbert transformed to obtain the analytical signal. Based on the phase function of the analytical signal, the instantaneous frequency value is calculated. The instantaneous frequency value is set as the center frequency, the bandwidth range is determined according to the gradient change value, and the corresponding bandpass filter is constructed. For example, if the instantaneous frequency is 50Hz and the gradient change value is 10Hz, a bandpass filter with a center frequency of 50Hz and a bandwidth of 20Hz is constructed. The real-time grid-connected operation data is filtered by the bandpass filter to obtain a filtered signal sequence.

[0094] The filter signal sequence is processed using the variational mode decomposition method, and the decomposition dimension is set to the length of the target frequency band sequence. For example, if the target frequency band sequence contains 3 frequency bands, the decomposition dimension is set to 3 to obtain 3 mode decomposition sequences.

[0095] The energy value of each modal component in the modal decomposition sequence is calculated separately to obtain the modal energy sequence. The modal energy proportion sequence is obtained by dividing each modal energy value in the modal energy sequence by the sum of the modal energy sequence.

[0096] Select the maximum modal component in the modal energy proportion sequence, extract the period of the maximum modal component as the reference period, and update the decomposition window width with the reference period. For example, if the period of the maximum modal component is 1 hour, update the decomposition window width to 1 hour.

[0097] Based on the updated decomposition window width, the real-time grid-connected operation data is segmented to obtain multiple groups of original feature sequences.

[0098] In this embodiment, the analysis accuracy of the grid-connected operation data of the new energy station is improved. Through adaptive decomposition window width and multi-scale wavelet analysis, the fluctuation information in the data can be captured more accurately, thereby improving the analysis accuracy; the monitoring capability of the grid-connected operation status of the new energy station is enhanced. By extracting key feature information, the grid-connected operation status of the new energy station can be more effectively monitored, and abnormal situations can be discovered in time; the grid-connected operation control strategy of the new energy station is optimized. Based on the analysis results, the grid-connected operation control strategy of the new energy station can be optimized to improve the grid efficiency and stability.

[0099] In an optional implementation, a recursive support model is constructed, and the feature dimensions are dynamically screened through information entropy for the multiple groups of original feature sequences to obtain the screened feature subsets, including:

[0100] Input the multiple groups of original feature sequences into a deep belief network, perform layer-by-layer feature transformation through a multi-layer restricted Boltzmann machine, obtain a feature hidden layer representation sequence, calculate the quantum information entropy of the feature hidden layer representation sequence, and construct a feature uncertainty matrix; construct an attention weight network, calculate the correlation scores between feature dimensions based on the feature uncertainty matrix, and generate a feature importance sequence;

[0101] Input the feature importance sequence into a pre-built recursive neural network, extract the feature temporal dependency, calculate the migration entropy value between feature dimensions, and build a conditional entropy matrix based on the migration entropy value;

[0102] Inputting the feature importance sequence and the conditional entropy matrix into a Monte Carlo tree search module, and iteratively optimizing the feature combination scheme using a policy gradient method;

[0103] Calculate the entropy increase rate of the feature combination scheme, determine the feature redundancy threshold based on the entropy increase rate, and eliminate the feature dimensions whose information gain rate is lower than the feature redundancy threshold; calculate the recursive support of each remaining feature dimension based on the feature importance sequence and the conditional entropy matrix, arrange the recursive support in descending order, and accumulate the recursive support values ​​in sequence to obtain the support accumulation value, select the feature dimensions whose support accumulation value reaches the preset ratio threshold, and construct the filtered feature subset.

[0104] The quantum information entropy specifically refers to a metric used to quantify information uncertainty in a quantum system. It describes the degree of disorder of the system or the uncertainty of information in a quantum state. Compared with classical information entropy, quantum information entropy can reflect the characteristics of superposition and entanglement in a quantum system, and is suitable for describing the correlation or degree of uncertainty between quantum states.

[0105] The transfer entropy value is specifically used to measure the information transfer intensity between two random processes. It describes how the state of one variable affects the state of another variable in time, with special attention to the directionality in the causal relationship. A high transfer entropy value indicates that one variable has a strong ability to drive or influence another variable, and is a key indicator for analyzing the dynamic dependency of the system;

[0106] The entropy increase rate specifically describes the rate of change of information entropy in the process of feature combination or information transmission, and is used to measure the contribution of new information to the reduction of overall uncertainty in the system. The higher the entropy increase rate, the more significant the improvement of the system description by the new feature or new information; a low entropy increase rate indicates that the feature contributes little to the overall information gain or is redundant;

[0107] First, multiple sets of original time series feature data are input into a deep belief network. The deep belief network is composed of multiple layers of restricted Boltzmann machines, which can perform nonlinear transformations on the input features layer by layer to learn higher-level feature representations. Each layer of the restricted Boltzmann machine consists of a visible layer and a hidden layer, and the features are transformed through the connection weights between the visible layer and the hidden layer. After multiple layers of transformation, a hidden layer representation sequence of the features is obtained. For example, suppose there are 100 sets of time series data, each set of data contains 50 time steps, and each time step has 10 feature dimensions. After transformation by the deep belief network, 100 sets of new time series data can be obtained, each set of data contains 50 time steps, and each time step has 20 hidden layer feature dimensions.

[0108] Next, the quantum information entropy of the feature hidden layer representation sequence is calculated to measure the uncertainty of each feature dimension. The higher the information entropy, the greater the uncertainty of the feature dimension. The uncertainties of all feature dimensions are combined into a matrix, called the feature uncertainty matrix. For example, the average information entropy of each hidden layer feature dimension over all time steps and all samples can be calculated to obtain a 20×20 feature uncertainty matrix.

[0109] Then, an attention weight network is constructed, with the input being the feature uncertainty matrix. The attention weight network can learn the degree of correlation between different feature dimensions. Based on the feature uncertainty matrix, the correlation scores between feature dimensions are calculated to generate a feature importance sequence. The higher the correlation score, the more important the feature dimension. For example, the attention weight network can be a simple fully connected neural network, with the output being the importance score for each feature dimension, thus obtaining a feature importance sequence containing 20 elements.

[0110] After that, the feature importance sequence is input into a pre-built recurrent neural network. Recurrent neural networks are good at processing time series data and can capture the temporal dependencies between feature dimensions. Through the recurrent neural network, the migration entropy value between feature dimensions is calculated to measure the influence of one feature dimension on the future value of another feature dimension. The conditional entropy matrix is ​​constructed based on the migration entropy value. For example, the influence of each feature dimension on the future value of other feature dimensions can be calculated to obtain a 20×20 conditional entropy matrix.

[0111] Next, the feature importance sequence and conditional entropy matrix are input into a Monte Carlo tree search module. Monte Carlo tree search is a search algorithm based on random simulation, which can be used to find the optimal feature combination scheme. The policy gradient method is used to iteratively optimize the feature combination scheme. For example, the feature dimension combination can be regarded as an action space, and the Monte Carlo tree search is used to find the best feature combination action sequence.

[0112] Then, the entropy increase rate of each feature combination scheme is calculated to measure the information gain brought by the scheme. The feature redundancy threshold is determined based on the entropy increase rate, and the feature dimensions whose information gain rate is lower than the feature redundancy threshold are eliminated. For example, an entropy increase rate threshold can be set, and feature dimensions below the threshold are considered redundant and need to be eliminated.

[0113] Finally, based on the feature importance sequence and conditional entropy matrix, the recursive support of each remaining feature dimension is calculated. The recursive support takes into account the importance, temporal dependency and redundancy of the feature. The recursive support is sorted in descending order, and the recursive support values ​​are accumulated in sequence to obtain the cumulative support value. The feature dimensions whose cumulative support values ​​reach the preset ratio threshold are selected, for example, the feature dimensions whose cumulative values ​​reach 80% of the total support are selected to construct the final filtered feature subset.

[0114] In this embodiment, through the deep belief network and attention mechanism, important feature dimensions can be identified quickly and effectively, reducing computational complexity; through the recursive neural network and transfer entropy, the temporal dependency between feature dimensions can be captured, so that the screened feature subset can better reflect the dynamic characteristics of the original data; through the Monte Carlo tree search and entropy increase rate, redundant features can be effectively eliminated, information duplication can be avoided, and the generalization ability of the model can be improved.

[0115] In an optional implementation, inputting the filtered feature subset into a manifold embedding network, constructing a high-dimensional feature space based on Riemannian geometry, and generating a reconstructed feature matrix includes:

[0116] Inputting the filtered feature subset into a feature mapping layer, processing the real component of the filtered feature subset using a hyperbolic tangent function, processing the imaginary component of the filtered feature subset using a cosine function, and generating a complex-valued feature map;

[0117] The complex-valued feature map is projected to the manifold tangent space through an exponential mapping function, and the minimum path integral between feature points is calculated to obtain the geodesic distance; based on the geodesic distance, a local metric tensor is constructed, and the second-order covariant derivative of the local metric tensor is calculated through coordinate transformation to obtain the manifold curvature tensor; based on the manifold curvature tensor, a manifold curvature constraint is constructed through a cross-sectional curvature criterion, and the manifold curvature constraint is input into the metric learning layer;

[0118] In the metric learning layer, a covariant derivative operator is constructed using an affine connection, and a parallel shift equation group is calculated through a third-order connection coefficient, and a characteristic manifold transmission matrix is ​​obtained by solving the equation. A variational functional is established according to the characteristic manifold transmission matrix, and a geodesic flow equation is obtained by minimizing the variational functional. The geodesic flow equation is solved to obtain the characteristic manifold evolution trajectory.

[0119] Performing wavelet decomposition on the evolution trajectory of the characteristic manifold, calculating the covering dimension at each scale, obtaining a fractal feature sequence, calculating the cumulative dimension of the fractal feature sequence at different scales, determining the dimension of the target feature space, and constructing a high-dimensional feature space of corresponding dimension;

[0120] The fractal feature sequence is transformed into the high-dimensional feature space by isometric mapping that maintains the geodesic distance, so as to generate a reconstructed feature matrix.

[0121] The affine connection specifically refers to a geometric structure used to describe how a vector field on a manifold moves locally in parallel. It defines a rule that allows tangent vectors at different points to be compared, or vectors to be "transferred" on a surface or manifold while maintaining certain properties. The affine connection provides a way to link the rate of change of geometric objects to the intrinsic structure of the manifold, especially in the absence of flat space. Through the affine connection, the "parallelism" and "direction preservation" properties on the manifold can be described. It is a basic tool in fields such as surface geometry, Riemannian geometry, and general relativity, used to study changes in paths (such as curvature, torsion) and the dynamic evolution of features on manifolds.

[0122] First, feature mapping is performed on the filtered feature subset. The real component of the feature subset is processed using the hyperbolic tangent function, and the imaginary component is processed using the cosine function to generate a complex-valued feature map. For example, assuming that the feature subset is [0.5, 0.8, -0.2, 0.1], the real component [0.5, -0.2] is processed by the hyperbolic tangent function to obtain a new value, and the imaginary component [0.8, 0.1] is processed by the cosine function to obtain a new value, and finally combined into a complex-valued feature map.

[0123] Next, the complex-valued feature map is projected to the manifold tangent space. The exponential mapping function is used for projection, and the minimum path integral between the feature points is calculated to obtain the geodesic distance. For example, suppose there are two feature points with coordinates [0.1, 0.2] and [0.3, 0.4] on the manifold. By calculating the minimum path integral between them, the geodesic distance between them can be obtained as 0.28. Based on the geodesic distance, a local metric tensor is constructed. The second-order covariant derivative of the local metric tensor is calculated by coordinate transformation to obtain the manifold curvature tensor. For example, the local metric tensor can be regarded as a matrix, and the manifold curvature tensor is obtained by calculating the second-order derivative of the matrix elements. Based on the manifold curvature tensor, the manifold curvature constraints are constructed using the section curvature criterion, and these constraints are input into the metric learning layer.

[0124] In the metric learning layer, the covariant derivative operator is constructed using the affine connection. The parallel move equations are calculated by the third-order connection coefficients, and the characteristic manifold transfer matrix is ​​obtained by solving them. For example, assuming that the third-order connection coefficient is a three-dimensional tensor, a matrix, namely the characteristic manifold transfer matrix, can be obtained by solving the parallel move equations. A variational functional is established based on the characteristic manifold transfer matrix, and the geodesic flow equation is obtained by minimizing the functional. The geodesic flow equation is solved to obtain the characteristic manifold evolution trajectory. For example, the characteristic manifold evolution trajectory can be regarded as a collection of points that describe the evolution process of the features on the manifold.

[0125] Perform wavelet decomposition on the evolution trajectory of the characteristic manifold, calculate the covering dimension at each scale, and obtain the fractal feature sequence. For example, the evolution trajectory of the characteristic manifold can be decomposed into wavelet coefficients of different scales, and the covering dimension at each scale is calculated to obtain a sequence, namely the fractal feature sequence. Calculate the cumulative dimension of the fractal feature sequence at different scales, determine the dimension of the target feature space, and construct a high-dimensional feature space of the corresponding dimension. For example, assuming that the fractal feature sequence is [1.2, 1.5, 1.8, 2.1], by calculating the cumulative dimension, it can be determined that the dimension of the target feature space is 3.

[0126] Finally, the fractal feature sequence is transformed into a high-dimensional feature space by isometric mapping that maintains the geodesic distance to generate a reconstructed feature matrix. For example, assuming the fractal feature sequence is [1.2, 1.5, 1.8], it can be mapped into a three-dimensional space through isometric mapping to generate a matrix, namely the reconstructed feature matrix.

[0127] In this embodiment, by mapping the original features to a high-dimensional feature space, the nonlinear relationship between features can be better captured and the expressiveness of the features can be improved; by utilizing the idea of ​​manifold learning, the noise and outliers in the data can be effectively processed and the robustness of the model can be enhanced; by constructing a high-dimensional feature space, the computational complexity can be effectively reduced and the computational efficiency can be improved.

[0128] In an optional implementation, the reconstructed feature matrix is ​​input into a dual-branch verification network, and the dual-branch verification network analyzes the features from a time series branch and a frequency domain branch respectively; the time series branch identifies periodic variation features and mutation features based on the correlation of data in adjacent time periods, and generates a time series feature vector; the frequency domain branch identifies distortion features based on energy density distribution, and generating a frequency domain feature vector includes:

[0129] Inputting the reconstructed feature matrix into a recursive attention unit, the recursive attention unit reads the feature data stream through an input gate, filters the historical feature information through a forget gate, and generates preliminary temporal features through an output gate;

[0130] Based on a preset time window length, a corresponding sliding time window is set for the preliminary time series feature, two adjacent sliding time windows are traversed in sequence, the feature sequence mutual correlation coefficients are calculated, the feature sequence mutual correlation coefficients are formed into a feature correlation matrix, the feature correlation matrix is ​​subjected to singular value decomposition to obtain a left singular matrix, a singular value matrix and a right singular matrix, and the column vector of the left singular matrix corresponding to the maximum singular value in the singular value matrix is ​​selected as a periodic variation feature;

[0131] Performing kernel density estimation on the preliminary time series features to obtain a probability density function of a feature sequence, calculating a deviation between the probability density function and a preset reference density function, and identifying a feature mutation point through an adaptive threshold judgment unit;

[0132] Through the feature fusion module, the periodic change feature and the feature mutation point are weightedly combined to generate a time series feature vector;

[0133] Performing multi-scale wavelet decomposition on the reconstructed feature matrix, calculating wavelet coefficients using orthogonal wavelet basis functions, and taking the sum of the squares of the wavelet coefficients in different frequency bands as the frequency band energy density;

[0134] Calculating a query matrix, a key matrix and a value matrix of the frequency band energy density, obtaining a frequency band attention weight by matrix multiplication, weightedly combining the frequency band attention weight with the frequency band energy density, and identifying a frequency band distortion feature based on an energy density threshold;

[0135] A state transfer equation is constructed, and the position information of the frequency band distortion feature in the time domain and the frequency domain is input into the state transfer equation. The optimal state path is solved by a preset dynamic programming algorithm to generate a frequency domain feature vector.

[0136] First, the input time series data is preprocessed, such as data cleaning and normalization, to obtain a reconstructed feature matrix. Here, a sensor data containing 100 time points is taken as an example. After cleaning and normalization, the sensor data forms a 100×1 reconstructed feature matrix.

[0137] Then, the reconstructed feature matrix is ​​input into the recursive attention unit. The recursive attention unit contains an input gate, a forget gate, and an output gate. The input gate controls the input of the feature information of the current time step, the forget gate controls the degree of retention of the historical feature information, and the output gate controls the preliminary time series features of the final output. For example, at the 50th time point, the input gate reads the data of the 50th row in the reconstructed feature matrix, the forget gate decides to retain or discard part of the historical information based on the previous feature information, and the output gate finally generates the preliminary time series features of the 50th time point.

[0138] Next, extract the periodic change features from the preliminary time series features. Set the time window length to 10, and set a sliding time window for the preliminary time series features. Traverse two adjacent sliding time windows in sequence, such as the 1st to 10th time points and the 2nd to 11th time points, and calculate the mutual correlation coefficients of the feature sequences in these two time windows. The mutual correlation coefficients of all adjacent time windows form a feature correlation matrix. Perform singular value decomposition on the matrix to obtain a left singular matrix, a singular value matrix, and a right singular matrix. Select the column vector of the left singular matrix corresponding to the maximum singular value in the singular value matrix as the periodic change feature.

[0139] At the same time, mutation feature extraction is performed on the preliminary time series features. The probability density function of the feature sequence is calculated using the kernel density estimation method. The deviation between the probability density function and the preset reference density function (such as the standard normal distribution) is calculated. The feature mutation point is identified through the adaptive threshold judgment unit. For example, if the deviation between the probability density function of a certain time point and the reference density function exceeds the preset threshold, the time point is identified as a mutation point.

[0140] The periodic change features and feature mutation points are weighted and combined through the feature fusion module to generate a time series feature vector. For example, different weights can be assigned according to the importance of the periodic change features and mutation features, and then they can be linearly combined into a time series feature vector.

[0141] On the other hand, multi-scale wavelet decomposition is performed on the reconstructed feature matrix. The wavelet coefficients are calculated using orthogonal wavelet basis functions, and the sum of the squares of the wavelet coefficients in different frequency bands is taken as the frequency band energy density. For example, the reconstructed feature matrix is ​​decomposed into three frequency bands, and the sum of the squares of the wavelet coefficients in each frequency band is calculated to obtain the energy density of the three frequency bands.

[0142] The query matrix, key matrix, and value matrix of the band energy density are calculated, and the band attention weights are obtained by matrix multiplication. The band attention weights are weighted and combined with the band energy density to identify the band distortion features based on the energy density threshold. For example, if the energy density of a certain band exceeds a preset threshold, the band is considered to have distortion features.

[0143] Finally, a state transfer equation is constructed, and the position information of the frequency band distortion characteristics in the time domain and the frequency domain is input into the state transfer equation. The optimal state path is solved by a dynamic programming algorithm (such as the Viterbi algorithm) to generate a frequency domain feature vector.

[0144] In this embodiment, by combining time series and frequency domain analysis, the periodic changes, mutations and distortion features in the data can be more comprehensively captured, thereby improving the accuracy of feature analysis; recursive attention units and multi-scale wavelet decomposition can effectively deal with problems such as noise and data missing, thereby enhancing the robustness of feature recognition; the dynamic programming algorithm can quickly solve the optimal state path, thereby improving the efficiency of feature extraction.

[0145] In an optional implementation, a topological persistence coherence analyzer is constructed, and the topological relationship between features is extracted by calculating the Betti number and the persistence graph, and a feature connection structure is established; the time series feature vector, the frequency domain feature vector, and the feature connection structure are input into a dynamic coupler, and the feature structure is kept invariant based on the symplectic geometry principle, and a multidimensional feature space is constructed, including:

[0146] Mapping the time series feature vector and the frequency domain feature vector to the feature space, and calculating the distance value between the feature points in the feature space based on the Euclidean distance calculation formula;

[0147] A first distance threshold is set to start constructing a first simplicial complex: when the distance value between the feature points is less than the first distance threshold, a first connecting edge is established between the corresponding feature points; when the three feature points all have the first connecting edge, a first two-dimensional simplicial complex is constructed; when any three feature points among the four feature points all constitute the first two-dimensional simplicial complex, a first three-dimensional simplicial complex is constructed;

[0148] Based on a preset distance increment step, the first distance threshold is increased to obtain a second distance threshold, a second simplicial complex is constructed, and iteration is repeated until a preset maximum distance threshold is reached to obtain a simplicial complex sequence;

[0149] Calculate the zero-order Betti number of the simplicial complex sequence to obtain the number of connected components, calculate the first-order Betti number of the simplicial complex sequence to obtain the number of ring structures, and calculate the second-order Betti number of the simplicial complex sequence to obtain the number of cavity structures; construct a continuous homology group based on the number of connected components, the number of ring structures, and the number of cavity structures, and extract the generation time and extinction time of the topological structure in the continuous homology group; calculate the time interval between the generation time and the extinction time to obtain the topological feature life cycle, set a life cycle threshold, and screen out stable topological features that are greater than the life cycle threshold;

[0150] A symmetric adjacency matrix is ​​constructed based on the stable topological features, and the connection strength between the feature points is used as the matrix element of the symmetric adjacency matrix; a first tensor product operation is performed on the symmetric adjacency matrix and the time series feature vector to obtain a time series feature connection structure, and a second tensor product operation is performed on the symmetric adjacency matrix and the frequency domain feature vector to obtain a frequency domain feature connection structure;

[0151] The time series feature connection structure is used as a generalized coordinate, and the frequency domain feature connection structure is used as a generalized momentum to construct a symplectic geometric system; based on the generalized coordinates, the potential energy is determined, based on the generalized momentum, the kinetic energy is determined, the potential energy and the kinetic energy are added together to construct the energy function of the symplectic geometric system, the canonical equation group is determined, the canonical equation group is solved to obtain the dynamic evolution trajectory of the time series feature connection structure and the frequency domain feature connection structure, a multidimensional feature space is constructed, and the dynamic evolution trajectory is mapped into the multidimensional feature space to realize the dynamic coupling of the feature structures.

[0152] The connected component is a concept used in topology to describe mutually independent parts in space. A connected component refers to a region in a graph or geometric space where every pair of points can be connected by a path. Multiple connected components indicate that the space is divided into several unconnected parts.

[0153] The ring structure is a one-dimensional closed path in the topological space, usually called a "hole". It represents a structure in the space that is not enclosed by a boundary but is still closed. The ring structure has a direct correspondence in the first-order Betti number, reflecting the characteristics of one-dimensional connectivity in the space.

[0154] The cavity structure is a two-dimensional or higher-dimensional closed body in topological space, which encloses a hole or three-dimensional space and represents the three-dimensional cavity characteristics in the geometric shape. The second-order Betti number is used to count such cavity structures.

[0155] The symplectic geometry is a geometric theory used to describe the mathematical structure of a dynamic system, and is particularly suitable for studying systems with conservative properties. Symplectic geometry is based on phase space, unifies the momentum and position (generalized momentum and generalized coordinates) of the system under a framework, and characterizes the evolution law of the system through energy functions or Hamiltonian functions. Symplectic geometry systems have symmetry and conservation characteristics. Among them, generalized coordinates are a concept in classical mechanics, which are used to describe the state of the system. In a symplectic geometry system, generalized coordinates are not limited to spatial positions, but can also be an abstract description of the dynamic behavior of the system, which is suitable for analyzing complex characteristic structures or system states. Generalized momentum is another set of variables corresponding to generalized coordinates in a symplectic geometry system, which usually reflects the kinetic energy or dynamic characteristics in the system. It describes the momentum characteristics of the system, including both momentum in the traditional sense and dynamic states in some abstract senses.

[0156] First, extract the time series feature vector and frequency domain feature vector from the original data. For example, for a piece of audio, the time series feature can be the volume of each frame, and the frequency domain feature can be the spectrum energy distribution of each frame. Assuming there are 100 frames of audio data, the time series feature vector and the frequency domain feature vector are both 100-dimensional vectors.

[0157] Next, a topologically persistent homology analyzer is constructed to extract the topological relationship between features and establish a feature connection structure. Starting from a small distance threshold, such as 0.1, the Euclidean distance between any two feature points in the feature space is calculated. If the distance between two feature points is less than this threshold, a connecting edge is established between the two feature points. If there are connecting edges between any two of the three feature points, a two-dimensional simplicial complex is formed. Similarly, higher-dimensional simplicial complexes can be constructed.

[0158] Then, the distance threshold is gradually increased, for example, from 0.1 to 1.0, and the above steps are repeated at each distance threshold to construct a series of simplicial complexes. For example, when the distance threshold is 0.2, more feature points may be connected to each other, thus forming a more complex simplicial complex.

[0159] For each simplicial complex at each distance threshold, calculate its zero-order, first-order, and second-order Betti numbers. Betti numbers can be understood as a quantitative description of topological structures of different dimensions. The zero-order Betti number represents the number of connected components, the first-order Betti number represents the number of ring structures, and the second-order Betti number represents the number of cavity structures. For example, at a certain distance threshold, the calculated zero-order Betti number is 5, the first-order Betti number is 2, and the second-order Betti number is 0, which indicates that there are 5 connected components and 2 ring structures in the feature space at this time.

[0160] According to the Betti numbers at different distance thresholds, we can construct a continuous homology group and extract the generation and extinction time of the topological structure. For example, if a ring structure appears at a distance threshold of 0.2 and disappears at a distance threshold of 0.8, its generation time is 0.2 and its extinction time is 0.8.

[0161] Calculate the time interval between the generation time and the extinction time to get the topological feature life cycle. Set a life cycle threshold, such as 0.5, to filter out stable topological features with a life cycle greater than the threshold. In the above example, the life cycle of the ring structure is 0.6, which is greater than the threshold 0.5, so it is considered a stable topological feature.

[0162] A symmetric adjacency matrix is ​​constructed based on the selected stable topological features. The elements of the matrix represent the connection strength between feature points, which can be defined as the inverse of the distance between feature points or other suitable metrics. For example, if the distance between two feature points is 0.2, the connection strength between them can be defined as 1 / 0.2=5.

[0163] The adjacency matrix is ​​subjected to tensor product operations with the time series feature vector and the frequency domain feature vector respectively to obtain the time series feature connection structure and the frequency domain feature connection structure. The tensor product operation can be understood as integrating the connection strength information into the feature vector.

[0164] The time series feature connection structure is used as the generalized coordinates and the frequency domain feature connection structure is used as the generalized momentum to construct a symplectic geometry system. Based on the generalized coordinates and generalized momentum, the potential energy and kinetic energy of the system are determined, and the energy function of the system is obtained by adding the two. Then, the canonical equations are determined based on the energy function, and the equations are solved to obtain the dynamic evolution trajectory of the time series feature connection structure and the frequency domain feature connection structure.

[0165] Finally, a multidimensional feature space is constructed and the dynamic evolution trajectory is mapped into the space to achieve dynamic coupling of feature structures.

[0166] In this embodiment, by fusing topological information, the relationship between features can be more comprehensively characterized, thereby improving the expressive power of features; the stable topological features extracted based on topological continuous homology are highly robust to noise and disturbances, thereby enhancing the stability of the model; by simulating the dynamic evolution process of features through a symplectic geometry system, the inherent dynamic evolution laws of the data can be revealed.

[0167] In an optional implementation, the source of the deviation is located and a propagation path is established by tracking and analyzing the feature propagation link in the multidimensional feature space. A fractional-order differential operator is constructed based on the propagation path. The long-range dependency relationship between parameters on the propagation path is characterized by non-integer-order derivatives. A dependency intensity matrix is ​​generated. A causal test is performed on the dependency intensity matrix. The causal link from the source of the deviation to each propagation node is identified. A dynamic causal graph is constructed. The contribution of each propagation node is quantified according to the dynamic causal graph. The evolution mechanism of the deviation on the propagation path is analyzed. The output deviation analysis results include:

[0168] In the multidimensional feature space, a differential sequence of feature values ​​at adjacent moments is calculated to establish a feature propagation link, wherein the feature propagation link includes nodes and edges, wherein the nodes represent features, the edges represent propagation relationships between features, and the weights of the edges are determined based on the temporal correlation of the differential sequence;

[0169] In the feature propagation link, an abnormal point that deviates from the expected value is identified as a deviation abnormal point through a threshold analysis method, and the earliest appearing deviation abnormal point is determined as the deviation source along the edge with the largest weight in the feature propagation link; starting from the deviation source, the edge with the largest weight in the feature propagation link is traced forward until the deviation abnormal point, and the path from the deviation source to the deviation abnormal point is determined as the propagation path;

[0170] A fractional-order differential operator is constructed based on the propagation path, the order of the fractional-order differential operator is between zero and one, the characteristic values ​​of adjacent nodes on the propagation path are substituted into the fractional-order differential operator, the prediction error is minimized by an iterative optimization method, the optimal order of the fractional-order differential operator is determined, and the long-range dependency relationship between the parameters on the propagation path is obtained;

[0171] Based on the long-range dependency, the square integral of the error between the output value of the fractional differential operator and the actual eigenvalue on the propagation path is calculated to generate the dependency strength between adjacent nodes on the propagation path and organize it into a dependency strength matrix;

[0172] An autoregressive model containing historical information is constructed for the dependency strength matrix, and the significance of the causal relationship between the nodes in the dependency strength matrix is ​​determined by the F test method, so as to identify the causal link from the deviation source to each node on the propagation path;

[0173] Constructing a dynamic causal graph, taking all nodes on the propagation path as a node set of the dynamic causal graph, and taking the causal links as an edge set of the dynamic causal graph;

[0174] Based on the dynamic causal graph, the path probability from the deviation source through each node on the propagation path to the target node is calculated, the path probability is normalized to obtain a normalized path probability, the normalized path probability is multiplied by the corresponding dependency strength in the dependency strength matrix and the sum is calculated to obtain the contribution of each node on the propagation path;

[0175] According to the contribution of each node on the propagation path and the time correlation relationship, the evolution mechanism of the deviation on the propagation path is analyzed, and the deviation analysis result is output based on the evolution mechanism.

[0176] First, key operating parameter data at the grid connection point is collected from the new energy station monitoring system. These parameters include active power, reactive power, voltage amplitude, current amplitude, power factor, frequency, etc. The collection process uses a high-precision sampling device, and the sampling period is to record data every 100 milliseconds. The collected time series data is preprocessed to remove obvious error data and communication interruption data to form a continuous and complete data sequence.

[0177] For the preprocessed data sequence, the difference between the parameter values ​​at two adjacent sampling moments is calculated to obtain the differential sequence. The correlation between the differential sequences is analyzed by the sliding time window method. The sliding window length is set to 10 seconds, and the window sliding step is 1 second. In each sliding window, the correlation coefficient between the differential sequences of different parameters is calculated to establish the correlation relationship between the parameters. The calculated correlation coefficient is used as the weight of the edge in the feature propagation link. The larger the correlation coefficient, the stronger the propagation correlation between the parameters.

[0178] In the feature propagation link, set an abnormal threshold based on the statistical analysis of historical operating data. Compare the real-time operating parameter value with the normal value required by the grid connection standard. When the parameter deviation exceeds the set threshold, mark the point as a deviation abnormal point. Starting from all the deviation abnormal points found, gradually trace back along the edge with the largest weight in the feature propagation link. Determine the earliest abnormal point by timestamp comparison and identify it as the source of the deviation. Starting from the source of the deviation, trace forward along the edge with the largest weight until all related abnormal points are traversed, thereby obtaining a complete deviation propagation path.

[0179] Based on the determined propagation path, a fractional differential operator is constructed to describe the nonlinear propagation relationship between parameters. The optimal fractional order is found by iteration. First, the initial order value is set, and the parameter values ​​of adjacent nodes on the propagation path are substituted into the fractional differential operator. The prediction error between the fractional differential result and the actual parameter value is calculated. The order value is continuously adjusted using the gradient descent algorithm until the prediction error is minimized. The optimal order finally determined reflects the long-range dependency characteristics between parameters.

[0180] The square integral of the error between the output value and the actual parameter value on the propagation path is calculated using the optimal order fractional differential operator. This integral value characterizes the degree of dependence between adjacent parameters. This calculation is performed for all adjacent parameter pairs on the propagation path to generate a complete dependency strength matrix. Each element in the matrix represents the dependency strength between the corresponding parameter pairs.

[0181] An autoregressive model is constructed for the dependency intensity matrix, and the historical change information of the parameters is introduced. The F test method is used to evaluate the significance of the causal relationship between the parameters. According to the set significance level, it is judged whether there is a significant causal relationship between the parameter pairs in the dependency intensity matrix. The parameter pairs that pass the significance test constitute a causal link.

[0182] A dynamic causal graph is constructed based on the identified causal links. The parameters on the propagation path are used as nodes in the graph, and directed edges are connected between pairs of parameters with significant causal relationships. The direction of the edge indicates the direction of the causal relationship, forming a complete dynamic causal relationship network.

[0183] In the dynamic causal graph, the probabilities of all possible paths from the deviation source node to the target node are calculated. These path probabilities are normalized to obtain the standardized path importance. The normalized path probability is multiplied by the corresponding strength value in the dependency strength matrix and summed to obtain the contribution of each node on the propagation path to the system deviation.

[0184] Analyze the change pattern of each node's contribution over time, and study the diffusion and evolution of the deviation on the propagation path. Based on the temporal change characteristics of the contribution, clarify the main paths and key nodes of deviation propagation, and provide decision-making basis for system regulation.

[0185] For example, during the grid-connected operation of a photovoltaic power station, the monitoring system detected abnormalities in multiple parameters such as voltage and frequency at 9:30. Through the above method, it was found that the active power first fluctuated at 9:28, and the fluctuation amplitude exceeded 15% of the rated value; the voltage parameter then shifted at 9:29, and the voltage amplitude decreased to 92% of the nominal value; finally, the frequency fluctuated at 9:30, deviating from the rated value by 0.3Hz. The analysis results show that the active power anomaly is the source, causing system disturbances through the propagation path of "active power-voltage-frequency". It was found that the active output of the photovoltaic inverter fluctuated violently due to the sudden change in light intensity, which triggered a series of parameter deviations. Based on the analysis results, the operation and maintenance personnel promptly optimized the active control strategy of the inverter and improved the dynamic response capability of the system.

[0186] In this embodiment, through feature propagation link backtracking and threshold analysis, the root cause of the deviation can be located quickly and accurately, avoiding blind troubleshooting in a complex system; by using fractional calculus to characterize the long-range dependency between parameters, combined with causality tests and dynamic causal graphs, the evolution of the deviation on the propagation path can be deeply analyzed, revealing the inherent mechanism of deviation propagation; by calculating the contribution of each node, the role of each node in the deviation propagation process can be quantitatively evaluated, providing a basis for targeted intervention and optimization.

[0187] Figure 2 Schematic diagram of the structure of the real-time deviation analysis system of the grid-connected operation data of the new energy station according to the embodiment of the present invention. Figure 2 As shown, the system comprises:

[0188] The first unit is used to collect real-time grid-connected operation data of new energy stations, input the real-time grid-connected operation data into an adaptive decomposition unit, dynamically adjust the decomposition window width based on the data fluctuation trend, and perform time domain decomposition on the real-time grid-connected operation data to obtain multiple groups of original feature sequences; construct a recursive support model, dynamically filter feature dimensions of the multiple groups of original feature sequences through information entropy, and obtain filtered feature subsets; input the filtered feature subsets into a manifold embedding network, construct a high-dimensional feature space based on Riemannian geometry, and generate a reconstructed feature matrix;

[0189] The second unit is used to input the reconstructed feature matrix into a dual-branch verification network, and the dual-branch verification network analyzes the features from the time series branch and the frequency domain branch respectively; the time series branch identifies the periodic change features and the mutation features based on the data correlation of adjacent time periods, and generates a time series feature vector; the frequency domain branch identifies the distortion features based on the energy density distribution, and generates a frequency domain feature vector; a topological persistence coherence analyzer is constructed, and the topological relationship between features is extracted by calculating the Betti number and the persistence diagram, and a feature connection structure is established; the time series feature vector, the frequency domain feature vector and the feature connection structure are input into a dynamic coupler, and the feature structure is kept invariant based on the symplectic geometry principle to construct a multidimensional feature space;

[0190] The third unit is used to track and analyze the feature propagation link in the multi-dimensional feature space, locate the source of the deviation, establish a propagation path, construct a fractional-order differential operator based on the propagation path, characterize the long-range dependency between parameters on the propagation path through non-integer-order derivatives, generate a dependency intensity matrix, perform a causal test on the dependency intensity matrix, identify the causal link from the source of the deviation to each propagation node, construct a dynamic causal graph, quantify the contribution of each propagation node according to the dynamic causal graph, analyze the evolution mechanism of the deviation on the propagation path, and output the deviation analysis result.

[0191] According to a third aspect of the embodiments of the present invention,

[0192] An electronic device is provided, comprising:

[0193] processor;

[0194] a memory for storing processor-executable instructions;

[0195] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0196] A fourth aspect of the embodiments of the present invention is:

[0197] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0198] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0199] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A real-time deviation analysis method for grid-connected operation data of new energy stations, characterized in that: include: Collecting real-time grid-connected operation data of new energy stations, inputting the real-time grid-connected operation data into an adaptive decomposition unit, dynamically adjusting the decomposition window width based on the data fluctuation trend, performing time domain decomposition on the real-time grid-connected operation data, and obtaining multiple groups of original feature sequences; A recursive support model is constructed to dynamically screen the feature dimensions of the multiple groups of original feature sequences through information entropy to obtain a screened feature subset; the screened feature subset is input into a manifold embedding network, a high-dimensional feature space is constructed based on Riemannian geometry, and a reconstructed feature matrix is ​​generated; The reconstructed feature matrix is ​​input into a dual-branch verification network, and the dual-branch verification network analyzes the features from the time series branch and the frequency domain branch respectively; the time series branch identifies the periodic change features and the mutation features based on the data correlation of adjacent time periods, and generates a time series feature vector; the frequency domain branch identifies the distortion features based on the energy density distribution, and generates a frequency domain feature vector; a topological persistence coherence analyzer is constructed, and the topological relationship between features is extracted by calculating the Betti number and the persistence diagram, and a feature connection structure is established; the time series feature vector, the frequency domain feature vector and the feature connection structure are input into a dynamic coupler, and the feature structure is kept invariant based on the symplectic geometry principle to construct a multidimensional feature space; In the multi-dimensional feature space, the source of the deviation is located and a propagation path is established by tracking and analyzing the feature propagation link. A fractional-order differential operator is constructed based on the propagation path. The long-range dependency relationship between parameters on the propagation path is characterized by non-integer-order derivatives. A dependency intensity matrix is ​​generated. A causal test is performed on the dependency intensity matrix. The causal link from the source of the deviation to each propagation node is identified, a dynamic causal graph is constructed, the contribution of each propagation node is quantified according to the dynamic causal graph, the evolution mechanism of the deviation on the propagation path is analyzed, and the deviation analysis result is output.

2. The method according to claim 1, characterized in that Collect the real-time grid-connected operation data of the new energy station, input the real-time grid-connected operation data into the adaptive decomposition unit, dynamically adjust the decomposition window width based on the data fluctuation trend, perform time domain decomposition on the real-time grid-connected operation data, and obtain multiple groups of original feature sequences including: Input the real-time grid-connected operation data into an orthogonal wavelet analysis unit, perform continuous wavelet transform using orthogonal wavelet basis functions, obtain a time-domain-frequency domain energy distribution diagram, calculate the energy integral value of each frequency band in the time-domain-frequency domain energy distribution diagram, sort the energy integral values ​​by size, generate a frequency band energy sequence, calculate the energy accumulation ratio based on the frequency band energy sequence, extract the frequency band information whose energy accumulation ratio meets a preset threshold, and construct a target frequency band sequence; map the target frequency band sequence to the time domain, establish a time domain mapping matrix, and calculate the gradient change value of the time domain mapping matrix; Performing Hilbert transform on the real-time grid-connected operation data to obtain an analytical signal, calculating an instantaneous frequency value based on a phase function of the analytical signal, setting the instantaneous frequency value as a center frequency, determining a bandwidth range according to the gradient change value, constructing a corresponding bandpass filter, filtering the real-time grid-connected operation data through the bandpass filter, and obtaining a filtered signal sequence; Processing the filtered signal sequence using a variational mode decomposition method, setting the decomposition dimension to the length of the target frequency band sequence, and obtaining a mode decomposition sequence; Calculating the energy value of each modal component in the modal decomposition sequence respectively to obtain a modal energy sequence, and dividing each modal energy value in the modal energy sequence by the sum of the modal energy sequence to obtain a modal energy proportion sequence; Selecting the maximum modal component in the modal energy proportion sequence, extracting the period of the maximum modal component as a reference period, and updating the decomposition window width with the reference period; Based on the updated decomposition window width, the real-time grid-connected operation data is segmented to obtain multiple groups of original feature sequences.

3. The method according to claim 1, characterized in that A recursive support model is constructed to dynamically filter feature dimensions of the multiple groups of original feature sequences through information entropy, and the filtered feature subsets are obtained, including: Input the multiple groups of original feature sequences into a deep belief network, perform layer-by-layer feature transformation through a multi-layer restricted Boltzmann machine, obtain a feature hidden layer representation sequence, calculate the quantum information entropy of the feature hidden layer representation sequence, and construct a feature uncertainty matrix; construct an attention weight network, calculate the correlation scores between feature dimensions based on the feature uncertainty matrix, and generate a feature importance sequence; Input the feature importance sequence into a pre-built recursive neural network, extract the feature temporal dependency, calculate the migration entropy value between feature dimensions, and build a conditional entropy matrix based on the migration entropy value; Inputting the feature importance sequence and the conditional entropy matrix into a Monte Carlo tree search module, and iteratively optimizing the feature combination scheme using a policy gradient method; Calculate the entropy increase rate of the feature combination scheme, determine the feature redundancy threshold based on the entropy increase rate, and eliminate the feature dimensions whose information gain rate is lower than the feature redundancy threshold; calculate the recursive support of each remaining feature dimension based on the feature importance sequence and the conditional entropy matrix, arrange the recursive support in descending order, and accumulate the recursive support values ​​in sequence to obtain the support accumulation value, select the feature dimensions whose support accumulation value reaches the preset ratio threshold, and construct the filtered feature subset.

4. The method according to claim 1, characterized in that The filtered feature subset is input into the manifold embedding network, a high-dimensional feature space is constructed based on Riemannian geometry, and a reconstructed feature matrix is ​​generated, including: Inputting the filtered feature subset into a feature mapping layer, processing the real component of the filtered feature subset using a hyperbolic tangent function, processing the imaginary component of the filtered feature subset using a cosine function, and generating a complex-valued feature map; The complex-valued feature map is projected to the manifold tangent space through an exponential mapping function, and the minimum path integral between feature points is calculated to obtain the geodesic distance; based on the geodesic distance, a local metric tensor is constructed, and the second-order covariant derivative of the local metric tensor is calculated through coordinate transformation to obtain the manifold curvature tensor; based on the manifold curvature tensor, a manifold curvature constraint is constructed through a cross-sectional curvature criterion, and the manifold curvature constraint is input into the metric learning layer; In the metric learning layer, a covariant derivative operator is constructed using an affine connection, and a parallel shift equation group is calculated through a third-order connection coefficient, and a characteristic manifold transmission matrix is ​​obtained by solving the equation. A variational functional is established according to the characteristic manifold transmission matrix, and a geodesic flow equation is obtained by minimizing the variational functional. The geodesic flow equation is solved to obtain the characteristic manifold evolution trajectory. Performing wavelet decomposition on the evolution trajectory of the characteristic manifold, calculating the covering dimension at each scale, obtaining a fractal feature sequence, calculating the cumulative dimension of the fractal feature sequence at different scales, determining the dimension of the target feature space, and constructing a high-dimensional feature space of corresponding dimension; The fractal feature sequence is transformed into the high-dimensional feature space by isometric mapping that maintains the geodesic distance, so as to generate a reconstructed feature matrix.

5. The method according to claim 1, characterized in that The reconstructed feature matrix is ​​input into a dual-branch verification network, and the dual-branch verification network analyzes the features from a time series branch and a frequency domain branch respectively; the time series branch identifies periodic variation features and mutation features based on the correlation of data in adjacent time periods, and generates a time series feature vector; The frequency domain branch identifies distortion features based on energy density distribution and generates a frequency domain feature vector including: Inputting the reconstructed feature matrix into a recursive attention unit, the recursive attention unit reads the feature data stream through an input gate, filters the historical feature information through a forget gate, and generates preliminary temporal features through an output gate; Based on a preset time window length, a corresponding sliding time window is set for the preliminary time series feature, two adjacent sliding time windows are traversed in sequence, the feature sequence mutual correlation coefficients are calculated, the feature sequence mutual correlation coefficients are formed into a feature correlation matrix, the feature correlation matrix is ​​subjected to singular value decomposition to obtain a left singular matrix, a singular value matrix and a right singular matrix, and the column vector of the left singular matrix corresponding to the maximum singular value in the singular value matrix is ​​selected as a periodic variation feature; Performing kernel density estimation on the preliminary time series features to obtain a probability density function of a feature sequence, calculating a deviation between the probability density function and a preset reference density function, and identifying a feature mutation point through an adaptive threshold judgment unit; Through the feature fusion module, the periodic change feature and the feature mutation point are weightedly combined to generate a time series feature vector; Performing multi-scale wavelet decomposition on the reconstructed feature matrix, calculating wavelet coefficients using orthogonal wavelet basis functions, and taking the sum of the squares of the wavelet coefficients in different frequency bands as the frequency band energy density; Calculating a query matrix, a key matrix and a value matrix of the frequency band energy density, obtaining a frequency band attention weight by matrix multiplication, weightedly combining the frequency band attention weight with the frequency band energy density, and identifying a frequency band distortion feature based on an energy density threshold; A state transfer equation is constructed, and the position information of the frequency band distortion feature in the time domain and the frequency domain is input into the state transfer equation. The optimal state path is solved by a preset dynamic programming algorithm to generate a frequency domain feature vector.

6. The method according to claim 1, characterized in that Constructing a topological persistence coherence analyzer, extracting the topological relationship between features and establishing a feature connection structure by calculating the Betti number and the persistence graph; inputting the time series feature vector, the frequency domain feature vector and the feature connection structure into a dynamic coupler, maintaining the invariance of the feature structure based on the principle of symplectic geometry, and constructing a multidimensional feature space, including: Mapping the time series feature vector and the frequency domain feature vector to the feature space, and calculating the distance value between the feature points in the feature space based on the Euclidean distance calculation formula; A first distance threshold is set to start constructing a first simplicial complex: when the distance value between the feature points is less than the first distance threshold, a first connecting edge is established between the corresponding feature points; when the three feature points all have the first connecting edge, a first two-dimensional simplicial complex is constructed; when any three feature points among the four feature points all constitute the first two-dimensional simplicial complex, a first three-dimensional simplicial complex is constructed; Based on a preset distance increment step, the first distance threshold is increased to obtain a second distance threshold, a second simplicial complex is constructed, and iteration is repeated until a preset maximum distance threshold is reached to obtain a simplicial complex sequence; Calculate the zero-order Betti number of the simplicial complex sequence to obtain the number of connected components, calculate the first-order Betti number of the simplicial complex sequence to obtain the number of ring structures, and calculate the second-order Betti number of the simplicial complex sequence to obtain the number of cavity structures; construct a continuous homology group based on the number of connected components, the number of ring structures, and the number of cavity structures, and extract the generation time and extinction time of the topological structure in the continuous homology group; calculate the time interval between the generation time and the extinction time to obtain the topological feature life cycle, set a life cycle threshold, and screen out stable topological features that are greater than the life cycle threshold; A symmetric adjacency matrix is ​​constructed based on the stable topological features, and the connection strength between the feature points is used as the matrix element of the symmetric adjacency matrix; a first tensor product operation is performed on the symmetric adjacency matrix and the time series feature vector to obtain a time series feature connection structure, and a second tensor product operation is performed on the symmetric adjacency matrix and the frequency domain feature vector to obtain a frequency domain feature connection structure; The time series feature connection structure is used as a generalized coordinate, and the frequency domain feature connection structure is used as a generalized momentum to construct a symplectic geometric system; based on the generalized coordinates, the potential energy is determined, based on the generalized momentum, the kinetic energy is determined, the potential energy and the kinetic energy are added together to construct the energy function of the symplectic geometric system, the canonical equation group is determined, the canonical equation group is solved to obtain the dynamic evolution trajectory of the time series feature connection structure and the frequency domain feature connection structure, a multidimensional feature space is constructed, and the dynamic evolution trajectory is mapped into the multidimensional feature space to realize the dynamic coupling of the feature structures.

7. The method according to claim 1, characterized in that In the multidimensional feature space, the source of the deviation is located by tracking and analyzing the feature propagation link, a propagation path is established, a fractional-order differential operator is constructed based on the propagation path, the long-range dependency relationship between parameters on the propagation path is characterized by non-integer-order derivatives, a dependency intensity matrix is ​​generated, a causal test is performed on the dependency intensity matrix, the causal link from the source of the deviation to each propagation node is identified, a dynamic causal graph is constructed, the contribution of each propagation node is quantified according to the dynamic causal graph, the evolution mechanism of the deviation on the propagation path is analyzed, and the output deviation analysis results include: In the multidimensional feature space, a differential sequence of feature values ​​at adjacent moments is calculated to establish a feature propagation link, wherein the feature propagation link includes nodes and edges, wherein the nodes represent features, the edges represent propagation relationships between features, and the weights of the edges are determined based on the temporal correlation of the differential sequence; In the feature propagation link, an abnormal point that deviates from the expected value is identified as a deviation abnormal point through a threshold analysis method, and the earliest appearing deviation abnormal point is determined as the deviation source along the edge with the largest weight in the feature propagation link; starting from the deviation source, the edge with the largest weight in the feature propagation link is traced forward until the deviation abnormal point, and the path from the deviation source to the deviation abnormal point is determined as the propagation path; A fractional-order differential operator is constructed based on the propagation path, the order of the fractional-order differential operator is between zero and one, the characteristic values ​​of adjacent nodes on the propagation path are substituted into the fractional-order differential operator, the prediction error is minimized by an iterative optimization method, the optimal order of the fractional-order differential operator is determined, and the long-range dependency relationship between the parameters on the propagation path is obtained; Based on the long-range dependency, the square integral of the error between the output value of the fractional differential operator and the actual eigenvalue on the propagation path is calculated to generate the dependency strength between adjacent nodes on the propagation path and organize it into a dependency strength matrix; An autoregressive model containing historical information is constructed for the dependency strength matrix, and the significance of the causal relationship between the nodes in the dependency strength matrix is ​​determined by the F test method, so as to identify the causal link from the deviation source to each node on the propagation path; Constructing a dynamic causal graph, taking all nodes on the propagation path as a node set of the dynamic causal graph, and taking the causal links as an edge set of the dynamic causal graph; Based on the dynamic causal graph, the path probability from the deviation source through each node on the propagation path to the target node is calculated, the path probability is normalized to obtain a normalized path probability, the normalized path probability is multiplied by the corresponding dependency strength in the dependency strength matrix and the sum is calculated to obtain the contribution of each node on the propagation path; According to the contribution of each node on the propagation path and the time correlation relationship, the evolution mechanism of the deviation on the propagation path is analyzed, and the deviation analysis result is output based on the evolution mechanism.

8. A real-time deviation analysis system for grid-connected operation data of a new energy station, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect real-time grid-connected operation data of new energy stations, input the real-time grid-connected operation data into an adaptive decomposition unit, dynamically adjust the decomposition window width based on the data fluctuation trend, perform time domain decomposition on the real-time grid-connected operation data, and obtain multiple groups of original feature sequences; Constructing a recursive support model, dynamically screening feature dimensions of the multiple groups of original feature sequences through information entropy, and obtaining a screened feature subset; Inputting the filtered feature subset into a manifold embedding network, constructing a high-dimensional feature space based on Riemannian geometry, and generating a reconstructed feature matrix; The second unit is used to input the reconstructed feature matrix into a dual-branch verification network, and the dual-branch verification network analyzes the features from the time series branch and the frequency domain branch respectively; the time series branch identifies the periodic change features and the mutation features based on the data correlation of adjacent time periods, and generates a time series feature vector; the frequency domain branch identifies the distortion features based on the energy density distribution, and generates a frequency domain feature vector; a topological persistence coherence analyzer is constructed, and the topological relationship between features is extracted by calculating the Betti number and the persistence diagram, and a feature connection structure is established; the time series feature vector, the frequency domain feature vector and the feature connection structure are input into a dynamic coupler, and the feature structure is kept invariant based on the symplectic geometry principle to construct a multidimensional feature space; The third unit is used to track and analyze the feature propagation link in the multi-dimensional feature space, locate the source of the deviation, establish a propagation path, construct a fractional-order differential operator based on the propagation path, characterize the long-range dependency between parameters on the propagation path through non-integer-order derivatives, generate a dependency intensity matrix, perform a causal test on the dependency intensity matrix, identify the causal link from the source of the deviation to each propagation node, construct a dynamic causal graph, quantify the contribution of each propagation node according to the dynamic causal graph, analyze the evolution mechanism of the deviation on the propagation path, and output the deviation analysis result.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images