A photovoltaic support control system based on neural network optimization

Abstract

The present invention belongs to the field of intelligent control systems, and in particular relates to a photovoltaic support control system based on neural network optimization, comprising a data acquisition module, a feature embedding module, a dynamic graph processing module, a feature interaction module and an intelligent control module; the present invention creatively extracts various parameter features, and constructs a dynamic adjacency matrix containing all parameter relationships through a dynamic adjacency matrix construction method based on a gated self-attention mechanism, calculates the corresponding relationship between various parameters according to the dynamic adjacency matrix, controls the angle of the photovoltaic support by using the corresponding relationship, and realizes intelligent adjustment of the optimal support angle; the present invention adopts a multi-head attention mechanism, measures the relationship between various parameters by an attention score, highlights the influence of important features, improves the model's attention to key features, allows the model to comprehensively consider the relationship between features, reduces model overfitting, and realizes accurate calculation of the angle corresponding to the maximum power generation.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent control systems, and in particular relates to a photovoltaic support control system based on neural network optimization. Background Art

[0002] Photovoltaic support control refers to controlling the angle and direction of the photovoltaic support in the photovoltaic power generation system so that the solar panels can better receive solar radiation, thereby improving the power generation efficiency of the photovoltaic power generation system. However, the existing photovoltaic support control system adopts a fixed control strategy and cannot be dynamically adjusted according to real-time environmental conditions and system status, resulting in low power generation efficiency and difficulty in achieving automated and intelligent control. Summary of the invention

[0003] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a photovoltaic bracket control system based on neural network optimization. In view of the problem that the existing photovoltaic bracket control system adopts a fixed control strategy and cannot be dynamically adjusted according to the real-time environmental conditions and system status, the present invention creatively designs a neural network architecture to extract power generation characteristics, spatial embedding characteristics, time domain embedding characteristics and weather embedding characteristics, and constructs a dynamic adjacency matrix containing the relationship of [spatial-temporal weather-bracket position and angle-power generation] through a dynamic adjacency matrix construction method based on a gated self-attention mechanism. According to the dynamic adjacency matrix, the entanglement degree of all data is calculated and the features are fused to obtain the correspondence between the various features. The corresponding relationship is used to adjust the bracket angle corresponding to the real-time weather data and the maximum power generation to realize intelligent control of the bracket angle; the present invention creatively adopts a multi-head attention mechanism to calculate the cross-attention scores of time features, spatial features, weather attributes and power generation features, and measures the relationship between [spatiotemporal weather-bracket position and angle-power generation] through attention scores. Compared with the simple and direct fusion of all features, it can better highlight the influence of important features, improve the model's attention to key features, and enable the model to comprehensively consider the relationship between [spatiotemporal weather-bracket position and angle-power generation] features, reduce the risk of model overfitting, and realize accurate calculation of the angle corresponding to the maximum power generation.

[0004] The present invention provides a photovoltaic support control system based on neural network optimization, including a data acquisition module, a feature embedding module, a dynamic graph processing module, a feature interaction module and an intelligent control module;

[0005] The data acquisition module collects the position data, photovoltaic panel angle, power generation data and corresponding weather data of all photovoltaic brackets in the historical photovoltaic bracket array, and sends all the collected data to the feature embedding module;

[0006] The feature embedding module constructs an undirected graph based on the location data, photovoltaic panel angles, power generation data and corresponding weather data of all photovoltaic brackets, and extracts power generation features, spatial embedding features, temporal embedding features and weather embedding features, and sends all extracted features to the dynamic graph processing module and the feature interaction module;

[0007] The dynamic graph processing module, through a dynamic adjacency matrix construction method based on a gated self-attention mechanism, cross-integrates the power generation features, spatial embedding features, temporal embedding features and weather embedding features to obtain a dynamic adjacency matrix containing the relationship between [temporal and spatial weather-bracket position and angle-power generation], and sends the dynamic adjacency matrix to the feature interaction module;

[0008] The feature interaction module performs convolution processing on the dynamic adjacency matrix to obtain dynamic adjacency features, and uses a multi-head attention mechanism to capture the entanglement degree and fusion features between the [spatiotemporal weather-bracket position and angle-power generation] relationship, and sends the fusion features to the intelligent control module;

[0009] The intelligent control module activates the fusion feature, obtains the corresponding relationship of [spatiotemporal weather-bracket position and angle-power generation], collects real-time weather data, and controls the angle of the photovoltaic brackets at each position in the real-time photovoltaic bracket array according to the bracket position and angle corresponding to the corresponding relationship [real-time weather data-maximum power generation].

[0010] Furthermore, in the feature embedding module, an undirected graph is constructed to extract power generation features, spatial embedding features, temporal embedding features and weather embedding features, specifically including the following steps:

[0011] Step S1: Undirected graph construction, construct an undirected graph according to the positions of n photovoltaic brackets and the angles of photovoltaic panels:

[0012] ;

[0013] In the formula, ∈[1,n], represents a graph node, including the location information of the photovoltaic bracket and the angle information of the photovoltaic panel, Represents the edges formed by connecting different photovoltaic brackets. Represents the adjacency matrix, which is used to store the connection information between nodes;

[0014] Step S2: Power generation information encoding, collecting and encoding the historical power generation of the photovoltaic brackets at n locations at a specific time to obtain power generation characteristics;

[0015] Step S3: Spatial feature extraction, using node2vec algorithm to extract spatial features from undirected graphs The feature embedding is converted and encoded with a two-layer fully connected neural network to obtain spatial embedding features.

[0016] Step S4: Time feature encoding, encoding the information of each node in the undirected graph at a specific time to obtain the time domain embedding feature;

[0017] Step S5: Encoding external parameter information, collecting weather data for each node at a specific time, and encoding it with an encoder to obtain weather embedding features;

[0018] Furthermore, in the dynamic graph processing module, a method for constructing a dynamic adjacency matrix based on a gated self-attention mechanism specifically includes the following steps:

[0019] Step Q1: Calculate the spatial weather embedding features and the temporal weather embedding features. Add the weather embedding features to the spatial embedding features and the temporal embedding features to obtain the spatial weather embedding features and the temporal weather embedding features:

[0020] ;

[0021] ;

[0022] In the formula, represents the time domain embedding feature, represents the spatial embedding feature, represents weather embedding features, represents the space weather embedding features, represents the temporal weather embedding features, Represents node location and specific time;

[0023] Step Q2: Dynamically adaptive spatiotemporal weather feature calculation, add the spatial embedding features, temporal embedding features and weather embedding features of each node to obtain the spatiotemporal weather embedding features:

[0024] ;

[0025] In the formula, represents the time domain embedding feature, represents the spatial embedding feature, represents weather embedding features, represents the spatiotemporal weather embedding features, Represents node location and specific time;

[0026] Step Q3: Calculate the final embedded features, perform feature connection on the power generation features and the spatiotemporal weather embedded features to obtain the final embedded features:

[0027] ;

[0028] In the formula, represents the final embedding feature, Represents the power generation characteristics, represents the sum of spatiotemporal weather embedding features of n nodes, Represents a feature connection operation;

[0029] Step Q4: Dynamic feature matrix calculation, perform full connection layer processing on the final embedded features, calculate the parameters of the self-attention mechanism, and calculate the dynamic feature matrix of the feature time according to the parameters of the self-attention mechanism:

[0030] ;

[0031] ;

[0032] ;

[0033] In the formula, Represents two nonlinear fully connected layer processing, is the intermediate parameter, , and is the learnable weight matrix, , and are the parameters of the self-attention mechanism, representing the query vector, key vector and value vector respectively. represents the calculation of the self-attention mechanism, represents the dynamic feature matrix;

[0034] Step Q5: Extract spatial topology information. Use the gating mechanism to extract spatial topology information using the dynamic feature matrix at a specific time and the dynamic feature matrix of the previous time step:

[0035] ;

[0036] In the formula, represents the activation function, represents the dynamic feature matrix of the previous time step, represents the time step, and represents the learnable transformation matrix;

[0037] Step Q6: Dynamic adjacency matrix construction:

[0038] ;

[0039] In the formula, represents the hardmard product operation, represents the dynamic adjacency matrix;

[0040] Furthermore, in the feature interaction module, the calculation of the entanglement degree and the fusion feature specifically includes the following steps:

[0041] Step P1: Graph convolutional neural network processing, stack multiple convolutional layers to build a graph convolutional neural network, and perform feature extraction to obtain dynamic adjacency features:

[0042] ;

[0043] In the formula, Represents graph convolutional neural network processing with multiple convolutional layers, Represents dynamic adjacency features;

[0044] Step P2: Multi-head attention processing, through multi-head attention, the node relationship of the space weather embedding features is calculated, the weight parameters are calculated, and the activation function is used for normalization to obtain the space weather attention score:

[0045] ;

[0046] ;

[0047] In the formula, and represents two different nonlinear projection functions, Represents the time period The space weather embedding features, represents the feature concatenation operation, represents the inner product operation, is the preset parameter, represents the weight parameter, represents the normalization of the activation function, is the activation function, represents the space weather attention score;

[0048] Step P3: Hidden state update calculation, calculate the space-weather hidden state of each node through the space weather attention score:

[0049] ;

[0050] In the formula, represents the number of heads of multi-head attention, represents the nonlinear projection function, Represents the space-weather hidden state;

[0051] Step P4: Repeat steps P2-P3, take the time-weather embedding features as input, calculate the attention score and hidden state between time and node, and obtain the time-weather hidden state:

[0052] ;

[0053] ;

[0054] ;

[0055] In the formula, Represents the time-weather hidden state;

[0056] Step P5: Feature fusion, weighted fusion of space-weather hidden state and time-weather hidden state to obtain feature fusion:

[0057] ;

[0058] ;

[0059] In the formula, For activation function processing, and is the preset weight parameter matrix, is the preset adjustment parameter, represents the hardmard product operation, represents the degree of entanglement, Represents fusion features.

[0060] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0061] (1) In view of the problem that the existing photovoltaic bracket control system adopts a fixed control strategy and cannot be dynamically adjusted according to the real-time environmental conditions and system status, the present invention creatively designs a neural network architecture to extract power generation features, spatial embedding features, temporal embedding features and weather embedding features, and constructs a dynamic adjacency matrix containing the relationship between [spatiotemporal weather-bracket position and angle-power generation] through a dynamic adjacency matrix construction method based on a gated self-attention mechanism. According to the dynamic adjacency matrix, the entanglement degree of all data is calculated and the features are fused to obtain the corresponding relationship between the various features. Using the corresponding relationship, the bracket angle corresponding to the real-time weather data and the maximum power generation is adjusted to achieve intelligent control of the bracket angle;

[0062] (2) The present invention creatively adopts a multi-head attention mechanism to calculate cross-attention scores for temporal features, spatial features, weather attributes, and power generation features. The attention score is used to measure the relationship between [spatiotemporal weather-bracket position and angle-power generation]. Compared with the simple and direct fusion of all features, it can better highlight the influence of important features, improve the model's attention to key features, and enable the model to comprehensively consider the relationship between [spatiotemporal weather-bracket position and angle-power generation] features, reduce the risk of model overfitting, and achieve accurate calculation of the angle corresponding to the maximum power generation. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 It is a module diagram of a photovoltaic support control system based on neural network optimization provided by the present invention;

[0064] Figure 2 It is a flow chart of a method for constructing a dynamic adjacency matrix based on a gated self-attention mechanism provided by the present invention.

[0065] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0066] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0067] Example 1, see Figure 1 , a photovoltaic support control system based on neural network optimization, including a data acquisition module, a feature embedding module, a dynamic graph processing module, a feature interaction module and an intelligent control module;

[0068] The data acquisition module collects the position data, photovoltaic panel angle, power generation data and corresponding weather data of all photovoltaic brackets in the historical photovoltaic bracket array, and sends all the collected data to the feature embedding module;

[0069] The feature embedding module constructs an undirected graph based on the location data, photovoltaic panel angles, power generation data and corresponding weather data of all photovoltaic brackets, and extracts power generation features, spatial embedding features, temporal embedding features and weather embedding features, and sends all extracted features to the dynamic graph processing module and the feature interaction module;

[0070] The dynamic graph processing module, through a dynamic adjacency matrix construction method based on a gated self-attention mechanism, cross-integrates the power generation features, spatial embedding features, temporal embedding features and weather embedding features to obtain a dynamic adjacency matrix containing the relationship between [temporal and spatial weather-bracket position and angle-power generation], and sends the dynamic adjacency matrix to the feature interaction module;

[0071] The feature interaction module performs convolution processing on the dynamic adjacency matrix to obtain dynamic adjacency features, and uses a multi-head attention mechanism to capture the entanglement degree and fusion features between the [spatiotemporal weather-bracket position and angle-power generation] relationship, and sends the fusion features to the intelligent control module;

[0072] The intelligent control module activates the fusion feature, obtains the corresponding relationship of [spatiotemporal weather-bracket position and angle-power generation], collects real-time weather data, and controls the angle of the photovoltaic brackets at each position in the real-time photovoltaic bracket array according to the bracket position and angle corresponding to the corresponding relationship [real-time weather data-maximum power generation].

[0073] Embodiment 2: This embodiment is based on the above embodiment. In the feature embedding module, an undirected graph is constructed to extract power generation features, spatial embedding features, time domain embedding features and weather embedding features. Specifically, the following steps are included:

[0074] Step S1: Undirected graph construction: construct an undirected graph based on the positions of 20 photovoltaic brackets and the angles of photovoltaic panels:

[0075] ;

[0076] In the formula, ∈[1,20], represents a graph node, including the location information of the photovoltaic bracket and the angle information of the photovoltaic panel, Represents the edges formed by connecting different photovoltaic brackets. Represents the adjacency matrix, which is used to store the connection information between nodes;

[0077] Step S2: Power generation information encoding, collecting and encoding the historical power generation of photovoltaic brackets at 20 locations at a specific time to obtain power generation characteristics;

[0078] Step S3: Spatial feature extraction, using node2vec algorithm to extract spatial features from undirected graphs The feature embedding is converted and encoded with a two-layer fully connected neural network to obtain spatial embedding features.

[0079] Step S4: Time feature encoding, encoding the information of each node in the undirected graph at a specific time to obtain the time domain embedding feature;

[0080] Step S5: Encoding the external parameter information, collecting weather data for each node at a specific time, and encoding it with an encoder to obtain weather embedding features.

[0081] Example 3, see Figure 2 This embodiment is based on the above embodiment. In the dynamic graph processing module, a method for constructing a dynamic adjacency matrix based on a gated self-attention mechanism specifically includes the following steps:

[0082] Step Q1: Calculate the spatial weather embedding features and the temporal weather embedding features. Add the weather embedding features to the spatial embedding features and the temporal embedding features to obtain the spatial weather embedding features and the temporal weather embedding features:

[0083] ;

[0084] ;

[0085] In the formula, represents the time domain embedding feature, represents the spatial embedding feature, represents weather embedding features, represents the space weather embedding features, represents the temporal weather embedding features, Represents node location and specific time;

[0086] Step Q2: Dynamically adaptive spatiotemporal weather feature calculation, add the spatial embedding features, temporal embedding features and weather embedding features of each node to obtain the spatiotemporal weather embedding features:

[0087] ;

[0088] In the formula, represents the time domain embedding feature, represents the spatial embedding feature, represents weather embedding features, represents the spatiotemporal weather embedding features, Represents node location and specific time;

[0089] Step Q3: Calculate the final embedded features, perform feature connection on the power generation features and the spatiotemporal weather embedded features to obtain the final embedded features:

[0090] ;

[0091] In the formula, represents the final embedding feature, Represents the power generation characteristics, represents the sum of spatiotemporal weather embedding features of n nodes, Represents a feature connection operation;

[0092] Step Q4: Dynamic feature matrix calculation, perform full connection layer processing on the final embedded features, calculate the parameters of the self-attention mechanism, and calculate the dynamic feature matrix of the feature time according to the parameters of the self-attention mechanism:

[0093] ;

[0094] ;

[0095] ;

[0096] In the formula, Represents two nonlinear fully connected layer processing, is the intermediate parameter, , and is the learnable weight matrix, , and are the parameters of the self-attention mechanism, representing the query vector, key vector and value vector respectively. represents the calculation of the self-attention mechanism, represents the dynamic feature matrix;

[0097] Step Q5: Extract spatial topology information. Use the gating mechanism to extract spatial topology information using the dynamic feature matrix at a specific time and the dynamic feature matrix of the previous time step:

[0098] ;

[0099] In the formula, represents the activation function, represents the dynamic feature matrix of the previous time step, represents the time step, and represents the learnable transformation matrix;

[0100] Step Q6: Dynamic adjacency matrix construction:

[0101] ;

[0102] In the formula, represents the hardmard product operation, Represents a dynamic adjacency matrix.

[0103] By executing the above operations, the power generation features, spatial embedding features, time domain embedding features and weather embedding features are extracted, and a dynamic adjacency matrix construction method based on a gated self-attention mechanism is used to construct a dynamic adjacency matrix containing the relationship between [spatiotemporal weather-bracket position and angle-power generation]. According to the dynamic adjacency matrix, the entanglement degree of all data is calculated and the features are fused to obtain the corresponding relationship between the various features. The corresponding relationship is used to adjust the bracket angle corresponding to the real-time weather data and the maximum power generation, thereby realizing intelligent control of the bracket angle.

[0104] Embodiment 4: This embodiment is based on the above embodiment. In the feature interaction module, the calculation of the entanglement degree and the fusion feature specifically includes the following steps:

[0105] Step P1: Graph convolutional neural network processing, stack multiple convolutional layers to build a graph convolutional neural network, and perform feature extraction to obtain dynamic adjacency features:

[0106] ;

[0107] In the formula, Represents graph convolutional neural network processing with multiple convolutional layers, Represents dynamic adjacency features;

[0108] Step P2: Multi-head attention processing, through multi-head attention, the node relationship of the space weather embedding features is calculated, the weight parameters are calculated, and the activation function is used for normalization to obtain the space weather attention score:

[0109] ;

[0110] ;

[0111] In the formula, and represents two different nonlinear projection functions, Represents the time period The space weather embedding features, represents the feature concatenation operation, represents the inner product operation, is the preset parameter, represents the weight parameter, represents the normalization of the activation function, is the activation function, represents the space weather attention score;

[0112] Step P3: Hidden state update calculation, calculate the space-weather hidden state of each node through the space weather attention score:

[0113] ;

[0114] In the formula, represents the number of heads of multi-head attention, represents the nonlinear projection function, Represents the space-weather hidden state;

[0115] Step P4: Repeat steps P2-P3, take the time-weather embedding features as input, calculate the attention score and hidden state between time and node, and obtain the time-weather hidden state:

[0116] ;

[0117] ;

[0118] ;

[0119] In the formula, Represents the time-weather hidden state;

[0120] Step P5: Feature fusion, weighted fusion of space-weather hidden state and time-weather hidden state to obtain feature fusion:

[0121] ;

[0122] ;

[0123] In the formula, For activation function processing, and is the preset weight parameter matrix, is the preset adjustment parameter, represents the hardmard product operation, represents the degree of entanglement, Represents fusion features.

[0124] By performing the above operations, the attention scores of time features, spatial features, weather attributes and power generation features are cross-calculated. The relationship between [spatiotemporal weather-bracket position and angle-power generation] is measured by the attention score. Compared with the simple and direct fusion of all features, it can better highlight the influence of important features, improve the model's attention to key features, and enable the model to comprehensively consider the relationship between [spatiotemporal weather-bracket position and angle-power generation] features, reduce the risk of model overfitting, and achieve accurate calculation of the angle corresponding to the maximum power generation.

[0125] Embodiment 5, this embodiment is based on the above embodiment, in step S3, the Node2Vec algorithm specifically includes the following steps:

[0126] Define the random walk parameters:

[0127] Define the parameters for performing random walks in the graph, including the walk length and the number of neighbors sampled for each node;

[0128] Random Walk:

[0129] Perform multiple random walks starting from each node, walk in the graph according to predefined parameters, and record the walk paths;

[0130] Generate a node sequence:

[0131] The path sequence obtained from each random walk is used as a node sequence to construct the contextual relationship between nodes;

[0132] Building the Skip-gram model:

[0133] Use the obtained node sequence to build a Skip-gram model, where Skip-gram is a model for learning word embeddings, which can be applied to node embedding learning;

[0134] Train the model:

[0135] The Skip-gram model is used to train node embedding. By learning the relationship between nodes, each node is mapped to a low-dimensional vector space to obtain feature embedding.

[0136] Embodiment 6: This embodiment is based on the above embodiment, and the details of the data acquisition module are as follows:

[0137] Weather data:

[0138] Real-time light intensity (unit: W / m²);

[0139] Outside temperature, humidity, wind speed;

[0140] The solar altitude and azimuth at the current time;

[0141] Location Data:

[0142] Geographical location information of all photovoltaic brackets;

[0143] Power generation data:

[0144] The output power, voltage and current of the photovoltaic panels;

[0145] Photovoltaic bracket angle:

[0146] The current tilt angle and azimuth of the photovoltaic panel;

[0147] Data acquisition equipment:

[0148] Light sensor, temperature and humidity sensor, wind speed and direction meter, GPS module, current / voltage sensor and angle meter.

[0149] Embodiment 7: This embodiment is based on the above embodiment and optimizes the system training. The details are as follows:

[0150] Training Data:

[0151] Use specific historical weather data and the location, angle, and power generation of photovoltaic panels as training data sets;

[0152] The data samples include input parameters, real-time weather data and target output (optimal angle);

[0153] Loss function:

[0154] Minimize the power prediction error by the mean square error;

[0155] Optimization algorithm:

[0156] Use the adaptive optimizer Adam to speed up training;

[0157] Dynamically adjust the learning rate to avoid overfitting.

[0158] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0159] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

[0160] The present invention and its embodiments are described above, and such description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if ordinary technicians in the field are inspired by it, without departing from the purpose of the invention, they can design a structure and embodiment similar to the technical solution without creativity, which should belong to the protection scope of the present invention.

Claims

1. A photovoltaic support control system based on neural network optimization, characterized in that: It includes data acquisition module, feature embedding module, dynamic graph processing module, feature interaction module and intelligent control module; The data acquisition module collects the position data, photovoltaic panel angle, power generation data and corresponding weather data of all photovoltaic brackets in the historical photovoltaic bracket array, and sends all the collected data to the feature embedding module; The feature embedding module constructs an undirected graph based on the location data, photovoltaic panel angles, power generation data and corresponding weather data of all photovoltaic brackets, and extracts power generation features, spatial embedding features, temporal embedding features and weather embedding features, and sends all extracted features to the dynamic graph processing module and the feature interaction module; The dynamic graph processing module, through a dynamic adjacency matrix construction method based on a gated self-attention mechanism, cross-integrates the power generation features, spatial embedding features, temporal embedding features and weather embedding features to obtain a dynamic adjacency matrix containing the relationship between [temporal and spatial weather-bracket position and angle-power generation], and sends the dynamic adjacency matrix to the feature interaction module; In the dynamic graph processing module, a method for constructing a dynamic adjacency matrix based on a gated self-attention mechanism specifically includes the following steps: Step Q1: Calculate the spatial weather embedding features and the temporal weather embedding features. Add the weather embedding features to the spatial embedding features and the temporal embedding features to obtain the spatial weather embedding features and the temporal weather embedding features: ； ； In the formula, represents the time domain embedding feature, represents the spatial embedding feature, represents weather embedding features, represents the space weather embedding features, represents the temporal weather embedding features, Represents node location and specific time; Step Q2: Dynamically adaptive spatiotemporal weather feature calculation, add the spatial embedding features, temporal embedding features and weather embedding features of each node to obtain the spatiotemporal weather embedding features: ； In the formula, represents the time domain embedding feature, represents the spatial embedding feature, represents weather embedding features, represents the spatiotemporal weather embedding features, Represents node location and specific time; Step Q3: Calculate the final embedded features, connect the power generation features with the spatiotemporal weather embedded features, and obtain the final embedded features: ； In the formula, represents the final embedding feature, Represents the power generation characteristics, represents the sum of spatiotemporal weather embedding features of n nodes, Represents a feature connection operation; Step Q4: Dynamic feature matrix calculation, perform full connection layer processing on the final embedded features, calculate the parameters of the self-attention mechanism, and calculate the dynamic feature matrix of the feature time according to the parameters of the self-attention mechanism: ； ； ； In the formula, Represents two nonlinear fully connected layer processing, is the intermediate parameter, , and is the learnable weight matrix, , and are the parameters of the self-attention mechanism, representing the query vector, key vector and value vector respectively. represents the calculation of the self-attention mechanism, represents the dynamic feature matrix; Step Q5: Extract spatial topology information. Use the gating mechanism to extract spatial topology information using the dynamic feature matrix at a specific time and the dynamic feature matrix of the previous time step: ； In the formula, represents the activation function, represents the dynamic feature matrix of the previous time step, represents the time step, and represents the learnable transformation matrix, Represents spatial topological information; Step Q6: Dynamic adjacency matrix construction: ； In the formula, represents the hardmard product operation, represents the dynamic adjacency matrix; The feature interaction module performs convolution processing on the dynamic adjacency matrix to obtain dynamic adjacency features, and uses a multi-head attention mechanism to capture the entanglement degree and fusion features between the [spatiotemporal weather-bracket position and angle-power generation] relationship, and sends the fusion features to the intelligent control module; The intelligent control module activates the fusion feature, obtains the correspondence between the spatiotemporal weather-bracket position and angle-power generation, collects real-time weather data, and controls the angle of the photovoltaic brackets at each position in the real-time photovoltaic bracket array according to the bracket position and angle corresponding to the corresponding relationship [real-time weather data-maximum power generation].

2. According to claim 1, a photovoltaic support control system based on neural network optimization is characterized in that: In the feature embedding module, an undirected graph is constructed to extract power generation features, spatial embedding features, temporal embedding features and weather embedding features, specifically including the following steps: Step S1: Undirected graph construction: construct an undirected graph based on the positions of n photovoltaic brackets and the angles of photovoltaic panels: ； In the formula, ∈[1,n], represents a graph node, including the location information of the photovoltaic bracket and the angle information of the photovoltaic panel, Represents the edges formed by connecting different photovoltaic brackets. Represents the adjacency matrix, which is used to store the connection information between nodes; Step S2: Power generation information encoding, collecting and encoding the historical power generation of the photovoltaic brackets at n locations at a specific time to obtain power generation characteristics; Step S3: Spatial feature extraction, using node2vec algorithm to extract spatial features from undirected graphs The feature embedding is converted and encoded with a two-layer fully connected neural network to obtain spatial embedding features. Step S4: Time feature encoding, encoding the information of each node in the undirected graph at a specific time to obtain the time domain embedding feature; Step S5: Encoding the external parameter information, collecting weather data for each node at a specific time, and encoding it with an encoder to obtain weather embedding features.

3. A photovoltaic support control system based on neural network optimization according to claim 2, characterized in that: In the feature interaction module, the calculation of the entanglement degree and the fusion feature specifically includes the following steps: Step P1: Graph convolutional neural network processing, stack multiple convolutional layers to build a graph convolutional neural network, and perform feature extraction to obtain dynamic adjacency features: ； In the formula, Represents graph convolutional neural network processing with multiple convolutional layers, Represents dynamic adjacency features; Step P2: Multi-head attention processing, through multi-head attention, the node relationship of the space weather embedding features is calculated, the weight parameters are calculated, and the activation function is used for normalization to obtain the space weather attention score: ； ； In the formula, and represents two different nonlinear projection functions, Represents the time period The space weather embedding features, represents the feature concatenation operation, represents the inner product operation, is the preset parameter, represents the weight parameter, represents the normalization of the activation function, is the activation function, represents the space weather attention score; Step P3: Hidden state update calculation, calculate the space-weather hidden state of each node through the space weather attention score: ； In the formula, represents the number of heads of multi-head attention, represents the nonlinear projection function, Represents the space-weather hidden state; Step P4: Repeat steps P2-P3, take the time-weather embedding features as input, calculate the attention score and hidden state between time and node, and obtain the time-weather hidden state: Step P5: Feature fusion, weighted fusion of space-weather hidden state and time-weather hidden state to obtain feature fusion: ； ； In the formula, For activation function processing, and is the preset weight parameter matrix, is the preset adjustment parameter, represents the hardmard product operation, represents the degree of entanglement, Represents fusion features.

Images