A deep learning-based pvt photovoltaic photothermal array output power dynamic prediction system

Abstract

This invention discloses a deep learning-based dynamic prediction system for the output power of a PVT photovoltaic-thermal array, comprising: a state processing module for collecting operational state data of the PVT photovoltaic-thermal array; a coupling construction module for generating array coupling state feature tensors; a spatial mapping module for generating array spatial correlation feature tensors; a dual-graph prediction module for inputting data into an improved MTGNN model, introducing a dual-graph parallel learning mechanism to generate electrical correlation graphs and thermal correlation graphs; performing multivariate time-series modeling to generate an output power prediction sequence; and a constraint output module for introducing array operational state reachability constraints to generate dynamic prediction results for output power. This invention achieves high-precision, strong consistency, and engineering reachability dynamic prediction of the output power of a PVT photovoltaic-thermal array under complex operating conditions.

Description

Technical Field

[0001] This invention relates to the field of new energy power generation and intelligent prediction technology, and in particular to a dynamic prediction system for the output power of a PVT photovoltaic thermal array based on deep learning. Background Technology

[0002] With the continuous expansion of the scale of new energy system integration and the increasing demand for refined dispatch of distributed energy, dynamic prediction technology for the output power of PVT photovoltaic thermal arrays has gradually become a research hotspot in the field of photovoltaic power generation and thermal energy integrated utilization. In actual operation, PVT photovoltaic thermal arrays are simultaneously affected by multiple factors such as irradiance, ambient temperature, wind speed, and array structure layout, and their electrical and thermal power outputs exhibit significant nonlinear, multivariate strong coupling, and time-series fluctuation characteristics.

[0003] Existing output power prediction methods are mostly based on single power types or simplified statistical models, making it difficult to characterize electrothermal coupling effects and spatial relationships between array units, thus limiting prediction accuracy under complex operating conditions. Furthermore, most methods neglect the constraints of array structure information on power evolution during modeling, failing to fully reflect the mutual influences of adjacent array units in terms of heat conduction and electrical connections, leading to discontinuities or abrupt jumps in prediction results in local areas. Simultaneously, existing deep learning prediction models typically focus on time-dimensional modeling, insufficiently considering the physical reachability constraints of the prediction results, easily generating power predictions that exceed actual operating boundaries, affecting system scheduling and operational safety.

[0004] Therefore, how to provide a dynamic prediction system for the output power of PVT photovoltaic-thermal arrays based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based dynamic prediction system for the output power of PVT photovoltaic-thermal arrays. This invention constructs an operating state input tensor, an array coupling state feature tensor, and an array spatial correlation feature tensor based on multi-source operating state data of the PVT photovoltaic-thermal array. An improved MTGNN model with a dual-graph parallel learning mechanism is introduced to achieve collaborative modeling of electrical and thermal correlation graphs. Simultaneously, the output power prediction results are constrained by array operating state reachability constraints, thereby completing the dynamic prediction of the PVT photovoltaic-thermal array output power. This invention can simultaneously characterize the coupling relationship between electrical and thermal power, the spatial correlation relationship between array units, and the dynamic evolution characteristics in the time dimension. It possesses advantages such as high prediction accuracy, strong adaptability to changes in operating conditions, good physical consistency of prediction results, and high feasibility for engineering applications.

[0006] A deep learning-based dynamic prediction system for the output power of a PVT photovoltaic-thermal array, according to an embodiment of the present invention, includes:

[0007] The status processing module is used to collect the operating status data of the PVT photovoltaic thermal array, perform time alignment and structured processing, and generate the array operating status input tensor.

[0008] The coupling construction module is used to construct array coupling state features based on the array operating state input tensor and generate an array coupling state feature tensor.

[0009] The spatial mapping module is used to construct array spatial relationships based on array structure information in the running status data, map the array coupling state feature tensor, and generate array spatial relationship feature tensor.

[0010] The dual-graph prediction module is used to input the array operating state input tensor, the array coupling state feature tensor, and the array spatial correlation feature tensor into the improved MTGNN model. The improved MTGNN model introduces a dual-graph parallel learning mechanism to generate an electrical correlation graph and a thermal correlation graph. Based on the electrical correlation graph and the thermal correlation graph, multivariate time series modeling is performed to generate an output power prediction sequence.

[0011] The constraint output module is used to introduce array operating state reachability constraints based on the output power prediction sequence, perform reachability constraint processing, and generate dynamic output power prediction results.

[0012] Optionally, modules can be integrated using the following methods:

[0013] S1. Collect the operating status data of the PVT photovoltaic thermal array, perform time alignment and structured processing, and generate the array operating status input tensor;

[0014] S2. Construct array coupling state features based on the array operating state input tensor, and generate array coupling state feature tensor;

[0015] S3. Construct array spatial correlation based on array structure information in the running status data, map array coupling state feature tensor, and generate array spatial correlation feature tensor;

[0016] S4. Input the array operating state input tensor, the array coupling state feature tensor, and the array spatial correlation feature tensor into the improved MTGNN model. The improved MTGNN model introduces a dual-graph parallel learning mechanism to generate an electrical correlation graph and a thermal correlation graph.

[0017] S5. Perform multivariate time series modeling based on the electrical correlation diagram and the thermal correlation diagram to generate an output power prediction sequence;

[0018] S6. Based on the output power prediction sequence, an array operating state reachability constraint is introduced, reachability constraint processing is performed, and a dynamic output power prediction result is generated.

[0019] Optionally, S1 specifically includes:

[0020] The operation status data of the PVT photovoltaic thermal array is collected, including electrical power data, thermal power data, temperature data, irradiance data, ambient temperature data, wind speed data, and array structure information.

[0021] The time alignment process is performed on the running status data. The time alignment process includes setting a unified time base and a unified sampling period, and aligning the timestamps of the running status data to generate aligned running status data.

[0022] The aligned running status data is subjected to structured processing, which includes determining the data field set and field order, and organizing the data according to the array cell identifier and time index to generate the array running status input tensor.

[0023] The tensor element corresponding to each time index in the array operation status input tensor is composed of the splicing result of electrical power data, thermal power data, temperature data, irradiance data, ambient temperature data, wind speed data and array structure information.

[0024] The splicing result is formed by sequentially splicing the data field values ​​corresponding to the same array unit identifier and the same time index to form a feature vector. The array running state input tensor is obtained by stacking the feature vectors corresponding to the time index in chronological order.

[0025] Optionally, S2 specifically includes:

[0026] Based on the array operating status input tensor, feature sequences corresponding to electric power, thermal power, temperature, irradiance, ambient temperature and wind speed are extracted in a manner that the array unit identifier and time index are consistent.

[0027] Under the same array cell identifier and the same time index, coupling operations are performed on the power characteristics and temperature characteristics. Under adjacent time indices, change operations are performed on the power characteristics and temperature characteristics to generate electrothermal coupling characteristics.

[0028] Under the same array cell identifier and the same time index, environmental coupling operations are performed on irradiance features, ambient temperature features and wind speed features to generate environmental coupling features;

[0029] The running status features, electrothermal coupling features and environmental coupling features are concatenated according to the preset field order to form an array coupling status feature vector;

[0030] Stack the array coupling state feature vectors corresponding to different time indices in chronological order to generate the array coupling state feature tensor.

[0031] Optionally, the step of constructing the array spatial association relationship based on the array structure information in the operating status data specifically involves:

[0032] Extract array cell identifiers, array cell arrangement order, and physical connection information between array cells from the operational status data;

[0033] The relative spatial positions of array elements in the array are determined based on the arrangement order of array elements, and the direct connection relationships between array elements are determined based on physical connection information.

[0034] An array cell association matrix is ​​constructed according to the array cell identifier. The values ​​of the matrix elements in the array cell association matrix are determined by whether there is a physical connection and spatial adjacency between the corresponding array cells.

[0035] Based on the array cell association matrix, a structured representation is generated to describe the spatial association between array cells.

[0036] Optionally, the mapping of the array coupling state feature tensor to generate the array spatial correlation feature tensor specifically involves:

[0037] Based on the array cell correlation matrix, spatial correlation mapping is performed on the array cell feature vectors under the same time index in the array coupling state feature tensor.

[0038] The spatial association mapping includes performing a weighted combination operation on the feature vectors of mutually associated array units according to the weight distribution of the corresponding rows in the array unit association matrix;

[0039] The feature results obtained by weighted combination operation are rearranged according to array cell identifier and time index to generate array space correlation feature tensor.

[0040] Optionally, the improved MTGNN model includes an input encoding structure, a dual-graph learning structure, a graph convolution update structure, a temporal modeling structure, and a prediction output structure, specifically:

[0041] The input encoding structure performs input concatenation and linear transformation on the array running state input tensor, the array coupling state feature tensor, and the array spatial correlation feature tensor to generate an array input feature representation.

[0042] A dual-graph parallel learning mechanism is introduced into the dual-graph learning structure. A set of node embedding vectors is constructed based on the array input feature representation. Similarity calculation and normalization processing are performed on the set of node embedding vectors to generate an electrical correlation matrix and a thermal correlation matrix. The similarity calculation includes performing an inner product operation on the node embedding vector pairs to obtain a similarity value. The normalization processing includes performing an exponential processing on the similarity value and normalizing it by row to obtain a weight distribution. An electrical correlation graph is generated based on the electrical correlation matrix, and a thermal correlation graph is generated based on the thermal correlation matrix.

[0043] In the graph convolutional update structure, neighborhood weighted aggregation is performed on the array input feature representation based on the electrical correlation graph to obtain the electrical graph feature representation, and neighborhood weighted aggregation is performed on the array input feature representation based on the thermal correlation graph to obtain the thermal graph feature representation. The neighborhood weighted aggregation operation includes performing a weighted summation operation on the feature vectors of the associated nodes according to the weight distribution; feature concatenation and linear transformation are performed on the electrical graph feature representation and the thermal graph feature representation to generate a fused graph feature representation.

[0044] In the time series modeling structure, multivariate time series modeling operations are performed based on the feature representation of the fusion graph to output the power prediction sequence. The multivariate time series modeling operations include performing sequence convolution operations on the time index dimension and performing gated combination operations on the sequence convolution output.

[0045] In the predicted output structure, a linear mapping is performed on the output power prediction sequence to generate a sequence of predicted values ​​for the output power prediction sequence.

[0046] Optionally, S4 specifically includes:

[0047] An input concatenation operation is performed on the array running state input tensor, the array coupling state feature tensor, and the array spatial association feature tensor. The input concatenation operation includes concatenating the three types of tensors in the feature dimension in a manner consistent with the array unit identifier and time index to generate an array input feature representation.

[0048] A set of node embedding vectors is generated based on the array input feature representation, and each node embedding vector in the set corresponds to an array cell identifier;

[0049] Perform similarity calculation on the set of node embedding vectors, wherein the similarity calculation includes performing an inner product operation on any two node embedding vectors to obtain a similarity value;

[0050] The similarity values ​​are indexed and then normalized by row to generate a weight distribution matrix.

[0051] The electrical correlation matrix and the thermal correlation matrix are generated based on the weight distribution matrix. The electrical correlation matrix and the thermal correlation matrix are two sets of independent weight distribution matrices.

[0052] An electrical correlation graph is generated based on the electrical correlation matrix, which consists of array cell identifiers and the electrical correlation matrix; a thermal correlation graph is generated based on the thermal correlation matrix, which consists of array cell identifiers and the thermal correlation matrix.

[0053] Optionally, S5 specifically includes:

[0054] Based on the electrical correlation graph and the thermal correlation graph, a neighborhood weighted aggregation operation is performed on the array input feature representation to obtain the electrical graph feature representation and the thermal graph feature representation.

[0055] Perform feature concatenation and linear transformation on the electrical map feature representation and the thermal map feature representation to generate a fused map feature representation;

[0056] Multivariate temporal modeling operations are performed based on the feature representation of the fused graph. The multivariate temporal modeling operations include performing sequence convolution operations and gating combination operations in the time index dimension to generate output power prediction sequences.

[0057] The sequence convolution operation is to perform a convolution summation operation on the fusion map feature representation and the convolution kernel weight in the time index dimension. The fusion map feature value and the convolution kernel weight value within the preset convolution window are multiplied and accumulated to obtain the convolution output value.

[0058] The gated combination operation is to calculate the gate weight value of the convolution output value obtained by the sequence convolution operation, and perform a weighted combination on the convolution output value based on the gate weight value. The gate weight value is obtained by the convolution output value through compression mapping and normalization mapping, and the weighted combination is obtained by multiplying the gate weight value and the convolution output value one by one and accumulating them.

[0059] Optionally, S6 specifically includes:

[0060] The array operation state reachability constraint set is calculated based on the output power prediction sequence and operation state data. The array operation state reachability constraint set includes power change rate constraint and power value range constraint.

[0061] The power change rate constraint is obtained by the difference between the output power prediction values ​​under adjacent time indices. The difference value is obtained by subtracting the output power prediction value corresponding to the previous time index from the output power prediction value corresponding to the later time index, and the power change rate value is obtained by dividing the difference value by the uniform sampling period.

[0062] The power value range constraint is obtained by comparing the predicted output power value with the range boundary value of the operating status data. The range boundary value is obtained by linearly mapping the irradiance value, ambient temperature value, and wind speed value in the operating status data. The irradiance value, ambient temperature value, and wind speed value are multiplied by a preset coefficient and accumulated to obtain the range boundary value.

[0063] Based on the array operating state reachability constraint set, reachability constraint processing is performed on the output power prediction sequence. The reachability constraint processing includes rate pruning for the output power prediction value corresponding to the time index where the power change rate exceeds the rate threshold, and interval pruning for the output power prediction value corresponding to the time index where the output power prediction value exceeds the interval boundary value, thereby generating dynamic output power prediction results.

[0064] The beneficial effects of this invention are:

[0065] (1) By constructing the array coupling state feature tensor, the coupling relationship between electric power, thermal power, temperature and environmental factors is characterized under the same array unit identifier and time index, and the operating state change characteristics are characterized under adjacent time indexes. This enhances the ability to express the electrothermal synergistic output characteristics of PVT photovoltaic solar thermal array and avoids the problem that the single power modeling method is insufficient to characterize the electrothermal coupling mechanism.

[0066] (2) By introducing array structure information to construct array unit correlation matrix and mapping array coupling state feature tensor to spatial correlation feature tensor, explicit modeling of spatial adjacency and physical connection relationship between array units is realized, which effectively improves the perception of the influence of power transfer and heat diffusion inside the array and overcomes the prediction bias caused by the traditional time series model ignoring the differences in array spatial structure.

[0067] (3) By introducing a dual-graph parallel learning mechanism in the improved MTGNN model, an electrical correlation graph and a thermal correlation graph are constructed respectively and multivariate time series modeling is performed to realize the synchronous learning of the dynamic correlation relationship of electrical power and the dynamic correlation relationship of thermal power. This avoids the problem that a single correlation graph cannot simultaneously characterize the heterogeneous correlation features of electrical and thermal power, and significantly improves the accuracy and stability of the output power prediction sequence.

[0068] (4) By introducing array operation state reachability constraints, power change rate constraints and power value range constraints are applied to the output power prediction sequence, which effectively suppresses abrupt changes and out-of-bounds values ​​in the prediction results that do not conform to the actual operating conditions, and enhances the physical consistency and executability of the dynamic prediction results of output power under engineering operating conditions. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0070] Figure 1 This is a schematic diagram of a deep learning-based dynamic prediction system for the output power of a PVT photovoltaic-thermal array proposed in this invention.

[0071] Figure 2This is a flowchart of a deep learning-based dynamic prediction system for the output power of a PVT photovoltaic-thermal array proposed in this invention.

[0072] Figure 3 This is a data flow diagram of a deep learning-based dynamic prediction system for the output power of a PVT photovoltaic-thermal array proposed in this invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0074] refer to Figure 1 A deep learning-based dynamic prediction system for the output power of a PVT photovoltaic-thermal array includes:

[0075] The status processing module is used to collect the operating status data of the PVT photovoltaic thermal array, perform time alignment and structured processing, and generate the array operating status input tensor.

[0076] The coupling construction module is used to construct array coupling state features based on the array operating state input tensor and generate an array coupling state feature tensor.

[0077] The spatial mapping module is used to construct array spatial relationships based on array structure information in the running status data, map the array coupling state feature tensor, and generate array spatial relationship feature tensor.

[0078] The dual-graph prediction module is used to input the array operating state input tensor, the array coupling state feature tensor, and the array spatial correlation feature tensor into the improved MTGNN model. The improved MTGNN model introduces a dual-graph parallel learning mechanism to generate an electrical correlation graph and a thermal correlation graph. Based on the electrical correlation graph and the thermal correlation graph, multivariate time series modeling is performed to generate an output power prediction sequence.

[0079] The constraint output module is used to introduce array operating state reachability constraints based on the output power prediction sequence, perform reachability constraint processing, and generate dynamic output power prediction results.

[0080] In this embodiment, the state processing module performs time alignment and structuring processing on the operating state data of the PVT photovoltaic-thermal array under a unified time reference, forming an array operating state input tensor. This ensures that different array units maintain consistency in both the time dimension and data structure dimension, further reducing the impact of differences in time scale and field organization of multi-source operating state data on the subsequent modeling process. The coupling construction module constructs an array coupling state feature tensor based on the array operating state input tensor. By uniformly expressing the coupling relationship between electrical power, thermal power, and environmentally related state features, it further enhances the ability to characterize the array operating state in the feature space. The spatial mapping module constructs array spatial relationships based on the array structure information in the operating state data. The array coupling state feature tensor is mapped to generate an array spatial correlation feature tensor, thereby characterizing the physical connection and spatial adjacency relationships between array units at the feature level. The dual-graph prediction module inputs the array operating state input tensor, the array coupling state feature tensor, and the array spatial correlation feature tensor into the improved MTGNN model. By introducing a dual-graph parallel learning mechanism, an electrical correlation graph and a thermal correlation graph are constructed, and multivariate time series modeling is performed based on the correlation graphs to further improve the ability of the output power prediction sequence to represent changes in the array operating state. The constraint output module introduces array operating state reachability constraints on the output power prediction sequence and performs constraint processing on the prediction results to ensure that the dynamic prediction results of output power remain consistent in terms of temporal continuity and state rationality.

[0081] refer to Figure 2 and Figure 3 In this embodiment, the modules are interconnected through the following method:

[0082] S1. Collect the operating status data of the PVT photovoltaic thermal array, perform time alignment and structured processing, and generate the array operating status input tensor;

[0083] S2. Construct array coupling state features based on the array operating state input tensor, and generate array coupling state feature tensor;

[0084] S3. Construct array spatial correlation based on array structure information in the running status data, map array coupling state feature tensor, and generate array spatial correlation feature tensor;

[0085] S4. Input the array operating state input tensor, the array coupling state feature tensor, and the array spatial correlation feature tensor into the improved MTGNN model. The improved MTGNN model introduces a dual-graph parallel learning mechanism to generate an electrical correlation graph and a thermal correlation graph.

[0086] S5. Perform multivariate time series modeling based on the electrical correlation diagram and the thermal correlation diagram to generate an output power prediction sequence;

[0087] S6. Based on the output power prediction sequence, an array operating state reachability constraint is introduced, reachability constraint processing is performed, and a dynamic output power prediction result is generated.

[0088] In this embodiment, S1 specifically refers to:

[0089] The system collects operational status data of the PVT photovoltaic thermal array. This data is acquired synchronously by the power sensor, thermal power sensor, temperature sensor, irradiance sensor, ambient temperature sensor, and wind speed sensor configured in each unit of the array. At the same time, the system reads the array structure information, which includes the array unit identifier, the array unit arrangement order, and the physical connection relationship between the array units.

[0090] Time alignment processing is performed on the running status data. Time alignment processing is completed by setting a unified time base and a unified sampling period. The unified time base is the system timestamp, and the unified sampling period is preferably set to 60 seconds. Interpolation or resampling operations are performed on the timestamps in the collected data that deviate from the unified sampling period, so that all running status data form an equally spaced sequence in the time dimension, generating aligned running status data.

[0091] The aligned operation status data is processed in a structured manner. The structured processing is completed by determining a fixed set of data fields and the order of the fields. The set of fields includes electrical power, thermal power, temperature, irradiance, ambient temperature, wind speed and array structure information. The data is organized in two dimensions according to the array unit identifier and time index. The time index corresponds to the sequence position under a unified sampling period, and the array operation status input tensor is generated.

[0092] The tensor element corresponding to each time index in the array operation status input tensor is formed by concatenating the electric power data, thermal power data, temperature data, irradiance data, ambient temperature data, wind speed data and array structure information under the same array unit identifier in the order of fields. The concatenation process is to arrange the values ​​corresponding to each field in sequence to form a one-dimensional feature vector.

[0093] The array running state input tensor is obtained by stacking the feature vectors corresponding to each time index in chronological order, with the stacking direction corresponding to the increasing direction of the time index, thus forming a three-dimensional input tensor that simultaneously contains the time dimension, the array unit dimension, and the feature dimension.

[0094] In this embodiment, S2 specifically refers to:

[0095] Based on the array operating status input tensor, and in accordance with the organization method of array unit identifier and time index being consistent, the electric power feature sequence, thermal power feature sequence, temperature feature sequence, irradiance feature sequence, ambient temperature feature sequence and wind speed feature sequence are separated from the tensor. Each feature sequence maintains the order consistency with the unified sampling period in the time dimension.

[0096] Under the same array cell identifier and the same time index, coupling operations are performed on the power characteristic value and the temperature characteristic value. The coupling operation is obtained by multiplying the two values ​​one by one and combining them with a linear weighting method. At the same time, change operations are performed on the power characteristic value and the temperature characteristic value under adjacent time indices. The change operation is obtained by subtracting the current time index value from the previous time index value. The electrothermal coupling feature is formed by splicing the coupling value under the same time index and the change amplitude under adjacent time indices.

[0097] Under the same array unit identifier and the same time index, environmental coupling operation is performed on the irradiance feature value, ambient temperature feature value and wind speed feature value. The environmental coupling operation obtains the environmental coupling value by weighting and summing the three values ​​according to preset weights. Preferably, the irradiance weight is 0.5, the ambient temperature weight is 0.3 and the wind speed weight is 0.2.

[0098] The operating status features, electrothermal coupling features, and environmental coupling features are concatenated according to the preset field order. The concatenation process involves arranging the corresponding feature values ​​under the same array unit identifier and the same time index in sequence to form a one-dimensional array coupling state feature vector. The array coupling state feature vector contains both the original operating status information and the coupling derived information.

[0099] The array coupling state feature vectors corresponding to different time indices are stacked in ascending order of time index. The stacking result forms a continuous sequence in the time dimension, thereby generating the array coupling state feature tensor.

[0100] In this embodiment, the construction of array spatial association based on array structure information in the operating status data specifically includes:

[0101] The array structure information is read from the operating status data. The array unit identifier, array unit arrangement order and physical connection information between array units are parsed. The array unit identifier is used to uniquely identify each unit in the array. The array unit arrangement order is used to describe the arrangement position of the array units in the row and column directions. The physical connection information is used to describe whether there are direct pipe, electrical or heat conduction connections between array units.

[0102] Based on the arrangement order of the array units, the relative spatial position relationship between the array units is calculated according to the row index and column index of the array units in the array. The relative spatial position relationship is determined by judging the adjacency of the array units in the row direction or column direction. Preferably, the adjacency judgment threshold is set to a row index difference of no more than 1 and a column index difference of no more than 1.

[0103] At the same time, the direct connection relationship between array units is determined based on the physical connection information. The direct connection relationship is determined by judging whether there is a connection record for the corresponding array unit pair in the array structure information. If a connection record exists, it is marked as a connected state; if no connection record exists, it is marked as a non-connected state.

[0104] The array unit association matrix is ​​constructed according to the array unit identifier order. The array unit association matrix is ​​a two-dimensional matrix structure, with the matrix rows and columns corresponding to the array unit identifiers. The value of any element in the matrix is ​​determined by the spatial adjacency relationship and the direct connection relationship between the corresponding array units. When the spatial adjacency relationship or the direct connection relationship is established, the matrix element value is set to 1. When neither the spatial adjacency relationship nor the direct connection relationship is established, the matrix element value is set to 0.

[0105] Based on the array cell association matrix, a structured representation is generated to describe the spatial association between array cells. The structured representation stores the association state between array cells in matrix form, which is used to constrain the propagation range of features in the array space dimension during the array coupling state feature mapping process, so that the array space association is clearly expressed at the feature level.

[0106] In this embodiment, the mapping of the array coupling state feature tensor to generate the array spatial correlation feature tensor specifically involves:

[0107] Based on the array unit correlation matrix, spatial correlation mapping is performed on the array unit feature vectors corresponding to the same time index in the array coupling state feature tensor. The spatial correlation mapping uses the correlation relationship recorded in the array unit correlation matrix as the mapping constraint, so that the array unit features only combine information between units that have spatial adjacency or physical connection.

[0108] According to the correlation weight distribution corresponding to each row in the array unit correlation matrix, the feature vectors of the array units that are correlated with the target array unit are subjected to weighted combination operation. The weighted combination operation is obtained by multiplying the value of each dimension of the feature vector of the correlated array unit with the corresponding weight value one by one and accumulating them. The weight value is determined by the value of the corresponding element in the array unit correlation matrix. When the matrix element value is 1, it participates in the combination operation; when the matrix element value is 0, it does not participate in the combination operation.

[0109] Preferably, in order to avoid excessively large combined feature amplitudes during spatial association mapping, the weight values ​​participating in the combined operation are normalized. The normalization method is to use the sum of all weight values ​​in the same row as the normalization denominator, so that the sum of the normalized weight values ​​is 1.

[0110] The feature result obtained by weighted combination operation is used as the spatial correlation feature representation of the array unit under the corresponding time index. The spatial correlation feature representation is consistent with the feature vector of the array coupling state in the feature dimension.

[0111] The spatial association feature representations are rearranged according to the array unit identifier and time index. The spatial association feature representations corresponding to all array units under each time index are stacked in the ascending direction of the time index to generate the array spatial association feature tensor, so that the array coupling state features have the ability to express association constraints in both the time and space dimensions.

[0112] In this embodiment, the improved MTGNN model includes an input encoding structure, a dual-graph learning structure, a graph convolution update structure, a temporal modeling structure, and a prediction output structure, specifically:

[0113] The input coding structure receives the array operating state input tensor, the array coupling state feature tensor, and the array spatial association feature tensor. It concatenates the three types of tensors along the feature dimensions to form a joint feature tensor and applies a linear transformation to the joint feature tensor. The linear transformation maps the values ​​of each feature dimension of the joint feature tensor to the weight matrix and the bias vector to generate the array input feature representation. The array input feature representation is consistent with the input tensor in terms of time index and array cell identifier dimensions.

[0114] The dual-graph learning structure constructs a set of node embedding vectors based on the array input feature representation. Each vector in the set corresponds to an array cell identifier. The node embedding vectors are obtained by linearly projecting the array input feature representation onto the feature dimension. Similarity calculation is performed on any two node embedding vectors in the set. The similarity value is obtained by multiplying the values ​​of each dimension of the two node embedding vectors and summing them. The similarity value is then exponentially processed and normalized row-wise along the node dimension to obtain a weight distribution. The normalization method is to use the sum of the exponentially similarity values ​​in the same row as the denominator, making the sum of the weight distributions equal to 1. Electrical and thermal correlation matrices are generated based on two independently initialized linear mapping parameters, each corresponding to a different weight distribution. An electrical correlation graph is constructed based on the electrical correlation matrix, and a thermal correlation graph is constructed based on the thermal correlation matrix. The correlation graph is jointly represented by the array cell identifier set and the corresponding correlation matrix.

[0115] The graph convolutional update structure performs neighborhood weighted aggregation on the array input feature representation based on the electrical correlation graph to obtain the electrical graph feature representation. The neighborhood weighted aggregation operation is obtained by multiplying and accumulating the feature vectors of array cells that are related to the target array cells according to the weight distribution in the electrical correlation graph. At the same time, a heatmap feature representation is obtained based on the thermal correlation graph in the same way. The electrical graph feature representation and the heatmap feature representation are concatenated in the feature dimension, and a linear transformation is applied to the concatenation result to generate a fused graph feature representation.

[0116] The temporal modeling structure performs multivariate temporal modeling operations on the time index dimension based on the fused graph feature representation. These operations include convolution summation on the fused graph feature representation and convolution kernel weights on the time index dimension. The convolution summation is achieved by multiplying and summing the feature values ​​within a preset convolution window with the convolution kernel weight values, obtaining the convolution output value. Simultaneously, gate weights are calculated on the convolution output values. These gate weights are obtained by linearly compressing and normalizing the convolution output values; preferably, the normalization method restricts the gate weight values ​​to between 0 and 1. A weighted combination of the convolution output values ​​is then performed based on the gate weights to generate an output power prediction sequence.

[0117] The predicted output structure applies a linear mapping to the output power prediction sequence, mapping the feature values ​​obtained from time series modeling to the output power prediction value sequence under the corresponding time index, so that the prediction result is consistent with the input running state sequence in the time dimension.

[0118] In this embodiment, the improved MTGNNGT model is based on the MTGNN model. It is derived from the fact that PVT photovoltaic thermal arrays have two types of output characteristics, namely electrical power and thermal power, during operation. Moreover, the two types of outputs have significant differences in spatial structure correlation and temporal evolution. The existing MTGNN model uses a single correlation graph to uniformly model the temporal relationship of multiple variables. It is difficult to simultaneously characterize the differentiated correlation characteristics of array units on the power transmission path and heat conduction path, and it is also difficult to constrain the prediction results to meet the requirements of the actual operating state of the array in terms of temporal continuity and physical accessibility. To address the aforementioned technical challenges, the improved MTGNNGT model, while retaining the multivariate graph time-series modeling framework of the MTGNN model, introduces a dual-graph parallel learning structure. This decomposes the relationships between array units into two structural representations: electrical correlation graphs and thermal correlation graphs. This allows electrical and thermal power features to aggregate neighborhood information under the constraints of their respective correlation graphs, thus avoiding interference between relationships under different physical mechanisms. Simultaneously, an array operating state reachability constraint is introduced in the model output stage, limiting the output power prediction results to the reachable range of the array operating state change rate and power value interval, ensuring consistency between the model prediction results over time and at the physical operation level. Without altering the core structure of the original MTGNN time-series modeling, the improved MTGNNGT model achieves differentiated modeling and joint prediction of the electrothermal dual-channel structural features of PVT photovoltaic-thermal arrays, forming a technical solution that combines structural correlation modeling with a prediction result constraint mechanism.

[0119] In this embodiment, S4 specifically refers to:

[0120] The array operation state input tensor, array coupling state feature tensor, and array spatial association feature tensor are jointly processed. The joint processing is performed by concatenating the three types of tensors in the feature dimension while keeping the one-to-one correspondence between array unit identifiers and time indices unchanged, to obtain the array input feature representation. The array input feature representation contains operation state information, coupling state information, and spatial association information at each time index.

[0121] A set of node embedding vectors is constructed based on the array input feature representation. The set of node embedding vectors is obtained by linearly projecting the array input feature representation onto the feature dimension. The linear projection process is to multiply each feature value of the array input feature representation with the projection weight value one by one and then sum them to form a new feature value. Each node embedding vector in the set of node embedding vectors is uniquely associated with an array cell identifier.

[0122] The similarity of each pair of node embedding vectors in the node embedding vector set is calculated. The similarity is obtained by multiplying the values ​​of the corresponding dimensions of the two node embedding vectors one by one and summing them. The similarity value is used to characterize the association strength of the array units in the feature space.

[0123] The similarity values ​​are indexed and normalized row by row in the node dimension to generate a weight distribution matrix. The indexation process is to take the natural exponent of the similarity values ​​to enhance the difference in similarity. The row normalization process is to use the sum of the indexed similarity values ​​in the same row as the normalization denominator so that the sum of the weight distributions in each row is 1.

[0124] Based on the weight distribution matrix, electrical correlation matrix and thermal correlation matrix are constructed respectively. The electrical correlation matrix and thermal correlation matrix are obtained by mapping the set of node embedding vectors using different linear projection parameters, so that the two sets of correlation matrices express different emphasis on the correlation relationship while having the same structural form.

[0125] An electrical correlation graph is generated based on the electrical correlation matrix. The electrical correlation graph is composed of the array cell identifier set and the electrical correlation matrix, and is used to represent the electrical correlation relationship between array cells. At the same time, a thermal correlation graph is generated based on the thermal correlation matrix. The thermal correlation graph is composed of the array cell identifier set and the thermal correlation matrix, and is used to represent the thermal correlation relationship between array cells. Thus, two parallel and independent array correlation representations are formed in the same modeling process, providing structured correlation input for multivariate time series modeling.

[0126] In this embodiment, S5 specifically refers to:

[0127] Based on electrical correlation graphs and thermal correlation graphs, the array input feature representation is subjected to structure-guided processing. According to the correlation weight distribution between array units in the electrical correlation graph, the array input feature values ​​that have an electrical correlation relationship with the target array unit are weighted and accumulated item by item to obtain the electrical graph feature representation corresponding to each array unit. At the same time, according to the correlation weight distribution between array units in the thermal correlation graph, the array input feature values ​​that have a thermal correlation relationship with the target array unit are weighted and accumulated item by item to obtain the thermal graph feature representation corresponding to each array unit. In the neighborhood weighted aggregation operation, the feature values ​​of the array unit itself participate in the weighted combination.

[0128] The electrical map feature representation and the heat map feature representation are concatenated along the feature dimension to form a joint feature vector. A linear transformation is then applied to the joint feature vector. The linear transformation maps the values ​​of each dimension of the joint feature vector linearly through the weight matrix and the bias vector to generate a fused graph feature representation. The fused graph feature representation remains consistent with the array input feature representation in the array cell dimension and the time index dimension.

[0129] Multivariate temporal modeling is performed on the time index dimension based on the fusion graph feature representation. The multivariate temporal modeling is completed by introducing a convolution window of fixed length in the time index dimension. Preferably, the length of the convolution window is set to 3. The fusion graph feature value at each time index in the convolution window is multiplied by the corresponding convolution kernel weight value and accumulated to obtain the convolution output value under the corresponding time index. The convolution output value is used to characterize the local change trend of the fusion graph feature in the time dimension.

[0130] The gating weights are calculated for the convolution output values. The gating weights are obtained by linearly compressing the convolution output values ​​and applying a normalization mapping. Preferably, the normalization mapping method is to restrict the gating weights to between 0 and 1, so that the gating weights reflect the importance of feature information under different time indices.

[0131] The convolution output values ​​are weighted and combined based on the gating weight values. The weighted combination process is obtained by multiplying the convolution output values ​​with the corresponding gating weight values ​​one by one and accumulating them to form an output power prediction sequence. The output power prediction sequence corresponds one-to-one with the input running state sequence in the time index dimension, and is used to characterize the dynamic change trend of the array output power in the time dimension.

[0132] In this embodiment, S6 specifically refers to:

[0133] Based on the output power prediction sequence and the operating status data, an array operating status reachability constraint set is constructed. The array operating status reachability constraint set is used to limit the range of change of the output power prediction sequence in terms of time continuity and physical feasibility. The array operating status reachability constraint set is composed of power change rate constraint and power value range constraint.

[0134] The power change rate constraint is obtained by differential calculation of the output power prediction sequence under adjacent time indices. The differential calculation method is to subtract the output power prediction value corresponding to the previous time index from the output power prediction value corresponding to the subsequent time index to obtain the difference value, and divide the difference value by the uniform sampling period to obtain the power change rate value. The uniform sampling period is consistent with the sampling period set in S1. Preferably, the uniform sampling period is 60 seconds. The power change rate threshold is set according to the actual operating capacity of the array. Preferably, the power change rate threshold is 0.05 per minute of the rated power.

[0135] The power value range constraint is obtained by comparing the predicted output power value with the range boundary value obtained by mapping the operating status data. The range boundary value is calculated by linearly mapping the irradiance value, ambient temperature value, and wind speed value in the operating status data. The linear mapping method is to multiply the irradiance value by a coefficient of 0.6, the ambient temperature value by a coefficient of 0.3, and the wind speed value by a coefficient of 0.1, and then sum the three results to form the range boundary reference value. At the same time, the range boundary reference value is scaled by combining the array's rated output power to obtain the upper and lower bound values ​​allowed for the output power prediction.

[0136] The output power prediction sequence is processed based on the array operation state reachability constraint set. The reachability constraint processing includes rate pruning of the output power prediction value corresponding to the time index where the power change rate exceeds the power change rate threshold. The rate pruning method is to limit the power change rate within the threshold range and correct the corresponding output power prediction value accordingly. At the same time, interval pruning is performed on the time index where the output power prediction value exceeds the upper or lower bound of the power value interval constraint. The interval pruning method is to limit the output power prediction value within the corresponding interval boundary range. This generates a dynamic output power prediction result that meets the array operation state reachability constraint conditions, so that the prediction result is consistent and feasible at the time evolution and physical operation constraint levels.

[0137] Example 1:

[0138] To verify the feasibility of this invention in practice, it was applied to a rooftop PVT photovoltaic-thermal array at a comprehensive energy station in a coastal manufacturing park. Forty-eight PVT modules were installed on the park's roof, arranged in a 6-string, 8-parallel array. A heat exchange loop was arranged on the back of each module and connected to a 2-cubic-meter buffer water tank. The rated power of the grid-connected electrical side was 20kW, and the thermal side supplied mixed hot water for production and domestic use. Two prominent problems persisted during operation: First, the array was affected by sea breezes, rapid cloud cover, and ambient temperature fluctuations, resulting in significant non-stationary coupling changes between electrical and thermal power. Traditional predictions based on a single power curve or a single correlation increased in deviation during periods of cloud cover or sudden wind speed changes. Second, the predicted sequence exhibited jumps that did not conform to physical reachability, such as a sudden power surge exceeding the array's reachable limit within 10 minutes, leading to misjudgments in energy management strategies, resulting in inverter power limiting and frequent heat pump start-ups and shutdowns.

[0139] During on-site deployment, the state processing module receives data from inverter power sampling, flow meter and inlet / outlet water temperature sampling, irradiance meter, ambient temperature meter, and anemometer, while simultaneously writing array structure information. A unified sampling period of 10 minutes and a unified time base set to the server's NTP clock are used to complete time alignment and structured processing, forming the array operating state input tensor. The coupling construction module extracts feature sequences of power, heat, temperature, irradiance, ambient temperature, and wind speed from the input tensor, generating electrothermal coupling features and environmental coupling features, which are then concatenated with the operating state features to form the array coupling state feature tensor. The spatial mapping module constructs an array unit association matrix based on the array unit arrangement order and physical connection information, and performs spatial association mapping on the array coupling state feature tensor to obtain the array spatial association feature tensor. The dual-graph prediction module inputs the three types of tensors into the improved MTGNN model, simultaneously generating an electrical association graph and a thermal association graph under a dual-graph parallel learning mechanism, and completes multivariate time-series modeling to output a predicted power sequence. The constraint output module calculates the array's operational status reachability constraint set based on the output power prediction sequence and operational status data. The preferred rate threshold is 0.25 kW / min, and the preferred linear mapping coefficients for the interval boundaries are: irradiance coefficient 0.020 kW / (W / ㎡), ambient temperature coefficient -0.060 kW / ℃, and wind speed coefficient -0.120 kW / (m / s). An intercept term of 2.0 kW is set to complete rate clipping and interval clipping, and output the dynamic prediction results of the output power, which can be directly used by the park's energy management system for rolling access.

[0140] During a 30-day continuous operation and verification, four typical weather days were selected to compare the prediction results of the traditional single-map prediction model with those of the present invention. The prediction accuracy of output power and the number of constraint violations were statistically analyzed, and the results are shown in Table 1.

[0141] Table 1. Comparison of Dynamic Power Output Prediction Accuracy and Availability Violations on Typical Weather Days

[0142] Indicators / Weather Types Sunny steady state day Cloudy with shadows Windy and cool day intermittent light rain days Daily average irradiance (W / m²) 720 510 560 340 Average wind speed (m / s) 3.1 4.8 8.6 5.2 Traditional single-image model MAPE (%) 8.7 14.9 13.2 16.5 MAPE (%) of this invention 5.3 8.6 7.9 10.4 Traditional single-image model RMSE (kW) 1.92 3.08 2.74 3.41 The RMSE (kW) of this invention 1.21 1.87 1.69 2.12 Number of reachability violations in traditional single-graph models (times / day) 12 37 29 41 Number of reachability violations of this invention (times / day) 1 4 3 6

[0143] As shown in Table 1, this invention achieves lower prediction errors on average across four typical weather days, with the advantage being more pronounced on cloudy days with shadows and intermittent light rain. This is because the parallel learning of electrical and thermal correlation graphs separates the rapid response of the electrical side from the lag inertia of the thermal side, avoiding misinterpreting the gradual changes in the thermal loop as instantaneous correlations of electrical power, thus reducing error amplification under rapid changes in cloud cover. Simultaneously, the array operating state reachability constraints explicitly remove two common violations: "rate cannot suddenly increase" and "interval cannot exceed limits," reducing the number of violations from 29–41 per day to 3–6. This results in a smoother prediction sequence received by the energy management system, and the rolling scheduling no longer frequently triggers inverter power limiting and heat pump protection logic.

[0144] To further demonstrate the actual benefits to the park's operation, the invention was integrated into the park's energy management system for daily scheduling. The scheduling and energy consumption indicators were compared for 30 days before and after its implementation, and the results are shown in Table 2.

[0145] Table 2 Comparison of Operating Indicators of Energy Stations in the Industrial Park

[0146] Indicator Name Before enabling this invention After enabling this invention range of change Inverter power generation limit (kWh) 612 328 -46.4% Dispatch power (kWh) caused by power-side forecasting 487 255 -47.6% Average daily start-up and shutdown frequency of heat pump (times / day) 18.6 11.2 -39.8% Auxiliary electric heating for heat replenishment (kWh_th) 1380 1035 -25.0% Average daily temperature fluctuation in the water tank (°C) 6.4 4.1 -35.9% Demand response assessment deduction (RMB) 2140 980 -54.2%

[0147] Table 2 shows that the benefits of this invention are reflected not only in "more accurate predictions" but also in "more stable operation." After the predicted sequence is processed by reachability constraints, the scheduling strategy's misjudgment of power surges is significantly reduced, the inverter's power generation limitation decreases, the scheduling deviation caused by the electrical side prediction decreases synchronously, and the demand response penalty is reduced. On the thermal side, parallel learning of the electrical and thermal correlation graphs improves the availability of the trend of thermal power and water tank temperature, the energy management system's allocation of heat pump load is smoother, the average daily start-up and shutdown frequency of heat pumps decreases, the auxiliary electric heating supplementary heat is reduced, water tank temperature fluctuations converge, and the supply of hot water for production is more stable. In summary, this invention achieves improved accuracy and executability of dynamic output power prediction results under complex meteorological disturbances and electrothermal coupling conditions, solving the scheduling instability problem caused by large prediction deviations and numerous prediction violations.

[0148] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based dynamic prediction system for the output power of a PVT photovoltaic-thermal array, characterized in that, include: The status processing module is used to collect the operating status data of the PVT photovoltaic thermal array, perform time alignment and structured processing, and generate the array operating status input tensor. The coupling construction module is used to construct array coupling state features based on the array operating state input tensor and generate an array coupling state feature tensor. The spatial mapping module is used to construct array spatial relationships based on array structure information in the running status data, map the array coupling state feature tensor, and generate array spatial relationship feature tensor. The dual-graph prediction module is used to input the array operating state input tensor, the array coupling state feature tensor, and the array spatial correlation feature tensor into the improved MTGNN model. The improved MTGNN model introduces a dual-graph parallel learning mechanism to generate an electrical correlation graph and a thermal correlation graph. Multivariate time series modeling is performed based on the electrical correlation diagram and the thermal correlation diagram to generate an output power prediction sequence; The constraint output module is used to introduce array operating state reachability constraints based on the output power prediction sequence, perform reachability constraint processing, and generate dynamic output power prediction results.

2. The deep learning-based PVT photovoltaic-thermal array output power dynamic prediction system according to claim 1, characterized in that, Inter-module communication is achieved through the following methods: S1. Collect the operating status data of the PVT photovoltaic thermal array, perform time alignment and structured processing, and generate the array operating status input tensor; S2. Construct array coupling state features based on the array operating state input tensor, and generate array coupling state feature tensor; S3. Construct array spatial correlation based on array structure information in the running status data, map array coupling state feature tensor, and generate array spatial correlation feature tensor; S4. Input the array operating state input tensor, the array coupling state feature tensor, and the array spatial correlation feature tensor into the improved MTGNN model. The improved MTGNN model introduces a dual-graph parallel learning mechanism to generate an electrical correlation graph and a thermal correlation graph. S5. Perform multivariate time series modeling based on the electrical correlation diagram and the thermal correlation diagram to generate an output power prediction sequence; S6. Based on the output power prediction sequence, an array operating state reachability constraint is introduced, reachability constraint processing is performed, and a dynamic output power prediction result is generated.

3. The deep learning-based PVT photovoltaic-thermal array output power dynamic prediction system according to claim 2, characterized in that, Specifically, S1 is: The operation status data of the PVT photovoltaic thermal array is collected, including electrical power data, thermal power data, temperature data, irradiance data, ambient temperature data, wind speed data, and array structure information. The time alignment process is performed on the running status data. The time alignment process includes setting a unified time base and a unified sampling period, and aligning the timestamps of the running status data to generate aligned running status data. The aligned running status data is subjected to structured processing, which includes determining the data field set and field order, and organizing the data according to the array cell identifier and time index to generate the array running status input tensor. The tensor element corresponding to each time index in the array operation status input tensor is composed of the splicing result of electrical power data, thermal power data, temperature data, irradiance data, ambient temperature data, wind speed data and array structure information. The splicing result is formed by sequentially splicing the data field values ​​corresponding to the same array unit identifier and the same time index to form a feature vector. The array running state input tensor is obtained by stacking the feature vectors corresponding to the time index in chronological order.

4. The deep learning-based PVT photovoltaic-thermal array output power dynamic prediction system according to claim 2, characterized in that, Specifically, S2 is: Based on the array operating status input tensor, feature sequences corresponding to electric power, thermal power, temperature, irradiance, ambient temperature and wind speed are extracted in a manner that the array unit identifier and time index are consistent. Under the same array cell identifier and the same time index, coupling operations are performed on the power characteristics and temperature characteristics. Under adjacent time indices, change operations are performed on the power characteristics and temperature characteristics to generate electrothermal coupling characteristics. Under the same array cell identifier and the same time index, environmental coupling operations are performed on irradiance features, ambient temperature features and wind speed features to generate environmental coupling features; The running status features, electrothermal coupling features and environmental coupling features are concatenated according to the preset field order to form an array coupling status feature vector; Stack the array coupling state feature vectors corresponding to different time indices in chronological order to generate the array coupling state feature tensor.

5. The deep learning-based PVT photovoltaic-thermal array output power dynamic prediction system according to claim 2, characterized in that, The process of constructing array spatial relationships based on array structure information in operational status data specifically involves: Extract array cell identifiers, array cell arrangement order, and physical connection information between array cells from the operational status data; The relative spatial positions of array elements in the array are determined based on the arrangement order of array elements, and the direct connection relationships between array elements are determined based on physical connection information. An array cell association matrix is ​​constructed according to the array cell identifier. The values ​​of the matrix elements in the array cell association matrix are determined by whether there is a physical connection and spatial adjacency between the corresponding array cells. Based on the array cell association matrix, a structured representation is generated to describe the spatial association between array cells.

6. The deep learning-based PVT photovoltaic-thermal array output power dynamic prediction system according to claim 2, characterized in that, The mapping of the array coupling state feature tensor to generate the array spatial correlation feature tensor is specifically as follows: Based on the array cell correlation matrix, spatial correlation mapping is performed on the array cell feature vectors under the same time index in the array coupling state feature tensor. The spatial association mapping includes performing a weighted combination operation on the feature vectors of mutually associated array units according to the weight distribution of the corresponding rows in the array unit association matrix; The feature results obtained by weighted combination operation are rearranged according to array cell identifier and time index to generate array space correlation feature tensor.

7. The deep learning-based PVT photovoltaic-thermal array output power dynamic prediction system according to claim 2, characterized in that, The improved MTGNN model includes an input encoding structure, a dual-graph learning structure, a graph convolution update structure, a temporal modeling structure, and a prediction output structure, specifically: The input encoding structure performs input concatenation and linear transformation on the array running state input tensor, the array coupling state feature tensor, and the array spatial correlation feature tensor to generate an array input feature representation. A dual-graph parallel learning mechanism is introduced into the dual-graph learning structure. A set of node embedding vectors is constructed based on the array input feature representation. Similarity calculation and normalization processing are performed on the set of node embedding vectors to generate electrical correlation matrix and thermal correlation matrix respectively. The similarity calculation includes performing inner product operation on the node embedding vector pairs to obtain similarity value. The normalization processing includes performing exponential processing on the similarity value and normalizing it by row to obtain weight distribution. Electric correlation graphs are generated based on electrical correlation matrices, and thermal correlation graphs are generated based on thermal correlation matrices. In the graph convolutional update structure, neighborhood weighted aggregation is performed on the array input feature representation based on the electrical correlation graph to obtain the electrical graph feature representation, and neighborhood weighted aggregation is performed on the array input feature representation based on the thermal correlation graph to obtain the thermal graph feature representation. The neighborhood weighted aggregation operation includes performing a weighted summation operation on the feature vectors of the associated nodes according to the weight distribution; feature concatenation and linear transformation are performed on the electrical graph feature representation and the thermal graph feature representation to generate a fused graph feature representation. In the time series modeling structure, multivariate time series modeling operations are performed based on the feature representation of the fusion graph to output the power prediction sequence. The multivariate time series modeling operations include performing sequence convolution operations on the time index dimension and performing gated combination operations on the sequence convolution output. In the predicted output structure, a linear mapping is performed on the output power prediction sequence to generate a sequence of predicted values ​​for the output power prediction sequence.

8. The deep learning-based PVT photovoltaic-thermal array output power dynamic prediction system according to claim 2, characterized in that, Specifically, S4 is: An input concatenation operation is performed on the array running state input tensor, the array coupling state feature tensor, and the array spatial association feature tensor. The input concatenation operation includes concatenating the three types of tensors in the feature dimension in a manner consistent with the array unit identifier and time index to generate an array input feature representation. A set of node embedding vectors is generated based on the array input feature representation, and each node embedding vector in the set corresponds to an array cell identifier; Perform similarity calculation on the set of node embedding vectors, wherein the similarity calculation includes performing an inner product operation on any two node embedding vectors to obtain a similarity value; The similarity values ​​are indexed and then normalized by row to generate a weight distribution matrix. The electrical correlation matrix and the thermal correlation matrix are generated based on the weight distribution matrix. The electrical correlation matrix and the thermal correlation matrix are two sets of independent weight distribution matrices. An electrical correlation graph is generated based on the electrical correlation matrix. The electrical correlation graph consists of array cell identifiers and the electrical correlation matrix. A thermal correlation graph is generated based on the thermal correlation matrix. The thermal correlation graph consists of array cell identifiers and the thermal correlation matrix.

9. The deep learning-based PVT photovoltaic-thermal array output power dynamic prediction system according to claim 2, characterized in that, Specifically, S5 is: Based on the electrical correlation graph and the thermal correlation graph, a neighborhood weighted aggregation operation is performed on the array input feature representation to obtain the electrical graph feature representation and the thermal graph feature representation. Perform feature concatenation and linear transformation on the electrical map feature representation and the thermal map feature representation to generate a fused map feature representation; Multivariate temporal modeling operations are performed based on the feature representation of the fused graph. The multivariate temporal modeling operations include performing sequence convolution operations and gating combination operations in the time index dimension to generate output power prediction sequences. The sequence convolution operation is to perform a convolution summation operation on the fusion map feature representation and the convolution kernel weight in the time index dimension. The fusion map feature value and the convolution kernel weight value within the preset convolution window are multiplied and accumulated to obtain the convolution output value. The gated combination operation is to calculate the gate weight value of the convolution output value obtained by the sequence convolution operation, and perform a weighted combination on the convolution output value based on the gate weight value. The gate weight value is obtained by the convolution output value through compression mapping and normalization mapping, and the weighted combination is obtained by multiplying the gate weight value and the convolution output value one by one and accumulating them.

10. The deep learning-based PVT photovoltaic-thermal array output power dynamic prediction system according to claim 2, characterized in that, Specifically, S6 is: The array operation state reachability constraint set is calculated based on the output power prediction sequence and operation state data. The array operation state reachability constraint set includes power change rate constraint and power value range constraint. The power change rate constraint is obtained by the difference between the output power prediction values ​​under adjacent time indices. The difference value is obtained by subtracting the output power prediction value corresponding to the previous time index from the output power prediction value corresponding to the later time index, and the power change rate value is obtained by dividing the difference value by the uniform sampling period. The power value range constraint is obtained by comparing the predicted output power value with the range boundary value of the operating status data. The range boundary value is obtained by linearly mapping the irradiance value, ambient temperature value, and wind speed value in the operating status data. The irradiance value, ambient temperature value, and wind speed value are multiplied by a preset coefficient and accumulated to obtain the range boundary value. Based on the array operating state reachability constraint set, reachability constraint processing is performed on the output power prediction sequence. The reachability constraint processing includes rate pruning for the output power prediction value corresponding to the time index where the power change rate exceeds the rate threshold, and interval pruning for the output power prediction value corresponding to the time index where the output power prediction value exceeds the interval boundary value, thereby generating dynamic output power prediction results.

Images