A risk assessment method and system for new energy photovoltaic power generation access to a power distribution network

Abstract

The application discloses a new energy photovoltaic power generation access power distribution network risk assessment method and system, belongs to the power distribution network operation safety technical field, including: collecting illumination resource data and power distribution network load data, through empirical mode decomposition, Copula function and improved Latin hypercube sampling to generate basic data set covering typical and extreme scenarios; a coupling evaluation model combining graph neural network and risk propagation theory is constructed, an improved CRITIC-entropy weight method is used to dynamically weight multi-dimensional risk indexes, and a time-varying comprehensive risk grade matrix is output; the risk characteristics are analyzed and the pre-constructed measure risk knowledge graph is matched, a dynamic self-adaptive response scheme is generated through conflict checking and priority sorting; the scheme is verified in a digital twin power distribution network simulation model, and when the scheme does not meet the standard, the model parameters and the basic data set are updated through incremental data feedback; the application improves the comprehensiveness and timeliness of risk assessment.

Description

Technical Field

[0001] This invention belongs to the field of power distribution network operation safety technology, specifically a risk assessment method and system for connecting new energy photovoltaic power generation to the power distribution network. Background Technology

[0002] The proportion of new energy photovoltaic (PV) power generation in the energy structure continues to increase, and the connection of a large number of PV power plants to the distribution network has become an inevitable trend in energy transformation. However, PV power generation is characterized by strong intermittency, volatility, and randomness. Its connection will change the power flow distribution, voltage level, and short-circuit current characteristics of the distribution network, bringing many risks to the safe and stable operation of the distribution network, such as voltage exceeding limits, power imbalance, and harmonic pollution. Existing risk assessment methods for PV connection to the distribution network have many shortcomings: First, traditional assessment methods are mostly based on static data and linear models, which are difficult to adapt to the dynamic fluctuation characteristics of PV power generation, resulting in a large deviation between the assessment results and the actual operating conditions; Second, existing models mostly consider the single factor of PV output or distribution network load in isolation, without constructing a coupling correlation mechanism between the two and the grid absorption capacity, and cannot accurately capture the risk evolution law under the synergistic effect of multiple factors; Third, risk response plans are mostly preset fixed strategies, lacking dynamic adaptive adjustment capabilities, and are difficult to match real-time changes in risks. Therefore, there is an urgent need for a method that can accurately couple the multi-dimensional dynamic characteristics of photovoltaic power generation and distribution network, reduce the risk of curtailment of photovoltaic power, and realize dynamic risk assessment and adaptive response, so as to ensure the safe and efficient access of new energy photovoltaic power generation and the stable operation of distribution network. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a risk assessment method and system for integrating new energy photovoltaic power generation into the distribution network. It collects solar resource data and distribution network load data, and generates a basic dataset covering typical and extreme scenarios through empirical mode decomposition, Copula function, and improved Latin hypercube sampling. A coupled assessment model integrating graph neural networks and risk propagation theory is constructed, and an improved CRITIC-entropy weighting method is used to dynamically weight multi-dimensional risk indicators, outputting a time-varying comprehensive risk level matrix. Risk characteristics are analyzed and matched with a pre-constructed risk knowledge graph of measures, and dynamic adaptive response schemes are generated through conflict verification and priority ranking. The schemes are verified in a digital twin distribution network simulation model, and if the targets are not met, incremental data feedback is used to update model parameters and the basic dataset. This invention improves the comprehensiveness and timeliness of risk assessment.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A risk assessment method for connecting new energy photovoltaic power generation to the power distribution network includes:

[0006] S1: Collect solar energy resource data from photovoltaic power plants and load data from distribution networks to form a basic dataset; the basic dataset includes photovoltaic output ranges, load levels, and corresponding grid absorption capacity boundaries under different operating scenarios;

[0007] S2: Based on the aforementioned basic dataset, construct a coupled assessment model that integrates graph neural networks and risk propagation theory, and output a time-varying comprehensive risk level matrix;

[0008] S3: Match the time-varying comprehensive risk level matrix with the pre-constructed risk knowledge graph of measures, and generate a dynamic adaptive response plan based on the matching results;

[0009] S4: Verify the dynamic adaptive response scheme in a digital twin-based power distribution network simulation model. If the verification result does not meet the preset standard, use the incremental data generated by the simulation verification as feedback to update the parameters of the coupled evaluation model and simultaneously optimize the basic dataset.

[0010] Specifically, the formation process of the basic dataset is as follows:

[0011] S1.1: Collect historical solar irradiance and solar power plant time series data as solar resource data, and simultaneously collect historical active and reactive load time series data of each node in the distribution network as load data. Perform empirical mode decomposition on the solar resource data and load data respectively, decomposing each data sequence into intrinsic mode functions and a residual component. The intrinsic mode functions are components extracted in descending order of instantaneous frequency to characterize the oscillation modes of different scales of the data. The residual component is the monotonic or constant sequence remaining after decomposition, and the residual component is defined as the trend component.

[0012] S1.2: Select components whose instantaneous frequency matches the daily and annual cycles from the intrinsic mode functions, and superimpose and reconstruct them into periodic components. Superimpose the remaining high-frequency intrinsic mode functions into random components. Based on the trend components, periodic components, and random components, use the Copula function to establish a joint probability distribution model of photovoltaic power output and load.

[0013] S1.3: Based on the joint probability distribution model, the Latin hypercube sampling technique is used to sample under the condition of satisfying the grid capacity constraint, and generate a set of probabilistic scenarios covering typical and extreme operating scenarios, i.e., the basic dataset. Each probabilistic scenario consists of a set of photovoltaic power output values ​​and load values, and has a corresponding probability of occurrence.

[0014] Specifically, the method of establishing a joint probability distribution model of photovoltaic output and load based on trend components, periodic components, and random components using the Copula function includes:

[0015] S1.2.1: The trend component and the periodic component are superimposed to form a deterministic sequence of photovoltaic power output and a deterministic sequence of load. The random component is defined as a random sequence of photovoltaic power output and a random sequence of load.

[0016] S1.2.2: Fit the marginal probability distributions of the photovoltaic output random sequence and the load random sequence respectively, determine their respective optimal probability distribution models and parameters, and thus obtain the marginal probability distributions of photovoltaic output randomness and load randomness.

[0017] S1.2.3: Based on the fitted photovoltaic output randomness marginal probability distribution and load randomness marginal probability distribution, the observed values ​​of the photovoltaic output randomness sequence and load randomness sequence are transformed into pseudo-observed values ​​that follow a uniform distribution of [0,1] through probability integral transformation. The pseudo-observed values ​​are used to select the optimal Copula function from the preset Copula function family by calculating the maximum likelihood value and estimate its parameters.

[0018] S1.2.4: By using the selected optimal Copula function to connect the random marginal probability distribution of photovoltaic power output and the random marginal probability distribution of load, a joint random probability distribution model of photovoltaic power output and load is constructed.

[0019] S1.2.5: Integrate the deterministic sequence of photovoltaic power output and the deterministic sequence of load with the established stochastic joint probability distribution model to finally establish a joint probability distribution model of photovoltaic power output and load.

[0020] Specifically, the specific steps of S1.3 include:

[0021] S1.3.1: Based on the joint probability distribution model of photovoltaic power output and load, the Latin hypercube sampling technique is used to generate initial sampling points that satisfy the marginal probability distribution but are independent of each other for the two variables of photovoltaic power output and load.

[0022] S1.3.2: The sorting of the sample points is adjusted by using the rank correlation coefficient matching technique, and the sample points adjusted by correlation are mapped back to the data space composed of the original photovoltaic power output value and load value through their respective inverse marginal distribution functions to obtain the original set of numerical sample points.

[0023] S1.3.3: In the original set of numerical sample points, the power grid capacity constraint is introduced as a filter. Each sample is checked to see if it meets the node voltage constraint, line power constraint, transformer capacity constraint and distributed power output constraint. All samples that violate any constraint are removed. The remaining scenario samples constitute the probabilistic scenario set, which is the basic dataset that ultimately covers typical and extreme scenarios.

[0024] Specifically, the calculation process of the coupling evaluation model includes:

[0025] S2.1: The generation nodes, load nodes, and substation nodes in the distribution network are modeled as graph nodes, and the transmission lines and transformer branches are modeled as edges. Each edge is configured with a weight reflecting the line impedance to construct a graph neural network model. The operation of each layer of the graph neural network is defined as follows: each node aggregates the feature information of its one-hop neighbor nodes through the connecting edges. Each node concatenates the aggregated neighbor information with its own feature representation from the previous layer and inputs it into a fully connected layer and activation function to update and generate its own feature representation in the current layer. The feature information includes the voltage, injected power, and risk state value of the neighbor nodes.

[0026] S2.2: By stacking multiple layers of graph convolutional layers, the final feature representation of each node is fused with the global information of its multi-hop neighbors;

[0027] S2.3: For each operating scenario in the generated basic dataset, use the trained graph neural network model to calculate multi-dimensional node risk indicators for each node in the distribution network; the multi-dimensional node risk indicators include curtailment risk indicators, power quality risk indicators, and power grid safety and stability risk indicators.

[0028] S2.4: For the multi-dimensional node risk indicators under each operating scenario, the improved CRITIC-entropy weighting method is used to dynamically assign weights to the curtailment risk indicators, power quality risk indicators and grid security and stability risk indicators, and the weighted summation of each node is used to obtain the comprehensive risk value.

[0029] S2.5: The collection of comprehensive risk values ​​of all nodes under all scenarios and at all times constitutes a time-varying comprehensive risk level matrix.

[0030] Specifically, the time-varying comprehensive risk level matrix is ​​a three-dimensional data structure. Its first dimension index represents the node number in the distribution network, the second dimension index represents the discretized time point, and the third dimension stores the comprehensive risk value of the node at the corresponding time point and the decomposed values ​​of various risk indicators.

[0031] The generation process of the time-varying comprehensive risk level matrix is ​​as follows: for each probabilistic scenario in the basic dataset, the risk value of all nodes in the scenario at all time points is calculated using the coupled evaluation model. Then, a weighted average is performed based on the probability of each scenario in the basic dataset to finally obtain the time-varying comprehensive risk level matrix that reflects the overall risk level.

[0032] Specifically, the steps of S3 include:

[0033] S3.1: Analyze the time-varying comprehensive risk level matrix, extract risk features from it, and transform it into a standardized graph query statement; the risk features include the risk type, risk level, and risk impact range characteristics at a specified time, at a specified node, or in a specified area;

[0034] S3.2: In the pre-constructed risk knowledge graph of measures, execute the graph query statement to obtain a set of candidate response and disposal measures; the graph traversal query first locates the historical risk case node with the highest similarity to the current risk feature, and then finds all response and disposal measure nodes connected to the historical risk case node along the adoption measure relationship edge in the risk knowledge graph of measures;

[0035] S3.3: Perform conflict verification and priority ranking on the set of candidate response measures, and generate a dynamic adaptive response plan that includes the execution order of specific measures, the timing of activation, the expected effect, and the resource allocation plan based on the priority ranking result; the conflict verification is to determine whether different measures will compete for the same control resources or produce offsetting effects when they are executed; the priority ranking is a comprehensive score based on the implementation cost of the measures, the speed of the expected effect decay risk, and the current real-time operating status of the power grid.

[0036] Specifically, the pre-constructed risk knowledge graph includes at least three types of entity nodes: risk event case nodes, response and handling measure nodes, and power grid equipment nodes. The attributes of the risk event case nodes include the case occurrence time, risk type, risk level, affected power grid area, and weather conditions. The attributes of the response and handling measure nodes include the measure type, required control resources, implementation cost, expected effect, and effective delay time. The attributes of the power grid equipment nodes include equipment ID, equipment type, and equipment capacity.

[0037] The entity nodes are connected by relational edges, including: the impact relationship between risk event cases and the affected power grid equipment; the adoption relationship between risk event cases and the response and handling measures taken; and the interaction relationship between the response and handling measures and the operated power grid equipment.

[0038] Specifically, the process of validating the dynamic adaptive response scheme in a digital twin-based distribution network simulation model includes:

[0039] S4.1: Synchronously inject the current operating scenario data and the measures and instructions in the dynamic adaptive response scheme from the basic dataset into the distribution network digital twin simulation model to complete the initialization;

[0040] S4.2: Run the initialized digital twin simulation model of the distribution network to simulate the dynamic response process of the distribution network within a preset time period after the execution of the dynamic adaptive response scheme. Record key performance indicators during the simulation process. The key performance indicators include system voltage deviation, frequency fluctuation, curtailed solar power, and equipment load rate.

[0041] S4.3: Compare the simulated key performance indicators with the preset standard thresholds for each indicator. If all key performance indicators are better than the standard thresholds, the verification is deemed successful. If any key performance indicator exceeds the standard threshold, the verification is deemed unsuccessful. When the verification fails, the incremental data generated by the simulation is recorded. The incremental data includes voltage over-limit records, power over-limit records, power quality indicators, equipment operating status, and system stability indicators.

[0042] A risk assessment system for integrating new energy photovoltaic power generation into the power distribution network, the system comprising: a data construction module, a coupled assessment model module, a scheme generation module, and a verification and optimization module;

[0043] The data construction module is used to collect solar resource data and power distribution network load data, and generate a probabilistic scenario set covering typical and extreme operating scenarios through data decomposition, joint probabilistic modeling, and constraint sampling operations.

[0044] The coupled assessment model module integrates graph neural networks and risk propagation theory to perform risk calculations on various operating scenarios in the basic dataset and finally outputs a time-varying comprehensive risk level matrix.

[0045] The solution generation module, based on the time-varying comprehensive risk level matrix, matches it with a pre-constructed risk knowledge graph of measures, and generates a dynamic adaptive response solution through conflict verification and priority sorting.

[0046] The verification and optimization module is used to verify the effectiveness of the dynamic adaptive response scheme through a digital twin simulation model. For schemes that do not meet the preset standards, the incremental data generated during the simulation process is used as feedback to update the parameters of the coupled evaluation model and optimize the basic dataset.

[0047] Compared with the prior art, the beneficial effects of the present invention are:

[0048] 1. By accurately extracting the trend, periodic and random components of solar resources and load data through empirical mode decomposition, a joint probability distribution model is constructed by combining the Copula function, and a Latin hypercube sampling technique is adopted to generate a probabilistic scenario set covering typical and extreme scenarios under the premise of satisfying grid capacity constraints. This ensures the comprehensiveness and representativeness of data coverage and effectively captures the complex correlation characteristics between photovoltaic output and load.

[0049] 2. By integrating graph neural networks and risk propagation theory, the nodes and branches of the distribution network are modeled as graph structures. Global information is aggregated through multi-layer convolution, and an improved CRITIC-entropy weighting method is used to dynamically assign weights to multi-dimensional risk indicators. This not only comprehensively considers various risks such as curtailment of solar power, power quality, and grid security and stability, but also accurately presents the risk status of different nodes at different times through a time-varying comprehensive risk level matrix, thereby improving the comprehensiveness, timeliness, and accuracy of risk assessment.

[0050] 3. Based on a pre-built risk knowledge graph, candidate measures are quickly matched by analyzing risk characteristics. After conflict verification, situations of resource competition or offsetting effects are eliminated. Priority is ranked by factors such as implementation cost and effect decay rate, clarifying the execution order, start time and resource allocation plan. It can dynamically adapt to different risk scenarios and effectively solve the problems of traditional response measures lacking pertinence and low execution efficiency.

[0051] 4. By simulating the dynamic response of the power grid after the implementation of the response plan through a digital twin model, the key performance indicators are accurately verified to ensure that they meet the standards. If they do not meet the standards, the parameters of the coupled evaluation model and the basic dataset are updated with incremental simulation data, which ensures the effectiveness of the response plan. At the same time, it promotes the continuous iteration and optimization of the evaluation model and dataset, enabling the entire risk assessment system to have self-improvement capabilities, improving its reliability and adaptability in long-term applications, and ensuring the safe and stable operation of new energy photovoltaic power generation after it is connected to the distribution network. Attached Figure Description

[0052] Figure 1 This is a schematic diagram of a risk assessment method for connecting new energy photovoltaic power generation to the power distribution network according to the present invention;

[0053] Figure 2 This is a flowchart illustrating the principle of a risk assessment method for connecting new energy photovoltaic power generation to the power distribution network according to the present invention.

[0054] Figure 3 This is a system architecture diagram of a risk assessment system for connecting new energy photovoltaic power generation to the power distribution network according to the present invention. Detailed Implementation

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0056] Example 1:

[0057] like Figure 1 and Figure 2As shown in the illustration, a specific embodiment of the present invention discloses a risk assessment method for connecting new energy photovoltaic power generation to a distribution network. This embodiment takes a 10kV distribution network in an industrial park in East China as the application object. This distribution network contains 30 nodes, of which node #1 is a substation outgoing node, nodes #25-26 are photovoltaic grid connection points, and the rest are load nodes. A distributed photovoltaic power station with a capacity of 20MW is connected. Here, kV represents kilovolts and MW represents megawatts. The region has a subtropical monsoon climate, and the solar resources have significant seasonal and diurnal cycle characteristics. The distribution network load is mainly industrial load, including some commercial load, and there are obvious peak-valley fluctuations. Based on this scenario, this embodiment elaborates on the complete implementation process of risk assessment.

[0058] The risk assessment method in this embodiment specifically includes:

[0059] S1: Collect solar energy resource data from photovoltaic power plants and load data from distribution networks to form a basic dataset; the basic dataset includes photovoltaic output ranges, load levels, and corresponding grid absorption capacity boundaries under different operating scenarios;

[0060] S2: Based on the aforementioned basic dataset, construct a coupled assessment model that integrates graph neural networks and risk propagation theory, and output a time-varying comprehensive risk level matrix;

[0061] S3: Match the time-varying comprehensive risk level matrix with the pre-constructed risk knowledge graph of measures, and generate a dynamic adaptive response plan based on the matching results;

[0062] S4: Verify the dynamic adaptive response scheme in a digital twin-based power distribution network simulation model. If the verification result does not meet the preset standard, use the incremental data generated by the simulation verification as feedback to update the parameters of the coupled evaluation model and simultaneously optimize the basic dataset.

[0063] Furthermore, the preset standards include, but are not limited to: all node voltages remaining within safe limits during the simulation process, all line and transformer load rates not exceeding their long-term current carrying capacity, and system frequency maintained within the rated range; if all simulation results meet the preset standards, the verification is deemed successful; if any indicator is not met, the verification is deemed unsuccessful, and the specific indicator that failed, the time of occurrence, and the location of the equipment are recorded. These records are collectively referred to as the incremental data generated by the simulation verification.

[0064] The formation process of the basic dataset is as follows:

[0065] S1.1: Collect historical solar irradiance and solar power plant time series data as solar resource data, and simultaneously collect historical active and reactive load time series data of each node in the distribution network as load data. Perform empirical mode decomposition on the solar resource data and load data respectively, decomposing each data sequence into intrinsic mode functions and a residual component. The intrinsic mode functions are components extracted in descending order of instantaneous frequency to characterize the oscillation modes of different scales of the data. The residual component is the monotonic or constant sequence remaining after decomposition, and the residual component is defined as the trend component.

[0066] Furthermore, the specific steps for performing empirical mode decomposition on the illumination resource data and load data, decomposing each data sequence into intrinsic mode functions and a residual component, include:

[0067] (1) The collected light resource data and load data are respectively used as time series signals to be decomposed, and the current signal to be processed is defined as the original signal. Traverse the original signal and identify all its local maxima and local minima.

[0068] (2) Using the cubic spline interpolation method, interpolation fitting is performed on the identified local maxima and local minima respectively to obtain the upper envelope and lower envelope of the signal. The upper envelope connects all local maxima and the lower envelope connects all local minima.

[0069] (3) Calculate the average value of the upper envelope and the lower envelope to obtain the mean envelope. Subtract the mean envelope from the original signal to obtain an intermediate signal.

[0070] (4) Determine whether the intermediate signal satisfies the two conditions of the intrinsic mode function:

[0071] The number of extreme points is equal to or differs from the number of zero-crossing points by at most one;

[0072] At any given time, the average of the upper envelope formed by local maxima and the lower envelope formed by local minima is zero.

[0073] (5) If the condition is not met, the intermediate signal is used as the new original signal, and (1)-(3) are repeated until the condition of the intrinsic mode function is met;

[0074] (6) When the intermediate signal satisfies the condition of the intrinsic mode function, it is taken as the first intrinsic mode function component. Then the first intrinsic mode function component is subtracted from the original signal to obtain the remaining signal. The remaining signal is then taken as the new original signal. (1)-(4) are repeated to extract the second intrinsic mode function component. The process is repeated iteratively until the remaining signal becomes a monotonic function or a constant function and no new intrinsic mode function can be extracted.

[0075] (7) Arrange all extracted intrinsic mode functions in descending order of instantaneous frequency. The component with the higher instantaneous frequency represents the fast fluctuation component of the signal, and the component with the lower instantaneous frequency represents the slow trend component of the signal. Take the remaining monotonic function or constant function as the residual component. This residual component represents the long-term trend or DC component of the signal. Finally, the original signal can be expressed as the sum of all intrinsic mode function components and residual components.

[0076] For example, this embodiment collects the solar irradiance data of the photovoltaic power station for the entire year of 2024. The sampling interval is 15 minutes, with a total of 35,040 data points. The irradiance data range is 0-1200W / m², and the solar irradiance data range is 0-1100W / m². Simultaneously, the active and reactive load time series data of 30 nodes in the distribution network for the entire year of 2024 are collected, with a sampling interval of 15 minutes. The active load range is 0.5-8MW, and the reactive load range is 0.2-3MW. Taking the irradiance time series data as an example, six intrinsic mode functions (IMF1-IMF6) and one residual component (Res) are obtained through empirical mode decomposition. The instantaneous frequencies of each IMF component are as follows: IMF1: 1 / 15min - ¹, High frequency; IMF2: 1 / 60min - ¹; IMF3: 1 / 1440min - ¹, Daily cycle; IMF4: 1 / 43200min - ¹, Monthly cycle; IMF5: 1 / 525600min - ¹, Quarterly cycle; IMF6: 1 / 1051200min - ¹, annual cycle; the residual component Res shows a slow upward trend followed by stabilization, characterizing the overall trend of annual light intensity variation, where min - ¹ indicates per minute, which is a unit of frequency, such as 1 / 15 min. - ¹ indicates that a complete oscillation cycle is completed every 15 minutes, meaning that the frequency of change of this component is 15 minutes / time.

[0077] It should also be explained that W / m² represents the solar radiation power received per square meter of area, which is the energy flow density per unit area.

[0078] S1.2: Select components whose instantaneous frequency matches the daily and annual cycles from the intrinsic mode functions, and superimpose and reconstruct them into periodic components. Superimpose the remaining high-frequency intrinsic mode functions into random components. Based on the trend components, periodic components, and random components, use the Copula function to establish a joint probability distribution model of photovoltaic power output and load.

[0079] S1.3: Based on the joint probability distribution model, the Latin hypercube sampling technique is used to sample under the condition of satisfying the grid capacity constraint, and generate a set of probabilistic scenarios covering typical and extreme operating scenarios, i.e., the basic dataset. Each probabilistic scenario consists of a set of photovoltaic power output values ​​and load values, and has a corresponding probability of occurrence.

[0080] The joint probability distribution model of photovoltaic power output and load, based on trend components, periodic components, and random components, and using the Copula function, includes:

[0081] S1.2.1: The trend component and the periodic component are superimposed to form a deterministic sequence of photovoltaic power output and a deterministic sequence of load. The random component is defined as a random sequence of photovoltaic power output and a random sequence of load.

[0082] For example, the Res obtained from the decomposition of light intensity is superimposed with IMF3 and IMF6 to obtain a deterministic sequence of light intensity. Combined with the photovoltaic module conversion efficiency, which is set to 22%, a deterministic sequence of photovoltaic output is calculated. Its range is 0-20MW; the random component of light intensity is obtained by superimposing IMF1 and IMF2, and then converted into a random sequence of photovoltaic power output. Its fluctuation range is -1.2MW to 1.5MW. Similarly, the trend component of the load data is superimposed with the daily and annual IMF components to obtain a deterministic load sequence. The load randomness sequence is obtained by superimposing the remaining high-frequency IMF components. .

[0083] S1.2.2: Fit the marginal probability distributions of the photovoltaic output random sequence and the load random sequence respectively, determine their respective optimal probability distribution models and parameters, and thus obtain the marginal probability distributions of photovoltaic output randomness and load randomness.

[0084] Furthermore, the calculation steps for the marginal probability distribution include:

[0085] (1) Obtain the photovoltaic output randomness sequence and load randomness sequence to be analyzed, and calculate the mean, variance, skewness and kurtosis of each sequence. The calculation formulas for mean, variance, skewness and kurtosis are all conventional means that can be understood and implemented by those skilled in the art. This application is not limited to specific partitioning methods.

[0086] (2) Based on the calculated statistics, select a set of candidate probability distribution models for each random sequence. The candidate models should cover different types of distribution forms, such as normal distribution, Weibull distribution, log-normal distribution, Gamma distribution, etc., to form a set of candidate models for the sequence.

[0087] (3) For each probability distribution model in the candidate model set, the maximum likelihood estimation method is used to maximize the likelihood function value of the observed data under the given model to obtain the optimal estimate of the model parameters, thereby obtaining a set of candidate models with specific parameter values. The maximum likelihood estimation method is a conventional means that can be understood and implemented by those skilled in the art. This application is not limited to specific partitioning methods.

[0088] (4) For each candidate model with parameter values, calculate its goodness-of-fit test statistic and evaluate the fit between the model and the data using a probability plot. At the same time, calculate its Akaike information criterion value. The Akaike information criterion can effectively penalize model complexity and prevent overfitting while evaluating the goodness-of-fit of the model. The Akaike information criterion is a conventional means that can be understood and implemented by those skilled in the art. This application is not limited to specific partitioning methods.

[0089] (5) Based on the quantitative evaluation results, select the optimal model from the candidate model set. The criteria for the optimal model is: among the models that pass the goodness-of-fit test with the significance level requirement, the Akaike Information Criterion value is the smallest. If the criterion values ​​of multiple models are similar, the model with a simpler structure and fewer parameters is preferred.

[0090] (6) The selected optimal model and its corresponding parameter estimates are finally determined as the marginal probability distribution of the random sequence.

[0091] For example, the KS test is used to analyze the randomness sequence of photovoltaic power output. Marginal distribution fitting was performed, with candidate distributions including normal, Weibull, log-normal, and Gamma distributions. By calculating the KS statistic and p-value for each distribution, the results showed that the Weibull distribution provided the best fit, with a KS statistic of 0.042 and a p-value of 0.86, both greater than the preset value of 0.05. Its parameters were: shape parameter k = 2.3, scale parameter... =0.8MW; for load stochastic sequence During the fitting process, the log-normal distribution showed the best fit, with a KS statistic of 0.038 and a p-value of 0.91, both greater than the preset value of 0.05. The parameter is the mean. =0.02, standard deviation =0.15, where KS test, normal distribution, Weibull distribution, log-normal distribution, and Gamma distribution are all conventional means that can be understood and implemented by those skilled in the art, and this application is not limited to specific partitioning methods.

[0092] S1.2.3: Based on the fitted photovoltaic output randomness marginal probability distribution and load randomness marginal probability distribution, the observed values ​​of the photovoltaic output randomness sequence and load randomness sequence are transformed into pseudo-observed values ​​that follow a uniform distribution of [0,1] through probability integral transformation. The pseudo-observed values ​​are used to select the optimal Copula function from the preset Copula function family by calculating the maximum likelihood value and estimating its parameters. The probability integral transformation is a conventional means that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0093] For example, the default Copula function family includes Gaussian Copula, t-Copula, Clayton Copula, and Gumbel Copula. For and The observed values ​​are subjected to probability integral transformation to obtain pseudo-observations u and v. The log-likelihood values ​​of each Copula function are calculated: Gaussian Copula, t-Copula, Clayton Copula, and Gumbel Copula. Since Gumbel Copula has the largest log-likelihood value, it is chosen as the optimal Copula function. Its parameters are obtained through maximum likelihood estimation. =1.8 indicates that there is a moderate positive correlation between the randomness of photovoltaic output and the randomness of load.

[0094] S1.2.4: By using the selected optimal Copula function to connect the random marginal probability distribution of photovoltaic power output and the random marginal probability distribution of load, a joint random probability distribution model of photovoltaic power output and load is constructed.

[0095] Furthermore, the specific steps in S1.2.4 include:

[0096] (1) Input each observation value in the photovoltaic output random sequence and load random sequence into the photovoltaic output random marginal probability distribution and load random marginal probability distribution determined in the marginal probability distribution fitting, respectively, and calculate the cumulative distribution function value corresponding to each observation value. These cumulative distribution function values ​​are the pseudo observation value sequence that follows a uniform distribution, and are used as the input data of the Copula function.

[0097] (2) The two sets of pseudo-observation sequences generated are used as sample data and input into the selected optimal Copula function. The maximum likelihood estimation method is used to estimate the parameters. The numerical optimization algorithm is used to iteratively search for the parameter value that maximizes the likelihood function value composed of the Copula density function and the marginal probability density function. This parameter value is the parameter of the optimal Copula function.

[0098] (3) According to Sklar's theorem, the optimal Copula function and the parameters estimated in (2) are used to connect the random marginal probability distribution of photovoltaic output and the random marginal probability distribution function of load respectively, and a joint probability distribution model of photovoltaic output and load is constructed. After the construction is completed, the empirical Copula function is calculated and compared with the constructed theoretical Copula function. By drawing a probability graph and calculating the statistics, the goodness of fit of the constructed model to the relevant structure of the sample data is tested. If the test fails, the Copula function is reselected. It should be explained that the joint distribution function is equal to the composite function of the Copula function on the random marginal probability distribution function. Sklar's theorem is a conventional means that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method.

[0099] (4) The output of the verified joint probability distribution model of photovoltaic power output and load is used as the input of the complete joint probability distribution model integration.

[0100] S1.2.5: Integrate the deterministic sequence of photovoltaic power output and the deterministic sequence of load with the established stochastic joint probability distribution model to finally establish a joint probability distribution model of photovoltaic power output and load.

[0101] The specific steps of S1.3 include:

[0102] S1.3.1: Based on the joint probability distribution model of photovoltaic power output and load, the Latin hypercube sampling technique is used to generate initial sampling points for the two variables of photovoltaic power output and load, which satisfy the marginal probability distribution but are independent of each other. The Latin hypercube sampling technique is a conventional method that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0103] For example, the sampling size is set to 1000, and the probability interval [0,1] is divided into 1000 sub-intervals. A sample point is randomly selected in each sub-interval, and the initial sampling sample point is obtained by mapping the marginal distribution inverse function of photovoltaic power output and load. There are a total of 1000 groups. and They are independent of each other.

[0104] S1.3.2: The sorting of the sample points is adjusted by using the rank correlation coefficient matching technique, and the sample points adjusted by correlation are mapped back to the data space composed of the original photovoltaic power output value and load value through their respective inverse marginal distribution functions to obtain the original set of numerical sample points.

[0105] Furthermore, the specific steps in S1.3.2 include:

[0106] (1) The sampling points of photovoltaic power output and load generated by Latin hypercube sampling technology are used as the input to be adjusted. The expected rank correlation coefficient between the random sequence of photovoltaic power output and load is obtained from the joint probability distribution model and set as the target correlation coefficient to be achieved in this adjustment.

[0107] (2) The photovoltaic output component and load component in the sampled sample points are sorted separately, and the Spearman rank correlation coefficient between them is calculated. The sorting of the sample points is adjusted by an iterative algorithm so that the rank correlation coefficient of the adjusted sampled sample points approaches the target correlation coefficient. The specific method is as follows: First, the difference between the current sampled sample point and the target correlation coefficient is calculated. Then, the difference is gradually reduced by swapping the sorting positions of the photovoltaic output and load components in the sampled sample points. Each swapping operation will change the relative order between the sample points, thereby changing the rank correlation coefficient. When the rank correlation coefficient reaches the preset tolerance range, the adjustment process is stopped, and the correlation matching sample points are obtained. The Spearman rank correlation coefficient is a conventional method that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method.

[0108] (3) Input the photovoltaic output component in the correlation matching sample points into the inverse function of the random marginal probability distribution of photovoltaic output load, and input the load component into the inverse function of the random marginal probability distribution of load, so that the probability value samples with values ​​ranging from zero to one are mapped back to the original physical dimension numerical space of photovoltaic output and load respectively, thereby converting the correlation matching sample points in the probability space into the original numerical sample points in the physical space.

[0109] (4) After completing the inverse mapping, the resulting data set containing the one-to-one correspondence between the physical values ​​of photovoltaic output and the physical values ​​of load is the original numerical sample point set after correlation adjustment.

[0110] For example, the rank correlation coefficient determined by the Gumbel Copula function. The initial sampling point sorting is adjusted so that the adjusted sample points It satisfies this rank correlation coefficient. For example, any group in the initial sample The sorting deviated from the relevance requirement and was adjusted accordingly. This ensures a positive correlation between photovoltaic output and load.

[0111] S1.3.3: In the original set of numerical sample points, the power grid capacity constraint is introduced as a filter. Each sample is checked to see if it meets the node voltage constraint, line power constraint, transformer capacity constraint and distributed power output constraint. All samples that violate any constraint are removed. The remaining scenario samples constitute the probabilistic scenario set, which is the basic dataset that ultimately covers typical and extreme scenarios.

[0112] Furthermore, node voltage constraints refer to the requirement that the voltage amplitude at each node in a power system must be maintained within specified upper and lower limits, typically ±5% or ±10% of the rated voltage, to ensure the normal operation of electrical equipment and power quality. Line power constraints refer to the requirement that the active and reactive power transmitted by transmission and distribution lines cannot exceed their thermal stability and stability limits, preventing line overload from causing overheating, excessive voltage drop, or system instability. Transformer capacity constraints refer to the requirement that the apparent power transmitted by transformers cannot exceed their rated capacity, including capacity limits for main transformers and distribution transformers, to avoid transformer overload affecting lifespan and system safety. Distributed power generation output constraints refer to the requirement that the actual output of distributed power sources such as photovoltaics and wind power must be within their technical output range, i.e., between minimum and maximum output, while also meeting dynamic constraints such as ramp rate. The minimum output is typically zero or the technical minimum output, while the maximum output is the rated capacity.

[0113] Furthermore, the power grid capacity constraint is implemented by calling an embedded fast power flow calculation module when screening scenarios: for each initial scenario generated by sampling, the built-in fast power flow calculation module simulates the steady-state operation of the power grid under the scenario and calculates the voltage of each node and the power of each branch; the calculation results are compared with the preset safety limits, and the scenario is retained in the probabilistic scenario set only when all electrical quantities are within the safety limits; otherwise, the scenario is marked as physically infeasible and discarded.

[0114] For example, among the 1000 initial adjustment samples, 32 samples violated the constraints, such as 25 samples with photovoltaic output exceeding 20MW and 7 samples with line power exceeding 30MW. After removing these, 968 valid scenario samples were obtained, forming the basic dataset. Among them, those with a probability of occurrence greater than or equal to 0.1% are typical scenarios, totaling 850, such as scenario 1: P=12.5MW, L=4.2MW, with a probability of 0.12%; those with a probability of occurrence less than 0.1% are extreme scenarios, totaling 118, such as scenario 968: P=19.8MW, representing extreme illumination, L=7.5MW, representing extreme load, with a probability of 0.08%.

[0115] The calculation process of the coupling evaluation model includes:

[0116] S2.1: The generation nodes, load nodes, and substation nodes in the distribution network are modeled as graph nodes, and the transmission lines and transformer branches are modeled as edges. Each edge is configured with a weight reflecting the line impedance to construct a graph neural network model. The operation of each layer of the graph neural network is defined as follows: each node aggregates the feature information of its one-hop neighbor nodes through the connecting edges. Each node concatenates the aggregated neighbor information with its own feature representation from the previous layer and inputs it into a fully connected layer and activation function to update and generate its own feature representation in the current layer. The feature information includes the voltage, injected power, and risk state value of the neighbor nodes.

[0117] Furthermore, a one-hop neighbor node refers to a node that is directly connected to the target node, and is typically used for feature aggregation and node representation learning.

[0118] Furthermore, the specific layer structure configuration of the graph neural network model is as follows: the model consists of an input layer, at least two graph convolutional layers, and a readout layer connected in sequence; the input layer is used to receive the initial feature vector of each graph node, the initial feature vector including node type, real-time active power, real-time reactive power, and voltage amplitude; each graph convolutional layer uses a Chebyshev polynomial based on the spectral domain for convolution operation to reduce computational complexity, and its activation function adopts the ReLU function; the readout layer performs a global average pooling operation on the final feature representation of all nodes to obtain a vector representing the global state of the entire distribution network, and then performs a secondary fusion with the features of each node to finally output multiple risk indicators for each node.

[0119] For example, in this embodiment, 30 nodes of the distribution network are modeled as graph nodes. Node 1 is a substation, nodes 25-26 are photovoltaic power generation nodes, and the rest are load nodes. 35 lines and 2 transformers are modeled as edges, and the edge weights are set to the line impedance values. For example, the impedance of line 1-2 is 0.05 + j0.12Ω, and the weight is taken as the impedance magnitude of 0.13Ω. A 3-layer graph convolutional neural network is constructed. The input layer has a feature dimension of 6, including voltage amplitude, voltage phase, active power injection, reactive power injection, node load rate, and historical risk value. The hidden layer has a dimension of 16, and the output layer has a dimension of 3. The ReLU function is used as the activation function. Taking node 25 as an example, its one-hop neighbors are nodes 24 and 26. The voltage, injected power, and other features of the neighbor nodes are aggregated, concatenated with the node's own features, and then input into the fully connected layer to update the feature representation of node 25.

[0120] S2.2: By stacking multiple layers of graph convolutional layers, the final feature representation of each node is fused with the global information of its multi-hop neighbors;

[0121] S2.3: For each operating scenario in the generated basic dataset, use the trained graph neural network model to calculate multi-dimensional node risk indicators for each node in the distribution network; the multi-dimensional node risk indicators include curtailment risk indicators, power quality risk indicators, and power grid safety and stability risk indicators.

[0122] Furthermore, the specific steps in S2.3 include:

[0123] (1) Obtain a running scenario from the basic dataset. This probabilistic scenario includes photovoltaic power output and load data of each node in the distribution network. Using the photovoltaic power output and load data contained in this probabilistic scenario as input, construct the corresponding graph structure data, including: taking the nodes of the distribution network as graph vertices and the electrical connections between nodes as edges, assigning a feature vector to each vertex, which includes the injected power, voltage amplitude, node type and capacity information of the node, expressing the connection relationship between vertices in the form of an adjacency matrix, and outputting the graph structure data representing the running scenario;

[0124] (2) Input the constructed graph structure data into the trained graph neural network model. The graph neural network model performs calculations through its internal multi-layer graph convolution and message passing mechanism. Each vertex aggregates the feature information of its one-hop neighbor vertex in each layer and updates its own features. After multiple iterations, the model outputs the final representation vector of each vertex.

[0125] (3) Input the final representation vector into a fully connected neural network layer. The parameters of the fully connected neural network layer have been learned during the model training phase. This neural network layer analyzes the features related to photovoltaic consumption in the final representation vector, such as line transmission margin, voltage level, node injected power, etc., and outputs a quantified value, which is the curtailment risk index of the node.

[0126] (4) Input the same final representation vector into another parallel fully connected neural network layer. This fully connected neural network layer parses the features related to voltage, waveform and frequency in the final representation vector, such as voltage amplitude, voltage imbalance, harmonic distortion level, etc., and outputs a quantified value, which is the power quality risk index of the node.

[0127] (5) Input the final representation vector into the third parallel fully connected neural network layer. The fully connected neural network layer analyzes the deep features related to the system stability margin in the final representation vector, such as static voltage stability margin, power angle stability index, short circuit capacity, etc., and outputs a quantified value, which is the power grid safety and stability risk index of the node.

[0128] (6) The risk indicators of curtailment of solar power, power quality and power grid safety and stability calculated for all nodes are summarized to form a set of multi-dimensional risk indicators for each node in the current scenario. Then, for each scenario in the basic dataset, (1)-(6) are repeated to complete the risk calculation for all scenarios.

[0129] For example, risk indicators are calculated for scenario 968 (P=19.8MW, L=7.5MW) in the basic dataset: the curtailment risk indicator is calculated by the difference between photovoltaic output and grid absorption capacity. The curtailment risk indicator for node 25 is 0.82, where the full score is 1.0, and the higher the score, the greater the risk; the power quality risk indicator is calculated based on voltage deviation and frequency fluctuation. For example, the voltage of node 25 is 10.8kV, the deviation is 8%, exceeding the ±7% threshold, and the frequency fluctuation is 0.2Hz, so the score is 0.76; the grid security and stability risk indicator is calculated based on the line load rate (the load rate of line 25-24 is 89%) and the transformer load rate (the load rate of main transformer 1 is 78%). For example, the load rate of line 25-24 is 89%, and the load rate of main transformer 1 is 78%, so the score is 0.65.

[0130] S2.4: For the multi-dimensional node risk indicators under each operating scenario, the improved CRITIC-entropy weighting method is used to dynamically assign weights to the curtailment risk indicators, power quality risk indicators and grid security and stability risk indicators, and the weighted summation of each node is used to obtain the comprehensive risk value.

[0131] Furthermore, the specific steps in S2.4 include:

[0132] (1) Obtain a set of multi-dimensional risk indicators for each node under all operating scenarios, standardize all indicator data, and map them uniformly to a value range of zero to one;

[0133] (2) Calculate the standard deviation of each risk indicator on the time series to characterize its contrast strength. At the same time, calculate the Pearson correlation coefficient between the indicators to measure their correlation. When calculating the traditional conflict indicators, a volatility factor based on the rate of change of the indicators between adjacent time points is introduced to correct it, and an improved conflict indicator is obtained. The contrast strength of each indicator is multiplied by the improved conflict indicator to obtain its information carrying capacity. Then, the weight based on the CRITIC method is calculated by normalization. The standard deviation and Pearson correlation coefficient calculation formulas are conventional means that can be understood and implemented by those skilled in the art. This application is not limited to a specific partitioning method.

[0134] (3) Based on the values ​​of each risk indicator in all operating scenarios, calculate the probability of occurrence in each scenario to form the probability distribution of each indicator. According to the definition of information entropy, calculate the information entropy of each indicator based on the probability distribution, and calculate the information utility value of each indicator accordingly. The weight based on the entropy weight method is calculated by normalization. The calculation formulas of information entropy and information utility are conventional means that can be understood and implemented by those skilled in the art. This application is not limited to specific partitioning methods.

[0135] (4) The weights calculated based on the CRITIC method and the weights based on the entropy weight method are linearly combined by using equal weight average to obtain the final comprehensive weights of the three risk indicators of curtailment of solar power, power quality and grid security and stability.

[0136] (5) For each node, standardize the three risk indicators of the node in any operating scenario, multiply them by the corresponding comprehensive weights, and then perform a weighted summation to obtain the comprehensive risk value of the node in that scenario. Repeat this weighted summation operation for each node and each operating scenario.

[0137] It is important to emphasize that when using the improved CRITIC-entropy weighting method to dynamically assign weights to the three risk indicators, the improvement lies in the fact that the calculation of the conflict indicators in the CRITIC method incorporates the volatility of the indicators over time, while the calculation of the information entropy in the entropy weighting method is based on the probability distribution of the risk indicators in different scenarios. The weighted risk indicators are summed to obtain the comprehensive risk value of each node at different times, thereby forming the time-varying comprehensive risk level matrix.

[0138] For example, the weights are calculated using the improved CRITIC-entropy weight method. First, the conflict and information content of each indicator are calculated. The conflict between curtailment and power quality indicators is 0.32, between curtailment and power safety and stability indicators is 0.28, and between power quality and power safety and stability indicators is 0.41. The entropy values ​​of their information content are 0.82, 0.75, and 0.78, respectively. The combined weights are: curtailment risk indicator 0.35, power quality risk indicator 0.30, and grid safety and stability risk indicator 0.35. For instance, the comprehensive risk value of node 25 is 0.82 × 0.35 + 0.76 × 0.30 + 0.65 × 0.35 = 0.74, indicating a high risk level. The risk levels are further divided into: 0-0.3 (low), 0.3-0.6 (medium), and 0.6-1.0 (high).

[0139] S2.5: The collection of comprehensive risk values ​​of all nodes under all scenarios and at all times constitutes a time-varying comprehensive risk level matrix.

[0140] The time-varying comprehensive risk level matrix is ​​a three-dimensional data structure. Its first dimension index represents the node number in the distribution network, i.e., 1-30. The second dimension index represents the discretized time point, such as 96 time points throughout the day at 15-minute intervals, corresponding to the numbers 1-96. The third dimension stores the comprehensive risk value of the node at the corresponding time point and the decomposed values ​​of various risk indicators.

[0141] The generation process of the time-varying comprehensive risk level matrix is ​​as follows: for each probabilistic scenario in the basic dataset, the risk value of all nodes in the scenario at all time points is calculated using the coupled evaluation model. Then, a weighted average is performed based on the probability of each scenario in the basic dataset to finally obtain the time-varying comprehensive risk level matrix that reflects the overall risk level.

[0142] For example, the generated time-varying comprehensive risk level matrix has dimensions of 30×96×4, including 30 nodes, 96 time points, and 4 data items: comprehensive risk value, curtailment risk value, power quality risk value, and safety and stability risk value. Specifically, node 25 at time point 48 (12:00) has a comprehensive risk value of 0.72, a curtailment risk value of 0.80, a power quality risk value of 0.71, and a safety and stability risk value of 0.65; at time point 72 (18:00), the comprehensive risk value is 0.68, the curtailment risk value is 0.25, the power quality risk value is 0.75, and the safety and stability risk value is 0.82.

[0143] The specific steps of S3 include:

[0144] S3.1: Analyze the time-varying comprehensive risk level matrix, extract risk features from it, and transform it into a standardized graph query statement; the risk features include the risk type, risk level, and risk impact range characteristics at a specified time, at a specified node, or in a specified area;

[0145] Furthermore, the specific steps of S3.1 include:

[0146] (1) Read the input time-varying comprehensive risk level matrix and set the risk level classification threshold. For example, set the risk value to be low risk if it is between 0 and 0.3, medium risk if it is between 0.3 and 0.6, and high risk if it is between 0.6 and 1.0.

[0147] (2) Locate the spatiotemporal position where the risk value exceeds the set threshold from the time-varying comprehensive risk level matrix, that is, determine the specific time and node number of the occurrence of medium and high risks. According to the dimension to which the risk value belongs at the location, determine whether the risk type is curtailment risk, power quality risk or grid security and stability risk. According to the size of the risk value, determine whether the risk level is medium or high by referring to the risk level classification threshold. Identify the electrical adjacent node or the preset area division of the node in the network topology to determine the scope of risk impact. Finally, combine them into a risk feature triplet containing risk type, risk level and affected node.

[0148] (3) The extracted risk feature triples are standardized and transformed, wherein the risk type is mapped to the predefined enumeration words in the knowledge graph, the risk level is mapped to the level label defined in the knowledge graph, and the scope of influence is converted into a standardized region code in the form of a list of node numbers or a region identifier.

[0149] (4) Prepare a query statement template according to the predefined graph query pattern. The query statement template includes the node type to be queried, the relationship type, and placeholders for receiving standardized risk characteristics as filtering conditions. The query target is usually set to find historical risk case nodes similar to the current risk characteristics.

[0150] (5) Use the obtained standardized risk type, risk level and impact range as parameters, fill them into the corresponding placeholders in the query statement template, and generate a specific executable standardized graph query statement.

[0151] For example, the time-varying comprehensive risk level matrix is ​​parsed to extract key risk features. For instance, nodes 25-26 have a high power quality risk level at time points 48-52, and the affected area covers nodes 24-27. The risk type is voltage deviation exceeding the standard. This is then converted into a standardized graph query statement: "MATCH (n: risk event case)-[r: impact]->(m: grid equipment) WHEREn.risk type='voltage deviation exceeding the standard' AND n.risk level='high' AND n.affected area='nodes 24-27'RETURN (n)-[r]->(m)".

[0152] S3.2: In the pre-constructed risk knowledge graph of measures, execute the graph query statement to obtain a set of candidate response and disposal measures; the graph traversal query first locates the historical risk case node with the highest similarity to the current risk feature, and then finds all response and disposal measure nodes connected to the historical risk case node along the adoption measure relationship edge in the risk knowledge graph of measures. The similarity calculation is a conventional means that can be understood and implemented by those skilled in the art, and this application is not limited to a specific partitioning method.

[0153] For example, the pre-built risk knowledge graph contains 120 risk event case nodes, 85 response and handling measure nodes, and 50 power grid equipment nodes. After executing the query, the historical case with the highest similarity is located, for example, case ID: C28, characteristics: excessive voltage deviation in photovoltaic area, high risk, affecting 4 nodes). Along the "adopted measures" relationship edge, 3 response measure nodes are found: measure A: put into operation SVG reactive power compensation device, measure B: adjust photovoltaic output limit, and measure C: optimize line switching, forming a set of candidate measures.

[0154] S3.3: Perform conflict verification and priority ranking on the set of candidate response measures, and generate a dynamic adaptive response plan that includes the execution order of specific measures, the timing of activation, the expected effect, and the resource allocation plan based on the priority ranking result; the conflict verification is to determine whether different measures will compete for the same control resources or produce offsetting effects when they are executed; the priority ranking is a comprehensive score based on the implementation cost of the measures, the speed of the expected effect decay risk, and the current real-time operating status of the power grid.

[0155] For example, conflict verification: Measure A and Measure B have no resource competition, and Measure C has no conflict with A or B; priority scoring: Measure A has an implementation cost of 20,000 yuan, a slow effect decay rate, a real-time adaptability score of 90, and a comprehensive score of 88; Measure B has an implementation cost of 10,000 yuan, a fast effect decay rate, a real-time adaptability score of 85, and a comprehensive score of 76; Measure C has an implementation cost of 50,000 yuan, a medium effect decay rate, a real-time adaptability score of 82, and a comprehensive score of 72. The ranking result is A>B>C, generating a dynamic adaptive response plan:

[0156] Start-up timing: Time point 47; Execution sequence: First, put the SVG reactive power compensation device near node 25 into operation. If the voltage deviation does not drop to within the threshold, then limit the photovoltaic output to 18MW; Expected effect: The voltage deviation will drop to within ±5%, and the power quality risk level will drop to medium; Resource allocation: Dispatch 1 group of SVG device operation and maintenance personnel and 1 group of photovoltaic power station control personnel to arrive before 11:30.

[0157] The pre-constructed risk knowledge graph includes at least three types of entity nodes: risk event case nodes, response and disposal measure nodes, and power grid equipment nodes. The attributes of the risk event case node include the case occurrence time, risk type, risk level, affected power grid area, and weather conditions. The attributes of the response and disposal measure node include the measure type, required control resources, implementation cost, expected effect, and effective delay time. The attributes of the power grid equipment node include the equipment ID, equipment type, and equipment capacity.

[0158] The entity nodes are connected by relational edges, including: the impact relationship between risk event cases and the affected power grid equipment; the adoption relationship between risk event cases and the response and handling measures taken; and the interaction relationship between the response and handling measures and the operated power grid equipment.

[0159] The process of validating the dynamic adaptive response scheme in a digital twin-based distribution network simulation model includes:

[0160] S4.1: Synchronously inject the current operating scenario data and the measures and instructions in the dynamic adaptive response scheme from the basic dataset into the distribution network digital twin simulation model to complete the initialization;

[0161] Furthermore, the digital twin simulation model of the distribution network includes a physical model layer, a data layer, a model layer, and an application layer; the physical model layer constructs a three-dimensional geometric model and a physical characteristic model of the distribution network equipment; the data layer stores real-time operating data, historical data, and simulation data; the model layer includes a power flow calculation model, a fault analysis model, and a risk assessment model; and the application layer provides visualization and interactive control functions.

[0162] S4.2: Run the initialized digital twin simulation model of the distribution network to simulate the dynamic response process of the distribution network within a preset time period after the execution of the dynamic adaptive response scheme. Record key performance indicators during the simulation process. The key performance indicators include system voltage deviation, frequency fluctuation, curtailed solar power, and equipment load rate.

[0163] S4.3: Compare the simulated key performance indicators with the preset standard thresholds for each indicator. If all key performance indicators are better than the standard thresholds, the verification is deemed successful. If any key performance indicator exceeds the standard threshold, the verification is deemed unsuccessful. When the verification fails, record the incremental data generated by the simulation. The incremental data includes voltage over-limit records, power over-limit records, power quality indicators, equipment operating status, and system stability indicators. The incremental data is stored in time series format and labeled with corresponding operating scenario tags and risk level tags.

[0164] For example, a digital twin simulation model of the distribution network is built based on PSCAD / EMTDC, and the operating data of scenario 968 (P=19.8MW, L=7.5MW) and response instructions are injected, such as putting a 2Mvar SVG into operation at 11:45. The simulation period is set from 11:00 to 14:00, and key performance indicators are recorded: before 11:45, the voltage deviation at node 25 is 8%; after the SVG is put into operation, the voltage deviation drops to 6.2% at 11:46, but still exceeds the preset standard threshold; the frequency fluctuates by 0.15Hz; the curtailed solar power is 0; and the line load rate is 85%. The verification is deemed unsuccessful, and incremental data is recorded: the failure time 11:45-13:20, the system operating status trajectory, such as voltage fluctuations between 6.2% and 6.8%, and the reason for the failure of the response plan, such as insufficient SVG compensation capacity.

[0165] Using the incremental data as feedback, the parameters of the coupled evaluation model were updated: the weights of the hidden layers in the graph neural network were adjusted, for example, the weights of voltage deviation-related features were increased from 0.25 to 0.32; the base dataset was optimized: samples from scenarios with insufficient SVG compensation were added, and the base dataset was regenerated. Based on the updated model and dataset, a new response plan was generated, and all key indicators met the requirements, thus the validation was successful.

[0166] This invention discloses a risk assessment method for integrating new energy photovoltaic (PV) power generation into distribution networks. By constructing a precise basic dataset, integrating a coupled assessment model of graph neural networks and risk propagation theory, implementing dynamic adaptive response schemes, and employing a digital twin verification and feedback mechanism, it overcomes the bottlenecks of traditional risk assessment methods, which suffer from weak targeting and poor dynamic adaptability. This embodiment uses the integration of a 20MW PV power plant into a 10kV distribution network in East China as a specific scenario, demonstrating the complete implementation process of the method in detail: Empirical Mode Decomposition (EMD) and Copula functions are used to accurately characterize the joint distribution characteristics of PV output and load; improved Latin hypercube sampling generates a basic dataset covering typical and extreme scenarios; a graph neural network is used to accurately calculate multi-dimensional risk indicators; an improved CRITIC-entropy weighting method is used to dynamically weight and obtain a time-varying comprehensive risk level matrix; a risk knowledge graph of measures is used to quickly match and generate the optimal response scheme; and digital twin simulation verification and feedback optimization ensure the effectiveness and reliability of the scheme. This method improves the accuracy, dynamic adaptability, and engineering operability of risk assessment for integrating new energy PV power generation into distribution networks, providing technical support for the safe and stable operation of distribution networks. It is applicable to various risk assessment scenarios for distributed PV power generation into distribution networks.

[0167] Example 2:

[0168] Please see Figure 3 Another embodiment of the present invention provides: a risk assessment system for connecting new energy photovoltaic power generation to the distribution network, comprising:

[0169] Data construction module, coupled evaluation model module, solution generation module, verification and optimization module;

[0170] The data construction module is used to collect solar resource data and power distribution network load data. Through data decomposition, joint probabilistic modeling, and constrained sampling operations, it generates a probabilistic scenario set covering typical and extreme operating scenarios.

[0171] The coupled assessment model module integrates graph neural networks and risk propagation theory to perform risk calculations on various operating scenarios in the basic dataset. Finally, it outputs a time-varying comprehensive risk level matrix that reflects the three-dimensional relationship between nodes, time, and risk values, thereby achieving accurate assessment of multi-dimensional risks in the distribution network.

[0172] The solution generation module, based on the time-varying comprehensive risk level matrix, matches it with the pre-built risk knowledge graph of measures. Through conflict verification and priority sorting, it generates dynamic response solutions that adapt to the real-time risk status, providing a clear basis for risk management.

[0173] The verification and optimization module is used to verify the effectiveness of the dynamic adaptive response scheme through the digital twin simulation model. For schemes that do not meet the preset standards, the incremental data generated during the simulation process is used as feedback to update the parameters of the coupled evaluation model and optimize the basic dataset, thereby continuously improving the accuracy of the evaluation model and the feasibility of the response scheme.

[0174] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments under the guidance of the present invention without departing from the spirit and scope of the present invention. All of these variations are within the protection scope of the present invention.

Claims

1. A risk assessment method for connecting new energy photovoltaic power generation to the power distribution network, characterized in that, include: S1: Collect solar energy resource data from photovoltaic power plants and load data from distribution networks to form a basic dataset; the basic dataset includes photovoltaic output ranges, load levels, and corresponding grid absorption capacity boundaries under different operating scenarios; S2: Based on the aforementioned basic dataset, construct a coupled assessment model that integrates graph neural networks and risk propagation theory, and output a time-varying comprehensive risk level matrix; S3: Match the time-varying comprehensive risk level matrix with the pre-constructed risk knowledge graph of measures, and generate a dynamic adaptive response plan based on the matching results; S4: The dynamic adaptive response scheme is verified in a digital twin-based power distribution network simulation model. If the verification result does not meet the preset standard, the incremental data generated by the simulation verification is used as feedback to update the parameters of the coupled evaluation model and simultaneously optimize the basic dataset. Based on the basic dataset, a graph structure of distribution network nodes and branches is constructed and multiple graph convolutional layers are stacked to calculate multi-dimensional risk indicators of curtailment of solar power, power quality, and grid security and stability. After dynamic weighting using the improved CRITIC-entropy weighting method, a comprehensive risk value is obtained, and a three-dimensional time-varying comprehensive risk level matrix of node-time-risk value is output. Based on the time-varying comprehensive risk level matrix, risk features are extracted and a graph query statement is generated. This query is matched with a risk knowledge graph containing risk event cases, response and disposal measures, and power grid equipment entity nodes. After conflict verification and priority sorting, a dynamic adaptive response plan containing execution order, initiation timing, expected effects, and resource allocation is output. The improved CRITIC-entropy weight method introduces the time volatility factor of the index in the calculation of the conflict index of the CRITIC method to obtain the improved CRITIC weight. The entropy weight method calculates the information entropy based on the probability distribution of the risk index scenario to obtain the entropy weight method weight. The improved CRITIC weight and the entropy weight method weight are linearly combined and weighted to obtain the comprehensive risk value.

2. The risk assessment method for connecting new energy photovoltaic power generation to the distribution network as described in claim 1, characterized in that, The formation process of the basic dataset is as follows: S1.1: Collect historical solar irradiance and solar power plant time series data as solar resource data, and simultaneously collect historical active and reactive load time series data of each node in the distribution network as load data. Perform empirical mode decomposition on the solar resource data and load data respectively, decomposing each data sequence into intrinsic mode functions and a residual component; the intrinsic mode functions are components extracted in descending order of instantaneous frequency to characterize the oscillation modes of different scales of the data. The residual component is the monotonic or constant sequence remaining after decomposition, and the residual component is defined as the trend component. S1.2: Select components whose instantaneous frequency matches the daily and annual cycles from the intrinsic mode functions, and superimpose and reconstruct them into periodic components. Superimpose the remaining high-frequency intrinsic mode functions into random components. Based on the trend components, periodic components, and random components, use the Copula function to establish a joint probability distribution model of photovoltaic power output and load. S1.3: Based on the joint probability distribution model, the Latin hypercube sampling technique is used to sample under the condition of satisfying the grid capacity constraint, and generate a set of probabilistic scenarios covering typical and extreme operating scenarios, i.e., the basic dataset. Each probabilistic scenario consists of a set of photovoltaic power output values ​​and load values, and has a corresponding probability of occurrence.

3. The risk assessment method for connecting new energy photovoltaic power generation to the distribution network as described in claim 2, characterized in that, The joint probability distribution model of photovoltaic power output and load, based on trend components, periodic components, and random components, and using the Copula function, includes: S1.2.1: The trend component and the periodic component are superimposed to form a deterministic sequence of photovoltaic power output and a deterministic sequence of load. The random component is defined as a random sequence of photovoltaic power output and a random sequence of load. S1.2.2: Fit the marginal probability distributions of the photovoltaic output random sequence and the load random sequence respectively, determine their respective optimal probability distribution models and parameters, and thus obtain the marginal probability distributions of photovoltaic output randomness and load randomness. S1.2.3: Based on the fitted photovoltaic output randomness marginal probability distribution and load randomness marginal probability distribution, the observed values ​​of the photovoltaic output randomness sequence and load randomness sequence are transformed into pseudo-observed values ​​that follow a uniform distribution of [0,1] through probability integral transformation. The pseudo-observed values ​​are used to select the optimal Copula function from the preset Copula function family by calculating the maximum likelihood value and estimate its parameters. S1.2.4: By using the selected optimal Copula function to connect the random marginal probability distribution of photovoltaic power output and the random marginal probability distribution of load, a joint random probability distribution model of photovoltaic power output and load is constructed. S1.2.5: Integrate the deterministic sequence of photovoltaic power output and the deterministic sequence of load with the established stochastic joint probability distribution model to finally establish a joint probability distribution model of photovoltaic power output and load.

4. The risk assessment method for connecting new energy photovoltaic power generation to the distribution network as described in claim 3, characterized in that, The specific steps of S1.3 include: S1.3.1: Based on the joint probability distribution model of photovoltaic power output and load, the Latin hypercube sampling technique is used to generate initial sampling points that satisfy the marginal probability distribution but are independent of each other for the two variables of photovoltaic power output and load. S1.3.2: The sorting of the sample points is adjusted by using the rank correlation coefficient matching technique, and the sample points adjusted by correlation are mapped back to the data space composed of the original photovoltaic power output value and load value through their respective inverse marginal distribution functions to obtain the original set of numerical sample points. S1.3.3: In the original set of numerical sample points, the power grid capacity constraint is introduced as a filter. Each sample is checked to see if it meets the node voltage constraint, line power constraint, transformer capacity constraint and distributed power output constraint. All samples that violate any constraint are removed. The remaining scenario samples constitute the probabilistic scenario set, which is the basic dataset that ultimately covers typical and extreme scenarios.

5. The risk assessment method for connecting new energy photovoltaic power generation to the distribution network as described in claim 1, characterized in that, The calculation process of the coupling evaluation model includes: S2.1: The generation nodes, load nodes, and substation nodes in the distribution network are modeled as graph nodes, and the transmission lines and transformer branches are modeled as edges. Each edge is configured with a weight reflecting the line impedance to construct a graph neural network model. The operation of each layer of the graph neural network is defined as follows: each node aggregates the feature information of its one-hop neighbor nodes through the connecting edges. Each node concatenates the aggregated neighbor information with its own feature representation from the previous layer and inputs it into a fully connected layer and activation function to update and generate its own feature representation in the current layer. The feature information includes the voltage, injected power, and risk state value of the neighbor nodes. S2.2: By stacking multiple layers of graph convolutional layers, the final feature representation of each node is fused with the global information of its multi-hop neighbors; S2.3: For each operating scenario in the generated basic dataset, use the trained graph neural network model to calculate multi-dimensional node risk indicators for each node in the distribution network; the multi-dimensional node risk indicators include curtailment risk indicators, power quality risk indicators, and power grid safety and stability risk indicators. S2.4: For the multi-dimensional node risk indicators under each operating scenario, the improved CRITIC-entropy weighting method is used to dynamically assign weights to the curtailment risk indicators, power quality risk indicators and grid security and stability risk indicators, and the weighted summation of each node is used to obtain the comprehensive risk value. S2.5: The collection of comprehensive risk values ​​of all nodes under all scenarios and at all times constitutes a time-varying comprehensive risk level matrix.

6. The risk assessment method for connecting new energy photovoltaic power generation to the distribution network as described in claim 5, characterized in that, The time-varying comprehensive risk level matrix is ​​a three-dimensional data structure. Its first dimension index represents the node number in the distribution network, the second dimension index represents the discretized time point, and the third dimension stores the comprehensive risk value of the node at the corresponding time point and the decomposed values ​​of various risk indicators. The generation process of the time-varying comprehensive risk level matrix is ​​as follows: for each probabilistic scenario in the basic dataset, the risk value of all nodes in the scenario at all time points is calculated using the coupled evaluation model. Then, a weighted average is performed based on the probability of each scenario in the basic dataset to finally obtain the time-varying comprehensive risk level matrix that reflects the overall risk level.

7. The risk assessment method for connecting new energy photovoltaic power generation to the distribution network as described in claim 1, characterized in that, The specific steps of S3 include: S3.1: Analyze the time-varying comprehensive risk level matrix, extract risk features from it, and transform it into a standardized graph query statement; the risk features include the risk type, risk level, and risk impact range characteristics at a specified time, at a specified node, or in a specified area; S3.2: In the pre-constructed risk knowledge graph of measures, execute the graph query statement to obtain a set of candidate response and disposal measures; the graph traversal query first locates the historical risk case node with the highest similarity to the current risk feature, and then finds all response and disposal measure nodes connected to the historical risk case node along the adoption measure relationship edge in the risk knowledge graph of measures; S3.3: Perform conflict verification and priority ranking on the set of candidate response measures, and generate a dynamic adaptive response plan that includes the execution order of specific measures, the timing of activation, the expected effect, and the resource allocation plan based on the priority ranking result; the conflict verification is to determine whether different measures will compete for the same control resources or produce offsetting effects when they are executed; the priority ranking is a comprehensive score based on the implementation cost of the measures, the speed of the expected effect decay risk, and the current real-time operating status of the power grid.

8. The risk assessment method for connecting new energy photovoltaic power generation to the distribution network as described in claim 7, characterized in that, The pre-constructed risk knowledge graph includes at least three types of entity nodes: risk event case nodes, response and disposal measure nodes, and power grid equipment nodes. The attributes of the risk event case node include the case occurrence time, risk type, risk level, affected power grid area, and weather conditions. The attributes of the response and disposal measure node include the measure type, required control resources, implementation cost, expected effect, and effective delay time. The attributes of the power grid equipment node include the equipment ID, equipment type, and equipment capacity. The entity nodes are connected by relational edges, including: the influence relationship between risk event cases and the affected power grid equipment, the adoption relationship between risk event cases and the response and handling measures taken, and the action relationship between the response and handling measures and the operated power grid equipment.

9. The risk assessment method for connecting new energy photovoltaic power generation to the distribution network as described in claim 1, characterized in that, The process of validating the dynamic adaptive response scheme in a digital twin-based distribution network simulation model includes: S4.1: Synchronously inject the current operating scenario data and the measures and instructions in the dynamic adaptive response scheme from the basic dataset into the distribution network digital twin simulation model to complete the initialization; S4.2: Run the initialized digital twin simulation model of the distribution network to simulate the dynamic response process of the distribution network within a preset time period after the execution of the dynamic adaptive response scheme. Record key performance indicators during the simulation process. The key performance indicators include system voltage deviation, frequency fluctuation, curtailed solar power, and equipment load rate. S4.3: Compare the simulated key performance indicators with the preset standard thresholds for each indicator. If all key performance indicators are better than the standard thresholds, the verification is deemed successful. If any key performance indicator exceeds the standard threshold, the verification is deemed unsuccessful. When the verification fails, the incremental data generated by the simulation is recorded. The incremental data includes voltage over-limit records, power over-limit records, power quality indicators, equipment operating status, and system stability indicators.

10. A risk assessment system for connecting new energy photovoltaic power generation to a distribution network, used to implement the risk assessment method for connecting new energy photovoltaic power generation to a distribution network as described in any one of claims 1-9, characterized in that, include: Data construction module, coupled evaluation model module, solution generation module, verification and optimization module; The data construction module is used to collect solar resource data and power distribution network load data, and generate a probabilistic scenario set covering typical and extreme operating scenarios through data decomposition, joint probabilistic modeling, and constraint sampling operations. The coupled assessment model module integrates graph neural networks and risk propagation theory to perform risk calculations on various operating scenarios in the basic dataset and finally outputs a time-varying comprehensive risk level matrix. The solution generation module, based on the time-varying comprehensive risk level matrix, matches it with a pre-constructed risk knowledge graph of measures, and generates a dynamic adaptive response solution through conflict verification and priority sorting. The verification and optimization module is used to verify the effectiveness of the dynamic adaptive response scheme through a digital twin simulation model. For schemes that do not meet the preset standards, the incremental data generated during the simulation process is used as feedback to update the parameters of the coupled evaluation model and optimize the basic dataset.

Images