Method and system for evaluating distributed photovoltaic access capability of power distribution network

Abstract

The invention discloses a distributed photovoltaic access capability assessment method and system for a power distribution network, and the method comprises the steps: building a photovoltaic output probability model based on irradiance distribution and a load probability model based on statistical characteristics, and carrying out the random sampling of the photovoltaic output probability model and the load probability model, and respectively generating a photovoltaic output data set and a load data set; calculating to obtain photovoltaic limit access capacity data sets corresponding to different loads and different topologies; an MPNN model is adopted as a state estimation model, and a network structure and parameters of the MPNN model are set; constructing graph data by using node injection power and branch load flow power results obtained by probabilistic load flow calculation; performing offline training on the MPNN model by using multi-section historical graph data; and inputting the real-time graph data set of the power distribution network into the trained MPNN model to obtain the real-time photovoltaic limit access capacity of the power distribution network. According to the method, the distributed photovoltaic access potential of the power distribution network can be effectively improved.

Description

Technical Field

[0001] This invention relates to the field of distributed photovoltaic (PV) grid connection capability assessment technology, and in particular to a method for assessing distributed PV grid connection capability. Background Technology

[0002] Compared to traditional power systems, new power systems have undergone significant changes in terms of main power sources, system characteristics, and objectives. Solar photovoltaic (PV) power generation, as a representative of renewable energy, possesses the characteristics of being clean, safe, reliable, and not limited by the geographical distribution of resources. However, the large-scale integration of distributed PV has significantly altered the energy structure, operating mechanism, and balance mode within the distribution network, potentially leading to problems such as voltage exceeding limits and line overload. Simultaneously, the uncertainty within the distribution network system has increased significantly. Furthermore, compared to transmission networks, distribution networks not only have more complex load types but also more flexible and variable topologies and parameters. In this context, the observability of the distribution network deteriorates, posing risks to its safe operation. Moreover, the integration capacity and methods of distributed PV still involve many uncertainties, making it impossible to guarantee that all indicators of the distribution network will meet the actual requirements of engineering projects under various circumstances. Therefore, optimizing the allocation of distribution network resources and maximizing the potential for distributed PV integration is a problem that needs to be addressed. Summary of the Invention

[0003] The technical problem to be solved by this invention is: how to optimize the allocation of distribution network resources and maximize the potential of distributed photovoltaic access in the distribution network.

[0004] To address the above problems, this invention provides a method for assessing the distributed photovoltaic (PV) grid connection capability of a distribution network, comprising:

[0005] Step 1: Establish a photovoltaic output probability model based on irradiance distribution and a load probability model based on statistical characteristics. The photovoltaic output probability model and the load probability model are randomly sampled to generate photovoltaic power output dataset and load dataset, respectively.

[0006] Using photovoltaic power output dataset and load dataset as input, the Monte Carlo simulation method is used to perform probabilistic power flow calculations on distribution networks under different topologies to obtain a complete power flow solution dataset. Then, the branch current dataset and total network loss dataset are calculated through circuit relationships. The power flow solution dataset includes node voltage, node injected power, and branch power flow.

[0007] The cumulative distribution function and corresponding upper limit value of the photovoltaic output dataset, node voltage dataset, branch current dataset, and total network loss dataset are calculated respectively. Under the condition that the voltage of each node, the current of each branch, and the network loss of the distribution network do not exceed the limit, the limit photovoltaic access capacity of the distribution network is determined, and the photovoltaic limit access capacity dataset corresponding to different loads and different topologies is obtained.

[0008] Step 2: Use the MPNN model as the state estimation model, and set the network structure and parameters of the MPNN model;

[0009] Step 3: Construct graph data using the node injection power and branch power flow results obtained from the probabilistic power flow calculation in Step 1;

[0010] Step 4: Use multi-section historical map data to train the MPNN model offline;

[0011] Step 5: Use the real-time load information and topology information of the distribution network to form the input space of the MPNN model, and input it into the trained MPNN model to obtain the real-time photovoltaic limit access capacity of the distribution network.

[0012] In the aforementioned method for assessing the distributed photovoltaic (PV) grid connection capability, in step 1, the output power of the PV power source is determined by the irradiance, and the distribution of irradiance over a day or multiple hours is considered as a Beta distribution, with the probability density function... as follows:

[0013] (1)

[0014] In the formula, It is a gamma function; This represents the actual light intensity. This represents the maximum light intensity. and These are the shape parameters one and two of the Beta distribution, respectively, which depend on the mean of the illumination intensity. and variance :

[0015] (2)

[0016] (3)

[0017] Neglecting the nonlinearity of the photovoltaic power output characteristics, the photovoltaic output is directly proportional to the light intensity, expressed as:

[0018] (4)

[0019] In the formula, This refers to the total area of ​​the photovoltaic panels; The photoelectric conversion efficiency of photovoltaic power sources. When the photovoltaic panel receives irradiance At that time, the electrical power that the photovoltaic module can output.

[0020] In the aforementioned method for assessing the distributed photovoltaic (PV) grid connection capacity of a distribution network, the load probability model assumes that the distribution network load is a continuous load and follows a normal distribution over a certain period of time, with the load's active power... reactive power probability model , They are respectively:

[0021] (5)

[0022] (6)

[0023] In the formula, , These are the average values ​​of the load's active power and reactive power, respectively. , These are the standard deviations of the active power and reactive power of the load, respectively. This indicates that the probability density decreases exponentially with deviation from the mean.

[0024] In the aforementioned method for assessing the distributed photovoltaic (PV) grid access capability, step 2 involves the MPNN model, where the forward propagation process includes two phases: message passing and reading. During the message passing phase, a graph is defined. ,in, Represents the set of nodes in the graph; This represents the set of edges in the graph, which undergoes multiple message passing processes during the message passing phase.

[0025] For a specific node , is represented as:

[0026] (7)

[0027] In the formula, For nodes In time The information received; It is a node All adjacent points; For nodes In time eigenvectors, Let ω be the node state at the t-th iteration; It is a node and nodes Edge features; For nodes in time The message function; Formula (7) represents the node The received information comes from the node. Original state and the status of surrounding nodes and the edge features connected to it ;

[0028] Update the nodes after generating the information:

[0029] (8)

[0030] In the formula, The node update function updates the original node state. and information As input, we obtain the new node state. ;

[0031] Based on the above node modeling, the edges in the graph are modeled using the following formula:

[0032] (9)

[0033] (10)

[0034] In the formula, This means fusing the edge's own features with the features of its two endpoints to generate the message vector that the edge will transmit in the next step. This means updating the hidden state of an edge using the edge features from the previous time step and the newly generated message, so that the edge representation gains new semantic information. To connect nodes in the t-th iteration The eigenvectors of the edges connected to node ω, Represents the edge message passing function Intermediate results obtained from the calculation This represents the connection node at the (t+1)th iteration. The eigenvectors of the edges connected to node ω.

[0035] In the aforementioned method for assessing the distributed photovoltaic (PV) grid connection capability of a distribution network, step 2 involves a node message passing function as shown in equation [equation missing]. As shown:

[0036] (11)

[0037] In the formula, This is the learnable parameter matrix that acts on the nodes of the graph structure; To represent nodes The aggregated message vector at the (t+1)th iteration. Let represent the function used to calculate the edge weights in the t-th iteration. For the t-th iteration, node The feature vector of the edge between node ω;

[0038] The edge message passing function is shown in equation (12):

[0039] (12)

[0040] In the formula, To achieve a learnable parameter matrix for edge message interaction in a graph structure;

[0041] The state update function is implemented using a gated recurrent unit (GRU), and the formula is as follows:

[0042] (13)

[0043] (14)

[0044] During the reading phase, a reading function is used to calculate the feature vector based on the entire image, represented as:

[0045] (15)

[0046] In the formula, For nodes After passing through GNN The final node features obtained after the iteration. This is the final output vector; The reading function reads the graph embedding vector from the last hidden state of all nodes in the graph; the vector read from the last GNN layer. As input to the subsequent LSTM network part.

[0047] In the aforementioned method for assessing the distributed photovoltaic access capability of a distribution network, in step 3, a graph dataset is constructed using a two-dimensional vector consisting of load active power and load reactive power as node features, and a six-dimensional vector consisting of active and reactive power at the beginning and end of the branch, branch conductance, and branch susceptance as edge features.

[0048] In the aforementioned method for assessing the distributed photovoltaic (PV) grid connection capability of a distribution network, step 3 involves constructing a graph dataset. The expression is:

[0049] (16)

[0050] (17)

[0051] (18)

[0052] In the formula, The feature set representing a node; The feature set representing a branch; and These represent the injected active power and reactive power at node i, respectively. and Representing branches The active and reactive power at the head end; and Representing branches Active and reactive power at the end; and Representing branches Its electrical conductivity and susceptivity.

[0053] In the aforementioned method for assessing the distributed photovoltaic (PV) grid access capability, in step 4, the constructed graph dataset is input into the MPNN model. The MPNN model automatically learns and aggregates the feature information of nodes and edges through the message passing and state update mechanism defined by multiple GNN layers, thereby extracting graph structure features that are highly related to PV grid access capability.

[0054] The graph structure features are ultimately integrated and mapped through a fully connected layer to output an estimated value of the photovoltaic limit access capacity of the distribution network.

[0055] The difference between the estimated value output by the model and the actual photovoltaic limit access capacity obtained in step 1 is calculated using the mean square error of the loss function (MSE).

[0056] Using the backpropagation algorithm, all learnable weight parameters in the MPNN model are updated based on the gradient calculated from the loss function.

[0057] By repeatedly executing the forward propagation, loss calculation, and backpropagation processes, the model weights are continuously optimized, ultimately yielding the state estimation model.

[0058] In the aforementioned method for assessing the distributed photovoltaic access capability of a distribution network, step 5 involves using graph convolution to extract structured features from the distribution network topology. The extracted structured features are used to characterize the structural role of nodes in the topology, their connectivity strength, and their influence on power distribution. These structured features, along with load information, serve as the structural input features of the MPNN model.

[0059] In the aforementioned method for assessing the distributed photovoltaic (PV) grid connection capability of a distribution network, step 5 involves selecting a determination coefficient δ. R2 and mean absolute error δ MSE The evaluation metric for the MPNN model is calculated using the following formula:

[0060] (19)

[0061] (20)

[0062] In the formula, yi This is the actual value; This is a predicted value; is the average of the true values; N is the total number of samples.

[0063] A computer system includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the above-described method for assessing the distributed photovoltaic access capability of a distribution network.

[0064] A computer-readable storage medium storing a computer program, which, when executed by a processor, implements the aforementioned method for assessing the distributed photovoltaic (PV) grid connection capability of a distribution network. The beneficial effects of this invention are as follows: The method of this invention, based on historical measurement data from multiple sections of the distribution network, utilizes deep learning to learn the potential relationship between the current state of the distribution network and the available PV capacity, ultimately obtaining a deep learning-based assessment model for the distributed PV grid connection capability. Under the premise of meeting the actual safety constraints of a complex and variable distribution network topology, it optimizes the allocation of distribution network resources by analyzing distributed PV connection bottlenecks, thereby effectively improving the distributed PV connection potential of the distribution network while ensuring system safety. Attached Figure Description

[0065] Figure 1 This is a flowchart of the method for assessing the distributed photovoltaic access capability of a distribution network in Embodiment 1 of the present invention;

[0066] Figure 2 This is a message passing graph neural network structure diagram in Embodiment 1 of the present invention;

[0067] Figure 3 This is the topology diagram of the IEEE 33-node computation example in Embodiment 1 of the present invention;

[0068] Figure 4 This refers to the load change trend over four consecutive days in Embodiment 1 of the present invention;

[0069] Figure 5 This is a graph of the cumulative distribution function of branch currents in the IEEE 33-node distribution network in Embodiment 1 of the present invention;

[0070] Figure 6 These are the fast evaluation results of MPNN under different topologies in Embodiment 1 of the present invention. Detailed Implementation

[0071] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0072] Example 1

[0073] like Figure 1 As shown in the figure, this embodiment provides a method for assessing the distributed photovoltaic (PV) grid connection capability of a distribution network, including:

[0074] Step 1: Obtain neural network training data, including:

[0075] A photovoltaic output probability model based on irradiance distribution and a load probability model based on statistical characteristics are established. The photovoltaic output probability model and the load probability model are randomly sampled to generate photovoltaic power output dataset and load dataset, respectively.

[0076] Using photovoltaic power output dataset and load dataset as input, the Monte Carlo simulation method is used to perform probabilistic power flow calculations on distribution networks under different topologies to obtain a complete power flow solution dataset. Then, the branch current dataset and total network loss dataset are calculated through circuit relationships. The power flow solution dataset includes node voltage, node injected power, branch power flow, etc.

[0077] Finally, the cumulative distribution function and corresponding upper limit value of the photovoltaic output dataset, node voltage dataset, branch current dataset, and total network loss dataset are calculated respectively. Under the condition that the voltage of each node, the current of each branch, and the network loss of the distribution network do not exceed the limit, the limit photovoltaic access capacity of the distribution network is determined, and the photovoltaic limit access capacity datasets corresponding to different loads and different topologies are obtained.

[0078] Step 2: Use the MPNN model as the state estimation model, and set the network structure and parameters of the MPNN model. The network structure includes the total number of layers in the model and the number of GNN layers. The hyperparameters include the optimizer type, learning rate, and number of iterations.

[0079] Step 3: Construct graph data using the node injection power and branch power flow results obtained from the probabilistic power flow calculation in Step 1;

[0080] A graph dataset is constructed using a two-dimensional vector consisting of active and reactive loads as node features, and a six-dimensional vector consisting of active and reactive loads at the beginning and end of branches, branch conductance, and branch susceptance as edge features.

[0081] Step 4: Use multi-section historical map data to train the MPNN model offline;

[0082] Step 5: Perform online evaluation. The real-time load information and topology information of the distribution network constitute the complete input space of the MPNN model. Input the model into the trained MPNN model to obtain the real-time photovoltaic limit access capacity of the distribution network.

[0083] Considering that graph convolution can fully utilize the correlation between nodes, edges, and network structures and has the advantage of modeling complex network relationships, graph convolution is used to extract structured features from the distribution network topology. The extracted structured features are used to characterize the structural role of nodes in the topology, connectivity strength, and influence on power distribution. These structured features, together with load information, serve as the structural input features of the MPNN model.

[0084] In step 1, in the photovoltaic output probability model, the output power of the photovoltaic power source is determined by the light intensity. The distribution of light intensity over a day or several hours is considered as a Beta distribution, and the probability density function is... as follows:

[0085] (1)

[0086] In the formula, It is a gamma function; This represents the actual light intensity. This represents the maximum light intensity. and These are the shape parameters one and two of the Beta distribution, respectively, which depend on the mean of the illumination intensity. and variance :

[0087] (2)

[0088] (3)

[0089] Neglecting the nonlinearity of the photovoltaic power output characteristics, the photovoltaic output is directly proportional to the light intensity, expressed as:

[0090] (4)

[0091] In the formula, This refers to the total area of ​​the photovoltaic panels; The photoelectric conversion efficiency of photovoltaic power sources. When the photovoltaic panel receives irradiance At that time, the electrical power that photovoltaic modules can output is extremely sensitive to light intensity, and the light intensity exhibits obvious non-negative, skewed, and boundary-type distribution characteristics. Considering that the Beta distribution has good flexibility, it can be used to determine the shape parameters. and Simultaneously fitting the skewness, kurtosis, and diurnal time-history characteristics of light intensity, and strictly ensuring that the variables are within the physically feasible range, the Beta distribution is chosen to characterize the intra-day fluctuation pattern of light intensity, thereby more realistically reflecting the uncertainty and randomness of photovoltaic power.

[0092] In the load probability model, the distribution network load exhibits strong volatility and randomness. Over a period of time, the distribution network load is a continuous load and follows a normal distribution. The active power of the load... reactive power probability model , They are respectively:

[0093] (5)

[0094] (6)

[0095] In the formula, , These are the average values ​​of the load's active power and reactive power, respectively. , These are the standard deviations of the active power and reactive power of the load, respectively. This indicates that the probability density decays exponentially with deviation from the mean, a characteristic of the normal distribution. The reason the normal distribution is chosen to describe the active power of the distribution network load is... With reactive power This is because power distribution load exhibits continuous fluctuations over short timescales and is influenced by a large number of independent random factors, exhibiting statistical characteristics consistent with the central limit theorem. Considering that the normal distribution can effectively characterize the central tendency and dispersion of load fluctuations with its mean and variance, and has advantages such as simple modeling, strong parameter measurability, and high consistency with actual load statistical laws, the normal distribution is chosen to construct the probabilistic model of the load to obtain stable inputs that can be used for subsequent risk assessment and ultimate capacity calculation.

[0096] like Figure 2 As shown, in step 2, ReLU is the activation function in the MPNN model, and the forward propagation process of the MPNN model includes two stages: message passing and message reading. In the message passing stage, a graph is defined. ,in, Represents the set of nodes in the graph; This represents the set of edges in the graph. Multiple message passing processes are executed during the message passing phase.

[0097] For a specific node , is represented as:

[0098] (7)

[0099] In the formula, For nodes In time The information received; It is a node All adjacent points; For nodes In time eigenvectors, Let ω be the node state at the t-th iteration; It is a node and nodes Edge features; For nodes in time The message function; Formula (7) represents the node The received information comes from the node. Original state and the status of surrounding nodes and the edge features connected to it .

[0100] Update the nodes after generating the information:

[0101] (8)

[0102] In the formula, The node update function updates the original node state. and information As input, we obtain the new node state. .

[0103] Based on the above node modeling, the edges in the graph are modeled using the following formula:

[0104] (9)

[0105] (10)

[0106] In the formula, This means fusing the edge's own features with the features of its two endpoints to generate the message vector that the edge will transmit in the next step. This means updating the hidden state of an edge using the edge features from the previous time step and the newly generated message, so that the edge representation gains new semantic information. To connect nodes in the t-th iteration The eigenvectors of the edges connected to node ω, Represents the edge message passing function The intermediate results obtained from the calculation are not the final features. This represents the connection node at the (t+1)th iteration. The eigenvectors of the edges connected to node ω.

[0107] The node message passing function is as follows: As shown:

[0108] (11)

[0109] In the formula, This is the learnable parameter matrix that acts on the nodes of the graph structure; To represent nodes The aggregated message vector at the (t+1)th iteration. Let represent the function used to calculate the edge weights in the t-th iteration. For the t-th iteration, node The feature vector of the edge between node ω;

[0110] The edge message passing function is shown in equation (12):

[0111] (12)

[0112] In the formula, To achieve a learnable parameter matrix for edge message interaction in a graph structure;

[0113] The state update function is implemented using a gated recurrent unit (GRU), and the formula is as follows:

[0114] (13)

[0115] (14)

[0116] During the reading phase, a reading function is used to calculate the feature vector based on the entire image, represented as:

[0117] (15)

[0118] In the formula, For nodes After passing through GNN The final node features obtained after the iteration. Let be a node in the graph. A graph, consisting of a set of nodes and a set of edges, serves as the input structure for the entire graph neural network computation. This is the final output vector; The reading function reads the graph embedding vector from the last hidden state of all nodes in the graph; the vector read from the last GNN layer. As input to the subsequent LSTM network section. Read function. It is a fully connected layer, and the weight parameters are learned through the backpropagation algorithm during the end-to-end training of the MPNN model. Considering that graph convolution can make full use of the correlation between nodes, edges and network structure and has the advantage of modeling complex network relationships, graph convolution is chosen to extract structured features of the power distribution network.

[0119] In step 3, the constructed graph dataset The expression is:

[0120] (16)

[0121] (17)

[0122] (18)

[0123] In the formula, The feature set representing a node; The feature set representing a branch; and These represent the injected active power and reactive power at node i, respectively. and Representing branches The active and reactive power at the head end; and Representing branches Active and reactive power at the end; and Representing branches Its electrical conductivity and susceptivity.

[0124] In step 4, the constructed graph dataset is input into the MPNN model. The MPNN model automatically learns and aggregates the feature information of nodes and edges through the message passing and state update mechanism defined by multiple GNN layers, thereby extracting graph structure features that are highly related to photovoltaic access capabilities.

[0125] The graph structure features are ultimately integrated and mapped through a fully connected layer to output an estimated value of the photovoltaic limit access capacity of the distribution network.

[0126] Next, the difference between the estimated value output by the model and the actual photovoltaic limit access capacity obtained in step 1 is calculated using the mean square error of the loss function (MSE).

[0127] Subsequently, using the backpropagation algorithm, all learnable weight parameters in the MPNN model are updated based on the gradient calculated from the loss function.

[0128] Finally, by iteratively executing the forward propagation, loss calculation, and backpropagation processes, the model weights are continuously optimized, ultimately resulting in an accurate, stable state estimation model that can adapt to different distribution network topologies.

[0129] In step 5, the graph dataset is used as the input to the MPNN model, and the photovoltaic limit access capacity of the distribution network is used as the output of the MPNN model.

[0130] Select the coefficient of determination δ R2 and mean absolute error δMSE The evaluation metric for the MPNN model is calculated using the following formula:

[0131] (19)

[0132] (20)

[0133] In the formula, y i This is the actual value; This is a predicted value; The average of the true values; N is the total number of samples, taking δ into account. R2 It can measure a model's ability to explain the variance of real data, reflect the model's fit to the overall trend, and has the advantages of being intuitive, having a fixed range (0-1), and being able to quantify the "goodness of fit" of a model. Therefore, δ is chosen. R2 To evaluate the overall prediction accuracy of the MPNN model, and at the same time, δ MSE Insensitive to outliers, it accurately reflects the average deviation of prediction errors, and possesses advantages such as stability, ease of interpretation, and better representation of the absolute magnitude of prediction deviations in engineering scenarios. Therefore, δ is chosen. MSE This is used to measure the error level of the model in the process of predicting the maximum access capacity.

[0134] To verify the superiority of the algorithm of this invention in the distribution network, this embodiment conducts a simulation test based on the classic IEEE 33-node distribution network test system. The topology diagram of the IEEE 33-node simulation is shown below. Figure 3 As shown, nodes 5, 12, 20, 24, and 30 are the access nodes for distributed photovoltaic (PV1~PV5) power sources, and node 1 is the balancing node connected to the main power grid. Five topologies are also set up, with specific topology parameters shown in Table 1. For ease of example analysis, it is assumed that the conductor type used in each line of IEEE 33 nodes is JKLYJ-240, and the maximum current carrying capacity of the lines is 425A. The daily load curve used in this invention is shown in Table 1. Figure 4 As shown.

[0135] The hardware configuration of the MPNN-based distribution network renewable energy access capability assessment model is an AMD Ryzen 7 5800H CPU and an Nvidia RTX3060 GPU. It is programmed in Python and implemented using the PyTorch library. This invention sets a graph convolutional network (GCN) model as a contrast model to MPNN. Since GCN struggles to fully account for the characteristics of distribution network branches, the node feature vectors in the input graph data are the same as those in MPNN. For branch features, only the connection relationships between nodes, i.e., the adjacency matrix, can be considered.

[0136] Since the voltage and network loss at each node are within acceptable limits after photovoltaic grid connection, only the cumulative distribution function of the branch current at IEEE 33 node is shown, such as... Figure 5 As shown, CDF is the cumulative probability density function. According to... Figure 5 It can be seen that the current of branch 1 has reached the limit of current carrying capacity. If the photovoltaic access capacity is further increased, the current of branch 1 will be overloaded. At this time, the limit of photovoltaic access capacity of the distribution network is 400kW.

[0137] Figure 6 This chart compares the rapid evaluation values ​​of the photovoltaic (PV) grid connection capacity of the distribution network based on the MPNN model with the actual values. Figure 6 It can be seen that the fast evaluation algorithm for the potential of new energy access in distribution networks proposed in this invention performs well in actual calculation examples.

[0138] Table 2 compares the accuracy of the MPNN model and the GCN model in evaluating the system's photovoltaic limit access capacity under the IEEE 33-node example. It can be seen that the average absolute error of the MPNN evaluation result is reduced by 4.541e-3MW compared with the GCN evaluation result, and the coefficient of determination is increased by 6.86%, thus proving that the MPNN-based distribution network renewable energy access capacity evaluation model proposed in this invention has better accuracy.

[0139] Table 1: Topology Parameter Table

[0140]

[0141] Table 2: Comparison of the accuracy of MPNN model and GCN model in evaluating the maximum photovoltaic grid connection capacity of the system under the IEEE 33-node example.

[0142]

[0143] Example 2

[0144] A computer system includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to perform the steps of the method as described in Embodiment 1.

[0145] Example 3

[0146] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the method as described in Example 1.

[0147] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0148] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0149] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0150] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for assessing the distributed photovoltaic (PV) grid connection capability of a distribution network, characterized in that, The method comprises the following steps: Step 1, establish a photovoltaic output probability model based on irradiance distribution and a load probability model based on statistical characteristics, and randomly sample the photovoltaic output probability model and the load probability model to generate a photovoltaic output data set and a load data set respectively; Step 2, use the photovoltaic output data set and the load data set as inputs, and use the Monte Carlo simulation method to perform probabilistic power flow calculation on the distribution network under different topologies to obtain a complete power flow solution data set, and then calculate the branch current data set and the total network loss data set through circuit relationship; the power flow solution data set comprises node voltage, node injection power and branch power flow; Step 3, calculate the cumulative distribution function and the corresponding upper limit value of the photovoltaic output data set, the node voltage data set, the branch current data set and the total network loss data set, and determine the limit photovoltaic access capacity of the distribution network under the condition that the node voltage, the branch current and the network loss of the distribution network do not exceed the limit, to obtain a photovoltaic limit access capacity data set corresponding to different loads and different topologies; Step 4, use the MPNN model as a state estimation model, and set the network structure and parameters of the MPNN model; Step 5, use the node injection power and the branch power flow obtained by the probabilistic power flow calculation to construct graph data; Step 6, use the multi-section historical graph data to perform offline training on the MPNN model; Step 7, input the real-time load information and the distribution network topology information of the distribution network into the input space of the MPNN model, and input them into the trained MPNN model to obtain the real-time photovoltaic limit access capacity of the distribution network.

2. The method of claim 1, wherein, In step 1, in the photovoltaic output probability model, the output power of the photovoltaic power source is determined by the light intensity, the distribution of the light intensity in a day or more hours is regarded as a Beta distribution, and the probability density function is as follows: f(x) = x(a-1)(b-1) / B(a, b) x(a, b) (1) As follows: （1） wherein is the gamma function; is the actual light intensity; is the maximum light intensity; and are the shape parameter one and parameter two, respectively, of the Beta distribution, depending on the mean and variance of the light intensity: （2） （3） Neglecting the nonlinear part of the photovoltaic power output characteristics, the photovoltaic output is proportional to the illumination intensity, which is represented as: （4） wherein is the total area of the photovoltaic panel; is the photoelectric conversion efficiency of the photovoltaic power source, is the electric power that the photovoltaic assembly is able to output when the photovoltaic panel receives an irradiance of 1000 W / m2.

3. The method of claim 1, wherein, In the load probability model, the distribution of the load active power and the load reactive power of the distribution network in a period of time is continuous and obeys normal distribution. , ​ （5） （6） wherein, , are the mean values of the load active power, reactive power, respectively; , are the standard deviations of the load active power, reactive power, respectively, denotes an exponential decay of the probability density with the deviation from the mean value.

4. The method of claim 1, wherein, In step 2, in the MPNN model, the forward propagation process of the MPNN model includes two stages of message passing and reading; in the message passing stage, a graph is defined wherein, denotes a set of nodes in the graph; denotes a set of edges in the graph, and the message passing process is performed multiple times in the message passing stage; For a particular node is represented as: （7） In the formula, For nodes In time The information received; It is a node All adjacent points; For nodes In time eigenvectors, Let ω be the node state at the t-th iteration; It is a node and nodes Edge features; For nodes in time The message function; Formula (7) represents the node The received information comes from the node. Original state and the status of surrounding nodes and the edge features connected to it ; After generating the information, the node is updated: （8） wherein is the node update function, taking the original node state and information as input, resulting in the new node state ; Based on the above modeling of the node, the edge in the graph is modeled, and the formula is as follows: （9） （10） In the formula, indicates that the edge itself features and the two end node features are fused to generate the message vector of the edge to be transmitted in the next step, indicates that the hidden state of the edge is updated with the edge features of the last time and the newly generated message, so that the representation of the edge obtains new semantic information, is connected to the node is the feature vector of the edge of the node ω, indicates that the edge message passing function is an intermediate result calculated, is connected to the node is the feature vector of the edge of the node ω.

5. The method of claim 4, wherein, In step 2, the node message passing function is as shown in equation : （11） wherein, is a learnable parameter matrix acting on the graph structure nodes; is a representation of a node is the aggregated message vector at iteration t+1, is a function used to compute the edge weight in iteration t, is the representation of a node is the feature vector of the edge between node ω and node The edge message passing function is shown in formula (12): （12） In the formula, To realize the graph structure edge message interaction of the learnable parameter matrix; The state update function is realized by using a gated recurrent unit GRU, and the formula is: （13） （14） In the reading phase, the reading function is used to calculate the feature vector based on the whole graph, which is represented as: （15） wherein is a node After passing through the GNN the final node features obtained after the is the final output vector; is a readout function that reads the graph embedding vector from the last layer of hidden states of the whole graph nodes; the last GNN layer passes the read vector as input to the following LSTM network part.

6. The method of claim 1, wherein, In step 3, a two-dimensional vector composed of load active power and load reactive power is used as node features, and a six-dimensional vector composed of branch active power and reactive power, branch conductance and branch susceptance is used as edge features to construct a graph data set.

7. The method of claim 6, wherein, In step 3, the constructed graph dataset The expression is: （16） （17） （18） wherein denotes a characteristic set of a node; denotes a characteristic set of a branch; and denote the injected active and reactive power of a node i, respectively; and denote the active and reactive power at the beginning of a branch , respectively; and denote the active and reactive power at the end of a branch , respectively; and denote the conductance and susceptance of a branch , respectively.

8. The method of claim 1, wherein, In step 4, the constructed graph data set is input into the MPNN model, and the MPNN model automatically learns and aggregates the feature information of nodes and edges through the message passing and state update mechanism defined by multiple GNN layers, so as to extract the graph structure features highly related to the photovoltaic access capacity; The graph structure features are finally integrated and mapped through a fully connected layer to output the estimated value of the photovoltaic limit access capacity of the distribution network; The difference between the estimated value output by the model and the actual photovoltaic limit access capacity obtained in step 1 is calculated through the loss function mean square error MSE; Using the back propagation algorithm, all learnable weight parameters in the MPNN model are updated according to the gradient calculated by the loss function; Through multiple iterations of the forward propagation, loss calculation and back propagation processes, the model weights are continuously optimized, and finally the state estimation model is obtained.

9. The method of claim 1, wherein, In the step 5, the power distribution network topology is structured feature extraction using graph convolution, and the structured feature extracted is used to represent the structural role of the node in the topology, the connectivity strength and the influence law on the power distribution, and the structured feature and the load information are jointly used as the structural input features of the MPNN model.

10. The method of claim 1, wherein, In the step 5, the coefficient of determination δ R2 and the mean absolute error δ MSE are selected as the evaluation indexes of the MPNN model, and the calculation formula is as follows: （19） （20） where y i is the actual value; is the predicted value; is the average of the true values; N is the total number of samples.

11. A computer system comprising: The memory, the processor and the computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the power distribution network distributed photovoltaic access capacity evaluation method of any one of claims 1-10.

12. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the power distribution network distributed photovoltaic access capacity evaluation method of any one of claims 1-10.

Images