A Modeling Method for Multi-Dimensional Information Fusion of a Photovoltaic Power Plant
By using multi-dimensional information fusion modeling, the problem of information silos in the digital modeling of photovoltaic power plants has been solved. This has enabled unified modeling of high-precision, real-time equipment status monitoring and dynamic operating status of photovoltaic power plants, improving the accuracy of equipment anomaly location and power prediction.
Patent Information
- Application Number
- CN202611285027.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-24
- Publication Date
- 2026-11-17
AI Technical Summary
Existing digital modeling of photovoltaic power plants is mostly single-dimensional, which cannot fully reflect the multi-dimensional coupled operation status of the plant. This leads to information silos in fault diagnosis and performance evaluation, making it impossible to achieve accurate equipment anomaly location and status assessment.
A multi-dimensional information fusion modeling method is adopted. By acquiring multi-dimensional data on geospatial, meteorological environment, electrical operation and equipment physical status, a unified multi-dimensional digital model is constructed by using comparative self-supervised pre-training, spatiotemporal dynamic graph convolution and physical law constraints. The model self-calibration and uncertainty quantification are achieved through error feedback loop.
It achieves high-precision, real-time, and reliable equipment status mapping and dynamic operation status monitoring in photovoltaic fields, improves the accuracy of equipment anomaly location and power prediction, reduces model calculation latency, and meets real-time operation and maintenance needs.
Smart Images

Figure CN123116808A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of digital modeling technology for photovoltaic power generation stations, and in particular to a modeling method for integrating multi-dimensional information of the entire photovoltaic power generation station. Background Technology
[0002] With the large-scale development of the photovoltaic industry, the scale of photovoltaic power plants is increasing daily, and the complexity of their operation and maintenance management is also increasing dramatically. A complete photovoltaic field includes tens of thousands of photovoltaic modules, a large number of electrical devices, and a complex terrain environment. Its operating status is affected by the coupling of multiple dimensions of factors, including geographical environment, meteorological conditions, physical condition of equipment (such as temperature and dust accumulation), and electrical connection relationships. Currently, digital modeling of photovoltaic fields usually focuses on a single dimension, such as using Geographic Information Systems (GIS) to build a three-dimensional geometric model, or using Supervisory Control and Data Acquisition (SCADA) systems to monitor electrical data. These isolated models cannot fully reflect the overall, dynamic, and multi-dimensional coupled operating status of the field, resulting in an "information silo" phenomenon in fault diagnosis, performance evaluation, and power prediction, making it impossible to achieve accurate equipment anomaly location and status assessment. Summary of the Invention
[0003] The purpose of this invention is to provide a modeling method for multi-dimensional information fusion of a photovoltaic power plant. This invention can deeply integrate multi-dimensional heterogeneous information of a photovoltaic power plant, thereby constructing a unified, real-time, and high-fidelity digital model, achieving a comprehensive and accurate mapping of the power plant's physical entities and dynamic operating status.
[0004] The technical solution provided by this invention is as follows: A modeling method for multi-dimensional information fusion of a photovoltaic power plant, comprising the following steps: Step S1: Obtain raw data from multiple dimensions of the photovoltaic field, including geospatial dimension, meteorological environment dimension, electrical operation dimension, and equipment physical status dimension; Step S2: Perform spatiotemporal registration and data cleaning on the original data to obtain an aligned multi-source dataset; Step S3: Adopt a contrastive self-supervised pre-training strategy, take data of different dimensions at the same time as positive sample pairs and multi-dimensional data of different times or mismatched as negative sample pairs, pre-train the feature extraction network through the contrastive loss function, so that the features of each dimension are aligned in the latent space, and then extract feature vectors from the data of each dimension in the multi-source dataset to obtain a multi-dimensional feature set. Step S4: A multimodal fusion network is used to perform cross-dimensional feature fusion on the multi-dimensional feature set to generate a fused feature tensor. The multimodal fusion network is embedded with physical constraints based on Kirchhoff's current law and the single diode model of photovoltaic modules, and adopts a spatiotemporal dynamic graph convolutional fusion architecture. It dynamically generates a graph adjacency matrix based on geographic spatial distance and electrical connection relationship, extracts spatial coupling and temporal evolution features between devices through spatiotemporal graph convolution, and then modulates the fused features with meteorological environmental features through a cross-modal attention mechanism. Step S5: Based on the fusion feature tensor, construct a three-dimensional geometric model and a dynamic operation status model of the entire photovoltaic power station to form a unified multi-dimensional information fusion digital model.
[0005] In the aforementioned modeling method for multi-dimensional information fusion of the entire photovoltaic power plant, the contrast loss function in step S3 is: ; In the formula, and These are the latent feature vectors of positive sample pairs. The cosine similarity function is used. For temperature coefficient, The number of samples in a training batch. For indicator functions, when The value is 1 if it is true and 0 otherwise.
[0006] In the aforementioned modeling method for multi-dimensional information fusion of the entire photovoltaic power plant, step S4, the spatiotemporal dynamic graph convolutional fusion architecture adopted by the multimodal fusion network includes: Dynamic graph construction unit, dynamically generating graph adjacency matrix based on geographic spatial distance and electrical connection relationships. The adjacency matrix changes with time step Update, its element is defined as: ; in, For equipment With equipment Geographical distance between them For scale parameters, For electrical connection edge set, For indicating functions, when the device With equipment The value is 1 if an electrical connection exists, otherwise it is 0. The spatiotemporal graph convolutional layer uses graph convolution to extract coupling features between devices in the spatial dimension and gated recurrent units to capture dynamic evolution patterns in the temporal dimension; the graph convolution operation is represented as: ; in, , It is the identity matrix. for The degree matrix, For the first Input features of layer graph convolution, For the first Layer-learnable weight matrix, For activation functions; A cross-modal attention bridging layer integrates meteorological and environmental features. As conditional information, the output of the spatiotemporal graph convolution Perform feature modulation: ; in, As a multi-head attention mechanism, , , These are query, key, and value matrices, respectively. It is a multilayer perceptron. This is the modulated feature tensor output by the cross-modal attention bridging layer.
[0007] In the aforementioned modeling method for multi-dimensional information fusion of the entire photovoltaic field, in step S4, the physical law constraint module calculates the residual constraint term of the electrical dimension based on Kirchhoff's current law and the single diode model of the photovoltaic module, and adds the residual constraint term as a regularization loss to the training process of the fusion network. The single diode model is represented as follows: ; In the formula, To output current for photovoltaic modules, For output voltage, For photocurrent, It is the reverse saturation current. For series resistance, For parallel resistors, As an ideal factor, Thermoelectric voltage; The Kirchhoff's current law constraint is: for any combiner box node ,satisfy ,in This is the index of the branch lines under the combiner box. For the first The current in each branch; The residual constraint term is defined as follows: ; in, The current value reconstructed from the fused feature tensor. This is the measured current value. and These are the weighting coefficients.
[0008] The aforementioned modeling method for multi-dimensional information fusion of the entire photovoltaic field also includes step S6: constructing a parallel digital twin error feedback loop; real-time acquisition of measured electrical data and infrared thermal imaging data of the photovoltaic field, performing point-by-point residual calculation on the measured electrical data and infrared thermal imaging data with the corresponding state variables output by the multi-dimensional information fusion digital model to obtain the residual field; using a differentiable physical information neural network to backpropagate the residual field to the fusion feature tensor and the intermediate layer of the model, updating some weight parameters of the multimodal fusion network online to achieve model self-calibration; The residual field is defined as follows: ; in, For spatial coordinates, Using time as the coordinate, The state variables output by the model. These are the corresponding measured state quantities; The update rule is as follows: ; in, These are some weight parameters in the multimodal fusion network that need to be updated online. For learning rate, For about The gradient operator.
[0009] The aforementioned modeling method for multi-dimensional information fusion of the entire photovoltaic field also includes step S7: quantifying the uncertainty of the output of the multi-dimensional information fusion digital model; the multi-modal fusion network adopts a Bayesian approximate inference architecture, and after fusing the feature tensor, it is connected to a Monte Carlo dropout layer to output the predicted mean and predicted variance of power, temperature and irradiance at each spatial location point, and the predicted variance is used as the model confidence level and labeled on the three-dimensional geometric model.
[0010] The aforementioned modeling method for multi-dimensional information fusion of the entire photovoltaic field includes the following data: the geospatial dimension data includes the digital elevation model, digital orthophoto model, and lidar point cloud data of the photovoltaic field; the meteorological environment dimension data includes real-time collected irradiance, ambient temperature, wind speed, humidity, and air pressure; the electrical operation dimension data includes the voltage, current, and power of each photovoltaic string, as well as the operating parameters of the combiner box and inverter; and the equipment physical state dimension data includes the surface temperature of the photovoltaic modules, the degree of dust accumulation, hot spot detection results, and the tilt angle of the support.
[0011] The aforementioned modeling method for multi-dimensional information fusion of the entire photovoltaic power plant includes the following steps in step S2: time synchronization of multi-source data at the same timestamp and mapping of all data to the same geographic coordinate system; data cleaning includes missing value imputation, outlier removal and filtering smoothing.
[0012] The aforementioned modeling method for multi-dimensional information fusion of the entire photovoltaic field includes a three-dimensional geometric model that comprises the topography of the photovoltaic field, the layout of the photovoltaic array, and the three-dimensional structure of buildings and roads; and a dynamic operating status model that comprises the real-time power generation distribution, temperature distribution, irradiance distribution, and inverter operating conditions of each photovoltaic string.
[0013] The aforementioned modeling method for multi-dimensional information fusion of the entire photovoltaic field is used for equipment anomaly location, power prediction, or component cleanliness assessment in the photovoltaic field.
[0014] Compared with the prior art, the present invention has the following significant advantages: 1. This invention integrates data from four dimensions: geospatial, meteorological environment, electrical operation, and equipment physical state. It also embeds a photovoltaic module single diode model and Kirchhoff's current law as regularization constraints in a multimodal fusion network, making the data-driven model strictly obey the laws of electrical physics. This overcomes the limitations of single-dimensional modeling and significantly improves the physical consistency and generalization ability of the model.
[0015] 2. This invention uses dynamic graph construction units to adaptively generate time-varying graph adjacency matrices based on geographical distance and electrical connections. It employs spatiotemporal graph convolutional layers to simultaneously capture the spatial coupling relationships and temporal evolution patterns between devices. Furthermore, it utilizes cross-modal attention bridging layers to modulate features using meteorological environmental characteristics as conditional information, thereby achieving high-precision modeling of complex spatiotemporal dynamics in photovoltaic fields.
[0016] 3. This invention constructs a parallel error feedback loop, calculates the residual field using measured electrical data and infrared thermal imaging data, and achieves online model self-calibration through backpropagation of a differentiable physical information neural network; at the same time, it uses Bayesian approximate inference to output the predicted mean and variance, and marks the model confidence on the three-dimensional geometric model, providing a reliable uncertainty boundary for operation and maintenance decisions such as equipment anomaly location and power prediction. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating steps S1 and S2 of the present invention. Figure 2 This is a flowchart illustrating step S3 of the present invention; Figure 3 This is a flowchart illustrating step S4 of the present invention; Figure 4 This is a flowchart illustrating step S5 of the present invention; Figure 5 This is a flowchart illustrating steps S6 and S7 of the present invention. Detailed Implementation
[0018] The present invention will be further described below with reference to embodiments, but these embodiments are not intended to limit the scope of the invention.
[0019] Example: A modeling method for multi-dimensional information fusion of a photovoltaic power plant. Its core lies in fusing heterogeneous data from four dimensions—geospatial, meteorological environment, electrical operation, and equipment physical status—through self-supervised pre-training, spatiotemporal dynamic graph convolution, physical constraint, and error feedback mechanisms to construct a unified, high-fidelity, and self-calibrating digital twin model. Specific steps are as follows: Step S1: Obtain raw data from the photovoltaic field across multiple dimensions, including geospatial dimensions, meteorological environment dimensions, electrical operation dimensions, and equipment physical status dimensions, such as... Figure 1 As shown; In this step, taking a 100MW photovoltaic power plant in actual operation as an example, four types of raw data are collected in the following ways: Geospatial dimension: UAVs equipped with LiDAR are used to acquire point cloud data of the site (accuracy 0.1m), which are combined with digital elevation model (DEM, resolution 0.5m) and digital orthophoto (DOM, resolution 0.1m) to construct the basic framework for terrain and equipment layout.
[0020] Meteorological environment dimension: Three small weather stations were deployed at the center and edge of the site to collect real-time data on irradiance (GHI, DNI, DHI), ambient temperature, wind speed, humidity, and air pressure at a sampling frequency of 1Hz.
[0021] Electrical operation dimension: The DC voltage, current, and power of each photovoltaic string (2000 strings in total) and the AC side parameters of the combiner boxes (80 in total) and inverters (5 in total) are obtained through the SCADA system. The sampling frequency is 10Hz, and the timestamp and device ID are marked.
[0022] Equipment physical status dimension: The surface temperature distribution map of the components is obtained by using an infrared thermal imaging drone (inspected every 2 hours), and the hot spot area is extracted by image recognition algorithm; the degree of dust accumulation (0-100%) of each string is estimated by using light reflection method or image grayscale analysis; the actual tilt angle of each row of brackets is obtained by tilt sensor.
[0023] Step S2: Perform spatiotemporal registration and data cleaning on the original data to obtain an aligned multi-source dataset.
[0024] In this step, time synchronization is based on the GPS clock, interpolating all data to a unified time axis (e.g., every 5 seconds). Linear interpolation is used for electrical data, spline interpolation for meteorological data, and nearest neighbor time matching for inspection data (such as thermal imaging). Spatial registration maps the spatial coordinates of all devices (strings, combiner boxes, inverters) to the UTM coordinate system (e.g., WGS84 / UTM 50N). A 3D mesh model is generated after registration of the LiDAR point cloud and DOM. Data cleaning uses the 3σ principle to remove abnormal jump points in the electrical data; missing values (e.g., communication interruption in a certain string) are filled using the mean of other strings in the same combiner box or time-series-based Kalman filtering; median filtering is used to smooth noise in thermal imaging. After the above processing, an aligned multi-source dataset is obtained, with each time point containing: the location coordinates (x, y, z) of each string, instantaneous irradiance, ambient temperature, string voltage / current / power, surface temperature, dust accumulation level, support tilt angle, etc.
[0025] Step S3: Employing a contrastive self-supervised pre-training strategy, different dimensions of data at the same time are used as positive sample pairs, while multi-dimensional data from different times or mismatched data are used as negative sample pairs. The feature extraction network is pre-trained using a contrastive loss function to align features across dimensions in the latent space. Then, feature vectors are extracted from each dimension of the multi-source dataset to obtain a multi-dimensional feature set, such as... Figure 2 As shown.
[0026] In this step, to avoid the cost of manual annotation and to achieve semantic alignment of multi-dimensional features, this embodiment adopts a contrastive self-supervised learning strategy for pre-training.
[0027] Positive and negative sample construction: Take the four-dimensional features (such as meteorological irradiance, electrical current, physical temperature, and geographic coordinate embedding) of the same string at the same time (e.g., t=10:00:00) as positive sample pairs; take the dimensional features of different times (e.g., morning and afternoon) or mismatches (e.g., the current of string A and the irradiance of string B) as negative sample pairs.
[0028] Feature extraction network: Each dimension uses an independent encoder (e.g., 1D-CNN for meteorology, LSTM for electrical engineering, MLP for physics, and GCN for geography), and outputs a fixed-length latent feature vector.
[0029] Contrastive Loss Function: The contrastive loss function in the form of NT-Xent (normalized temperature-scale cross-entropy loss) is used for training, and its expression is as follows: ; In the formula, and These are the latent feature vectors of positive sample pairs. The cosine similarity function is used. This is the temperature coefficient, with a value of 0.07. The number of samples in a training batch. For indicator functions, when The value is 1 if it is true and 0 otherwise.
[0030] By minimizing this loss, features of different dimensions of the same physical entity are forced to move closer together in the latent space, while features of different types move further away.
[0031] Feature extraction: After pre-training, the encoder's underlying parameters are frozen, and feature vectors of each dimension are extracted from all samples to form a multi-dimensional feature set.
[0032] Step S4: A multimodal fusion network is used to perform cross-dimensional feature fusion on the multi-dimensional feature set to generate a fused feature tensor. The multimodal fusion network is embedded with physical constraints based on Kirchhoff's current law and the single-diode model of photovoltaic modules, and adopts a spatiotemporal dynamic graph convolutional fusion architecture. It dynamically generates a graph adjacency matrix based on geographic spatial distance and electrical connection relationships, extracts spatial coupling and temporal evolution features between devices through spatiotemporal graph convolution, and then modulates the fused features with meteorological environmental features through a cross-modal attention mechanism, such as... Figure 3 As shown.
[0033] In this step, the spatiotemporal dynamic graph convolutional fusion architecture adopted by the multimodal fusion network includes: Dynamic graph construction unit, dynamically generating graph adjacency matrix based on geographic spatial distance and electrical connection relationships. The adjacency matrix changes with time step Update, its element is defined as: ; in, For equipment With equipment Geographical distance (Euclidean distance) between them. This is a scale parameter, with a value of 10m (which can be adjusted according to the size of the site). For electrical connection edge collection (such as string group under the same combiner box, combiner box and inverter). For indicating functions, when the device With equipment The value is 1 if an electrical connection exists, and 0 otherwise. This adjacency matrix is dynamically updated every 30 seconds to adapt to weather conditions or changes in electrical topology (such as circuit breaker switching).
[0034] The spatiotemporal graph convolutional layer uses graph convolution to extract coupling features between devices in the spatial dimension and gated recurrent units (GRUs) to capture dynamic evolution patterns in the temporal dimension. Specifically, the spatial feature sequence (temporal dimension) of each node after graph convolution is input into a GRU to capture the evolution of current, temperature, and other parameters over time (such as power fluctuations caused by sunrise / sunset and cloud cover). The graph convolution operation is represented as follows: ; in, , It is the identity matrix. for The degree matrix, For the first Input features of layer graph convolution, For the first Layer-learnable weight matrix, For activation functions; A cross-modal attention bridging layer integrates meteorological and environmental features. (e.g., irradiance, ambient temperature) are used as conditional information to influence the output of the spatiotemporal graph convolution. Perform feature modulation: ; in, It is a multi-head attention mechanism (8 heads). , , These are query, key, and value matrices, respectively (the query comes from device features, and the keys and values come from meteorological features). It is a multilayer perceptron. This is the modulated feature tensor output by the cross-modal attention bridging layer. This cross-modal attention bridging layer enables the model to dynamically learn "which strings' outputs should be enhanced or suppressed under the current irradiance".
[0035] Furthermore, the multimodal fusion network is embedded with physical constraints based on Kirchhoff's current law and the single diode model of photovoltaic modules. The residual constraint term of the electrical dimension is calculated and added as a regularization loss to the training process of the fusion network to ensure that the output conforms to the actual physical laws.
[0036] The single diode model is represented as follows: ; In the formula, To output current for photovoltaic modules, For output voltage, For photocurrent, It is the reverse saturation current. For series resistance, For parallel resistors, As an ideal factor, Thermoelectric voltage; The Kirchhoff's current law constraint is: for any combiner box node ,satisfy ,in This is the index of the branch lines under the combiner box. For the first The current in each branch; The residual constraint term is defined as follows: ; in, The current value reconstructed from the fused feature tensor. This is the measured current value. and These are the weighting coefficients, taken as 0.7 and 0.3 respectively.
[0037] Finally, the multimodal fusion network outputs a fused feature tensor. Where N is the number of devices and C is the feature dimension (e.g., 256).
[0038] Step S5: Based on the fused feature tensor, construct a three-dimensional geometric model and a dynamic operation status model of the entire photovoltaic power plant, forming a unified multi-dimensional information fusion digital model, such as... Figure 4 As shown.
[0039] In this step, the 3D geometric model is generated by overlaying LiDAR point clouds, DOM, and device coordinates using a GIS engine. This creates a 3D scene including terrain, roads, buildings, and photovoltaic arrays (each string is an independent entity), supporting scaling, rotation, and panoramic navigation. The dynamic operating status model maps the real-time power, surface temperature, irradiance, and dust accumulation of each string to color scales (e.g., red for high power / high temperature, blue for low power), dynamically updating the 3D model to form a "heatmap" overlay. Inverter operating conditions (e.g., efficiency, alarm status) are displayed as icons at corresponding locations. The final result is a unified, multi-dimensional information fusion digital model, allowing users to view all historical and real-time data curves of any string by clicking on it.
[0040] Furthermore, this embodiment may also include step S6, such as Figure 5 As shown, this step involves constructing a parallel digital twin error feedback loop; real-time acquisition of measured electrical data and infrared thermal imaging data from the photovoltaic field; and point-by-point residual calculation of the measured electrical data and infrared thermal imaging data with the corresponding state variables output by the multi-dimensional information fusion digital model every 10 minutes to obtain the residual field; the residual field is defined as: ; in, For spatial coordinates, Using time as the coordinate, The state variables output by the model (such as the current in a certain string). These are the corresponding measured state quantities; A differentiable physical information neural network is used to backpropagate the residual field to the fusion feature tensor and intermediate layers of the model, updating some weight parameters of the multimodal fusion network online to achieve model self-calibration. The update rule is as follows: ; in, These are some weight parameters in the multimodal fusion network that need to be updated online. The learning rate is set to 0.001. For about The gradient operator. After feedback updates, the model can automatically adapt to characteristic drift caused by increased dust accumulation, component aging, or sudden weather changes.
[0041] Furthermore, this embodiment may also include step S7, which involves quantifying the uncertainty of the output of the multi-dimensional information fusion digital model. The multimodal fusion network adopts a Bayesian approximate inference architecture, and after fusing the feature tensors, it connects to a Monte Carlo dropout layer (MC Dropout). During each inference, 20% of the neurons are randomly dropped, and 10 forward propagations are performed. The predicted mean and predicted variance of power, temperature, and irradiance at each spatial location point are output, and the predicted variance is used as the model confidence level, which is marked on the three-dimensional model in a semi-transparent cloud map manner: low variance (high confidence) areas are clearly displayed, while high variance areas are indicated by a warning color to remind maintenance personnel to conduct key checks.
[0042] Based on the digital model constructed above, the system can automatically implement the following in this embodiment: Anomaly localization: When the residual difference between the predicted power and the actual power of a certain string exceeds the 3σ threshold, and there are no obvious anomalies in the adjacent strings, the system highlights the string in the 3D model and prompts "Possible hot spot" or "Bypass diode failure".
[0043] Cleanliness assessment: Combining the characteristics of dust accumulation with historical meteorological irradiance data, the model can output a recommended cleaning time window when "current power loss is 12% due to dust accumulation".
[0044] In terms of technical performance, in an actual test at a 50MW photovoltaic power plant, the accuracy of module-level anomaly location was improved to 94.2% (compared to 71.5% for the traditional method) after adopting this method; the power prediction RMSE was reduced by 31.6%; and the model calculation latency was 87ms / frame on a GPGPU (such as NVIDIA A10), which meets the real-time requirements.
[0045] Through the detailed embodiments described above, those skilled in the art can fully implement the multi-dimensional information fusion modeling method for photovoltaic fields proposed in this invention, and understand the specific meaning and function of each technical feature in the claims. This embodiment is only illustrative; the parameters, network structure, and data sources in actual applications can be adjusted according to the specific field size and equipment model.
Claims
1. A modeling method for multi-dimensional information fusion of a photovoltaic power plant area, characterized in that, Includes the following steps: Step S1: Obtain raw data from multiple dimensions of the photovoltaic field, including geospatial dimension, meteorological environment dimension, electrical operation dimension, and equipment physical status dimension; Step S2: Perform spatiotemporal registration and data cleaning on the original data to obtain an aligned multi-source dataset; Step S3: Adopt a contrastive self-supervised pre-training strategy, take data of different dimensions at the same time as positive sample pairs and multi-dimensional data of different times or mismatched as negative sample pairs, pre-train the feature extraction network through the contrastive loss function, so that the features of each dimension are aligned in the latent space, and then extract feature vectors from the data of each dimension in the multi-source dataset to obtain a multi-dimensional feature set. Step S4: A multimodal fusion network is used to perform cross-dimensional feature fusion on the multi-dimensional feature set to generate a fused feature tensor. The multimodal fusion network is embedded with physical constraints based on Kirchhoff's current law and the single diode model of photovoltaic modules, and adopts a spatiotemporal dynamic graph convolutional fusion architecture. It dynamically generates a graph adjacency matrix based on geographic spatial distance and electrical connection relationship, extracts spatial coupling and temporal evolution features between devices through spatiotemporal graph convolution, and then modulates the fused features with meteorological environmental features through a cross-modal attention mechanism. Step S5: Based on the fusion feature tensor, construct a three-dimensional geometric model and a dynamic operation status model of the entire photovoltaic power station to form a unified multi-dimensional information fusion digital model.
2. The modeling method for multi-dimensional information fusion of the entire photovoltaic power plant area according to claim 1, characterized in that, In step S3, the contrast loss function is: ; In the formula, and These are the latent feature vectors of positive sample pairs. The cosine similarity function is used. For temperature coefficient, The number of samples in a training batch. For indicator functions, when The value is 1 if it is true and 0 otherwise.
3. The modeling method for multi-dimensional information fusion of the entire photovoltaic power plant area according to claim 1, characterized in that, In step S4, the spatiotemporal dynamic graph convolutional fusion architecture adopted by the multimodal fusion network includes: Dynamic graph construction unit, dynamically generating graph adjacency matrix based on geographic spatial distance and electrical connection relationships. The adjacency matrix changes with time step Update, its element is defined as: ; in, For equipment With equipment Geographical distance between them For scale parameters, For electrical connection edge set, For indicator functions, when the device With equipment The value is 1 if an electrical connection exists, otherwise it is 0. The spatiotemporal graph convolutional layer uses graph convolution to extract coupling features between devices in the spatial dimension and gated recurrent units to capture dynamic evolution patterns in the temporal dimension; the graph convolution operation is represented as: ; in, , It is the identity matrix. for The degree matrix, For the first Input features of layer graph convolution, For the first Layer-learnable weight matrix, For activation functions; A cross-modal attention bridging layer integrates meteorological and environmental features. As conditional information, the output of the spatiotemporal graph convolution Perform feature modulation: ; in, As a multi-head attention mechanism, , , These are query, key, and value matrices, respectively. It is a multilayer perceptron. This is the modulated feature tensor output by the cross-modal attention bridging layer.
4. The modeling method for multi-dimensional information fusion of the entire photovoltaic power plant area according to claim 3, characterized in that, In step S4, the physical law constraint module calculates the residual constraint term of the electrical dimension based on Kirchhoff's current law and the single diode model of the photovoltaic module, and adds the residual constraint term as a regularization loss to the training process of the fusion network. The single diode model is represented as follows: ; In the formula, To output current for photovoltaic modules, For output voltage, For photocurrent, It is the reverse saturation current. For series resistance, For parallel resistors, As an ideal factor, Thermoelectric voltage; The Kirchhoff's current law constraint is: for any combiner box node ,satisfy ,in This is the index of the branch lines under the combiner box. For the first The current in each branch; The residual constraint term is defined as follows: ; in, The current value reconstructed from the fused feature tensor. This is the measured current value. and These are the weighting coefficients.
5. The modeling method for multi-dimensional information fusion of the entire photovoltaic power plant area according to claim 1, characterized in that, It also includes step S6: constructing a parallel digital twin error feedback loop; real-time acquisition of measured electrical data and infrared thermal imaging data of the photovoltaic field, performing point-by-point residual calculation on the measured electrical data and infrared thermal imaging data with the corresponding state variables output by the multi-dimensional information fusion digital model to obtain the residual field; using a differentiable physical information neural network to backpropagate the residual field to the fusion feature tensor and the intermediate layer of the model, updating some weight parameters of the multimodal fusion network online to achieve model self-calibration; The residual field is defined as follows: ; in, For spatial coordinates, Using time as the coordinate, The state variables output by the model. These are the corresponding measured state quantities; The update rule is as follows: ; in, These are some weight parameters in the multimodal fusion network that need to be updated online. For learning rate, For about The gradient operator.
6. The modeling method for multi-dimensional information fusion of the entire photovoltaic power plant area according to claim 1, characterized in that, It also includes step S7: performing uncertainty quantification on the output of the multi-dimensional information fusion digital model; the multimodal fusion network adopts a Bayesian approximate inference architecture, and after fusing the feature tensor, it is connected to the Monte Carlo dropout layer to output the predicted mean and predicted variance of power, temperature and irradiance at each spatial location point, and the predicted variance is used as the model confidence level to label the three-dimensional geometric model.
7. The modeling method for multi-dimensional information fusion of the entire photovoltaic power plant area according to claim 1, characterized in that, The geospatial data includes digital elevation models, digital orthophoto models, and lidar point cloud data of the photovoltaic field; the meteorological environment data includes real-time collected irradiance, ambient temperature, wind speed, humidity, and air pressure; the electrical operation data includes the voltage, current, and power of each photovoltaic string, as well as the operating parameters of the combiner box and inverter; and the equipment physical state data includes the surface temperature of the photovoltaic modules, the degree of dust accumulation, hot spot detection results, and the tilt angle of the support.
8. The modeling method for multi-dimensional information fusion of the entire photovoltaic power plant area according to claim 1, characterized in that, The spatiotemporal registration in step S2 includes: synchronizing the time of multi-source data under the same timestamp and mapping all data to the same geographic coordinate system; the data cleaning includes missing value imputation, outlier removal and filtering smoothing.
9. The modeling method for multi-dimensional information fusion of the entire photovoltaic power plant area according to claim 1, characterized in that, The three-dimensional geometric model includes the topography of the photovoltaic field, the layout of the photovoltaic array, and the three-dimensional structure of buildings and roads; the dynamic operation status model includes the real-time power generation distribution, temperature distribution, irradiance distribution of each photovoltaic string, and the operating conditions of the inverter.
10. The modeling method for multi-dimensional information fusion of the entire photovoltaic power plant area according to claim 1, characterized in that, The multi-dimensional information fusion digital model is used for equipment anomaly location, power prediction, or component cleanliness assessment in photovoltaic fields.
