A digital power distribution network power time series data prediction method and system based on multi-scale dynamic decomposition

Abstract

This invention discloses a method and system for predicting power time-series data in digital distribution networks based on multi-scale dynamic decomposition, aiming to solve the problem that existing prediction methods are unable to accurately capture the complex dynamic characteristics of power system time-series data. The method includes: performing multi-scale decomposition on time-series data; applying dynamic mode decomposition to sequences at different scales to extract mode and time-domain evolution information characterizing their dynamic characteristics; constructing a two-dimensional spatiotemporal representation matrix for each scale sequence based on the mode and evolution information; extracting time-series features containing intra-modal time-domain evolution features and inter-modal correlation features, performing cross-scale fusion to obtain fused features, and generating predicted values ​​for future power time-series data based on these features. This invention utilizes dynamic mode decomposition to capture the inherent dynamic patterns of data at various scales and constructs a unique two-dimensional spatiotemporal representation matrix for feature extraction and fusion, effectively improving the prediction accuracy of power system time-series data.

Description

Technical Field

[0001] This invention belongs to the field of time series forecasting technology, specifically relating to a method and system for forecasting digital distribution network power time series data based on multi-scale dynamic decomposition. Background Technology

[0002] Power system time series data forecasting is a fundamental aspect of power system planning, dispatching, operation, and market trading, and its forecasting accuracy directly impacts the safe, stable operation and economic benefits of the power grid. With the increasing demand for electricity and the large-scale integration of new loads / power sources such as distributed energy and electric vehicles, power system time series data exhibits stronger nonlinearity, non-stationarity, multi-periodicity, and higher volatility and uncertainty, placing higher demands on forecasting accuracy and timeliness. As a key component supporting these complex loads and distributed resources, the refined operation and intelligent control of digital distribution networks rely heavily on high-precision forecasting.

[0003] Traditional forecasting methods are primarily based on statistical models, such as the Autoregressive Moving Average (ARIMA) model and exponential smoothing. These methods are simple in principle and easy to implement, but they often struggle to achieve satisfactory forecast accuracy when dealing with modern power data characterized by complex nonlinearity and non-stationarity. In recent years, with the rapid development of artificial intelligence, machine learning, especially deep learning methods, has demonstrated powerful capabilities in time series forecasting. Recurrent Neural Networks (RNNs) and their variant, Long Short-Term Memory (LSTM), can effectively handle temporal dependencies in sequence data, but they still suffer from problems such as vanishing or exploding gradients and low training efficiency in capturing long-term dependencies. Convolutional Neural Networks (CNNs), especially one-dimensional convolutional networks (1D CNNs), extract local patterns in time series through convolutional kernels and expand the receptive field by stacking layers, but for very long-spanning temporal dependencies, excessively deep network structures may be necessary.

[0004] Transformer models and their variants, such as Informer, Autoformer, and FEDformer, utilize self-attention mechanisms to directly compute the dependencies between any two time points in a sequence, achieving significant progress in long-term time series prediction tasks. However, the computational complexity of standard self-attention mechanisms is quadratic with the sequence length, leading to enormous computational and memory overhead when processing extremely long time series. To address this, researchers have proposed various improvement strategies, such as sparse attention and low-rank approximation, but these often come with increased model complexity or dependence on specific data assumptions.

[0005] To improve model performance and address the complexity of time series data, researchers have introduced various auxiliary techniques. Time series decomposition is a common strategy, aiming to break down the original sequence into more easily modelable components such as trend, seasonal, and residual terms. For example, Autoformer and DLinear use moving averages for decomposition. TimeMixer++ proposes using biaxial attention in the latent space for seasonal and trend decomposition. However, moving average-based decomposition can be too rigid, while attention-based methods, while flexible, increase model complexity. Multiscale analysis is another important approach, aiming to simultaneously capture patterns in time series at different temporal resolutions. For example, TimeMixer and TimeMixer++ generate multiscale sequences through downsampling and design specific mixing mechanisms (such as bottom-up mixing of seasonal terms and top-down mixing of trend terms) to fuse cross-scale information. Despite the success of these methods, effectively extracting and fusing core dynamic features across different scales remains a challenge.

[0006] Dynamic Mode Decomposition (DMD), as a data-driven analysis method, can extract dynamic modes with specific frequencies and growth / decay rates from complex data, revealing the intrinsic dynamic characteristics of a system. Existing techniques have explored the application of DMD to power grid data analysis, such as identifying low-frequency oscillation patterns in the power grid for stability assessment. Some studies have also attempted to combine DMD with machine learning methods for prediction, for example, using features extracted by DMD as state inputs for reinforcement learning. However, these methods primarily treat DMD as a feature extraction tool or an analytical means for specific phenomena.

[0007] On the other hand, constructing effective feature representations for use by deep learning models is also crucial. TimesNet proposed a novel method that identifies sequence periodicity through Fast Fourier Transform (FFT) and reshapes one-dimensional time series into two-dimensional representations. This allows the use of powerful two-dimensional convolutional networks (2D CNNs) to simultaneously capture intra- and inter-periodic variation patterns and process long sequences with near-linear complexity. This approach of converting time series into two-dimensional structures for analysis is inspiring, but its method of constructing two-dimensional representations relies entirely on the periodicity assumption based on FFT. For sequences containing complex transient dynamics and non-periodic fluctuations, it may not be able to fully capture their essential features.

[0008] In summary, existing power system time series data forecasting methods, whether traditional statistical models or advanced deep learning models, still have limitations in accurately capturing the highly nonlinear, non-stationary, and multi-scale dynamic characteristics of modern distribution network time series data. In particular, how to effectively utilize the intrinsic dynamic information revealed by data-driven methods (such as DMD) and transform it into structured input suitable for advanced feature extraction models (such as models handling two-dimensional representations) to improve prediction accuracy and model efficiency remains a pressing technical problem. Therefore, it is necessary to develop a new power system time series data forecasting method that can better integrate multi-scale analysis, intrinsic dynamic pattern extraction, and efficient feature representation and learning mechanisms. Summary of the Invention

[0009] To address the shortcomings of existing technologies, this invention provides a predictive solution that addresses the limitations of current power system time-series data prediction methods in handling the increasingly complex load characteristics in digital distribution networks. By innovatively combining Dynamic Mode Decomposition (DMD) with multi-scale analysis, a novel two-dimensional spatiotemporal representation is constructed to more effectively extract and fuse the inherent dynamic pattern information of power load at different time scales, thereby significantly improving prediction accuracy and supporting the refined operation and intelligent control of digital distribution networks.

[0010] The present invention adopts the following technical solution: This invention protects, in one aspect, a method for predicting power time-series data in digital distribution networks based on multi-scale dynamic decomposition, comprising the following steps: Step 1: Obtain target time series data, which includes historical observations of power system measurement data; Step 2: Perform multi-scale decomposition on the target time series data. By recursively processing the original time series according to a preset downsampling operator, multiple sequence representations at different time scales are obtained. Step 3: Perform dynamic mode decomposition on the sequence representations at multiple different time scales to obtain dynamic modes characterizing the dynamic characteristics of the sequence at each scale and the corresponding feature values ​​of each dynamic mode. Calculate the temporal evolution information of each dynamic mode based on the dynamic modes, feature values, and the initial state of the time series segment. Step 4: Based on the dynamic modes and their corresponding temporal evolution information, construct a two-dimensional spatiotemporal representation matrix, wherein the first dimension represents different dynamic modes and the second dimension represents the temporal evolution information of the dynamic modes; Step 5: Extract features from the two-dimensional spatiotemporal representation matrix at each time scale to obtain time series features at each scale. The time series features include intramodal temporal evolution features and / or intermodal correlation features. Step 6: Perform cross-scale fusion on the time series features at each scale. The fusion includes the transfer of features from the fine time scale to the coarse time scale and / or the transfer from the coarse time scale to the fine time scale, to obtain fused features. Step 7: Based on the fusion features, generate future predicted values ​​for the target time series.

[0011] Furthermore, the power system measurement data includes at least one of the following: historical load data, voltage data, current data, frequency data, electricity price data, weather data, and renewable energy output data.

[0012] Furthermore, in step 2, the target time series data is decomposed into multiple scales by using a scale division method that increases in power of 2, so that the time resolution between adjacent scales is proportional.

[0013] Furthermore, the downsampling operator is one of the following: average pooling operation, max pooling operation, or one-dimensional convolution operation with a preset stride. The average pooling operation involves dividing the sequence into windows along the time dimension and calculating the average value of the elements within the window. The max pooling operation involves dividing the sequence into windows along the time dimension and selecting the maximum value of the elements within the window. The one-dimensional convolution operation involves performing convolution operations on the sequence along the time dimension using a preset one-dimensional convolution kernel with a preset stride value.

[0014] Preferably, an embedding operation is included before step 3: An embedding operation is applied to the sequence representation for each scale index, mapping the original sequence to a high-dimensional sequence with a preset embedding dimension through a one-dimensional convolutional network or a fully connected linear layer.

[0015] Furthermore, step 3, which involves performing dynamic mode decomposition on the multiple sequence representations at different time scales, includes: 1) By constructing a data matrix of adjacent time snapshots; 2) Perform singular value decomposition on the data matrix; 3) Calculate low-dimensional dynamic operators based on singular value decomposition results; 4) Perform eigenvalue decomposition on the low-dimensional dynamic operator to obtain the dynamic modes and eigenvalues; 5) Reconstruct the high-dimensional dynamic mode based on the low-dimensional feature vector matrix.

[0016] Furthermore, step 3, which calculates the time-domain evolution information of each dynamic mode based on the dynamic mode, eigenvalues, and the initial state of the time series segment, includes: i) Using the initial state of the time series as the initial condition, calculate the initial amplitude coefficient of each dynamic mode; ii) Based on the initial amplitude coefficients and dynamic modal characteristic values, calculate the temporal evolution information of each dynamic mode within the time series segment, wherein the temporal evolution information includes at least information on the change of amplitude of each mode over time.

[0017] Furthermore, in step 4, a two-dimensional spatiotemporal representation matrix is ​​constructed based on the dynamic modes and their corresponding temporal evolution information, specifically including: Step 4.1 Based on the preset mode selection criteria, select a preset number of target dynamic modes and their corresponding target time-domain evolution information from the extracted dynamic modes and their corresponding time-domain evolution information; Step 4.2 Construct a two-dimensional matrix as the two-dimensional spatiotemporal representation matrix; Step 4.3 Calculate the scalar value based on the target temporal evolution information and fill the two-dimensional matrix to form a two-dimensional spatiotemporal representation matrix that simultaneously associates the modal dimension and the temporal dimension.

[0018] Furthermore, the calculation method for obtaining the scalar value based on the target time-domain evolution information in step 4.3 is selected from at least one of the following: 1) Extract the real part of the target time-domain evolution information; 2) Extract the imaginary part of the target time-domain evolution information; 3) Calculate the modulus of the target time-domain evolution information; 4) Calculate the phase angle of the target time-domain evolution information; 5) Combine the real and imaginary parts.

[0019] Furthermore, the feature extraction in step 5 employs a two-dimensional model, including: One of the following: a two-dimensional convolutional neural network architecture, a Transformer architecture based on self-attention mechanism and suitable for two-dimensional input, or a multilayer perceptron architecture.

[0020] Furthermore, step 6, which involves cross-scale fusion of the extracted time-series features at each scale, specifically includes: A preset cross-scale information interaction mechanism is used to process the time series features at each scale, aggregate pattern information at different time scales, and generate the fused features. The cross-scale information interaction mechanism includes at least one of a bottom-up feature aggregation path and a top-down feature guidance path; The bottom-up feature aggregation path involves processing the detailed pattern information in the time series features extracted at a fine time scale using a preset uplink mixing operator, and then integrating it into the time series features at a coarser time scale. The top-down feature involves processing the macroscopic pattern information in the time series features extracted from the coarser time scale using a preset downlink mixing operator and then integrating it into the time series features at the finer time scale. The uplink hybrid operator and the downlink hybrid operator are each independently selected from one of the following: linear layer, convolutional layer, or attention module.

[0021] Furthermore, the specific process of generating future predicted values ​​based on the fused features in step 7 is as follows: The fused features are input into a preset prediction head, and the prediction head maps the fused features to the predicted values ​​at future time steps. The prediction head includes at least one of a linear layer and a multilayer perceptron. The linear layer directly transforms the fused features linearly to the required prediction dimension. The multilayer perceptron further learns prediction patterns from the fused features through nonlinear mapping of at least one hidden layer and then outputs prediction values.

[0022] This invention also protects a digital distribution network power time-series data prediction system based on multi-scale dynamic decomposition, the system comprising: The data acquisition module is used to collect and store target time series data containing historical observations of power system measurement data; The multi-scale decomposition module is used to perform multi-scale decomposition operations on the target time series data. By recursively processing the original time series according to a preset downsampling operator, multiple sequence representations at different time scales are obtained. The dynamic mode decomposition module is used to perform dynamic mode decomposition on the sequence representations of the multiple different time scales respectively, to obtain the dynamic modes characterizing the dynamic characteristics of the sequence at each scale and the feature values ​​corresponding to each dynamic mode, and to calculate the time-domain evolution information of each dynamic mode based on the dynamic modes, feature values ​​and the initial state of the time series segment. The two-dimensional spatiotemporal representation construction module is used to construct a two-dimensional spatiotemporal representation matrix based on dynamic modes and their corresponding temporal evolution information; The feature extraction module is used to extract features from the two-dimensional spatiotemporal representation matrix at each time scale to obtain time series features at each scale. The time series features include intramodal temporal evolution features and / or intermodal correlation features. The cross-scale fusion module is used to perform cross-scale fusion operations on time series features at various scales. The fusion includes the transfer of features from a fine time scale to a coarse time scale and / or the transfer of features from a coarse time scale to a fine time scale, generating fused features. The prediction module is used to generate and output future predicted values ​​of the target time series based on the fused features.

[0023] The present invention also protects a terminal, including a processor and a storage medium; The storage medium is used to store instructions; The processor is used to operate according to the instructions to execute the digital distribution network power time series data prediction method based on multi-scale dynamic decomposition.

[0024] This invention also protects a computer-readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method for predicting digital distribution network power time-series data based on multi-scale dynamic decomposition.

[0025] The beneficial effects of this invention are that, compared with the prior art, 1. This invention innovatively introduces Dynamic Mode Decomposition (DMD) at various scales. This mechanism can extract the true physical modes with specific oscillation frequencies and growth or decay rates from messy measurement data. This not only effectively reduces prediction distortion caused by random fluctuations in renewable energy output and sudden load changes, but also makes the prediction results highly interpretable, facilitating insights into the intrinsic causes of grid fluctuations.

[0026] 2. Innovative dimensionality reduction and reconstruction mechanisms enhance the extraction efficiency of complex nonlinear features. Addressing the challenge of extracting deep patterns from one-dimensional long-series signals, this invention constructs a two-dimensional spatiotemporal representation, structurally orthogonally organizing the originally flat "modal information" and "temporal evolution." This transformation enables the model to simultaneously capture the evolutionary patterns of a single mode over time (intra-modal features) and the mutual coupling and interference relationships between different modes (inter-modal features), much like image processing, thus improving the accuracy of perceiving complex operating states of the distribution network.

[0027] 3. The bidirectional interaction of long- and short-cycle characteristics effectively overcomes the limitations of single-scale prediction. Real-world power operation data includes both long-term trends within a day / week and short-term high-frequency disturbances such as cloud cover affecting photovoltaic power. The bottom-up aggregation and top-down guided cross-scale fusion mechanism designed in this invention allows the model to use both macroscopic long-cycle trends as a "benchmark anchor" to prevent prediction divergence and microscopic short-cycle details as "corrective supplements" to capture sudden fluctuations, thus improving the robustness and stability of predictions under extreme conditions.

[0028] 4. Provides a high-fidelity data engine for digital distribution network twin models and refined control. The high-precision, multi-scale prediction data output by this invention can be directly used as underlying input to support the construction and derivation of high-fidelity digital distribution network twin models. High-quality prediction results help the dispatching system break free from overly conservative control strategies forced by "inaccuracy," thereby releasing system margins and improving the overall economic efficiency of resource allocation while ensuring the safe operation of the power grid. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the digital distribution network power time series data prediction method based on multi-scale dynamic decomposition in this invention; Figure 2 This is a schematic diagram of the structure of the digital distribution network power time series data prediction system based on multi-scale dynamic decomposition in this invention. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. The embodiments described in this application are merely some embodiments of this invention, and not all embodiments. Based on the spirit of this invention, other embodiments obtained by those skilled in the art without creative effort are all within the protection scope of this invention.

[0031] To achieve the aforementioned objectives, this invention provides a method and system for predicting power time-series data in digital distribution networks based on multi-scale dynamic decomposition. The following will describe this solution in detail with reference to the accompanying drawings.

[0032] Example 1 Figure 1 The method for predicting power time-series data of digital distribution networks based on multi-scale dynamic decomposition in Embodiment 1 of the present invention includes the following method steps: Step 1: Obtain target time series data, which includes historical observations; the target time series data includes at least one type of power system measurement data, such as historical load data, voltage data, current data, frequency data, electricity price data, weather data, or renewable energy output data.

[0033] Step 2: Perform multi-scale decomposition on the target time series data to obtain multiple sequence representations at different time scales, specifically including the following steps: In this invention, finer and coarser time scales are defined relative to the temporal resolution of each scale sequence after multi-scale decomposition. A finer time scale is defined as one with higher temporal resolution, more sampling points retained per unit time, and a better ability to reflect short-term rapid fluctuations; a coarser time scale is defined as one with lower temporal resolution, formed after more time aggregation, and a better ability to reflect long-term slow changes.

[0034] The target time series data is defined as a sequence representation at the 0th scale, denoted as... ; in, The initial sequence length is... This represents the number of variables in the time series. For scale indexing... Incrementing from 1 to the preset maximum scale number By using the preset downsampling operator Applied to the Sequence representation at each scale Recursively generate the first Sequence representation at each scale Mathematically, it is represented as:

[0035] in, , For the first The factor for level downsampling, and ,make sure Time resolution is lower than , T m That is, the first m The length of the sequence of items, m =0 represents the length of the original sequence; m =1, then T 0 / 2 means half the original length, as... m The growth rates are 2, 3, 4..., corresponding to sequence lengths of one-quarter, one-eighth... of the original.

[0036] The downsampling operator Selected from, but not limited to, average pooling operations: Divide the data into non-overlapping or partially overlapping windows along the time dimension, and calculate the average value of the elements within each window; Max pooling operation: Divide the data into non-overlapping or partially overlapping windows along the time dimension, and select the maximum value of each element within each window; step size. One-dimensional convolution operation: using a pre-defined one-dimensional convolution kernel along the time dimension with a stride of 1 / 2. right Perform convolution operations.

[0037] Specifically, let the original target time series be the series at the 0th scale. Its sampling interval is For scale index By applying a preset downsampling operator (in this embodiment, the downsampling factor is set to 2 for each iteration) to the first... A sequence of scale n is recursively generated to produce the nth scale. Sequences of several scales Therefore, a scale division method with powers of 2 is adopted between each scale, that is, the scale division of the first scale is the power of 2. The equivalent sampling interval corresponding to each scale is: , Accordingly, the first The length of a scale sequence can be expressed as: in, The length of the original sequence. For the first The number of time steps contained in a scale sequence This indicates rounding down. Therefore, it can be seen that with the scale index... As the number of points increases, the temporal resolution of the sequence gradually decreases, while the time range of each covered point gradually increases. If... Then the first The first scale is relative to the first The first scale is a finer time scale, while the second... The first scale is relative to the first The first scale is a coarser time scale. For example, when the sampling interval of the original sequence is 15 minutes, the 0th scale corresponds to a 15-minute resolution, the 1st scale corresponds to a 30-minute resolution, and so on. In this case, the 0th and 1st scales can be regarded as finer time scales, suitable for characterizing detailed patterns such as load spikes, short-term disturbances, and local fluctuations; the 2nd and 3rd scales can be regarded as coarser time scales, suitable for characterizing intraday trends, slow-changing periodic characteristics, and macroscopic evolutionary laws.

[0038] Step 3: Apply dynamic mode decomposition (DMD) to the sequence representations at multiple different time scales to extract the dynamic modes and corresponding temporal evolution information that characterize the dynamic properties of the sequence representations at each scale.

[0039] Before applying dynamic mode decomposition to the sequence representations at the multiple different time scales, the method further includes: For each scale index (in , The sequence representation corresponding to the maximum scale index. Application Embedded Operations This is transformed to obtain the embedded high-dimensional sequence representation:

[0040] in, Indicates the first The time length of the sequence at each scale This represents the number of variables or channels in the target time series data. This indicates the preset embedding dimension.

[0041] The embedding operation This includes using one-dimensional convolutional networks (1D CNNs) or fully connected linear layers to represent the sequence. The numerical values ​​or fragments thereof are mapped to the embedding dimension. .

[0042] The Dynamic Mode Decomposition (DMD) algorithm is applied to sequence representations at multiple different time scales to extract dynamic modes characterizing the dynamic properties of the sequence representations at each scale and the corresponding temporal evolution information. Specifically, this includes: For each scale index (in ): i. From the first High-dimensional sequence representation after embedding at each scale Select at least one time series segment; ii. Apply a preset DMD to at least one selected time series segment to calculate a set of dynamic modes corresponding to the time series segment. and their corresponding eigenvalues , respectively represented as:

[0043]

[0044] in, Indicates the first The number of dynamic modes retained at each scale For modal indexing, Indicates the first The first scale One dynamic mode, Indicates the first Eigenvalues ​​corresponding to each dynamic mode; iii. Based on the dynamic mode eigenvalues Given the initial state of the time series segment, calculate the temporal evolution information of each dynamic mode within the time series segment. The temporal evolution information includes at least the change in amplitude of each mode over time. .

[0045] The application's preset dynamic mode decomposition algorithm specifically adopts the standard DMD algorithm, which is applied to each scale index. The calculation steps for the corresponding time series segments are as follows: (1) Let the first one be the first one. High-dimensional sequence representation after embedding at each scale The selected time series segments include The embedding vectors for each of the consecutive time snapshots are represented as follows:

[0046] in, , .

[0047] Construct the first data matrix E m Second data matrix :

[0048]

[0049] in, and Each is composed of an embedding vector that is staggered at adjacent time points.

[0050] (2) For the first data matrix Performing singular value decomposition (SVD) yields:

[0051] in, It is a left singular vector matrix. It is a singular value diagonal matrix. It is a right singular vector matrix, with the symbol... This indicates the conjugate transpose. The rank retained after singular values ​​are truncated.

[0052] (3) Due to the original high-dimensional embedding space (dimension is Often, there is a lot of redundant information and extremely high computational dimensionality, making it difficult to directly solve for its evolution operators. (satisfy The computational load is enormous; therefore, this invention maps the high-dimensional space to a form derived from the previous method using singular value decomposition (SVD). A low-dimensional subspace composed of dominant singular vectors, and That is, the primitive full-order operator Similarity transformations or projections on this low-dimensional subspace. Based on the singular value decomposition results, calculate the low-dimensional dynamic operator:

[0053] In the formula, -1 in the upper right corner refers to this Finding the inverse of a matrix. It's not a summation; it represents a special singular value diagonal matrix. Physically speaking, it means the system possesses the current data E. m And the next moment E m ', connected via A, E m 'Approximately equal to AE m However, directly calculating A is difficult, so we need to calculate... .

[0054] (4) For the low-dimensional dynamic operator Eigenvalue decomposition yields:

[0055] in, The eigenvalue diagonal matrix of DMD. It is a low-dimensional eigenvector matrix.

[0056] (5) Based on the eigenvalue decomposition results, reconstruct the first... High-dimensional dynamic mode matrix at various scales:

[0057] in, Its column vector Represents the first in the embedding space One dynamic mode.

[0058] (6) The initial state of the time series segment As initial conditions, calculate the initial amplitude coefficients for each dynamic mode:

[0059] in, , Representation matrix The false rebellion, Indicates the first The initial amplitude coefficients of each dynamic mode.

[0060] (7) Based on the initial amplitude coefficient and DMD characteristic value, calculate the first... Temporal evolution information of each dynamic mode within the time series segment:

[0061] in, Indicates the first The first scale A dynamic mode at time... The time-domain evolution term.

[0062] (8) The first The change of the amplitude of each dynamic mode over time is expressed as:

[0063] This yields the temporal evolution information of each dynamic mode within the time series segment.

[0064] Step 4: Based on the extracted dynamic modes and time-domain evolution information at each scale, construct two-dimensional spatiotemporal representations for the sequence representations at each scale, wherein the first dimension of the two-dimensional spatiotemporal representation is associated with the dynamic modes and the second dimension is associated with the time-domain evolution information.

[0065] The method described above, based on the extracted dynamic modes and temporal evolution information at each scale, constructs two-dimensional spatiotemporal representations for the sequence representations at each scale. The method is characterized by constructing two-dimensional spatiotemporal representations for the sequence representations at each scale, specifically including: For each scale index (in From 0 to ): Step 4.1 Based on the preset modality selection criteria, select from the extracted dynamic modes and its corresponding temporal evolution information Select The dynamic modes of each target and their corresponding temporal evolution information ; Step 4.2, construct a two-dimensional matrix As the two-dimensional spatiotemporal representation: For the Let there be a set of scales, and let the modal selection criteria be used to select a total of [number] scales. The target dynamic modes are denoted as follows:

[0066] Suppose the time series segment contains Each time step is denoted as:

[0067] Then construct a two-dimensional matrix:

[0068] or

[0069] or

[0070] As the first Two-dimensional spatiotemporal representation at various scales.

[0071] in: For scale indexing, Indicates the first Number of target dynamic modes selected at each scale Indicates the first The index of a target dynamic mode in the set of all dynamic modes at this scale. , This indicates the number of time steps contained in the time series segment. Represents the first in the time series segment The time corresponding to each time step , Representing a two-dimensional matrix In the line, number The element values ​​of the column; The first dimension of the two-dimensional matrix is ​​used to characterize different target dynamic modes, and the second dimension is used to characterize the temporal evolution information of the target dynamic modes as time steps change.

[0072] Step 4.3, fill the two-dimensional matrix. : For any as well as First determine the first The dynamic mode of the target at time... The time-domain evolution term:

[0073] in, Indicates the first The first scale The initial amplitude coefficients of each target dynamic mode. This represents the eigenvalue corresponding to the dynamic mode of the target.

[0074] The two-dimensional matrix In the line, number Column element values It is based on the aforementioned time-domain evolution term The calculated scalar value is obtained by means of at least one of the following methods: 1) Get the real part:

[0075] 2) Take the imaginary part:

[0076] 3) Taking the mold:

[0077] 4) Take the phase angle:

[0078] 5) Combining the real and imaginary parts: When the two-dimensional matrix is ​​set to At that time, the real and imaginary parts of the time-domain evolution term are respectively filled into the range of the first term. In the two rows corresponding to each target dynamic mode, that is:

[0079]

[0080] Although these two formulas are the same as the formulas for taking the real and imaginary parts above, the row indices on the left side of the equation are different. Methods 1) and 2) above are single-choice, and the matrix dimension is J. m L m This means that if only the real part or only the imaginary part is used, then the first... k The modality is placed in the k-th row, i.e., Z. m ( k , l The fifth method here is to select all and interleave the matrix: the matrix dimension doubles, becoming 2J. m L m In order for the neural network to learn the complete information of complex numbers simultaneously (because conventional CNNs cannot directly process complex numbers), it is necessary to... k The real and imaginary parts of each modality are separated, with the real part placed in the odd-numbered rows (2). k -1), the imaginary part is placed immediately next to the even-numbered row (2). k ).

[0081] In some implementations, if the target temporal evolution information is represented in amplitude form, the element values ​​of the two-dimensional matrix can also be written as:

[0082] in,

[0083] Therefore, a two-dimensional matrix The The line represents the first The evolution trajectory of the target dynamic mode throughout the entire time series segment, the first The column represents the dynamic modes of all targets at time t. The joint state is thus formed to create a two-dimensional spatiotemporal representation that simultaneously associates the modal dimension and the temporal dimension.

[0084] Step 5: Use a preset two-dimensional feature extraction model to process the two-dimensional spatiotemporal representation at each scale and extract the time series features at each scale.

[0085] The pre-set two-dimensional feature extraction model is used. The two-dimensional spatiotemporal representations at each scale are processed separately. (Step e) specifically includes: using the aforementioned two-dimensional feature extraction model For the two-dimensional spatiotemporal representation Processing is performed to extract features that characterize the intrinsic dynamic patterns of the time series segments. The features It must contain at least one or a combination of the following two types of features: (1) Intramodal temporal evolution characteristics, through analysis of the two-dimensional spatiotemporal representation The variation law along its time step dimension (second dimension) is extracted to characterize the characteristics of a single dynamic mode itself as it evolves over time. The characteristics include at least one of the following: the oscillation frequency, amplitude variation law, phase information, and growth or decay rate of the mode. (2) Intermodal correlation characteristics, through analysis of the two-dimensional spatiotemporal representation The two-dimensional feature extraction model is obtained by extracting the variation pattern along its modal dimension (first dimension) or the correlation across different modal dimensions, and is used to characterize the interrelationships between different dynamic modes. These interrelationships include, for example, at least one of the following: relative amplitude ratio, relative energy ratio, phase difference, and temporal evolution correlation between different modes. The selection includes, but is not limited to: two-dimensional convolutional neural network (2DCNN) architecture; Transformer architecture based on self-attention mechanism and suitable for processing two-dimensional input; and multilayer perceptron architecture.

[0086] Step 6: Perform cross-scale fusion on the extracted time-series features at each scale to obtain fused features. Specifically, the cross-scale fusion of the extracted time-series features at each scale includes: A pre-defined cross-scale information interaction mechanism is used to process the time series features at each scale. This allows for the aggregation of pattern information at different time scales to generate the fused features. ; The cross-scale information interaction mechanism includes information flow paths selected from at least one or a combination of the following: 1. Bottom-up feature aggregation path: This will be determined by a finer time scale (e.g., scale) m =0,1 The extracted time series features The detailed pattern information contained within is processed through a preset uplink hybrid operator. After processing, it is incorporated into a coarser timescale (e.g., scale). The time series features of ) In order to enhance the perception of details by coarse-scale features; 2. Top-down feature guidance path: This will be determined by a coarser time scale (e.g., scale). The extracted time series features The macroscopic pattern information contained within is processed through a pre-defined downlink hybrid operator. After processing, it is incorporated into a finer time scale (e.g., scale). The time series features of ) In this context, macroscopic information is used to guide the representation of fine-scale features.

[0087] The uplink hybrid operator and downlink hybrid operators Each layer can be independently selected from linear layers, convolutional layers, or attention modules.

[0088] In this invention, finer and coarser time scales are defined relative to the temporal resolution of each scale sequence after multi-scale decomposition. A finer time scale is defined as one with higher temporal resolution, more sampling points retained per unit time, and a better ability to reflect short-term rapid fluctuations; a coarser time scale is defined as one with lower temporal resolution, formed after more time aggregation, and a better ability to reflect long-term slow changes.

[0089] Coarser and finer time scales involve scaling a time series, such as a year, by powers of 2^m. For example, m=1 transforms the series into half-yearly segments, which are then averaged. m=2 transforms it into quarterly segments, which are also averaged. Finer time scales represent the original and newly averaged time series, such as m=0, 1, 2. Coarser time scales are m=M, M-1, M-2... where M represents the maximum scale number, the maximum number of averaging operations, which is determined manually and can be 3, 4, 5...

[0090] Step 7: Based on the fusion features, generate future predicted values ​​for the target time series.

[0091] The specific steps of generating future predicted values ​​for the target time series based on the fusion features include: The fusion features Input into a pre-defined prediction head. The prediction head Predictions used to map the fused features to future time steps ,in The preset prediction length, The number of output variables; the prediction head A neural network layer comprising at least one or a combination of the following: Linear Layer: directly applying the fused features Linear transformation to the desired prediction dimension; Multi-Layer Perceptron (MLP): a non-linear mapping containing at least one hidden layer for extracting features from the fused data. After further learning the prediction pattern, the predicted value is output.

[0092] Example 2 Figure 2 The present invention provides a power system time series data prediction system based on dynamic mode decomposition and multi-scale hybridization in a digital distribution network, comprising: The data acquisition module is used to collect and store target time series data containing historical observations of power system measurement data; The multi-scale decomposition module is used to perform multi-scale decomposition operations on the target time series data to obtain multiple sequence representations at different time scales; The DMD feature extraction module integrates the embedding submodule and the DMD computation submodule. It first performs embedding operations on the sequence representations at each scale to obtain high-dimensional sequence representations, and then applies dynamic mode decomposition to the high-dimensional sequence representations to extract dynamic mode and time-domain evolution information at each scale. The two-dimensional spatiotemporal representation construction module is used to construct corresponding two-dimensional spatiotemporal representations for sequence representations at each scale based on dynamic modalities and time-domain evolution information at each scale. The feature extraction module is equipped with a preset two-dimensional feature extraction model, which is used to process two-dimensional spatiotemporal representations at various scales and extract time series features; The cross-scale fusion module is used to perform cross-scale fusion operations on time series features at various scales to generate fused features; The prediction module, equipped with a preset prediction head, is used to map the fused features into future predicted values ​​of the target time series and output them.

[0093] Embodiment 3 of the present invention provides a terminal, including a processor and a storage medium; the storage medium is used to store instructions; the processor is used to operate according to the instructions to execute the method steps provided according to Embodiment 1.

[0094] Embodiment 4 of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method steps provided according to Embodiment 1.

[0095] This disclosure can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of this disclosure.

[0096] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0097] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0098] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0099] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the protection scope of the claims of the present invention.

Claims

1. A digital power distribution network power time series data prediction method based on multi-scale dynamic decomposition, characterized in that, Includes the following steps: Step 1: Obtain target time series data, which includes historical observations of power system measurement data; Step 2: Perform multi-scale decomposition on the target time series data. By recursively processing the original time series according to a preset downsampling operator, multiple sequence representations at different time scales are obtained. Step 3: Perform dynamic mode decomposition on the sequence representations at multiple different time scales to obtain dynamic modes characterizing the dynamic characteristics of the sequence at each scale and the corresponding feature values ​​of each dynamic mode. Calculate the temporal evolution information of each dynamic mode based on the dynamic modes, feature values, and the initial state of the time series segment. Step 4: Based on the dynamic modes and their corresponding temporal evolution information, construct a two-dimensional spatiotemporal representation matrix, wherein the first dimension represents different dynamic modes and the second dimension represents the temporal evolution information of the dynamic modes; Step 5: Extract features from the two-dimensional spatiotemporal representation matrix at each time scale to obtain time series features at each scale. The time series features include intramodal temporal evolution features and / or intermodal correlation features. Step 6: Perform cross-scale fusion on the time series features at each scale. The fusion includes the transfer of features from the fine time scale to the coarse time scale and / or the transfer from the coarse time scale to the fine time scale, to obtain fused features. Step 7: Based on the fusion features, generate future predicted values ​​for the target time series.

2. The method for digital power distribution network power time series data prediction based on multi-scale dynamic decomposition according to claim 1, characterized in that, The power system measurement data includes at least one of the following: historical load data, voltage data, current data, frequency data, electricity price data, weather data, and renewable energy output data.

3. The method of claim 1, wherein, In step 2, the target time series data is decomposed into multiple scales by using a scale division method that increases in power of 2, so that the time resolution between adjacent scales is proportional.

4. The method for predicting digital distribution network power time series data based on multi-scale dynamic decomposition according to claim 1, characterized in that, The downsampling operator is one of the following: average pooling, max pooling, or one-dimensional convolution with a preset stride. The average pooling operation involves dividing the sequence into windows along the time dimension and calculating the average value of the elements within the window. The max pooling operation involves dividing the sequence into windows along the time dimension and selecting the maximum value of the elements within the window. The one-dimensional convolution operation involves performing convolution operations on the sequence along the time dimension using a preset one-dimensional convolution kernel with a preset stride value.

5. The method for digital power distribution network power time series data prediction based on multi-scale dynamic decomposition according to claim 1, characterized in that, Step 3 is preceded by an embedding operation: An embedding operation is applied to the sequence representation for each scale index, mapping the original sequence to a high-dimensional sequence with a preset embedding dimension through a one-dimensional convolutional network or a fully connected linear layer.

6. The method for digital power distribution network power time series data prediction based on multi-scale dynamic decomposition according to claim 1, characterized in that, Step 3 involves performing dynamic mode decomposition on the multiple sequence representations at different time scales, including: 1) By constructing a data matrix of adjacent time snapshots; 2) Perform singular value decomposition on the data matrix; 3) Calculate low-dimensional dynamic operators based on singular value decomposition results; 4) Perform eigenvalue decomposition on the low-dimensional dynamic operator to obtain the dynamic modes and eigenvalues; 5) Reconstruct the high-dimensional dynamic mode based on the low-dimensional feature vector matrix.

7. The method for predicting digital distribution network power time series data based on multi-scale dynamic decomposition according to claim 5, characterized in that, Step 3, based on the dynamic modes, eigenvalues, and the initial states of the time series segments, calculates the temporal evolution information of each dynamic mode, including: i) Using the initial state of the time series as the initial condition, calculate the initial amplitude coefficient of each dynamic mode; ii) Based on the initial amplitude coefficients and dynamic modal characteristic values, calculate the temporal evolution information of each dynamic mode within the time series segment, wherein the temporal evolution information includes at least information on the change of amplitude of each mode over time.

8. The method for predicting digital distribution network power time series data based on multi-scale dynamic decomposition according to claim 1, characterized in that, Step 4 involves constructing a two-dimensional spatiotemporal representation matrix based on the dynamic modes and their corresponding temporal evolution information, specifically including: Step 4.1 Based on the preset mode selection criteria, select a preset number of target dynamic modes and their corresponding target time-domain evolution information from the extracted dynamic modes and their corresponding time-domain evolution information; Step 4.2 Construct a two-dimensional matrix as the two-dimensional spatiotemporal representation matrix; Step 4.3 Calculate the scalar value based on the target temporal evolution information and fill the two-dimensional matrix to form a two-dimensional spatiotemporal representation matrix with associated modal dimension and time dimension.

9. The method for predicting digital distribution network power time series data based on multi-scale dynamic decomposition according to claim 8, characterized in that, The calculation method for obtaining the scalar value based on the target time-domain evolution information in step 4.3 is selected from at least one of the following: 1) Extract the real part of the target time-domain evolution information; 2) Extract the imaginary part of the target time-domain evolution information; 3) Calculate the modulus of the target time-domain evolution information; 4) Calculate the phase angle of the target time-domain evolution information; 5) Combine the real and imaginary parts.

10. The method for predicting power time-series data of digital distribution networks based on multi-scale dynamic decomposition according to claim 1, characterized in that, The feature extraction in step 5 uses a two-dimensional model, including: One of the following: a two-dimensional convolutional neural network architecture, a Transformer architecture based on self-attention mechanism and suitable for two-dimensional input, or a multilayer perceptron architecture.

11. The method for predicting power time-series data of digital distribution networks based on multi-scale dynamic decomposition according to claim 1, characterized in that, Step 6, which involves cross-scale fusion of the extracted time-series features at each scale, includes: A preset cross-scale information interaction mechanism is used to process the time series features at each scale, aggregate pattern information at different time scales, and generate fused features. The cross-scale information interaction mechanism includes at least one of a bottom-up feature aggregation path and a top-down feature guidance path; The bottom-up feature aggregation path involves processing the detailed pattern information in the time series features extracted at a fine time scale using a preset uplink mixing operator, and then integrating it into the time series features at a coarser time scale. The top-down feature involves processing the macroscopic pattern information in the time series features extracted from the coarser time scale using a preset downlink mixing operator, and then integrating it into the time series features at the finer time scale. The uplink and downlink hybrid operators are each independently selected from linear layers, convolutional layers, or attention modules.

12. The method for predicting digital distribution network power time series data based on multi-scale dynamic decomposition according to claim 1, characterized in that, The specific process of generating future predicted values ​​based on the fused features in step 7 is as follows: The fused features are input into a preset prediction head, and the prediction head maps the fused features to the predicted values ​​of future time steps. The prediction head includes at least one of a linear layer and a multilayer perceptron. The linear layer directly transforms the fused features linearly to the required prediction dimension. The multilayer perceptron further learns prediction patterns from the fused features through nonlinear mapping of at least one hidden layer and then outputs prediction values.

13. A digital distribution network power time-series data prediction system based on multi-scale dynamic decomposition, characterized in that, The system for implementing the method of any one of claims 1-12, the system comprising: The data acquisition module is used to collect and store target time series data containing historical observations of power system measurement data; The multi-scale decomposition module is used to perform multi-scale decomposition operations on the target time series data. By recursively processing the original time series according to a preset downsampling operator, multiple sequence representations at different time scales are obtained. The dynamic mode decomposition module is used to perform dynamic mode decomposition on the sequence representations of the multiple different time scales respectively, to obtain the dynamic modes characterizing the dynamic characteristics of the sequence at each scale and the feature values ​​corresponding to each dynamic mode, and to calculate the time-domain evolution information of each dynamic mode based on the dynamic modes, feature values ​​and the initial state of the time series segment. The two-dimensional spatiotemporal representation construction module is used to construct a two-dimensional spatiotemporal representation matrix based on dynamic modes and their corresponding temporal evolution information; The feature extraction module is used to extract features from the two-dimensional spatiotemporal representation matrix at each time scale to obtain time series features at each scale. The time series features include intramodal temporal evolution features and / or intermodal correlation features. The cross-scale fusion module is used to perform cross-scale fusion operations on time series features at various scales. The fusion includes the transfer of features from a fine time scale to a coarse time scale and / or the transfer of features from a coarse time scale to a fine time scale, generating fused features. The prediction module is used to generate and output future predicted values ​​of the target time series based on the fused features.

14. A terminal, comprising a processor and a storage medium; characterized in that: The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the digital distribution network power time series data prediction method based on multi-scale dynamic decomposition according to any one of claims 1-12.

15. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the digital distribution network power time series data prediction method based on multi-scale dynamic decomposition as described in any one of claims 1-12.

Images