A user-side post-meter light photovoltaic power decomposition method

Abstract

The application provides a user-side post-meter photovoltaic power decomposition method, which comprises the following steps: obtaining the net load power curve of a user in a plurality of continuous first preset time periods to generate an extreme value curve, and predicting photovoltaic capacity C according to the extreme value curve by a capacity estimation model; receiving meteorological data curves including horizontal irradiance curve, direct irradiance curve, scattered irradiance curve, temperature curve and relative humidity curve in a second preset time period and performing capacity normalization processing, receiving the net load power curve of the user in the second preset time period and performing capacity normalization processing, decomposing an initial photovoltaic power curve according to the capacity-normalized net load power curve and the meteorological data curves by a photovoltaic decomposition model, and performing capacity reverse normalization processing on the initial photovoltaic power to generate a predicted photovoltaic power curve. The application can decompose the photovoltaic power curve of the photovoltaic equipment on the meter side according to the net load power curve of the user meter, and facilitate subsequent fine management of the power grid.

Description

Technical Field

[0001] This invention relates to the field of power grid and distributed energy management technology, specifically to a method for decomposing photovoltaic power after the user-side meter. Background Technology

[0002] With the rapid development of distributed photovoltaic (PV), the penetration rate of post-meter PV systems (i.e., PV equipment installed after the user's electricity meter) has significantly increased. While these systems reduce users' electricity costs, their fluctuating and uncontrollable power generation pose challenges to the power balance, load forecasting, and dispatching of the distribution network. To achieve refined grid management and user-side energy optimization, it is necessary to accurately decompose PV power generation and actual user electricity load from the net load data of the electricity meter (i.e., the superposition of grid power supply and user-generated and consumed PV power). However, current technologies still have the following limitations:

[0003] Current physical model-based photovoltaic (PV) power decomposition methods rely on the geographical location of PV arrays (such as roof tilt angle and orientation) and external meteorological data to construct accurate power curves. However, most PV systems are located on private rooftops, and their geographical information involves sensitive privacy. This lack of information directly limits the application scope and reliability of physical models. Furthermore, existing physical models are typically designed based on standardized PV module parameters (such as conversion efficiency and degradation coefficient). In real-world scenarios, different users have significantly different PV equipment models, installation years, and maintenance conditions, and there are even cases where multiple types of modules are used interchangeably. These individualized characteristics make it difficult for a uniform physical model to accurately characterize the dynamic characteristics of a single-point PV system, especially under conditions of equipment aging, partial shading, or abnormal operating conditions, further amplifying the model output error.

[0004] Many traditional data-driven algorithms assume that the relationships between data are linear or expressed as simple functions. This assumption is often inaccurate when dealing with highly nonlinear and complex systems like photovoltaics, leading to poor model fit. Furthermore, traditional algorithms typically require extensive manual feature engineering—extracting and selecting features from raw data. This process is not only time-consuming but may also miss important influencing factors due to a lack of domain knowledge or experience, thus reducing model effectiveness.

[0005] Therefore, in order to solve the above problems, a photovoltaic power decomposition method based on user net load data is needed, which can mine the nonlinear characteristics between data. Summary of the Invention

[0006] In view of this, the problem to be solved by the present invention is to provide a method for decomposing photovoltaic power after the user-side meter, which can decompose the photovoltaic power curve of the photovoltaic equipment on the meter side based on the net load power curve of the user's electricity meter, so as to facilitate the subsequent refined management of the power grid.

[0007] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0008] A method for decomposing photovoltaic power after a user's meter includes predicting photovoltaic capacity, obtaining net load power curves of users within several consecutive first preset time periods, generating an extreme value curve with the sequence number of the first preset time periods as the horizontal axis and the minimum value of the net load power curve as the vertical axis, and a capacity estimation model predicting the photovoltaic capacity C of the user's meter-side photovoltaic equipment based on the extreme value curve.

[0009] The photovoltaic power is decomposed. Meteorological data curves including horizontal irradiance curve, direct irradiance curve, diffuse irradiance curve, temperature curve and relative humidity curve are received within the second preset time period and normalized for capacity. The net load power curve of users within the second preset time period is received and normalized for capacity. The photovoltaic decomposition model decomposes the initial photovoltaic power curve based on the capacity-normalized net load power curve and meteorological data curve. The initial photovoltaic power is then inversely normalized for capacity to generate the predicted photovoltaic power curve.

[0010] Furthermore, each day includes two pre-set time periods: nighttime and daytime.

[0011] Generating the meteorological data curves includes: obtaining the average values ​​of horizontal irradiance, direct irradiance, diffuse irradiance, temperature, and relative humidity for each hour of the day, and generating horizontal irradiance curves, direct irradiance curves, diffuse irradiance curves, temperature curves, and relative humidity curves with time as the horizontal axis and the average value as the vertical axis.

[0012] Furthermore, the capacity normalization process includes dividing the net load power curve or meteorological data curve by the photovoltaic capacity C;

[0013] The capacity denormalization process includes multiplying the initial photovoltaic power curve by the photovoltaic capacity C.

[0014] Furthermore, after capacity normalization of the net load power curve and meteorological data curve, it is then normalized using Z-score.

[0015] A method for expanding the training set of a capacity estimation model includes obtaining the photovoltaic power curves of several known users' photovoltaic power generation systems. The actual load curve L of its electrical equipment and the true photovoltaic capacity of its photovoltaic system. ;

[0016] The expansion method includes standardizing the photovoltaic power curve G, and the standardization formula is:

[0017] ,

[0018] in, and These are the minimum and maximum values ​​of the photovoltaic power curve within a day. It is a standardized photovoltaic power generation curve;

[0019] Actual photovoltaic capacity of several users By estimating kernel density to generate a smooth probability density function, Monte Carlo sampling generates several virtual capacities based on the probability density function. Randomly combined virtual capacity Actual load curve L and standardized photovoltaic power generation curve And generate several virtual capacities The corresponding virtual net load curve is calculated using the following formula:

[0020] ,

[0021] ,

[0022] in, P represents the virtual photovoltaic power curve, and P represents the virtual net load curve.

[0023] Several virtual capacities will be generated The corresponding virtual net load curve is added to the training set.

[0024] A photovoltaic decomposition model comprises, in sequence, an input layer, a CNN layer, a first BILSTM layer, a second BILSTM layer, an attention layer, a fully connected layer, and an output layer. The input layer receives the capacity-normalized net load power curve and meteorological data curve. The CNN layer is used for local feature extraction. Both the first and second BILSTM layers are used to extract the temporal features of the data curve. The attention layer is used to capture key information. The fully connected layer and the output layer are used to output the initial photovoltaic power.

[0025] Furthermore, both the first BILSTM layer and the second BILSTM layer are followed by Dropout layers for randomly dropping out a number of neurons.

[0026] The advantages and positive effects of this invention are:

[0027] By setting up a capacity estimation model, which predicts the photovoltaic capacity C of the user-side photovoltaic equipment based on the extreme value curve, and then normalizing the meteorological data curve and the net load power curve through the photovoltaic capacity C normalization process, the capacity normalization process can convert both the meteorological data curve and the net load power curve into a ratio relationship curve with the photovoltaic capacity C. This establishes a strong correlation between the photovoltaic capacity C and the net load power, and a strong correlation between the photovoltaic capacity C and the meteorological data. This facilitates the extraction of correlation features and time series correlation features between the photovoltaic capacity C, the net load power, or the meteorological data by the photovoltaic decomposition model. Based on these correlation features, the photovoltaic power curve of the meter-side photovoltaic equipment can be predicted more accurately, which is conducive to the subsequent refined management of the power grid. Attached Figure Description

[0028] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0029] Figure 1 This is a schematic diagram of a method for decomposing photovoltaic power after metering on the user side.

[0030] Figure 2 This is a structural diagram of a photovoltaic decomposition model;

[0031] Figure 3 This is a comparison chart of the photovoltaic power prediction curve and the actual photovoltaic curve for a user-side photovoltaic power decomposition method. Detailed Implementation

[0032] 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 embodiments of the present invention, and not all embodiments. 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.

[0033] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed terms.

[0034] This invention provides a method for decomposing photovoltaic power after the user-side meter, such as... Figure 1As shown, the method includes predicting photovoltaic capacity, obtaining the net load power curves of users within a number of consecutive first preset time periods, generating an extreme value curve with the number of first preset time periods as the horizontal axis and the minimum value of the net load power curve as the vertical axis, and the capacity estimation model predicting the photovoltaic capacity C of the photovoltaic equipment on the meter side of the user based on the extreme value curve.

[0035] Each day includes two first preset time periods: night and day. One embodiment of this application is: the first preset time period is set to 12 hours, with 7 a.m. to 7 p.m. being daytime and 7 p.m. to 7 a.m. the next day being nighttime.

[0036] The net load power curve is generated by continuously collecting power values ​​from user meters and plotting time on the horizontal axis and power values ​​on the vertical axis. When the power generation of the user-side photovoltaic equipment is higher than the user's electricity consumption, the photovoltaic equipment supplies power to the grid, and the power value of the user's meter is negative; when the power generation of the user-side photovoltaic equipment is lower than the user's electricity consumption, the grid supplies power to the user, and the power value of the user's meter is positive.

[0037] One embodiment of this application is as follows: If the user's electricity consumption is 0, the photovoltaic equipment is exactly at its maximum power generation state. At this time, the power value of the meter is at its minimum and is equal to the photovoltaic capacity C. The daytime minimum value corresponds to the maximum power provided by the user-side photovoltaic equipment to the grid, and the nighttime minimum value represents the user's minimum power consumption (e.g., a refrigerator plugged in all day). The difference between the nighttime minimum value and the daytime minimum value is close to the photovoltaic capacity C of the photovoltaic equipment.

[0038] The process of generating extreme value curves involves: acquiring the net load power curves for the user during nighttime and daytime (60 consecutive first preset time periods) over a 30-day period; obtaining the minimum value for each net load power curve; and generating extreme value curves with the sequence number of the first preset time periods as the horizontal axis and the minimum value as the vertical axis. The capacity estimation model receives the extreme value curves and predicts the photovoltaic capacity C of the user-side photovoltaic equipment based on the extreme value curves.

[0039] The photovoltaic power is decomposed. Meteorological data curves including horizontal irradiance curve, direct irradiance curve, diffuse irradiance curve, temperature curve and relative humidity curve are received within the second preset time period and normalized for capacity. The net load power curve of users within the second preset time period is received and normalized for capacity. The photovoltaic decomposition model decomposes the initial photovoltaic power curve based on the capacity-normalized net load power curve and meteorological data curve. The initial photovoltaic power is then inversely normalized for capacity to generate the predicted photovoltaic power curve.

[0040] One embodiment of this application is as follows: the second preset time period is 1 day, and the net load power curve and meteorological data curve are fixedly captured within 24 hours by setting a sliding window. The generation of meteorological data curves includes: obtaining the average values ​​of horizontal irradiance, direct irradiance, diffuse irradiance, temperature and relative humidity for each hour within a day, and generating horizontal irradiance curve, direct irradiance curve, diffuse irradiance curve, temperature curve and relative humidity curve with time as the horizontal axis and the average value as the vertical axis.

[0041] Capacity normalization processing includes receiving the generated horizontal irradiance curve, direct irradiance curve, diffuse irradiance curve, temperature curve, relative humidity curve, and user net load power curve. All six curves are divided by the (predicted) user photovoltaic capacity C to complete capacity normalization. This normalization process establishes a strong correlation between the six curves and the photovoltaic capacity C, facilitating the extraction of correlation features between the net load power curve, climate data curve, and photovoltaic capacity C and photovoltaic power curve by the photovoltaic decomposition model. It also extracts the time-series distribution characteristics of the net load power curve based on photovoltaic capacity C and the climate data curve based on photovoltaic capacity C, improving the accuracy of photovoltaic power curve prediction.

[0042] The capacity inverse normalization process includes multiplying the initial photovoltaic power curve output by the photovoltaic decomposition model by the photovoltaic capacity C to generate the predicted photovoltaic power curve.

[0043] One embodiment of this application is as follows: after capacity normalization processing of the net load power curve and meteorological data curve, the net load power curve and meteorological data curve are then processed by Z-score standardization. Z-score standardization can convert the net load power curve and meteorological data curve into a distribution with a mean of 0 and a standard deviation of 1, which can eliminate the dimensional differences between different curves, make different curves of the same order of magnitude, and improve the convergence speed of the neural network when extracting features in the photovoltaic decomposition model.

[0044] A method for expanding the training set of a capacity estimation model is presented. The photovoltaic capacity model receives 60 sample points (one month's eigenvalue curves) as input and outputs a single capacity estimate as output. The number of hidden layer neurons used for feature extraction in the photovoltaic capacity model is 128, 256, 128, 64, and 32, respectively. Each hidden layer uses a Rectified Linear Unit (ReLU) activation function. The first three hidden layers incorporate a dropout mechanism with a dropout rate of 0.1. The optimizer is Adaptive Moment Estimation (Adam), with 500 iterations, a learning rate of 0.001, mean squared error (MSE) loss, and a batch size of 64.

[0045] Capacity estimation process: Extract the minimum values ​​from the net load curves of each night and day within a month for the user, and generate extreme value curves as input features. The input vector dimension is 60 (30 days × 2 features / day). The capacity estimation model uses the user's photovoltaic capacity C as the output.

[0046] Training the capacity estimation model involves obtaining a training set, which is selected from several users of the Ausgrid dataset from the Australian power grid company, and testing is conducted using photovoltaic power curves. The time resolution of the actual load curve L is reduced to 1 hour.

[0047] The Ausgrid dataset includes the net load curves of several known users' electricity meters and the actual photovoltaic capacitance of their photovoltaic systems. Photovoltaic power curve of photovoltaic power generation system The actual load curve L of the electrical equipment. The Ausgrid dataset is expanded using an augmentation method to increase the data strength within the training set. This augmentation method includes obtaining the photovoltaic power curves of several known users' photovoltaic power generation systems. The actual load curve L of its electrical equipment and the true photovoltaic capacity of its photovoltaic system. The expansion method includes standardizing the photovoltaic power curve G, and the standardization formula is:

[0048] ,

[0049] in, and These are the minimum and maximum values ​​of the photovoltaic power curve within a day. It is a standardized photovoltaic power generation curve. and The difference between C and G is the photovoltaic capacity. The difference between them represents the actual power generation of the photovoltaic equipment. The value is between 0 and 1, and is used to represent the power generation efficiency of photovoltaic equipment. The closer it is to 1, the higher the power generation efficiency.

[0050] Actual photovoltaic capacity of several known users By estimating kernel density to generate a smooth probability density function, Monte Carlo sampling generates several virtual capacities based on the probability density function. In this invention, a Gaussian kernel is selected as the KDE kernel function, and the optimal bandwidth is selected from the range of 0 to 10 using a grid search method, with the optimal bandwidth being 0.26.

[0051] Randomly combined virtual capacity Actual load curve L and standardized photovoltaic power generation curve And generate several virtual capacities The corresponding virtual net load curve is calculated using the following formula:

[0052] ,

[0053] ,

[0054] in, P represents the virtual photovoltaic power curve, and P represents the virtual net load curve.

[0055] Specifically, when two photovoltaic (PV) devices are installed in close proximity, their irradiance and climate data are essentially the same; therefore, the corresponding virtual capacity of the two PV devices is similar. The curves with similar or identical shapes will affect the virtual capacity. With standardized photovoltaic curves Multiplication can generate a virtual photovoltaic power curve. Randomly select actual load curves and with A virtual net load curve is generated by overlaying the generated virtual capacities. The corresponding virtual net load curve is added to the training set to expand the user scenarios within the dataset and achieve data augmentation.

[0056] A photovoltaic decomposition model, such as Figure 2 As shown, the photovoltaic decomposition model sequentially includes an input layer, a CNN layer, a first BILSTM layer, a second BILSTM layer, an attention layer, a fully connected layer, and an output layer. The input layer receives the capacity-normalized net load power curve and meteorological data curve. The CNN layer is used for local feature extraction. Both the first and second BILSTM layers are used to extract the temporal features of the data curves. The attention layer is used to capture key information. The fully connected layer and the output layer are used to output the initial photovoltaic power. The first and second BILSTM layers are both followed by a Dropout layer for randomly discarding a number of neurons.

[0057] The photovoltaic decomposition model first uses a one-dimensional convolutional neural network (CNN) to extract local features from the time series of net load power curves and meteorological data curves. The input channel number is 6, the output channel number is 24, the kernel size is 3, and the padding is 1. Then, it models the model using a two-layer bidirectional long short-term memory (BiLSTM) neural network. Each BiLSTM layer has 48 hidden units in both the forward and backward directions, allowing it to fully learn the temporal dependencies of local features. A dropout mechanism with a dropout rate of 0.3 is added to each BiLSTM layer to randomly deactivate network units, increasing the model's generalization ability and preventing overfitting. Next, a multi-head attention mechanism is used to better capture key information. The embedding dimension is 96, and the number of attention heads is 8. Finally, the model outputs through a fully connected layer with 96 input features and 24 output features, outputting the initial photovoltaic power.

[0058] The output of the photovoltaic decomposition model is then inversely normalized by capacity to predict the photovoltaic power curve for unknown users over 24 hours. The CNN-BiLSTM-Attention hybrid model for photovoltaic power decomposition uses Adam as the optimizer, 100 iterations, a learning rate of 0.001, MSE as the loss function, and a batch size of 128.

[0059] One embodiment of this application involves randomly selecting 260 users, choosing one-third of them as known users (as the training set), and the remaining two-thirds as unknown users (used as the test set). Meteorological data is sourced from the U.S. National Solar Radiation Database, and meteorological data from the areas where the 260 users are located are selected to represent the average weather information level for the entire region. Figure 3 As shown, the photovoltaic power curve (predicted value) predicted by the capacity estimation model is highly similar to the actual photovoltaic power curve (true value).

[0060] The embodiments of the present invention have been described in detail above, but the content described is only a preferred embodiment of the present invention and should not be considered as limiting the scope of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the present invention.

Claims

1. A method for decomposing photovoltaic power after a user-side meter, characterized in that, include, To predict photovoltaic capacity, the net load power curves of users within several consecutive first preset time periods are obtained, and an extreme value curve is generated with the sequence number of the first preset time period as the horizontal axis and the minimum value of the net load power curve as the vertical axis. The capacity estimation model predicts the photovoltaic capacity C of the photovoltaic equipment on the meter side of the user based on the extreme value curve. The photovoltaic power is decomposed, and meteorological data curves including horizontal irradiance curve, direct irradiance curve, diffuse irradiance curve, temperature curve and relative humidity curve are received within the second preset time period and normalized by capacity. The net load power curve of users within the second preset time period is received and normalized by capacity. The photovoltaic decomposition model decomposes the initial photovoltaic power curve based on the capacity-normalized net load power curve and meteorological data curve. The initial photovoltaic power is inversely normalized by capacity to generate the predicted photovoltaic power curve. The capacity normalization process includes dividing the net load power curve or meteorological data curve by the photovoltaic capacity C; the capacity denormalization process includes multiplying the initial photovoltaic power curve by the photovoltaic capacity C. The method for expanding the training set of the capacity estimation model includes: obtaining the photovoltaic power curves of several known users' photovoltaic power generation systems. The actual load curve L of its electrical equipment and the true photovoltaic capacity of its photovoltaic system. ; The expansion method includes standardizing the photovoltaic power curve G, and the standardization formula is: ， in, and These are the minimum and maximum values ​​of the photovoltaic power curve within a day. It is a standardized photovoltaic power generation curve; Actual photovoltaic capacity of several known users By estimating kernel density to generate a smooth probability density function, Monte Carlo sampling generates several virtual capacities based on the probability density function. Randomly combined virtual capacity Actual load curve L and standardized photovoltaic power generation curve And generate several virtual capacities The corresponding virtual net load curve is calculated using the following formula: ， ， in, P represents the virtual photovoltaic power curve, and P represents the virtual net load curve. Several virtual capacities will be generated The corresponding virtual net load curve is added to the training set.

2. The method for decomposing photovoltaic power after a user-side meter as described in claim 1, characterized in that, Each day includes two preset time periods: night and day. Generating the meteorological data curves includes: obtaining the average values ​​of horizontal irradiance, direct irradiance, diffuse irradiance, temperature, and relative humidity per hour within a second preset time period, and generating horizontal irradiance curves, direct irradiance curves, diffuse irradiance curves, temperature curves, and relative humidity curves with time as the horizontal axis and the average value as the vertical axis.

3. The method for decomposing photovoltaic power after a user-side meter as described in claim 1, characterized in that, After normalizing the net load power curve and meteorological data curve, the data is then normalized using Z-score.

4. A method for decomposing photovoltaic power after the user-side meter according to any one of claims 1 to 3, characterized in that, The photovoltaic decomposition model sequentially includes an input layer, a CNN layer, a first BILSTM layer, a second BILSTM layer, an attention layer, a fully connected layer, and an output layer. The input layer is used to receive the capacity-normalized net load power curve and meteorological data curve. The CNN layer is used for local feature extraction. Both the first and second BILSTM layers are used to extract the temporal features of the data curve. The attention layer is used to capture key information. The fully connected layer and the output layer are used to output the initial photovoltaic power.

5. The method for decomposing photovoltaic power after the user-side meter as described in claim 4, characterized in that, Both the first BILSTM layer and the second BILSTM layer are followed by Dropout layers for randomly dropping out a number of neurons.

Images