A distributed photovoltaic grid-connected point voltage prediction method

Abstract

The application provides a distributed photovoltaic grid-connected point voltage prediction method, relates to the technical field of distributed photovoltaic power generation, and comprises the following steps: acquiring voltage influence parameters of a distributed photovoltaic grid-connected point; inputting the voltage influence parameters into a preset grid-connected point voltage prediction network model to obtain multiple predicted voltage sub-signal components of the distributed photovoltaic grid-connected point; the model takes sample voltage influence parameters as model input, takes multiple sample voltage sub-signal components corresponding to sample voltage as model output, and takes sample voltage sub-signal component features obtained based on the multiple sample voltage sub-signal components as model constraints, and the network model is trained to obtain the multiple predicted voltage sub-signal components; and superimposing the multiple predicted voltage sub-signal components of the distributed photovoltaic grid-connected point to obtain predicted voltage of the distributed photovoltaic grid-connected point. Through capturing the change rule of the voltage of the distributed photovoltaic grid-connected point, the voltage of the grid-connected point is predicted according to the voltage influence parameters, and the accuracy of voltage prediction is improved.

Description

Technical Field

[0001] This invention relates to the field of distributed photovoltaic power generation, and in particular to a method for predicting the voltage at the grid connection point of distributed photovoltaic power generation. Background Technology

[0002] With the rapid development of distributed photovoltaic (PV) power generation and the large-scale integration of distributed PV into the distribution network, the safe and stable operation of the distribution network has been significantly impacted. Voltage exceeding limits at the grid connection point is a key issue. Because the power output and consumption of distributed PV are constantly changing, the voltage at the grid connection point is difficult to predict accurately and control effectively. Currently, the main methods for predicting the voltage at the grid connection point of distributed PV include traditional statistical methods and some machine learning-based methods. Traditional statistical methods, such as time series analysis, typically assume that the data has specific statistical properties such as stationarity. However, in actual distributed PV grid-connected systems, due to the complexity and variability of environmental factors, the non-stationarity and nonlinear characteristics of the data are significant, limiting the prediction accuracy of traditional statistical methods. Machine learning-based methods, such as artificial neural networks, can handle nonlinear problems to some extent, but when dealing with complex time series data such as voltage data, they struggle to fully extract the forward and backward characteristics of the data, thus limiting the model's generalization ability and prediction accuracy.

[0003] Therefore, there is an urgent need for a method that can accurately and effectively predict the voltage at the grid connection point of distributed photovoltaic power generation in order to improve the operational stability and reliability of the distribution network and ensure the safe and efficient operation of the distributed photovoltaic power generation system. Summary of the Invention

[0004] The purpose of this invention is to address the above problems by providing a method for predicting the voltage at the grid connection point of distributed photovoltaic systems.

[0005] This invention provides a method for predicting voltage at the grid connection point of a distributed photovoltaic system, comprising: S1. Obtain the voltage impact parameters of the distributed photovoltaic grid connection point; S2. Input the voltage influence parameters into a preset grid connection point voltage prediction network model to obtain multiple predicted voltage sub-signal components of the distributed photovoltaic grid connection point; the preset grid connection point voltage prediction network model is obtained by training the network model with the sample voltage influence parameters of the distributed photovoltaic grid connection point as the model input, multiple sample voltage sub-signal components corresponding to the sample voltage of the distributed photovoltaic grid connection point as the model output, and the sample voltage sub-signal component features obtained based on the multiple sample voltage sub-signal components as the model constraints. S3. The predicted voltage sub-signal components of the distributed photovoltaic grid-connected point are superimposed to obtain the predicted voltage of the distributed photovoltaic grid-connected point.

[0006] According to the technical solutions provided in some embodiments of the present invention, the method further includes: S0-1. Obtain the sample voltage dataset of distributed photovoltaic grid-connected points; the sample voltage dataset of distributed photovoltaic grid-connected points includes multiple sets of one-to-one corresponding sample voltages and sample voltage influencing parameters; S0-2. Perform intrinsic mode separation on each of the sample voltages to obtain multiple sample voltage sub-signal components corresponding to each of the sample voltages; S0-3. Based on the multiple sample voltage sub-signal components, and using a bidirectional temporal convolutional network model, obtain the sample voltage sub-signal component features corresponding to each of the sample voltage sub-signal components. S0-4. Using the sample voltage influence parameters as model input, multiple sample voltage sub-signal components as model output, and the characteristics of the sample voltage sub-signal components as model constraints, perform network model training to obtain the preset grid connection point voltage prediction network model.

[0007] According to the technical solutions provided in certain embodiments of the present invention, S0-2, intrinsic mode separation is performed on each of the sample voltages to obtain a plurality of sample voltage sub-signal components corresponding to each of the sample voltages, including: S0-2-1, Use the sample voltage as the voltage to be separated; S0-2-2. Add noise to the voltage to be separated to obtain a noisy voltage; S0-2-3. Perform empirical mode decomposition on the noisy voltage to obtain multiple intrinsic mode functions and residual component functions; S0-2-4. If the residual component function is a monotonic function, then output the intrinsic mode function as the sample voltage sub-signal component; otherwise, proceed to step S0-2-5. S0-2-5. Calculate the average intrinsic mode function based on the multiple intrinsic mode functions, update the voltage to be separated based on the average intrinsic mode function, and then execute steps S0-2-2 to S0-2-4 until the residual component function is a monotonic function.

[0008] According to the technical solutions provided in certain embodiments of the present invention, S0-3, based on a bidirectional temporal convolutional network model, the sample voltage sub-signal component features corresponding to each of the sample voltage sub-signal components are obtained, including: S0-3-1. Determine the optimal combination of hyperparameters for the bidirectional temporal convolutional network model; S0-3-2. Based on the sample voltage sub-signal components, obtain the features of the sample voltage sub-signal components using a bidirectional temporal convolutional network model with optimal hyperparameter combination.

[0009] According to the technical solutions provided in certain embodiments of the present invention, S0-3-1, determining the optimal hyperparameter combination of a bidirectional temporal convolutional network model includes: Determine the initial search interval for the optimal hyperparameters; An initial hyperparameter combination is generated based on the spatial density distribution function and the initial search interval of the optimal hyperparameters, and the initial hyperparameter combination is configured into the bidirectional temporal convolutional network model. The hyperparameters of the bidirectional temporal convolutional network model with the initial hyperparameter combination are iteratively optimized until the iteration termination condition is met, thus obtaining the optimal hyperparameter combination.

[0010] According to certain embodiments of the present invention, a bidirectional temporal convolutional network model configured with an initial hyperparameter combination is subjected to iterative hyperparameter optimization until the iteration termination condition is met, thereby obtaining the optimal hyperparameter combination, including: The sample training dataset is divided into a first training set and a first test set; the sample training dataset includes sample voltage influence parameters and sample voltage sub-signal components. Using the sample voltage influence parameters of the first training set as model input and the sample voltage sub-signal components of the first training set as model output, the bidirectional temporal convolutional network model configured with the initial hyperparameter combination is trained to obtain an intermediate convolutional network model. The sample voltage influence parameters of the first test set are input into the intermediate convolutional network model to obtain the first predicted voltage sub-signal component; The hyperparameter fitness function is calculated based on the first predicted voltage sub-signal component and the sample voltage sub-signal component of the first test set. The correlation coefficients between hyperparameters and the network model are calculated based on the hyperparameter fitness function and the initial hyperparameter combination. The hyperparameter combination of the bidirectional temporal convolutional network model is dynamically adjusted based on the correlation coefficient between the hyperparameters and the network model until the hyperparameter fitness function value is less than or equal to a preset fitness threshold, or the number of iterations is greater than or equal to a preset number, thus obtaining the optimal hyperparameter combination.

[0011] According to the technical solutions provided in certain embodiments of the present invention, features of the sample voltage sub-signal components are obtained based on a bidirectional temporal convolutional network model with optimal hyperparameter combinations, including: The sample voltage sub-signal component is input into the bidirectional temporal convolutional network model with the optimal hyperparameter combination. The forward and backward features of the sample voltage sub-signal component are obtained through the forward convolutional layer and the backward convolutional layer of the bidirectional temporal convolutional network model with the optimal hyperparameter combination, respectively. Calculate the forward and reverse feature weights of the sample voltage sub-signal components based on the sample voltage sub-signal components; Based on the forward feature weights and inverse feature weights, and using a variable weighted collaborative feature aggregation strategy, the forward features and inverse features are fused to obtain the sample voltage sub-signal component features.

[0012] According to certain embodiments of the present invention, calculating the forward feature weight and the reverse feature weight of the sample voltage sub-signal component based on the sample voltage sub-signal component includes: Calculate the mean and variance of the sample voltage sub-signal components; Based on the hyperbolic tangent function, a fully connected operation is performed on the mean and variance to calculate the weighted intermediate variables; The intermediate weight variables are normalized to obtain an adaptive weight vector; The forward and reverse feature weights of the sample voltage sub-signal components are determined based on the adaptive weight vector.

[0013] According to the technical solutions provided in certain embodiments of the present invention, S0-4, using the sample voltage influence parameters as model input, multiple sample voltage sub-signal components as model output, and the characteristics of the sample voltage sub-signal components as model constraints, a network model is trained to obtain the preset grid connection point voltage prediction network model, including: S0-4-1. Divide the sample training dataset into a second training set and a second test set; the sample training dataset includes sample voltage influence parameters and sample voltage sub-signal components. S0-4-2. Using the sample voltage influence parameters of the second training set as the model input, the sample voltage sub-signal components of the second training set as the model output, and the sample voltage sub-signal component features output by the bidirectional temporal convolutional network model with the optimal hyperparameter combination as the constraint condition, the model is trained to obtain the intermediate prediction model. S0-4-3. Input the sample voltage influence parameters of the second test set into the intermediate prediction model to obtain the second predicted voltage sub-signal component; S0-4-4 Calculate the signal error between the second predicted voltage sub-signal component and the sample voltage sub-signal component of the second test set. If the signal error is greater than the signal error threshold, then proceed to step S0-4-5. S0-4-5. Adjust the model parameters of the intermediate prediction model according to the signal error and update the optimal hyperparameter combination of the intermediate convolutional network model. Then execute steps S0-4-2 to S0-4-4 until the signal error is less than or equal to the signal error threshold to obtain the preset grid connection point voltage prediction network model.

[0014] According to the technical solutions provided in certain embodiments of the present invention, the hyperparameter combination includes: the learning rate of the bidirectional temporal convolutional network, the number of convolutional kernels of the bidirectional temporal convolutional network, and the L2 regularization coefficient of the bidirectional temporal convolutional network.

[0015] In summary, this invention provides a method for predicting the voltage of a distributed photovoltaic (PV) grid-connected point, comprising: obtaining voltage influence parameters of the distributed PV grid-connected point; inputting the voltage influence parameters into a preset grid-connected point voltage prediction network model to obtain multiple predicted voltage sub-signal components of the distributed PV grid-connected point; the preset grid-connected point voltage prediction network model is obtained by training a network model with sample voltage influence parameters of the distributed PV grid-connected point as model input, multiple sample voltage sub-signal components corresponding to the sample voltage of the distributed PV grid-connected point as model output, and sample voltage sub-signal component features obtained based on the multiple sample voltage sub-signal components as model constraints; and superimposing the multiple predicted voltage sub-signal components of the distributed PV grid-connected point to obtain the predicted voltage of the distributed PV grid-connected point. This invention constructs a grid-connected point voltage prediction network model with sample voltage influence parameters (including grid-connected point irradiance, temperature, PV output power, and load power consumption) as input, sample voltage sub-signal components as output, and sample voltage sub-signal component features as constraints. The voltage influence parameters of the distributed PV grid-connected point are input into this grid-connected point voltage prediction network model to obtain multiple predicted voltage sub-signal components, which are then superimposed to obtain the final predicted voltage. Among them, the voltage sub-signal component characteristics can capture the voltage variation pattern of distributed photovoltaic grid connection point, which solves the problem that the existing technology is difficult to capture the voltage variation pattern of grid connection point, resulting in inaccurate voltage prediction results and inability to adapt to complex voltage variation characteristics. It improves the accuracy and reliability of voltage prediction, realizes accurate prediction of grid connection point voltage based on voltage influence parameters of distributed photovoltaic grid connection point, improves the stability and reliability of distribution network operation, and thus ensures the safe and efficient operation of distributed photovoltaic power generation system.

[0016] It should be understood that the descriptions of technical features, technical solutions, beneficial effects, or similar language in this invention do not imply that all features and advantages can be achieved in any single embodiment. Rather, it is understood that the description of a feature or beneficial effect means that a specific technical feature, technical solution, or beneficial effect is included in at least one embodiment. Therefore, the descriptions of technical features, technical solutions, or beneficial effects in this specification do not necessarily refer to the same embodiment. Furthermore, the technical features, technical solutions, and beneficial effects described in this embodiment can be combined in any suitable manner. Those skilled in the art will understand that embodiments can be implemented without one or more specific technical features, technical solutions, or beneficial effects of a particular embodiment. In other embodiments, additional technical features and beneficial effects may be identified in specific embodiments that do not embody all embodiments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 A flowchart illustrating a method for predicting voltage at a distributed photovoltaic grid connection point, provided in an embodiment of the present invention; Figure 2 This is a flowchart illustrating step S0 as provided in an embodiment of the present invention; Figure 3 This is a flowchart illustrating step S0-2 provided in an embodiment of the present invention; Figure 4 This is a flowchart illustrating steps S0-3 provided in an embodiment of the present invention; Figure 5 This is a flowchart illustrating step S0-3-1 provided in an embodiment of the present invention; Figure 6 This is a flowchart illustrating step S0-3-1-3 provided in an embodiment of the present invention; Figure 7 This is a flowchart illustrating step S0-3-2 provided in an embodiment of the present invention; Figure 8 This is a flowchart illustrating step S0-3-2-2 provided in an embodiment of the present invention; Figure 9 This is a flowchart illustrating steps S0-4 provided in an embodiment of the present invention; Figure 10 This is a schematic diagram of the intrinsic mode decomposition results of the sample voltage sub-signal components provided in an embodiment of the present invention; Figure 11 This is a schematic diagram illustrating the change of the fitness function with the number of iterations provided in an embodiment of the present invention; Figure 12 This is a schematic diagram comparing different voltage prediction models provided in the embodiments of the present invention; Figure 13 This is a schematic diagram illustrating the evaluation of voltage prediction results provided in an embodiment of the present invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the technical solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. This description is merely illustrative and explanatory, and should not be construed as limiting the scope of protection of the present invention in any way. Specifically, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort should fall within the scope of protection of the present invention.

[0020] It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or apparatus.

[0021] With the rapid development of distributed photovoltaic (PV) power generation (PV systems that are distributed on the user side, connected to the distribution network nearby, and generated and consumed locally), large-scale distributed PV access to the distribution network has placed higher demands on the safe and stable operation of the distribution network. Among these demands, the voltage at the connection point between the PV equipment and the distribution network (the distributed PV grid connection point) has become a crucial factor affecting the safe and stable operation of the distribution network. Due to the strong randomness and volatility of PV power generation (e.g., PV output power fluctuates significantly when sunlight intensity and temperature change rapidly), and the inherent impedance of the distribution network lines, the power generated by the PV equipment causes a voltage drop on the distribution network lines. This directly leads to a rapid rise or fall at the connection point between the PV equipment and the distribution network (the distributed PV grid connection point). Simultaneously, the user-side load power also changes in real time (e.g., there are peak and off-peak electricity consumption periods within the same day), further exacerbating the voltage fluctuations at the distributed PV grid connection point. Therefore, the voltage at the distributed PV grid connection point is highly susceptible to exceeding its limits.

[0022] Currently, the main methods for predicting voltage at grid-connected points of distributed photovoltaic (PV) systems include traditional statistical methods and some machine learning-based methods. Traditional statistical methods, such as time series analysis, typically assume that the data possesses specific statistical properties such as stationarity. However, in actual distributed PV grid-connected systems, due to the complexity and variability of environmental factors, the non-stationarity and nonlinearity of the data are significant, limiting the prediction accuracy of traditional statistical methods. Machine learning-based methods, such as artificial neural networks, can handle nonlinear problems to some extent, but they have shortcomings in feature extraction and model parameter optimization. When processing complex time series data such as voltage data, it is difficult to fully extract the forward and backward features of the data, and the hyperparameters of the model are often difficult to determine accurately, thus limiting the model's generalization ability and prediction accuracy. Furthermore, some existing methods do not fully consider the intrinsic modal characteristics of the data when processing grid-connected point voltage data, failing to effectively decompose and analyze the voltage data, thereby affecting the accuracy of the prediction.

[0023] As mentioned above, regarding the problems in the existing technology, such as Figure 1 As shown, this embodiment provides a method for predicting the voltage at the grid connection point of a distributed photovoltaic system, including: S1. Obtain the voltage impact parameters of the distributed photovoltaic grid connection point; Specifically, photovoltaic (PV) power generation exhibits strong randomness and volatility (e.g., PV output power fluctuates significantly with changes in irradiance and temperature), leading to voltage fluctuations at the connection point between PV equipment and the distribution network (distributed PV grid-connected voltage). Simultaneously, user-side load power also changes in real time (e.g., peak and off-peak electricity consumption occur within the same day), further exacerbating voltage fluctuations at the distributed PV grid-connected point. Therefore, the voltage at the distributed PV grid-connected point is highly susceptible to exceeding its limits. As mentioned above, since the voltage at the distributed PV grid-connected point is simultaneously affected by irradiance, ambient temperature, PV power output, and user electricity load, irradiance, temperature, PV output power, and load power consumption can be selected as voltage influence parameters for the distributed PV grid-connected point. These parameters can then be used to predict the grid-connected voltage. However, directly modeling and predicting the original voltage can lead to difficulties in distinguishing noise, random fluctuations, and effective variation patterns, resulting in high model convergence difficulty, weak generalization ability, and unreliable prediction accuracy. Therefore, this invention first decomposes the original non-stationary voltage signal into multiple relatively stationary voltage sub-signal components with different frequencies and time scales through intrinsic mode separation. Each voltage sub-signal component has a clear physical meaning and can correspond to voltage components caused by different disturbance sources such as sudden changes in illumination, power fluctuations, and load changes. This makes the characteristics of the voltage sub-signal components clearer and their patterns easier to capture, thereby improving the accuracy of subsequent predictions.

[0024] S2. Input the voltage influence parameters into the preset grid connection point voltage prediction network model to obtain multiple predicted voltage sub-signal components of the distributed photovoltaic grid connection point. The preset grid connection point voltage prediction network model is obtained by training the network model with the sample voltage influence parameters of the distributed photovoltaic grid connection point as the model input, the multiple sample voltage sub-signal components corresponding to the sample voltage of the distributed photovoltaic grid connection point as the model output, and the sample voltage sub-signal component features obtained based on the multiple sample voltage sub-signal components as the model constraints. Specifically, after obtaining the voltage influence parameters of the distributed photovoltaic (PV) grid-connected point, these parameters are input into a preset grid-connected point voltage prediction network model. This yields multiple predicted voltage sub-signal components for the distributed PV grid-connected point. The predicted voltage of the distributed PV grid-connected point can then be obtained based on these sub-signal components. The preset grid-connected point voltage prediction network model uses historical data on sample voltage influence parameters (irradiance, temperature, PV output power, and load power consumption) as input, multiple sample voltage sub-signal components corresponding to the sample voltage (historical data) as output, and the features of these sample voltage sub-signal components as model constraints. The network model is trained using these features, which are compact representations of the time-series information of the sample voltage sub-signal components. These features guide the model training process, allowing the preset grid-connected point voltage prediction network model to better reflect the actual voltage variation characteristics of the distributed PV grid-connected point, thereby improving the prediction accuracy of the voltage sub-signal components.

[0025] S3. Superimpose multiple predicted voltage sub-signal components of the distributed photovoltaic grid-connected point to obtain the predicted voltage of the distributed photovoltaic grid-connected point.

[0026] Specifically, after obtaining multiple predicted voltage sub-signal components of the distributed photovoltaic grid-connected point, the predicted voltage of the distributed photovoltaic grid-connected point can be obtained by superimposing the predicted voltage sub-signal components, thus realizing the prediction of the grid-connected point voltage based on the voltage influence parameters of the distributed photovoltaic grid-connected point.

[0027] This invention constructs a grid-connected point voltage prediction network model. This model takes voltage influence parameters (grid-connected point irradiance, temperature, photovoltaic output power, and load power consumption) as input, outputs voltage sub-signal components, and uses the characteristics of sample voltage sub-signal components as constraints for training. Inputting relevant voltage influence parameters of the distributed photovoltaic grid-connected point into this prediction network outputs multiple predicted voltage sub-signal components, which are then superimposed to obtain the final predicted grid-connected point voltage. By leveraging the characteristics of the voltage sub-signal components, this invention can effectively capture the complex variation patterns of distributed photovoltaic grid-connected point voltage, overcoming the shortcomings of traditional methods that struggle to adapt to dynamic voltage characteristics and lack sufficient prediction accuracy. This effectively improves the accuracy and reliability of voltage prediction, achieving precise prediction of distributed photovoltaic grid-connected point voltage, thereby enhancing the stability and security of the distribution network and ensuring the safe, efficient, and stable operation of the distributed photovoltaic power generation system.

[0028] In a preferred embodiment, such as Figure 2 As shown, the method also includes: S0-1. Obtain the sample voltage dataset of distributed photovoltaic grid-connected points; the sample voltage dataset of distributed photovoltaic grid-connected points includes multiple sets of one-to-one corresponding sample voltages and sample voltage influencing parameters. Specifically, to predict the corresponding voltage using the voltage influence parameters of the grid connection point, a grid connection point voltage prediction network model (i.e., a pre-defined grid connection point voltage prediction network model) needs to be established in advance. This pre-defined model requires historical data of the grid connection point voltage influence parameters (sample voltage influence parameters) as input and historical data of the sub-signal components of the sample voltage corresponding to the sample voltage influence parameters (sample voltage sub-signal components) as output for training. Therefore, the first step is to obtain the training data, namely, the distributed photovoltaic grid connection point sample voltage dataset. This dataset includes multiple sets of one-to-one corresponding sample voltages and sample voltage influence parameters. The sample voltage sub-signal components required for training can then be obtained by processing the sample voltages.

[0029] S0-2. Perform intrinsic mode separation on each sample voltage to obtain multiple sample voltage sub-signal components corresponding to each sample voltage; Specifically, to obtain the sample voltage sub-signal components required for model training, intrinsic mode separation of the sample voltage is necessary. This decomposes the original non-stationary voltage signal into multiple relatively stationary voltage sub-signal components with different frequencies and time scales. Each voltage sub-signal component has a clear physical meaning and can correspond to voltage components caused by different disturbance sources such as grid-connected irradiance, temperature, photovoltaic output power, and load power consumption. This lays the foundation for subsequently capturing the variation patterns of distributed photovoltaic grid-connected voltage.

[0030] S0-3. Based on multiple sample voltage sub-signal components, obtain the sample voltage sub-signal component features corresponding to each sample voltage sub-signal component using a bidirectional temporal convolutional network model. Specifically, after obtaining the sample voltage sub-signal components, to further explore the temporal variation characteristics and inherent correlations contained within these components, this invention employs a bidirectional temporal convolutional network model to extract features from each sample voltage sub-signal component. This yields sample voltage sub-signal component features with strong characterization capabilities, corresponding to each component. These features reflect the impact of factors such as grid-connected point irradiance, temperature, photovoltaic output power, and load power consumption on the grid-connected point voltage. This provides support for subsequent training of a pre-defined grid-connected point voltage prediction network model using the sample voltage sub-signal component features as constraints, further enhancing the model's learning ability and prediction accuracy regarding complex voltage variation patterns.

[0031] S0-4. Using the sample voltage influence parameters as model input, multiple sample voltage sub-signal components as model output, and the characteristics of the sample voltage sub-signal components as model constraints, the network model is trained to obtain the preset grid connection point voltage prediction network model.

[0032] Specifically, after performing intrinsic mode decomposition (IMD) on the sample voltage and extracting features from the sample voltage sub-signal components, the sample voltage influence parameters are used as input to the network model, and the multiple sample voltage sub-signal components obtained through IMD decomposition are used as outputs. The extracted sample voltage sub-signal component features are then used as constraints for model training to optimize the prediction network. By introducing sample voltage sub-signal component features as model constraints, the model can be guided to focus more on the variation patterns of the voltage signal during training, avoiding excessive attention to noise and redundant information, and improving the model's feature learning ability and generalization performance. After iterative training, a pre-defined grid-connected point voltage prediction network model with accurate prediction capabilities is obtained, laying the foundation for subsequent voltage prediction of distributed photovoltaic grid-connected points.

[0033] In a preferred embodiment, such as Figure 3 As shown, S0-2, intrinsic mode separation is performed on each sample voltage to obtain multiple sample voltage sub-signal components corresponding to each sample voltage, including: S0-2-1, Use the sample voltage as the voltage to be separated; Specifically, in this embodiment, the complete empirical mode decomposition method with adaptive noise is used to perform intrinsic mode separation on the sample voltage. Since the sample voltage is the object requiring intrinsic mode separation, it is used as the voltage to be separated. .

[0034] S0-2-2. Add noise to the voltage to be separated to obtain a noisy voltage; Specifically, the voltage to be separated is obtained. Then, the voltage to be separated is determined according to the following formula (1). Add Gaussian white noise Noisy voltage is obtained : Formula (1) in, Noisy voltage; It is Gaussian white noise; The voltage to be separated; Noise number, , It represents the total number of times noise was added. It is a time variable.

[0035] By introducing Gaussian white noise to assist in the intrinsic mode decomposition of sample voltage, mode aliasing can be suppressed, numerical singularities can be avoided during the decomposition process, the accuracy and stability of intrinsic mode separation can be improved, and the physical characteristics of each mode component can be made clearer.

[0036] S0-2-3. Perform empirical mode decomposition on the noisy voltage to obtain multiple intrinsic mode functions and residual component functions; Specifically, when obtaining the noisy voltage Then, the noisy voltage is processed according to the following formula (2). Empirical mode decomposition is performed to obtain multiple intrinsic mode functions. and residual component functions : Formula (2) in, Noisy voltage; Add to The noise of the first One intrinsic mode function; Noise number; It is the index of the intrinsic mode function; It is the total number of intrinsic mode functions obtained through decomposition; This is the residual component function.

[0037] S0-2-4. If the residual component function is a monotonic function, then output the intrinsic mode function as the sample voltage sub-signal component; otherwise, proceed to step S0-2-5. S0-2-5. Calculate the average intrinsic mode function based on multiple intrinsic mode functions, update the voltage to be separated based on the average intrinsic mode function, and then execute steps S0-2-2 to S0-2-4 until the residual component function is a monotonic function.

[0038] Specifically, after obtaining the residual component function Next, it is necessary to determine its monotonicity. If the residual component function... If the voltage increases or decreases monotonically with time, it indicates that the voltage to be separated is... There are no longer any intrinsic mode functions that can be further decomposed, satisfying the termination condition of decomposition. The decomposed intrinsic mode functions are the sample voltage sub-signal components.

[0039] If the residual component function If the voltage is not monotonic with time, it indicates that the voltage to be separated is... The intrinsic mode functions still contain further extractable intrinsic mode functions, requiring continued iterative decomposition. At this point, the average intrinsic mode function needs to be calculated according to the following formula (3). : Formula (3) in, The average intrinsic mode function; It represents the total number of times noise was added. Noise number; It is the index of the intrinsic mode function; Add to The noise of the first One intrinsic mode function.

[0040] After obtaining the average intrinsic mode function Then, update the voltage to be separated according to the following formula (4): Formula (4) in, The updated voltage to be separated; The voltage to be separated; is the average intrinsic mode function.

[0041] Next, noise is added to the updated voltage to be separated, resulting in a new noisy voltage. Empirical Mode Decomposition (EMD) is then performed on this new noisy voltage, yielding multiple new intrinsic mode functions (IMFs) and new residual component functions. It is then determined whether the new residual component functions are monotonic. If they are monotonic, the new IMFs are output as voltage sub-signal components, completing the IMF decomposition process. If they are not monotonic, the voltage to be separated is updated, and the above steps are repeated until the latest decomposition yields a monotonic residual component function. In this case, the IMFs are output as voltage sub-signal components, ending the IMF decomposition process.

[0042] In a preferred embodiment, S0-3, based on multiple sample voltage sub-signal components and a bidirectional temporal convolutional network model, the features of the sample voltage sub-signal components corresponding to each sample voltage sub-signal component are obtained, including: S0-3-1. Determine the optimal combination of hyperparameters for the bidirectional temporal convolutional network model; S0-3-2. Based on the sample voltage sub-signal components, obtain the characteristics of the sample voltage sub-signal components using a bidirectional temporal convolutional network model with optimal hyperparameter combination.

[0043] Specifically, such as Figure 4 As shown, this step first determines the optimal hyperparameter combination of the bidirectional temporal convolutional network model through a dynamic constraint optimization regulator, maximizing the feature extraction capability of the bidirectional temporal convolutional network. Then, the bidirectional temporal convolutional network with the optimal hyperparameters is used to extract the features of the sample voltage sub-signal components, obtaining the variation law of the distributed photovoltaic grid-connected point voltage. This guides the model training process, making the preset grid-connected point voltage prediction network model more closely match the actual variation characteristics of the distributed photovoltaic grid-connected point voltage, thereby improving the prediction accuracy of the preset grid-connected point voltage prediction network model for voltage sub-signal components.

[0044] Furthermore, such as Figure 5 As shown, in a preferred embodiment, S0-3-1, determining the optimal hyperparameter combination of the bidirectional temporal convolutional network model includes: S0-3-1-1. Determine the initial search interval for the optimal hyperparameters; S0-3-1-2. Generate an initial hyperparameter combination based on the spatial density distribution function and the initial search interval of the optimal hyperparameters, and configure the initial hyperparameter combination into the bidirectional temporal convolutional network model; S0-3-1-3. Perform iterative optimization of the hyperparameters of the bidirectional temporal convolutional network model with the initial hyperparameter combination until the iteration termination condition is met, and obtain the optimal hyperparameter combination.

[0045] Specifically, the process of determining the optimal hyperparameter combination for the bidirectional temporal convolutional network model is as follows: For the learning rate l and the number of convolutional kernels N of the bidirectional temporal convolutional network... conv The three hyperparameters, L1, L2 regularization coefficient, and L2, are first set to the initial search intervals of the dynamic constraint optimization controller as

[10] . -4 10 -1 ]、[16,64]、 [10 -4 10 -1 Then, based on the spatial density distribution function, an initial hyperparameter combination is generated within the interval. The expression for the spatial density distribution function is shown in the following formula (5): Formula (5) in, Let be the spatial density distribution function; For hyperparameters; The mean of the hyperparameter range; denoted as the standard deviation of the hyperparameter.

[0046] Next, the initial hyperparameters will be combined. The input configuration is fed into the bidirectional temporal convolutional network model, and the hyperparameters of the bidirectional temporal convolutional network model with the initial hyperparameter combination are iteratively optimized until the iteration termination condition is met, and the optimal hyperparameter combination is obtained.

[0047] Furthermore, such as Figure 6 As shown, in a preferred embodiment, S0-3-1-3, the bidirectional temporal convolutional network model configured with the initial hyperparameter combination is subjected to iterative optimization of hyperparameters until the iteration termination condition is met to obtain the optimal hyperparameter combination, including: S0-3-1-3-1. Divide the sample training dataset into a first training set and a first test set; the sample training dataset includes sample voltage influence parameters and sample voltage sub-signal components. S0-3-1-3-2. Using the sample voltage influence parameters of the first training set as the model input and the sample voltage sub-signal components of the first training set as the model output, train the bidirectional temporal convolutional network model with the initial hyperparameter combination to obtain the intermediate convolutional network model. S0-3-1-3-3: Input the sample voltage influence parameters of the first test set into the intermediate convolutional network model to obtain the first predicted voltage sub-signal component; S0-3-1-3-4. Calculate the hyperparameter fitness function based on the first predicted voltage sub-signal component and the sample voltage sub-signal component of the first test set. S0-3-1-3-5. Calculate the correlation coefficient between hyperparameters and network model based on the hyperparameter fitness function and the initial hyperparameter combination. S0-3-1-3-6. Dynamically adjust the hyperparameter combination of the bidirectional temporal convolutional network model according to the correlation coefficient between the hyperparameters and the network model until the hyperparameter fitness function value is less than or equal to the preset fitness threshold, or the number of iterations is greater than or equal to the preset number, to obtain the optimal hyperparameter combination.

[0048] Specifically, the process of iteratively optimizing the hyperparameters of a bidirectional temporal convolutional network model with an initial hyperparameter combination is as follows: First, the sample training dataset, containing sample voltage influence parameters (irradiance, temperature, photovoltaic output power, and load power consumption) and sample voltage sub-signal components, is split into a first training set and a first test set. These sets serve as the first training set for model training and the first test set for hyperparameter performance verification, respectively, providing a standardized verification benchmark for subsequent iterative hyperparameter optimization and avoiding overfitting issues caused by using a single dataset. Then, using the sample voltage influence parameters from the first training set as model input and the sample voltage sub-signal components as model output, a bidirectional temporal convolutional network model with an initial hyperparameter combination is trained, resulting in an intermediate convolutional network model with basic feature extraction capabilities. Finally, the sample voltage influence parameters from the first test set are input into the intermediate convolutional network model to obtain the first predicted voltage sub-signal component. Next, based on the first predicted voltage sub-signal component... and the sample voltage sub-signal components of the first test set The hyperparameter fitness function is calculated according to the following formula (6). : Formula (6) in, The hyperparameter fitness function; This is the first predicted voltage sub-signal component; The sample voltage sub-signal components of the first test set.

[0049] Next, based on the hyperparameter fitness function Combination of initial hyperparameters The correlation coefficient between hyperparameters and the network model is calculated according to the following formula (7). : Formula (7) in, It is the correlation coefficient between hyperparameters and the network model; It is the number of iterations; It is the total number of iterations; It is the first Hyperparameter values ​​at the next iteration; It is the mean of the hyperparameters; No. The fitness function value at the next iteration; It is the mean of the fitness function values.

[0050] Based on the correlation coefficient between hyperparameters and network model The weight factors of the dynamic constraint optimization controller are calculated according to the following formula (8). : Formula (8) in, The weighting factors for the dynamic constraint optimization controller; These are the initial weighting factors for the dynamic constraint optimization controller; For adjustment coefficients; It is the correlation coefficient between hyperparameters and the network model.

[0051] Based on the adjusted weighting factor The dynamic constraint optimization regulator can optimize the search direction and step size of hyperparameters, continuously iterate to generate new hyperparameter combinations and repeat the performance evaluation process until the hyperparameter fitness function value is less than or equal to the preset fitness threshold, or the number of iterations is greater than or equal to the preset number, and obtain the optimal hyperparameter combination. The hyperparameter combination obtained at this time is the optimal hyperparameter combination of the bidirectional temporal convolutional network model.

[0052] In a preferred embodiment, S0-3-2, based on the sample voltage sub-signal components, and using a bidirectional temporal convolutional network model with optimal hyperparameter combinations, the features of the sample voltage sub-signal components are obtained, including: S0-3-2-1. Input the sample voltage sub-signal component into the bidirectional temporal convolutional network model with the optimal hyperparameter combination, and obtain the forward and reverse features of the sample voltage sub-signal component through the forward convolutional layer and the reverse convolutional layer of the bidirectional temporal convolutional network model with the optimal hyperparameter combination, respectively. S0-3-2-2, Calculate the forward and reverse feature weights of the sample voltage sub-signal components based on the sample voltage sub-signal components; S0-3-2-3. Based on the forward feature weights and inverse feature weights, and using a variable weighted collaborative feature aggregation strategy, the forward and inverse features are fused to obtain the sample voltage sub-signal component features.

[0053] Specifically, such as Figure 7 As shown, after obtaining the optimal hyperparameter combination, the bidirectional temporal convolutional network model is configured with the optimal hyperparameter combination. At this time, the sample voltage sub-signal component is input into the bidirectional temporal convolutional network model with the optimal hyperparameter combination. The forward features of the sample voltage sub-signal component can be obtained through the forward convolutional layer of the bidirectional temporal convolutional network model with the optimal hyperparameter combination according to the following formula (9): Formula (9) in, The output features of the forward convolutional layer are called forward features. This represents the size of the convolutional kernel's time window, i.e., the length of the convolutional kernel in the time dimension. This refers to the element in the β-th row of the convolution kernel of the forward convolutional layer; The first component of the sample voltage sub-signal Voltage component values ​​at each time step; For the current time step, ; This represents the offset of the convolution kernel within the time window, and its value ranges from 0 to... ; It is an activation function; This is the bias term for the forward convolutional layer.

[0054] Through this operation, the forward convolutional layer extracts the temporal features of the sample voltage sub-signal components from the starting position to the ending position of the sample voltage sub-signal components, capturing the forward evolution pattern of the data in the time dimension.

[0055] Similarly, the forward features of the sample voltage sub-signal components can be obtained from the inverse convolutional layer of the bidirectional temporal convolutional network model with optimal hyperparameter combination according to the following formula (10): Formula (10) in, The output features of the inverse convolutional layer are called inverse features. This represents the size of the convolutional kernel's time window, i.e., the length of the convolutional kernel in the time dimension. This represents the element in the β-th row of the convolution kernel of the inverse convolutional layer; The first component of the sample voltage sub-signal Voltage component values ​​at each time step; For the current time step, ; This represents the offset of the convolution kernel within the time window, and its value ranges from 0 to... ; It is an activation function; This is the bias term for the inverse convolutional layer.

[0056] This process extracts temporal features from the end position of the sample voltage sub-signal component to the beginning position of the sample voltage sub-signal component, mines the inverse time dependence in the sample voltage sub-signal component, and captures the inverse evolution pattern of the data in the time dimension.

[0057] Next, the forward feature weights of the sample voltage sub-signal components are calculated based on the sample voltage sub-signal components. and inverse feature weights Then, based on the forward feature weights... and inverse feature weights Based on the variable weight collaborative feature aggregation strategy, the forward features are aggregated according to the following formula (11). and reverse features By fusing the samples, the characteristics of the voltage sub-signal components can be obtained. : Formula (11) in, Sample voltage sub-signal component characteristics; This indicates a dimensional concatenation operation; Forward feature weights; These are the inverse feature weights; The output features of the forward convolutional layer are called forward features. These are the output features of the inverse convolutional layer, i.e., the inverse features.

[0058] Furthermore, such as Figure 8 As shown, in a preferred embodiment, S0-3-2-2, calculating the forward and reverse feature weights of the sample voltage sub-signal components based on the sample voltage sub-signal components, includes: S0-3-2-2-1. Calculate the mean and variance of the sample voltage sub-signal components; S0-3-2-2-2: Based on the hyperbolic tangent function, a fully connected operation is performed on the mean and variance to calculate the weighted intermediate variables; S0-3-2-2-3. Normalize the intermediate weight variables to obtain the adaptive weight vector; S0-3-2-2-4. Determine the forward and reverse feature weights of the sample voltage sub-signal components based on the adaptive weight vector.

[0059] Specifically, to calculate the forward characteristic weights of the sample voltage sub-signal components and inverse feature weights The mean value of the sample voltage sub-signal component should first be calculated according to the following formula (12). The variance of the sample voltage sub-signal component is calculated according to the following formula (13). : Formula (12) Formula (13) in, This is the mean of the sample voltage sub-signal components; The variance of the sample voltage sub-signal component; The sample voltage sub-signal component.

[0060] Then, based on the hyperbolic tangent function, the mean is calculated according to the following formula (14). and variance Perform a fully connected operation to calculate the intermediate weight variables. : Formula (14) in, is the intermediate variable for weights; tanh is the hyperbolic tangent function; FC is the fully connected operation; This is the mean of the sample voltage sub-signal components; Let V be the variance of the sample voltage sub-signal components.

[0061] Then, the intermediate variables are weighted according to the following formula (15). Normalization is performed to obtain the adaptive weight vector. : Formula (15) in, It is an adaptive weight vector, that is , Forward feature weights; These are the inverse feature weights; These are intermediate variables for weighting.

[0062] Due to the adaptive weight vector Therefore, in obtaining the adaptive weight vector The forward feature weights can then be obtained from this. and inverse feature weights .

[0063] In a preferred embodiment, S0-4, using sample voltage influence parameters as model input, multiple sample voltage sub-signal components as model output, and the characteristics of the sample voltage sub-signal components as model constraints, a network model is trained to obtain a preset grid connection point voltage prediction network model, including: S0-4-1. Divide the sample training dataset into a second training set and a second test set; the sample training dataset includes sample voltage influence parameters and sample voltage sub-signal components. S0-4-2. Using the sample voltage influence parameters of the second training set as the model input, the sample voltage sub-signal components of the second training set as the model output, and the sample voltage sub-signal component features output by the bidirectional temporal convolutional network model with the optimal hyperparameter combination as the constraint condition, the model is trained to obtain the intermediate prediction model. S0-4-3. Input the sample voltage influence parameters of the second test set into the intermediate prediction model to obtain the second predicted voltage sub-signal component; S0-4-4 Calculate the signal error between the second predicted voltage sub-signal component and the sample voltage sub-signal component of the second test set. If the signal error is greater than the signal error threshold, proceed to step S0-4-5. S0-4-5. Adjust the model parameters of the intermediate prediction model according to the signal error and update the optimal hyperparameter combination of the intermediate convolutional network model. Then execute steps S0-4-2 to S0-4-4 until the signal error is less than or equal to the signal error threshold to obtain the preset grid connection point voltage prediction network model.

[0064] Specifically, such as Figure 9 As shown, after obtaining the sample voltage influence parameters, sample voltage sub-signal components, and sample voltage sub-signal component features, network model training can be performed to obtain the preset grid connection point voltage prediction network model. The specific process is as follows: First, the sample training dataset is divided into a second training set and a second test set, which respectively serve the functions of model training and model performance verification. The sample training dataset includes sample voltage influence parameters and sample voltage sub-signal components. Then, using the sample voltage influence parameters from the second training set as model input, the sample voltage sub-signal components from the second training set as model output, and the sample voltage sub-signal component features output by the bidirectional temporal convolutional network model with the optimal hyperparameter combination as constraints, model training is performed to obtain an intermediate prediction model. Afterwards, the sample voltage influence parameters from the second test set are input into the intermediate prediction model to obtain the second predicted voltage sub-signal component, and the signal error between the second predicted voltage sub-signal component and the sample voltage sub-signal component from the second test set is calculated. If the signal error is greater than the signal error threshold, it indicates that the performance of the current intermediate prediction model is substandard. The model parameters of the intermediate prediction model should be adjusted according to the signal error, and the optimal hyperparameter combination of the intermediate convolutional network model should be updated. The update process of the intermediate convolutional network model has been described in steps S0-3-1-3-1 to S0-3-1-3-6 and will not be repeated here. Then, based on the sample voltage sub-signal component features output by the updated intermediate convolutional network model, the adjusted intermediate prediction model, and the second training set and the second test set, model training continues until the signal error is less than or equal to the signal error threshold, thus obtaining the preset grid connection point voltage prediction network model.

[0065] In a preferred embodiment, the hyperparameter combination includes: the learning rate of the bidirectional temporal convolutional network, the number of convolutional kernels of the bidirectional temporal convolutional network, and the L2 regularization coefficient of the bidirectional temporal convolutional network.

[0066] Specifically, in this embodiment, the hyperparameter combination includes: the learning rate of the bidirectional temporal convolutional network, the number of convolutional kernels of the bidirectional temporal convolutional network, and the L2 regularization coefficient of the bidirectional temporal convolutional network. These three parameters are the core key parameters that determine the network training efficiency, feature extraction capability, and generalization stability. The learning rate regulates the model parameter update rate, the number of convolutional kernels determines the dimension and richness of feature extraction, and the L2 regularization coefficient is used to suppress model overfitting. The synergistic effect of the three directly affects the feature extraction effect of the bidirectional temporal convolutional network on sample voltage sub-signal components and the overall performance of the subsequent prediction network. Integrating them into a combination for synergistic iterative optimization can achieve the optimal combination of each parameter, so that the comprehensive performance of the bidirectional temporal convolutional network reaches the optimal level.

[0067] Furthermore, the advantages of this invention have been demonstrated through a series of experiments, specifically including: Experiment 1 Figure 10 This diagram illustrates the intrinsic mode decomposition (IMF) results of the sample voltage sub-signal components. A complete empirical mode decomposition (EMF) method based on adaptive noise was used to separate the intrinsic modes of the grid-connected voltage data, resulting in multiple voltage sub-signal components: nine IMF components and one residual component. This decomposes the highly nonlinear and unstable grid-connected voltage data into several weakly nonlinear and unstable components. The results show that each component is relatively independent, and the correlation between the mode components is weak, effectively reducing the volatility and instability of the grid-connected voltage data.

[0068] Experiment 2 Figure 11 This diagram illustrates the change of the fitness function with the number of iterations, reflecting the performance trend of the dynamic constraint optimization regulator in optimizing the hyperparameters of the bidirectional temporal convolutional network. As the number of iterations gradually increases, the fitness function value gradually converges. When the iteration approaches the termination condition, the fitness function value stabilizes near its minimum value, at which point the hyperparameter combination reaches its optimum, and the network performance reaches its optimal state. The results demonstrate the effectiveness of the dynamic constraint optimization regulator designed in this invention, providing a solid guarantee for high-precision prediction of grid connection point voltage.

[0069] Experiment 3 Figure 12 This diagram illustrates a comparison of different voltage prediction models. The figure compares the prediction results of this invention with those of Bidirectional Temporal Convolutional Network (BiTCN), Dynamically Constrained Optimization Bidirectional Temporal Convolutional Network (DCRO-BiTCN), and Bidirectional Temporal Convolutional Network Based on Adaptive Noise Complete Empirical Mode Decomposition (CEEMDAN-BiTCN). The results show that the grid-connected voltage predicted by this invention is closest to the actual value and almost coincides with the true voltage curve. This invention exhibits the highest accuracy in voltage prediction for distributed photovoltaic grid-connected points.

[0070] Experiment 4, such as Figure 13This diagram illustrates the evaluation of voltage prediction results. The present invention is quantitatively evaluated against other methods, including Bidirectional Temporal Convolutional Network (BiTCN), Dynamically Constrained Optimized Bidirectional Temporal Convolutional Network (DCRO-BiTCN), and Bidirectional Temporal Convolutional Network Based on Adaptive Noise Complete Empirical Mode Decomposition (CEEMDAN-BiTCN), by calculating the mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (R²). The MSE value ranges from [0, +∞), with a smaller value indicating a smaller difference between the predicted and actual values, and higher prediction accuracy. The MAE value also ranges from [0, +∞), reflecting the magnitude of the mean absolute deviation between the predicted and actual values; a smaller MAE indicates a prediction closer to the actual value. The R² value ranges from [0, 1], with a value closer to 1 indicating a better fit to the data and superior prediction network performance. The results show that the present invention exhibits the best performance in voltage prediction, with MSE, MAE, and R² values ​​of 0.1035, 0.2472, and 0.9966, respectively. Compared with the DCRO-BiTCN method, the MSE and MAE of this invention are reduced by 89.57% and 63.85%, respectively, and R² is increased by 2.96%. Compared with the CEEMDAN-BiTCN method, the MSE and MAE of this invention are reduced by 86.80% and 71.22%, respectively, and R² is increased by 2.23%.

[0071] In summary, the distributed photovoltaic (PV) grid-connected point voltage prediction method provided by this invention has significant advantages in the field of distributed PV grid-connected point voltage prediction. By extracting the characteristics of voltage sub-signal components (the variation law of distributed PV grid-connected point voltage) to guide model training and optimization, the accuracy and reliability of voltage prediction are greatly improved, with prediction accuracy significantly superior to existing methods. By performing mode decomposition on the raw voltage data, the problems of strong nonlinearity and poor stability of voltage data are effectively solved. By constructing a dynamic constraint optimization regulator, adaptive optimization of hyperparameter combinations is achieved, ensuring that the network is in the optimal prediction state. This method effectively overcomes the technical bottlenecks of insufficient feature extraction, poor hyperparameter adaptability, and insufficient prediction accuracy in traditional prediction methods, effectively improving the accuracy and reliability of voltage prediction, achieving accurate prediction of distributed PV grid-connected point voltage, providing reliable technical support for improving the voltage stability and regulation accuracy of distributed PV grid-connected systems, and providing strong technical guarantee for the efficient and safe operation of the power grid.

[0072] This article uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. The above are only preferred embodiments of the present invention. It should be noted that due to the limitations of textual expression, while there are objectively infinite specific structures, those skilled in the art can make several improvements, modifications, or changes without departing from the principles of the present invention, and can also combine the above technical features in an appropriate manner; these improvements, modifications, changes, or combinations, or the direct application of the inventive concept and technical solution to other occasions without modification, should all be considered within the scope of protection of the present invention.

Claims

1. A method for predicting voltage at a distributed photovoltaic grid connection point, characterized in that, include: S1. Obtain the voltage impact parameters of the distributed photovoltaic grid connection point; S2. Input the voltage influence parameters into a preset grid connection point voltage prediction network model to obtain multiple predicted voltage sub-signal components of the distributed photovoltaic grid connection point; the preset grid connection point voltage prediction network model is obtained by training the network model with the sample voltage influence parameters of the distributed photovoltaic grid connection point as the model input, multiple sample voltage sub-signal components corresponding to the sample voltage of the distributed photovoltaic grid connection point as the model output, and the sample voltage sub-signal component features obtained based on the multiple sample voltage sub-signal components as the model constraints. S3. The predicted voltage sub-signal components of the distributed photovoltaic grid-connected point are superimposed to obtain the predicted voltage of the distributed photovoltaic grid-connected point.

2. The method for predicting voltage at the grid connection point of distributed photovoltaic power according to claim 1, characterized in that, The method further includes: S0-1. Obtain the sample voltage dataset of distributed photovoltaic grid-connected points; the sample voltage dataset of distributed photovoltaic grid-connected points includes multiple sets of one-to-one corresponding sample voltages and sample voltage influencing parameters; S0-2. Perform intrinsic mode separation on each of the sample voltages to obtain multiple sample voltage sub-signal components corresponding to each of the sample voltages; S0-3. Based on the multiple sample voltage sub-signal components, and using a bidirectional temporal convolutional network model, obtain the sample voltage sub-signal component features corresponding to each of the sample voltage sub-signal components. S0-4. Using the sample voltage influence parameters as model input, multiple sample voltage sub-signal components as model output, and the characteristics of the sample voltage sub-signal components as model constraints, perform network model training to obtain the preset grid connection point voltage prediction network model.

3. The method for predicting voltage at the grid connection point of distributed photovoltaic power according to claim 2, characterized in that, S0-2. Perform intrinsic mode separation on each of the sample voltages to obtain multiple sample voltage sub-signal components corresponding to each sample voltage, including: S0-2-1, Use the sample voltage as the voltage to be separated; S0-2-2. Add noise to the voltage to be separated to obtain a noisy voltage; S0-2-3. Perform empirical mode decomposition on the noisy voltage to obtain multiple intrinsic mode functions and residual component functions; S0-2-4. If the residual component function is a monotonic function, then output the intrinsic mode function as the sample voltage sub-signal component; otherwise, proceed to step S0-2-5. S0-2-5. Calculate the average intrinsic mode function based on the multiple intrinsic mode functions, update the voltage to be separated based on the average intrinsic mode function, and then execute steps S0-2-2 to S0-2-4 until the residual component function is a monotonic function.

4. The method for predicting voltage at the grid connection point of distributed photovoltaic power according to claim 2, characterized in that, S0-3. Based on the multiple sample voltage sub-signal components, and using a bidirectional temporal convolutional network model, obtain the sample voltage sub-signal component features corresponding to each of the sample voltage sub-signal components, including: S0-3-1. Determine the optimal combination of hyperparameters for the bidirectional temporal convolutional network model; S0-3-2. Based on the sample voltage sub-signal components, obtain the features of the sample voltage sub-signal components using a bidirectional temporal convolutional network model with optimal hyperparameter combination.

5. The method for predicting voltage at the grid connection point of distributed photovoltaic power according to claim 4, characterized in that, S0-3-1. Determine the optimal hyperparameter combination for the bidirectional temporal convolutional network model, including: Determine the initial search interval for the optimal hyperparameters; An initial hyperparameter combination is generated based on the spatial density distribution function and the initial search interval of the optimal hyperparameters, and the initial hyperparameter combination is configured into the bidirectional temporal convolutional network model. The hyperparameters of the bidirectional temporal convolutional network model with the initial hyperparameter combination are iteratively optimized until the iteration termination condition is met, thus obtaining the optimal hyperparameter combination.

6. The method for predicting voltage at the grid connection point of distributed photovoltaic power according to claim 5, characterized in that, The hyperparameters of a bidirectional temporal convolutional network model with an initial hyperparameter combination are iteratively optimized until the iteration termination condition is met to obtain the optimal hyperparameter combination, including: The sample training dataset is divided into a first training set and a first test set; the sample training dataset includes sample voltage influence parameters and sample voltage sub-signal components. Using the sample voltage influence parameters of the first training set as model input and the sample voltage sub-signal components of the first training set as model output, the bidirectional temporal convolutional network model configured with the initial hyperparameter combination is trained to obtain an intermediate convolutional network model. The sample voltage influence parameters of the first test set are input into the intermediate convolutional network model to obtain the first predicted voltage sub-signal component; The hyperparameter fitness function is calculated based on the first predicted voltage sub-signal component and the sample voltage sub-signal component of the first test set. The correlation coefficients between hyperparameters and the network model are calculated based on the hyperparameter fitness function and the initial hyperparameter combination. The hyperparameter combination of the bidirectional temporal convolutional network model is dynamically adjusted based on the correlation coefficient between the hyperparameters and the network model until the hyperparameter fitness function value is less than or equal to a preset fitness threshold, or the number of iterations is greater than or equal to a preset number, thus obtaining the optimal hyperparameter combination.

7. The method for predicting voltage at the grid connection point of distributed photovoltaic power generation according to claim 4, characterized in that, Based on the sample voltage sub-signal components, and using a bidirectional temporal convolutional network model with optimal hyperparameter combinations, the features of the sample voltage sub-signal components are obtained, including: The sample voltage sub-signal component is input into the bidirectional temporal convolutional network model with the optimal hyperparameter combination. The forward and backward features of the sample voltage sub-signal component are obtained through the forward convolutional layer and the backward convolutional layer of the bidirectional temporal convolutional network model with the optimal hyperparameter combination, respectively. Calculate the forward and reverse feature weights of the sample voltage sub-signal components based on the sample voltage sub-signal components; Based on the forward feature weights and inverse feature weights, and using a variable weighted collaborative feature aggregation strategy, the forward features and inverse features are fused to obtain the sample voltage sub-signal component features.

8. The method for predicting voltage at the grid connection point of distributed photovoltaic power according to claim 7, characterized in that, Calculating the forward and reverse feature weights of the sample voltage sub-signal components based on the sample voltage sub-signal components includes: Calculate the mean and variance of the sample voltage sub-signal components; Based on the hyperbolic tangent function, a fully connected operation is performed on the mean and variance to calculate the weighted intermediate variables; The intermediate weight variables are normalized to obtain an adaptive weight vector; The forward and reverse feature weights of the sample voltage sub-signal components are determined based on the adaptive weight vector.

9. The method for predicting voltage at the grid connection point of distributed photovoltaic power according to claim 6, characterized in that, S0-4. Using the sample voltage influence parameters as model input, multiple sample voltage sub-signal components as model output, and the characteristics of the sample voltage sub-signal components as model constraints, perform network model training to obtain the preset grid connection point voltage prediction network model, including: S0-4-1. Divide the sample training dataset into a second training set and a second test set; the sample training dataset includes sample voltage influence parameters and sample voltage sub-signal components. S0-4-2. Using the sample voltage influence parameters of the second training set as the model input, the sample voltage sub-signal components of the second training set as the model output, and the sample voltage sub-signal component features output by the bidirectional temporal convolutional network model with the optimal hyperparameter combination as the constraint condition, the model is trained to obtain the intermediate prediction model. S0-4-3. Input the sample voltage influence parameters of the second test set into the intermediate prediction model to obtain the second predicted voltage sub-signal component; S0-4-4 Calculate the signal error between the second predicted voltage sub-signal component and the sample voltage sub-signal component of the second test set. If the signal error is greater than the signal error threshold, then proceed to step S0-4-5. S0-4-5. Adjust the model parameters of the intermediate prediction model according to the signal error and update the optimal hyperparameter combination of the intermediate convolutional network model. Then execute steps S0-4-2 to S0-4-4 until the signal error is less than or equal to the signal error threshold to obtain the preset grid connection point voltage prediction network model.

10. The method for predicting voltage at the grid connection point of distributed photovoltaic power generation according to claim 6, characterized in that, The hyperparameter combination includes: the learning rate of the bidirectional temporal convolutional network, the number of convolutional kernels of the bidirectional temporal convolutional network, and the L2 regularization coefficient of the bidirectional temporal convolutional network.

Images