A photovoltaic power probability prediction method and system based on a graph neural network

Abstract

The application discloses a photovoltaic power probability prediction method and system based on a graph neural network, and relates to the field of photovoltaic power prediction.The method comprises the following steps: acquiring basic observation data of a photovoltaic power station; constructing a graph structure between photovoltaic power stations based on the basic observation data and generating node features; inputting the node features into a meteorological time sensing module to generate a dynamic control vector; in a prediction stage, inputting the graph structure and the node features into a graph neural network, performing multiple benchmark forward predictions based on the dynamic control vector, and obtaining multiple benchmark prediction results; performing statistical analysis on the multiple benchmark prediction results, constructing a probability density distribution of photovoltaic power, and outputting a prediction confidence interval.The application can quantify prediction uncertainty, improve prediction robustness in sudden weather scenarios, provide quantitative risk support for power grid dispatching, and ensure safe and optimized operation of a high photovoltaic penetration power grid.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power prediction, and in particular to a photovoltaic power probabilistic prediction method and system based on graph neural networks. Background Technology

[0002] As renewable energy continues to increase its share in the global energy mix, photovoltaic (PV) power generation systems are being widely deployed due to their cleanliness and sustainability. However, because PV power output is heavily dependent on weather conditions, it exhibits strong volatility and high uncertainty. Therefore, accurate and stable prediction of PV power has become a key technical aspect for ensuring the safe operation and optimized dispatch of the power grid. In recent years, data-driven methods have gradually become mainstream, especially deep learning models, which have demonstrated strong feature extraction and nonlinear fitting capabilities in PV prediction. Meanwhile, graph neural networks, due to their advantages in handling spatial relationships and structured data, have been gradually introduced into PV power prediction tasks involving multiple sites and strong spatial correlations, achieving some success.

[0003] While graph neural networks can effectively model the spatial relationships between photovoltaic power plants, most current research focuses primarily on the static representation of node features and the transmission of structural information, lacking sufficient modeling of temporal dynamics and meteorological environmental changes. Particularly in the Dropout operation between nodes and edges, static or stochastic strategies are commonly used, failing to reflect the dynamic changes in node reliability under complex weather conditions. Furthermore, existing technologies mostly rely on single-prediction outputs, ignoring the uncertainty of predicted values. This makes it difficult to provide quantifiable risk assessment criteria for grid dispatching and to generate important indicators reflecting prediction stability, such as confidence intervals and probability distributions. These limitations are particularly pronounced when facing highly uncertain scenarios such as rapid changes in cloud cover and sudden weather events. Summary of the Invention

[0004] The purpose of this application is to provide a photovoltaic power probability prediction method and system based on graph neural networks, which can quantify prediction uncertainty, improve prediction robustness in sudden weather scenarios, provide quantitative risk support for power grid dispatch, and ensure the safe and optimized operation of power grids with high photovoltaic penetration.

[0005] To achieve the above objectives, this application provides the following solution: Firstly, this application provides a photovoltaic power probability prediction method based on graph neural networks, including: Acquire basic observation data for photovoltaic power plants; the basic observation data includes historical meteorological data, geographic coordinate information, and historical power data; Based on the aforementioned basic observation data, a graph structure is constructed between photovoltaic power stations, and node features are generated; The node features are input into the meteorological time sensing module to generate a dynamic control vector; In the prediction phase, the graph structure and the node features are input into the graph neural network, and multiple baseline forward predictions are made based on the dynamic control vector rows to obtain multiple baseline prediction results. Statistical analysis is performed on multiple benchmark prediction results to construct the probability density distribution of photovoltaic power and output the prediction confidence interval.

[0006] Secondly, this application provides a photovoltaic power probability prediction system based on graph neural networks, the system being used to execute the aforementioned photovoltaic power probability prediction method based on graph neural networks, the system comprising: The data acquisition module is used to acquire basic observation data of the photovoltaic power station; the basic observation data includes historical meteorological data, geographic coordinate information and historical power data; The data processing module is used to construct a graph structure between photovoltaic power stations based on the basic observation data and generate node features; The dynamic control vector generation module is used to input the node features into the meteorological time sensing module to generate a dynamic control vector; The prediction module is used to input the graph structure and the node features into the graph neural network during the prediction phase, and make multiple baseline forward predictions based on the dynamic control vector row to obtain multiple baseline prediction results. The results output module is used to perform statistical analysis on multiple benchmark prediction results, construct the probability density distribution of photovoltaic power, and output the prediction confidence interval.

[0007] According to the specific embodiments provided in this application, this application has the following technical effects: This application first integrates meteorological, geographical, and power multi-source observation data to construct the graph structure and node characteristics of photovoltaic power plants. Then, through a meteorological time-sensing module, it mines the temporal evolution patterns of meteorological data to generate dynamic control vectors, replacing the traditional static and random graph dropout strategy. This allows for adaptive adjustment of the credibility of each photovoltaic node and spatially related edges according to meteorological changes, simultaneously completing the joint modeling of the spatial correlation and temporal dynamic characteristics of photovoltaic data. This effectively suppresses sudden meteorological noise interference and improves the accuracy of photovoltaic power prediction for multiple sites. Furthermore, relying on the dynamic control vectors, it performs multiple baseline forward predictions using a graph neural network. The probability density distribution and prediction confidence interval of photovoltaic power are statistically obtained from the multiple baseline prediction results, enabling a quantifiable expression of prediction uncertainty.

[0008] This application addresses the shortcomings of existing graph neural networks, which cannot dynamically adapt to changes in meteorological time series and can only output single-point prediction values. It has stronger prediction robustness in highly volatile scenarios such as sudden changes in cloud cover and extreme weather, and can provide quantitative risk assessment basis for power grid dispatch, supporting the safe operation and dispatch optimization of high-penetration photovoltaic power grids. Attached Figure Description

[0009] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0010] Figure 1 This is a flowchart illustrating a photovoltaic power probability prediction method based on a graph neural network, provided in one embodiment of this application. Detailed Implementation

[0011] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0012] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0013] In one exemplary embodiment, such as Figure 1 As shown, a photovoltaic power probability prediction method based on graph neural network is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server. In this embodiment, the method is described using a server as an example, and includes the following steps S1 to S5.

[0014] S1: Obtain basic observation data of the photovoltaic power station; the basic observation data includes historical meteorological data, geographic coordinate information and historical power data.

[0015] S2: Construct a graph structure between photovoltaic power stations based on the basic observation data and generate node features.

[0016] S3: Input the node features into the meteorological time sensing module to generate a dynamic control vector.

[0017] S4: In the prediction phase, the graph structure and the node features are input into the graph neural network, and multiple baseline forward predictions are made based on the dynamic control vector row to obtain multiple baseline prediction results.

[0018] S5: Perform statistical analysis on multiple benchmark prediction results, construct the probability density distribution of photovoltaic power, and output the prediction confidence interval.

[0019] By implementing steps S1 to S5 above, and by introducing a meteorological time-sensing module and a dynamic control vector, and combining multiple baseline forward prediction output probability distributions, the modeling of power prediction uncertainty and risk interpretability analysis were achieved, enhancing the model's adaptability to prediction fluctuations and scheduling response capabilities under complex environments.

[0020] Furthermore, in step S1, the sampling frequency of historical power data is 5 minutes; the historical power data includes at least the active power value of each photovoltaic power station at each sampling time, in MW.

[0021] Historical meteorological data includes: solar irradiance, temperature, humidity, wind speed, and cloud cover; geographic coordinate information includes latitude and longitude. The unit for solar irradiance is W / m². 2 Temperature is expressed in °C, humidity is represented as relative humidity from 0 to 100%, wind speed is expressed in m / s, and cloud cover is represented as a value between 0 and 1, where 0 represents no clouds and 1 represents complete cloud cover. Using these data with clearly defined physical dimensions as input avoids treating "meteorological conditions" as merely an unquantifiable abstract description.

[0022] Before constructing the graph structure and training the model, the data from different photovoltaic power plants were aligned according to a unified timestamp. Historical power data retained a 5-minute sampling granularity for calculating power correlation between sites; meteorological data and power labels used for node feature construction were aligned with a 1-hour time step. For cases where multiple power sampling values ​​exist within the same hour, the arithmetic mean of the effective power sampling values ​​within that hour was taken as the hourly power value; for cases where multiple meteorological sampling values ​​exist within the same hour, the arithmetic mean of each meteorological parameter was taken as the corresponding hourly meteorological parameter.

[0023] For meteorological data with a continuous missing period of no more than 2 hours, local weighted linear interpolation between the two valid sampling times before and after the missing interval is used for filling, with valid sampling values ​​closer to the missing time having a larger weight. For meteorological data with a continuous missing period of more than 2 hours, the median of the meteorological parameters corresponding to the same hour in the most recent 7 valid dates for the photovoltaic power station is used for filling. For historical power data with a continuous missing period of no more than 30 minutes, linear interpolation is used for filling; for historical power data with a continuous missing period of more than 30 minutes, the missing interval is not used in the calculation of the power correlation coefficient between two photovoltaic power stations. If the number of common valid power sampling points for two photovoltaic power stations in the past 7 days is less than 80% of all sampling points, the power correlation rule between the two photovoltaic power stations is marked as invalid, and this rule is not used as the basis for establishing an edge connection relationship between them.

[0024] For data exceeding the physical range of photovoltaic equipment, boundary correction is performed: power values ​​less than 0 are set to 0, and power values ​​greater than the rated installed capacity of the corresponding photovoltaic power station are set to that rated installed capacity; solar irradiance less than 0 is set to 0; cloud cover less than 0 is set to 0, and cloud cover greater than 1 is set to 1; humidity less than 0% is set to 0%, and humidity greater than 100% is set to 100%. This boundary correction is used to eliminate non-physical values ​​caused by momentary sensor malfunctions without altering normal fluctuations within the physical range.

[0025] The basic observation data, after time alignment, missing data filling, and boundary correction, serves as the common input for graph structure construction, node feature generation, and model training, ensuring that the data time reference and physical meaning used in different steps remain consistent.

[0026] Further, in step S2, constructing the graph structure between photovoltaic power stations based on the basic observation data specifically includes: constructing similarity rules based on the basic observation data; the similarity rules include spatial proximity rules, power correlation rules, and meteorological similarity rules; taking each photovoltaic power station as a node in the graph structure, and constructing edge connection relationships between any two nodes based on the similarity rules to obtain the graph structure between photovoltaic power stations.

[0027] in: Spatial proximity rule: If the geographical distance between any two nodes is less than 50 kilometers, the corresponding two photovoltaic power stations are considered to be spatially adjacent.

[0028] Specifically, the spherical distance between the two photovoltaic power stations is obtained using the Haversine geographic distance calculation method, based on their longitude and latitude. .in, and They represent the first The photovoltaic power station node and the first There are 10 photovoltaic power station nodes, hereinafter referred to as nodes. A spatial proximity threshold of 50 kilometers is used because photovoltaic power stations within this distance are typically affected by the combined influence of cloud movement and local weather systems of similar scale, while avoiding the introduction of numerous distant power stations with weaker meteorological correlations into the graph structure due to an excessively large connection range. At that time, the spatial proximity indicator value ;otherwise, .

[0029] Power correlation rule: If the historical power correlation coefficient of two nodes over the past week, i.e. the Pearson correlation coefficient of the historical power series over the past 7 days, is greater than 0.85, then the corresponding two photovoltaic power plants are considered to be related in power.

[0030] For a common effective power sampling point, the first power sampling point is calculated according to the following formula. The node and the first Pearson correlation coefficient between nodes: in, This indicates the sequence number of the common effective power sampling point between the two photovoltaic power plants over the past 7 days. This represents the total number of common effective power sampling points. and They represent the first The node and the first The node at the th The power value of a common valid sampling point and These represent the average values ​​of the corresponding power sequences. When At that time, power correlation indicator value ;otherwise, A variance of 0.85 can filter out sites with strong consistency in power generation trends, avoiding connecting sites with only weak or incidental correlations to the same information propagation path. If the variance of any power series is 0, an effective correlation coefficient cannot be calculated; in this case, [the following will be considered]. Set to 0.

[0031] Meteorological similarity rule: The average values ​​of multiple meteorological parameters of two nodes within the nearest time window are compared using multidimensional features. If the differences are within a set range, the meteorological features are considered similar.

[0032] To avoid the dominance of parameters with large numerical ranges, such as irradiance, in similarity calculations due to the different dimensions of various meteorological parameters, standardization is first performed based on the average and standard deviation of each meteorological parameter: in, Indicates the first Each node at time... The Classified meteorological parameters, Indicates the first Each node at time... The Meteorological parameters, The five meteorological parameters are, in order: solar irradiance, temperature, humidity, wind speed, and cloud cover. and They represent the first Mean and standard deviation of meteorological parameters; This is used to avoid division by zero when the standard deviation is 0. Standardized parameters and Saved after training is complete and remains unchanged during the prediction phase.

[0033] In this embodiment, the 6 hours preceding the current time are selected as the meteorological nearest neighbor time window. The average vectors of the five standardized meteorological parameters for the two nodes within this window are calculated, and the meteorological difference value is calculated according to the following formula: in, and They represent the first The node and the first Each node at time... The five-dimensional standardized meteorological vector, This indicates the number of hours that have been traveled back from the current time. This represents the Euclidean norm. When... At that time, the meteorological similarity indicator value ;otherwise, A 6-hour window can reduce the noise of meteorological measurements at a single moment while preserving recent characteristics of cloud movement and wind field changes; a standardized difference threshold of 0.8 is used to exclude stations with significant differences in multidimensional meteorological conditions.

[0034] When any two similarity rules are satisfied, an edge connection is established between the corresponding two nodes, generating an adjacency matrix. .

[0035] Specifically, the adjacency matrix Generate according to the following formula: in, This represents the total number of photovoltaic power plant nodes. Adopting the requirement of "at least two rules" instead of just one rule ensures that each inter-station edge has at least two independent criteria, reducing the probability of incorrect connections arising solely from proximity but inconsistent power variations, or solely from accidental correlations in historical power. Let... This means that each node retains its own self-connections, so that a node can still make predictions using its own node features even when it has not established edge connections with other nodes.

[0036] Further, in step S2, generating node features specifically includes: standardizing the historical meteorological data, dividing it into time segments according to a fixed sliding window, and splicing each time segment into meteorological features; encoding the timestamp information to obtain time-coded features; and splicing the meteorological features and the time-coded features to obtain node features.

[0037] Specifically, by sliding with a time step of 1 hour and a window size of 24 hours, the meteorological characteristics of each node over the past 24 hours are obtained: Among them, among them, Indicates the first Each node at time... Corresponding meteorological characteristics Indicates the first The five-dimensional standardized meteorological feature vector of each photovoltaic power station node at the current moment. The five-dimensional meteorological features include irradiance, temperature, humidity, wind speed and cloud cover.

[0038] The timestamp information includes hour, day of the week, and holiday markers. For each timestamp, a time encoding feature is constructed.

[0039] Hour encoding, represented by sine and cosine: in, For a moment Hourly sine code, For a moment Hourly cosine encoding, For a moment The corresponding number of hours ranges from 0 to 23.

[0040] The weekday code also uses sine and cosine representation: in, Indicates time The corresponding weekday number ranges from 0 to 6. Holiday codes. Take 0 or 1, take 1 for statutory holidays, and take 0 for non-holidays.

[0041] Concatenate the hour code, weekday code, and holiday code to form a time code feature: in, For the first Each node at time... Node characteristics, This represents expanding a matrix into a vector in chronological order, using the symbol... This indicates a feature concatenation operation. Because... It includes 5-dimensional meteorological features at 24 time points, so its unfolded form is a 120-dimensional vector. After concatenation with the 5-dimensional time code, the node features are obtained. It has 125 dimensions. Defining the node feature dimensions allows the input layer of the meteorological coding network to be directly configured accordingly.

[0042] Node features are aligned and updated synchronously in chronological order during graph structure construction; if meteorological data is missing, it is filled using the aforementioned local weighted linear interpolation or the median of the same hour in history.

[0043] In this embodiment, a graph structure is constructed based on the basic observation data of photovoltaic power plants, and time features are introduced to generate node features, realizing joint modeling of photovoltaic power generation behavior from three dimensions: spatial heterogeneity, meteorological dynamics, and temporal periodicity. By establishing edge connections between nodes through a triple standard of spatial proximity rules, power correlation rules, and meteorological similarity rules, the physical proximity, power generation trend correlation, and environmental synchronicity between photovoltaic power plants can be comprehensively characterized, effectively avoiding the problem of poor structural robustness under traditional single connection standards. The introduction of sine and cosine time coding and sliding window meteorological sequences into the node features helps the model capture the periodic variation characteristics such as diurnal rhythms, weekdays and holidays, and compensates for the lack of explicit time perception in the original meteorological data.

[0044] Furthermore, the meteorological time sensing module includes a meteorological coding network and a gating mechanism network.

[0045] Step S3 specifically includes: constructing a meteorological coding network based on a multilayer perceptron, and using the node features as input to output a meteorological state embedding vector; inputting the meteorological state embedding vector into the gating mechanism network, outputting the Dropout probability corresponding to each node respectively, and forming a dynamic control vector from the Dropout probabilities corresponding to each node.

[0046] Specifically, for each node The node features, including meteorological and temporal coding features, are input into the meteorological coding network, which outputs a meteorological state embedding vector. A gating mechanism network is then constructed, which takes the meteorological state embedding vector as input to generate a dynamic control vector.

[0047] The structure of the meteorological coding network is as follows: Input layer: Fully connected, inputting 125-dimensional node features, outputting 64-dimensional features; Hidden layer 1: Fully connected layer, ReLU activation, output 64 dimensions; Hidden layer 2: Fully connected layer, ReLU activation, output 32 dimensions; Output layer: Outputs a 32-dimensional weather state embedding vector. ; Indicates the first Each node at time... Recent weather trends and time context.

[0048] The structure of the gating mechanism network is as follows: Input layer: 32-dimensional meteorological state embedding vector; Fully connected layer 1: Output 16-dimensional, ReLU activation; Fully connected layer 2: Output 1D, Sigmoid activation, to obtain the 1st... Dropout probability of each node .

[0049] The Dropout probabilities of all nodes form the dynamic control vector: in, This indicates the total number of nodes in a photovoltaic power station.

[0050] To enable the determination of "severe weather conditions" directly from collectable data, this embodiment constructs severe weather indication values ​​based on the changes in cloud cover and wind speed between the current time and the previous hour: in, and They represent the first Cloud cover at the current moment and in the previous hour for each node. and These represent the wind speed at the corresponding time point. A cloud cover change threshold of 0.4 indicates that the cloud cover ratio changes by more than 40% within one hour, and a wind speed change threshold of 3 m / s is used to characterize significant wind field changes over a short period. When either condition is met, [the following will occur]. Set it to 1, otherwise set it to 0.

[0051] The node dropout probability is obtained according to the following formula: in, and These represent the weight vector and bias of the output layer of the gating mechanism network, respectively. This represents the Sigmoid function. This represents the intensity factor of severe weather, which is taken as 0.3 in this embodiment; This indicates a numerical truncation operation. , .

[0052] set up To preserve a minimum level of random structural perturbation and prevent all prediction propagations from always using the same graph structure; [Settings] This is used to prevent individual nodes from being permanently discarded during severe weather, allowing them to still have a certain probability of participating in message aggregation. The gating mechanism outputs the "discard probability," therefore the node's retention probability is... .

[0053] This application constructs a meteorological coding network, integrating local meteorological change trends and time-cycle characteristics of nodes to generate meteorological state embedding vectors, which are then mapped to node dropout control probabilities through a gating mechanism. These probabilities are not only optimized through prediction errors during training but also dynamically increased under conditions such as sudden strong winds or rapid increases in cloud cover, thereby proactively reducing the frequency of unstable node information participating in graph message propagation.

[0054] To avoid the inability to train the gating mechanism network due to the non-direct differentiability of hard Bernoulli sampling, this embodiment uses binary Concrete continuous relaxation to generate differentiable node retention values ​​during the model training phase.

[0055] For the 1 node, generate uniform random numbers And obtain the continuous retention values ​​according to the following formula: in, Continuous retained values ​​between 0 and 1 For the first Uniformly random numbers corresponding to each node This is the temperature parameter. It will be used at the start of training. Set to 1.0, and linearly decrease to 0.1 with each training epoch, so that the continuously retained values ​​gradually approach 0 or 1. During the model's forward computation, The corresponding node is set to a reserved state, and The corresponding node is set to a discarded state; during backpropagation, it is used. Gradients are computed for the gating mechanism network. This pass-through estimation process ensures that the discrete node states remain consistent with the differentiable training process.

[0056] This embodiment uses a two-layer graph convolutional network for power prediction. The features of all nodes are used to construct an initial node representation matrix. , No. Layered graph convolution is calculated as follows: in, Indicates the sequence number of the graph convolutional layer. This represents the adjacency matrix after adding self-connections. express An identity matrix of order 1. express The degree matrix, and They represent the first Weights and biases in graph convolution. The first graph convolutional layer outputs a 64-dimensional node representation, and the second graph convolutional layer outputs a 32-dimensional node representation. When a node is set to the discard state, the node representation of its corresponding row is set to zero, and the weight of its associated edge in this graph convolution is also set to zero.

[0057] The node representations of the convolutional layer in the second figure are converted into power prediction values ​​for each node at the next prediction time by the fully connected output layer. For the implementation of predicting the total power of the region, the power prediction values ​​of each node are summed to obtain the region's predicted power. For the implementation of predicting the power of a single target power plant, the output of the corresponding target node is read as... This embodiment takes predicting the total regional power for the next hour as an example, training labels... This is the sum of the actual power of each node in the next hour.

[0058] The model training uses a loss function composed of mean squared error and Dropout probability constraints: in, This represents the number of samples in a training batch; in this example, it is set to 64. and They represent the first The predicted power and actual power of each training sample; Indicates the first In the training samples, the th The probability of Dropout for each node; For probability constraint coefficients, in this embodiment, we take... This is used to suppress the degradation where the Dropout probability of all nodes approaches the upper limit simultaneously.

[0059] The Adam optimizer is used to update the parameters of the meteorological coding network, the gating mechanism network, the graph neural network, and the output layer. The initial learning rate is set to... The maximum number of training epochs is set to 200. Training data is divided into training, validation, and test sets in chronological order, with proportions of 70%, 15%, and 15%, respectively, without disrupting the chronological order between sets. After each training epoch, the validation set loss is calculated. Training stops when the validation set loss does not decrease for 20 consecutive epochs, and all model parameters corresponding to the epoch with the lowest validation set loss are saved. During the prediction phase, the saved parameters are used consistently, and the model is not updated based on the baseline prediction results.

[0060] Further, step S4 specifically includes: in the prediction stage, the graph structure and the node features are input into the graph neural network, and the graph neural network performs multiple baseline forward predictions; in each baseline forward prediction, Bernoulli sampling is performed based on the dynamic control vector to generate different subgraph structures and perform power prediction to obtain multiple baseline prediction results.

[0061] The graph structure is input into the graph neural network in the form of an adjacency matrix, and the node features and edge weights are encoded in tensor form and updated synchronously.

[0062] In each forward propagation of the graph neural network, for each node... Bernoulli distribution sampling is performed according to its Dropout probability: in, Indicates the first In the next baseline forward prediction The state of each node is preserved. ;like , indicating the first Each node was preserved in this propagation; if , indicating the first Each node is discarded during this propagation. After a node is discarded, its output is set to zero, and the edges associated with that node do not participate in this message aggregation.

[0063] In this embodiment, the multiple predictions based on the original dynamic control vector are referred to as baseline sampling, and the number of baseline forward predictions is... Set to 100 samples. 100 samples can generate a sample set for judging the central region and preliminary tail region of the baseline prediction result, while keeping computational resources under control; in devices with limited computing resources, It can be set to no less than 50 times.

[0064] No. The node-preserved states corresponding to the next baseline forward prediction form the baseline subgraph mask: Based on the baseline subgraph mask, the nodes in the discarded state are represented as zero, and their associated edges are masked, generating the first... The first baseline subgraph structure is input into the trained graph neural network to obtain the second baseline subgraph structure. Benchmark prediction results If all nodes are in a discarded state during a sampling, the sampling result is abandoned and Bernoulli sampling is re-executed until at least one node is retained.

[0065] Since each node independently performs Bernoulli sampling under a given dynamic control vector, the... The generation probability of each reference sub-image mask under the original sampling distribution is: in, Indicates the dynamic control vector The original sampling distribution is defined. The server associates and records the mask for each baseline submap. The corresponding baseline prediction results and generation probability The generated probability is used to perform probability correction on the tail subgraph samples that are added in a targeted manner, so as to avoid the misinterpretation of "being actively sampled multiple times" as "having a higher probability of occurring in the original model".

[0066] When the number of nodes is large, the product of multiple generation probabilities may approach the lower limit of floating-point numbers in computers. Therefore, in actual calculations, the natural logarithm of the generation probability is saved. in, Represents the mask of any subgraph. Indicates the first The state of each node is preserved. Exponential operations are performed only when the generation probability is needed. When summing multiple sampling distributions, logarithmic summation with exponential calculation is used to reduce numerical underflow.

[0067] Furthermore, step S5 specifically includes: S51: Determine the tail candidate set based on multiple benchmark prediction results and corresponding benchmark subgraph masks; the tail candidate set includes a low-power tail candidate set and a high-power tail candidate set.

[0068] Ordinary multiple stochastic baseline forward predictions tend to concentrate most samples in the central region of the probability distribution, potentially resulting in too few samples representing extremely low or high power levels that truly impact energy storage and standby unit scheduling. Directly increasing the total number of baseline forward predictions would require a significant amount of computation to repeatedly generate samples in the central region. Therefore, this embodiment first uses the baseline sampling results to determine the tail candidate subgraph, and then adjusts the node sampling probability based on the specific impact of node states on the tail baseline prediction results, allowing the limited additional baseline forward predictions to prioritize coverage of extreme power regions.

[0069] First, the 100 baseline prediction results Sort by power value from smallest to largest. Then sort by the lowest power value. The baseline prediction results and corresponding baseline subgraph masks were determined as the low-power tail candidate set. , will the highest The baseline prediction results and corresponding baseline subgraph masks were determined as the high-power tail candidate set. ,in: in, This indicates rounding up to the nearest integer. For In this scenario, the low-power tail candidate set and the high-power tail candidate set each include 10 baseline sampling results. Selecting the lowest and highest 10% allows the candidate sets to be concentrated at both ends of the initial distribution, while ensuring that each set has at least 5 samples, avoiding the judgment of node contribution based solely on a single extreme sampling result. Even if multiple baseline prediction results have the same value, the candidate sets are still determined according to their sorting position, thus ensuring a fixed number of candidate sets.

[0070] S52: Calculate the average counterfactual contribution of the photovoltaic power plant in the tail candidate set.

[0071] For each reference subgraph mask in the low-power tail candidate set and the high-power tail candidate set, a counterfactual state switch is performed on each node. The counterfactual state switch means: keeping the reserved states of other nodes in the same reference subgraph unchanged, setting the node to be analyzed to a reserved state and performing a power prediction once, then setting that node to a discarded state and performing a power prediction once more. This yields the... The candidate subgraph of the th... Preservation status baseline prediction results for each node and discarded state baseline prediction results This operation does not retrain the graph neural network; it only changes the state of a single node in the input graph structure.

[0072] Calculate the first one respectively The average counterfactual contribution of each node in the low-power tail candidate set and the high-power tail candidate set: in, Indicates the first The average counterfactual contribution of each node in the low-power tail candidate set. This represents the average counterfactual contribution of the node in the high-power tail candidate set. and These represent the number of samples in the two candidate sets, respectively. A positive contribution value indicates that retaining the node will increase the prediction power compared to discarding it; a negative contribution value indicates that retaining the node will decrease the prediction power; the larger the absolute value of the contribution value, the stronger the influence of the node's retention status on the corresponding tail baseline prediction result.

[0073] S53: Adjust the Dropout probability in the dynamic control vector based on the average counterfactual contribution value to obtain the tail-guided Dropout probability.

[0074] Based on the above contribution direction and contribution level, low-power tail-guided Dropout probabilities and high-power tail-guided Dropout probabilities are generated respectively: in, and They represent the first The probability of low-power tail-guided dropout and the probability of high-power tail-guided dropout for each node. This represents the maximum magnitude of a single directional adjustment, which is 0.15 in this embodiment; This represents a power scale that converts the power contribution value into a dimensionless value. In this embodiment, the rated installed capacity of the photovoltaic power station group is used. 5% of ; This is used to limit the adjusted Dropout probability to between 0.02 and 0.98; Used to limit the probability adjustment range when the contribution value increases abnormally.

[0075] For low-power tail section, if Then retain the first Each node will increase the predicted power, and to increase the chance of low-power subgraphs appearing, the Dropout probability of that node will be increased accordingly; if If the node is not properly configured, retaining it will reduce the predicted power and consequently decrease its Dropout probability. For high-power tails, the opposite adjustment direction is applied. These rules provide a clear basis for understanding the power impact of changes in node sampling probabilities, rather than uniformly increasing or decreasing the Dropout probability of all nodes.

[0076] S54: Based on the tail-guided Dropout probability, Bernoulli resampling is performed to generate a tail-guided subgraph structure, and power prediction is performed to obtain the tail-guided baseline prediction result.

[0077] Based on respectively and Perform Bernoulli resampling. In this embodiment, the number of low-power tail-guided samplings is [number missing]. Set to 50 times, high-power tail-guide sampling number Set to 50 times. In each low-power tail-guided sampling, the first... Each node The probability is preserved; in each high-power tail-guided sampling, the first... Each node The probability is preserved. Correspondingly, low-power tail steering subgraphs and high-power tail steering subgraphs are generated, and the same graph neural network is used to obtain the low-power tail steering baseline prediction results and high-power tail steering baseline prediction results, respectively. If all nodes are discarded in a single sampling, resampling is performed.

[0078] S55: Construct a hybrid sampling distribution based on the generation probability of the baseline prediction result and the generation probability of the tail-guided baseline prediction result.

[0079] Any subgraph mask The generation probabilities under the low-power tail-guided sampling distribution and the high-power tail-guided sampling distribution are as follows: in, This indicates the generation probability of the low-power tail-steering baseline prediction result. This represents the generation probability of the high-power tail-guided reference prediction result.

[0080] The baseline sampling results, low-power tail steering sampling results, and high-power tail steering sampling results are combined, resulting in a total number of predicted samples: In this embodiment The mixing coefficient is determined based on the proportion of the three types of sampling results in the total sample. A hybrid sampling distribution is constructed from the original sampling distribution, the low-power tail-guided sampling distribution, and the high-power tail-guided sampling distribution: S56: Determine the importance weight of the benchmark prediction result based on the ratio of the generation probability of the benchmark prediction result to the mixed sampling distribution.

[0081] The mixed sampling distribution represents the actual probability of obtaining any sub-map mask from the merging of the three sampling processes. For the merged... Sub-image mask and corresponding benchmark prediction results The importance weights are determined according to the following formula: in, Indicates the first The importance weights of each baseline prediction result are calculated. The numerator represents the probability of the target occurring in the subgraph under the original dynamic control vector, and the denominator represents the actual probability of the subgraph occurring during the three-class sampling merging process. Because tail subgraphs are subject to targeted additional sampling, their mixed sampling probability is usually higher than the original sampling probability; therefore, their corresponding importance weights are reduced to avoid overestimating the true probability of tail events.

[0082] To determine whether the importance weights are overly concentrated, the effective sample size is calculated: in, This represents the effective sample size after importance weighting. When This indicates that the weight of a minority sample is too large, and the tail-guided sampling distribution deviates too much from the original sampling distribution. In this case, the probability adjustment range should be adjusted. Halve the power and re-perform low-power and high-power tail-guided sampling until... or The value is reduced to 0.03. This treatment prevents the probability density estimation from becoming unstable due to excessively large variances in the importance weights.

[0083] S57: Based on the baseline prediction results and their corresponding importance weights, a weighted kernel density estimation is performed to construct the probability density distribution of photovoltaic power.

[0084] The weighted forecast mean and weighted standard deviation are calculated based on all baseline forecast results and their corresponding importance weights: in, This represents the predicted mean after importance weighting. This represents the weighted standard deviation. The weighted forecast mean is used for high-power dispatching decisions in the power grid, while the weighted standard deviation is used to determine the kernel density estimation bandwidth.

[0085] A kernel density estimation method is used to smooth the distribution of multiple benchmark prediction results, constructing a probability density distribution of photovoltaic power. This embodiment employs a Gaussian kernel function: in, This represents the difference in predicted power after standardization. The kernel density bandwidth is determined based on the weighted effective sample size. in, This represents the kernel density estimation bandwidth. To avoid the bandwidth approaching zero when the baseline prediction results are too concentrated, or the bandwidth being too wide when the sample dispersion is too large, the bandwidth is... Limited to to between.

[0086] The weighted probability density function is expressed as: in, This represents the photovoltaic power value to be calculated. Indicates power value The probability density at that point. Since the physical range of photovoltaic power is 0 to... Set the probability density outside this range to 0, and set the probability density between 0 and 0. The probability density within the range is renormalized so that its integral is 1.

[0087] S58: Integrate the probability density distribution to obtain the cumulative distribution, and determine the prediction confidence interval based on the cumulative distribution.

[0088] Integrating the probability density distribution yields the cumulative distribution: in, Indicates that the photovoltaic power is not greater than Cumulative probability, integral variable Used only for probability density integrals.

[0089] In this embodiment, the preset signal level is... With a confidence level of 95%, the prediction confidence interval is expressed as follows: in, This represents the power value corresponding to a cumulative probability of 2.5%. This represents the power value corresponding to a cumulative probability of 97.5%.

[0090] The prediction confidence interval is determined by using cumulative distribution quantiles. It does not require multiple baseline prediction results to follow a normal distribution, and is suitable for skewed or multimodal distributions that may occur under conditions such as sudden changes in cloud cover.

[0091] After step S5, the method further includes: extracting the tail cumulative probability value of the probability density distribution in the preset extreme value region, and determining whether there is a power deviation risk based on the tail cumulative probability value.

[0092] Specifically, the tail accumulation probability value includes a low-power tail accumulation probability value and a high-power tail accumulation probability value. The low-power tail accumulation probability value and the high-power tail accumulation probability value are obtained by integrating the probability density distribution within the preset low-power extreme value region and the preset high-power extreme value region, respectively.

[0093] In this embodiment, the rated installed capacity of the photovoltaic power station group For 100MW, the low power threshold will be... Setting it to 10MW will set the high power threshold. Set to 90MW. The low-power threshold corresponds to 10% of the rated installed capacity, characterizing significantly insufficient photovoltaic output; the high-power threshold corresponds to 90% of the rated installed capacity, characterizing photovoltaic output approaching the grid connection or dispatch limit. Low-power tail cumulative probability value. and high-power tail cumulative probability value They are respectively: in, This indicates the probability that the predicted power output falls within the range of 0 to 10 MW. This represents the probability that the predicted power falls within the 90MW to 100MW range. Using the regional cumulative probability instead of the probability density value at a single point avoids the risk assessment being directly affected by changes in kernel density bandwidth and power measurement units.

[0094] When the cumulative probability value of the low-power tail exceeds the preset low-power risk threshold, it is determined that there is a potential extremely low power deviation risk; when the cumulative probability value of the high-power tail exceeds the preset high-power risk threshold, it is determined that there is a potential extremely high power deviation risk.

[0095] In this embodiment, the low-power risk threshold is set to 0.15, and the high-power risk threshold is set to 0.25. Low-power events directly involve insufficient power supply and backup capacity requirements, therefore a relatively low trigger probability is used; high-power events can usually be absorbed by energy storage or limited in operation, therefore a relatively high trigger probability is used. At that time, it was determined that there was a potential risk of extremely low power deviation; when When there is a potential extremely high power deviation risk, the extreme power risk state is not triggered and the current scheduling strategy is maintained when neither of the two exceeds the corresponding threshold.

[0096] After step S5, the method further includes: connecting the predicted confidence interval, the probability density distribution, and the tail cumulative probability value to the power grid control system to perform strategy optimization and resource scheduling for standby units and energy storage devices.

[0097] Specifically, when the cumulative probability value of the low-power tail is detected to be higher than a preset first threshold and the lower boundary of the prediction confidence interval is lower than a preset power supply safety threshold, the energy storage device is triggered to discharge and the standby unit is placed in standby mode; the low-power region is the power range where the power is less than the preset low-power threshold.

[0098] In this embodiment, the preset first threshold is 0.15, and the preset power supply safety threshold is... It is 15MW. When simultaneously satisfying as well as At this time, the grid control system sends a discharge command to the energy storage device and switches the standby unit from shutdown to standby status. The discharge power of the energy storage device does not exceed its rated discharge power, nor does it exceed the difference between the preset power supply safety threshold and the lower boundary of the predicted confidence interval. After the standby unit enters the standby state, it does not immediately connect to the grid for power generation. Grid connection is only executed when the real-time photovoltaic power is lower than 15MW for two consecutive 5-minute sampling periods. This dual-condition triggering method can avoid frequent activation of the energy storage and standby units due to a single low-power predicted sample.

[0099] When the cumulative probability value of the high-power tail is detected to be higher than the preset second threshold, and the weighted prediction mean determined according to each benchmark prediction result and its importance weight exceeds the preset scheduling power limit, the energy storage absorption, power generation limitation or adjustable load enhancement strategy is executed.

[0100] In this embodiment, the second threshold is preset to 0.25, and the upper limit of the scheduling power is preset. It is 95MW. When simultaneously satisfying as well as When the energy storage device is charged, priority is given to absorbing photovoltaic power; when the remaining rechargeable capacity of the energy storage device is insufficient to absorb photovoltaic power... When the adjustable load increases power consumption, the power consumption is controlled to increase; if the predicted power still exceeds 95MW after the energy storage and adjustable load increases, photovoltaic power curtailment is implemented on the remaining excess power. The sum of the energy storage charging power, the increase in adjustable load, and the curtailed power shall not be less than [amount missing]. And each shall not exceed the rated adjustment capacity of the corresponding equipment.

[0101] When the probability density exhibits a distinct bimodal pattern, simultaneously satisfying both low-power and high-power triggering conditions, priority is given to low-power dispatch related to power supply security. This keeps energy storage devices in a dischargeable state and puts standby units into standby mode. The high-power absorption strategy remains a reserve instruction, to be implemented based on the next forecast cycle. , The real-time power is reassessed. This priority is used to prevent the same energy storage device from receiving both charging and discharging commands simultaneously.

[0102] When the effective proportion of basic observation data within the current prediction window is less than 80%, or the effective sample size after importance correction is less than [a certain percentage], Even after being reduced to 0.03, it is still less than If a low-confidence prediction is made, it is marked as such. For low-confidence results, the grid limiting or standby unit grid connection operation is not automatically executed based on a single prediction. Instead, the prediction confidence interval, tail cumulative probability value, and low-confidence mark are output to the grid control system, which then reconfirms the prediction in conjunction with real-time power measurements.

[0103] It should be noted that by performing multiple baseline forward predictions in the graph neural network and sampling the subgraphs based on dynamic control vectors to output the photovoltaic power prediction distribution, uncertainty modeling and probabilistic risk expression of the model's baseline prediction results are achieved. Based on this, by switching the node retention states one by one in the baseline tail subgraph, the extreme baseline prediction results can be correlated with the information participation states of specific photovoltaic nodes; adjusting the node dropout probability according to this correspondence allows additional computational resources to be concentrated in low-power and high-power regions. Since directional sampling changes the actual sampling frequency of various subgraphs, this embodiment further calculates the importance weights using the ratio of the original sampling probability to the mixed sampling probability, increasing the number of tail samples without causing a systematic shift in probability estimation.

[0104] Independent sampling and inference are performed using the subgraph structure generated in each propagation. Based on importance weights, statistical prediction values ​​are constructed to establish prediction confidence intervals and probability density functions, reflecting the fluctuation characteristics of prediction values ​​under current meteorological inputs and information pathways at different nodes. A continuous probability distribution of predicted power is built through weighted kernel density estimation, and the cumulative probability of low-power and high-power regions is analyzed. This not only identifies potential extremely low or extremely high power generation risks but also provides power range outputs with credibility indicators for grid dispatch. The graph neural network is extended from a traditional single-point numerical prediction model to a probabilistic output model, enabling it to quantify risks. This allows it to support operational strategies such as energy storage activation, standby unit dispatch, and power limiting in large-scale photovoltaic grid-connected scenarios, improving the smart grid's responsiveness to uncertainties in new energy access and enhancing dispatch security.

[0105] Based on the same inventive concept, this application also provides a system for implementing the above-described graph neural network-based photovoltaic power probability prediction method. The solution provided by this system is similar to the implementation described in the above method. Therefore, the specific limitations of one or more graph neural network-based photovoltaic power probability prediction system embodiments provided below can be found in the above-described limitations of the graph neural network-based photovoltaic power probability prediction method, and will not be repeated here.

[0106] In one exemplary embodiment, a photovoltaic power probability prediction system based on graph neural networks is provided, including the following modules.

[0107] The data acquisition module is used to acquire basic observation data of the photovoltaic power station; the basic observation data includes historical meteorological data, geographic coordinate information and historical power data.

[0108] The data processing module is used to construct a graph structure between photovoltaic power stations based on the basic observation data and generate node features.

[0109] The dynamic control vector generation module is used to input the node features into the meteorological time sensing module to generate dynamic control vectors.

[0110] The prediction module is used to input the graph structure and the node features into the graph neural network during the prediction phase, and to obtain multiple benchmark prediction results based on the dynamic control vector row for multiple benchmark forward predictions.

[0111] The results output module is used to perform statistical analysis on multiple benchmark prediction results, construct the probability density distribution of photovoltaic power, and output the prediction confidence interval.

[0112] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0113] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0114] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0115] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0116] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A photovoltaic power probabilistic prediction method based on graph neural networks, characterized in that, include: Obtain basic observation data for photovoltaic power plants; The basic observation data includes historical meteorological data, geographic coordinate information, and historical power data; Based on the aforementioned basic observation data, a graph structure is constructed between photovoltaic power stations, and node features are generated; The node features are input into the meteorological time sensing module to generate a dynamic control vector; In the prediction phase, the graph structure and the node features are input into the graph neural network, and multiple baseline forward predictions are made based on the dynamic control vector rows to obtain multiple baseline prediction results. Statistical analysis is performed on multiple benchmark prediction results to construct the probability density distribution of photovoltaic power and output the prediction confidence interval.

2. The photovoltaic power probabilistic prediction method based on graph neural networks according to claim 1, characterized in that, Based on the aforementioned fundamental observation data, a graph structure is constructed between photovoltaic power plants, specifically including: Similarity rules are constructed based on the aforementioned basic observation data; these similarity rules include spatial proximity rules, power correlation rules, and meteorological similarity rules. Each photovoltaic power station is treated as a node in the graph structure, and the edge connection relationship between any two nodes is constructed based on the similarity rule to obtain the graph structure between photovoltaic power stations.

3. The photovoltaic power probabilistic prediction method based on graph neural networks according to claim 1, characterized in that, Generate node features, specifically including: The historical meteorological data is standardized, divided into time segments by a fixed sliding window, and the time segments are spliced ​​together to form meteorological features. Encode the timestamp information to obtain time-coded features; The meteorological features and the time-coded features are concatenated to obtain node features.

4. The photovoltaic power probabilistic prediction method based on graph neural networks according to claim 1, characterized in that, The meteorological time sensing module includes a meteorological coding network and a gating mechanism network; Specifically, the node features are input into the meteorological time sensing module to generate a dynamic control vector, including: A meteorological coding network is constructed based on a multilayer perceptron, and the node features are used as input to output a meteorological state embedding vector. The weather state embedding vector is input into the gating mechanism network, and the Dropout probability corresponding to each node is output. The Dropout probability corresponding to each node is used to form a dynamic control vector.

5. The photovoltaic power probabilistic prediction method based on graph neural networks according to claim 1, characterized in that, In the prediction phase, the graph structure and the node features are input into the graph neural network, and multiple baseline forward predictions are performed based on the dynamic control vector to obtain multiple baseline prediction results, specifically including: During the prediction phase, the graph structure and the node features are input into the graph neural network, which performs multiple baseline forward predictions. In each baseline forward prediction, Bernoulli sampling is performed based on the dynamic control vector to generate different subgraph structures and perform power prediction, resulting in multiple baseline prediction results.

6. The photovoltaic power probabilistic prediction method based on graph neural networks according to claim 1, characterized in that, Statistical analysis is performed on multiple benchmark prediction results to construct the probability density distribution of photovoltaic power and output the prediction confidence interval, specifically including: The tail candidate set is determined based on multiple benchmark prediction results and corresponding benchmark subgraph masks; the tail candidate set includes a low-power tail candidate set and a high-power tail candidate set; Calculate the average counterfactual contribution of the photovoltaic power plant in the tail candidate set; The Dropout probability in the dynamic control vector is adjusted based on the average counterfactual contribution value to obtain the tail-guided Dropout probability. Based on the tail-guided Dropout probability, Bernoulli resampling is performed to generate a tail-guided submap, and power prediction is performed to obtain the tail-guided baseline prediction result. A hybrid sampling distribution is constructed based on the generation probability of the baseline prediction result and the generation probability of the tail-guided baseline prediction result; The importance weight of the baseline prediction result is determined based on the ratio of the generation probability of the baseline prediction result to the mixed sampling distribution. Based on the baseline prediction results and their corresponding importance weights, a weighted kernel density estimation is performed to construct the probability density distribution of photovoltaic power. The cumulative distribution is obtained by integrating the probability density distribution, and the prediction confidence interval is determined based on the cumulative distribution.

7. The photovoltaic power probabilistic prediction method based on graph neural networks according to claim 1, characterized in that, Following the output of the prediction confidence interval, the following is also included: Extract the tail cumulative probability value of the probability density distribution in the preset extreme value region. When the tail cumulative probability value exceeds the preset risk threshold, it is determined that there is a potential extreme power deviation risk. The tail cumulative probability value includes low power tail cumulative probability value and high power tail cumulative probability value.

8. The photovoltaic power probabilistic prediction method based on graph neural networks according to claim 7, characterized in that, Following the output of the prediction confidence interval, the following is also included: The predicted confidence interval, the probability density distribution, and the tail cumulative probability value are connected to the power grid control system to optimize the strategy and schedule resources for standby units and energy storage devices.

9. The photovoltaic power probabilistic prediction method based on graph neural networks according to claim 8, characterized in that, By connecting the predicted confidence interval, the probability density distribution, and the tail cumulative probability value to the power grid control system, strategy optimization and resource scheduling are performed on standby units and energy storage devices, specifically including: When the cumulative probability value of the low-power tail is detected to be higher than the preset first threshold and the lower boundary of the prediction confidence interval is lower than the preset power supply safety threshold, the energy storage device is triggered to discharge and the standby unit is placed in standby mode. When the cumulative probability value of the high-power tail is detected to be higher than the preset second threshold, and the weighted average of multiple benchmark prediction results exceeds the preset scheduling power limit, the energy storage absorption, power generation limitation or adjustable load enhancement strategy is executed.

10. A photovoltaic power probabilistic prediction system based on graph neural networks, characterized in that, The system is used to execute the photovoltaic power probability prediction method based on graph neural networks as described in any one of claims 1-9, and the system comprises: The data acquisition module is used to acquire basic observation data of the photovoltaic power station; the basic observation data includes historical meteorological data, geographic coordinate information and historical power data; The data processing module is used to construct a graph structure between photovoltaic power stations based on the basic observation data and generate node features; The dynamic control vector generation module is used to input the node features into the meteorological time sensing module to generate a dynamic control vector; The prediction module is used to input the graph structure and the node features into the graph neural network during the prediction phase, and make multiple baseline forward predictions based on the dynamic control vector row to obtain multiple baseline prediction results. The results output module is used to perform statistical analysis on multiple benchmark prediction results, construct the probability density distribution of photovoltaic power, and output the prediction confidence interval.

Images