Photovoltaic grid-connected risk assessment method based on multi-gated double convolutional neural network

Abstract

The present application relates to the technical field of intelligent power grid, especially to a photovoltaic grid-connected risk assessment method based on a multi-gate-double convolutional neural network, comprising the following steps: S1: obtaining power grid data: obtaining power grid data comprehensively covering dynamic characteristics of the power grid, and photovoltaic power station output power, load data and meteorological data; S2: constructing a feature engineering: preprocessing and feature extraction are performed on each data in S1 to form complete time series input data, and input time series data is defined; S3: constructing a power grid grid-connected risk assessment data model: features are extracted by using a double convolutional network, the extracted features include local features and global features, and the double convolutional network comprises two parallel convolutional paths; S4: constructing an assessment model of dynamic changes of risks, and the risks are divided into three levels of high, medium and low, and the present application improves the risk assessment precision.

Description

Technical Field

[0001] This invention relates to the field of smart grid technology, and in particular to a photovoltaic grid-connection risk assessment method based on a multi-gated dual convolutional neural network. Background Technology

[0002] With the advancement of global energy structure transformation, distributed photovoltaic (PV) power generation has gradually become an important component of renewable energy development. Distributed PV power generation systems are characterized by flexibility, environmental friendliness, and sustainability; however, their grid connection also faces a series of technical challenges, particularly risks related to grid stability, load fluctuations, and power quality. Therefore, accurate risk assessment and dynamic monitoring are crucial to ensuring the safe and efficient grid connection of distributed PV power generation systems.

[0003] Chinese invention patent application CN118798641A discloses a method, device, and electronic device for evaluating the operational status of a distributed photovoltaic grid-connected system, comprising: collecting a set of status indicator data from a distributed photovoltaic grid-connected system based on a preset grid-connected safety operation indicator system; preprocessing the status indicator data in the set of status indicator data to obtain a preprocessed set of status indicator data; constructing a status indicator vector based on the preprocessed set of status indicator data and obtaining a weight vector for the status indicator; inputting the status indicator vector and the weight vector into a neural network model to output the risk level of the distributed photovoltaic grid-connected system, wherein the neural network model is a pre-trained model for evaluating the operational risk of the distributed photovoltaic grid-connected system, and the neural network model consists of at least a self-attention network, a temporal convolutional network, and a long short-term memory network; and generating an operational status evaluation result of the distributed photovoltaic grid-connected system based on the risk level. The aforementioned disclosed solution can deeply mine the features in the data and mark the importance of each feature based on the weight values ​​configured for each feature, thereby improving the accuracy of the status assessment. This solves the technical problem in related technologies where the accuracy of the assessment results based on mathematical methods for evaluating the operating status of photovoltaic systems is low. This invention discloses a solution that is different from the above solution for improving the accuracy of photovoltaic grid connection risk assessment. Summary of the Invention

[0004] Technical objective: In order to overcome the shortcomings of the existing technology, this invention provides a photovoltaic grid connection risk assessment method based on multi-gated dual convolutional neural networks.

[0005] Technical Solution: To achieve the above objectives, this invention discloses a photovoltaic grid-connection risk assessment method based on a multi-gated dual convolutional neural network, comprising the following steps:

[0006] S1: Obtain grid data: Obtain grid data that comprehensively covers the dynamic characteristics of the power grid, as well as photovoltaic power plant output power, load data, and meteorological data;

[0007] S2: Feature Engineering: Preprocess and extract features from the data in S1 to form complete time series input data, defined as X∈R. T*F Where T is the time step, F is the feature dimension, and R is the power grid dynamic data;

[0008] S3: Construct a data model for grid connection risk assessment: Use a dual convolutional network to extract features, including extracting local features and extracting global features. The dual convolutional network includes two parallel convolutional paths.

[0009] S4: Construct a dynamic risk assessment model, classifying risks into three levels: high, medium, and low.

[0010] Furthermore, the extraction of local features in S3 includes the following steps:

[0011] A1: Extracting Local Time Series Features: Local time series features are extracted from the input time series data through one-dimensional convolution calculation. The output feature of the one-dimensional convolution calculation is H∈R. T*C The calculation formula is:

[0012]

[0013] in, For the convolution kernel, k t Where C is the time window size, b is the number of output channels, and c is the number of output channels. c σ is the bias of channel c, and σ(·) is the nonlinear activation function;

[0014] A2: Extracting Local Feature Information: Local feature information of the data is extracted through multi-layer convolution operations. The calculation formula is as follows:

[0015] H l =σ(W l *X+b l ),

[0016] Among them, W l The kernel is b, * represents the convolution operation, and b represents the convolution kernel. l Here, σ is the bias term, and σ is the activation function.

[0017] A3: Introduction of gating mechanism: After each convolutional layer in A2, a gating mechanism is introduced to control the output of each convolutional layer. The calculation formula is as follows:

[0018]

[0019] Among them, G l It is the output of the gating unit. It is the weight matrix of the gating unit. σ is the gating bias term, and σ is the activation function.

[0020] Furthermore, the method for extracting global features in S3 includes the following steps:

[0021] B1: Dilated Convolution and Global Pooling: The dilated convolution introduces a scaling factor d to increase the receptive field while preserving temporal resolution. The global pooling performs global pooling along the temporal dimension to extract global feature representations. The calculation formula for the dilated convolution is:

[0022]

[0023] Where H is the convolution calculation, d is the dilation factor that determines the convolution kernel step size, σ is the activation function, W is the weight matrix in the time dimension, X is the feature value of the power grid data, T is the time step, F is the feature dimension, and c is the output channel.

[0024] The formula for calculating global pooling is:

[0025] H global =Pooling(H),

[0026] Among them, H global The operation is a global convolution, and the pooling operation is either max pooling or average pooling, used to aggregate features in the time dimension. H represents a convolutional layer.

[0027] B2: Introduce multiple gating mechanism units to fuse and generate the final feature representation. The multiple gating mechanism units include time gating, spatial gating and environmental gating.

[0028] Furthermore, the time gating dynamically adjusts the time weights of features for important time points in the time series data. These important time points include the photovoltaic system startup time, the photovoltaic system shutdown time, and peak fluctuation periods. The formula for the time gating is:

[0029] G t =σ(W t ·X+b t ),

[0030] Where, X∈R T×F For input time series data, W t ∈R F×F b t ∈R F Let G be the weight matrix and bias vector for time gating, respectively, and σ be the sigmoid function used to constrain the weights to [0,1].t ∈R T×F Time-gated weights;

[0031] The adjusted time characteristics are:

[0032]

[0033] Where ⊙ represents element-wise multiplication;

[0034] The spatial gating is used to dynamically adjust the weights of different power station characteristic channels based on the interaction information between distributed photovoltaic power stations, thereby strengthening the correlation between power stations. The formula for the spatial gating is:

[0035] G s =σ(W s ·X+b s ),

[0036] Among them, W s ∈R F×F b s ∈R F These are the weight matrix and bias of the spatial gating, respectively, G t ∈R T×F Spatial gating weights;

[0037] The adjusted spatial characteristics are as follows:

[0038]

[0039] Where ⊙ represents element-wise multiplication;

[0040] The environmental gating system dynamically adjusts the weights of features to adapt to different environmental changes in response to the influence of environmental conditions. The environmental gating formula is as follows:

[0041] G e =σ(W e ·E+b e ),

[0042] Where, E∈R T×Q For environmental data, W e ∈R Q×F b e ∈R F G represents the weight matrix and bias for environment gating. e ∈R T ×F Environmental gating weights;

[0043] The adjusted environmental characteristics are as follows:

[0044]

[0045] Here, ⊙ represents element-wise multiplication.

[0046] Furthermore, the multi-gating mechanism unit fuses the features resulting from the time gating, spatial gating, and environmental gating to obtain the final adjusted feature representation. The multi-gating mechanism formula is as follows:

[0047]

[0048] Where, α t ,α s ,α e The weighting coefficients for time, space, and environment gating, respectively, X final ∈R T×F This represents the final feature representation after the action of multiple gating mechanisms.

[0049] Further, the risk level in step S4 is defined as Y:

[0050] Y∈{0,1,2}, where:

[0051] Constructing an assessment model for dynamic risk changes includes the following steps:

[0052] C1: Based on the operating status of the power grid, multi-dimensional indicators are selected as the input feature vector. These multi-dimensional indicators include voltage stability x1, frequency fluctuation x2, and power flow deviation x3. The input feature vector is:

[0053] X = {x1, x2, ..., x} n};

[0054] C2: A weighted cross-entropy loss function L is introduced to accurately evaluate high-risk samples. The formula for calculating the weighted cross-entropy loss function L is as follows:

[0055]

[0056] Where N is the total number of samples, y i For sample i, As the weight of the risk level, P(Y=y i |X i ;θ) represents the probability of model evaluation;

[0057] C3: Determine the model evaluation metrics to evaluate the model's performance. The model evaluation metrics include accuracy, recall for high-risk assessments, F1 score, comprehensive index, and objective function.

[0058] The accuracy rate is the proportion of all correctly predicted samples out of the total number of samples;

[0059] The recall formula for the high-risk assessment is:

[0060]

[0061] The F1 score is a weighted average of precision and recall, and the formula for calculating the F1 score is as follows:

[0062]

[0063] The comprehensive index incorporates evaluation weights for voltage stability, frequency fluctuation, and power flow deviation, generating a weighted score S, calculated using the following formula:

[0064] S = α·voltage stability score + β·frequency fluctuation score + γ·power flow deviation score

[0065] Where α, β, and χ are weight allocation coefficients, and α + β + γ = 1;

[0066] The objective function formula is:

[0067]

[0068] The present invention also discloses an apparatus for implementing the above-described photovoltaic grid-connected risk assessment method based on a multi-gated dual convolutional neural network, comprising:

[0069] The data acquisition module is configured as a power grid time-series data input module to acquire power grid data;

[0070] The network model building module is configured as a distributed power grid dynamic risk assessment model, and constructs a dual convolutional neural network model.

[0071] The gating weight optimization module is configured as the model weight optimization module, based on the distributed power grid risk network weight optimization module.

[0072] The risk analysis module is configured as a distributed power grid assessment model to construct an assessment model for dynamic risk changes.

[0073] The present invention also discloses a computer-readable storage medium for a photovoltaic grid-connection risk assessment method based on a multi-gated dual convolutional neural network, which stores multiple instructions, which are loaded and executed by the processor of a terminal device.

[0074] The present invention also discloses a terminal device for a photovoltaic grid-connected risk assessment method based on a multi-gated dual convolutional neural network, comprising a processor for implementing various instructions and a computer-readable storage medium for storing multiple instructions. When the processor executes the instructions in the computer-readable storage medium, it can complete the photovoltaic grid-connected risk assessment method based on the multi-gated dual convolutional neural network described above.

[0075] The beneficial effects of this invention are:

[0076] 1. The distributed photovoltaic grid-connected risk dynamic assessment method based on multi-gated dual convolutional neural network proposed in this invention can dynamically capture spatiotemporal features and improve assessment accuracy; the multi-gated mechanism can accurately and dynamically screen key features through time, space and feature fusion gating, which significantly enhances the ability to model spatiotemporal dependencies.

[0077] 2. The distributed photovoltaic grid-connected risk dynamic assessment method based on multi-gated dual convolutional neural networks of the present invention improves the assessment capability by focusing on key risks; introduces a weighted loss function to focus on the assessment accuracy of high-risk samples, thereby effectively identifying potential high-risk scenarios in photovoltaic grid-connected risk assessment and providing early warning support for grid operation; and dynamically adjusts the importance of features through feature fusion gating to avoid information redundancy and further improve the model's sensitivity to key risk factors.

[0078] 3. The distributed photovoltaic grid-connected risk dynamic assessment method based on multi-gated dual convolutional neural networks of the present invention has high computational efficiency and is suitable for real-time assessment; the parallel computing characteristics of convolutional networks improve the running speed of the model and can support real-time risk assessment of large-scale photovoltaic grid connection. Attached Figure Description

[0079] Figure 1 This is a schematic diagram of the overall method of the present invention;

[0080] Figure 2 This is a schematic diagram of the multi-gating mechanism unit in this invention;

[0081] Figure 3 This is a comparison chart of the training loss convergence speed of the model of this invention and other models;

[0082] Figure 4 The curves showing the mean squared error of the dual convolutional network algorithm in this invention are compared with those of other algorithms.

[0083] Figure 5 This is a comparison curve of the root mean square error between the dual convolutional network algorithm in this invention and other algorithms. Detailed Implementation

[0084] The following is in conjunction with the appendix Figure 1 To be continued Figure 5 The principles and features of the present invention are described, and the examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0085] Example 1: A photovoltaic grid-connection risk assessment method based on multi-gated dual convolutional neural networks, such as... Figure 1 As shown, it includes the following steps:

[0086] S1: Obtain power grid data: Obtain power grid data that comprehensively covers the dynamic characteristics of the power grid, as well as the output power, load data, and meteorological data of photovoltaic power plants; specifically, the power grid data includes real-time and historical power grid operation data obtained from the monitoring systems of enterprises such as State Grid and China Southern Power Grid, such as voltage, current, and frequency; the load data is the historical record of load power demand; and the meteorological data includes irradiance, temperature, and wind speed.

[0087] S2: Feature Engineering: Preprocessing and feature extraction are performed on the data in S1 to form complete time series input data, defined as X∈R. T*F Where T is the time step, F is the feature dimension, and R is the power grid dynamic data;

[0088] S3: Construct a data model for grid connection risk assessment: Use a dual convolutional network (DCN) to extract features, including extracting local features and extracting global features.

[0089] Extracting local features from input data can effectively capture short-term fluctuations, sudden changes, and local anomalies in power output, ensuring a complete expression of fine-grained dynamic characteristics and providing accurate basic data support for subsequent risk assessment.

[0090] Furthermore, extracting local features in S3 includes the following steps:

[0091] A1: Extracting Local Time Series Features: Local time series features are extracted from the time series input data through one-dimensional convolution. The output feature of the one-dimensional convolution is H∈R. T*C The calculation formula is:

[0092]

[0093] in, For convolution kernel; k t C is the time window size; C is the number of output channels; b c σ is the bias of channel c; σ(·) is a nonlinear activation function, such as ReLU;

[0094] Dual convolutional networks consist of two parallel convolutional paths, designed to extract local features. The purpose is to capture short-term dynamic changes in the input time series, such as short-term fluctuations and sudden changes in power output. Core design: One-dimensional convolution (Conv1D) is used to extract local time series features; small convolutional kernels (e.g., k...) are used... t =3) Extract fine-grained features; extract more complex local dynamic characteristics layer by layer through multi-layer stacking. Activation functions: ReLU or Swish, etc., are used to enhance nonlinear characteristics.

[0095] A2: Extracting Local Feature Information: Local feature information of the data is extracted through multi-layer convolution operations. The calculation formula is as follows:

[0096] H l =σ(W l *X+b l ),

[0097] Among them, W l It is the convolution kernel, * indicates the convolution operation, b l σ is the bias term, and σ is the activation function; this module extracts local feature information from the data through multi-layer convolution operations.

[0098] A3: Introduction of gating mechanism: After each convolutional layer in A2, a gating mechanism is introduced to control the output of each convolutional layer through gating units. The calculation formula is as follows:

[0099]

[0100] Among them, G l For the output of the gating unit, Here is the weight matrix of the gated unit. σ is the gated bias term, and σ is the activation function;

[0101] Gating mechanisms enable each layer to automatically select whether to activate specific convolutional features based on the different characteristics of the input data, thereby improving the expressive power of the model.

[0102] Furthermore, the purpose of extracting global features is to capture long-term trends and global patterns in time series, such as long-term power variations and seasonality. Core design elements include: Dilated Convolution: Increasing the receptive field by introducing a dilation factor while preserving temporal resolution; and Global Pooling: Performing global pooling across the temporal dimension to extract global feature representations.

[0103] Specifically, the method for extracting global features in S3 includes the following steps:

[0104] B1: Dilated Convolution and Global Pooling: Dilated convolution introduces a scaling factor d to increase the receptive field while preserving temporal resolution. Global pooling performs global pooling along the temporal dimension to extract global feature representations. The formula for calculating dilated convolution is:

[0105]

[0106] Where H represents convolution computation, d is the dilation factor that determines the convolution kernel jump step size, σ is the activation function, W is the weight matrix in the time dimension, X is the feature value of the power grid data, T is the time step size, F is the feature dimension, and c represents the output channel.

[0107] The formula for calculating global pooling is:

[0108] H global =Pooling(H),

[0109] Among them, H global represents global convolution operation, Pooling operation is max pooling or average pooling, used to aggregate features in the time dimension, and H represents convolutional layer.

[0110] B2: To extract and adjust features more precisely, multiple gating mechanisms are introduced and fused to generate the final feature representation. These gating mechanisms include temporal gate, spatial gate, and environmental gate. These gating mechanisms dynamically adjust the weights of features from the temporal, spatial, and environmental dimensions, respectively, and finally generate the final feature representation through fusion.

[0111] like Figure 2 The diagram illustrates the feature extraction process for various input data types, including temporal data (which may include periodic features such as circadian rhythms), environmental data (which may include temperature and humidity changes), and spatial data (which may include location distribution or proximity). The core mechanism is represented by a "gating unit" or "dynamic adjustment module," which internally demonstrates how different data characteristics are filtered or weighted through weight allocation or conditional judgments. The output is the processed comprehensive weights or decision information for downstream applications.

[0112] Furthermore, time gating dynamically adjusts the time weights of features for important time points in time series data. These important time points include the photovoltaic system startup time, the photovoltaic system shutdown time, and peak fluctuation periods. The formula for time gating is:

[0113] G t =σ(W t ·X+b t ),

[0114] Where, X∈R T×F For input time series data, W t ∈R F×F b t ∈R F Let G be the weight matrix and bias vector for time gating, respectively, and σ be the sigmoid function used to constrain the weights to [0,1]. t ∈R T×F Time-gated weights;

[0115] The adjusted time features are dynamically adjusted using gating weights to control the input features of the time dimension.

[0116]

[0117] Where ⊙ represents element-wise multiplication;

[0118] Furthermore, spatial gating dynamically adjusts the weights of characteristic channels of different distributed photovoltaic power stations to enhance the correlation between them, based on the interaction information between the power stations. The formula for spatial gating is:

[0119] G s =σ(W s ·X+b s ),

[0120] Among them, W s ∈R F×F b s ∈R F These are the weight matrix and bias of the spatial gating, respectively, G t ∈R T×F Spatial gating weights;

[0121] The adjusted spatial characteristics are dynamically weighted according to spatial gating weights for different power plant characteristics:

[0122]

[0123] Where ⊙ represents element-wise multiplication;

[0124] Furthermore, environmental gating dynamically adjusts the weights of features to adapt to different environmental changes, taking into account the impact of environmental conditions such as meteorological data and weather variations. The environmental gating formula is as follows:

[0125] G e =σ(W e ·E+b e ),

[0126] Where, E∈R T×Q For environmental data, such as temperature, light intensity, wind speed, etc.; W e ∈R Q×F b e ∈R F G represents the weight matrix and bias for environment gating. e ∈R T×F Environmental gating weights;

[0127] The adjusted environmental features are then used to correct the feature extraction results through environmental gating weights.

[0128]

[0129] Where ⊙ represents element-wise multiplication;

[0130] Furthermore, the multi-gating mechanism unit fuses the features resulting from the time gating, spatial gating, and environmental gating to obtain the final adjusted feature representation. The multi-gating mechanism formula is as follows:

[0131]

[0132] Where, α t ,α s ,α e These are the weighting coefficients for time, space, and environment gating, respectively. These weighting coefficients can be set via attention mechanisms or hyperparameters; X final ∈R T×F This represents the final feature representation after the action of multiple gating mechanisms.

[0133] S4: Construct a dynamic risk assessment model, classifying risks into three levels: high, medium, and low. Assessment indicators: Combine multi-dimensional indicators such as grid voltage stability, frequency fluctuation, and power flow deviation for comprehensive evaluation.

[0134] Define the risk level in step S4 as Y, where risk level Y is a multi-class classification problem:

[0135] Y∈{0,1,2}, where:

[0136] Constructing an assessment model for dynamic risk changes includes the following steps:

[0137] C1: Based on the operating status of the power grid, multi-dimensional indicators are selected as the input feature vector. These multi-dimensional indicators include voltage stability x1, frequency fluctuation x2, and power flow deviation x3. The input feature vector is:

[0138] X = {x1, x2, ..., x} n};

[0139] Among them, voltage stability x1 is quantified using voltage stability margin (VSM) or maximum load capacity; frequency fluctuation x2 is the mean or standard deviation of frequency deviation; and power flow deviation x3 is the difference between power flow and the optimal distribution. The goal is to learn the dynamic patterns of risk changes from these characteristics and evaluate the classification of risk level Y using a multi-classification algorithm.

[0140] C2: Introduce the weighted cross-entropy loss function L to accurately evaluate high-risk samples. The formula for calculating the weighted cross-entropy loss function L is:

[0141]

[0142] Where N is the total number of samples; y i This represents the true risk level of sample i. Assuming w2 > w1 > w0 as the weights for risk levels, high-risk samples have the largest weights, for example: w0 = 1, w1 = 2, w2 = 5; P(Y = y i |X i ;θ) represents the probability of model evaluation;

[0143] By adjusting the weights The model will tend to improve the accuracy of assessments for high-risk samples.

[0144] C3: Determine the model evaluation metrics to evaluate the model's performance. To evaluate the model's performance, the model evaluation metrics include accuracy, recall for high-risk assessments, F1 score, overall performance metrics, and objective function.

[0145] Furthermore, the accuracy rate is the classification accuracy rate for all risk levels, i.e., the proportion of all correctly predicted samples out of the total number of samples;

[0146] The formula for recall in high-risk assessment is:

[0147]

[0148] Furthermore, the F1 score is a weighted average of precision and recall. The formula for calculating the F1 score (F1-5 cores) is as follows:

[0149]

[0150] Furthermore, the comprehensive index incorporates evaluation weights for voltage stability, frequency fluctuation, and power flow deviation, generating a weighted score S, calculated using the following formula:

[0151] S = α·voltage stability score + β·frequency fluctuation score + γ·power flow deviation score

[0152] Where α, β, and χ are weight allocation coefficients, and α + β + γ = 1;

[0153] Furthermore, the objective function of the model is the weighted cross-entropy loss function for risk assessment. Combining the aforementioned multi-dimensional indicators, the goal is to output the predicted risk level based on a comprehensive assessment, and the model performance is validated through analysis of various dimensions.

[0154] The objective function formula is:

[0155]

[0156] By training and optimizing the objective function, the model will be able to dynamically assess the risk level of the power grid and provide corresponding risk warnings based on the grid's operating status, such as voltage stability, frequency fluctuations, and power flow deviations.

[0157] like Figure 3 — Figure 5 As shown, the risk assessment method proposed in this invention can effectively improve the accuracy of risk assessment results.

[0158] Example 2: An apparatus for implementing a photovoltaic grid-connected risk assessment method based on a multi-gated dual convolutional neural network, comprising:

[0159] The data acquisition module is configured as a power grid time-series data input module to acquire power grid data;

[0160] The model building module is configured as a distributed power grid dynamic risk assessment model, and constructs a dual convolutional neural network model.

[0161] The gating weight optimization module is configured as the model weight optimization module, based on the distributed power grid risk network weight optimization module.

[0162] The risk analysis module is configured as a distributed power grid assessment model to construct an assessment model for dynamic risk changes.

[0163] Example 3: A computer-readable storage medium for a photovoltaic grid-connection risk assessment method based on a multi-gated dual convolutional neural network, which stores multiple instructions, which are loaded and executed by the processor of a terminal device.

[0164] Example 4: A terminal device for a photovoltaic grid-connection risk assessment method based on a multi-gated dual convolutional neural network includes a processor for implementing each instruction and a computer-readable storage medium for storing multiple instructions. When the processor runs the instructions in the computer-readable storage medium, it can complete the above-mentioned photovoltaic grid-connection risk assessment method based on a multi-gated dual convolutional neural network, which has greater versatility and flexibility.

[0165] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A photovoltaic grid-connection risk assessment method based on multi-gated dual convolutional neural networks, characterized in that: Includes the following steps: S1: Obtain grid data: Obtain grid data that comprehensively covers the dynamic characteristics of the power grid, as well as photovoltaic power plant output power, load data, and meteorological data; S2: Feature Engineering: Preprocess and extract features from the data in S1 to form complete time series input data. Define the input time series data as... Where T is the time step, F is the feature dimension, and R is the power grid dynamic data; S3: Construct a data model for grid connection risk assessment: Use a dual convolutional network to extract features, including extracting local features and extracting global features. The dual convolutional network includes two parallel convolutional paths. The extraction of local features in S3 includes the following steps: A1: Extracting Local Time Series Features: Local time series features are extracted from the input time series data using one-dimensional convolution calculation. The output features of the one-dimensional convolution calculation are... The calculation formula is: ， in, For convolution kernel, The size of the time window. Number of output channels aisle The bias, It is a non-linear activation function; A2: Extracting Local Feature Information: Local feature information of the data is extracted through multi-layer convolution operations. The calculation formula is as follows: ， in, For convolution kernel, For convolution operations, For bias terms, For activation functions; A3: Introduction of gating mechanism: After each convolutional layer in A2, a gating mechanism is introduced to control the output of each convolutional layer. The calculation formula is as follows: ， in, It is the output of the gating unit. It is the weight matrix of the gated unit. It is a gating bias term. It is an activation function; The method for extracting global features in S3 includes the following steps: B1: Dilated Convolution and Global Pooling: The dilated convolution introduces a scaling factor d to increase the receptive field while preserving temporal resolution. The global pooling performs global pooling along the temporal dimension to extract global feature representations. The calculation formula for the dilated convolution is: ， Where H represents convolution calculation, The expansion factor determines the stride of the convolution kernel. The activation function is W, the weight matrix in the time dimension is X, the feature value of the power grid data is T, the time step is F, the feature dimension is c, and the output channel is c. The formula for calculating global pooling is: ， Among them, H global The operation is a global convolution, and the pooling operation is either max pooling or average pooling, used to aggregate features in the time dimension. H is a convolutional layer. B2: Introduce multiple gating mechanism units to fuse and generate the final feature representation. The multiple gating mechanism units include time gating, spatial gating, and environmental gating. The time gating dynamically adjusts the time weights of features for important time points in time series data. These important time points include the photovoltaic system startup time, the photovoltaic system shutdown time, and peak fluctuation periods. The formula for the time gating is: ， in, To input time series data, These are the weight matrix and bias vector for time gating, respectively. The sigmoid function is used to constrain the weights to [0,1]. Time-gated weights; The adjusted time characteristics are: , in, This indicates element-wise multiplication; The spatial gating is used to dynamically adjust the weights of different power station characteristic channels based on the interaction information between distributed photovoltaic power stations, thereby strengthening the correlation between power stations. The formula for the spatial gating is: , in, These are the weight matrix and bias of the spatial gating, respectively. Spatial gating weights; The adjusted spatial characteristics are as follows: , in, This indicates element-wise multiplication; The environmental gating system dynamically adjusts the weights of features to adapt to different environmental changes in response to the influence of environmental conditions. The environmental gating formula is as follows: , in, For environmental data, For the weight matrix and bias of the environmental gating, Environmental gating weights; The adjusted environmental characteristics are as follows: , in, This indicates element-wise multiplication; The multi-gating mechanism unit fuses the features resulting from the time-gating, spatial-gating, and environmental-gating effects to obtain the final adjusted feature representation. The multi-gating mechanism formula is as follows: , in, These are the weighting coefficients for time, space, and environment gating, respectively. This represents the final feature representation after the action of multiple gating mechanisms; S4: Construct a dynamic risk assessment model, classifying risks into three levels: high, medium, and low.

2. The photovoltaic grid-connection risk assessment method based on a multi-gated dual convolutional neural network according to claim 1, characterized in that, Define the risk level in step S4 as Y: ; Constructing an assessment model for dynamic risk changes includes the following steps: C1: Based on the operating status of the power grid, multi-dimensional indicators are selected as input feature vectors, including voltage stability. Frequency fluctuation , flow deviation The input feature vector is: ; C2: A weighted cross-entropy loss function L is introduced to accurately evaluate high-risk samples. The formula for calculating the weighted cross-entropy loss function L is as follows: , in, The total number of samples, For the sample The true risk level, As the weight of risk level, The probability evaluated by the model; C3: Determine the model evaluation metrics to evaluate the model's performance. The model evaluation metrics include accuracy, recall for high-risk assessments, F1 score, comprehensive index, and objective function. The accuracy rate is the proportion of all correctly predicted samples out of the total number of samples; The recall formula for the high-risk assessment is: ; The F1 score is a weighted average of precision and recall, and the formula for calculating the F1 score is as follows: ; The comprehensive index incorporates evaluation weights for voltage stability, frequency fluctuation, and power flow deviation, generating a weighted score S, calculated using the following formula: ,in, For weighting coefficients, ; The objective function formula is: .

3. An apparatus for implementing the photovoltaic grid-connected risk assessment method based on a multi-gated dual convolutional neural network as described in any one of claims 1-2, characterized in that, include: The data acquisition module is configured as a power grid time-series data input module to acquire power grid data; The network model building module is configured as a distributed power grid dynamic risk assessment model, and constructs a dual convolutional neural network model. The gating weight optimization module is configured as the model weight optimization module, based on the distributed power grid risk network weight optimization module. The risk analysis module is configured as a distributed power grid assessment model to construct an assessment model for dynamic risk changes.

4. A computer-readable storage medium for a photovoltaic grid-connection risk assessment method based on a multi-gated dual convolutional neural network, characterized in that, It stores multiple instructions, which are loaded and executed by the processor of the terminal device as described in any one of claims 1-2, for the photovoltaic grid-connected risk assessment method based on a multi-gated dual convolutional neural network.

5. A terminal device for a photovoltaic grid-connected risk assessment method based on a multi-gated dual convolutional neural network, characterized in that, The method includes a processor for implementing the instructions and a computer-readable storage medium for storing multiple instructions. When the processor executes the instructions in the computer-readable storage medium, it can complete the photovoltaic grid-connected risk assessment method based on a multi-gated dual convolutional neural network as described in any one of claims 1-2.

Images