A post-table photovoltaic power decoupling method and device based on unsupervised capacity estimation

Abstract

The application discloses a post-table photovoltaic power decoupling method and device based on unsupervised capacity estimation. The method collects original data of target area user net load time sequence, meteorological data, and observable user paired load and photovoltaic measured data in the same area. After time sequence alignment, data cleaning, feature screening and standardization, the basic data set is generated. Based on the load data in the data set, effective samples are screened, and the basic load is fitted to obtain the user photovoltaic peak capacity. A multi-scale load matrix is constructed and spliced with meteorological features, and an optimized feature matrix is obtained through multi-head self-attention weighting. A hybrid structure decoupling model is built based on the matrix, and the pre-training of the model is completed by using the standardized observable user data. The photovoltaic capacity is adapted to the pre-trained model, the optimized features are input to obtain the standardized photovoltaic power, and the user photovoltaic time sequence decoupling power is output after inverse standardization. The application solves the problem of insufficient feature extraction capability in the prior art in a complex scene.

Description

Technical Field

[0001] This invention relates to the fields of photovoltaic power generation and smart electricity technology, specifically to a post-meter photovoltaic power decoupling method and apparatus based on unsupervised capacity estimation. Background Technology

[0002] With the continuous advancement of the construction of new power systems based on new energy sources, distributed photovoltaic (PV) power generation, with its advantages of low carbon emissions, flexibility, and local consumption, has become a core component of the global energy transition, leading to rapid growth in the installed capacity of behind-the-meter (BTM) PV systems. To support the participation of distributed PV in the electricity market, grids in Europe and the United States generally implement net metering policies, measuring only the net load power of users through a single smart meter, i.e., the coupling result of PV power generation and user load consumption. However, due to limitations in deployment costs and user privacy protection, most BTM PV systems are not equipped with independent PV power generation metering devices. The grid side can only obtain the net load data of users and cannot directly separate the actual PV output from the actual user load. When BTM PV power generation exceeds the actual load demand of users, it poses a severe challenge to the operation, control, and maintenance of the distribution network. For example, unregistered PV systems may experience unexpected reverse feed-in during grid isolation maintenance, which, in scenarios without anti-islanding relays, home energy storage, or other protective devices, poses serious safety hazards to on-site maintenance personnel. Therefore, accurately decoupling BTM photovoltaic power from user load and obtaining actual photovoltaic output data has become a core task for ensuring the safe and stable dispatch of the distribution network and improving the level of refined management and control of distributed photovoltaics.

[0003] Current research on decoupling photovoltaic (BTM) power mainly falls into two categories: physical modeling methods and data-driven methods. Physical modeling methods are based on the photoelectric conversion principle and the law of conservation of energy. They construct a power generation model by integrating solar irradiance, ambient temperature, and inherent parameters of the photovoltaic system. However, these methods require complete physical parameters such as the tilt angle, azimuth angle, and module conversion efficiency of the photovoltaic system, making them extremely impractical in distributed residential photovoltaic scenarios.

[0004] Data-driven methods rely on advanced measurement systems and construct mapping models by mining correlation patterns in historical data. They exhibit significant advantages in nonlinear modeling and adaptability to complex operating conditions and can be further divided into two categories: The first category is methods independent of the region. These methods mainly achieve photovoltaic decoupling by mining the inherent patterns in the net load data of the target user. However, these methods still require obtaining independent metering historical photovoltaic data of the target user during the training phase, which contradicts the core application scenario of BTM photovoltaic without independent metering. Moreover, they are highly sensitive to user load fluctuations and are prone to misinterpreting load changes as changes in photovoltaic power generation, leading to a significant decrease in estimation accuracy. The second category is transfer learning methods based on the region. These methods use observable users with dual metering in the same region as a reference, train the decoupling model, and then transfer it to the target user. However, these methods have an unsolvable core defect: when the photovoltaic installed capacity of the observable user and the target user are inconsistent, direct transfer of the model will result in significant accuracy degradation, and the reliability of cross-user scenario adaptation is extremely poor.

[0005] In summary, current data-driven BTM photovoltaic power decoupling methods still face three key unresolved challenges: First, methods independent of observable users within the region still require acquiring independent historical photovoltaic power generation data of the target user during the training phase, which contradicts the core application scenario of BTM photovoltaics without independent metering. Second, transfer learning methods based on observable users exhibit significant decoupling errors when the photovoltaic installed capacity of the observable user and the target user are inconsistent, making direct transfer unreliable and lacking an effective solution to address the core bottleneck of capacity mismatch. Third, existing decoupling models have insufficient feature extraction capabilities, easily losing key temporal features of photovoltaic output in complex scenarios such as cloudy skies, overcast days, and sudden load changes, resulting in a sharp drop in decoupling accuracy and failing to meet engineering application requirements. These shortcomings severely restrict the improvement of accuracy and large-scale implementation of BTM photovoltaic power decoupling technology, failing to meet the core requirements of new power systems for observable, measurable, and controllable distributed photovoltaics.

[0006] Therefore, there is an urgent need for a post-table photovoltaic power decoupling method based on unsupervised capacity estimation to solve the problems of existing technologies, such as reliance on prior photovoltaic data of target users, model transfer accuracy decay when there is capacity mismatch across users, and insufficient feature extraction capability in complex scenarios. Summary of the Invention

[0007] To address these issues, this invention provides a table-based photovoltaic power decoupling method and apparatus based on unsupervised capacity estimation, which solves the problems in existing technologies such as reliance on prior photovoltaic data of target users, model transfer accuracy decay when there is a capacity mismatch across users, and insufficient feature extraction capability in complex scenarios.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a post-table photovoltaic power decoupling method based on unsupervised capacity estimation, characterized in that it includes:

[0009] By collecting net load time series data of target users within the target area, meteorological monitoring data, and paired net load and photovoltaic measured data of observable users in the same area, multi-source raw data are obtained; the multi-source raw data are then subjected to time axis alignment, outlier removal, missing value imputation, feature filtering, and standardization to obtain a standardized basic dataset.

[0010] Based on the net load data in the standardized basic dataset, the daily net load extreme value feature is extracted, the high irradiance qualified sample is screened, the short-term load fluctuation is corrected and the user's basic load is dynamically fitted, a photovoltaic peak capacity candidate set is constructed and the optimal value is solved, and the photovoltaic peak capacity estimation result of the target user is output.

[0011] Based on the standardized basic dataset, a net load input matrix with multiple sampling rates and multiple time scales is constructed and spliced ​​and fused with meteorological features to obtain fused features; the fused features are then subjected to feature-adaptive weighted enhancement through a feature-level multi-head self-attention mechanism to obtain an optimized feature matrix;

[0012] Based on the optimized feature matrix, an end-to-end decoupling model integrating bidirectional gated recurrent units and two-layer multi-head self-attention is constructed; the peak power standardization processing of the paired net load data of observable users in the same area and the photovoltaic measured power data is performed to obtain training data; the end-to-end decoupling model is pre-trained using the training data to obtain a pre-trained decoupling model;

[0013] Based on the photovoltaic peak capacity estimation results, the pre-trained decoupling model is adaptively adapted to obtain the adapted decoupling model; the optimized feature matrix is ​​input into the adapted decoupling model for decoupling processing, and the standardized photovoltaic power decoupling result is output; after the standardized photovoltaic power decoupling result is inversely standardized, the time-series decoupling result of the photovoltaic power generation of the target user is obtained.

[0014] As a preferred scheme for a post-meter photovoltaic power decoupling method based on unsupervised capacity estimation, the formula for extracting the absolute value of the daily net load minimum is as follows during the processes of extracting extreme features of daily net load, screening qualified samples with high irradiance, correcting short-term load fluctuations, and dynamically fitting user baseline load:

[0015]

[0016] In the formula, For the first The absolute value of the minimum net load per day; For the first sky Net load value at any given time; For the first The time series interval of 1 day; The total number of days for user data;

[0017] The formula for screening qualified samples with high irradiance is:

[0018]

[0019]

[0020]

[0021]

[0022] In the formula, For the first The absolute value of the minimum net load for a sample of sunny days; This represents the total number of sunny day samples. This was the highest global horizontal irradiance of the day. This represents the daily rainfall.

[0023] The formula for correcting short-term load fluctuations is:

[0024]

[0025] In the formula, This is the corrected net load extreme value; This is an estimate of short-term load fluctuations; For the first The set of net load data points 30 minutes before and after the extreme value of each sample; for Net load value at any given time; This represents the average net load data within this time window; This represents the total number of data points within the window.

[0026] The estimated user base load is output through dynamic fitting of the user base load; the expression for the estimated user base load is:

[0027]

[0028]

[0029]

[0030] In the formula, This represents the logarithmic rate of change of the estimated base load values ​​between two consecutive days. These are dynamic weighting coefficients; This is the final output base load estimate; The newly fitted baseline load value; This is the historical baseline load value; The baseline load value for Huber regression estimation.

[0031] As a preferred embodiment of a post-table photovoltaic power decoupling method based on unsupervised capacity estimation, the expression for the photovoltaic peak capacity candidate set is:

[0032]

[0033] In the formula, A candidate set for peak photovoltaic capacity; This represents the total number of extreme net load samples after correction. The total number of samples for the basic load value; This is the corrected net load extreme value; The baseline load value obtained through fitting;

[0034] In the process of solving the photovoltaic peak capacity candidate set, the optimal tangent point of the candidate estimation space is solved by the tangent method, and the photovoltaic peak capacity estimation result is output.

[0035] The formula for calculating the photovoltaic peak capacity estimation result is as follows:

[0036]

[0037] In the formula, This is the final output estimate of the photovoltaic peak capacity; These are candidate values ​​for peak photovoltaic capacity in the candidate set. These are the coordinates of the starting point of the photovoltaic peak estimation curve; These are the coordinates of the curve's endpoint under the theoretical lower limit constraint.

[0038] As a preferred embodiment of a post-table photovoltaic power decoupling method based on unsupervised capacity estimation, in the process of constructing the multi-sampling-rate, multi-time-scale net load input matrix, for the net load data in the standardized basic dataset, a high-frequency net load sequence with a 30-minute sampling interval and a low-frequency net load sequence with a 1-hour sampling interval are constructed respectively to form the multi-sampling-rate, multi-time-scale net load input matrix; the high-frequency net load sequence is used to capture the instantaneous fluctuation characteristics of photovoltaic power generation; the low-frequency net load sequence is used to capture the daily global trend characteristics of photovoltaic power generation.

[0039] The expression for the optimized feature matrix is:

[0040]

[0041]

[0042] In the formula, The input is the fused feature matrix; , , For each attention head, there is a trainable mapping matrix; , , These are the query vector, key vector, and value vector of the i-th attention head, respectively. Let be the attention weight of the i-th attention head; For multi-head self-attention output function; This is the weighted and enhanced feature matrix.

[0043] As a preferred embodiment of a post-table photovoltaic power decoupling method based on unsupervised capacity estimation, the expression for the time-series decoupling result of the photovoltaic power generation of the target user is as follows:

[0044]

[0045]

[0046]

[0047] In the formula, This refers to the time-series decoupling result of the photovoltaic power generation; To standardize the photovoltaic power decoupling results; Learnable weight parameters for fully connected layers; These are the bias parameters for the fully connected layer. This is the temporal feature matrix output by the BiGRU layer; , , This represents the trainable mapping matrix for each head of the temporal attention; , , The first The query vector, key vector, and value vector of each attention head; This is the temporal attention weight matrix; For multi-head self-attention output function; This is the final time series feature matrix after weighted optimization.

[0048] This invention also provides a post-metering photovoltaic power decoupling device based on unsupervised capacity estimation, employing the above-mentioned post-metering photovoltaic power decoupling method based on unsupervised capacity estimation, comprising:

[0049] The raw data acquisition and processing module is used to acquire multi-source raw data by collecting net load time series data of target users in the target area, meteorological monitoring data, and paired net load and photovoltaic measured data of observable users in the same area; and to perform time axis alignment, outlier removal, missing value imputation, feature filtering and standardization processing on the multi-source raw data to obtain a standardized basic dataset.

[0050] The photovoltaic peak capacity estimation result acquisition module is used to extract the extreme features of daily net load, screen qualified samples of high irradiance, correct short-term load fluctuations and dynamically fit the user's basic load based on the net load data in the standardized basic dataset, construct a photovoltaic peak capacity candidate set and solve for the optimal value, and output the photovoltaic peak capacity estimation result of the target user.

[0051] The optimized feature matrix acquisition module is used to construct a multi-sampling rate, multi-timescale net load input matrix based on the standardized basic dataset, and to concatenate and fuse it with meteorological features to obtain fused features; the fused features are then subjected to feature adaptive weighting enhancement through a feature-level multi-head self-attention mechanism to obtain an optimized feature matrix.

[0052] An end-to-end decoupling model construction and training module is used to construct an end-to-end decoupling model that integrates a bidirectional gated recurrent unit and a two-layer multi-head self-attention based on the optimized feature matrix; to perform peak power standardization processing on the paired net load data of observable users in the same area and the measured photovoltaic power data to obtain training data; and to pre-train the end-to-end decoupling model using the training data to obtain a pre-trained decoupling model.

[0053] The photovoltaic power generation time-series decoupling result acquisition module is used to adaptively adapt the parameters of the pre-trained decoupling model based on the photovoltaic peak capacity estimation result to obtain the adapted decoupling model; input the optimized feature matrix into the adapted decoupling model for decoupling processing, and output the standardized photovoltaic power decoupling result; after performing inverse standardization processing on the standardized photovoltaic power decoupling result, the photovoltaic power generation time-series decoupling result of the target user is obtained.

[0054] As a preferred embodiment of a post-meter photovoltaic power decoupling device based on unsupervised capacity estimation, the photovoltaic peak capacity estimation result acquisition module uses the following formula to extract the absolute value of the minimum daily net load during the processes of daily net load extreme value feature extraction, high irradiance qualified sample screening, short-term load fluctuation correction, and dynamic fitting of user base load:

[0055]

[0056] In the formula, For the first The absolute value of the minimum net load per day; For the first sky Net load value at any given time; For the first The time series interval of 1 day; The total number of days for user data;

[0057] The formula for screening qualified samples with high irradiance is:

[0058]

[0059]

[0060]

[0061]

[0062] In the formula, For the first The absolute value of the minimum net load for a sample of sunny days; This represents the total number of sunny day samples. This was the highest global horizontal irradiance of the day. This represents the daily rainfall.

[0063] The formula for correcting short-term load fluctuations is:

[0064]

[0065] In the formula, This is the corrected net load extreme value; This is an estimate of short-term load fluctuations; For the first The set of net load data points 30 minutes before and after the extreme value of each sample; for Net load value at any given time; This represents the average net load data within this time window; This represents the total number of data points within the window.

[0066] The estimated user base load is output through dynamic fitting of the user base load; the expression for the estimated user base load is:

[0067]

[0068]

[0069]

[0070] In the formula, This represents the logarithmic rate of change of the estimated base load values ​​between two consecutive days. These are dynamic weighting coefficients; This is the final output base load estimate; The newly fitted baseline load value; This is the historical baseline load value; The baseline load value for Huber regression estimation.

[0071] As a preferred embodiment of a post-table photovoltaic power decoupling device based on unsupervised capacity estimation, the expression for the photovoltaic peak capacity candidate set in the photovoltaic peak capacity estimation result acquisition module is as follows:

[0072]

[0073] In the formula, A candidate set for peak photovoltaic capacity; This represents the total number of extreme net load samples after correction. The total number of samples for the basic load value; This is the corrected net load extreme value; The baseline load value obtained through fitting;

[0074] In the process of solving the photovoltaic peak capacity candidate set, the optimal tangent point of the candidate estimation space is solved by the tangent method, and the photovoltaic peak capacity estimation result is output.

[0075] The formula for calculating the photovoltaic peak capacity estimation result is as follows:

[0076]

[0077] In the formula, This is the final output estimate of the photovoltaic peak capacity; These are candidate values ​​for peak photovoltaic capacity in the candidate set. These are the coordinates of the starting point of the photovoltaic peak estimation curve; These are the coordinates of the curve's endpoint under the theoretical lower limit constraint.

[0078] As a preferred embodiment of a post-table photovoltaic power decoupling device based on unsupervised capacity estimation, in the optimized feature matrix acquisition module, during the construction of the multi-sampling rate multi-timescale net load input matrix, for the net load data in the standardized basic dataset, a high-frequency net load sequence with a 30-minute sampling interval and a low-frequency net load sequence with a 1-hour sampling interval are constructed respectively to form the multi-sampling rate multi-timescale net load input matrix; the high-frequency net load sequence is used to capture the instantaneous fluctuation characteristics of photovoltaic power generation; the low-frequency net load sequence is used to capture the daily global trend characteristics of photovoltaic power generation;

[0079] The expression for the optimized feature matrix is:

[0080]

[0081]

[0082] In the formula, The input is the fused feature matrix; , , For each attention head, there is a trainable mapping matrix; , , These are the query vector, key vector, and value vector of the i-th attention head, respectively. Let be the attention weight of the i-th attention head; For multi-head self-attention output function; This is the weighted and enhanced feature matrix.

[0083] As a preferred embodiment of a post-table photovoltaic power decoupling device based on unsupervised capacity estimation, the expression for the photovoltaic power generation time-series decoupling result of the target user in the photovoltaic power generation time-series decoupling result acquisition module is as follows:

[0084]

[0085]

[0086]

[0087] In the formula, This refers to the time-series decoupling result of the photovoltaic power generation; To standardize the photovoltaic power decoupling results; Learnable weight parameters for fully connected layers; These are the bias parameters for the fully connected layer. This is the temporal feature matrix output by the BiGRU layer; , , This represents the trainable mapping matrix for each head of the temporal attention; , , The first The query vector, key vector, and value vector of each attention head; This is the temporal attention weight matrix; For multi-head self-attention output function; This is the final time series feature matrix after weighted optimization.

[0088] The present invention has the following advantages:

[0089] First, this invention designs an unsupervised photovoltaic peak capacity estimation process that relies solely on net load data. It can accurately estimate photovoltaic capacity without obtaining prior information such as target user photovoltaic nameplate parameters or independent metering data, providing a reliable capacity benchmark for cross-user model migration and adapting to the large-scale deployment scenario of photovoltaics after household meters.

[0090] Second, this invention constructs a net load input matrix with multiple sampling rates and multiple time scales to simultaneously capture the instantaneous fluctuation characteristics and daily global trend characteristics of photovoltaic power generation. At the same time, through a feature-level multi-head self-attention mechanism, it adaptively strengthens the core features that are strongly correlated with photovoltaic output and suppresses irrelevant load noise interference. It can still maintain stable feature extraction capabilities under complex operating conditions such as cloudy days, overcast days, and sudden load changes, thereby improving the effectiveness of the input features of the decoupled model from the source.

[0091] Third, this invention constructs a fusion architecture of bidirectional gated recurrent unit and two-layer multi-head self-attention: the bidirectional gated recurrent unit comprehensively captures the bidirectional temporal dependencies of the net load sequence, and then the temporal-level multi-head self-attention mechanism quantifies the feature importance of different temporal nodes. Compared with the traditional unidirectional recurrent network, it realizes deep feature extraction of photovoltaic power output mode in coupled data. In scenarios with drastic fluctuations in user load and nonlinear changes in photovoltaic power output, the decoupling accuracy and operational stability are significantly improved.

[0092] Fourth, this invention designs a transfer learning framework of "observable user standardized pre-training - target user capacity adaptive adaptation": first, the model is pre-trained on observable user data through peak power standardization, and then, based on the estimated capacity of the target user, the backbone weights of the model are frozen and only the output layer is incrementally fine-tuned to complete the adaptive adaptation of the model to the target user. This effectively alleviates the accuracy decay caused by capacity mismatch and can still maintain high decoupling accuracy in cross-user and cross-capacity scenarios, greatly improving the generalization ability and engineering feasibility of the model.

[0093] Fifth, this invention deeply embeds the physical rules of photovoltaic output into the entire algorithm process: through high irradiance sample screening, zero power generation at night constraints, and dynamic fitting of user base load, strict physical boundaries are set for the algorithm, which can effectively avoid the risk of model overfitting and avoid outputting invalid results that violate the physical laws of photoelectric conversion. It can still maintain stable output under unconventional scenarios such as extreme weather and photovoltaic module aging, and the reliability and robustness of operation are significantly enhanced.

[0094] Sixth, this invention designs an adaptive dynamic optimization mechanism for algorithm parameters: through methods such as dynamic sliding window, adaptive adjustment of window length driven by the coefficient of variation, and dynamic weighted update of base load, it can adapt online to the long-term changes in user power consumption characteristics and photovoltaic power output characteristics, effectively alleviate the cumulative effect of decoupling error, reduce the frequency of manual parameter calibration, extend the stable operation cycle of the system, and reduce engineering operation and maintenance costs.

[0095] Seventh, the decoupling results of the photovoltaic power output by this invention have the characteristics of no prior dependence, strong generalization and high robustness. They can be directly used as the core data foundation for observable and measurable distributed photovoltaic power in distribution networks, implicit grid connection monitoring, and grid safety and stability scheduling. They can effectively identify the implicit grid connection behavior of unregistered photovoltaic power, reduce the probability of reverse feed-in accidents during grid operation and maintenance, and provide data support for distribution network demand response optimization and local consumption of distributed photovoltaic power. They have significant engineering application value and socio-economic benefits. Attached Figure Description

[0096] To more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are merely exemplary, and those skilled in the art can derive other embodiments based on the provided drawings without creative effort.

[0097] The structures, proportions, sizes, etc. illustrated in this specification are only for the purpose of assisting those skilled in the art in understanding and reading the content disclosed herein, and are not intended to limit the conditions under which the present invention can be implemented. Therefore, they have no substantial technical significance. Any modifications to the structure, changes in the proportions, or adjustments to the size, without affecting the effects and objectives that the present invention can produce, should still fall within the scope of the technical content disclosed in the present invention.

[0098] Figure 1 This is a flowchart illustrating a post-table photovoltaic power decoupling method based on unsupervised capacity estimation provided in Embodiment 1 of the present invention.

[0099] Figure 2 This is a schematic diagram illustrating the specific implementation process of a post-table photovoltaic power decoupling method based on unsupervised capacity estimation provided in Embodiment 1 of the present invention.

[0100] Figure 3 This is a schematic diagram illustrating the components of net load and their interrelationships in one possible embodiment of Embodiment 1 of the present invention.

[0101] Figure 4This is a schematic diagram illustrating the influence of different installed capacities on the amplitude of the photovoltaic output power curve in one possible embodiment of Embodiment 1 of the present invention;

[0102] Figure 5 This is a schematic diagram of Huber regression results in one possible embodiment provided in Embodiment 1 of the present invention;

[0103] Figure 6 This is a schematic diagram illustrating a three-stage growth pattern of the photovoltaic peak estimation curve in one possible embodiment of Embodiment 1 of the present invention;

[0104] Figure 7 This is a schematic diagram of the basic computing unit (GRU) structure of the decoupled model in one possible embodiment of the present invention, provided in Embodiment 1 of the present invention;

[0105] Figure 8 This is a schematic diagram of the overall structure of the MTS-BiGRU-DMHSA decoupling model in one possible embodiment of Embodiment 1 of the present invention;

[0106] Figure 9 This is a schematic diagram illustrating the decoupling effect of the MTS-BiGRU-DMHSA model on a typical target user for one week in a possible embodiment of Embodiment 1 of the present invention.

[0107] Figure 10 This is a schematic diagram of the architecture of a post-table photovoltaic power decoupling device based on unsupervised capacity estimation provided in Embodiment 2 of the present invention. Detailed Implementation

[0108] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0109] Example 1

[0110] See Figure 1 and Figure 2 Embodiment 1 of the present invention provides a post-table photovoltaic power decoupling method based on unsupervised capacity estimation, comprising the following steps:

[0111] S1. By collecting net load time series data of target users in the target area, meteorological monitoring data, and paired net load and photovoltaic measured data of observable users in the same area, multi-source raw data is obtained; the multi-source raw data is subjected to time axis alignment, outlier removal, missing value imputation, feature filtering and standardization processing to obtain a standardized basic dataset;

[0112] S2. Based on the net load data in the standardized basic dataset, extract the extreme features of daily net load, screen qualified samples of high irradiance, correct short-term load fluctuations and dynamically fit the user's basic load, construct a candidate set of photovoltaic peak capacity and solve for the optimal value, and output the photovoltaic peak capacity estimation result of the target user.

[0113] S3. Based on the standardized basic dataset, construct a net load input matrix with multiple sampling rates and multiple time scales, and concatenate and fuse it with meteorological features to obtain fused features; perform feature adaptive weighting enhancement on the fused features through a feature-level multi-head self-attention mechanism to obtain an optimized feature matrix;

[0114] S4. Based on the optimized feature matrix, construct an end-to-end decoupling model that integrates a bidirectional gated recurrent unit and a two-layer multi-head self-attention; standardize the peak power of the paired net load data of observable users in the same area and the measured photovoltaic power data to obtain training data; pre-train the end-to-end decoupling model using the training data to obtain a pre-trained decoupling model.

[0115] S5. Based on the photovoltaic peak capacity estimation result, perform parameter adaptive adaptation on the pre-trained decoupling model to obtain the adapted decoupling model; input the optimized feature matrix into the adapted decoupling model for decoupling processing, and output the standardized photovoltaic power decoupling result; after performing inverse standardization processing on the standardized photovoltaic power decoupling result, obtain the photovoltaic power generation time-series decoupling result of the target user.

[0116] In this embodiment, in step S1, multi-source raw data is obtained by collecting net load time series data of target users in the target area, meteorological monitoring data, and paired net load and photovoltaic measured data of observable users in the same area; the multi-source raw data is then subjected to time axis alignment, outlier removal, missing value imputation, feature filtering and standardization processing to obtain a standardized basic dataset.

[0117] Specifically, the net load time series data of target users in the target area, meteorological monitoring data of the corresponding area, and paired net load time series data and photovoltaic power generation measurement data of observable users with dual metering devices in the same area are collected. Data time axis alignment, outlier removal and missing value filling are completed to build a basic dataset.

[0118] The study employed a local sliding window 3σ criterion to remove outliers from net load and meteorological data, marking them as invalid. Linear interpolation was then used to fill in the missing data corresponding to these invalid values. A two-layer screening mechanism was constructed using Pearson correlation coefficient and symmetric uncertainty to verify the linear and nonlinear correlations of meteorological characteristics, filtering out redundant features weakly correlated with photovoltaic power generation and retaining core meteorological features.

[0119] For the basic dataset, the Min-Max standardization method is used to map all input features to the [0,1] interval, eliminate the difference in units, and construct a standardized basic dataset.

[0120] In this embodiment, in step S2, based on the net load data in the standardized basic dataset, the extreme features of daily net load are extracted, qualified samples with high irradiance are screened, short-term load fluctuations are corrected and the user's basic load is dynamically fitted, a candidate set of photovoltaic peak capacity is constructed and the optimal value is solved, and the photovoltaic peak capacity estimation result of the target user is output.

[0121] Specifically, by extracting the absolute value of the minimum daily net load, a basic feature set of the photovoltaic peak capacity is constructed:

[0122]

[0123] In the formula, For the first The absolute value of the minimum net load per day; For the first sky Net load value at any given time; For the first The time series interval of 1 day; The total number of days for user data.

[0124] The following constraints are used to filter low irradiance data during cloudy and rainy weather, and to select qualified samples with high irradiance:

[0125]

[0126]

[0127]

[0128]

[0129] In the formula, For the first The absolute value of the minimum net load for a sample of sunny days; This represents the total number of sunny day samples. This was the highest global horizontal irradiance of the day. This represents the daily rainfall.

[0130] By performing short-term load fluctuation correction on the extreme values ​​of net load for qualified samples, the interference of short-term load surges on the extreme values ​​can be eliminated.

[0131]

[0132] In the formula, This is the corrected net load extreme value; This is an estimate of short-term load fluctuations; For the first The set of net load data points 30 minutes before and after the extreme value of each sample; for Net load value at any given time; This represents the average net load data within this time window; This represents the total number of data points within the window.

[0133] In this embodiment, during the dynamic fitting process of user base load, the minimum net load at night is extracted to construct the basic sequence for fitting user base load:

[0134]

[0135] In the formula, This represents the minimum net load for the night of day i. For the first The nighttime hours during which photovoltaic power generation is not conducted; The total number of days for user data.

[0136] A dynamic sliding window combined with the Huber regression algorithm is used to fit a stable user base load value:

[0137]

[0138] In the formula, The baseline load value for Huber regression estimation; This is the regression calculation function for Huber; For the first Minimum net load during the day and night; The sliding window length is initially set to 7 days.

[0139] The coefficient of variation of the data within the sliding window is calculated using the following formula to achieve dynamic window adjustment:

[0140]

[0141] In the formula, The coefficient of variation; The standard deviation of the sequence of minimum nighttime net load values ​​within the sliding window; The mean of the sequence of minimum nighttime net load values ​​within the sliding window; when When the value is greater than 0.15, the sliding window length is shortened from 7 days to 3 days.

[0142] By performing a dynamic weighted update of the base load, the final estimated user base load is output:

[0143]

[0144]

[0145]

[0146] In the formula, This represents the logarithmic rate of change of the estimated base load values ​​between two consecutive days. These are dynamic weighting coefficients; This is the final output base load estimate; The newly fitted baseline load value; This is the historical baseline load value; The baseline load value for Huber regression estimation.

[0147] In this embodiment, the corrected net load extreme value and the fitted base load value are used to perform a Cartesian product operation to construct a candidate estimation space for photovoltaic peak capacity:

[0148]

[0149] In the formula, A candidate set for peak photovoltaic capacity; This represents the total number of extreme net load samples after correction. The total number of samples for the basic load value; This is the corrected net load extreme value; The base load value obtained by fitting.

[0150] The theoretical lower limit of photovoltaic peak capacity is:

[0151]

[0152] In the formula, This represents the theoretical lower limit of photovoltaic peak capacity. This is the corrected net load extreme value.

[0153] The optimal tangent point in the candidate estimation space is determined using the tangent method, and the final photovoltaic peak capacity estimation result is output.

[0154]

[0155] In the formula, This is the final output estimate of the photovoltaic peak capacity; These are candidate values ​​for peak photovoltaic capacity in the candidate set. These are the coordinates of the starting point of the photovoltaic peak estimation curve; These are the coordinates of the curve's endpoint under the theoretical lower limit constraint.

[0156] In this embodiment, in step S3, a multi-sampling-rate, multi-time-scale net load input matrix is ​​constructed based on the standardized basic dataset, and spliced ​​and fused with meteorological features to obtain fused features; the fused features are then subjected to feature-adaptive weighted enhancement through a feature-level multi-head self-attention mechanism to obtain an optimized feature matrix.

[0157] Specifically, for the net load data in the standardized basic dataset, a high-frequency net load sequence with a 30-minute sampling interval and a low-frequency net load sequence with a 1-hour sampling interval are constructed to form the multi-sampling rate, multi-time-scale net load input matrix; the high-frequency net load sequence is used to capture the instantaneous fluctuation characteristics of photovoltaic power generation; the low-frequency net load sequence is used to capture the daily global trend characteristics of photovoltaic power generation.

[0158] After completing the multi-feature splicing and fusion of high-frequency net load sequence, low-frequency net load sequence and screened core meteorological features, the optimized feature matrix is ​​output through feature-level multi-head self-attention weighted enhancement.

[0159] The expression for the optimized feature matrix is:

[0160]

[0161]

[0162] In the formula, The input is the fused feature matrix; , , For each attention head, there is a trainable mapping matrix; , , These are the query vector, key vector, and value vector of the i-th attention head, respectively. Let be the attention weight of the i-th attention head; For multi-head self-attention output function; This is the weighted and enhanced feature matrix.

[0163] The formula for calculating the multi-head self-attention output function is as follows:

[0164]

[0165] In the formula, This is the output of the i-th attention head; Specify the vector dimension; To output the transformation matrix; This is a vector concatenation operation.

[0166] In this embodiment, in step S4, based on the optimized feature matrix, an end-to-end decoupling model integrating a bidirectional gated cyclic unit and a two-layer multi-head self-attention is constructed; the peak power standardization processing of the paired net load data of observable users in the same area and the photovoltaic measured power data is performed to obtain training data; the end-to-end decoupling model is pre-trained using the training data to obtain a pre-trained decoupling model.

[0167] Specifically, the basic computational unit of the decoupling model is the gated recurrent unit (GRU), and its core calculation formula is:

[0168]

[0169] In the formula: To update the gate output; To reset the gate output; , , The weight matrix is ​​trainable. Use the Sigmoid activation function; This is the hidden state from the previous moment; Input features for the current time step; This is the candidate hidden state; This represents the final hidden state at the current moment.

[0170] Temporal feature extraction is performed using a Bidirectional Gated Recurrent Unit (BiGRU), and the calculation formula is as follows:

[0171]

[0172] In the formula, Let be the hidden state of the forward GRU at time t; Let be the hidden state of the backward GRU at time t; The input features at time t; , These are the weight coefficients for the bidirectional hidden states; For bias terms; This is the temporal feature matrix of the final output of the BiGRU layer.

[0173] In this embodiment, the core calculation formula of the self-attention mechanism is:

[0174]

[0175] In the formula, This is the query vector, representing the current information to be processed. For the input key vector; The input value vector; Specify the vector dimension.

[0176] A temporal multi-head self-attention layer is set at the end of the BiGRU layer, and the adaptive weighted optimization of temporal features is completed by the following formula:

[0177]

[0178] In the formula, This is the temporal feature matrix output by the BiGRU layer; , , This represents the trainable mapping matrix for each head of the temporal attention; , , The first The query vector, key vector, and value vector of each attention head; This is the temporal attention weight matrix; For multi-head self-attention output function; This is the final time series feature matrix after weighted optimization.

[0179] In this embodiment, during the pre-training process of the end-to-end decoupled model, the Adam optimizer is used, the loss function is set to mean squared error (MSE), the learning rate is set to 0.001, the maximum number of iterations is set to 1000, the batch size is set to 64, and the ratio of training set to validation set is 8:2.

[0180] In this embodiment, in step S5, the pre-trained decoupling model is adaptively adapted based on the photovoltaic peak capacity estimation result to obtain the adapted decoupling model; the optimized feature matrix is ​​input into the adapted decoupling model for decoupling processing, and the standardized photovoltaic power decoupling result is output; after the standardized photovoltaic power decoupling result is inversely standardized, the photovoltaic power generation time-series decoupling result of the target user is obtained.

[0181] Specifically, based on the estimated peak photovoltaic capacity of the target user, the pre-trained decoupled model is fine-tuned. During fine-tuning, the pre-trained weights of the feature-level multi-head self-attention layer, BiGRU layer, and time-series multi-head self-attention layer are frozen, and incremental training is performed only on the weights and bias parameters of the fully connected output layer. The learning rate of incremental training is set to 10% of the pre-training learning rate.

[0182] After adaptation, the standardized multi-timescale net load feature matrix of the target user is input into the adapted decoupled model, and a standardized photovoltaic power generation estimation sequence is output. Using the photovoltaic peak capacity estimated by the target user, the standardized photovoltaic power generation estimation sequence is inversely standardized to output the final real-scale photovoltaic power generation time series result for the target user.

[0183] The expression for the time-series decoupling result of the photovoltaic power generation of the target user is as follows:

[0184]

[0185]

[0186] In the formula, This refers to the time-series decoupling result of the photovoltaic power generation; To standardize the photovoltaic power decoupling results; Learnable weight parameters for fully connected layers; These are the bias parameters for the fully connected layer. This is the final time series feature matrix after weighted optimization.

[0187] In one possible implementation, a verification example is provided as follows:

[0188] This invention applies to off-meter (BTM) distributed photovoltaic (PV) power users without energy storage devices, whose net load satisfies the following physical relationship: Net load power = User's actual electricity load power - Net load of PV power generation. The principle of construction is as follows... Figure 3 As shown.

[0189] This embodiment uses a publicly available dataset released by Australian energy supplier Ausgrid for verification. The dataset covers actual electricity load and photovoltaic power generation records of 300 residential users in Sydney and surrounding areas from July 2010 to June 2013, with a data sampling interval of 30 minutes. To ensure the rigor of the experiment, 30 users with abnormally high capacity were removed, leaving 270 users for testing. These users were divided into an observable user group and a target user group at a 1:2 ratio. The observable user group obtained complete net load and measured photovoltaic data, while the target user group only provided net load data and meteorological parameters. The meteorological data was obtained from the National Solar Radiation Database (NSRDB), and records from the geographic center point of 151°30'E, 33°27'S were selected to characterize the overall weather features of the region.

[0190] To quantify the technical effects of this invention, the following industry-standard evaluation indicators are used:

[0191] Photovoltaic capacity estimation task: Mean Absolute Percentage Error (MAPE), used to quantify the accuracy of capacity estimation, is calculated using the following formula:

[0192]

[0193] In the formula, For the first Estimated photovoltaic capacity for each user; For the first The actual photovoltaic capacity of each user; This represents the total number of users.

[0194] Photovoltaic power decoupling task: Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and Normalized Root Mean Square Error (NRMSE) are used to quantify the accuracy of power decoupling. The calculation formulas are as follows:

[0195]

[0196] In the formula, for Real value of photovoltaic power generation at any given time; for Estimated photovoltaic power generation at any given time; This represents the total number of time steps. , These represent the maximum and minimum values ​​of the photovoltaic power generation sequence, respectively.

[0197] The specific implementation steps of this embodiment are as follows:

[0198] T1. Multi-source data acquisition and preprocessing

[0199] Collect net load time series data of target users in the target area, meteorological monitoring data of the corresponding area, and paired net load time series data and photovoltaic power generation measurement data of observable users with dual metering devices in the same area. Complete data time axis alignment, outlier removal and missing value imputation to build a basic dataset.

[0200] Specifically, outlier removal and missing value imputation: The local sliding window 3σ criterion is used to remove outliers from the net load and meteorological data and mark them as invalid values. Linear interpolation is used to imput the missing data corresponding to the invalid values.

[0201] Two-layer correlation feature screening: A two-layer screening mechanism is constructed using Pearson correlation coefficient and symmetric uncertainty to verify the linear and nonlinear correlation of meteorological features, filter out redundant features that are weakly correlated with photovoltaic power generation, and finally retain 6 core meteorological features: global horizontal irradiance (GHI), direct normal irradiance (DNI), diffuse horizontal irradiance (DHI), solar zenith angle, temperature, and relative humidity.

[0202] Feature standardization: The Min-Max standardization method is used to map all input features to the [0,1] interval to eliminate the difference in units and build a standardized basic dataset.

[0203] T2, Unsupervised peak photovoltaic capacity estimation

[0204] Among the structural parameters of a photovoltaic (PV) system, installed capacity is the core dominant factor affecting the amplitude of the PV output power curve. Changes in azimuth angle primarily cause phase shifts in the power curve, while tilt angle and capacity mainly determine the amplitude modulation of the power curve, with capacity having a stronger impact on peak power. Residential users within the same geographical area generally use near-optimal azimuth and tilt angles; therefore, the differences in the shape of the power curves among users mainly stem from differences in installed capacity. The impact of different installed capacities on the PV output power curve is as follows: Figure 4 As shown. Therefore, accurately estimating the peak photovoltaic capacity of the target user is a core prerequisite for realizing cross-user model transfer. This step relies solely on the net load data of the target user to complete the unsupervised peak photovoltaic capacity estimation.

[0205] Specifically, the absolute value of the daily minimum net load is extracted to construct the basic feature set of photovoltaic peak capacity. Under the constraints of global horizontal irradiance ≥800W / m² and no precipitation, valid samples from sunny days with high irradiance are screened, while low-quality data from cloudy and rainy days are filtered out to construct a feature parameter set for qualified samples. Short-term load fluctuation correction is applied to the 30-minute time window before and after the extreme value point to eliminate the interference of sudden load increases on the extreme value.

[0206] The minimum net load during non-generation periods of nighttime photovoltaic power generation is extracted to construct a basic sequence for fitting user base load. A 7-day dynamic sliding window combined with the Huber regression algorithm is used to fit the stable user base load value. The coefficient of variation (CV) of the data within the sliding window is calculated. When CV > 0.15, the sliding window length is shortened from 7 days to 3 days. The base load fitting result is then optimized through dynamic weighting and updating, and the final estimated user base load value is output. The Huber regression result is as follows: Figure 5 As shown.

[0207] The corrected net load extremum is multiplied by the fitted base load value using a Cartesian product to construct a candidate estimation space for photovoltaic peak capacity. The theoretical lower limit of the photovoltaic peak capacity is determined to be the maximum value of the corrected net load extremum. The optimal tangent point in the candidate estimation space is solved using the tangent method, and the final photovoltaic peak capacity estimation result is output. The photovoltaic peak capacity estimation curve exhibits a three-stage growth pattern, as shown below. Figure 6 As shown.

[0208] This step uses user data from 1 month, 1 quarter, and 1 year to conduct tests. The results show that as the available data duration increases, the capacity estimation error shows a significant downward trend. The median MAPE of annual data can be reduced to 4.8%, the error distribution converges significantly, and the model stability is greatly improved. Test results in different seasons show that the estimation accuracy is higher in summer, spring, and autumn, while the error increases slightly in winter due to weaker sunlight conditions, but the overall accuracy still remains at an engineering usable level.

[0209] T3, Multi-timescale Feature Construction and Enhancement

[0210] Based on the basic dataset, a multi-timescale net load sequence input matrix with multiple sampling rates is constructed. After multi-feature splicing and fusion, feature adaptive weighting enhancement is achieved through a feature-level multi-head self-attention mechanism, and the optimized feature matrix is ​​output.

[0211] Specifically, for the preprocessed net load data, a high-frequency net load sequence with a 30-minute sampling interval and a low-frequency net load sequence with a 1-hour sampling interval are constructed to form a dual-sampling-rate multi-time-scale input matrix. The high-frequency net load sequence is used to capture the instantaneous fluctuation characteristics of photovoltaic power generation, while the low-frequency net load sequence is used to capture the daily global trend characteristics of photovoltaic power generation.

[0212] After completing the multi-feature splicing and fusion of high-frequency net load sequence, low-frequency net load sequence and screened core meteorological features, feature adaptive weighting enhancement is completed through feature-level multi-head self-attention mechanism, which assigns higher weights to high-impact features, suppresses irrelevant noise interference, and outputs the optimized feature matrix.

[0213] Construction and Pre-training of T4, MTS-BiGRU-DMHSA Decoupling Model

[0214] An end-to-end power decoupling model integrating a bidirectional gated recurrent unit (BiGRU) and a two-layer multi-head self-attention DMHSA was constructed. Paired net load data of observable users in the same area and measured photovoltaic power data were used. After peak power standardization, the model was pre-trained and the pre-trained model weights were solidified.

[0215] The basic computational unit of the decoupling model is the gated recurrent unit (GRU), whose structure is as follows: Figure 7 As shown in the diagram, temporal feature extraction is achieved through a bidirectional gated recurrent unit (BiGRU). The BiGRU stacks forward and backward GRU layers to simultaneously capture bidirectional contextual dependencies in the sequence data. Its structure is as follows: Figure 8 As shown.

[0216] The MTS-BiGRU-DMHSA decoupled model constructed in this embodiment adopts a five-layer architecture, consisting of a multi-scale input layer, a feature-level multi-head self-attention layer, a BiGRU layer, a temporal-level multi-head self-attention layer, and a fully connected output layer. The overall structure of the model is as follows: Figure 8 As shown. The model training uses the Adam optimizer, with the loss function set to mean squared error (MSE), the learning rate set to 0.001, the maximum number of iterations set to 1000, the batch size set to 64, and the training set to validation set ratio set to 8:2. After the model pre-training is complete, the backbone network weights are fixed.

[0217] T5, Target User Model Adaptation and Decoupling Output

[0218] Based on the estimated peak photovoltaic power generation capacity of the target user, the parameters of the pre-trained decoupling model are adaptively adapted. The optimized feature matrix of the target user is input into the adapted model, and after inverse standardization, the decoupling result of the photovoltaic power generation time series of the target user is output.

[0219] Specifically, based on the photovoltaic peak capacity estimated by the target user, the pre-trained decoupled model is fine-tuned. During fine-tuning, the pre-trained weights of the feature-level multi-head self-attention layer, BiGRU layer, and time-series multi-head self-attention layer are frozen, and incremental training is performed only on the weights and bias parameters of the fully connected output layer. The learning rate for incremental training is set to 10% of the pre-training learning rate. The standardized multi-timescale net load feature matrix of the target user is input into the adapted decoupled model, and a standardized photovoltaic power generation estimation sequence is output. Using the photovoltaic peak capacity estimated by the target user, the standardized photovoltaic power generation estimation sequence is inversely standardized to output the final real-scale photovoltaic power generation time series result for the target user.

[0220] like Figure 9 As shown, the continuous periodic decoupling effect verification for typical target users indicates that the photovoltaic power curve obtained by decoupling exhibits a highly consistent periodic fluctuation with the true value on a long periodic scale. The midday power peak and waveform characteristics have a good match, the residual has no systematic deviation, and the decoupling result is stable and reliable.

[0221] The application scenarios of this invention are as follows:

[0222] In the scenario of distribution network line loss accounting, this invention relies on user net load and regional meteorological data to achieve unsupervised photovoltaic power disassembly, stripping the distributed photovoltaic power generation component after the table, and assisting in the statistical sorting of distribution network line loss data.

[0223] In the scenario of distribution network load forecasting modeling, this invention separates user electricity load from photovoltaic self-generated and self-consumed power, eliminates photovoltaic power output interference, improves the basic load dataset, and supports the short-term load modeling of distribution networks.

[0224] In the scenario of electricity consumption inspection for industrial and commercial users, this invention decouples the photovoltaic power generation time-series data hidden in the user's net load, and verifies the user's actual electricity consumption based on the decoupling results, thereby assisting in the investigation of abnormal electricity consumption behavior.

[0225] In the scenario of reactive power optimization and control in distribution networks, this invention obtains the actual output timing of distributed photovoltaic systems for individual households, providing a reference for photovoltaic output for the switching configuration of reactive power compensation equipment in distribution networks.

[0226] In the scenario of household distributed photovoltaic survey without the installation of metering devices, the present invention does not require on-site photovoltaic measurement meters, but only relies on electricity consumption and meteorological data to complete the calculation of photovoltaic power for a single household, thereby reducing the data collection cost of scattered household photovoltaic surveys.

[0227] It should be noted that the method of this disclosure embodiment can be executed by a single device, such as a computer or server. The method of this embodiment can also be applied to a distributed scenario, where multiple devices cooperate to complete the task. In such a distributed scenario, one of these devices may execute only one or more steps of the method of this disclosure embodiment, and the multiple devices will interact with each other to complete the method described.

[0228] It should be noted that the above description describes some embodiments of this disclosure. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in a different order than that shown in the above embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0229] Example 2

[0230] See Figure 10 Embodiment 2 of the present invention also provides a post-table photovoltaic power decoupling device based on unsupervised capacity estimation, comprising:

[0231] The raw data acquisition and processing module 001 is used to acquire multi-source raw data by collecting net load time series data of target users in the target area, meteorological monitoring data, and paired net load and photovoltaic measured data of observable users in the same area; and to perform time axis alignment, outlier removal, missing value imputation, feature filtering and standardization processing on the multi-source raw data to obtain a standardized basic dataset.

[0232] The photovoltaic peak capacity estimation result acquisition module 002 is used to extract the extreme features of daily net load, screen qualified samples of high irradiance, correct short-term load fluctuations and dynamically fit the user's basic load based on the net load data in the standardized basic dataset, construct a photovoltaic peak capacity candidate set and solve for the optimal value, and output the photovoltaic peak capacity estimation result of the target user.

[0233] The optimized feature matrix acquisition module 003 is used to construct a multi-sampling rate, multi-time scale net load input matrix based on the standardized basic dataset, and to splice and fuse it with meteorological features to obtain fused features; the fused features are then subjected to feature adaptive weighting enhancement through a feature-level multi-head self-attention mechanism to obtain an optimized feature matrix;

[0234] The end-to-end decoupling model construction and training module 004 is used to construct an end-to-end decoupling model that integrates bidirectional gated recurrent units and two-layer multi-head self-attention based on the optimized feature matrix; to perform peak power standardization processing on the paired net load data of observable users in the same area and the measured photovoltaic power data to obtain training data; and to pre-train the end-to-end decoupling model using the training data to obtain a pre-trained decoupling model.

[0235] The photovoltaic power generation time-series decoupling result acquisition module 005 is used to perform parameter adaptive adaptation on the pre-trained decoupling model based on the photovoltaic peak capacity estimation result to obtain the adapted decoupling model; input the optimized feature matrix into the adapted decoupling model for decoupling processing, and output the standardized photovoltaic power decoupling result; after performing inverse standardization processing on the standardized photovoltaic power decoupling result, obtain the photovoltaic power generation time-series decoupling result of the target user.

[0236] In this embodiment, in the photovoltaic peak capacity estimation result acquisition module 002, during the processes of extracting extreme features of daily net load, screening qualified samples for high irradiance, correcting short-term load fluctuations, and dynamically fitting user basic load, the formula for extracting the absolute value of the minimum daily net load is as follows:

[0237]

[0238] In the formula, For the first The absolute value of the minimum net load per day; For the first sky Net load value at any given time; For the first The time series interval of 1 day; The total number of days for user data;

[0239] The formula for screening qualified samples with high irradiance is:

[0240]

[0241]

[0242]

[0243]

[0244] In the formula, For the first The absolute value of the minimum net load for a sample of sunny days; This represents the total number of sunny day samples. This was the highest global horizontal irradiance of the day. This represents the daily rainfall.

[0245] The formula for correcting short-term load fluctuations is:

[0246]

[0247] In the formula, This is the corrected net load extreme value; This is an estimate of short-term load fluctuations; For the first The set of net load data points 30 minutes before and after the extreme value of each sample; for Net load value at any given time; This represents the average net load data within this time window; This represents the total number of data points within the window.

[0248] The estimated user base load is output through dynamic fitting of the user base load; the expression for the estimated user base load is:

[0249]

[0250]

[0251]

[0252] In the formula, This represents the logarithmic rate of change of the estimated base load values ​​between two consecutive days. These are dynamic weighting coefficients; This is the final output base load estimate; The newly fitted baseline load value; This is the historical baseline load value; The baseline load value for Huber regression estimation.

[0253] In this embodiment, in the photovoltaic peak capacity estimation result acquisition module 002, the expression for the photovoltaic peak capacity candidate set is:

[0254]

[0255] In the formula, A candidate set for peak photovoltaic capacity; This represents the total number of extreme net load samples after correction. The total number of samples for the basic load value; This is the corrected net load extreme value; The baseline load value obtained through fitting;

[0256] In the process of solving the photovoltaic peak capacity candidate set, the optimal tangent point of the candidate estimation space is solved by the tangent method, and the photovoltaic peak capacity estimation result is output.

[0257] The formula for calculating the photovoltaic peak capacity estimation result is as follows:

[0258]

[0259] In the formula, This is the final output estimate of the photovoltaic peak capacity; These are candidate values ​​for peak photovoltaic capacity in the candidate set. These are the coordinates of the starting point of the photovoltaic peak estimation curve; These are the coordinates of the curve's endpoint under the theoretical lower limit constraint.

[0260] In this embodiment, in the optimized feature matrix acquisition module 003, during the construction of the multi-sampling rate multi-time scale net load input matrix, for the net load data in the standardized basic dataset, a high-frequency net load sequence with a 30-minute sampling interval and a low-frequency net load sequence with a 1-hour sampling interval are constructed respectively to form the multi-sampling rate multi-time scale net load input matrix; the high-frequency net load sequence is used to capture the instantaneous fluctuation characteristics of photovoltaic power generation; the low-frequency net load sequence is used to capture the daily global trend characteristics of photovoltaic power generation;

[0261] The expression for the optimized feature matrix is:

[0262]

[0263]

[0264] In the formula, The input is the fused feature matrix; , , For each attention head, there is a trainable mapping matrix; , , These are the query vector, key vector, and value vector of the i-th attention head, respectively. Let be the attention weight of the i-th attention head; For multi-head self-attention output function; This is the weighted and enhanced feature matrix.

[0265] In this embodiment, in the photovoltaic power generation time-series decoupling result acquisition module 005, the expression for the photovoltaic power generation time-series decoupling result of the target user is:

[0266]

[0267]

[0268]

[0269] In the formula, This refers to the time-series decoupling result of the photovoltaic power generation; To standardize the photovoltaic power decoupling results; Learnable weight parameters for fully connected layers; These are the bias parameters for the fully connected layer. This is the temporal feature matrix output by the BiGRU layer; , , This represents the trainable mapping matrix for each head of the temporal attention; , , The first The query vector, key vector, and value vector of each attention head; This is the temporal attention weight matrix; For multi-head self-attention output function; This is the final time series feature matrix after weighted optimization.

[0270] It should be noted that the information interaction and execution process between the modules of the above-mentioned device are based on the same concept as the method embodiment in Embodiment 1 of this application, and the resulting technical effects are the same as those in the method embodiment of this application. For details, please refer to the description in the method embodiment shown above in this application, and it will not be repeated here.

[0271] Example 3

[0272] Embodiment 3 of the present invention provides a non-transitory computer-readable storage medium storing program code for a post-table photovoltaic power decoupling method based on unsupervised capacity estimation. The program code includes instructions for executing the post-table photovoltaic power decoupling method based on unsupervised capacity estimation of Embodiment 1 or any possible implementation thereof.

[0273] Computer-readable storage media can be any available medium that a computer can access, or a data storage device such as a server or data center that integrates one or more available media. The available medium can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0274] Example 4

[0275] Embodiment 4 of the present invention provides an electronic device, including: a memory and a processor;

[0276] The processor and the memory communicate with each other via a bus; the memory stores program instructions that can be executed by the processor, and the processor can call the program instructions to execute a table-based photovoltaic power decoupling method based on unsupervised capacity estimation, as described in Embodiment 1 or any possible implementation thereof.

[0277] Specifically, a processor can be implemented in hardware or software. When implemented in hardware, the processor can be a logic circuit, an integrated circuit, etc. When implemented in software, the processor can be a general-purpose processor that reads software code stored in memory. This memory can be integrated into the processor or located outside the processor and exist independently.

[0278] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable system. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0279] It is obvious to those skilled in the art that the modules or steps of the present invention described above can be implemented using general-purpose computing systems. They can be centralized on a single computing system or distributed across a network of multiple computing systems. Optionally, they can be implemented using program code executable by a computing system, thereby storing them in a storage system for execution by the computing system. In some cases, the steps shown or described can be performed in a different order than those presented herein, or they can be fabricated as separate integrated circuit modules, or multiple modules or steps can be fabricated as a single integrated circuit module. Thus, the present invention is not limited to any particular combination of hardware and software.

[0280] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A post-table photovoltaic power decoupling method based on unsupervised capacity estimation, characterized in that, include: By collecting net load time series data of target users within the target area, meteorological monitoring data, and paired net load and photovoltaic measured data of observable users in the same area, multi-source raw data are obtained; the multi-source raw data are then subjected to time axis alignment, outlier removal, missing value imputation, feature filtering, and standardization to obtain a standardized basic dataset. Based on the net load data in the standardized basic dataset, the daily net load extreme value feature is extracted, the high irradiance qualified sample is screened, the short-term load fluctuation is corrected and the user's basic load is dynamically fitted, a photovoltaic peak capacity candidate set is constructed and the optimal value is solved, and the photovoltaic peak capacity estimation result of the target user is output. Based on the standardized basic dataset, a net load input matrix with multiple sampling rates and time scales is constructed and spliced ​​and fused with meteorological features to obtain the fused features; The fused features are adaptively weighted and enhanced using a feature-level multi-head self-attention mechanism to obtain an optimized feature matrix. Based on the optimized feature matrix, an end-to-end decoupling model integrating bidirectional gated recurrent units and two-layer multi-head self-attention is constructed; the peak power standardization processing of the paired net load data of observable users in the same area and the photovoltaic measured power data is performed to obtain training data; the end-to-end decoupling model is pre-trained using the training data to obtain a pre-trained decoupling model; Based on the photovoltaic peak capacity estimation results, the pre-trained decoupling model is adaptively adapted to obtain the adapted decoupling model; the optimized feature matrix is ​​input into the adapted decoupling model for decoupling processing, and the standardized photovoltaic power decoupling result is output; after the standardized photovoltaic power decoupling result is inversely standardized, the time-series decoupling result of the photovoltaic power generation of the target user is obtained.

2. The method for decoupling photovoltaic power based on unsupervised capacity estimation according to claim 1, characterized in that, In the process of extracting extreme features of daily net load, screening qualified samples for high irradiance, correcting short-term load fluctuations, and dynamically fitting user basic load, the formula for extracting the absolute value of the minimum daily net load is as follows: ； In the formula, For the first The absolute value of the minimum net load per day; For the first sky Net load value at any given time; For the first The time series interval of 1 day; The total number of days for user data; The formula for screening qualified samples with high irradiance is: ； ； ； ； In the formula, For the first The absolute value of the minimum net load for a sample of sunny days; This represents the total number of sunny day samples. This was the highest global horizontal irradiance of the day. This represents the daily rainfall. The formula for correcting short-term load fluctuations is: ； In the formula, This is the corrected net load extreme value; This is an estimate of short-term load fluctuations; For the first The set of net load data points 30 minutes before and after the extreme value of each sample; for Net load value at any given time; This represents the average net load data within this time window; This represents the total number of data points within the window. The estimated user base load is output through dynamic fitting of the user base load; the expression for the estimated user base load is: ； ； ； In the formula, This represents the logarithmic rate of change of the estimated base load values ​​between two consecutive days. These are dynamic weighting coefficients; This is the final output base load estimate; The newly fitted baseline load value; This is the historical baseline load value; The baseline load value for Huber regression estimation.

3. The method for decoupling photovoltaic power based on unsupervised capacity estimation according to claim 2, characterized in that, The expression for the photovoltaic peak capacity candidate set is: ； In the formula, A candidate set for peak photovoltaic capacity; This represents the total number of extreme net load samples after correction. The total number of samples for the basic load value; This is the corrected net load extreme value; The baseline load value obtained through fitting; In the process of solving the photovoltaic peak capacity candidate set, the optimal tangent point of the candidate estimation space is solved by the tangent method, and the photovoltaic peak capacity estimation result is output. The formula for calculating the photovoltaic peak capacity estimation result is as follows: ； In the formula, This is the final output estimate of the photovoltaic peak capacity; These are candidate values ​​for peak photovoltaic capacity in the candidate set. These are the coordinates of the starting point of the photovoltaic peak estimation curve; These are the coordinates of the curve's endpoint under the theoretical lower limit constraint.

4. The method for decoupling photovoltaic power based on unsupervised capacity estimation according to claim 3, characterized in that, In constructing the multi-sampling-rate, multi-time-scale net load input matrix, for the net load data in the standardized basic dataset, a high-frequency net load sequence with a 30-minute sampling interval and a low-frequency net load sequence with a 1-hour sampling interval are constructed respectively to form the multi-sampling-rate, multi-time-scale net load input matrix; the high-frequency net load sequence is used to capture the instantaneous fluctuation characteristics of photovoltaic power generation; the low-frequency net load sequence is used to capture the daily global trend characteristics of photovoltaic power generation. The expression for the optimized feature matrix is: ； ； In the formula, The input is the fused feature matrix; , , For each attention head, there is a trainable mapping matrix; , , These are the query vector, key vector, and value vector of the i-th attention head, respectively. Let be the attention weight of the i-th attention head; For multi-head self-attention output function; This is the weighted and enhanced feature matrix.

5. The method for decoupling photovoltaic power based on unsupervised capacity estimation according to claim 4, characterized in that, The expression for the time-series decoupling result of the photovoltaic power generation of the target user is as follows: ； ； ； In the formula, This refers to the time-series decoupling result of the photovoltaic power generation; To standardize the photovoltaic power decoupling results; Learnable weight parameters for fully connected layers; These are the bias parameters for the fully connected layer. This is the temporal feature matrix output by the BiGRU layer; , , This represents the trainable mapping matrix for each head of the temporal attention; , , The first The query vector, key vector, and value vector of each attention head; This is the temporal attention weight matrix; For multi-head self-attention output function; This is the final time series feature matrix after weighted optimization.

6. A post-metering photovoltaic power decoupling device based on unsupervised capacity estimation, employing the post-metering photovoltaic power decoupling method based on unsupervised capacity estimation as described in any one of claims 1-5, characterized in that, include: The raw data acquisition and processing module is used to acquire multi-source raw data by collecting net load time series data of target users in the target area, meteorological monitoring data, and paired net load and photovoltaic measured data of observable users in the same area; and to perform time axis alignment, outlier removal, missing value imputation, feature filtering and standardization processing on the multi-source raw data to obtain a standardized basic dataset. The photovoltaic peak capacity estimation result acquisition module is used to extract the extreme features of daily net load, screen qualified samples of high irradiance, correct short-term load fluctuations and dynamically fit the user's basic load based on the net load data in the standardized basic dataset, construct a photovoltaic peak capacity candidate set and solve for the optimal value, and output the photovoltaic peak capacity estimation result of the target user. The optimized feature matrix acquisition module is used to construct a multi-sampling rate, multi-time-scale net load input matrix based on the standardized basic dataset, and to splice and fuse it with meteorological features to obtain fused features. The fused features are adaptively weighted and enhanced using a feature-level multi-head self-attention mechanism to obtain an optimized feature matrix. An end-to-end decoupling model construction and training module is used to construct an end-to-end decoupling model that integrates a bidirectional gated recurrent unit and a two-layer multi-head self-attention based on the optimized feature matrix; to perform peak power standardization processing on the paired net load data of observable users in the same area and the measured photovoltaic power data to obtain training data; and to pre-train the end-to-end decoupling model using the training data to obtain a pre-trained decoupling model. The photovoltaic power generation time-series decoupling result acquisition module is used to adaptively adapt the parameters of the pre-trained decoupling model based on the photovoltaic peak capacity estimation result to obtain the adapted decoupling model; input the optimized feature matrix into the adapted decoupling model for decoupling processing, and output the standardized photovoltaic power decoupling result; after performing inverse standardization processing on the standardized photovoltaic power decoupling result, the photovoltaic power generation time-series decoupling result of the target user is obtained.

7. A post-table photovoltaic power decoupling device based on unsupervised capacity estimation according to claim 6, characterized in that, In the photovoltaic peak capacity estimation result acquisition module, during the processes of extracting extreme features of daily net load, screening qualified samples for high irradiance, correcting short-term load fluctuations, and dynamically fitting user basic load, the formula for extracting the absolute value of the minimum daily net load is as follows: ； In the formula, For the first The absolute value of the minimum net load per day; For the first sky Net load value at any given time; For the first The time series interval of 1 day; The total number of days for user data; The formula for screening qualified samples with high irradiance is: ； ； ； ； In the formula, For the first The absolute value of the minimum net load for a sample of sunny days; This represents the total number of sunny day samples. This was the highest global horizontal irradiance of the day. This represents the daily rainfall. The formula for correcting short-term load fluctuations is: ； In the formula, This is the corrected net load extreme value; This is an estimate of short-term load fluctuations; For the first The set of net load data points 30 minutes before and after the extreme value of each sample; for Net load value at any given time; This represents the average net load data within this time window; This represents the total number of data points within the window. The estimated user base load is output through dynamic fitting of the user base load; the expression for the estimated user base load is: ； ； ； In the formula, This represents the logarithmic rate of change of the estimated base load values ​​between two consecutive days. These are dynamic weighting coefficients; This is the final output base load estimate; The newly fitted baseline load value; This is the historical baseline load value; The baseline load value for Huber regression estimation.

8. A post-table photovoltaic power decoupling device based on unsupervised capacity estimation according to claim 7, characterized in that, In the photovoltaic peak capacity estimation result acquisition module, the expression for the photovoltaic peak capacity candidate set is: ； In the formula, A candidate set for peak photovoltaic capacity; This represents the total number of extreme net load samples after correction. The total number of samples for the basic load value; This is the corrected net load extreme value; The baseline load value obtained through fitting; In the process of solving the photovoltaic peak capacity candidate set, the optimal tangent point of the candidate estimation space is solved by the tangent method, and the photovoltaic peak capacity estimation result is output. The formula for calculating the photovoltaic peak capacity estimation result is as follows: ； In the formula, This is the final output estimate of the photovoltaic peak capacity; These are candidate values ​​for peak photovoltaic capacity in the candidate set. These are the coordinates of the starting point of the photovoltaic peak estimation curve; These are the coordinates of the curve's endpoint under the theoretical lower limit constraint.

9. A post-table photovoltaic power decoupling device based on unsupervised capacity estimation according to claim 8, characterized in that, In the optimized feature matrix acquisition module, during the construction of the multi-sampling rate multi-timescale net load input matrix, for the net load data in the standardized basic dataset, a high-frequency net load sequence with a 30-minute sampling interval and a low-frequency net load sequence with a 1-hour sampling interval are constructed respectively to form the multi-sampling rate multi-timescale net load input matrix; the high-frequency net load sequence is used to capture the instantaneous fluctuation characteristics of photovoltaic power generation; the low-frequency net load sequence is used to capture the daily global trend characteristics of photovoltaic power generation. The expression for the optimized feature matrix is: ； ； In the formula, The input is the fused feature matrix; , , For each attention head, there is a trainable mapping matrix; , , These are the query vector, key vector, and value vector of the i-th attention head, respectively. Let be the attention weight of the i-th attention head; For multi-head self-attention output function; This is the weighted and enhanced feature matrix.

10. A post-table photovoltaic power decoupling device based on unsupervised capacity estimation according to claim 9, characterized in that, In the photovoltaic power generation time-series decoupling result acquisition module, the expression for the photovoltaic power generation time-series decoupling result of the target user is: ； ； ； In the formula, This refers to the time-series decoupling result of the photovoltaic power generation; To standardize the photovoltaic power decoupling results; Learnable weight parameters for fully connected layers; These are the bias parameters for the fully connected layer. This is the temporal feature matrix output by the BiGRU layer; , , This represents the trainable mapping matrix for each head of the temporal attention; , , The first The query vector, key vector, and value vector of each attention head; This is the temporal attention weight matrix; For multi-head self-attention output function; This is the final time series feature matrix after weighted optimization.

Images