Digital twin modeling method and system for photovoltaic power station

Abstract

The invention relates to the technical field of photovoltaic power station digital twinning modeling, and discloses a photovoltaic power station digital twinning modeling method and system, and the method comprises the steps: obtaining the monitoring data of each node of a distributed photovoltaic power station, and generating a temperature-power feature vector; analyzing the temperature gradient influence by using a graph convolutional network, and generating a node specificity temperature compensation matrix; calculating a deviation coefficient based on a standard temperature reference, and generating a temperature normalization mapping function; correcting model parameters by using an environment normalization method, and generating a model update vector after environment correction; and performing federal aggregation on the model update vector after environment correction, and outputting a global digital twinborn model. According to the method, high-precision distributed photovoltaic power station digital twinborn modeling under the environment heterogeneous condition is realized, and the accuracy and robustness of cross-climate region data integration are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic power station digital twin modeling, more specifically, it relates to a photovoltaic power station digital twin modeling method and system. BACKGROUND

[0002] In the digital twin modeling scenario of distributed photovoltaic power stations, it is necessary to integrate local models of multiple geographically dispersed nodes through federated learning technology to form a unified global digital twin model. However, due to the distribution of photovoltaic power station nodes in different climate regions, there are significant differences in environmental temperature. The output power of photovoltaic modules is negatively correlated with temperature, and for every 1℃ increase in temperature, the output power of crystalline silicon photovoltaic modules decreases by about 0.4%-0.5%. The performance difference caused by temperature is misinterpreted as model bias in the traditional federated learning framework, resulting in systematic errors when aggregating model parameters.

[0003] In addition, some nodes in the distributed photovoltaic power station network may provide low-quality data and model updates due to sensor failures, equipment aging, or improper maintenance. Traditional federated learning aggregates updates from all nodes with equal weights, so that the incorrect updates from these abnormal nodes directly affect the accuracy of the global model, and cannot guarantee the credibility of distributed modeling. SUMMARY

[0004] The present application provides a photovoltaic power station digital twin modeling method and system, which realizes high-precision distributed photovoltaic power station digital twin modeling under environmental heterogeneity, and improves the accuracy and robustness of data integration across climate regions.

[0005] The present application provides a photovoltaic power station digital twin modeling method, comprising: obtaining monitoring data of each node of a distributed photovoltaic power station, and generating a temperature-power feature vector; using a graph convolution network to analyze the influence of temperature gradient and generate a node-specific temperature compensation matrix; calculating a bias coefficient based on a standard temperature reference and generating a temperature normalization mapping function; using an environmental normalization method to correct model parameters and generating an environment-corrected model update vector; federated aggregation of the environment-corrected model update vector is performed to output a global digital twin model; wherein the environmental normalization method transforms the original parameters through the temperature normalization mapping function and fine-tunes them in combination with the node-specific temperature compensation coefficient.

[0006] Further, generating a temperature-power feature vector comprises:

[0007] calculating a temperature coefficient correction factor for the photovoltaic module, wherein the temperature coefficient correction factor is equal to 1 plus the product of the power temperature coefficient and the difference between the module temperature and the standard test condition temperature;

[0008] Based on the temperature coefficient correction factor and the actual power output, a normalized temperature-power characteristic is calculated, which is equal to the actual power divided by the product of the standard condition power, the temperature coefficient correction factor and the irradiance ratio;

[0009] The temperature-power characteristics in each time window are combined into a feature vector.

[0010] Further, the influence of the temperature gradient is analyzed using a graph convolution network, which includes:

[0011] The electrical topology of the photovoltaic power station is represented as a graph, where the nodes represent photovoltaic strings or inverters, and the edges represent electrical connection relationships;

[0012] The node feature data is standardized for preprocessing to eliminate differences in the dimensions of different physical quantities;

[0013] The preprocessed temperature distribution data and electrical topology are input into the graph convolution network, which captures the propagation and influence patterns of temperature in the electrical network through multiple layers of graph convolution operations;

[0014] The graph convolution network outputs the temperature influence weights of each node, forming a node-specific temperature compensation matrix.

[0015] Further, generating a temperature normalization mapping function includes:

[0016] A standard temperature reference is set, and the deviation of the ambient temperature of each node relative to the standard temperature is calculated;

[0017] Based on the temperature deviation and the temperature characteristic curve of the photovoltaic module, a temperature normalization mapping function is generated;

[0018] The temperature normalization mapping function standardizes the influence of temperature deviation on model parameters through a nonlinear transformation, which includes applying a hyperbolic tangent function to the temperature deviation to limit the temperature influence range.

[0019] Further, the environmental normalization method includes:

[0020] The temperature deviation vector of each node is calculated;

[0021] Based on the temperature deviation vector and the temperature coefficient of the photovoltaic module, a parameter correction matrix is constructed;

[0022] The local model update vector is linearly transformed by the parameter correction matrix to obtain the temperature-normalized parameters;

[0023] The corresponding row of the node-specific temperature compensation matrix is extracted as a node compensation vector, which is multiplied element by element with the normalized parameters to generate the final environmental correction vector.

[0024] Further, before federated aggregation, it also includes:

[0025] Obtain the local verification indicators and historical update records of each node, identify abnormal updates using a statistical anomaly detection method, and generate an abnormal label matrix;

[0026] Based on the historical performance of the block chain record node, calculate the time-decaying reputation points, and generate a reputation weight vector;

[0027] The model update vector after environmental correction and the reputation weight vector are input into a weighted federated aggregation module.

[0028] Further, the statistical anomaly detection method includes:

[0029] Calculate the statistical characteristics of each node model update, including mean, variance, skewness, kurtosis and time series autocorrelation coefficient;

[0030] Use the Isolation Forest method to identify abnormal updates deviating from the group distribution;

[0031] For the detected abnormal nodes, mark 1 in the corresponding position of the abnormal label matrix, and mark 0 for normal nodes.

[0032] Further, the weighted federated aggregation includes:

[0033] Calculate the effective node set and filter out the nodes that are not marked as abnormal;

[0034] The reputation weight of the effective node is normalized;

[0035] Based on the adjusted weight and the environmental correction vector, calculate the weighted update amount;

[0036] Add the weighted update amount to the current global model parameters to generate new global model parameters.

[0037] Further, it also includes:

[0038] Evaluate the performance of the global model using validation data sets in different climate zones;

[0039] Update the temperature compensation coefficient and reputation score parameter according to the prediction error feedback to realize adaptive optimization.

[0040] The present application provides a kind of photovoltaic power station digital twin modeling system, including:

[0041] Data acquisition module, for obtaining the monitoring data of each node of distributed photovoltaic power station and generating temperature-power characteristic vector;

[0042] Temperature compensation analysis module, for analyzing temperature gradient influence using graph convolution network and generating node-specific temperature compensation matrix;

[0043] an environment normalization module configured to generate a temperature normalization mapping function and correct model parameters based on a standard temperature reference;

[0044] a quality screening module configured to identify abnormal nodes and calculate reputation weights;

[0045] a federated aggregation module configured to perform weighted federated aggregation to generate a global digital twin model.

[0046] The present application has the following beneficial effects:

[0047] The present application uses a graph convolution network to analyze temperature gradient influence and generate a node-specific temperature compensation matrix through an environment normalization federated protocol, and realizes standardization of model parameters under different temperature conditions by combining a temperature normalization mapping function, thereby overcoming systematic bias factors caused by temperature differences and solving the problem of errors introduced by traditional federated learning ignoring environmental factors.

[0048] Further, by introducing a quality screening method based on statistical anomaly detection and a reputation scoring system based on a blockchain, effective suppression of low-quality node contributions is realized, and model performance degradation factors caused by equal aggregation are overcome, solving the problem of malicious or faulty nodes affecting the credibility of the global model.

[0049] Therefore, the present application realizes high-precision distributed photovoltaic power station digital twin modeling under environmental heterogeneous conditions, and improves the accuracy and robustness of data integration across climate regions. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 is a photovoltaic power station digital twin modeling method flowchart of the present application;

[0051] Figure 2 is a node temperature compensation matrix thermodynamic diagram of the present application;

[0052] Figure 3 is a performance comparison diagram of different aggregation methods of the present application;

[0053] Figure 4 is a model update vector comparison diagram before and after environmental correction of the present application; DETAILED DESCRIPTION

[0054] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that discussions of these implementations are merely provided to enable those skilled in the art to better understand and thus implement the subject matter described herein, and are not intended to limit the scope of the disclosure. Various modifications can be made to the function and arrangement of elements discussed without departing from the scope of the subject matter described herein. Each example can omit, substitute, or add various processes or components also in accordance. In addition, features described in relation to one example can also be combined in other examples.

[0055] A photovoltaic power station digital twin modeling method is disclosed in at least one embodiment of the present application, as shown, comprising the following steps: Figures 1-4

[0056] Step 100: Obtain the monitoring data of each node of the distributed photovoltaic power station, and generate a temperature-power feature vector.

[0057] Specifically, the temperature distribution data of the photovoltaic array, the electrical connection topology information and the update vector of the local digital twin model are collected from each photovoltaic power station node. Among them, the temperature distribution data includes component surface temperature, ambient temperature and irradiance data; the electrical connection topology describes the connection relationship of photovoltaic strings and inverters; the local model update vector is the model weight increment obtained by each node based on local data training. Based on the collected data, the temperature-power correlation coefficient at each time is calculated, and the temperature-power feature vector is extracted.

[0058] It should be noted that the extraction of the temperature-power feature vector includes the following sub-steps:

[0059] Step 101: Calculate the temperature coefficient correction factor of the photovoltaic module , wherein is the power temperature coefficient, is the component temperature, is the standard test condition temperature (25℃).

[0060] Further, the value range of the power temperature coefficient is [-0.005, -0.004] / ℃, and the value range of the power temperature coefficient is based on the industry standard of crystalline silicon photovoltaic modules. The effective range of the component temperature is [-40℃, 85℃], which corresponds to the working temperature range of the photovoltaic module. The constraint range of the temperature coefficient correction factor is [0.6, 1.4], and when the constraint range of the temperature coefficient correction factor is exceeded, it indicates that the component is working abnormally and needs to be handled abnormally.

[0061] Step 102: Calculate the normalized temperature-power feature based on the temperature coefficient correction factor and the actual power output , wherein is the actual power, is the standard condition power, is the actual irradiance, is the standard irradiance (1000W / m²).

[0062] Further, the time dimension of the normalized temperature-power feature formula is embodied as:

[0063] ​

[0064] where is the normalized temperature-power feature (normalized temperature-power feature value), normal range [0, 1.2], time index denotes the sampling time, sampling interval The constraint range of is [1 minute, 15 minutes]. When , the data noise is large and affects the stability of the feature, and when , the fast response of the light change cannot be captured. All parameters , , are functions of time .

[0065] Further, the effective range of the actual irradiance is [0, 1200] W / m², corresponding to from night to strong light conditions. The normal range of the normalized temperature-power feature is [0, 1.2], when , it indicates that there is a measurement anomaly or the component is overheated, and when , it indicates that the performance of the component is seriously degraded or blocked.

[0066] Step 103: Combine the temperature-power features in each time window into a feature vector , where denotes the feature vector dimension (time series length / feature number).

[0067] Further, the time series of the feature vector is expressed as:

[0068]

[0069]

[0070] where denotes the transpose operator, is the current time, is the time step, and the superscript denotes the feature value of the time window corresponding to the time . The constraint range of the time window length is [1 hour, 24 hours], which ensures the capture of complete daily periodic variation patterns.

[0071] Further, the value range of the feature vector dimension is [12, 288], corresponding to time resolutions from the hour level (12 hour windows) to the 5 minute level (288 5 minute windows). When When data is insufficient to capture the diurnal variation pattern, when The computational complexity is too high and the gain is limited.

[0072] Step 200: Analyze the temperature gradient impact using a graph convolution network to generate a node-specific temperature compensation matrix.

[0073] The electrical topology of the photovoltaic power station is represented as a graph , where the nodes represent photovoltaic strings or inverters, and the edges represent electrical connection relationships. Before inputting the graph convolution network, the node feature data is standardized for preprocessing: the temperature data, electrical parameters and geographical location information are respectively standardized by Z-score to eliminate the influence of different physical quantity dimension differences on network training. The preprocessed temperature distribution data and electrical topology are input into the graph convolution network to capture the propagation and influence pattern of temperature in the electrical network through multi-layer graph convolution operation. The graph convolution network outputs the temperature influence weight of each node to form a node-specific temperature compensation matrix .

[0074] Further, the construction method of the node feature matrix is as follows: for node , its feature vector is:

[0075]

[0076] where denotes the transpose operator, is the normalized temperature, is the normalized power, and are the normalized voltage and current, and are the normalized geographical coordinates, and are the mean and standard deviation of the corresponding parameters respectively. The construction method of the adjacency matrix is as follows: if nodes and are directly connected in the electrical network, then , otherwise , the diagonal element .

[0077] It should be noted that the graph convolution network belongs to the general application of graph neural network (GNN), and its input layer receives the node feature matrix and the adjacency matrix , where denotes the number of nodes involved in the calculation, This represents the node feature dimension. The input node features include temperature data, electrical parameters, and geographical location information of the photovoltaic string. The output layer generates a temperature influence weight matrix. This indicates the temperature compensation relationship between the nodes.

[0078] The aforementioned graph convolutional networks include For each graph convolutional layer, the forward propagation formula is: ,in For the first Hidden states of layer (the first layer) (layer output features) To add self-connected adjacency matrices, This is the corresponding degree matrix. For the first The learnable weight matrix of the layer, This is the ReLU activation function.

[0079] Furthermore, the degree matrix The calculation method is as follows: That is, the first The degree of a node is equal to the number of nodes connected to it (including self-joins), and the degree of off-diagonal elements is equal to the number of nodes connected to it. ( ).in Indicates from arrive Sum of the nodes. Weight matrix. The Xavier initialization method is used, with initial values ​​starting from a uniform distribution. Random sampling.

[0080] Furthermore, the number of layers in a graph convolutional network The value range of is [2, 6]. When The network's expressive power is insufficient to capture complex temperature propagation patterns. Overfitting and gradient vanishing problems can easily occur. Feature dimensions of each layer. The constraint range is [16, 128], which ensures that the network has sufficient expressive power and controllable computational complexity.

[0081] The aforementioned graph convolutional network was trained using a supervised learning model, employing historical temperature-compensated data as labels. Loss function... for:

[0082]

[0083] in The predicted temperature compensation coefficient, This is the actual compensation coefficient. For regularization parameters, is the Frobenius norm. Where and denote the summation of nodes from to and from to respectively, denotes the summation of network layers from to . The optimization strategy adopts Adam optimizer with an initial learning rate of 0.001, and uses cosine annealing strategy for adjustment.

[0084] Further, the time constraint form of the loss function is:

[0085]

[0086]

[0087]

[0088] where is the training time window length, is the time decay weight function ( denotes the natural exponential function), days is the decay coefficient, is the historical time index. Where and denote the summation of nodes from to and from to respectively, denotes the summation of time windows from to . The constraint conditions include: days, (normalization condition), (predicted value physical constraint).

[0089] Further, the calculation method of the real temperature compensation coefficient is: based on the historical operation data, the power output ratio of nodes and under the same irradiance condition is calculated, where and are the standard condition power of nodes and respectively, which are calculated by sliding window (window size is 30 days). When the historical data is insufficient, the similarity method based on geographical distance is used to estimate: where is the reference value, is the geographical distance between nodes, km is the characteristic distance.

[0090] Further, the regularization parameter is in the range of [1×10 -5 ,1×10 -2 ], when , the regularization effect is insufficient and prone to overfitting, when , over-constraint leads to underfitting. The effective range of the temperature compensation coefficient is [0.5, 2.0], indicating the multiple relationship of the temperature influence between nodes. The decay range of the learning rate is [0.001, 1×10 -6 ], and the training termination condition is that the loss function changes less than 1×10 -4 for 10 consecutive epochs or reaches the maximum training number of 500 times.

[0091] Step 300: Calculate the deviation coefficient based on the standard temperature reference, and generate the temperature normalization mapping function.

[0092] Set the standard temperature reference (such as 25℃), calculate the deviation of each node environment temperature relative to the standard temperature. Based on the temperature deviation and the temperature characteristic curve of the photovoltaic module, generate the temperature normalization mapping function , which is used to map the model parameters under different temperature conditions to a unified standard temperature reference.

[0093] Further, the method for obtaining the temperature characteristic curve of the photovoltaic module is: collecting the power output data of each node under different temperature conditions, and performing statistics according to the temperature interval [every 5℃ as a grade], and fitting the power-temperature relationship curve by the least square method , where is the first-order temperature coefficient, is the second-order temperature coefficient. The fitting data amount requires no less than 100 sample points, and the fitting correlation coefficient is considered valid.

[0094] Further, the setting range of the standard temperature reference is [20℃, 30℃], and 25℃ is usually selected as the international standard test condition. The effective range of the temperature deviation is [-65℃, 60℃], corresponding to the environmental temperature difference from the polar region to the desert region. When , the abnormal processing mechanism needs to be started.

[0095] It should be noted that the specific form of the temperature normalization mapping function is:

[0096]

[0097] in The input model parameter vector, This is a dimensionless temperature-sensitive adjustment parameter. The temperature response coefficient is in °C. Ensure It is a dimensionless quantity. The node temperature deviation (in °C) is the unit of measurement. The hyperbolic tangent function is used to limit the range of temperature influence. The temperature normalization mapping function standardizes the effect of temperature deviation on model parameters through nonlinear transformation.

[0098] Furthermore, the time-dependent form of the mapping function for:

[0099]

[0100] All parameters are time. The function. Time constraints include: parameter update frequency. The time correlation of temperature deviation is satisfied (times / day). ℃ (smoothness constraint), parameter adjustment cycle This ensures the timeliness of adaptive responses.

[0101] Furthermore, temperature sensitivity adjustment parameters The value range is [0.1, 0.8]. The temperature compensation effect is not obvious when Overcompensation may introduce new biases. Temperature response coefficient. The value range is [0.01, 0.05]℃ The range of values ​​for the temperature response coefficient ensures that the nonlinear characteristics of the temperature effect function are reasonable. The output multiple range of the mapping function is constrained to [0.2, 1.8] to prevent overcompensation under extreme temperature conditions.

[0102] Furthermore, parameters and The adaptive adjustment method is as follows: based on the prediction error on the validation set, the optimal value is determined by a grid search method. The search step size is 0.005℃. , The search step size is 0.1, and the evaluation metric is the root mean square error (RMSE). The search terminates when the RMSE changes by less than 0.001 for three consecutive iterations. The initial value is set to... ℃ , .

[0103] Step 400: Correcting model parameters with environment normalization method to generate environment corrected model update vector.

[0104] The local model update vector of each node and the node-specific temperature compensation matrix are input into the environment normalization processing module. The environment normalization processing module first applies a temperature normalization mapping function to transform the original parameters, and then combines the node-specific temperature compensation coefficients for fine adjustment, outputting the environment corrected model update vector , where represents the element-wise product.

[0105] Further, the algorithm flow of the environment normalization processing module is as follows: input check: verify that the dimension of is consistent with the dimension of the model parameters, and check the numerical range of ; temperature normalization: calculate , where is the temperature normalized local model update vector; spatial compensation: execute ; numerical stability check: if , then adopt the truncation processing .

[0106] The aforementioned environment normalization method includes the following processing steps:

[0107] Step 4001: Calculate the temperature deviation vector of each node , where is the temperature vector of node , and is the standard reference temperature vector;

[0108] Step 4002: Based on the temperature deviation vector and the temperature coefficient of the photovoltaic module, construct the parameter correction matrix , where is the temperature coefficient, is the diagonalization operation;

[0109] Further, the constraint range of the temperature coefficient is [-0.006, -0.003]℃ , and the constraint range of the temperature coefficient covers the temperature characteristics of different types of photovoltaic modules. The diagonal elements of the parameter correction matrix are constrained in the range [0.4, 1.6], ensuring that the correction factor is within a reasonable range.

[0110] Step 4003: Linearly transform the local model update vector through the parameter correction matrix to obtain the temperature normalized parameter ;

[0111] Step 4004: Extract the first column of the node-specific temperature compensation matrix as the node compensation vector, and perform element-wise multiplication with the normalization parameter to generate the final environment correction vector . .

[0112] Step 500: Perform federated aggregation on the environment-corrected model update vectors to output the global digital twin model.

[0113] Collect all node environment-corrected model update vectors and perform aggregation using the federated aggregation method: where denotes the number of nodes participating in federated aggregation, denotes the summation of nodes from to . After aggregation, the environment-robust global digital twin model adapted to multiple climate regions is obtained.

[0114] Further, the synchronization control method for federated aggregation is as follows: set the aggregation round interval seconds, and start aggregation when the number of participating nodes , where is the set of active nodes. For delayed node updates, adopt the expired discard strategy: if the update delay exceeds , discard the node update. Before aggregation, perform numerical stability check: calculate the L2 norm of the update vector , if it exceeds the threshold , perform truncation processing.

[0115] The input of the aforementioned federated aggregation method includes: the set of environment-corrected model update vectors of each node and the current global model parameter ; the output is the updated global model parameter .

[0116] It should be noted that the output decoding process of the global digital twin model includes the following steps: input the global model parameter into the photovoltaic power generation prediction neural network, the network receives the current environmental conditions (temperature, irradiance, humidity, etc.) as input features, and calculates the power output prediction value, generation efficiency prediction value, and equipment health status evaluation result of the photovoltaic array through forward propagation. These prediction results can be directly used for real-time monitoring, power scheduling, and maintenance decision-making of photovoltaic power stations.

[0117] Further, the specific structure of the photovoltaic power generation prediction neural network is: the dimension of the input layer is 8 (including temperature, irradiance, humidity, wind speed, altitude, azimuth angle, tilt angle, and time coding); the hidden layer adopts a 3-layer full connection structure, the number of neurons is [64, 32, 16] respectively, the activation function uses ReLU, and the dimension of the output layer is 3 (power prediction, efficiency prediction, and health status).

[0118] The total number of network parameters is about:

[0119] The forward propagation formula is: .

[0120] In the embodiments of the present application, in order to further improve the model quality, the step 500 further includes the following steps before the step 500:

[0121] Step 401: Obtain the local verification indicators and historical update records of each node, identify abnormal updates by using a statistical anomaly detection method, and generate an abnormal marking matrix.

[0122] Specifically, first, the input data is normalized and preprocessed: the local verification indicators (usually percentage data) and the statistical features of the historical update records are standardized to eliminate the influence of the dimension difference of different data types. Then, the statistical features (mean, variance) of the model update of each node are calculated, and the isolated forest method or the density-based anomaly detection technology is used to identify the abnormal updates deviating from the group distribution. For the detected abnormal nodes, the corresponding positions in the abnormal marking matrix are marked as 1, and the normal nodes are marked as 0.

[0123] Further, the construction method of the statistical feature matrix is: for a node , the feature vector , where denotes the transpose operator, is the average module length of the update vector, is the variance, and are skewness and kurtosis respectively, is the time series autocorrelation coefficient. Wherein denotes the summation of the time window from to . The time window is set to the last 30 updates.

[0124] The input of the aforementioned isolated forest method is the statistical feature matrix of the node model update vector, where denotes the feature dimension (the number of statistical features); and the output is an abnormal score vector , the abnormality label is generated by threshold judgment.

[0125] Further, the specific implementation steps of the isolation forest algorithm are as follows: random sampling: randomly sampling samples from the feature matrix to construct a training subset; random segmentation: for each isolated tree, randomly selecting a feature dimension and a segmentation threshold , recursively segmenting until the samples are isolated or the maximum depth is reached ; path length calculation: calculating the isolation path length in each tree for the test sample ; abnormal score : calculation , where is the average path length, is the normalization constant, is the harmonic number. The number of trees is set to 100. Where represents the feature dimension (the number of statistical features).

[0126] Further, the feature dimension ranges from 5 to 15, including mean, variance, skewness, kurtosis, and time series features. The abnormal score ranges from 0 to 1, and the abnormal detection threshold is set to 0.6, which is marked as an abnormal node when . The abnormal detection threshold is set based on experience and can be adjusted within the range of [0.5, 0.8] according to the system operation.

[0127] Step 402: Based on the historical performance of the blockchain record nodes, calculate the time-decaying reputation points and generate the reputation weight vector.

[0128] Record the historical model contribution quality, data integrity and update consistency of each node on the blockchain to ensure the non-tamperability of the record. When calculating the reputation points, use the time-decaying function , where is the quality score of the th contribution (dimensionless), is the decay coefficient and has the same dimension as the time reciprocal (e.g., days ) to ensure that the exponential part of the exponential function is dimensionless, is the contribution time, represents the sum of the contribution records from to . Based on the reputation points, generate a normalized reputation weight vector.

[0129] Further, the complete constraint condition of the reputation decay function is: time window constraint Days (history records beyond time window constraint are not involved in the calculation), contribution record number constraint (ensure statistical validity), time interval constraint Days (avoid noise from frequent updates), decay weight normalization condition , and numerical stability condition (terms less than this value can be ignored). Wherein represents the timestamp of the th contribution record, represents the contribution record index.

[0130] Further, the quality score ranges from [0, 1], where 0 represents complete untrustworthiness and 1 represents complete trustworthiness, which is usually evaluated according to the model validation accuracy, data integrity and update consistency. The decay coefficient ranges from [0.001, 0.1] days , when , the historical information decays too slowly, affecting real-time performance, and when , the decay is too fast, losing effective historical information. The element constraint of the reputation weight vector is and .

[0131] Step 403: input the environment-corrected model update vector and the reputation weight vector into the weighted federated aggregation module, and output the environment-robust global digital twin model. Modify the federated aggregation formula as: , wherein is used to shield the contribution of abnormal nodes.

[0132] Further, the time series optimization objective function of federated aggregation is :

[0133]

[0134] Wherein is the target model parameter, is the evaluation time window, is the regularization coefficient. Wherein represents the summation from to , represents the summation from to . The constraint conditions include: convergence constraint , stability constraint , and time consistency constraint .

[0135] The aforementioned weighted federated aggregation module comprises the following processing steps:

[0136] Step 4031: Calculate the set of valid nodes , filter out the nodes not marked as abnormal;

[0137] Step 4032: Normalize the reputation weight of the valid nodes, and calculate the adjusted weight ;

[0138] Further, the adjusted weight satisfies and . When the number of valid nodes , the system enters the protection mode and suspends the federated learning to wait for more trusted nodes to join.

[0139] Step 4033: Calculate the weighted update amount based on the adjusted weight and the environment correction vector , wherein represents the sum of all in the set of valid nodes ;

[0140] Step 4034: Add the weighted update amount to the current global model parameters to generate new global model parameters .

[0141] In the embodiments of the present application, in order to realize continuous optimization, the following steps are further included:

[0142] Step 600: Evaluate the performance of the global model using the validation data sets of different climate zones, and update the temperature compensation coefficient and reputation score parameters according to the prediction error feedback.

[0143] Independent validation data sets are collected from each climate zone, and the prediction error of the global model under different temperature conditions is calculated. Based on the error analysis results, the weight parameters of the graph convolution network and the decay coefficient of the reputation score are updated using the gradient descent method to realize adaptive optimization.

[0144] Further, the construction method of the validation data set is as follows: 10 representative nodes are selected from each climate zone (tropical, temperate, frigid, arid, humid), and each node provides 60 days of operation data as a validation set, including environmental parameters (temperature, irradiance, humidity) and actual power generation. The validation set is divided into a parameter adjustment set and a test set in a ratio of 7:3. The data preprocessing adopts the same standardization method as the training set to ensure data distribution consistency.

[0145] The input of the aforementioned gradient descent method is the loss function , and the gradient of the parameter is ; output is updated parameter wherein is the learning rate.

[0146] Further, the time series update rule of gradient descent is wherein the time index is in days, is the time of the model parameter, is the time of the model parameter, the learning rate time decay strategy is , the decay coefficient is in days . The time constraints include: the maximum optimization time is in days, the parameter change rate constraint (prevent drastic changes), and the performance monotonicity constraint within the time window wherein is the allowable growth rate.

[0147] Further, the specific method of gradient calculation is: the loss function is defined as wherein is the number of validation samples, is the true power generation, is the predicted power generation, denotes the summation of validation samples from to . The gradient is calculated using the backpropagation algorithm: for the layer parameter , its gradient is wherein denotes the transpose operator, is the output of the layer. The adaptive learning rate adjustment uses the Adam algorithm:

[0148]

[0149]

[0150]

[0151] wherein is the first moment estimate of the Adam algorithm (the first moment estimate of the gradient in the Adam optimization algorithm), is the first moment estimate of the last time, is the second moment estimate of the Adam algorithm (the second moment estimate of the gradient in the Adam optimization algorithm), is the second moment estimate of the last time, is the updated parameter (the model parameter of the next time) is a bias-corrected first moment estimate (bias-corrected first moment estimate), is a bias-corrected second moment estimate (bias-corrected second moment estimate), , , .

[0152] Further, the learning rate is in the range of [1×10 -5 , 0.01], when convergence is too slow to affect real-time performance, and when may lead to parameter oscillation and non-convergence. The prediction error is evaluated using the mean absolute percentage error (MAPE), and when MAPE>15%, the parameter update mechanism is triggered. The parameter convergence criterion is that the change in the parameter after 5 consecutive updates is less than 1×10 -4 or the maximum number of updates is reached, which is 100 times.

[0153] The following is a real application scenario of the present application:

[0154] A distributed digital twin modeling system is constructed by selecting 5 photovoltaic power station nodes distributed in different climate regions, including A city (temperate continental climate), B city (subtropical monsoon climate), C city (temperate continental arid climate), D city (temperate monsoon climate), and E city (tropical marine monsoon climate). The installed capacity of each node is 50 MW, and single-crystal silicon photovoltaic modules are used, with a power temperature coefficient / ℃.

[0155] Taking the period from 12:00 to 14:00 on July 2024 as an example, the original monitoring data collected by each node is shown in Table 1:

[0156] Table 1: Original monitoring data of each node

[0157]

[0158] Based on the above data, the temperature-power feature vector of each node is calculated according to the method of step 100:

[0159] For Node_1 (A city node), the temperature coefficient correction factor is calculated:

[0160]

[0161] The normalized temperature-power feature is calculated:

[0162]

[0163] The eigenvalues of other nodes are calculated as well, and the features in each time window are combined into a vector according to step 103. Taking the time window (1-hour sampling) as an example, the temperature-power feature vectors of each node are obtained as shown in Table 2 below:

[0164] Table 2: Temperature-power feature vectors (12:00-14:00 time window)

[0165]

[0166] According to step 200, the temperature gradient influence is analyzed using a graph convolution network. The node feature matrix is constructed, where the normalized node features are shown in Table 3 below:

[0167] Table 3: Normalized node feature matrix

[0168]

[0169] After processing by the graph convolution network, the generated node-specific temperature compensation matrix has some elements shown in Table 4 below:

[0170] Table 4: Temperature compensation matrix (partial elements)

[0171]

[0172] According to steps 300-400, the environmental normalization method is applied. First, according to sub-step 4001, the temperature deviation vector of each node is calculated:

[0173] For Node_2 (B city node), the environmental temperature is 38.5°C, and the standard temperature reference :

[0174]

[0175] According to sub-step 4002, the parameter correction matrix is constructed, with the temperature coefficient °C :

[0176]

[0177] According to sub-step 4003, the temperature-normalized parameters are:

[0178]

[0179] According to sub-step 4004, the first row of the temperature compensation matrix is extracted for element-wise multiplication to generate the final environmental correction vector:

[0180]

[0181]

[0182] As shown in the following table 5:

[0183] Table 5: Comparison of model update vectors before and after environmental correction (processed according to sub-steps 4001-4004)

[0184]

[0185] According to steps 500 and step 403, weighted federated aggregation is performed. The reputation weight of each node is , the abnormal flag matrix (no abnormal node), is calculated according to the aggregation formula:

[0186]

[0187] According to the processing procedure of steps 4031-4034:

[0188] Step 4031: Effective node set (no abnormal node)

[0189] Step 4032: The weight has been normalized,

[0190] Step 4033: Calculate the weighted update amount

[0191]

[0192]

[0193] Step 4034:

[0194] As shown in the following table 6:

[0195] Table 6: Comparison of federated aggregation results (based on step 403 weighted aggregation formula)

[0196]

[0197] As can be seen from the implementation results, after the environmental normalization federated protocol processing, the model prediction accuracy is improved from 12.8% of the traditional method to 7.1%, and the convergence speed is improved by 37.8%, effectively solving the problem of systematic deviation caused by temperature difference, and realizing high-precision distributed photovoltaic power station digital twin modeling across climate regions.

[0198] It can be understood that the data preprocessing method known to those skilled in the art includes data cleaning, data conversion, data reduction, wherein the data conversion includes type conversion and normalization and standardization, although the dimension and type of the data are ignored in the foregoing embodiment description, the data preprocessing is the technical knowledge known to those skilled in the art and the preprocessing step of data processing, and therefore the foregoing does not independently describe the known data preprocessing step.

[0199] The foregoing describes the embodiments of the present application, but the embodiments are not limited to the specific implementation described above, and the specific implementation described above is only illustrative but not restrictive, and those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which all belong to the protection of the embodiments.

Claims

1. A method for digital twin modeling of a photovoltaic power station, characterized in that, Includes the following steps: Acquire monitoring data from each node of a distributed photovoltaic power station and generate a temperature-power feature vector; The influence of temperature gradient is analyzed using graph convolutional networks, generating node-specific temperature compensation matrices. The deviation coefficient is calculated based on the standard temperature reference, and a temperature normalization mapping function is generated. The model parameters are corrected using the environment normalization method, and an environment-corrected model update vector is generated. Federated aggregation of the environmentally corrected model update vectors is performed to output a global digital twin model; The environmental normalization method transforms the original parameters through a temperature normalization mapping function and performs fine-tuning by combining a node-specific temperature compensation coefficient.

2. The method for digital twin modeling of photovoltaic power plants according to claim 1, characterized in that, The generated temperature-power feature vector includes: Calculate the temperature coefficient correction factor for the photovoltaic module, which is equal to 1 plus the product of the power temperature coefficient and the difference between the module temperature and the standard test condition temperature. Based on the temperature coefficient correction factor and the actual power output, the normalized temperature-power characteristic is calculated. The normalized temperature-power characteristic is equal to the actual power divided by the product of the standard condition power, the temperature coefficient correction factor, and the irradiance ratio. The temperature-power features within each time window are combined into a feature vector.

3. The method for digital twin modeling of photovoltaic power plants according to claim 1, characterized in that, The analysis of the effect of temperature gradient using graph convolutional networks includes: The electrical topology of a photovoltaic power station is represented as a graph, where nodes represent photovoltaic strings or inverters, and edges represent electrical connections. Standardize and preprocess the node feature data to eliminate differences in the dimensions of different physical quantities; The preprocessed temperature distribution data and electrical topology are input into a graph convolutional network, and the propagation and influence patterns of temperature in the electrical network are captured through multi-layer graph convolution operations. The graph convolutional network outputs the temperature influence weights of each node, forming a node-specific temperature compensation matrix.

4. The method for digital twin modeling of photovoltaic power plants according to claim 1, characterized in that, The generated temperature normalization mapping function includes: Set a standard temperature reference and calculate the deviation of the ambient temperature at each node from the standard temperature; A temperature normalization mapping function is generated based on the temperature deviation and the temperature characteristic curve of the photovoltaic module. The temperature normalization mapping function standardizes the influence of temperature deviation on model parameters through nonlinear transformation. The nonlinear transformation includes applying a hyperbolic tangent function to the temperature deviation to limit the range of temperature influence.

5. The method for digital twin modeling of photovoltaic power plants according to claim 1, characterized in that, The environment normalization method includes: Calculate the temperature deviation vector for each node; A parameter correction matrix is ​​constructed based on the temperature deviation vector and the temperature coefficient of the photovoltaic module; The local model update vector is linearly transformed using the parameter correction matrix to obtain the temperature-normalized parameters. The corresponding row of the node-specific temperature compensation matrix is ​​extracted as the node compensation vector, and multiplied element-wise with the normalization parameter to generate the final environmental correction vector.

6. The method for digital twin modeling of photovoltaic power plants according to claim 1, characterized in that, This also includes the following before performing federated aggregation: Obtain local verification metrics and historical update records for each node, identify abnormal updates using statistical anomaly detection methods, and generate an anomaly marker matrix; Based on the historical performance of nodes recorded in the blockchain, a reputation score with time decay is calculated to generate a reputation weight vector. The environmentally corrected model update vector and reputation weight vector are input into the weighted federated aggregation module.

7. The method for digital twin modeling of photovoltaic power plants according to claim 6, characterized in that, The statistical anomaly detection method includes: Calculate the statistical characteristics of the model updates at each node, including mean, variance, skewness, kurtosis, and time-series autocorrelation coefficient; Identify anomalous updates that deviate from the population distribution using the isolated forest method; For detected abnormal nodes, the corresponding position in the abnormal marking matrix is ​​marked as 1, and normal nodes are marked as 0.

8. The method for digital twin modeling of photovoltaic power plants according to claim 6, characterized in that, The weighted federated aggregation includes: Calculate the set of valid nodes and filter out nodes that have not been marked as abnormal; The reputation weights of valid nodes are normalized. Calculate the weighted update amount based on the adjusted weights and environmental correction vector; The weighted update amount is added to the current global model parameters to generate new global model parameters.

9. The method for digital twin modeling of photovoltaic power plants according to claim 1, characterized in that, Also includes: Evaluate the global model performance using validation datasets from different climate zones; The temperature compensation coefficient and credit score parameters are updated based on the prediction error feedback to achieve adaptive optimization.

10. A photovoltaic power plant digital twin modeling system, used to execute the photovoltaic power plant digital twin modeling method according to any one of claims 1-9, characterized in that, include: The data acquisition module is used to acquire monitoring data from each node of the distributed photovoltaic power station and generate temperature-power feature vectors; The temperature compensation analysis module is used to analyze the effects of temperature gradients using graph convolutional networks and generate node-specific temperature compensation matrices. The environmental normalization module is used to generate a temperature normalization mapping function based on a standard temperature reference and to correct model parameters. The quality screening module is used to identify abnormal nodes and calculate reputation weights; The federated aggregation module is used to perform weighted federated aggregation to generate a global digital twin model.

Images