Unsupervised decomposition method and system for photovoltaic and baseline load

Abstract

This invention discloses an unsupervised decomposition method and system for photovoltaic (PV) and baseline load. The method includes: collecting and processing net load and related exogenous data of a target substation to obtain preprocessing results; creating input features for PV and baseline load based on the preprocessing results to obtain PV features and baseline load features; inputting the PV features and baseline load features into a time-series neural network to obtain PV trend curves and baseline load trend curves; using a linear combination layer to weightedly merge the PV trend curves and baseline load trend curves to reconstruct the net load sequence; and outputting the net load sequence. By implementing this method, it is possible to achieve self-learning without explicit guidance and ensure that the decomposed results possess both identifiability and consistency in physical meaning, thereby better serving the diverse needs of smart grids.

Description

Technical Field

[0001] This invention relates to the field of power system technology, and in particular to a method and system for unsupervised decomposition of photovoltaic and baseline loads. Background Technology

[0002] With the widespread application of distributed photovoltaic (PV) systems in distribution networks and at the user end, actual measurements often only yield net load data, which is the result of the offsetting of user electricity consumption and local PV power generation. In this situation, without separate measurement methods, accurately extracting the PV output curve and baseline load curve from the net load becomes a challenging task. This is fundamental and crucial for applications such as demand response assessment, load forecasting, and optimizing dispatching equipment.

[0003] Current methods can be mainly divided into three categories: The first category is based on physical mechanisms and rules for decomposition. These methods rely heavily on detailed equipment parameters and weather forecast data, making it difficult to adapt to the complex electricity consumption patterns of different users. The second category is purely data-driven supervised learning methods. Although they can provide a certain degree of accuracy, they require a large number of labeled samples as support, which limits their generalization ability. The third category uses frequency domain analysis or statistical modeling. Although they can process time series data to a certain extent, they have limitations in utilizing multiple time-varying features simultaneously.

[0004] Therefore, it is necessary to design a new method that can not only learn on its own without explicit guidance, but also ensure that the decomposed results have identifiability and consistency in physical meaning, so as to better serve the multifaceted needs of the smart grid. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method and system for unsupervised decomposition of photovoltaic and baseline loads.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: an unsupervised decomposition method for photovoltaic and baseline load, comprising:

[0007] Collect and process the net load and related exogenous data of the target substation to obtain preprocessed results;

[0008] Based on the preprocessing results, input features for photovoltaic and baseline loads are created to obtain photovoltaic features and baseline load features;

[0009] The photovoltaic characteristics and the baseline load characteristics are respectively input into a time-series neural network to obtain the photovoltaic trend curve and the baseline load trend curve;

[0010] The photovoltaic trend curve and the baseline load trend curve are weighted and merged using a linear combination layer to reconstruct the net load sequence;

[0011] Output the net load sequence.

[0012] The further technical solution is as follows: the temporal neural network includes a temporal neural network structure with two parallel neural network channels.

[0013] The further technical solution is as follows: the collection and processing of the net load and related exogenous data of the target substation to obtain preprocessing results includes:

[0014] Obtain the net load and related exogenous data of the target substation to obtain initial data;

[0015] The initial data was completed using the piecewise cubic Hermitian interpolation PCHIP method.

[0016] Align the data time intervals, select a uniform sampling period, and perform full-window aggregation and take the arithmetic mean on the initial data window with the required resolution.

[0017] The net load time series was normalized by maximum value, and the exogenous features were normalized by minimum-maximum value to obtain the preprocessing results.

[0018] The further technical solution is as follows: The time-series neural network is obtained by collecting and processing the historical net load and related exogenous data of the target substation, and building input features and training labels of photovoltaic and baseline loads as a sample set to train the time-series neural network structure.

[0019] Its further technical solution is: the photovoltaic feature is represented as follows: Baseline load characteristics are expressed as Training labels are represented as ;in, Represents the temperature at time t; This represents the wind speed at time t; Represents the global horizontal radiation at time t; This represents the horizontal scattered radiation at time t; Represents the direct normal radiation at time t; , , These represent the time feature codes for the day of the week, hour, and minute at time t, including one-hot coding and sincos coding; This indicates the type of holiday at time t, where 1 represents a holiday and 0 represents a non-holiday. This represents the relative humidity at time t; This represents the net load at time t.

[0020] The further technical solution is as follows: the training process of the temporal neural network includes:

[0021] Construct a day / night mask based on irradiation triples, and read the data at each time t. , , Set small tolerance Used to suppress measurement noise, a preset day / night determination signal is defined. When the triplet is not greater than the small tolerance It was determined to be nighttime and ordered When the triplet is not simultaneously greater than the small tolerance The time was determined to be daytime. During the training phase, the input features of photovoltaics are... and Partial multiplication;

[0022] Training samples are constructed using a non-overlapping sliding window method. The window length is set to L. K non-overlapping windows are obtained by sequentially dividing the time series from the starting point, and the tail segments with a length less than L are discarded.

[0023] The further technical solution is as follows: the training process of the temporal neural network also includes:

[0024] Two parallel neural network channels, one for photovoltaic and one for baseline load, were constructed. Each channel was configured with an input projection layer and a time-series coding backbone, and softplus activation was configured at the end of the channel to output a non-negative trend sequence.

[0025] The neural network channel for processing photovoltaic data is connected in series with a day / night mask module after softplus, and the nighttime data is cleared bit by bit according to the preset day / night determination signal.

[0026] A further technical solution is as follows: The photovoltaic characteristics and the baseline load characteristics are respectively input into a time-series neural network to obtain the photovoltaic trend curve and the baseline load trend curve, including:

[0027] The photovoltaic features and the baseline load features are respectively input into the corresponding neural network channels of the time-series neural network. The time-series neural network extracts their respective time-series patterns in parallel and outputs the corresponding trend sequences. At the same time, a normalization scale is determined based on the full sample of the training set, and maximum value normalization is performed on the trend sequences of photovoltaic and baseline load to obtain the photovoltaic trend curve and the baseline load trend curve.

[0028] A further technical solution is as follows: the method of using a linear combination layer to weight and merge the photovoltaic trend curve and the baseline load trend curve to reconstruct the net load sequence includes:

[0029] Perform element-wise weighted summation on the photovoltaic trend curve and the baseline load trend curve at the same time index to reconstruct the net load sequence. The formulas used include ; ; ;in, , This represents the weighting coefficients that are automatically updated during training; , These are the photovoltaic trend curve and the baseline load trend curve, respectively.

[0030] The present invention also provides an unsupervised decomposition system for photovoltaic and baseline loads, comprising:

[0031] The preprocessing unit is used to collect and process the net load and related exogenous data of the target substation to obtain preprocessing results;

[0032] A creation unit is used to create input features for photovoltaic and baseline load based on the preprocessing results, so as to obtain photovoltaic features and baseline load features;

[0033] The input unit is used to input the photovoltaic features and the baseline load features into the time-series neural network to obtain the photovoltaic trend curve and the baseline load trend curve, respectively.

[0034] The reconstruction unit is used to reconstruct the net load sequence by weighting and merging the photovoltaic trend curve and the baseline load trend curve using a linear combination layer.

[0035] Output unit, used to output net load sequence.

[0036] The advantages of this invention compared to existing technologies are as follows: This invention creates input features for photovoltaic (PV) and baseline loads by collecting and processing net load and related exogenous data of the target substation. These features are then input into a specially designed time-series neural network to learn the time-series patterns of PV output and baseline load. Subsequently, a linear combination layer is used to weightedly merge the trend curves of both to reconstruct the net load sequence, achieving self-learning capability without explicit guidance. A key feature of this method is the use of a softplus non-negative activation function to ensure physical plausibility and a nighttime masking mechanism to guarantee zero PV output at night. This ensures that the decomposition results not only possess high discriminative power but also maintain physical consistency. Ultimately, this method can effectively serve multiple aspects of smart grids, including demand response assessment, load forecasting, and operational analysis, improving the flexibility and efficiency of the power system.

[0037] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0038] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] Figure 1 This is a flowchart illustrating the unsupervised decomposition method for photovoltaic and baseline load provided in an embodiment of the present invention.

[0040] Figure 2 This is a schematic diagram of the processing flow of a temporal neural network provided in an embodiment of the present invention;

[0041] Figure 3 A schematic diagram of a trend curve provided for an embodiment of the present invention;

[0042] Figure 4 This is a training convergence result diagram provided in an embodiment of the present invention;

[0043] Figure 5 A schematic diagram of the net load sequence of the output provided in an embodiment of the present invention;

[0044] Figure 6 A schematic block diagram of an unsupervised decomposition system for photovoltaic and baseline loads provided in an embodiment of the present invention;

[0045] Figure 7 A schematic block diagram of a computer device provided for an embodiment of the present invention. Detailed Implementation

[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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.

[0047] It should be understood that, when used in this specification and the appended claims, the terms "comprising" and "including" indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0048] It should also be understood that the terminology used in this specification is for the purpose of describing particular embodiments only and is not intended to limit the invention. As used in this specification and the appended claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise.

[0049] It should also be further understood that the term "and / or" as used in this specification and the appended claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0050] Please see Figure 1 , Figure 1 This is a flowchart illustrating the unsupervised decomposition method for photovoltaic and baseline load provided in an embodiment of the present invention. This unsupervised decomposition method for photovoltaic and baseline load is applied in a server. Please refer to... Figure 2 This method collects and processes substation net load and related exogenous data to create input features targeting photovoltaic (PV) and baseline load characteristics. A time-series neural network model with two parallel neural network channels is then used to predict PV trend curves and baseline load trend curves, respectively. The method employs a softplus activation function to ensure output non-negativity and uses day / night masking techniques to enhance the understanding of PV characteristics, effectively suppressing the impact of measurement noise and improving the accuracy of nighttime data processing. Furthermore, the original net load sequence is reconstructed by weighting and merging the trend curves of PV and baseline loads through a linear combination layer, ensuring the discriminability and consistency of the decomposition results' physical meaning. The entire process requires no explicit guidance, achieving self-learning and optimization, thus more accurately serving the needs of load forecasting, energy management, and system planning in smart grids.

[0051] Figure 1 This is a schematic flowchart of the unsupervised decomposition method for photovoltaic and baseline load provided in an embodiment of the present invention. Figure 1 As shown, the method includes the following steps S110 to S150.

[0052] S110. Collect and process the net load and related exogenous data of the target substation to obtain preprocessing results.

[0053] In this embodiment, the preprocessing result refers to the dataset obtained after cleaning, correcting and transforming the initial data, which is used for further analysis or as input to the algorithm.

[0054] In one embodiment, step S110 described above may include steps S111 to S114.

[0055] S111. Obtain the net load and related exogenous data of the target substation to obtain initial data.

[0056] In this embodiment, initial data refers to the unprocessed set of information directly obtained from sensors, monitoring devices, or other data sources.

[0057] This step involves collecting time-series data on the net load (i.e., total power consumption minus any locally generated electricity, such as that from solar power) of the target substation. In addition, a range of load-related exogenous data needs to be collected, including but not limited to meteorological data such as temperature, wind speed, global horizontal radiation, horizontal diffuse radiation, and direct normal radiation, as well as time-specific encodings (e.g., day of the week, hour, minute), holiday types, and relative humidity. This data collectively forms the basis of the model input, helping to identify and separate solar power output from the baseline load.

[0058] S112. Use the piecewise cubic Hermitian interpolation PCHIP method to complete the initial data.

[0059] In practice, due to various reasons (such as equipment failure, maintenance, etc.), the original data may contain missing or outlier values. To ensure data integrity and consistency, the Piecewise Cubic Hermite Interpolating Polynomial (PCHIP) method is used to fill in these missing data points. This method can maintain the trend and shape of the data and avoid inaccurate fluctuations that may be introduced by simple interpolation.

[0060] S113. Align the data time interval, select a uniform sampling period, perform full window aggregation and take the arithmetic mean on the initial data window with the required resolution.

[0061] This step aims to ensure that all data are aligned over the same time intervals for consistency in subsequent analyses. If the original sampling rate of the data is higher than the required analytical resolution, the resolution is reduced by full-window aggregation (e.g., every N consecutive samples as a window) and calculating the arithmetic mean of the data within that window. This not only helps reduce the amount of data but also eliminates high-frequency noise and improves data quality.

[0062] S114. The net load time series is normalized by maximum value, and the exogenous features are normalized by minimum-maximum value to obtain the preprocessing results.

[0063] In this embodiment, the formula for normalizing the maximum value is as follows: In the formula, Indicates the original sequence number 1 i One value, Indicates the normalized sequence number of the first... i One value, This represents the maximum value in the entire sequence, and N=1488 represents the sequence length; the min-max normalization formula is as follows: In the formula, This represents the minimum value of the entire sequence.

[0064] Normalization is a crucial step in data preprocessing. It transforms data from different scales to a common scale, facilitating effective training of subsequent machine learning models. For net load time series, maximum normalization is used to ensure all values ​​fall within the [0, 1] interval. For exogenous features, minimum-maximum normalization is employed, similarly mapping to the [0, 1] interval. This approach ensures comparability between different features and helps accelerate model convergence.

[0065] By following the above four steps, we can obtain carefully processed preprocessed results, which lay a solid foundation for the subsequent construction of input features of photovoltaic and baseline load and training of time-series neural networks.

[0066] S120. Based on the preprocessing results, input features for photovoltaic and baseline load are created to obtain photovoltaic features and baseline load features.

[0067] In this embodiment, the photovoltaic feature is represented as: Baseline load characteristics are expressed as ;in, Represents the temperature at time t; This represents the wind speed at time t; Represents the global horizontal radiation at time t; This represents the horizontal scattered radiation at time t; Represents the direct normal radiation at time t; , , These represent the time feature codes for the day of the week, hour, and minute at time t, including one-hot coding and sincos coding; This indicates the type of holiday at time t, where 1 represents a holiday and 0 represents a non-holiday. This represents the relative humidity at time t. All the above data have been normalized.

[0068] Photovoltaic Features: This feature set includes multiple variables directly related to photovoltaic power generation, such as temperature, wind speed, global horizontal radiation, horizontal diffuse radiation, direct normal radiation, and other time-related features (time codes for days of the week, hours, and minutes), as well as holiday types. These variables are crucial for predicting photovoltaic output because they directly affect the energy conversion efficiency of photovoltaic panels. For example, high radiation intensity generally means higher power generation, while high temperatures may reduce the efficiency of photovoltaic cells. Furthermore, time feature encoding helps the model capture daily and weekly periodic patterns of change.

[0069] Baseline load characteristics: In contrast, baseline load characteristics primarily consider factors that influence electricity consumption, such as temperature, wind speed, relative humidity, and the same time-related encoding (weekday, hour, minute), including holiday types. These factors help the model understand load demand patterns over different time periods, such as differences in electricity consumption between weekdays and weekends, and between day and night.

[0070] All the features mentioned above have been normalized to ensure that the data are compared on the same scale, which helps improve the stability and efficiency of model training. In particular, net load size is also included, although it is not directly part of the photovoltaic or baseline load features, it plays a key role in the construction of label Y, namely, reconstructing the net load by combining the trend sequences of photovoltaic and baseline loads.

[0071] Furthermore, a day / night masking mechanism based on irradiation triples was introduced when constructing photovoltaic features. This mechanism suppresses measurement noise by setting a small tolerance value and determines whether it is day or night based on whether the triples are simultaneously below this threshold. If it is night, the corresponding photovoltaic features are cleared, thus avoiding invalid features at night from interfering with the model's learning process, thereby improving the model's accuracy and physical consistency.

[0072] In summary, step S120 involves not only extracting relevant features from the raw data, but also performing appropriate preprocessing and transformation on these features to facilitate effective learning by the subsequent temporal neural network. This process is one of the key steps in achieving unsupervised decomposition of photovoltaic and baseline load.

[0073] S130. The photovoltaic characteristics and the baseline load characteristics are respectively input into the time-series neural network to obtain the photovoltaic trend curve and the baseline load trend curve.

[0074] In this embodiment, as Figure 3 As shown, Figure 3 The photovoltaic power output trend curve was captured over a period of time. and baseline load trend curve .

[0075] A photovoltaic trend curve refers to the trend curve of photovoltaic power generation over time, predicted by a trained time-series neural network model based on input photovoltaic characteristics. This curve reflects the power generation capacity of the photovoltaic system in different time periods.

[0076] Baseline load trend curve: This refers to the trend curve of basic electricity load demand over time, predicted by a trained time-series neural network model based on the input baseline load characteristics. This curve reflects the basic electricity demand within the substation area without considering the integration of renewable energy sources (such as photovoltaics).

[0077] Specifically, the photovoltaic features and the baseline load features are respectively input into the corresponding neural network channels of the time-series neural network. The time-series neural network extracts their respective time-series patterns in parallel and outputs the corresponding trend sequences. At the same time, a normalization scale is determined based on the full sample of the training set, and maximum value normalization is performed on the trend sequences of photovoltaic and baseline load to obtain the photovoltaic trend curve and the baseline load trend curve.

[0078] The temporal neural network comprises a structure with two parallel neural network channels. One channel is dedicated to processing photovoltaic characteristics, while the other is used to process baseline load characteristics.

[0079] The time-series neural network is obtained by collecting and processing historical net load and related exogenous data of the target substation, and constructing input features and training labels of photovoltaic and baseline loads as a sample set to train the time-series neural network structure. The training labels are represented as... ; This represents the net load at time t.

[0080] The training process of the temporal neural network includes:

[0081] Construct a day / night mask based on irradiation triples, and read the data at each time t. , , Set small tolerance Used to suppress measurement noise, a preset day / night determination signal is defined. When the triplet is not greater than the small tolerance It was determined to be nighttime and ordered When the triplet is not simultaneously greater than the small tolerance The time was determined to be daytime. During the training phase, the input features of photovoltaics are... and Partial multiplication; avoid interference from invalid features at night on the model;

[0082] Training samples are constructed using a non-overlapping sliding window method. The window length is set to L. K non-overlapping windows are obtained by sequentially dividing the time series from the starting point, and the tail segments with a length less than L are discarded.

[0083] like Figure 4 As shown, the training process of the temporal neural network further includes:

[0084] Two parallel neural network channels, one for photovoltaic and one for baseline load, were constructed. Each channel was configured with an input projection layer and a time-series coding backbone, and softplus activation was configured at the end of the channel to output a non-negative trend sequence.

[0085] The neural network channel for processing photovoltaic data is connected in series with a day / night mask module after softplus, and the nighttime data is cleared bit by bit according to the preset day / night determination signal.

[0086] In this embodiment, the input projection layer and the time-transformer are used: each channel first transforms the input features through the input projection layer to better capture latent patterns in the data. Subsequently, the time-transformer is used to further extract the temporal dependencies and dynamic characteristics of the input sequence.

[0087] Activation function selection: A softplus activation function is configured at the end of each channel to ensure that the output trend sequence is non-negative, which meets the requirements of photovoltaic output and baseline load in actual physical scenarios.

[0088] Application of the day / night masking module: Specifically, in the photovoltaic channel, a day / night masking module is introduced after softplus is activated. This module, based on a preset day / night determination signal, will clear the corresponding data bit by bit when it is determined to be nighttime, thereby avoiding invalid nighttime features from interfering with the model's learning process and improving the model's accuracy and physical consistency.

[0089] Normalization: Based on the maximum value determined from the entire training set samples, the trend sequences of photovoltaic and baseline loads are normalized to the maximum value, so that the output trend curves are on the same scale, which facilitates subsequent analysis and comparison.

[0090] Training process: During training, a non-overlapping sliding window method is used to construct training samples. The window length is set to L, and K non-overlapping windows are obtained sequentially from the starting point of the time series. The tail segments with a length of less than L are discarded.

[0091] The difference between the reconstructed payload and the actual payload label is used as a monitoring signal. End-to-end parameter updates are performed by minimizing the reconstruction loss until a preset convergence condition is met.

[0092] In summary, the core of step S130 lies in utilizing a carefully designed dual-channel time-series neural network structure to effectively extract the respective time-series patterns from photovoltaic features and baseline load features, and generate corresponding photovoltaic output trend curves and baseline load trend curves. This method not only improves the accuracy of the decomposition results but also enhances the interpretability and application value of the model.

[0093] Specifically, a regression training method is adopted, using the reconstructed net load obtained from the linear combination layer of the model. Compared with the actual net load label The difference is used as the sole monitoring signal for end-to-end parameter updates, minimizing the reconstruction loss. This drives both photovoltaic and baseline loads to learn trend representations consistent with their physical drivers, thus reducing reconstruction loss. The definition is as follows: .

[0094] The training process is considered convergent and updates are stopped. The convergence criterion is as follows: ;

[0095] in Here, 'e' represents the preset threshold, and 'e' represents the current training round.

[0096] By performing forward inference with fixed model parameters, the photovoltaic power output is directly output. and baseline load As a result.

[0097] like Figure 4 As shown, the model only accepts net load data as labels during training. However, the net load data in this embodiment is generated through simulation. Therefore, in the result analysis, the photovoltaic output of the model can be compared. and baseline load and actual photovoltaic output and baseline load ,like Figure 5 As shown.

[0098] S140. The photovoltaic trend curve and the baseline load trend curve are weighted and merged using a linear combination layer to reconstruct the net load sequence.

[0099] In this embodiment, the net load sequence refers to the sequence that reflects the change of the actual electricity demand of the substation over time, which is reconstructed by weighting the photovoltaic trend curve and the baseline load trend curve by elements.

[0100] Perform element-wise weighted summation on the photovoltaic trend curve and the baseline load trend curve at the same time index to reconstruct the net load sequence. The formulas used include ; ; ;in, , This represents the weighting coefficients that are automatically updated during training; , These are the photovoltaic trend curve and the baseline load trend curve, respectively.

[0101] This process not only relies on the trend curves obtained during the model training phase, but also requires precise weighted summation of these trend curves based on specific weight coefficients.

[0102] During the training phase, the values ​​of are optimized by minimizing the reconstruction loss function. The reconstruction loss is defined as the difference between the reconstructed net load output by the model and the actual net load label.

[0103] To ensure that different trend curves are compared and weighted on the same scale, maximum normalization is typically performed on the photovoltaic trend curve and the baseline load trend curve. This step helps improve the stability and accuracy of the model.

[0104] S150, Output Net Load Sequence.

[0105] After completing step S140, the reconstructed net load sequence is obtained. ,include , and The next step is to output this result as the final output of the entire decomposition method. Specifically:

[0106] Output format: Net load sequence It can be directly used as the output of the model for subsequent application scenarios, such as demand response assessment, load forecasting, or scenario generation.

[0107] Validation and Analysis: In experiments or practical applications, the effectiveness of the model can be validated by comparing the net load sequence output by the model with the actual data. Especially in simulation environments, since the net load data is pre-generated, the model output can be directly compared with the "true" values ​​to evaluate model performance.

[0108] Application scenarios: The obtained net load sequence is not only an effective way to decompose the original data, but also provides strong support for understanding the power consumption patterns in the substation area, which is of great significance for the integration and management of renewable energy (such as photovoltaic).

[0109] In summary, step S140 achieves accurate merging of the photovoltaic trend curve and the baseline load trend curve through the linear combination layer, while S150 is responsible for outputting this merged result in the form of a net load sequence, providing a solid data foundation for various power system-related analyses and decisions.

[0110] The key feature of this method is that it achieves effective decomposition solely based on substation-level net load and related exogenous data, without relying on specific equipment-level data or component parameters. First, net load and exogenous data such as meteorological, irradiance, and calendar data are acquired and preprocessed. Then, photovoltaic output-related and baseline load-related input features are constructed on the preprocessed results. A time-series neural network is established, with parallel extraction of time-series patterns for both photovoltaic and baseline load channels, and softplus non-negative activation is applied at the end. The photovoltaic channel uses a nighttime mask determined by irradiance triples to force zero activation at night. A scale is determined based on the full training set, and amplitude normalization is applied to the outputs of both channels. The net load is reconstructed by element-weighted superposition on a unified time index through a linear combination layer. Without actual decomposition annotations, regression training is performed using the mean square error between the reconstructed and actual net loads as the sole loss. After training convergence, the decomposed photovoltaic output trend and baseline load trend are output. This method requires no equipment-level annotations or component parameters and can obtain stable and interpretable results under conditions of aggregated heterogeneity and topological perturbations, making it suitable for applications such as demand response assessment, load forecasting, and operational analysis.

[0111] The method in this embodiment can obtain an interpretable decomposition of photovoltaic and baseline loads by relying solely on substation-level net load and multi-source exogenous features, even without real decomposition annotation. Mechanistic decoupling is achieved through feature physical separation and temporal neural networks. Softplus non-negative activation is introduced at the channel end, and a day / night mask based on irradiance determination is applied to the photovoltaic system to satisfy nighttime zero boundaries and physical consistency. Full-sample normalization is used to establish a comparable scale, and end-to-end optimization of linear combination reconstruction and reconstruction loss is combined to improve the identifiability and stability of the decomposition. The method does not rely on equipment-level measurements and component parameters, and has cross-seasonal / cross-station migration capabilities and engineering deployability. The output photovoltaic output and baseline load can directly support applications such as demand response assessment, load forecasting, and scenario generation.

[0112] The aforementioned unsupervised decomposition method for photovoltaic (PV) and baseline loads creates input features for PV and baseline loads by collecting and processing net load data and related exogenous data from the target substation. These features are then fed into a specially designed time-series neural network to learn the time-series patterns of PV output and baseline load. Subsequently, a linear combination layer is used to weightedly merge the trend curves of both to reconstruct the net load sequence, achieving self-learning capability without explicit guidance. A key feature of this method is the use of a softplus non-negative activation function to ensure physical plausibility and a nighttime masking mechanism to guarantee zero PV output at night. This ensures that the decomposition results are not only highly discriminative but also maintain physical consistency. Ultimately, this method can effectively serve multiple aspects of smart grids, including demand response assessment, load forecasting, and operational analysis, improving the flexibility and efficiency of the power system.

[0113] Figure 6 This is a schematic block diagram of an unsupervised decomposition system 300 for photovoltaic and baseline load provided in an embodiment of the present invention. Figure 6 As shown, corresponding to the above-described unsupervised decomposition method of photovoltaic and baseline load, the present invention also provides a photovoltaic and baseline load unsupervised decomposition system 300. This photovoltaic and baseline load unsupervised decomposition system 300 includes a unit for executing the above-described photovoltaic and baseline load unsupervised decomposition method, and the system can be configured in a server. Specifically, please refer to... Figure 6 The unsupervised decomposition system 300 for photovoltaic and baseline load includes a preprocessing unit 301, a creation unit 302, an input unit 303, a reconstruction unit 304, and an output unit 305.

[0114] The preprocessing unit 301 is used to collect and process the net load and related exogenous data of the target substation to obtain preprocessing results; the creation unit 302 is used to create input features of photovoltaic and baseline load based on the preprocessing results to obtain photovoltaic features and baseline load features; the input unit 303 is used to input the photovoltaic features and the baseline load features into a time-series neural network to obtain photovoltaic trend curves and baseline load trend curves; the reconstruction unit 304 is used to use a linear combination layer to weighted merge the photovoltaic trend curves and baseline load trend curves to reconstruct the net load sequence; and the output unit 305 is used to output the net load sequence.

[0115] In one embodiment, the preprocessing unit 301 includes:

[0116] The data acquisition subunit is used to acquire the net load and related exogenous data of the target substation to obtain initial data; the completion subunit is used to complete the initial data using the piecewise cubic Hermitian interpolation PCHIP method; the averaging subunit is used to align the data time intervals, select a uniform sampling period, perform full-window aggregation on the initial data window with the required resolution load, and take the arithmetic mean; the normalization subunit is used to normalize the net load time series with the maximum value and to normalize the exogenous features with the minimum-maximum value to obtain the preprocessed results.

[0117] In one embodiment, the input unit 303 is used to input the photovoltaic features and the baseline load features into the corresponding neural network channels of the time-series neural network. The time-series neural network extracts their respective time-series patterns in parallel and outputs the corresponding trend sequences. At the same time, a normalization scale is determined based on the full sample of the training set, and maximum value normalization is performed on the trend sequences of photovoltaic and baseline load to obtain the photovoltaic trend curve and the baseline load trend curve.

[0118] In one embodiment, the merging unit is configured to perform element-wise weighted summation on the photovoltaic trend curve and the baseline load trend curve at the same time index to reconstruct the net load sequence. The formulas used include ; ; ;in, , This represents the weighting coefficients that are automatically updated during training; , These are the photovoltaic trend curve and the baseline load trend curve, respectively.

[0119] It should be noted that those skilled in the art can clearly understand that the specific implementation process of the above-mentioned unsupervised decomposition system 300 for photovoltaic and baseline load and each unit can be referred to the corresponding description in the foregoing method embodiments. For the sake of convenience and brevity, it will not be repeated here.

[0120] The aforementioned unsupervised decomposition system 300 for photovoltaic and baseline load can be implemented as a computer program, which can, for example... Figure 7 It runs on the computer device shown.

[0121] Please see Figure 7 , Figure 7 This is a schematic block diagram of a computer device provided in an embodiment of this application. The computer device 500 can be a server, wherein the server can be a standalone server or a server cluster composed of multiple servers.

[0122] See Figure 7 The computer device 500 includes a processor 502, a memory, and a network interface 505 connected via a system bus 501. The memory may include a non-volatile storage medium 503 and internal memory 504.

[0123] The non-volatile storage medium 503 may store an operating system 5031 and a computer program 5032. The computer program 5032 includes program instructions that, when executed, cause the processor 502 to perform an unsupervised decomposition method of photovoltaic and baseline load.

[0124] The processor 502 provides computing and control capabilities to support the operation of the entire computer device 500.

[0125] The internal memory 504 provides an environment for the execution of the computer program 5032 in the non-volatile storage medium 503. When the computer program 5032 is executed by the processor 502, the processor 502 can execute a photovoltaic and baseline load unsupervised decomposition method.

[0126] This network interface 505 is used for network communication with other devices. Those skilled in the art will understand that... Figure 7 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device 500 to which the present application is applied. The specific computer device 500 may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0127] The processor 502 is used to run a computer program 5032 stored in a memory to implement all the steps of the unsupervised decomposition method of photovoltaic and baseline load.

[0128] It should be understood that in the embodiments of this application, the processor 502 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.

[0129] It will be understood by those skilled in the art that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program includes program instructions and can be stored in a storage medium, which is a computer-readable storage medium. The program instructions are executed by at least one processor in the computer system to implement the process steps of the embodiments of the above methods.

[0130] Therefore, the present invention also provides a storage medium. This storage medium may be a computer-readable storage medium. The storage medium stores a computer program, wherein when executed by a processor, the computer program causes the processor to perform all steps of the unsupervised decomposition method of photovoltaic and baseline load.

[0131] The storage medium can be any computer-readable storage medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0132] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0133] In the embodiments provided by this invention, it should be understood that the disclosed systems and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative. For example, the division of each unit is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0134] The steps in the method of this invention can be adjusted, merged, or reduced in order according to actual needs. The units in the system of this invention can be merged, divided, or reduced according to actual needs. Furthermore, the functional units in the various embodiments of this invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0135] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a terminal, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0136] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An unsupervised decomposition method for photovoltaic and baseline load, characterized in that, include: Collect and process the net load and related exogenous data of the target substation to obtain preprocessed results; Based on the preprocessing results, input features for photovoltaic and baseline loads are created to obtain photovoltaic features and baseline load features; The photovoltaic characteristics and the baseline load characteristics are respectively input into a time-series neural network to obtain the photovoltaic trend curve and the baseline load trend curve; The net load sequence is reconstructed by weighting and merging the photovoltaic trend curve and the baseline load trend curve using a linear combination layer; this includes performing element-wise weighted summation on the photovoltaic trend curve and the baseline load trend curve at the same time index to reconstruct the net load sequence. The formulas used include ; ; ;in, , This represents the weighting coefficients that are automatically updated during training; , These are the photovoltaic trend curve and the baseline load trend curve, respectively. Output net load sequence; The time-series neural network is obtained by collecting and processing historical net load and related exogenous data of the target substation, and building input features and training labels of photovoltaic and baseline load as a sample set to train the time-series neural network structure. The training process of the temporal neural network includes: Construct a day / night mask based on irradiation triples, and read the data at each time t. , , Set small tolerance Used to suppress measurement noise, a preset day / night determination signal is defined. When the triplet is not greater than the small tolerance It was determined to be nighttime and ordered When the triplet is not simultaneously greater than the small tolerance The time was determined to be daytime. During the training phase, the input features of photovoltaics are... and Partial multiplication; Training samples are constructed using a non-overlapping sliding window method. The window length is set to L. K non-overlapping windows are obtained by sequentially dividing the time series from the starting point, and the tail segments with a length of less than L are discarded. The training process of the temporal neural network also includes: Two parallel neural network channels, one for photovoltaic and one for baseline load, were constructed. Each channel was configured with an input projection layer and a time-series coding backbone, and softplus activation was configured at the end of the channel to output a non-negative trend sequence. The neural network channel for processing photovoltaic data is connected in series with a day / night mask module after softplus, and the nighttime data is cleared bit by bit according to the preset day / night determination signal.

2. The unsupervised decomposition method for photovoltaic and baseline load according to claim 1, characterized in that, The temporal neural network includes a temporal neural network structure with two parallel neural network channels.

3. The unsupervised decomposition method for photovoltaic and baseline load according to claim 2, characterized in that, The collection and processing of the net load and related exogenous data of the target substation to obtain preprocessing results includes: Obtain the net load and related exogenous data of the target substation to obtain initial data; The initial data was completed using the piecewise cubic Hermitian interpolation PCHIP method. Align the data time intervals, select a uniform sampling period, and perform full-window aggregation and take the arithmetic mean on the initial data window with the required resolution. The net load time series was normalized by maximum value, and the exogenous features were normalized by minimum-maximum value to obtain the preprocessing results.

4. The unsupervised decomposition method for photovoltaic and baseline load according to claim 1, characterized in that, The photovoltaic characteristics are represented as follows: Baseline load characteristics are expressed as Training labels are represented as ;in, This represents the temperature at time t; This represents the wind speed at time t; Represents the global horizontal radiation at time t; This represents the horizontal scattered radiation at time t; Represents the direct normal radiation at time t; , , These represent the time feature codes for the day of the week, hour, and minute at time t, including one-hot coding and sincos coding; This indicates the type of holiday at time t, where 1 represents a holiday and 0 represents a non-holiday. This represents the relative humidity at time t; This represents the net load at time t.

5. The unsupervised decomposition method for photovoltaic and baseline load according to claim 1, characterized in that, The step of inputting the photovoltaic characteristics and the baseline load characteristics into a time-series neural network to obtain the photovoltaic trend curve and the baseline load trend curve includes: The photovoltaic features and the baseline load features are respectively input into the corresponding neural network channels of the time-series neural network. The time-series neural network extracts their respective time-series patterns in parallel and outputs the corresponding trend sequences. At the same time, a normalization scale is determined based on the full sample of the training set, and maximum value normalization is performed on the trend sequences of photovoltaic and baseline load to obtain the photovoltaic trend curve and the baseline load trend curve.

6. An unsupervised decomposition system for photovoltaic and baseline load, characterized in that, The system uses the unsupervised decomposition method of photovoltaic and baseline load as described in any one of claims 1 to 5, including: The preprocessing unit is used to collect and process the net load and related exogenous data of the target substation to obtain preprocessing results; A creation unit is used to create input features for photovoltaic and baseline load based on the preprocessing results, so as to obtain photovoltaic features and baseline load features; The input unit is used to input the photovoltaic features and the baseline load features into the time-series neural network to obtain the photovoltaic trend curve and the baseline load trend curve, respectively. The reconstruction unit is used to reconstruct the net load sequence by weighting and merging the photovoltaic trend curve and the baseline load trend curve using a linear combination layer. Output unit, used to output net load sequence.

Images