Annual load scene generation method fusing STL decomposition and WK-GP

Abstract

The invention discloses an annual load scene generation method fusing STL decomposition and WK-GP. According to the method provided by the invention, a load sequence is decomposed into three components with definite physical significance, namely a trend component, a period component and a residual component, and a time sequence extrapolation method based on a clustering and Gaussian process and a probability modeling method based on joint distribution sampling are respectively adopted aiming at different characteristics of the periodic component and the residual component; and finally, the generated components are superposed and synthesized, so that the defect that the load time sequence dependence characteristic and the random uncertainty are difficult to consider in the related technology is effectively overcome, the generated annual load scene can accurately reflect the intra-day fluctuation and randomness while keeping the long-term seasonal and periodic rule, and the method is suitable for large-scale popularization and application. And the time sequence authenticity, the statistical accuracy and the integrity of uncertainty expression of the scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the technical field of annual load scenario generation, and specifically to a method for generating annual load scenarios that integrates STL decomposition and WK-GP. Background Technology

[0002] With the continuous increase in installed capacity of new energy sources and the large-scale grid connection of new loads, the power system exhibits strong uncertainties on both the source and load sides. In recent years, a large number of studies have been conducted on the uncertainties of wind and solar power, revealing the multi-timescale output characteristics of wind and solar power from different dimensions. As the process of electricity substitution accelerates and new loads are widely connected, the load side is also showing increasingly strong time-varying and stochastic characteristics, and traditional static load models are gradually becoming ineffective.

[0003] In the process of generating annual load scenarios for power systems, there is a technical problem that the generated scenarios are distorted because the relevant technologies cannot simultaneously and accurately capture the load time-series dependence characteristics (seasonality, periodicity) and random uncertainty. Summary of the Invention

[0004] The technical problem this invention aims to solve is that, in the process of generating annual load scenarios for power systems, related technologies often fail to accurately capture both the time-series dependence characteristics (seasonality and periodicity) and random uncertainties of load simultaneously, leading to distortion in the generated scenarios. The purpose is to provide a method for generating annual load scenarios that integrates STL decomposition and WK-GP, thereby resolving the problem of distorted generated scenarios.

[0005] This invention is achieved through the following technical solution:

[0006] In a first aspect, the present invention provides a method for generating annual load scenarios that integrates STL decomposition and WK-GP, comprising the following steps:

[0007] Acquire the load data sequence of the target area within a preset time period and at a preset time resolution;

[0008] The load data sequence is decomposed based on a preset signal decomposition algorithm to obtain a first component, a second component, and a third component that represent different physical characteristics respectively; wherein, the first component is used to represent the long-term variation trend of the load, the second component is used to represent the periodic fluctuation of the load, and the third component is used to represent the random fluctuation of the load.

[0009] A preset first probability modeling is performed on the second component to generate an annual second component curve; wherein, the first probability modeling is expressed as clustering of typical periodic patterns and temporal extrapolation of pattern evolution;

[0010] A preset second probability modeling is performed on the third component to generate an annual third component curve; wherein, the second probability modeling is represented by random sampling based on the joint probability distribution of the third component at multiple time points;

[0011] The first component, the second annual component curve, and the third annual component curve are superimposed according to corresponding time points to synthesize the annual load scenario.

[0012] Further, the step of decomposing the load data sequence based on a preset signal decomposition algorithm to obtain a first component, a second component, and a third component representing different physical characteristics includes:

[0013] Based on a preset period length, the load data sequence is initialized with periodic components to obtain an initial second component; wherein, the period length corresponds to the load fluctuation period;

[0014] The following sub-steps are executed alternately using an iterative smoothing filtering algorithm until the convergence condition is met: the second component is removed from the load data sequence to obtain a deperiodic sequence; local weighted regression smoothing is performed on the deperiodic sequence to estimate the first component; the first component is subtracted from the load data sequence to obtain a temporary residual sequence; the temporary residual sequence is combined with the second component to reconstruct a periodic sequence, and intra-periodic smoothing is performed on the reconstructed sequence to update the second component.

[0015] After the iteration is completed, the third component is derived from the load data sequence based on the final first and second components.

[0016] Further, the step of performing a preset first probability modeling for the second component to generate an annual second component curve includes:

[0017] Cluster analysis is performed on the periodic patterns included in the second component to identify at least one typical periodic pattern;

[0018] Based on the results of cluster analysis, a pattern weight sequence corresponding to the time series is generated;

[0019] The pattern weight sequence is modeled using a stochastic process model, and a simulated weight sequence is generated.

[0020] The simulated weight sequence is combined with the corresponding typical periodic pattern to reconstruct the annual second component curve.

[0021] Further, the step of performing cluster analysis on the periodic patterns included in the second component to determine at least one typical periodic pattern includes:

[0022] The optimal number of clusters for the cluster analysis is determined based on the silhouette coefficient or variance analysis.

[0023] Based on the optimal number of clusters, the periodic patterns included in the second component are divided to obtain each typical periodic pattern and its corresponding cluster center.

[0024] Furthermore, the step of modeling the pattern weight sequence using a stochastic process model and generating a simulated weight sequence includes:

[0025] Based on the time-series observations of the pattern weight sequence within the historical time period, the kernel function parameters of the stochastic process model are determined;

[0026] The covariance matrix between time points is constructed using the kernel function to characterize the correlation strength of the weight values ​​at different times.

[0027] Random sampling is performed based on the covariance matrix to generate a simulated weight sequence for the whole year that satisfies the correlation characteristics;

[0028] The generated simulated weight sequence is normalized so that the sum of the weights of each mode is one for each day.

[0029] Further, the step of performing a preset second probability modeling for the third component to generate an annual third component curve includes:

[0030] Determine the marginal probability distribution of the third component at multiple time points;

[0031] Based on the determined marginal probability distribution, a joint probability distribution model at multiple time points is constructed through a correlation coupling function, wherein the correlation coupling function is used to characterize the dependency structure of the residual values ​​at different time points;

[0032] A random sequence satisfying the dependency structure is generated from the joint probability distribution model using a random sampling algorithm, forming the annual third component curve.

[0033] Further, the step of superimposing the first component, the second annual component curve, and the third annual component curve according to corresponding time points to synthesize the annual load scenario includes:

[0034] Time point alignment processing is performed on the first component, the second annual component curve, and the third annual component curve to ensure that the three component values ​​with the same timestamp have a corresponding relationship.

[0035] The three aligned components are superimposed at each time point to obtain the composite load value at that time point;

[0036] Repeat the overlay process for all time points to generate a complete annual load scenario sequence;

[0037] Physical rationality checks are performed on the generated annual load scenarios to ensure that the load values ​​are within a preset reasonable range.

[0038] Secondly, this invention provides an annual load scenario generation device that integrates STL decomposition and WK-GP, comprising:

[0039] The acquisition module is used to acquire the load data sequence of the target area within a preset time period and at a preset time resolution;

[0040] The decomposition module is used to decompose the load data sequence based on a preset signal decomposition algorithm to obtain a first component, a second component, and a third component that respectively characterize different physical characteristics; wherein, the first component is used to characterize the long-term variation trend of the load, the second component is used to characterize the periodic fluctuation of the load, and the third component is used to characterize the random fluctuation of the load.

[0041] A first generation module is used to perform a preset first probability modeling on the second component to generate an annual second component curve; wherein the first probability modeling is expressed as clustering of typical periodic patterns and temporal extrapolation of pattern evolution.

[0042] The second generation module is used to perform a preset second probability modeling for the third component to generate an annual third component curve, wherein the second probability modeling is represented by random sampling based on the joint probability distribution of the third component at multiple time points;

[0043] The synthesis module is used to superimpose the first component, the annual second component curve, and the annual third component curve according to the corresponding time points to synthesize the annual load scenario.

[0044] Thirdly, the present invention provides an electronic device, comprising: a memory, and one or more processors communicatively connected to the memory; the memory stores instructions executable by the one or more processors, the instructions being executed by the one or more processors to cause the one or more processors to implement the method described above.

[0045] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0047] The method provided by this invention decomposes the load sequence into three components with clear physical meaning: trend, period, and residual. It then employs time-series extrapolation based on clustering and Gaussian processes and probabilistic modeling based on joint distribution sampling, respectively, to address the different characteristics of the periodic and residual components. Finally, the generated components are superimposed and synthesized. This method effectively overcomes the shortcomings of related technologies that struggle to simultaneously consider the time-series dependence of load and random uncertainty. The generated annual load scenario maintains long-term seasonality and periodicity while accurately reflecting intraday fluctuations and randomness, significantly improving the time-series realism, statistical accuracy, and completeness of uncertainty representation. Attached Figure Description

[0048] To more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be considered as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort. In the drawings:

[0049] Figure 1 This is one of the flowcharts for a method to generate an annual load scenario that integrates STL decomposition and WK-GP, as provided in the embodiments of this specification.

[0050] Figure 2 The second flowchart is a method for generating annual load scenarios that integrates STL decomposition and WK-GP, as provided in the embodiments of this specification.

[0051] Figure 3 This is a schematic diagram of the annual load trend components provided in the embodiments of this specification;

[0052] Figure 4 This is a schematic diagram of the generated annual load curve provided in the embodiments of this specification;

[0053] Figure 5 This is a schematic diagram of the quantitative indicators of the generated curves provided in the embodiments of this specification;

[0054] Figure 6 A schematic diagram of ACF coefficient kernel density estimation at lag time provided in the embodiments of this specification;

[0055] Figure 7 This is a schematic diagram of the KL divergence statistics of the ACF coefficients provided in the embodiments of this specification. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0057] In related technologies, with the continuous growth of new energy installed capacity and the large-scale integration of new loads (such as electric vehicles and distributed energy storage), modern power systems exhibit significant randomness and volatility on both the source and load sides. Against this backdrop, the uncertainty of annual load variations must be fully considered when conducting medium- and long-term planning, reliability assessments, and operational strategy optimization for power systems. Therefore, generating a large number of annual load scenarios that reflect both the inherent variation patterns of loads and characterize their random fluctuations has become a key fundamental task in this technical field.

[0058] In related technologies, the generation of annual load scenarios often relies on the following technical solutions:

[0059] The first category is based on historical data selection and splicing. This method directly selects representative daily or monthly load curves from historical data and splices them to form a typical annual curve. Although this method is simple and intuitive, it is essentially a repetition of historical data and cannot effectively represent the random fluctuations in future loads that may deviate from historical records, thus failing to meet the needs of uncertainty analysis.

[0060] The second category is random sampling methods based on parameterized distributions. This method assumes that the load follows a known probability distribution (e.g., a normal distribution) and generates scenarios through parameter estimation and random sampling. However, actual load changes are driven by multiple complex factors such as climate, economy, and social activities, and their true distribution is often difficult to accurately describe by any pre-defined simple mathematical model. This model mismatch caused by "distribution assumption bias" significantly limits the accuracy and adaptability of the generated scenarios.

[0061] The third category is the method that combines Markov chains and Generative Adversarial Networks (GANs). This method first uses Markov chains to simulate the transfer process of daily load types throughout the year, then uses GANs to generate daily load curves for each type, and finally splices them together. This method improves flexibility to some extent, but its core drawback is that the way it constructs the annual scene by independently generating and splicing daily curves disrupts the natural, long-term temporal dependence structure of the load (e.g., weekly correlation, holiday effects, etc.), which may cause the generated annual scene to have "sponge-up distortion" in terms of time series, failing to accurately reflect the smooth transition and long-term dynamics of the load from one day to the next.

[0062] In summary, the core technical problem in the generation of annual load scenarios for power systems can be summarized as follows: existing methods cannot simultaneously and accurately capture the complex temporal dependence characteristics of load (including seasonality, periodicity, and holiday effects) and real random uncertainty within the same framework, resulting in distortion of the generated annual load scenarios.

[0063] The root of this technical problem lies in the fact that the relevant technologies are often static and deterministic, lacking the representation of uncertainty, or when introducing randomness, they ignore or destroy the inherent multi-timescale temporal patterns of the load, or simplify the generation of long-term scenarios into a simple splicing of short-term scenarios, failing to establish a coherent and holistic temporal model.

[0064] The core inventive concept of this invention lies in: firstly, decoupling the original load sequence into components representing different physical characteristics using signal decomposition technology; then, employing the most suitable probabilistic modeling method for the characteristics of different components; and finally, solving the aforementioned technical problems through reconstruction and synthesis techniques. Specifically, after decomposing the load sequence into three components—trend, period, and residual—the periodic component is generated using a time-series extrapolation method based on clustering and Gaussian processes, and the residual component is generated using a joint distribution sampling method based on coupling functions. These components are then superimposed to form an annual scene. This concept ensures that the generated scene possesses both accurate temporal patterns and realistic random fluctuation characteristics.

[0065] like Figure 1 As shown, this embodiment provides a method for generating annual load scenarios that integrates STL decomposition and WK-GP. The execution entity of this embodiment can be a central server or a cloud computing platform. Alternatively, the execution entity can be equipment in a power system control center, specifically a control server including an Energy Management System (EMS), a Supervisory Control and Data Acquisition (SCADA) system, or a Distribution Management System (DMS). These systems can be directly connected to data acquisition devices such as smart meters and sensors, enabling real-time acquisition of load data and embedding this method as an advanced application function. The execution entity of this embodiment can also be an edge computing device or a smart gateway, or a dedicated analysis terminal device, etc.

[0066] The method may include the following steps:

[0067] Step S10: Obtain the load data sequence of the target area within a preset time period and at a preset time resolution.

[0068] In this embodiment, the acquisition action can be represented as the behavior of the executing entity receiving or extracting load data from the data source to ensure that the data can be used for subsequent processing.

[0069] In this embodiment, the load data sequence can be represented as a series of power load values ​​arranged in chronological order, thereby representing the power consumption of the target area over a certain period of time.

[0070] In this embodiment, the target area can be represented as a geographical or logical power network range, such as a province, city, or distribution area. The preset time period can be one year, or multiple months or years, etc.

[0071] In this embodiment, the preset time resolution can be expressed as the time interval for data acquisition, such as 1 hour, 30 minutes, or 15 minutes, etc.

[0072] In one specific implementation, the data source may include smart meters, sensors, SCADA systems, or historical databases, etc.

[0073] Step S12: Decompose the load data sequence based on a preset signal decomposition algorithm to obtain a first component, a second component, and a third component that represent different physical characteristics respectively; wherein, the first component is used to represent the long-term change trend of the load, the second component is used to represent the periodic fluctuation of the load, and the third component is used to represent the random fluctuation of the load.

[0074] In this embodiment, during this step, the executing entity can split the load data sequence into three components, each capturing different physical characteristics.

[0075] In this embodiment, the preset signal decomposition algorithm can be any method capable of decomposing a time series into trend, periodic, and residual components. The first component can be used to characterize the long-term trend of load changes, such as seasonal growth or holiday effects. The second component can be used to characterize periodic fluctuations, such as daily or weekly cycles. The third component can be used to characterize random fluctuations, such as noise or sudden events.

[0076] In a specific implementation, the signal decomposition algorithm may include the STL (Seasonal-Trend Decomposition using Loess) algorithm, wavelet transform algorithm, or empirical mode decomposition algorithm, etc. The STL algorithm can achieve decomposition through iterative smoothing.

[0077] Specifically, the periodic components are first initialized, and then trend estimation and periodic updates are performed alternately until convergence. In the implementation, the period length can be set to 24 hours (daily cycle) or 168 hours (weekly cycle). The first component (trend component) can reflect annual load level changes, the second component (periodic component) can be refined into weekday and weekend patterns, and the third component (residual component) can include weather effects or random faults. The decomposition process can also adjust parameters, such as the smoothing window size, to control the smoothness of the components.

[0078] Step S14: Perform a preset first probability modeling for the second component to generate an annual second component curve; wherein, the first probability modeling is represented as clustering of typical periodic patterns and temporal extrapolation of pattern evolution.

[0079] In this embodiment, this step can be based on historical cyclical components (the second component), generating annual cyclical component curves with reasonable temporal evolution through probabilistic modeling. The first probabilistic modeling can be represented as a hybrid modeling framework combining unsupervised learning and time series prediction. The clustering of typical cyclical patterns can be represented as a process of grouping historical daily cyclical curves (i.e., daily segments of the second component) according to morphological similarity. For example, the K-Means clustering algorithm can be used, or other clustering methods such as hierarchical clustering and DBSCAN can be employed. The typical cyclical pattern can be represented as the center curve of each cluster, representing a typical daily load cyclical pattern.

[0080] In this embodiment, the temporal extrapolation of pattern evolution can be represented as modeling and predicting the weight sequence of the occurrence and transformation of the aforementioned typical patterns throughout the year. Specifically, firstly, based on the clustering results, a weight vector representing the pattern to which each day in history belongs can be assigned. Then, this weight vector sequence is regarded as a multivariate time series and can be modeled using a Gaussian process (GP). The Gaussian process can be a non-parametric Bayesian probabilistic model, the specific characteristics of which can be determined by the mean function and the kernel function. The kernel function can be a radial basis function (RBF), a Marton kernel, etc., to capture the smoothness and correlation of the weight sequence over time. Finally, a simulated weight sequence for the whole year is generated by sampling from the trained Gaussian process, and the simulated weight of each day is linearly combined with the corresponding typical periodic pattern (cluster center) to reconstruct the annual second component curve.

[0081] Step S16: Perform a preset second probability modeling for the third component to generate an annual third component curve, wherein the second probability modeling is represented by random sampling based on the joint probability distribution of the third component at multiple time points.

[0082] In this embodiment, the second probability modeling can be achieved by using a joint probability distribution to model the multi-time point relationship of the third component and generating curves through random sampling.

[0083] In this embodiment, the joint probability distribution can be represented as a probability model describing the interdependence of residual values ​​at multiple time points, and random sampling can be represented as drawing a random sequence from the distribution.

[0084] Specifically, in one implementation, the joint probability distribution can be achieved using a Copula function (e.g., Gaussian Copula or t-Copula) or a nonparametric method. More specifically, first, the marginal distribution of the third component at each time point can be estimated, for example, using kernel density estimation or a parametric distribution (normal distribution, Gamma distribution). Then, a Copula model is constructed to capture the correlations between time points, such as the temporal dependence of hourly residuals. Random sampling can be performed using Monte Carlo methods, Latin hypercube sampling, or quasi-Monte Carlo methods, and the generated sequences must satisfy marginal distribution and correlation constraints.

[0085] Step S18: Superimpose the first component, the annual second component curve, and the annual third component curve according to the corresponding time points to synthesize the annual load scenario.

[0086] In this embodiment, the superposition can be achieved by arithmetically adding the three components at the same time point to generate the final load scenario.

[0087] The method provided in this embodiment decomposes the load sequence into three components with clear physical meaning: trend, period, and residual. It then employs time-series extrapolation based on clustering and Gaussian processes and probabilistic modeling based on joint distribution sampling, respectively, to address the different characteristics of the periodic and residual components. Finally, the generated components are superimposed and synthesized. This effectively overcomes the shortcomings of related technologies that struggle to simultaneously consider the time-series dependence of load and random uncertainty. The generated annual load scenario maintains long-term seasonality and periodicity while accurately reflecting intraday fluctuations and randomness, significantly improving the time-series realism, statistical accuracy, and completeness of uncertainty representation.

[0088] In some embodiments, the step of decomposing the load data sequence based on a preset signal decomposition algorithm to obtain a first component, a second component, and a third component representing different physical characteristics includes:

[0089] Step S122: Based on a preset period length, initialize the load data sequence with periodic components to obtain an initial second component; wherein, the period length corresponds to the load fluctuation period.

[0090] In this embodiment, this step can provide an initial estimate of the periodic components for the subsequent iterative decomposition process.

[0091] In this embodiment, the preset cycle length can be a predefined duration parameter that characterizes the main repetitive fluctuation pattern of the load data. For example, it can be set to 24 hours to correspond to the most significant intraday cycle characteristics, but in some application scenarios, it can also be set to 168 hours, etc., to correspond to the weekly cycle characteristics, or set to other cycle lengths according to the load type (e.g., industrial load).

[0092] The initialization of the load data sequence by periodic components to obtain the initial second component can be represented as dividing the entire load data sequence into several continuous periodic segments based on the aforementioned period length, and calculating the typical value of each time point within the period by aggregating the periodic segments, thereby forming a reference period curve.

[0093] In one specific implementation, the average value of all load data at the same location within each cycle (e.g., 0:00, 1:00, and even 23:00 each day) can be calculated, and the average values ​​can be arranged in order to form the initial second component.

[0094] In a specific implementation plan, the median or other robust statistics can also be used for aggregation to reduce the impact of outliers.

[0095] Step S124: Using an iterative smoothing filtering algorithm, alternately execute the following sub-steps until the convergence condition is met: remove the second component from the load data sequence to obtain a deperiodic sequence; perform local weighted regression smoothing on the deperiodic sequence to estimate the first component; subtract the first component from the load data sequence to obtain a temporary residual sequence; combine the temporary residual sequence with the second component to reconstruct a periodic sequence, and perform intra-periodic smoothing on the reconstructed sequence to update the second component.

[0096] In this embodiment, the pure trend component (first component) and periodic component (second component) can be gradually optimized and separated through multiple iterations.

[0097] In this implementation, the following sub-steps are executed alternately in each iteration:

[0098] Step S1242: Remove the second component from the load data sequence to obtain a deperiodic sequence.

[0099] Specifically, this sub-step can be used to subtract the corresponding value of the currently estimated periodic component (the initial second component in the first iteration, and the updated second component in subsequent iterations) point by point from the original load data sequence. The purpose of this subtraction operation is to remove periodic fluctuations in the data, thereby obtaining an intermediate sequence that mainly contains trend components and random noise, which is the deperiodic sequence.

[0100] Step S1244: Perform local weighted regression smoothing on the deperiodic sequence to estimate the first component.

[0101] Specifically, this sub-step can apply a low-pass filter to the aforementioned de-periodic sequence to extract its slowly changing trend component. The local weighted regression smoothing process can be performed using the Loess (Locally Estimated ScatterplotSmoothing) algorithm. More specifically, in this algorithm, for each target data point in the sequence, a local neighborhood window centered on that point can be selected. Within this window, a low-order polynomial (which can be first or second order) can be fitted using weighted least squares, where the weights decrease with distance from the target point. The value of this fitted polynomial at the target point is then estimated as the value of the trend component at that point. By traversing all points in the sequence, the first component updated in this iteration can be obtained.

[0102] Step S1246: Subtract the first component from the load data sequence to obtain a temporary residual sequence.

[0103] Specifically, this sub-step involves subtracting the first component, obtained from the current iteration, point by point from the original load data sequence. This operation removes the influence of trends, resulting in a mixed sequence that mainly contains periodic fluctuations and random noise, namely the temporary residual sequence.

[0104] Step S1248: Combine the temporary residual sequence with the second component to reconstruct the periodic sequence, and perform intra-period smoothing on the reconstructed sequence to update the second component.

[0105] Specifically, this sub-step can utilize a temporary residual sequence to update the periodic component. First, the temporary residual sequence can be grouped according to the preset period length to form multiple periodic segments. Then, for each periodic segment, the local weighted regression smoothing (Loess) is applied again for intra-period smoothing, the purpose of which is to eliminate random noise within each periodic segment and obtain a smoother periodic shape. Finally, all smoothed periodic segments are reassembled in chronological order to form the updated second component for the next iteration.

[0106] In this embodiment, the condition of "until convergence is met" is the termination condition for the loop. That is, when the change in the residual sequence or periodic component obtained from the previous two iterations is less than a preset threshold, the decomposition result is considered to be stable. Alternatively, a maximum number of iterations can be set as a safeguard to prevent infinite loops.

[0107] Step S126: After the iteration is completed, the third component is derived from the load data sequence based on the final first component and the second component.

[0108] In this embodiment, the final determined first component (trend) and second component (period) can be simultaneously subtracted from the original load data sequence. That is, for each time point in the sequence, the following calculation is performed: Third component = Original load value - First component - Second component. The resulting sequence obtained through this calculation is the third component, which represents all the remaining parts of the original data that cannot be explained by the trend and period, i.e., random fluctuations and noise.

[0109] In some implementations, the step of performing a preset first probability modeling for the second component to generate an annual second component curve includes:

[0110] Step S142: Perform cluster analysis on the periodic patterns included in the second component to determine at least one typical periodic pattern.

[0111] In this embodiment, the periodic component (second component) in historical data can be summarized to extract representative typical daily periodic patterns. The periodic pattern can be represented as a load change curve within a complete cycle (which can be one day), that is, the daily segment obtained by cutting the second component according to the preset cycle length (24 hours).

[0112] Step S144: Based on the results of cluster analysis, generate a pattern weight sequence corresponding to the time series.

[0113] In this embodiment, the membership relationship between the cyclical pattern and various typical cyclical patterns of each historical day can be quantified, thereby forming a weighted sequence that changes over time. The pattern weighted sequence can be a multivariate time series, where each time point (each day) corresponds to a weight vector, which describes the proportion in which the actual cyclical pattern of that day is composed of various typical cyclical patterns.

[0114] In one specific implementation, for each historical day, the similarity distance between its daily cycle pattern and each typical cycle pattern (cluster center) obtained in step S142 is calculated. The similarity distance can be Euclidean distance, cosine similarity, etc. Based on the similarity distance, a weight vector can be assigned to each day.

[0115] For example, it could be one-hot encoding, which means it belongs entirely to the nearest typical pattern, with a weight of 1 for the rest.

[0116] Alternatively, a Gaussian kernel function can be used to convert distance into similarity and then normalized so that the sum of the weights for all typical patterns for each day is 1. Thus, for a dataset containing K typical patterns and T historical dates, a weight matrix of size T×K can be generated. This matrix, arranged chronologically, constitutes the pattern weight sequence. This pattern weight sequence characterizes the appearance and evolution of typical periodic patterns over time.

[0117] Step S146: Model the pattern weight sequence using a stochastic process model and generate a simulated weight sequence.

[0118] In this embodiment, the step may further include: First, determining the kernel function parameters of the stochastic process model based on the time-series observations of the model weight sequence within a historical time period. Then, constructing a covariance matrix between time points using the kernel function to characterize the correlation strength of the weight values ​​at different times. Next, performing random sampling based on the covariance matrix to generate a simulated weight sequence for the entire year that satisfies the correlation characteristics. Finally, normalizing the generated simulated weight sequence so that the sum of the model weights for each day is one.

[0119] Step S148: Combine the simulated weight sequence with the corresponding typical periodic pattern to reconstruct the annual second component curve.

[0120] In this embodiment, the simulated weight sequence randomly generated in the previous step can be synthesized with the typical periodic pattern to finally construct the periodic component curve (the second component curve of the year) for the whole year.

[0121] In one specific implementation, each typical cycle pattern can be multiplied by the simulated weight of that day, and then all the weighted curves can be summed to obtain the synthesized daily cycle curve for that day.

[0122] The reconstruction to generate the annual second component curve can be represented as stitching together the daily cycle curves synthesized by the above method for each day of the year in strict chronological order, ultimately forming an annual cycle component curve with a length equal to the number of hours throughout the year.

[0123] In some embodiments, the step of performing cluster analysis on the periodic patterns included in the second component to determine at least one typical periodic pattern includes:

[0124] Step S1422: Determine the optimal number of clusters for the cluster analysis based on the silhouette coefficient or analysis of variance.

[0125] In this embodiment, the optimal number of clusters for the cluster analysis can be determined based on the silhouette coefficient.

[0126] Specifically, the silhouette coefficient can be expressed as a metric used to evaluate whether the assignment of each sample point in the clustering results is reasonable. It simultaneously considers the clustering tightness (cohesion) of samples of the same class and the degree of separation between samples of different classes.

[0127] In a specific implementation, for each periodic pattern sample, the average distance (a) between it and all other samples in the same cluster is first calculated to reflect cohesion. Then, the average distance (b) between this sample and all samples in the nearest other cluster is calculated to reflect separation. The silhouette coefficient of this sample is (ba) / max(a, b). The silhouette coefficient can range from [-1, 1], with values ​​closer to 1 indicating better clustering.

[0128] In practice, the executing entity can preset a candidate range for the number of clusters (e.g., from 2 to 10). For each candidate number K, a clustering algorithm (K-Means) is executed, and the average silhouette coefficient of all samples is calculated. Finally, the candidate K value with the highest average silhouette coefficient is determined as the optimal number of clusters.

[0129] In this embodiment, the optimal number of clusters for the cluster analysis can be determined based on analysis of variance.

[0130] Specifically, the analysis of variance can be expressed as determining the optimal value in clustering problems by analyzing the relationship between the within-cluster variance and the number of clusters.

[0131] In a specific implementation, firstly, within the same candidate range of cluster sizes, for each K value, a clustering algorithm is executed and the total sum of within-cluster variances is calculated. Then, a curve showing the relationship between the K value and the total sum of within-cluster variances can be generated. This curve often exhibits a clear inflection point. Before the inflection point, increasing K significantly reduces the within-cluster variance; after the inflection point, the improvement in efficiency from increasing K slows down. Therefore, the K value corresponding to this inflection point can be determined as the optimal number of clusters.

[0132] Step S1424: Based on the optimal number of clusters, divide the periodic patterns included in the second component to obtain each typical periodic pattern and its corresponding cluster center.

[0133] In this embodiment, the K-Means algorithm can be used for partitioning. The executing entity can treat each daily cycle pattern as a data point, using the optimal number K determined in step S1422 as the input parameter, and run the K-Means algorithm. This algorithm can be iteratively optimized to ultimately partition all daily cycle patterns into K mutually exclusive sets (i.e., clusters).

[0134] In this embodiment, the cluster center is the core representative of each cluster. In the K-Means algorithm, the cluster center is defined as the average value of all data points within the cluster. Therefore, each typical periodic pattern is one of the calculated cluster centers, which can be a smooth curve representing the average shape and common characteristics of all daily periodic patterns belonging to that cluster.

[0135] In some implementations, the step of modeling the pattern weight sequence using a stochastic process model and generating a simulated weight sequence includes:

[0136] Step S1462: Determine the kernel function parameters of the stochastic process model based on the time-series observations of the pattern weight sequence within the historical time period.

[0137] In this embodiment, the stochastic process model can be a Gaussian process stochastic process model. The kernel function is a core component of the Gaussian process; it can be a mathematical function used to define the covariance between observations at any two time points in the stochastic process, thereby quantifying the strength of their similarity or correlation. For example, the kernel function can be a squared exponential kernel, which assumes that the closer the points are in time, the stronger the correlation between their function values.

[0138] In this embodiment, determining the kernel function parameters of the stochastic process model based on the time-series observations of the pattern weight sequence within a historical time period can be achieved by using the historical pattern weight sequence as training data input, and for each dimension of the weight sequence (i.e., the weight time series corresponding to each typical periodic pattern), finding a set of optimal hyperparameters for the kernel function through an optimization algorithm. These hyperparameters control the scale of the correlation; for example, they can control the length scale parameter of the correlation decay rate with distance.

[0139] In this embodiment, the maximum likelihood estimation method can be used, that is, to find the hyperparameter that maximizes the probability of the historical weight sequence appearing under the Gaussian process model.

[0140] Step S1464: Construct a covariance matrix between time points using the kernel function to characterize the correlation strength of weight values ​​at different times.

[0141] In this embodiment, a kernel function with determined parameters can be used to construct a mathematical structure describing the correlation between all time points for the entire time span (year) to be simulated. The covariance matrix can be a symmetric positive semi-definite matrix, with its rows and columns corresponding to each time point in the year (e.g., day 1, day 2, ..., day 365). The element in the i-th row and j-th column of the matrix, whose value is calculated by the kernel function determined in step S1462, represents the expected covariance between the pattern weight values ​​of day i and day j.

[0142] In one specific implementation, a vector containing all time points throughout the year is first generated. Then, this vector is used as two inputs to a kernel function, which calculates the covariance value between each pair of time point combinations. Finally, all the calculated covariance values ​​are filled in according to their corresponding positions to form the covariance matrix.

[0143] Step S1466: Random sampling is performed based on the covariance matrix to generate a simulated weight sequence for the whole year that satisfies the correlation characteristics.

[0144] In this embodiment, a weighted sequence for the entire year that conforms to the temporal characteristics can be randomly generated from the constructed Gaussian process model.

[0145] In a specific implementation, the Cholesky decomposition method can be used. Specifically, first, the covariance matrix is ​​decomposed into the product of a lower triangular matrix and its transpose. Then, a vector with the same dimension as the mean vector, consisting of independent random numbers from a standard normal distribution, is generated. Finally, this random vector is multiplied by the decomposed lower triangular matrix, and then added to the mean vector. The resulting vector is a random sample satisfying the specified covariance structure. This random sample can be a continuous and smoothly varying numerical sequence that fully covers every day of the year and maintains the temporal correlation characteristics defined by the kernel function and the covariance matrix. This sequence constitutes the preliminary annual simulated weighted sequence.

[0146] Step S1468: Normalize the generated simulated weight sequence so that the sum of the weights of each mode is one for each day.

[0147] In this implementation, the initially generated simulated weight sequence can be processed day by day. For day t, its original weight vector may include K numerical values. The execution entity can input these K numerical values ​​into a transformation function. This transformation function can be a Softmax function, which can compress a vector of arbitrary real values ​​into a new vector whose sum of elements is 1, and each element is in the range (0, 1), thus which can be interpreted as a probability distribution.

[0148] In this embodiment, the sum of the original weight vectors for that day can also be calculated, and then each element in the vector can be divided by this sum to achieve normalization.

[0149] In some implementations, the step of performing a preset second probability modeling for the third component to generate an annual third component curve includes:

[0150] Step S162: Determine the marginal probability distribution of the third component at multiple time points.

[0151] In this embodiment, the probability distribution law of the residual value at each individual time point (e.g., each hour of a day) can be established by statistical methods.

[0152] In this embodiment, the multiple time points can correspond to various positions within a complete cycle. For example, in the case of a 24-hour daily cycle, it refers to 0:00, 1:00, up to 23:00 every day, a total of 24 time points.

[0153] In this embodiment, determining the marginal probability distribution can be expressed as fitting a one-dimensional probability distribution to describe the occurrence pattern of the residual value at each time point based on historical data.

[0154] In this embodiment, historical residual data can be aligned and grouped according to the stated time points. For each time point (e.g., the h-th hour of each day), all its historical residual values ​​are extracted to form a dataset. Then, probability distribution fitting is performed on this dataset.

[0155] In a specific implementation, a parametric method can be used to fit the probability distribution. It can be assumed that the dataset follows a specific parametric distribution, such as a normal distribution, a t-distribution, or a generalized error distribution. Then, the optimal parameters of this distribution (e.g., mean, variance, degrees of freedom, etc.) are calculated using parameter estimation techniques such as maximum likelihood estimation.

[0156] In a specific implementation, non-parametric methods can also be used for probability distribution fitting. For example, kernel density estimation can be used.

[0157] In this embodiment, for each time point, a definite probability distribution function can be obtained, which is used to describe the random behavior of the residual value at that time.

[0158] Step S164: Based on the determined marginal probability distribution, construct a joint probability distribution model at multiple time points through a correlation coupling function, wherein the correlation coupling function is used to characterize the dependency structure of the residual values ​​at different time points.

[0159] In this embodiment, a comprehensive probabilistic model capable of simultaneously describing the correlation between residual values ​​at all the aforementioned time points can be established. In a specific implementation, firstly, the marginal probability distributions of each time point determined in step S162 are used as input. Then, the marginal distributions are combined with the correlation coupling function to construct the joint probability distribution model. Specifically, the correlation coupling function can be a Copula function, for example, a Gaussian Copula function, which can define the linear dependence between multiple variables through a correlation matrix (which can be a Pearson correlation coefficient matrix). This correlation matrix can be estimated from historical residual data, and each element in the correlation matrix quantifies the correlation strength between residual values ​​at any two different time points (e.g., hour h and hour k). This comprehensive probabilistic model not only carries the individual fluctuation pattern of the residuals at each time point (described by the marginal distribution) but also accurately characterizes the cooperative or dependent relationships between residual fluctuations at different times (described by the Copula function).

[0160] Step S166: Generate a random sequence satisfying the dependency structure from the joint probability distribution model using a random sampling algorithm to form the annual third component curve. In this embodiment, random sampling can first be performed from the joint distribution constructed in step S164. Since this joint distribution is based on Copula, the sampling process can be carried out in two steps: First, a random vector is extracted from the Copula function (Gaussian Copula). The components of this vector are uniformly distributed in the interval [0, 1] and have corresponding correlation structures. Then, using the marginal probability distributions of each time point determined in step S162, each component in the uniform distribution vector obtained by Copula sampling is converted into the actual residual value that conforms to the marginal distribution through the inverse function (quantile function) of the marginal distribution. Next, after completing the above sampling once, a daily residual vector including 24 correlated random numbers can be obtained, which fully represents the joint random fluctuation of the hourly residuals within a day. Finally, the above process is repeated 365 (or 366) times to generate the daily residual vector for the whole year. The daily residual vectors are spliced ​​together strictly in chronological order to form the annual third component curve.

[0161] In some implementations, the step of superimposing the first component, the annual second component curve, and the annual third component curve according to corresponding time points to synthesize an annual load scenario includes:

[0162] Step S1822: Perform time point alignment processing on the first component, the second annual component curve, and the third annual component curve so that the three component values ​​with the same timestamp have a corresponding relationship.

[0163] In this embodiment, since the first component is often obtained from the decomposition of original historical data and is a sequence with the same length as the original sequence, while the second and third annual component curves are newly generated, covering a whole year's worth of sequences, the alignment process first needs to ensure that the three sequences have the same time length and the same time resolution. The executing entity can create a uniform timestamp sequence in memory, covering the entire year to be generated, with a resolution consistent with the resolution of the original data (e.g., 1 hour). Then, the executing entity can map the first component (as a baseline curve), the second annual component curve, and the third annual component curve to this uniform timestamp sequence.

[0164] In this embodiment, if any component's data is missing at a certain point in time, numerical filling algorithms such as linear interpolation and nearest neighbor interpolation can be used.

[0165] Step S1824: The three aligned components are superimposed at each time point to obtain the composite load value at that time point.

[0166] In this embodiment, for each independent time point in the unified time index sequence established in step S1822, the first component value (representing the long-term trend), the second annual component curve value (representing the generated periodic fluctuations), and the third annual component curve value (representing the generated random fluctuations) at that time point can be read respectively, and then these three values ​​are added together. The result of the calculation is to generate a single value, namely the composite load value at that time point, which represents a possible load level at a certain moment under the modeled uncertainty.

[0167] Step S1826: Repeat the overlay process for all time points to generate a complete annual load scenario sequence.

[0168] In this implementation, load scenarios for all times of the year can be constructed through cyclical operations.

[0169] Step S1828: Perform a physical rationality check on the generated annual load scenario to ensure that the load value is within the preset reasonable range.

[0170] In this embodiment, the generated annual load scenario sequence can be post-verified and corrected as necessary to ensure that it conforms to the basic physical laws of the power system.

[0171] In this embodiment, the preset reasonable range can be an absolute boundary. For example, the load value must be greater than zero (it is physically impossible to consume negative power) and should be less than a theoretical upper limit determined by the capacity of the power grid infrastructure (such as transformers and lines).

[0172] In this embodiment, the preset reasonable range can be a relative boundary. For example, the sudden change in load values ​​at adjacent time points should not exceed the maximum allowable rate of change set based on the generator ramp rate and load inertia.

[0173] In this implementation, the executing entity can iterate through the generated annual load scenario sequence and check whether each load value is within the reasonable range. If all values ​​meet the constraints, the scenario passes the verification and can be used as the final output.

[0174] In this embodiment, load values ​​that exceed the range can be directly truncated to their nearest boundary value (e.g., negative values ​​can be set to 0).

[0175] In this embodiment, a smoothing filter algorithm can also be applied to smooth out regions with excessively large abrupt changes.

[0176] This embodiment provides a method for generating annual load scenarios that integrates STL decomposition and WK-GP.

[0177] With the continuous increase in the installed capacity of new energy and the large-scale grid connection of new loads, the power system exhibits strong uncertainty on both the source and load sides. In recent years, a large number of studies have been conducted on the uncertainty of wind and solar power, revealing the multi-timescale output characteristics of wind and solar power from different dimensions. With the acceleration of the process of electricity substitution and the widespread access of new loads, the load side also exhibits increasingly strong time-varying and randomness, and the traditional static load model is gradually becoming ineffective. There are usually three approaches to generating annual load scenarios. (1) Direct selection or clustering method. Select representative monthly load curves from the annual historical data and splice the curves of 12 months to obtain the final typical annual curve. Because they are all historical data, the load curves obtained by this method are difficult to characterize their uncertainty. (2) Random sampling scenario generation method. Establish a probability model that follows the law of load distribution through a large amount of historical data, and generate random scenarios that satisfy the distribution through random sampling. This method relies on the pre-assumed probability distribution, and the load curves often cannot be accurately described by mathematical models, which greatly limits the adaptability of this method. (3) Method of integrating Markov chains and generative adversarial networks. First, a load type transfer chain for the whole year is generated based on a Markov model. For different types of load scenarios, a generator generates random daily load curves. Finally, the load curve for the whole year is formed by splicing them together.

[0178] This implementation plan proposes a method for generating annual load scenarios that integrates STL and WK-GP. First, the annual load curve is decomposed based on STL. Then, based on the characteristics of different components, corresponding annual curves are generated using WK-GP and Copula respectively. Finally, the annual load curve is generated by summing the components moment by moment.

[0179] S01: The STL algorithm is used to decompose the annual load scenario into trend components, periodic components, and residual components. The STL algorithm flow is as follows: Figure 2 As shown.

[0180] S02: The load cycle duration is set to 24, for each position within the cycle. Calculate the average value at this position as the initial value for the periodic component, using the following formula: In the formula This represents the initial value of the periodic component at time i. This represents the raw load data value collected at timestamp t, where i is the location index within the complete cycle, and p is the cycle length of the load data. Indicates position The number of samples on Indicates the number of iterations.

[0181] S03: Remove periodic components from the original data, using the following formula: , This represents the raw load data value collected at timestamp t. To subtract the currently estimated periodic component from the original load data value The intermediate sequence obtained later.

[0182] S04: Using Loess Smoothing filtering. Its core idea is to smooth the filter at each target point. Nearby, a small segment of neighborhood data is taken to construct a local low-order polynomial model, and then the weighted least squares method is used to fit this model to estimate the trend value of that point.

[0183] S05: After obtaining the trend components, the annual load... Subtracting the trend component yields the residual component. .

[0184] S06: Transfer residual components With periodic components After adding, we get .

[0185] S07: Reassemble the noisy periodic components into Each row of the matrix corresponds to a position within a period (e.g., row 0 corresponds to the 0th hour of each day, row 1 corresponds to the 1st hour of each day, and so on). Each column of the matrix represents a complete period (e.g., day 1, day 2, etc.). After applying Loess smoothing to each row individually, the resulting periodic components are reassembled according to the time series to obtain the updated matrix. .

[0186] S08: Will Replacement of step S02 Repeat steps S02 to S07 until the residual components are obtained. The process continues until convergence or the maximum number of iterations is reached.

[0187] S09: Classify the periodic components using the K-Means algorithm, and determine the number of classifications K using the "elbow method".

[0188] S10: For the center curve of each cluster, construct its annual weight time series. .

[0189] S11: Model the weighted time series using a Gaussian random process, defining the weighted time series as... GP represents a Gaussian process. Represents the kernel function.

[0190] S12: Generate K weighted Gaussian random sequences based on a Gaussian random process.

[0191] S13: Combine the sampling weights of each day with the corresponding cluster center curves to reconstruct the annual cycle component curves.

[0192] S14: The joint distribution of the residual components is modeled using Gaussian Copula, and the formula is as follows: In the formula It is the first The marginal distribution of hours, where C is a Gaussian Copula function.

[0193] S15: Generate the annual residual component curves using random sampling.

[0194] S16: Using the trend component of the load as the baseline curve, the generated annual cycle component and annual residual component curves are superimposed on the trend component moment by moment to generate the annual load curve. For example... Figure 3 As shown, the STL algorithm is used to decompose the annual load data into trend components, periodic components, and residual components. The trend component and the annual load curve are shown in the figure. Figure 4It is evident that the load exhibits a relatively pronounced seasonal characteristic, with summer load remaining at a high level throughout the year. Furthermore, the load is significantly affected by holidays, with a noticeable decrease in load levels during holidays lasting more than three days, the largest drop occurring around the Spring Festival. This implementation scheme is compared with traditional methods to verify its effectiveness in generating annual load scenarios. The comparative experiment first uses a K-means clustering algorithm based on the RV coefficient to cluster the daily load curves. Then, a daily state transition chain for the entire year is established using a Markov model Monte Carlo sampling method. Subsequently, daily load curves under different states are generated based on a temporal GAN, and finally, the entire year's load curve is formed by splicing these curves together.

[0195] from Figure 5 As can be seen, the proposed method outperforms the comparative method in terms of the reliability of interval prediction. Regarding distribution fitting accuracy, the Energy Score (ES) is reduced by 60.6% compared to the comparative method, indicating that the generated samples are closer to the true distribution and have higher distribution consistency and diversity control capabilities. In quantile prediction, the Pinball Loss of the proposed method is approximately 0.016 under different confidence intervals, about 50% lower than the comparative method (approximately 0.032), demonstrating higher quantile prediction accuracy. Furthermore, the interval width index decreases by an average of about 50%, resulting in a more compact predicted interval while maintaining coverage, significantly improving the model's effectiveness in expressing uncertainty. The autocorrelation coefficient is used as an indicator to compare the curve shapes of the generated scenarios. On a daily basis, the ACF coefficients for the first 12 hours are statistically analyzed, and the kernel density estimation curves of the ACF coefficients of the curves generated by the proposed method and the comparative method are calculated. The ACF kernel density estimation curve with a lag of 6 is shown below. Figure 6 As shown. Further, the KL divergence of the ACF kernel density estimation curves generated by different methods relative to the real samples was calculated, and the statistical results are shown below. Figure 7 As shown. From Figure 7As can be seen, the proposed method outperforms the comparative methods in kernel density estimation curves at all 11 lag times, compared to the KL divergence of the real samples. The improvement is most significant at time L1, with a reduction of approximately 40% in KL divergence, while the improvement is least significant at time L7, with a reduction of approximately 18% in KL divergence. These results demonstrate that the proposed method is superior to the comparative methods in modeling and characterizing temporal correlations. According to an embodiment of the present invention, an annual load scenario generation device integrating STL decomposition and WK-GP is provided, comprising: an acquisition module for acquiring a load data sequence collected from a target area within a preset time period and at a preset time resolution; a decomposition module for decomposing the load data sequence based on a preset signal decomposition algorithm to obtain a first component, a second component, and a third component representing different physical characteristics respectively; wherein the first component represents the long-term trend of load variation, the second component represents the periodic fluctuation of load, and the third component represents the random fluctuation of load; a first generation module for performing a preset first probability modeling on the second component to generate an annual second component curve; wherein the first probability modeling is represented by clustering of typical periodic patterns and temporal extrapolation of pattern evolution; a second generation module for performing a preset second probability modeling on the third component to generate an annual third component curve, wherein the second probability modeling is represented by random sampling based on the joint probability distribution of the third component at multiple time points; and a synthesis module for superimposing the first component, the annual second component curve, and the annual third component curve according to corresponding time points to synthesize an annual load scenario. According to an embodiment of the present invention, an electronic device is also provided. The electronic device in this embodiment may include one or more of the following components: a processor, a network interface, memory, non-volatile memory, and one or more application programs. The one or more application programs may be stored in the non-volatile memory and configured to be executed by one or more processors. The one or more programs are configured to perform the methods described in the foregoing method embodiments. According to embodiments of the present invention, a computer-readable storage medium is provided, on which a computer program is stored. When executed by a computer, the computer causes the computer to perform the methods described in any of the above embodiments. The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for generating annual load scenarios that integrates STL decomposition and WK-GP, characterized in that, Includes the following steps: Acquire the load data sequence of the target area within a preset time period and at a preset time resolution; The load data sequence is decomposed based on a preset signal decomposition algorithm to obtain a first component, a second component, and a third component that represent different physical characteristics respectively; wherein, the first component is used to represent the long-term variation trend of the load, the second component is used to represent the periodic fluctuation of the load, and the third component is used to represent the random fluctuation of the load. A preset first probability modeling is performed on the second component to generate an annual second component curve; wherein, the first probability modeling is expressed as clustering of typical periodic patterns and temporal extrapolation of pattern evolution; A preset second probability modeling is performed on the third component to generate an annual third component curve; wherein, the second probability modeling is represented by random sampling based on the joint probability distribution of the third component at multiple time points; The first component, the second annual component curve, and the third annual component curve are superimposed according to corresponding time points to synthesize the annual load scenario.

2. The method according to claim 1, characterized in that, The step of decomposing the load data sequence based on a preset signal decomposition algorithm to obtain a first component, a second component, and a third component representing different physical characteristics includes: Based on a preset period length, the load data sequence is initialized with periodic components to obtain an initial second component; wherein, the period length corresponds to the load fluctuation period; The following sub-steps are executed alternately using an iterative smoothing filtering algorithm until the convergence condition is met: the second component is removed from the load data sequence to obtain a deperiodic sequence; local weighted regression smoothing is performed on the deperiodic sequence to estimate the first component; the first component is subtracted from the load data sequence to obtain a temporary residual sequence; the temporary residual sequence is combined with the second component to reconstruct a periodic sequence, and intra-periodic smoothing is performed on the reconstructed sequence to update the second component. After the iteration is completed, the third component is derived from the load data sequence based on the final first and second components.

3. The method according to claim 1, characterized in that, The step of performing a preset first probability modeling on the second component to generate an annual second component curve includes: Cluster analysis is performed on the periodic patterns included in the second component to identify at least one typical periodic pattern; Based on the results of cluster analysis, a pattern weight sequence corresponding to the time series is generated; The pattern weight sequence is modeled using a stochastic process model, and a simulated weight sequence is generated. The simulated weight sequence is combined with the corresponding typical periodic pattern to reconstruct the annual second component curve.

4. The method according to claim 3, characterized in that, The step of performing cluster analysis on the periodic patterns included in the second component to determine at least one typical periodic pattern includes: The optimal number of clusters for the cluster analysis is determined based on the silhouette coefficient or variance analysis. Based on the optimal number of clusters, the periodic patterns included in the second component are divided to obtain each typical periodic pattern and its corresponding cluster center.

5. The method according to claim 3, characterized in that, The step of modeling the pattern weight sequence using a stochastic process model and generating a simulated weight sequence includes: Based on the time-series observations of the pattern weight sequence within the historical time period, the kernel function parameters of the stochastic process model are determined; The covariance matrix between time points is constructed using the kernel function to characterize the correlation strength of the weight values ​​at different times. Random sampling is performed based on the covariance matrix to generate a simulated weight sequence for the whole year that satisfies the correlation characteristics; The generated simulated weight sequence is normalized so that the sum of the weights of each mode is one for each day.

6. The method according to claim 1, characterized in that, The step of performing a preset second probability modeling on the third component to generate an annual third component curve includes: Determine the marginal probability distribution of the third component at multiple time points; Based on the determined marginal probability distribution, a joint probability distribution model at multiple time points is constructed through a correlation coupling function, wherein the correlation coupling function is used to characterize the dependency structure of the residual values ​​at different time points; A random sequence satisfying the dependency structure is generated from the joint probability distribution model using a random sampling algorithm, forming the annual third component curve.

7. The method according to claim 1, characterized in that, The step of superimposing the first component, the second annual component curve, and the third annual component curve according to corresponding time points to synthesize the annual load scenario includes: Time point alignment processing is performed on the first component, the second annual component curve, and the third annual component curve to ensure that the three component values ​​with the same timestamp have a corresponding relationship. The three aligned components are superimposed at each time point to obtain the composite load value at that time point; Repeat the overlay process for all time points to generate a complete annual load scenario sequence; Physical rationality checks are performed on the generated annual load scenarios to ensure that the load values ​​are within a preset reasonable range.

8. A device for generating annual load scenarios that integrates STL decomposition and WK-GP, characterized in that, include: The acquisition module is used to acquire the load data sequence of the target area within a preset time period and at a preset time resolution; The decomposition module is used to decompose the load data sequence based on a preset signal decomposition algorithm to obtain a first component, a second component, and a third component that respectively characterize different physical characteristics; wherein, the first component is used to characterize the long-term variation trend of the load, the second component is used to characterize the periodic fluctuation of the load, and the third component is used to characterize the random fluctuation of the load. A first generation module is used to perform a preset first probability modeling on the second component to generate an annual second component curve; wherein the first probability modeling is expressed as clustering of typical periodic patterns and temporal extrapolation of pattern evolution. The second generation module is used to perform a preset second probability modeling for the third component to generate an annual third component curve, wherein the second probability modeling is represented by random sampling based on the joint probability distribution of the third component at multiple time points; The synthesis module is used to superimpose the first component, the annual second component curve, and the annual third component curve according to the corresponding time points to synthesize the annual load scenario.

9. An electronic device, characterized in that, include: A memory, and one or more processors communicatively connected to the memory; The memory stores instructions that can be executed by the one or more processors to cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1 to 7.

Images