Distribution transformer load prediction method based on multi-scale fluctuation mode adaptive decomposition

Abstract

The invention discloses a distribution transformer load prediction method and system based on multi-scale fluctuation mode adaptive decomposition, and relates to the technical field of power system load prediction and power grid safe operation. According to the method, efficient prediction of the load of the power distribution network is realized through technologies of improved maximum overlapping discrete wavelet packet transformation, adaptive density estimation, a multi-level clustering strategy, dynamic time warping, anomaly detection optimization and the like. According to the method, distribution of load signal energy is optimized, a fluctuation mode is finely represented, an efficient index structure is constructed, and prediction accuracy and robustness are improved through an online correction mechanism. According to the whole process, the precision and efficiency of load prediction are remarkably improved, the method adapts to complex changes of load data, the beneficial effects of reducing the calculation amount, improving the prediction stability and capturing nonlinear correlation characteristics are achieved, and reliable support is provided for optimized operation of the power distribution network.

Description

Technical Field

[0001] The present invention relates to the technical field of power system load forecasting and power grid safe operation, and in particular to a distribution transformer load forecasting method based on multi-scale fluctuation mode adaptive decomposition. Background Art

[0002] In the field of power system load forecasting, related technologies have developed rapidly, and traditional forecasting methods such as time series analysis and regression analysis have been widely used. However, with the complexity of distribution network structure and the diversification of load characteristics, these traditional methods have shown certain limitations in forecasting accuracy and adaptability. In recent years, data-driven forecasting technology, especially methods combining signal processing and machine learning, has gradually become a research hotspot, providing a new perspective for load forecasting.

[0003] Although the existing technology has made progress in some aspects, it still has deficiencies in the accuracy and adaptability of distribution transformer load forecasting. First, traditional methods often have difficulty in dealing with nonlinear characteristics and volatility in load data, resulting in inaccurate forecast results. Second, the existing technology still has room for improvement in feature extraction and characterization of load data, especially when dealing with multi-time scale fluctuation patterns. In addition, outliers and noise have a greater impact on the forecast results, and the existing technology has limited means of processing in this regard.

[0004] At present, the main technical defects of distribution transformer load prediction are as follows:

[0005] Traditional time series forecasting methods (such as ARIMA, SARIMA, etc.) cannot effectively handle the multi-scale characteristics of load data, and the forecasting accuracy is limited.

[0006] Although deep learning methods (such as LSTM, Transformer, etc.) have powerful nonlinear fitting capabilities, they have poor adaptability to small sample scenarios and insufficient model interpretability.

[0007] Existing decomposition prediction methods (such as EMD, VMD, etc.) have problems such as mode aliasing and endpoint effect, and lack the ability to adaptively identify fluctuation characteristics of different scales.

[0008] Existing methods generally ignore the handling of abnormal fluctuations in load data, resulting in poor stability of prediction results.

[0009] In summary, the present invention has significant advantages in improving the accuracy, adaptability and robustness of distribution transformer load forecasting, and is expected to provide strong support for the safe operation and optimized scheduling of power systems. Summary of the invention

[0010] In view of the above-mentioned existing problems, the present invention aims to solve the problems of how to achieve adaptive multi-scale decomposition of load data to avoid mode aliasing and endpoint effects; how to construct a scientific fluctuation feature characterization system to achieve accurate characterization of fluctuation patterns of different scales; how to establish a dynamic pattern matching mechanism to improve the adaptability of the prediction model to load changes; and how to effectively handle abnormal fluctuations in load data to improve the reliability of prediction results.

[0011] In order to solve the above technical problems, a distribution transformer load forecasting method based on multi-scale fluctuation pattern adaptive decomposition is proposed, including:

[0012] The original data is input, and the improved maximum overlap discrete wavelet packet transform is used to perform multi-scale decomposition of the load signal, optimize the energy distribution, perform feature extraction and characterization optimization, and build a feature fluctuation library to store features; through adaptive density estimation and improved multi-level clustering strategy, the fluctuation pattern is finely characterized from different time scales, an efficient index structure is constructed, and prediction is performed through adaptive matching; through improved dynamic time warping, the distance metric and search strategy are optimized, and through the improved particle swarm optimization algorithm, the weights are dynamically adjusted to build a combined prediction framework; combined with integrated isolation forests and local anomaly factors, anomaly detection is optimized, and the online correction mechanism is optimized through adaptive windows and robust estimation.

[0013] As a preferred solution of the distribution transformer load prediction method based on multi-scale fluctuation mode adaptive decomposition described in the present invention, wherein: the improved maximum overlap discrete wavelet packet transform performs multi-scale decomposition of the load signal, including using the sym8 wavelet basis function with good symmetry as the basic decomposition tool, the support is set to 16, and the optimal wavelet basis is selected by minimizing the reconstruction error, and the error function E is:

[0014]

[0015] Among them, x is the original signal, x' is the reconstructed signal;

[0016] Set the reconstruction error threshold to 10 -6 , when the reconstruction error is less than 10 -6 When it is considered that the wavelet basis is appropriately selected;

[0017] At the same time, the wavelet basis function must satisfy three key properties at the same time, the first key property is orthogonality, the second key property is compact support, and the third key property is vanishing moment;

[0018] The endpoint effect problem in load data is processed by adaptive extension algorithm, and AR(p) autoregressive model is established to describe the time correlation of load data. The model order p is dynamically determined by minimizing the AIC criterion, and the extension length L is adaptively determined according to the data length N and the signal-to-noise ratio SNR. Denoising is performed by the improved soft threshold shrinkage method, and the threshold parameter λ j A scale-dependent adaptive form was adopted, robust estimation was performed using the median absolute deviation method, threshold selection was optimized using the SURE criterion, and the optimal threshold was determined by minimizing the unbiased risk estimate;

[0019] The energy distribution optimization includes, for the j-th layer wavelet decomposition coefficient, calculating the wavelet energy ratio of the j-th layer, setting an adaptive energy threshold, calculating the mutual information between adjacent decomposition layers through a constraint mechanism based on mutual information, when the mutual information value exceeds a threshold value θ=0.1, it indicates that there is information redundancy and a hierarchical adjustment is required, dynamically determining the optimal hierarchical level through an improved information gain calculation method, defining a distance function d(i,j) through a hierarchical merging strategy based on dynamic programming, merging when d(i,j)<δ, δ is determined by a cross-validation method, wherein d(i,j) is the distance between adjacent layers.

[0020] As a preferred solution of the distribution transformer load prediction method based on multi-scale fluctuation pattern adaptive decomposition described in the present invention, wherein: the feature extraction and characterization optimization includes deep mining and enhancement processing of distribution transformer load data to construct a feature fluctuation library;

[0021] The load data includes an amplitude feature family and a shape feature family, wherein the amplitude feature family includes a maximum amplitude, a root mean square amplitude, a peak factor, a margin factor, and a waveform factor; and the shape feature family includes a skewness, a kurtosis, a waveform index, and a pulse index;

[0022] The frequency domain features are enhanced by an improved power spectrum density estimation method, and the final power spectrum density is obtained by weighted superposition of multiple independent spectrum estimation results using multi-window spectrum estimation technology.

[0023] The energy characteristics of the frequency bands extracted through energy distribution characterization include sub-band energy ratio, energy entropy, spectrum centroid, and spectrum broadening. The sub-band energy ratio includes calculating the proportion of each sub-band energy to the total energy; obtaining the spectrum centroid through energy-weighted averaging to obtain the center position of the energy distribution; and obtaining the spectrum broadening by calculating the second-order moment of the frequency relative to the centroid.

[0024] As a preferred solution of the distribution transformer load forecasting method based on adaptive decomposition of multi-scale fluctuation patterns described in the present invention, wherein: the adaptive density estimation and improved multi-level clustering strategy include characterizing data distribution characteristics through improved kernel density estimation, selecting Epanechnikov kernel function, using adaptive method to select bandwidth, and determining density threshold based on dynamic determination method of density threshold of silhouette coefficient;

[0025] After characterization, the fluctuation pattern recognition under different time scales is completed based on the improved multi-level clustering strategy. According to the inherent law of load change in the distribution network, the time scale is divided into four levels including intra-day layer, daily layer, weekly layer and monthly layer:

[0026] The intra-day layer covers the fluctuation characteristics of 15 minutes to 4 hours, reflecting the short-term electricity consumption behavior of users; the daily layer covers the change pattern of 4 hours to 24 hours, reflecting the characteristics of the daily load curve; the weekly layer spans 1 day to 7 days, describing the difference between weekdays and weekends; the monthly layer ranges from 7 days to 30 days, capturing the seasonal change characteristics;

[0027] In each time scale layer, an improved density clustering algorithm is used to identify the fluctuation pattern. In the process of boundary point absorption, the density sensitivity factor dsf is introduced to determine the attribution of the boundary point. The judgment condition is:

[0028] d(p,q)≤eps·(ρ p / ρ q ) dsf

[0029] Among them, d)p,q) is the distance between point p and point q, ρ p and ρ q are the corresponding density values, eps is the neighborhood radius, where dsf∈[0.2,0.5]. When the dsf value is less than 0.35, a relatively tight clustering is formed; when the dsf value is greater than 0.35, points with large density differences are allowed to be classified into the same category;

[0030] The fine characterization of the fluctuation pattern includes using an improved Gaussian mixture model for probability distribution modeling and estimating parameters through an optimized expectation maximization algorithm:

[0031] Q(θ,θ t ) = E[logL(θ;X,Z)|X,θ t ]

[0032] Among them, Q(θ,θ t ) is the conditional expectation, θ is the model parameter, θ t is the current iteration parameter, L(θ; X, Z) is the likelihood function containing latent variables, and E[·] is the expectation operator;

[0033] Through the adaptive learning rate adjustment mechanism, the parameter update step size is dynamically adjusted according to the gradient change of the likelihood function in each iteration. At the same time, the Bayesian information criterion is used to adaptively determine the optimal number of components K:

[0034] K = arg min (BIC)

[0035] BIC=-2Log+KlogM

[0036] Among them, BIC is the Bayesian information criterion, Log is the log-likelihood value, and M is the number of samples;

[0037] At the same time, the nonlinear correlation characteristics in the load data of distribution equipment are characterized by the copula function framework:

[0038] C(u1,...,u d )=P(U1≤u1,...,U d ≤u d )

[0039] Among them, u d represents the cumulative distribution function value of the dth random variable, C represents the Copula joint distribution function, u1,...,u d represents the cumulative distribution function value of the marginal distribution, U1,...,U d Represents a random variable, and d represents the dimension. Through the copula function, the joint distribution of multidimensional random variables is decomposed into two parts: marginal distribution and dependency structure. According to the characteristics of load data, t-copula is selected as the basic dependency structure, and the relevant parameters are estimated by the maximum likelihood method. At the same time, the selection of marginal distribution adopts an adaptive strategy.

[0040] By describing the nonlinear correlation characteristics in the load data of power distribution equipment, multi-dimensional features are extracted, including constructing statistical feature system and morphological feature system;

[0041] The statistical moment feature system includes mean, variance, skewness and kurtosis. By calculating the mean

[0042] σ 4 Characterize the asymmetry and peak characteristics of load distribution;

[0043] The morphological feature system includes an opening operation to suppress peaks in a signal, a closing operation to fill valleys in a signal, and a mutation feature in a morphological gradient detection load curve.

[0044] As a preferred solution of the distribution transformer load forecasting method based on multi-scale fluctuation pattern adaptive decomposition described in the present invention, wherein: the construction of an efficient index structure includes building a multi-layer hybrid index and retrieval mechanism and supporting fast fluctuation pattern retrieval including multi-layer hybrid index, feature dimension index, and supporting fast retrieval of fluctuation patterns;

[0045] The multi-layer hybrid index is constructed based on the time dimension index structure of the B+ tree, and the timestamp of the load data is organized as a keyword. Each non-leaf node stores the boundary value of the time interval, and the leaf nodes are linked in chronological order. At the same time, a cache mechanism is constructed in the B+ tree to cache the index information of the hot time interval in the memory, and a pre-reading strategy is adopted to load the data of the adjacent time periods in advance.

[0046] The feature dimension index adopts a hybrid indexing strategy combining local sensitive hashing and KD tree to convert the feature vector Construct a hash function:

[0047]

[0048] Among them, a is a random vector, b is a random offset, ω is a bucket width parameter, and the feature space is divided using a KD tree. At each node, the dimension with the largest variance is selected as the partition axis to perform adaptive segmentation of the feature space.

[0049] The fast retrieval of the fluctuation pattern is supported by an index structure based on an inverted linked list. Each fluctuation pattern is assigned a unique ID. A mapping relationship between the pattern ID and the feature vector is established through the inverted linked list. Considering the high-dimensional characteristics of the feature vector, a compression storage strategy including feature quantization and sparse representation is adopted to reduce storage overhead. At the same time, the inverted linked list adopts a block storage method to store patterns with similar features in adjacent physical spaces.

[0050] The optimization of the retrieval mechanism after indexing includes multi-way parallel retrieval strategy and multi-level pruning strategy;

[0051] The multi-path parallel search strategy filters the time dimension, locates the data in the target time window, performs similarity matching in the feature space, finds candidate similar fluctuation patterns for pattern verification through the cooperation of LSH and KD tree, and determines the final matching result by calculating the accurate similarity measurement;

[0052] The multi-level pruning strategy includes pruning based on a lower distance bound, which quickly excludes areas that are unlikely to contain target patterns by calculating the minimum distance between a query vector and a feature space partition; density-based pruning, which utilizes the density distribution characteristics of a fluctuation pattern to prioritize high-density areas and skip sparse areas; and similarity threshold-based pruning, which sets a dynamic similarity threshold and terminates the verification of candidate patterns that do not meet the requirements.

[0053] As a preferred solution of the method for load forecasting of distribution transformers based on adaptive decomposition of multi-scale fluctuation patterns described in the present invention, the improved dynamic time warping includes constructing an efficient index structure and performing adaptive matching prediction, and optimizing the distance metric and search strategy through an improved dynamic time warping algorithm;

[0054] The optimized distance metric includes constructing a feature distance vector o=[o1, o2, ..., o m ], where each component o m Represents the distance measurement on different feature dimensions, and the adaptive weight allocation mechanism based on the entropy weight method. For features i and j, the information entropy Ie i Calculated as

[0055]

[0056] Among them, f IE，ij For the normalized eigenvalues, calculate the feature weights:

[0057]

[0058] Among them, w IE is the weight of the ith indicator, IE i is the entropy value of the ith indicator, while ensuring

[0059] Σw IE =1

[0060] The final fusion distance is expressed as d = ∑w IE o m ;

[0061] At the same time, the DTW path constraints are optimized and designed, and the bandwidth parameter bp is dynamically adjusted according to the periodic characteristics of the sequence:

[0062] bp=ap x SL exp (-ap2 per)

[0063] Among them, SL is the sequence length, per is the periodicity strength of the sequence, and ap x and ap2 are adjustment parameters. At the same time, the slope constraint |ij|≤τ is used to control the inclination of the path, where T is optimally selected in the range of [0.1, 0.5] through cross-validation;

[0064] The search strategy uses an improved A* heuristic algorithm to search for the optimal path and defines the evaluation function Wo(node) of the node node:

[0065] Wo(node)=g(node)+d E (node)

[0066] d E (node)=min_dist(node,goal)

[0067] Among them, g(node) represents the actual cumulative distance from the starting point to the node node, d E (node) is the estimated distance from node node to the end point, and the heuristic function based on the lower bound estimate is specified:

[0068]

[0069] Where Hf(node) is the heuristic estimated distance from node n to the target, x k is the kth element of the target sequence, y i is the i-th element of the reference sequence, w is the search window width, and Nc is the sequence length;

[0070] In addition, a multi-level pruning strategy is used to quickly filter the envelopes U and L of the sequence x using the LB_Keogh lower bound. When the distance between the candidate sequence x and the envelope exceeds the current optimal value, it is directly pruned. The Earlyabandoning strategy is adopted. During the cumulative distance calculation process, the calculation is stopped immediately when the threshold is exceeded. The lower bound estimation is gradually tightened through the Cascadinglowerbounds mechanism, and multi-level filtering from coarse to fine is performed.

[0071] As a preferred solution of the distribution transformer load forecasting method based on multi-scale fluctuation mode adaptive decomposition described in the present invention, the anomaly detection optimization includes using an improved integrated isolation forest algorithm for anomaly detection, optimizing the feature space partitioning strategy combined with the feature selection mechanism based on information gain, and calculating the information gain value IG(f) of each feature for the feature selection of each node:

[0072] IG(f)=CE(D)-CE(D|f)

[0073] Among them, CE(D) is the information entropy of data set D, CE(D|f) is the conditional entropy under the condition of feature f, and the feature with the largest information gain is selected as the split feature to capture the abnormal pattern in the load data of distribution equipment;

[0074] The splitting threshold is selected by the dynamic splitting threshold determination method. According to the distribution characteristics of the current node data, the probability density function of the data distribution is obtained by the kernel density estimation method, and the optimal splitting point is selected in the local minimum point of the density function.

[0075] The calculation accuracy is improved by using an improved local anomaly factor algorithm, and the distance metric is optimized by using the generalized Minkowski distance;

[0076] In the process of calculating the local anomaly factor, the reach-distance is optimized to consider the local density information of the sample points. In the density estimation link, the adaptive bandwidth selection mechanism is combined, and the bandwidth parameter is determined by minimizing the mean square error criterion. The improved isolation forest and local anomaly factor algorithms are integrated, and the detection results of the two algorithms are fused by weighted voting. The weight value is dynamically updated through online learning.

[0077] The optimized online correction mechanism includes adaptive window design and robust estimation optimization;

[0078] The adaptive window design establishes a basic window length of 24 hours × 4 sampling times per hour × 7 days, that is, L = 672 sampling points. The window length is adjusted by prediction error feedback. When the continuous prediction error exceeds the set threshold, the window length is adjusted according to L' = L·(1±δ), where δ is a dynamic adjustment coefficient, which is positively correlated with the prediction error.

[0079] Use exponential decay weighting strategy to optimize window data weighting:

[0080] w i =(1-α) i

[0081] Among them, α is the forgetting factor, i is the time sequence number of the data point; the forgetting factor α adopts an adaptive adjustment mechanism, with the initial value set to 0.1, and is dynamically adjusted according to the change of the prediction error; when the prediction error increases, increasing the α value accelerates the forgetting of historical data; when the prediction error decreases, reducing the α value maintains the stability of the model, and the adjustment range of α is limited to [0.05, 0.3];

[0082] The robust estimation optimization includes improving the M estimation method and enhancing the S estimation method, and processing outliers by improving the M estimation method and combining the Huber loss function;

[0083] The enhanced S estimation method optimizes the upper limit of the abnormal point ratio tolerated by the estimator. Considering the characteristics of the distribution network load data, the breaking point is set to:

[0084] ε * =(np) / (2n)

[0085] Among them, n is the number of samples, p is the number of parameters, and in the weight iteration process, the improved weight function w is used i =W(r i / s), where r iis the residual, s is the scale estimate, and during the iteration process, when the Euclidean distance between two adjacent parameter estimates is less than 10 -6 Convergence is determined when .

[0086] Another object of the present invention is to provide a distribution transformer load forecasting system based on adaptive decomposition of multi-scale fluctuation patterns. The present invention more accurately captures the fluctuation characteristics of load data through multi-scale fluctuation decomposition and feature extraction; realizes the adaptability and robustness of load forecasting: through adaptive matching prediction and exception processing, the system can adapt to fluctuation patterns of different time scales and process abnormal data; through real-time monitoring and prediction of distribution transformer loads, provides safety warnings for power grid operation and ensures stable operation of the power grid.

[0087] As a preferred solution of the distribution transformer load prediction system based on multi-scale fluctuation pattern adaptive decomposition described in the present invention, it is characterized by comprising a multi-scale fluctuation decomposition module, a characteristic fluctuation library construction module, an adaptive matching prediction module, and an abnormality processing module;

[0088] The multi-scale fluctuation decomposition module inputs the original load data, uses the improved maximum overlap discrete wavelet packet transform to perform multi-scale decomposition on the load signal, optimizes the energy distribution, and transmits the decomposed load signal characteristics to the characteristic fluctuation library construction module;

[0089] The characteristic fluctuation library construction module inputs the load signal characteristics after multi-scale decomposition, performs feature extraction and characterization optimization, and constructs a characteristic fluctuation library to provide data support for the adaptive matching prediction module;

[0090] The adaptive matching prediction module uses adaptive density estimation and improved multi-level clustering strategy to perform fine characterization of fluctuation patterns, build an efficient index structure, and perform predictions, while passing the prediction results to the exception handling module;

[0091] The anomaly processing module inputs the prediction results and combines them with the integrated isolation forest and the local anomaly factor to perform anomaly detection optimization, optimizes the online correction mechanism through adaptive window and robust estimation, and the corrected prediction results are used for online monitoring of the power grid safety margin.

[0092] A computer device comprises a memory and a processor, wherein the memory stores a computer program, and is characterized in that when the processor executes the computer program, the steps of a distribution transformer load prediction method based on adaptive decomposition of multi-scale fluctuation patterns are implemented.

[0093] A computer-readable storage medium having a computer program stored thereon, characterized in that when the computer program is executed by a processor, the steps of a distribution transformer load forecasting method based on adaptive decomposition of multi-scale fluctuation patterns are implemented.

[0094] Beneficial effects of the present invention: By inputting the original data and performing multi-scale decomposition, the present invention optimizes the distribution of load signal energy and extracts key features, thereby improving the accuracy and efficiency of load forecasting. Subsequently, by using adaptive density estimation and improved multi-level clustering strategies, we finely characterize the fluctuation pattern, construct an efficient index structure, and make predictions through adaptive matching, which not only identifies the time scale characteristics in the load data, but also improves the accuracy and adaptability of the prediction. The improved dynamic time warping and particle swarm optimization algorithms further optimize the distance metric and search strategy, construct a combined prediction framework, and enhance the flexibility and robustness of the prediction. The deep mining and enhanced processing of feature extraction and characterization optimization, as well as the construction of multi-layer hybrid indexing and retrieval mechanisms, comprehensively capture the intrinsic characteristics of the load data, and improve the accuracy of the prediction and system performance. Finally, the introduction of anomaly detection optimization and online correction mechanisms, by integrating isolation forest and local anomaly factor algorithms, as well as adaptive window design and robust estimation optimization, reduces the prediction error, improves the adaptability and stability of the prediction model, and achieves the beneficial effect of overall improvement of prediction efficiency and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0095] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work, among which:

[0096] Figure 1 An overall flow chart of a distribution transformer load forecasting method based on adaptive decomposition of multi-scale fluctuation patterns provided by an embodiment of the present invention.

[0097] Figure 2 A system solution module diagram of a distribution transformer load forecasting system based on adaptive decomposition of multi-scale fluctuation patterns provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0098] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0099] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0100] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor is it an embodiment that is mutually exclusive with other embodiments, either individually or selectively.

[0101] The present invention is described in detail with reference to schematic diagrams. When describing the embodiments of the present invention, for the sake of convenience, the cross-sectional diagrams showing the device structure will not be partially enlarged according to the general scale, and the schematic diagrams are only examples, which should not limit the scope of protection of the present invention. In addition, in actual production, the three-dimensional dimensions of length, width and depth should be included.

[0102] At the same time, in the description of the present invention, it should be noted that the directions or positional relationships indicated by the terms "upper, lower, inner and outer" are based on the directions or positional relationships shown in the drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific direction, be constructed and operated in a specific direction, and therefore cannot be understood as limiting the present invention. In addition, the terms "first, second or third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance.

[0103] In the present invention, unless otherwise clearly specified and limited, the terms "install, connect, connect" should be understood in a broad sense, for example: it can be a fixed connection, a detachable connection or an integral connection; it can also be a mechanical connection, an electrical connection or a direct connection, or it can be indirectly connected through an intermediate medium, or it can be the internal communication of two components. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0104] Example 1, reference Figure 1 , which is the first embodiment of the present invention, provides a distribution transformer load forecasting method based on multi-scale fluctuation mode adaptive decomposition, comprising:

[0105] S1: Input the original data, use the improved maximum overlap discrete wavelet packet transform to perform multi-scale decomposition of the load signal, optimize the energy distribution, perform feature extraction and characterization optimization, and build a feature fluctuation library to store features.

[0106] Furthermore, the improved maximum overlap discrete wavelet packet transform performs multi-scale decomposition of the load signal, including using the sym8 wavelet basis function with good symmetry as the basic decomposition tool, and the compact support characteristic with the support degree set to 16, which can effectively capture the multi-scale characteristics in the distribution network load data.

[0107] The optimal wavelet basis is selected by minimizing the reconstruction error. The error function E is:

[0108]

[0109] Among them, x is the original signal, x′ is the reconstructed signal;

[0110] Set the reconstruction error threshold to 10 -6 , when the reconstruction error is less than 10 -6 When it is considered that the wavelet basis is appropriately selected;

[0111] The wavelet basis function must satisfy three key properties at the same time, the first of which is orthogonality:

[0112] ∫ψ j，k (t)ψ j，l (t)dt=δ k，l

[0113] The second key property is tight support:

[0114] supp(ψ)=[0,2N-1]

[0115] The third key property is the vanishing moment:

[0116] ∫t k ψ(t)dt=0;k=0,1,...,N-1

[0117] Among them, ψ j，k (t) is the wavelet basis function of scale j and k translation, δ k，l is the Kronecker delta function, t is the time variable, supp(ψ) is the support set of the wavelet basis function ψ, N is the length parameter of the wavelet basis function, t k is the kth power of time, ψ(t) is the wavelet basis function, and N′ is the order of the vanishing moment;

[0118] To ensure the stability and reliability of the wavelet transform, the selected wavelet basis function must simultaneously satisfy three key properties. The first is orthogonality, that is, the wavelet functions at different scales and translations remain orthogonal, which can be verified by integral operations to ensure that there is no redundant information between the decomposed coefficients. The second is compact support, which requires the wavelet function to be non-zero on a finite interval [0, 2N-1], which helps to improve computational efficiency and reduce the impact of boundary effects. The third is the vanishing moment characteristic of sufficient order, that is, the integral of the wavelet function and the polynomial is zero, which ensures a good approximation of the polynomial trend in the load data.

[0119] In view of the common endpoint effect problem in load data processing, the present invention processes the endpoint effect problem in load data through an adaptive continuation algorithm and establishes an AR(p) autoregressive model to describe the time correlation of load data:

[0120] x t =∑φ i x t-i +ε t

[0121] Among them, x t is the time series value at time t, φ i is the autoregressive coefficient, ε t is the white noise term;

[0122] The model order p is determined dynamically by minimizing the AIC criterion:

[0123]

[0124] The AIC criterion comprehensively considers the model's goodness of fit (represented by the residual sum of squares RSS) and complexity (represented by the number of model parameters p), and strikes a balance between the two.

[0125] The extension length L is adaptively determined according to the data length N and the signal-to-noise ratio SNR:

[0126] L=min(N / 4,256·[log2(SNR)+1])

[0127] It not only ensures the adequacy of the extension, but also avoids the computational burden caused by excessive extension.

[0128] Denoising is performed using an improved soft threshold shrinkage method, with the threshold parameter λ j Using a scale-dependent adaptive form:

[0129]

[0130] Among them, λ jis the threshold parameter of the jth layer, α is the scale adjustment factor to balance the noise level of different decomposition scales, α∈[0.1,0.5]; σ j is the noise standard deviation, N″ is the data length, and is robustly estimated using the median absolute deviation method:

[0131]

[0132] The SURE criterion is used to optimize the threshold selection and determine the optimal threshold by minimizing the unbiased risk estimate:

[0133]

[0134] Where n is the number of samples, d i is the wavelet coefficient, λ is the threshold parameter, RSS is the residual sum of squares, and SNR is the signal-to-noise ratio.

[0135] It is possible to achieve optimal noise suppression while maintaining signal characteristics.

[0136] The above technical solution is particularly suitable for distribution network equipment-level load forecasting scenarios.

[0137] Since the distribution transformer load has strong randomness and multi-scale characteristics, the optimized wavelet transform can effectively separate the load fluctuation components of different scales. At the same time, considering the abnormal fluctuations and measurement noise that may exist in the load data, the endpoint effect suppression and threshold optimization scheme of the present invention can ensure the reliability of the preprocessing results and lay the foundation for subsequent feature extraction and predictive modeling.

[0138] Specifically, optimizing the energy distribution includes: for the j-th layer wavelet decomposition coefficient wd jk , calculate the wavelet energy proportion WE of the jth layer j :

[0139] Where SL is the signal length, and the relative energy ratio is calculated:

[0140]

[0141] Among them, ∑WE i The optimal decomposition level is determined dynamically as the sum of the energy values ​​of all layers;

[0142] Set the adaptive energy threshold η:

[0143]

[0144] This setting can effectively adapt to the energy distribution characteristics of different types of load data. Where J is the total number of decomposition layers; the mutual information I(X; Y) between adjacent decomposition layers is calculated through the constraint mechanism based on mutual information:

[0145]

[0146] Among them, JP(X, Y) represents the joint probability distribution, JP(X) and JP(Y) represent the marginal probability distribution respectively; when the mutual information value exceeds the threshold value θ = 0.1, it indicates that there is information redundancy and the level adjustment is required. This constraint mechanism is particularly suitable for handling complex load situations in which different user types (such as residential, commercial, and industrial) are mixed in the distribution network.

[0147] Dynamically determine the optimal level through an improved information gain calculation method:

[0148] IG(j)=H(j)-H(j+1)-bf·R(j)

[0149] Among them, H(j) and H(j+1) are the Shannon entropy values ​​of the j-th layer and the j+1-th layer, R(j) is the regularization term to balance the model complexity, bf is the balance factor, bf∈[0.1, 0.3]; this design can effectively avoid excessive decomposition and improve the model's adaptability to load mutations.

[0150] The distance function is defined by a hierarchical merging strategy based on dynamic programming:

[0151] d(i, j) = || S i -S j ||2+βKL(P i ||P j )

[0152] When d(i, j)<δ, merging is performed, and δ is determined by the cross-validation method; this merging strategy is particularly suitable for dealing with the multi-cycle superposition phenomenon in the distribution network equipment-level load.

[0153] Among them, d(i, j) is the distance between adjacent layers, β is the weight coefficient to balance the importance of the two distance metrics, and S i is the i-th layer signal, S j is the j-th layer signal, ||S i -S j ||2 is the Euclidean distance of the signal, KL is the probability distribution, P i is the probability distribution of the i-th layer, P j is the probability distribution of the jth layer.

[0154] The energy distribution optimization decomposition hierarchical method of the present invention fully considers the characteristics of the distribution network load data. Through adaptive energy analysis and dynamic hierarchical optimization, it can effectively extract the load change characteristics on different time scales, providing a reliable basis for subsequent load forecasting. Practice shows that this method has good adaptability and stability when dealing with different types of distribution transformer loads such as residential areas, commercial areas and industrial parks.

[0155] It should be noted that the load data of distribution transformers are deeply mined and enhanced to build a characteristic fluctuation library;

[0156] The load data includes an amplitude feature family and a shape feature family, wherein the amplitude feature family includes a maximum amplitude, a root mean square amplitude, a peak factor, a margin factor, and a waveform factor; and the shape feature family includes a skewness, a kurtosis, a waveform index, and a pulse index;

[0157] The maximum amplitude is A max =max|x(t)|, reflecting the extreme value characteristics of load fluctuation;

[0158] The RMS amplitude is Can effectively represent the effective value level of load;

[0159] The peak factor is K f =A max / A rms , which can reflect the peak characteristics of the load;

[0160] The margin factor is K m =A max / A m ;

[0161] The shape factor is K s =A rms / A m , can characterize the morphological characteristics of load fluctuations from different angles;

[0162] The skewness is It can characterize the asymmetry of load distribution;

[0163] The kurtosis is Used to describe the sharpness of load distribution;

[0164] The waveform index is It can effectively reflect the severity of load changes;

[0165] The pulse index is IF = A max / |x(t)| mean , the pulse characteristics of the load are measured by the ratio of the maximum amplitude to the average absolute value;

[0166] Where x′(t) is the square of the first-order derivative of the reconstructed load signal within time t, X is a random variable, μ SK , μ KU is the mean of skewness and kurtosis, v SK , σ KU are the standard deviations of skewness and kurtosis, respectively;

[0167] The frequency domain features are enhanced by improving the power spectrum density estimation method, and the multi-window spectrum estimation technology is adopted to weightedly superimpose multiple independent spectrum estimation results X k (f) Get the final power spectrum density S w (f):

[0168] S w (f) = ∑α k |X k (f)| 2

[0169] Using the improved Hamming window w(n):

[0170] w(n)=0.54-0.46cos(2πn / N)

[0171] It not only ensures better frequency resolution, but also effectively suppresses spectrum leakage.

[0172] Dynamically adjust the overlap rate mechanism:

[0173] or = 0.5 + 0.1log (SNR)

[0174] Among them, or is the overlap rate, α k is the weight coefficient of the kth window function, X k (f) is the complex number representation of the kth frequency component, |X k (f)| 2 is the power spectrum of the frequency component, n′ is the window function sampling point index, and Nu is the total length of the window function, which ensures the stability of the spectrum estimation;

[0175] The energy distribution is used to characterize the extraction of frequency band energy features, including sub-band energy ratio, energy entropy, spectrum centroid, and spectrum broadening. The sub-band energy ratio includes calculating the proportion of each sub-band energy to the total energy Se i / Se total , reflects the distribution characteristics of energy in the frequency domain;

[0176] The energy entropy is Energy entropy = -∑P i logP i , which can measure the uniformity of energy distribution.

[0177] The spectrum centroid f is obtained by energy-weighted averaging c =∑f·P(f) / ∑P(f), and the center position of the energy distribution is obtained;

[0178] The spectrum broadening B is obtained by calculating the second-order moment of the frequency relative to the center of mass. Among them, P iis the energy probability of the i-th frequency band, f is the frequency value, and P(f) is the power spectrum density corresponding to the frequency; it reflects the discrete degree of energy distribution. These features together constitute a comprehensive characterization of the frequency domain characteristics of the load.

[0179] The above-mentioned extraction process of time domain and frequency domain features fully considers the characteristics of distribution network load data and can effectively capture the load characteristics of different types of users (such as residential, commercial, and industrial users). These features not only retain the important indicators in traditional load analysis, but also introduce new feature representation methods, providing rich feature support for subsequent load forecasting. Through this multi-dimensional feature extraction and representation optimization, the model's ability to perceive the law of load changes has been significantly improved.

[0180] S2: Through adaptive density estimation and improved multi-level clustering strategy, the fluctuation patterns are finely characterized from different time scales, an efficient index structure is constructed, and prediction is performed through adaptive matching.

[0181] Furthermore, the data distribution characteristics are characterized by improved kernel density estimation, and the Epanechnikov kernel function is selected:

[0182] K(x)=(1-||x|| 2 )·I(||x||≤1)

[0183] It can effectively reduce the computational complexity and ensure the smoothness of the estimation results.

[0184] Adaptive method is used to select bandwidth:

[0185] h=σ h (4 / 3n) 1 / 5

[0186] This selection method not only takes into account the dispersion of the data, but also can be dynamically adjusted with the sample size, ensuring the accuracy of the density estimation.

[0187] The density threshold is determined by the dynamic determination method of the density threshold based on the silhouette coefficient:

[0188] ρ th =μ ρ +γσ p

[0189] Among them, σ is the standard deviation of the data, μ ρ , σ ρ are the density mean and standard deviation respectively, γ is optimized by the silhouette coefficient, γ∈[0.8,1.2] ensures the compactness and separation of the clustering results; I(||x||≤1) is the indicative function, and n is the sample size;

[0190] After characterization, the fluctuation pattern recognition under different time scales is completed based on the improved multi-level clustering strategy. According to the inherent law of load change in the distribution network, the time scale is divided into four levels including intra-day layer, daily layer, weekly layer and monthly layer:

[0191] The intra-day layer covers the fluctuation characteristics of 15 minutes to 4 hours, reflecting the short-term electricity consumption behavior of users; the daily layer covers the change pattern of 4 hours to 24 hours, reflecting the characteristics of the daily load curve; the weekly layer spans 1 day to 7 days, describing the difference between weekdays and weekends; the monthly layer ranges from 7 days to 30 days, capturing the seasonal change characteristics;

[0192] In each time scale layer, an improved density clustering algorithm is used to identify the fluctuation pattern, in which the core point is determined in an adaptive manner:

[0193] MinPts=4+[2·log10(n1)]

[0194] Among them, nl is the amount of data in the current layer. In the process of boundary point absorption, the density sensitivity factor dsf is introduced to determine the ownership of the boundary point. The judgment condition is:

[0195] d(p,q)≤eps·(ρ p / ρ q ) dsf

[0196] Where d(p,q) is the distance between point p and point q, ρ p and ρ q are the corresponding density values, eps is the neighborhood radius;

[0197] Among them, dsf∈[0.2, 0.5], when the dsf value is less than 0.35, a relatively tight cluster is formed; when the dsf value is greater than 0.35, points with large density differences are allowed to be classified into the same category;

[0198] Through the above improved hierarchical density clustering method, the present invention can accurately capture the fluctuation characteristics of distribution transformer load data on different time scales, and provide a reliable pattern matching basis for subsequent load forecasting. Practice shows that this method shows good adaptability and stability when processing different types of distribution transformer load data such as residential areas, commercial areas and industrial parks.

[0199] The fine characterization of the fluctuation pattern includes using an improved Gaussian mixture model for probability distribution modeling and estimating parameters through an optimized expectation maximization algorithm:

[0200] Q(θ,θ t ) = E[logL(θ;X,Z)|X,θ t ]

[0201] Among them, Q(θ,θ t ) is the conditional expectation, θ is the model parameter, θ t is the current iteration parameter, L(θ; X, Z) is the likelihood function containing latent variables, and E[·] is the expectation operator;

[0202] Through the adaptive learning rate adjustment mechanism, the parameter update step size is dynamically adjusted according to the gradient change of the likelihood function in each iteration. At the same time, the Bayesian information criterion is used to adaptively determine the optimal number of components K:

[0203] K = arg mmin (BIC)

[0204] BIC=-2Log+KlogM

[0205] Among them, BIC is the Bayesian information criterion, Log is the log-likelihood value, and M is the number of samples; this method can effectively balance the relationship between model complexity and fitting accuracy.

[0206] At the same time, the nonlinear correlation characteristics in the load data of distribution equipment are characterized by the copula function framework:

[0207] C(u1,...,u d )=P(U1≤u1,...,U d ≤u d )

[0208] Among them, u d represents the cumulative distribution function value of the dth random variable, C represents the Copula joint distribution function, u1,...,u d represents the cumulative distribution function value of the marginal distribution, U1,...,U d Represents a random variable, d represents the dimension. Through the copula function, the joint distribution of multidimensional random variables can be decomposed into two parts: marginal distribution and dependency structure. According to the characteristics of the load data, t-copula is selected as the basic dependency structure, and the relevant parameters are estimated by the maximum likelihood method. At the same time, the marginal distribution is selected using an adaptive strategy, and the normal distribution, lognormal distribution or generalized Pareto distribution can be flexibly selected according to the data characteristics.

[0209] By describing the nonlinear correlation characteristics in the load data of power distribution equipment, multi-dimensional features are extracted, including constructing statistical feature system and morphological feature system;

[0210] The statistical moment feature system includes mean, variance, skewness and kurtosis. By calculating the mean σ 4Characterize the asymmetry and peak characteristics of load distribution; these statistical moment characteristics can reflect the probability distribution characteristics of distribution equipment load from different angles, providing an important basis for subsequent load forecasting.

[0211] The morphological feature system includes an opening operation, a closing operation and a morphological gradient;

[0212] The opening operation is Suppress spikes in the signal;

[0213] The closing operation is φ B (X) = X·B, filling the valleys in the signal;

[0214] The morphological gradient is Detect sudden changes in load curves;

[0215] Among them, B is the structural element, X is the signal, δ B (X) and ε B (X) represents dilation and erosion operations respectively;

[0216] In practical applications, the method for fine characterization of fluctuation patterns of the present invention can effectively handle the diversity and complexity of load data of distribution network equipment. For example, for distribution transformers in commercial areas, this method can accurately capture the load fluctuation patterns during business hours; for distribution equipment in industrial parks, it can effectively identify the load mutation characteristics when switching between different production shifts. This refined feature characterization provides a reliable data basis for subsequent load forecasting, significantly improving the accuracy and robustness of the forecasting model.

[0217] It should also be noted that, in response to the demand for fluctuation pattern retrieval in distribution network equipment-level load forecasting, building an efficient index structure includes building a multi-layer hybrid index and retrieval mechanism and supporting fast fluctuation pattern retrieval, including multi-layer hybrid index, feature dimension index, and supporting fast retrieval of fluctuation patterns;

[0218] The multi-layer hybrid index is constructed based on the time dimension index structure of the B+ tree, and the timestamp of the load data is organized as a keyword. Each non-leaf node stores the boundary value of the time interval, and the leaf nodes are linked in chronological order. At the same time, a cache mechanism is constructed in the B+ tree to cache the index information of the hot time interval in the memory, and a pre-reading strategy is adopted to load the data of the adjacent time periods in advance.

[0219] The feature dimension index adopts a hybrid indexing strategy combining local sensitive hashing and KD tree to convert the feature vector Construct a hash function:

[0220]

[0221] Among them, a is a random vector, b is a random offset, ω is a bucket width parameter, and the feature space is divided using a KD tree. At each node, the dimension with the largest variance is selected as the partition axis to perform adaptive segmentation of the feature space.

[0222] In order to support the fast retrieval of fluctuation patterns, the fast retrieval uses an index structure based on an inverted linked list. Each fluctuation pattern is assigned a unique ID. The mapping relationship between the pattern ID and the feature vector is established through the inverted linked list. Considering the high-dimensional characteristics of the feature vector, a compression storage strategy including feature quantization and sparse representation is adopted to reduce storage overhead. At the same time, the inverted linked list adopts a block storage method to store patterns with similar features in adjacent physical spaces, which improves access locality.

[0223] After indexing, the retrieval mechanism is optimized, including multi-way parallel retrieval and multi-level pruning strategies;

[0224] The multi-path parallel retrieval strategy filters the time dimension, locates the data within the target time window, performs similarity matching in the feature space, finds candidate similar fluctuation patterns for pattern verification through the cooperation of LSH and KD trees, and determines the final matching result by calculating the precise similarity measurement; the entire retrieval process adopts multi-threaded parallel processing to make full use of the parallel computing capabilities of modern processors.

[0225] The multi-level pruning strategy includes pruning based on the lower bound of distance, which quickly excludes areas that are unlikely to contain the target pattern by calculating the minimum distance between the query vector and the feature space partition; pruning based on density, which uses the density distribution characteristics of the fluctuation pattern to prioritize high-density areas and skip sparse areas; and pruning based on similarity threshold, which sets a dynamic similarity threshold and terminates the verification of candidate patterns that do not meet the requirements. The combined use of these pruning strategies significantly reduces the computational complexity of retrieval.

[0226] The design of the above index structure and retrieval mechanism fully considers the characteristics of distribution network load data and can effectively support the retrieval requirements of large-scale load data fluctuation patterns. Experiments show that the scheme has achieved good results in terms of retrieval efficiency and accuracy, which is of great significance for improving the performance of load forecasting.

[0227] S3: Through improved dynamic time warping, distance metrics and search strategies are optimized, and through improved particle swarm optimization algorithm, weights are dynamically adjusted to build a combined prediction framework.

[0228] Furthermore, we build an efficient index structure and perform adaptive matching prediction, and optimize the distance metric and search strategy through an improved dynamic time warping algorithm;

[0229] The optimized distance metric includes constructing a feature distance vector o=[o1, o2, ..., o m ], where each component o m Represents the distance measurement on different feature dimensions, and the adaptive weight allocation mechanism based on the entropy weight method. For features i and j, the information entropy Ie i Calculated as

[0230]

[0231] Among them, f IE，ij For the normalized eigenvalues, calculate the feature weights:

[0232]

[0233] Among them, w IE is the weight of the ith indicator, IE i is the entropy value of the ith indicator, while ensuring:

[0234] ∑w IE =1

[0235] The final fusion distance is expressed as d = ∑W IE o m ;

[0236] At the same time, the DTW path constraints are optimized and designed, and the bandwidth parameter bp is dynamically adjusted according to the periodic characteristics of the sequence:

[0237] bp=ap x SL exp (-ap2 per)

[0238] Among them, SL is the sequence length, per is the periodicity strength of the sequence, and ap x and ap2 are adjustment parameters. At the same time, the slope constraint |ij|≤τ is used to control the inclination of the path, where T is optimally selected in the range of [0.1, 0.5] through cross-validation;

[0239] The search strategy uses an improved A* heuristic algorithm to search for the optimal path and defines the evaluation function Wo(node) of the node node:

[0240] Wo(node)=g(node)+d E (node)

[0241] d E (node)=min_dist(node,goal)

[0242] Among them, g(node) represents the actual cumulative distance from the starting point to the node node, dE (node) is the estimated distance from node node to the end point, and specifies the heuristic function based on the lower bound estimate

[0243]

[0244] Where Hf(node) is the heuristic estimated distance from node n to the target, x k is the kth element of the target sequence, y i is the i-th element of the reference sequence, w is the search window width, and Nc is the sequence length.

[0245] In addition, a multi-level pruning strategy is used to quickly filter the envelopes U and L of the sequence x using the LB_Keogh lower bound. When the distance between the candidate sequence x and the envelope exceeds the current optimal value, it is directly pruned. The Earlyabandoning strategy is used. During the cumulative distance calculation process, the calculation is stopped immediately when the threshold is exceeded. The lower bound estimation is gradually tightened through the Cascadinglower bounds mechanism, and multi-level filtering from coarse to fine is performed.

[0246] Experiments show that in a case study of load forecasting for a distribution transformer in a residential area, the improved DTW algorithm improves the matching accuracy by 15.3% compared to the traditional method, while the computational efficiency is increased by about 3.5 times. The algorithm has good adaptability to the time elastic deformation in the load sequence, can effectively identify similar load patterns, and provide a reliable pattern matching basis for subsequent forecasts.

[0247] It should be noted that the dynamic adjustment of weights includes firstly using an improved particle swarm optimization algorithm to optimize weights. The algorithm introduces a nonlinear decreasing inertia weight strategy to enable particles to have a stronger global exploration capability in the early stage of search. The particle speed is updated using an improved speed v update formula.

[0248] V=w v ·v+C1r1(p best -x)+c2r2(g best -x)

[0249] w=w max -(w max -w min )(t / T) 2

[0250] c i =c max -(c max -c min )(t / T)

[0251] Among them, w is the nonlinear decreasing inertia weight, w max and w min are the maximum and minimum values ​​of the inertia weight, t is the current number of iterations, T is the maximum number of iterations, c1 and c2 are learning factors, c max and c min are the upper and lower limits of the learning factor, respectively, best is the optimal position of the individual, g best is the global optimal position; this nonlinear decreasing strategy can make the inertia weight decrease slowly in the early stage of optimization and decrease rapidly in the later stage, thereby better balancing the global search and local development.

[0252] At the same time, the particle position is constrained by the Sigmoid mapping function Sig(V), and the weight normalization w Sig Ensure that the sum of the weights is 1:

[0253]

[0254]

[0255] The time decay update mechanism is used to adapt to the dynamic changes of load characteristics. The time decay factor is calculated by λ(t)=exp(-ζt), where ζ=-ln(0.5) / t 1 / 2 , t 1 / 2 This mechanism enables newer samples to have larger weights, while the weights of older samples gradually decrease, thereby timely reflecting changes in load characteristics.

[0256] Incremental learning is performed through a sliding window mechanism. The window length Len is set to 24·4·7, which is the number of sampling points in a week. For samples in the window, the contribution to the model update is measured by the sample importance weight:

[0257]

[0258] Among them, d Len is the distance between the sample and the current prediction target, σ Len is the scale parameter;

[0259] Through the abnormal penalty mechanism:

[0260]

[0261] Among them, s i is the abnormal score of the sample, pf is the penalty coefficient, which reduces the impact of abnormal samples on model update;

[0262] This adaptive weight mechanism fully considers the characteristics of distribution network load data, can effectively handle the differences in load characteristics of different types of users (such as residential, commercial, and industrial users), and can adapt to the dynamic changes of load over time, seasons, etc. Through experimental verification, this mechanism has shown good prediction performance in multiple actual scenarios, and the prediction accuracy is significantly better than traditional methods.

[0263] It should be noted that the construction of the combined prediction framework includes a deep learning architecture of a multi-head attention mechanism, a deep residual network structure, and a multi-level feature fusion mechanism;

[0264] The deep learning architecture of the multi-head attention mechanism includes two key components: the self-attention layer and the temporal attention layer;

[0265] The self-attention layer is transformed by the linear matrix W q , W u and W v Map the input feature x to the feature space of query vector Q, key vector U and value vector V:

[0266] Q=W q X;U=W u X; V = W v X

[0267] Among them, the query vector Q is used to capture the load characteristics at the current moment, and the key vector U and value vector V are used to store the key information and specific values ​​of the historical load pattern respectively;

[0268] By calculating the dot product of Q and U and normalizing it by the scale factor sf, we get the attention weight matrix, which reflects the correlation strength between load features at different times:

[0269]

[0270] The multi-head attention mechanism is used to enhance the model's ability to express multi-dimensional features in the load sequence. That is, Ω different attention heads are run in parallel. Each attention head has an independent parameter matrix to capture feature associations in different subspaces. The outputs of the Ω heads are concatenated through the concat operation and then fused through the linear transformation matrix WO to obtain the final multi-head attention output:

[0271] MultiHead=Concat(Ωead1,...,Ωead Ω )W O

[0272] This design enables the model to simultaneously focus on the dependencies of the load sequence in different feature dimensions, thus improving the comprehensiveness of feature extraction.

[0273] In the temporal attention layer, the temporal dependency in the load sequence is characterized by the relative position encoding mechanism, and the temporal information is modeled by calculating the relative distance between any two moments. Different from the traditional absolute position encoding, the relative position encoding models the temporal information by calculating the relative distance between any two moments. This design is more in line with the periodic characteristics of load changes.

[0274] At the same time, a local-global dual attention structure is added. Local attention captures short-term load fluctuation characteristics, while global attention captures long-term change trends. The combination of the two improves the model's ability to capture multi-scale temporal features.

[0275] In order to enhance the model's learning ability of the periodic pattern of the load sequence, a periodic enhancement module is integrated in the time attention layer to perform periodic decomposition on the input sequence, extracting periodic components of multiple scales such as daily, weekly and monthly cycles. Through the specified periodic attention sublayer, the association between different periodic components is established, thereby improving the model's ability to recognize and predict periodic load patterns.

[0276] The deep residual network structure establishes a direct connection between shallow features and deep features through identity mapping. For each residual unit, the output can be expressed as F(x)+x, where F(x) is the residual mapping and x is the identity mapping. After outputting the residual mapping, a pre-activation design is used to optimize the residual learning effect, and the batch normalization layer and the activation function layer are placed before the convolution layer. This design not only improves the convergence performance of the model, but also enhances the stability of feature propagation.

[0277] Through the identity mapping protection mechanism, residual information is transmitted in the deep network to reduce the loss; this structural design effectively alleviates the gradient vanishing problem in the deep network training process while maintaining the model's sensitivity to the original features.

[0278] In terms of feature fusion, the present invention constructs a multi-level feature fusion mechanism. First, the direct transmission of features of different depth layers is achieved through jump connections, avoiding information loss in deep networks. Secondly, a dense connection strategy is adopted to cumulatively fuse the features of each layer, making full use of feature information at different abstract levels. Finally, through the multi-scale feature fusion module, adaptive fusion of features of different time scales is achieved, improving the model's ability to perceive load changes.

[0279] S4: Combine ensemble isolation forests and local anomaly factors to optimize anomaly detection, and optimize the online correction mechanism through adaptive windows and robust estimation.

[0280] Furthermore, an improved ensemble isolation forest algorithm is used for anomaly detection. By optimizing the feature space partitioning strategy and combining the feature selection mechanism based on information gain, the information gain value IG(f) of each feature is calculated for the feature selection of each node:

[0281] IG(f)=CE(D)-CE(D|f)

[0282] Among them, CE(D) is the information entropy of data set D, CE(D|f) is the conditional entropy under the condition of feature f, and the feature with the largest information gain is selected as the split feature to capture the abnormal pattern in the load data of distribution equipment;

[0283] The split threshold is selected by the dynamic split threshold determination method. According to the distribution characteristics of the current node data, the kernel density estimation method is used to obtain the probability density function of the data distribution, and the optimal split point is selected from the local minimum point of the density function. At the same time, the tree depth maxDepth is determined in an adaptive way:

[0284]

[0285] in, For the number of samples, this setting can effectively balance the computational efficiency and model expression ability;

[0286] Improvements to the calculation of anomaly degree:

[0287]

[0288] Among them, z(x) represents the sample , Dp(z(x)) is the expected path length, c(n) is the normalization factor, c(n)=2Har(n-1)-(2(n-1) / n), Har(n-1) is the harmonic number; this calculation method takes into account the degree of isolation of samples in the feature space and can more accurately identify abnormal points in distribution load data.

[0289] The calculation accuracy is improved by improving the local anomaly factor algorithm. In determining the k nearest neighbors, an adaptive k value selection strategy is adopted:

[0290] k=[log2n]+(μ±σ)

[0291] Among them, n is the number of samples, μ and σ are the mean and standard deviation of the historical optimal k value, respectively. This dynamic adjustment mechanism can better adapt to the data distribution of different load characteristics;

[0292] In terms of distance measurement, the generalized Minkowski distance is used:

[0293] d(xf,yf)=(∑|xf i-yf i | pf ) 1 / pf

[0294] Among them, xf i ,yf i is the i-th component of x and y, and pf is the norm parameter; this distance measurement method can better characterize the similarity relationship of distribution load data in high-dimensional feature space. For different types of load characteristics (such as power consumption, power factor, etc.), a feature weighting strategy is adopted, and the weight is optimized and determined in a data-driven manner.

[0295] In the calculation process of local anomaly factors, the reach-distance is optimized to consider the local density information of the sample points. In the density estimation link, the bandwidth parameter is determined by minimizing the mean square error criterion in combination with the adaptive bandwidth selection mechanism. The present invention integrates the improved isolation forest and local anomaly factor algorithms, and uses weighted voting to fuse the detection results of the two algorithms. The weight value is dynamically updated through online learning, and the importance of different algorithms in different scenarios can be adaptively adjusted, thereby achieving a more robust anomaly detection effect. This integration strategy is particularly suitable for complex abnormal pattern recognition problems in distribution network load data.

[0296] It should be noted that the online correction mechanism optimization is a key module in the present invention, which mainly includes two parts: adaptive window design and robust estimation optimization. This mechanism can effectively handle abnormal fluctuations in the load forecasting of distribution network equipment level and improve the forecasting accuracy and stability;

[0297] The adaptive window design establishes a basic window length of 24 hours × 4 (number of samples per hour) × 7 days, that is, L = 672 sampling points. The window length is adjusted by prediction error feedback. When the continuous prediction error exceeds the set threshold, the window length is adjusted according to L' = L·(1±δ), where δ is a dynamic adjustment coefficient, which is positively correlated with the prediction error.

[0298] For distribution transformers in residential areas, the value range of δ is [0.1, 0.3]; for distribution transformers in commercial areas, the value range of δ is [0.2, 0.4]; for distribution transformers in industrial areas, the value range of δ is [0.3, 0.5].

[0299] Use exponential decay weighting strategy to optimize window data weighting:

[0300] w i =(1-α) i

[0301] Among them, α is the forgetting factor, i is the time sequence number of the data point; the forgetting factor α adopts an adaptive adjustment mechanism, with the initial value set to 0.1, and is dynamically adjusted according to the change of the prediction error. When the prediction error increases, increasing the α value accelerates the forgetting of historical data; when the prediction error decreases, reducing the α value maintains the stability of the model, and the adjustment range of α is limited to [0.05, 0.3];

[0302] The robust estimation optimization includes improving the M estimation method and enhancing the S estimation method. By improving the M estimation method and combining it with the Huber loss function to process outliers, square loss is used when the absolute value of the residual does not exceed the threshold k, and linear loss is used when it exceeds k:

[0303]

[0304] The threshold k is estimated by the median absolute deviation MAD method, k = 1.345 × MAD, where MAD = median(|x i -median(x)|);

[0305] The influence function design adopts the Huber type function, Θ(x) = x·min(1,k / |x|);

[0306] In order to further improve the robustness of the estimation, the enhanced S estimation method optimizes the upper limit of the abnormal point ratio tolerated by the estimator. Considering the characteristics of the distribution network load data, the breaking point is set to:

[0307] ε * =(np) / (2n)

[0308] Among them, n is the number of samples, p is the number of parameters, and in the weight iteration process, the improved weight function w is used i =W(r i / s), where r i is the residual, s is the scale estimate, and during the iteration process, when the Euclidean distance between two adjacent parameter estimates is less than 10 -6 Convergence is determined when .

[0309] The online correction mechanism of the present invention has shown excellent performance in practical applications. Taking a distribution transformer in a commercial area as an example, after adopting this mechanism, the mean absolute percentage error (MAPE) of load forecasting was reduced from 3.8% to 2.8%, and the root mean square error (RMSE) was reduced from 8.5kW to 6.2kW. At the same time, the reliability of the forecast results was significantly improved, and the coverage rate of the 95% confidence interval reached 96.2%. This fully proves that the mechanism has significant advantages in dealing with abnormal fluctuations in distribution network load forecasting.

[0310] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

[0311] Embodiment 2 is the second embodiment of the present invention, which provides a distribution transformer load forecasting method based on adaptive decomposition of multi-scale fluctuation patterns. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through experiments.

[0312] Example 1: Load prediction of a distribution transformer in a residential area, the specific contents are as follows:

[0313] Table 1 Input data

[0314]

[0315]

[0316] Table 2 Parameter configuration

[0317] MODWPT parameters Decomposition level: 5; wavelet basis: sym8; threshold: λ = 0.05 DBSCAN parameters Eps=0.5; MinPts=5; Density factor: α=0.3 DTW parameters Bandwidth: bp=48; Step mode: Symmetric; Distance metric: Euclidean distance PSO parameters Population size: 50; Maximum iterations: 100; W: 0.9 → 0.4; c1 = c2 = 2.0

[0318] Table 3 Data preprocessing

[0319] Missing Value Handling Short-term: cubic spline interpolation; medium-term: similar day substitution; long-term: model estimation Outlier handling 3σ rule initial screening; LOF algorithm fine screening; expert rule verification Standardization Z-score normalization; range: [0,1]

[0320] Feature engineering includes time feature construction, meteorological feature extraction, and load feature calculation; model training includes parameter optimization, model integration, and prediction verification;

[0321] Table 4 Prediction results

[0322] MAPE 3.2％ RMSE 4.5kW Coverage 95.8％

[0323] Example 2: Load prediction of distribution transformers in a commercial area

[0324] Table 5 Input data

[0325]

[0326] Table 6 Parameter configuration

[0327] MODWPT parameters Decomposition level: 6; wavelet basis: sym20; threshold: λ = 0.06 Clustering parameters Eps=0.6; MinPts=6; Density factor: α=0.4 DTW parameters Bandwidth: bp=72; Step mode: Symmetric; Distance metric: Euclidean distance Optimization parameters Population size: 60; Maximum iterations: 120; W: 0.9 → 0.3; c1 = c2 = 2.5

[0328] Table 7 Prediction results

[0329]

[0330]

[0331] Example 3: Load prediction of distribution transformers in an industrial park

[0332] Table 8 Input data

[0333]

[0334] Table 9 Parameter configuration

[0335] MODWPT parameters Decomposition level: 7; wavelet basis: sym2; threshold: λ = 0.04 Clustering parameters Eps=0.4; MinPts=8; Density factor: α=0.5 DTW parameters Bandwidth: bp=96; Step mode: Symmetric; Distance metric: Mahalanobis distance Optimization parameters Population size: 80; Maximum iterations: 150; W: 0.9→0.35; c1=c2=2.2

[0336] Table 10 Prediction results

[0337] MAPE 2.5％ RMSE 8.3kW Coverage 96.5％

[0338] Example 4: Load prediction of a distribution transformer in a large shopping mall

[0339] Table 11 Input data

[0340]

[0341] Table 12 Parameter configuration

[0342] MODWPT parameters Decomposition level: 6; wavelet basis: sym9; threshold: λ = 0.055 Clustering parameters Eps=0.55; MinPts=7; Density factor: α=0.35 DTW parameters Bandwidth: bp=60; Step mode: asymmetric; Distance metric: cosine distance Optimization parameters Population size: 70; Maximum iterations: 130; W: 0.85→0.3; c1=c2=2.3

[0343] Table 13 Prediction results

[0344]

[0345]

[0346] Embodiment 3, the third embodiment of the present invention, is different from the first two embodiments in that:

[0347] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the methods described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, and other media that can store program codes.

[0348] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0349] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic device), a portable computer disk case (magnetic device), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0350] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, a plurality of steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logic function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0351] Example 4, reference Figure 2 , which is the fourth embodiment of the present invention, and provides a distribution transformer load prediction system based on multi-scale fluctuation pattern adaptive decomposition, including a multi-scale fluctuation decomposition module 10, a characteristic fluctuation library construction module 20, an adaptive matching prediction module 30, and an abnormality processing module 40;

[0352] The multi-scale fluctuation decomposition module 10 inputs the original load data, uses the improved maximum overlap discrete wavelet packet transform to perform multi-scale decomposition on the load signal, optimizes the energy distribution, and transmits the decomposed load signal characteristics to the characteristic fluctuation library construction module 10;

[0353] The characteristic fluctuation library construction module 20 inputs the load signal characteristics after multi-scale decomposition, performs feature extraction and characterization optimization, and constructs a characteristic fluctuation library to provide data support for the adaptive matching prediction module 30;

[0354] The adaptive matching prediction module 30 uses adaptive density estimation and an improved multi-level clustering strategy to perform a fine characterization of the fluctuation pattern, construct an efficient index structure, and perform predictions, while passing the prediction results to the exception handling module 40;

[0355] The abnormality processing module 40 inputs the prediction results and combines them with the integrated isolation forest and the local abnormality factor to optimize the abnormality detection. The online correction mechanism is optimized through the adaptive window and robust estimation. The corrected prediction results are used for online monitoring of the safety margin of the power grid.

[0356] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A distribution transformer load forecasting method based on adaptive decomposition of multi-scale fluctuation patterns, characterized by: include, Input the original data, use the improved maximum overlap discrete wavelet packet transform to perform multi-scale decomposition of the load signal, optimize the energy distribution, perform feature extraction and characterization optimization, and build a feature fluctuation library to store features; Through adaptive density estimation and improved multi-level clustering strategy, the fluctuation pattern is finely characterized from different time scales, an efficient index structure is constructed, and prediction is performed through adaptive matching; By improving dynamic time warping, optimizing distance metrics and search strategies, and dynamically adjusting weights through an improved particle swarm optimization algorithm, a combined prediction framework is constructed; Anomaly detection is optimized by combining integrated isolation forests and local anomaly factors, and the online correction mechanism is optimized through adaptive windows and robust estimation.

2. A method for distribution transformer load forecasting based on multi-scale fluctuation mode adaptive decomposition as claimed in claim 1, characterized in that: The improved maximum overlap discrete wavelet packet transform performs multi-scale decomposition of the load signal, including using the sym8 wavelet basis function with good symmetry as the basic decomposition tool, setting the support degree to 16, selecting the optimal wavelet basis by minimizing the reconstruction error, and the error function E is: Among them, x is the original signal, x' is the reconstructed signal; Set the reconstruction error threshold to 10 -6 , when the reconstruction error is less than 10 -6 When it is considered that the wavelet basis is appropriately selected; At the same time, the wavelet basis function must satisfy three key properties at the same time, the first key property is orthogonality, the second key property is compact support, and the third key property is vanishing moment; The endpoint effect problem in load data is processed by adaptive extension algorithm, and AR(p) autoregressive model is established to describe the time correlation of load data. The model order p is dynamically determined by minimizing the AIC criterion, and the extension length L is adaptively determined according to the data length N and the signal-to-noise ratio SNR. Denoising is performed by the improved soft threshold shrinkage method, and the threshold parameter λ j A scale-dependent adaptive form was adopted, robust estimation was performed using the median absolute deviation method, threshold selection was optimized using the SURE criterion, and the optimal threshold was determined by minimizing the unbiased risk estimate; The energy distribution optimization includes, for the j-th layer wavelet decomposition coefficient, calculating the wavelet energy ratio of the j-th layer, setting an adaptive energy threshold, calculating the mutual information between adjacent decomposition layers through a constraint mechanism based on mutual information, when the mutual information value exceeds a threshold value θ=0.1, it indicates that there is information redundancy and a hierarchical adjustment is required, dynamically determining the optimal hierarchical level through an improved information gain calculation method, defining a distance function d(i,j) through a hierarchical merging strategy based on dynamic programming, merging when d(i,j)<δ, δ is determined by a cross-validation method, wherein d(i,j) is the distance between adjacent layers.

3. A method for distribution transformer load forecasting based on multi-scale fluctuation mode adaptive decomposition as claimed in claim 2, characterized in that: The feature extraction and characterization optimization includes deep mining and enhancement processing of distribution transformer load data to build a feature fluctuation library; The load data includes an amplitude feature family and a shape feature family, wherein the amplitude feature family includes a maximum amplitude, a root mean square amplitude, a peak factor, a margin factor, and a waveform factor; and the shape feature family includes a skewness, a kurtosis, a waveform index, and a pulse index; The frequency domain features are enhanced by an improved power spectrum density estimation method, and the final power spectrum density is obtained by weighted superposition of multiple independent spectrum estimation results using multi-window spectrum estimation technology. The energy characteristics of the frequency bands extracted through energy distribution characterization include sub-band energy ratio, energy entropy, spectrum centroid, and spectrum broadening. The sub-band energy ratio includes calculating the proportion of each sub-band energy to the total energy; obtaining the spectrum centroid through energy-weighted averaging to obtain the center position of the energy distribution; and obtaining the spectrum broadening by calculating the second-order moment of the frequency relative to the centroid.

4. A method for distribution transformer load forecasting based on multi-scale fluctuation mode adaptive decomposition as claimed in claim 3, characterized in that: The adaptive density estimation and improved multi-level clustering strategy includes: characterizing data distribution characteristics through improved kernel density estimation, selecting Epanechnikov kernel function, using adaptive method to select bandwidth, and determining density threshold based on dynamic determination method of density threshold of silhouette coefficient; After characterization, the fluctuation pattern recognition under different time scales is completed based on the improved multi-level clustering strategy. According to the inherent law of load change in the distribution network, the time scale is divided into four levels including intra-day layer, daily layer, weekly layer and monthly layer: The intra-day layer covers the fluctuation characteristics of 15 minutes to 4 hours, reflecting the short-term electricity consumption behavior of users; the daily layer covers the change pattern of 4 hours to 24 hours, reflecting the characteristics of the daily load curve; The weekly layer spans from 1 day to 7 days, describing the difference between weekdays and weekends; the monthly layer ranges from 7 days to 30 days, capturing seasonal variation characteristics; In each time scale layer, an improved density clustering algorithm is used to identify the fluctuation pattern. In the process of boundary point absorption, the density sensitivity factor dsf is introduced to determine the attribution of the boundary point. The judgment condition is: d(p,q)≤eps·(ρ p / r q ) dsf Among them, d)p,q) is the distance between point p and point q, ρ p and ρ q are the corresponding density values, eps is the neighborhood radius, where dsf∈[0.2,0.5]. When the dsf value is less than 0.35, a relatively tight clustering is formed; when the dsf value is greater than 0.35, points with large density differences are allowed to be classified into the same category; The fine characterization of the fluctuation pattern includes using an improved Gaussian mixture model for probability distribution modeling and estimating parameters through an optimized expectation maximization algorithm: Q(θ,θ t )=E[log L(θ;X,Z)|X,θ t ] Among them, Q(θ,θ t ) is the conditional expectation, θ is the model parameter, θ t is the current iteration parameter, L(θ; X, Z is the likelihood function containing latent variables, and E[·] is the expectation operator; Through the adaptive learning rate adjustment mechanism, the parameter update step size is dynamically adjusted according to the gradient change of the likelihood function in each iteration. At the same time, the Bayesian information criterion is used to adaptively determine the optimal number of components K: K = arg min (BIC) BIC=-2Log+KlogM Among them, BIC is the Bayesian information criterion, Log is the log-likelihood value, and M is the number of samples; At the same time, the nonlinear correlation characteristics in the load data of distribution equipment are characterized by the copula function framework: C(u1,...,u d )=P(U1≤u1,...,U d ≤u d ) Among them, u d represents the cumulative distribution function value of the dth random variable, C represents the Copula joint distribution function, u1,...,u d represents the cumulative distribution function value of the marginal distribution, U1,...,U d Represents a random variable, and d represents the dimension. Through the copula function, the joint distribution of multidimensional random variables is decomposed into two parts: marginal distribution and dependency structure. According to the characteristics of load data, t-copula is selected as the basic dependency structure, and the relevant parameters are estimated by the maximum likelihood method. At the same time, the selection of marginal distribution adopts an adaptive strategy. By describing the nonlinear correlation characteristics in the load data of power distribution equipment, multi-dimensional features are extracted, including constructing statistical feature system and morphological feature system; The statistical moment characteristic system includes mean, variance, skewness and kurtosis. The central tendency of the load level is reflected by calculating the mean μ=E[X]. The variance is calculated by σ 2 =E[(X-μ) 2 ] describes the discreteness of load fluctuations, and calculates the skewness γ = E[(X-μ) 3 ] / σ 3 and the kurtosis κ = E[X-μ) 4 ] / σ 4 Characterize the asymmetry and peak characteristics of load distribution; The morphological feature system includes an opening operation to suppress peaks in a signal, a closing operation to fill valleys in a signal, and a mutation feature in a morphological gradient detection load curve.

5. A method for distribution transformer load forecasting based on multi-scale fluctuation mode adaptive decomposition as claimed in claim 4, characterized in that: The said construction of efficient index structure includes construction of multi-layer hybrid index and retrieval mechanism and support for fast fluctuation pattern retrieval including multi-layer hybrid index, feature dimension index, support for fast retrieval of fluctuation pattern; The multi-layer hybrid index is constructed based on the time dimension index structure of the B+ tree, and the timestamp of the load data is organized as a keyword. Each non-leaf node stores the boundary value of the time interval, and the leaf nodes are linked in chronological order. At the same time, a cache mechanism is constructed in the B+ tree to cache the index information of the hot time interval in the memory, and a pre-reading strategy is adopted to load the data of the adjacent time periods in advance. The feature dimension index adopts a hybrid indexing strategy combining local sensitive hashing and KD tree to convert the feature vector Construct a hash function: Among them, a is a random vector, b is a random offset, ω is a bucket width parameter, and the feature space is divided using a KD tree. At each node, the dimension with the largest variance is selected as the partition axis to perform adaptive segmentation of the feature space. The fast retrieval of the fluctuation pattern is supported by an index structure based on an inverted linked list. Each fluctuation pattern is assigned a unique ID. A mapping relationship between the pattern ID and the feature vector is established through the inverted linked list. Considering the high-dimensional characteristics of the feature vector, a compression storage strategy including feature quantization and sparse representation is adopted to reduce storage overhead. At the same time, the inverted linked list adopts a block storage method to store patterns with similar features in adjacent physical spaces. The optimization of the retrieval mechanism after indexing includes multi-way parallel retrieval strategy and multi-level pruning strategy; The multi-path parallel search strategy filters the time dimension, locates the data in the target time window, performs similarity matching in the feature space, finds candidate similar fluctuation patterns for pattern verification through the cooperation of LSH and KD tree, and determines the final matching result by calculating the accurate similarity measurement; The multi-level pruning strategy includes pruning based on a lower distance bound, which quickly excludes areas that are unlikely to contain target patterns by calculating the minimum distance between a query vector and a feature space partition; density-based pruning, which utilizes the density distribution characteristics of a fluctuation pattern to prioritize high-density areas and skip sparse areas; and similarity threshold-based pruning, which sets a dynamic similarity threshold and terminates the verification of candidate patterns that do not meet the requirements.

6. A method for distribution transformer load forecasting based on multi-scale fluctuation mode adaptive decomposition as claimed in claim 5, characterized in that: The improved dynamic time warping includes constructing an efficient index structure and performing adaptive matching prediction, and optimizing distance metrics and search strategies through an improved dynamic time warping algorithm; The optimized distance metric includes constructing a feature distance vector o=[o1,o2,...,o m ], where each component o m Represents the distance measurement on different feature dimensions, and the adaptive weight allocation mechanism based on the entropy weight method. For features i and j, the information entropy IE i Calculated as Among them, f IE,ij For the normalized eigenvalues, calculate the feature weights: Among them, w IE is the weight of the ith indicator, IE i is the entropy value of the ith indicator, while ensuring The final fusion distance is expressed as d = Σw IE o m ; At the same time, the DTW path constraints are optimized and designed, and the bandwidth parameter bp is dynamically adjusted according to the periodic characteristics of the sequence: bp=ap1·SL·exp(-ap2·per) Where SL is the sequence length, per is the periodicity strength of the sequence, ap1 and ap2 are adjustment parameters, and the slope constraint |ij|≤τ is used to control the slope of the path, where τ is optimally selected in the range of [0.1, 0.5] through cross-validation; The search strategy uses an improved A* heuristic algorithm to search for the optimal path and defines the evaluation function Wo(node) of the node node: Where(node)=g(node)+d E (node) the E (node)=min_dist(node,goal) Among them, g(node) represents the actual cumulative distance from the starting point to the node node, d E (node) is the estimated distance from node node to the end point, and the heuristic function based on the lower bound estimate is specified: Where Hf(node) is the heuristic estimated distance from node n to the target, x k is the kth element of the target sequence, y i is the i-th element of the reference sequence, w is the search window width, and Nc is the sequence length; In addition, a multi-level pruning strategy is used to quickly filter the envelopes U and L of the sequence x using the LB_Keogh lower bound. When the distance between the candidate sequence x and the envelope exceeds the current optimal value, it is directly pruned. The Early abandoning strategy is adopted. During the cumulative distance calculation process, the calculation is stopped immediately when the threshold is exceeded. The lower bound estimation is gradually tightened through the Cascadinglowerbounds mechanism, and multi-level filtering from coarse to fine is performed.

7. A method for distribution transformer load forecasting based on multi-scale fluctuation mode adaptive decomposition as claimed in claim 6, characterized in that: The anomaly detection optimization includes using an improved integrated isolation forest algorithm for anomaly detection, optimizing the feature space partitioning strategy combined with the feature selection mechanism based on information gain, and calculating the information gain value IG(f) of each feature for the feature selection of each node: IG(f)=CE(D)-CE(D|f) Among them, CE(D) is the information entropy of data set D, CE(D|f) is the conditional entropy under the condition of feature f, and the feature with the largest information gain is selected as the split feature to capture the abnormal pattern in the load data of distribution equipment; The splitting threshold is selected by the dynamic splitting threshold determination method. According to the distribution characteristics of the current node data, the probability density function of the data distribution is obtained by the kernel density estimation method, and the optimal splitting point is selected in the local minimum point of the density function. The calculation accuracy is improved by using an improved local anomaly factor algorithm, and the distance metric is optimized by using the generalized Minkowski distance; In the process of calculating the local anomaly factor, the reach-distance is optimized to consider the local density information of the sample points. In the density estimation link, the adaptive bandwidth selection mechanism is combined, and the bandwidth parameter is determined by minimizing the mean square error criterion. The improved isolation forest and local anomaly factor algorithms are integrated, and the detection results of the two algorithms are fused by weighted voting. The weight value is dynamically updated through online learning. The optimized online correction mechanism includes adaptive window design and robust estimation optimization; The adaptive window design establishes a basic window length of 24 hours × 4 sampling times per hour × 7 days, that is, L = 672 sampling points. The window length is adjusted by prediction error feedback. When the continuous prediction error exceeds the set threshold, the window length is adjusted according to L' = L·(1±δ), where δ is a dynamic adjustment coefficient, which is positively correlated with the prediction error. Use exponential decay weighting strategy to optimize window data weighting: w i ＝(1-α) i Among them, α is the forgetting factor, i is the time sequence number of the data point; the forgetting factor α adopts an adaptive adjustment mechanism, with the initial value set to 0.1, and is dynamically adjusted according to the change of the prediction error; when the prediction error increases, increasing the α value accelerates the forgetting of historical data; when the prediction error decreases, reducing the α value maintains the stability of the model, and the adjustment range of α is limited to [0.05, 0.3]; The robust estimation optimization includes improving the M estimation method and enhancing the S estimation method, and processing outliers by improving the M estimation method and combining the Huber loss function; The enhanced S estimation method optimizes the upper limit of the abnormal point ratio tolerated by the estimator. Considering the characteristics of the distribution network load data, the breaking point is set to: ε * =(n-p) / (2n) Among them, n is the number of samples, p is the number of parameters, and in the weight iteration process, the improved weight function w is used i =W(r i / s), where r i is the residual, s is the scale estimate, and during the iteration process, when the Euclidean distance between two adjacent parameter estimates is less than 10 -6 Convergence is determined when .

8. A system using a distribution transformer load forecasting method based on multi-scale fluctuation mode adaptive decomposition as claimed in any one of claims 1 to 7, characterized in that: It includes multi-scale fluctuation decomposition module, characteristic fluctuation library construction module, adaptive matching prediction module and exception handling module; The multi-scale fluctuation decomposition module inputs the original load data, uses the improved maximum overlap discrete wavelet packet transform to perform multi-scale decomposition on the load signal, optimizes the energy distribution, and transmits the decomposed load signal characteristics to the characteristic fluctuation library construction module; The characteristic fluctuation library construction module inputs the load signal characteristics after multi-scale decomposition, performs feature extraction and characterization optimization, and constructs a characteristic fluctuation library to provide data support for the adaptive matching prediction module; The adaptive matching prediction module uses adaptive density estimation and improved multi-level clustering strategy to perform fine characterization of fluctuation patterns, build an efficient index structure, and perform predictions, while passing the prediction results to the exception handling module; The anomaly processing module inputs the prediction results and combines them with the integrated isolation forest and the local anomaly factor to perform anomaly detection optimization, optimizes the online correction mechanism through adaptive window and robust estimation, and the corrected prediction results are used for online monitoring of the power grid safety margin.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of a distribution transformer load forecasting method based on multi-scale fluctuation pattern adaptive decomposition as described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of a distribution transformer load forecasting method based on multi-scale fluctuation pattern adaptive decomposition as described in any one of claims 1 to 7 are implemented.

Images