A smart city street lamp group coordination control method

Abstract

The present application relates to a kind of smart city street lamp group coordination control method, its core technical scheme is using big data analysis, based on the connection event of street lamp node in continuous time window, construct the difference matrix of the topological variation and node activity, by topological evolution coding and symbol perception gate processing, extract topological change characteristics and map it as local parameter correction coefficient, realize the dynamic adjustment of model, improve the immediacy and accuracy of energy balance decision-making.The present application effectively improves the adaptability of street lamp network model to topological dynamic change and energy efficiency control performance.

Description

Technical Field

[0001] This invention relates to the field of big data analysis technology for energy consumption of smart city street light groups, and in particular to a coordinated control method for smart city street light groups. Background Technology

[0002] Current smart city street light cluster energy balance control in engineering applications largely relies on the assumption of low-frequency changes in the road network structure. This involves static topology data collection for main roads, secondary roads, and plazas to establish a baseline adjacency graph, followed by model training and optimization using long-term historical data. However, when topology changes occur, such as node additions / removals, power outages due to construction, temporary light pole deployment, or secondary road reconstruction, the computational overhead of traditional solutions increases significantly. Furthermore, frequent adjustments can easily lead to model instability and a rapid decline in energy balance prediction accuracy.

[0003] However, with the digital transformation of cities and the intelligent upgrading of infrastructure, the physical connections of street light clusters exhibit high-frequency dynamic evolution characteristics. The urban road network structure is constantly adjusted due to factors such as construction, temporary tasks, and functional expansion, resulting in significant daily changes in node adjacency relationships and local connectivity. In the context of massive data changes, existing graph neural network models, relying on static or low-change topologies, lack the ability to adapt to high-frequency structural changes in real time. Summary of the Invention

[0004] This application provides a smart city street light group coordinated control method, which aims to solve one of the problems or issues of the existing technology mentioned in the background.

[0005] This application provides a method for coordinated control of smart city street light groups, specifically including: S1: Obtain node connection status change events of urban street light groups within a continuous time window, and generate a topology difference matrix based on the node connection status change events; S2: Input the topological difference matrix into the pre-constructed topological evolution encoder, and use multi-layer sparse convolution and symbolic gate unit to perform feature compression processing on the topological difference matrix, and output the topological evolution feature vector. S3: Based on the topology evolution feature vector and the key layer structure parameters of the preset energy consumption balance control strategy model, perform a weight mapping transformation operation to generate local parameter correction coefficients for correcting the original energy consumption balance control strategy model. S4: The energy consumption balance control strategy model is dynamically fine-tuned using the local parameter correction coefficients to construct an online adaptive energy consumption balance graph neural network model. S5: Calculate the output deviation value of inputting historical stable topology sample data into the online adaptive energy consumption balance graph neural network model and the original energy consumption balance control strategy model for inference verification; S6: If the output deviation value exceeds the preset topology consistency constraint threshold, a local rollback mechanism is triggered to restore the original historical parameter state of the energy consumption balance control strategy model and output control commands; if the output deviation value does not exceed the topology consistency constraint threshold, the online adaptive energy consumption balance graph neural network model is maintained and control commands are output.

[0006] The smart city street light group coordination control method provided in this application has the following beneficial effects: (1) The novel graph neural network adaptation method provided in this application for energy consumption balance calculation function in collaborative decision-making of street light groups in smart cities has the following beneficial effects: In view of the problems of model rigidity, high retraining cost and large response delay that are common in traditional graph neural networks in dynamic topology scenarios, this solution models the structural evolution of urban road network as a time-series snapshot sequence and designs a lightweight topology evolution encoder to extract the difference features between adjacent snapshots, thus realizing for the first time a quantifiable and compressible representation of the topology change process. Compared with the existing technology that relies on offline retraining or full graph reconstruction, this method avoids frequent data backhaul and global parameter updates. It only needs to transmit the sparse difference matrix and complete the local model adaptation by the edge side, which significantly reduces the communication bandwidth occupation and the computing load of the central server. At the same time, since the evolution feature vector can accurately capture the change type of local connection relationship and its spatial influence range, the backbone GNN model can be quickly adjusted online without sacrificing the stability of historical knowledge. This effectively overcomes the technical bottleneck that static graph models are difficult to adapt to the dynamic evolution of urban infrastructure and greatly improves the continuous service capability and robustness of the system in complex real-world environments.

[0007] (2) Furthermore, by constructing a pluggable lightweight adaptation module containing a structure-sensitive weight mapper, a topology consistency constraint, and an energy consumption gradient guide, this scheme realizes a triple collaborative mechanism of structural change perception, historical performance assurance, and actual energy consumption target-driven model parameter update. Among them, the structure-sensitive weight mapper dynamically generates local correction coefficients of key GNN layers based on evolutionary characteristics, enabling the model to specifically enhance its expressive ability for newly added branches or disconnected regions; the topology consistency constraint effectively prevents global inference drift caused by local updates by monitoring and overshooting the output deviation of historical stable samples, ensuring the temporal continuity and credibility of model behavior; and the energy consumption gradient guide constructs a lightweight supervision signal using only the real-time power data of street light nodes in the current topology change area, driving the adaptation process to focus on the real energy consumption balance target, achieving closed-loop optimization without external annotation. The lightweight closed-loop mechanism formed by the coupling of these three elements not only avoids the high-dimensional computational overhead brought about by traditional multi-view fusion or hypergraph expansion methods, but also breaks through the application limitations of attention recalibration strategies in label-free scenarios. This enables the entire adaptation process to have high precision, low latency, and self-consistency, with a single response time controlled within 800 milliseconds, meeting the dual requirements of real-time performance and reliability for edge computing scenarios in smart cities.

[0008] (3) Furthermore, this method achieves a paradigm shift of "change as input" at the system architecture level, transforming the topological dynamics, which was originally considered an interference factor, into an effective signal driving model evolution. This fundamentally changes the design logic of passively responding to network changes in existing technologies, and constructs a graph neural network evolution system with autonomous growth capabilities. This system can adapt to urban operating environments with up to 47 topological changes per day without complex parameter tuning. In actual tests, it reduced the fluctuation range of energy consumption balance calculation errors by 61.3% compared to traditional methods, significantly improving the accuracy and stability of collaborative decision-making for street light groups. At the same time, it greatly reduces the computing power burden on edge gateways and has good deployability and scalability. The overall solution is not only applicable to smart street light systems, but can also be migrated to other urban IoT scenarios with spatiotemporal evolution characteristics, such as traffic signal collaborative control and charging pile network scheduling, demonstrating broad applicability and engineering value. The above-mentioned technical advantages together constitute a new paradigm of efficient, energy-saving, and adaptive urban-level graph model operation and maintenance, effectively promoting the leap of artificial intelligence technology from "static intelligence" to "continuously evolving intelligence". Attached Figure Description

[0009] Figure 1 This is the main flowchart of a smart city street light group coordination and control method; Figure 2 This is a sub-flowchart of a smart city street light group coordinated control method; Figure 3This is another sub-flowchart of a smart city street light group coordinated control method; Figure 4 This is an application environment diagram of a smart city street light group coordination control method in one embodiment; Figure 5 This is a schematic diagram of the structure of a computer device provided in one embodiment of this application. Detailed Implementation

[0010] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0011] The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the invention. Furthermore, reference numerals and / or letters may be repeated in different examples; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed.

[0012] like Figure 1 As shown, this application provides a smart city street light group coordinated control method, specifically including: S1: Obtain node connection status change events of smart city street light groups within a continuous time window, and based on the topology difference matrix of the node connection status change events, wherein the topology difference matrix includes edge addition / deletion identifiers and node activity changes.

[0013] S2: Input the topological difference matrix into the pre-constructed topological evolution encoder, and use multi-layer sparse convolution and symbolic gating unit to perform feature compression processing on the topological difference matrix, and output a topological evolution feature vector that represents the current change type and its spatial influence range.

[0014] S3: Based on the topology evolution feature vector and the key layer structure parameters of the preset energy consumption balance control strategy model, perform a weight mapping transformation operation to generate local parameter correction coefficients for correcting the original energy consumption balance control strategy model. These local parameter correction coefficients are used to correct the first hop aggregation layer and attention weight layer of the energy consumption balance control strategy model.

[0015] S4: The corresponding level parameters of the energy consumption balance control strategy model are dynamically fine-tuned using the local parameter correction coefficients to construct an online adaptive energy consumption balance graph neural network model with real-time adaptability.

[0016] S5: Calculate the output deviation value of inputting historical stable topology sample data into the online adaptive energy consumption balance graph neural network model and the original energy consumption balance control strategy model for inference verification.

[0017] S6: If the output deviation value exceeds the preset topology consistency constraint threshold, a local rollback mechanism is triggered to restore the original historical parameter state of the energy consumption balance control strategy model and output control commands; if the output deviation value does not exceed the topology consistency constraint threshold, the online adaptive energy consumption balance graph neural network model is maintained and control commands are output.

[0018] Step S1: Obtain node connection state change events of the smart city street light group within a continuous time window, and base the results on the topology difference matrix of the node connection state change events, wherein the topology difference matrix includes edge addition / deletion identifiers and node activity changes. Specifically, this includes: S1.1: Obtain the node connection status change event stream reported by each street light node within a continuous time window. Based on the timestamp and node identification information in the node connection status change event stream, perform time sequence alignment and deduplication on the original event data to generate a standardized node connection status change event sequence.

[0019] Dynamic monitoring signals of the connection status of smart city street light group nodes are centrally aggregated to form a raw event dataset containing multi-source acquisition results, which serves as the input object. Based on the timestamp field in the raw event dataset, a unified timeline mapping process is performed on the event occurrence times of different acquisition nodes using time-series calibration processing. This aligns events generated across devices into a unified reference time frame to eliminate time drift caused by differences in acquisition latency. Based on the node identifier field in the raw event dataset, hash encoding and uniqueness verification rules are applied to perform deduplication operations on duplicate event records, eliminating multiple records caused by broadcast redundancy or repeated reporting. The event sequence after time-series alignment and deduplication is input into a standardization encoding module. According to preset format specifications, timestamps are converted into millisecond-precision integer values, node identifiers are converted into fixed-length binary codes, and event attribute fields are renamed using field mapping rules to ensure semantic and structural consistency of data from different sources. A sequence integrity detector is invoked to perform continuity and integrity checks on the standardized event sequence, filling in missing records or marking invalid status bits caused by transmission packet loss to ensure the accuracy of the input for subsequent topology difference matrix construction. Through the above-described layer-by-layer processing method, the multi-source asynchronous event stream generated during the acquisition phase is transformed into a sequence of node connection state change events with a unified time base, unique node identifier, and standard format structure, thereby enabling unbiased data input in the subsequent topology parsing phase.

[0020] S1.2: Parse the operation type field in the standardized node connection state change event sequence, and use the adjacency relationship mapping rules to convert operations such as adding branches, connecting temporary light poles, or power outages during construction into edge addition identifiers and edge deletion identifiers, generating an original edge change identifier set containing topology change direction information.

[0021] Based on the standardized node connection state change event sequence output by step S1.1, structured parsing is performed on the operation type field to clarify the topology change category and its directionality information contained in the event. For each parsed event record, a preset adjacency mapping rule base is invoked to map the new branch event to the edge increment identifier between the corresponding node pairs, and the target node index position is recorded in the mapping table. For temporary light pole access events, based on the node identifier of its connected object, the connection relationship is also mapped to an edge increment identifier, and a bidirectional association storage is established to ensure the symmetry of the adjacency matrix. For construction power outage events, combined with the connection relationship of the affected nodes, they are mapped to the edge deletion identifier between the corresponding node pairs, and the triggering source node of the deletion operation is marked in the identifier set. A differentiated encoding strategy is adopted to assign different symbol codes to the edge increment identifier and the edge deletion identifier, forming an identifier structure that can reflect the topology change direction information. The edge increment identifiers and edge deletion identifiers obtained from all event mappings are merged in chronological order to generate a set data structure, and the timestamp and node index metadata corresponding to each identifier are appended to the set. By using the above processing method, the standardized event sequence of the previous step is transformed into a set of original edge change identifiers that can be used for subsequent activity calculation and adjacency matrix update, thereby achieving a clear quantitative representation of the direction of topological change.

[0022] S1.3: Count the number of active state changes of all street light nodes involved in the original edge change identifier set, calculate the online probability offset of each node based on the preset activity threshold, and generate a node activity change vector representing the fluctuation of local network connectivity.

[0023] The system receives the original set of edge change identifiers as a statistical object, calls the node index parsing module to extract the unique identifiers of all street light nodes involved in the set, and constructs a node active state count table to record the number of state changes for each node.

[0024] Based on the number of changes field in the node activity status count table, and combined with a preset activity threshold, the online probability calculation module is invoked to map the number of changes to a node online probability offset. The offset is calculated using the following expression: Where, ΔP i C represents the online probability offset of node i. i T represents the number of times node i changes its active state within the current time window.i This represents the activity threshold of node i.

[0025] The online probability offsets of all nodes are arranged in order of node identifier index to form a preliminary sequence of node activity changes that characterizes the fluctuations in local network connectivity. The numerical smoothing module is then called to perform a sliding median filter operation to eliminate extreme value interference.

[0026] Normalization is performed on the smoothed node activity change sequence to map the probability values ​​to a preset standardized interval for uniform scaling, and the node activity change vector is output for subsequent difference matrix generation steps.

[0027] By statistically analyzing the number of active changes, calculating the online probability offset, and smoothing and normalizing the data, the original set of edge change identifiers from the previous step is transformed into a quantifiable node activity change vector that can be directly embedded into the difference matrix, thus achieving a high-fidelity characterization of local network connectivity fluctuations.

[0028] S1.4: Construct a baseline version of the global street light group adjacency matrix at the current moment, map the edge addition identifiers and edge deletion identifiers in the original edge change identifier set to the corresponding row and column positions of the baseline version of the global street light group adjacency matrix, perform matrix element update operation, and generate a real-time topology adjacency matrix that reflects the latest physical connection relationship.

[0029] Using the original edge change identifier set generated in the previous step and the baseline version of the global street light group adjacency matrix as input objects, the matrix structure initialization module is called to read the two-dimensional data structure of the baseline version of the global adjacency matrix and construct a writable tensor copy in memory. Edge addition identifiers in the original edge change identifier set are located to the corresponding row and column coordinates of the tensor according to the node index mapping rules, and the matrix element at that position is modified from zero to one using the assignment operator to reflect the newly added physical connection. Similarly, edge deletion identifiers in the original edge change identifier set are located to the corresponding row and column coordinates of the tensor according to the node index mapping rules, and the matrix element at that position is modified from one to zero using the assignment operator to eliminate the broken physical connection. During the matrix element update process, a symmetry check function is called to update the horizontal and vertical elements of the matrix simultaneously for each edge addition or deletion operation, ensuring that the adjacency matrix maintains symmetry. A sparsity constraint check is performed on the matrix after all edge addition and deletion operations. If unnecessary non-zero elements are found, a sparsification function is called to remove them, and the final matrix is ​​compressed into a sparse matrix storage format to reduce memory usage. By mapping the positions of matrix elements and performing sparsification, the set of edge change identifiers from the previous step is transformed into a real-time topological adjacency matrix that reflects the latest physical connection relationships, thereby achieving standardized updates of the topology and improving its computability.

[0030] S1.5: Calculate the difference matrix between the real-time topological adjacency matrix and the baseline version of the global street light group adjacency matrix, embed the node activity change vector into the auxiliary channel dimension of the difference matrix, perform multi-source data fusion encoding processing, and output a topological difference matrix containing edge addition / deletion identifiers and node activity changes.

[0031] Step S2: Input the topological difference matrix into the pre-built topological evolution encoder, and use multi-layer sparse convolution and symbolic gating unit to perform feature compression processing on the topological difference matrix, and output the topological evolution feature vector that represents the current change type and its spatial influence range.

[0032] In this embodiment, the topology evolution encoder consists of a sparsity preprocessing unit, a multi-scale structural feature extraction module, a symbol-aware gating unit, and a feature fusion and compression module. The collaborative workflow of the topology evolution encoder is as follows: First, the sparsification preprocessing unit receives the topology difference matrix, removes invalid connection noise through channel rearrangement and zero-value masking, and outputs a standardized sparse adjacency tensor. Second, the multi-scale structural feature extraction module uses a multi-layer sparse convolution kernel group to extract the local connectivity offset of the tensor, aggregates the structural variation patterns of the node's neighbors, and generates a multi-scale topology residual feature map. Next, the symbol-aware gating unit performs nonlinear weight modulation based on the positive and negative sign distribution in the feature map to separate the anisotropic effects of edge addition (expansion) and edge deletion (contraction), generating a symbol-decoupled feature stream, thereby solving the defect of traditional convolution in distinguishing the direction of change. Finally, the feature fusion and compression module performs cross-channel attention fusion and global pooling compression on the symbol-decoupled feature stream, integrates multi-scale spatial influence information and eliminates redundant dimensions, outputs a topology evolution feature vector representing the change type and influence range, and completes the conversion from high-dimensional graph structure data to low-dimensional evolution driving signals.

[0033] Specifically, it includes: S2.1: Obtain the topological difference matrix as the initial input data, and perform channel rearrangement and zero-value mask filtering based on the discrete distribution characteristics of the edge addition and deletion indicators in the matrix to remove invalid connection noise and generate a normalized sparse adjacency tensor, ensuring that subsequent convolution operations are only performed on the effective topological change region.

[0034] S2.2: Using the standardized sparse adjacency tensor as the input object, the predefined multi-layer sparse convolution kernel group is called to perform the local connectivity offset extraction operation, so as to aggregate the structural variation patterns between the nodes and output the multi-scale topological residual feature map, realizing the initial abstraction from the original connection state to the local structural evolution pattern.

[0035] The normalized sparse adjacency tensor after channel rearrangement and zero-value masking is initialized by multi-scale convolution kernel calls. The convolution kernel size, stride and sparse connection mode of each scale are determined according to the preset sparse convolution kernel group structure parameters, so that the convolution calculation only covers the nodes marked as valid topology changes and their neighbors.

[0036] The first layer of sparse convolution is performed on the standardized sparse adjacency tensor. The connectivity offset is extracted in the neighborhood of the local node using the convolution kernel weight parameters. The spatial position index that corresponds one-to-one with the input topological difference matrix is ​​retained in the output features to ensure that the structural variation source can be accurately aligned during subsequent multi-scale fusion.

[0037] The first layer of convolution output features are used as input, and the second and subsequent multi-layer sparse convolution kernel groups are called. By expanding the receptive field coverage, the connectivity offset information of higher-order neighbors is aggregated layer by layer. In the calculation of the convolution amplitude of each layer, a normalization factor is introduced to eliminate the feature imbalance caused by the difference in node degree at different scales.

[0038] By performing interpolation operations between multi-layer convolution results, the incremental changes in structural variation patterns in convolution features at each scale are revealed by constructing topological residual feature maps. The convolution output is subtracted element-wise from the convolution result of the previous scale to obtain a multi-scale sequence of local connectivity evolution trends.

[0039] The multi-scale topological residual feature maps are merged into a unified tensor structure by using feature channel splicing, and the weight coefficients of each scale in the residual features are retained so that subsequent symbol-aware processing can separate and modulate the spatial effects at different scales. Through the above processing method, the sparse adjacency tensor of the previous step is transformed into a multi-scale topological residual feature map that represents the local structural evolution mode, realizing the initial abstraction of topological changes.

[0040] S2.3: Based on the positive and negative sign distribution information in the multi-scale topological residual feature map, activate the sign-aware gating unit to perform nonlinear weight modulation processing, so as to separate the expansion effect caused by edge addition and the contraction effect caused by edge deletion and generate the sign decoupled feature flow, thus solving the technical defect of traditional convolution that cannot distinguish the directionality of topological changes.

[0041] Based on the positive and negative sign distribution information in the multi-scale topological residual feature map, the initial state matrix of the symbol-aware gating unit is loaded as the execution object of weight modulation. The symbol polarity analysis component is called to perform symbol polarity detection on each channel of the input feature map and generate a polarity indicator matrix.

[0042] Element-wise multiplication is performed between the polarity indicator matrix and the magnitude matrix of the multi-scale topological residual feature map to form a magnitude-symbol coupling matrix that includes differences in edge addition and deletion, ensuring that the symbol-aware gating unit can simultaneously receive directional and intensity information.

[0043] A nonlinear weighted modulation operator is invoked to perform dual-channel modulation based on the amplitude-sign coupling matrix. The positive channel applies an expanding weight function, while the negative channel applies a contracting weight function. This function is defined as follows: Where A is the magnitude matrix, S is the sign matrix, α and β are the modulation coefficients, and σ is the sigmoid function, which realizes the weighted nonlinear mapping of topological changes in different directions.

[0044] The expansion weight function output and the contraction weight function output are applied to the feature components of the positive and negative polarity channels respectively, and element-level fusion within the channel is performed to eliminate the weight inversion effect caused by sign conflict.

[0045] Differential operator operations are performed on the fused positive polarity feature components and the fused negative polarity feature components to isolate the two types of spatial effects and generate a symbolically decoupled feature flow with directional decoupling.

[0046] Through the aforementioned nonlinear weight modulation and directional isolation processing, the multi-scale topological residual feature map of the previous step is transformed into a symbolically decoupled feature flow that can independently characterize the expansion and contraction spatial patterns, thus achieving the technical effect that traditional convolution cannot distinguish the directionality of topological changes.

[0047] S2.4: Perform cross-channel attention fusion and global pooling compression on the symbol decoupled feature stream to integrate spatial influence range information at different scales and eliminate redundant dimensions. Finally, output a topological evolution feature vector that represents the current change type and its spatial influence range, completing the conversion from high-dimensional graph structure data to low-dimensional evolution driving signal.

[0048] like Figure 2 As shown, step S3: Based on the topology evolution feature vector and the key layer structure parameters of the preset energy consumption balance control strategy model, a weight mapping transformation operation is performed to generate local parameter correction coefficients for correcting the original energy consumption balance control strategy model. These local parameter correction coefficients are used to correct the first hop aggregation layer and attention weight layer of the energy consumption balance control strategy model. Specifically, this includes: S3.1: Obtain the initialization configuration parameters of the predefined structure-sensitive weight mapper and the topological evolution feature vector output from the previous steps. Use a fully connected neural network layer to perform nonlinear projection transformation on the topological evolution feature vector to extract a high-dimensional latent space mapping code that represents the current topological change type, and output a sequence of latent variables of topological change with semantic expressive ability.

[0049] The inputs include the initialization configuration parameters of the structure-sensitive weight mapper and the topological evolution feature vectors output from the preceding steps.

[0050] The topological evolution feature vector is loaded into the first layer weight array of the fully connected neural network according to the preset input dimension, and a linear transformation is performed to generate the initial hidden representation vector.

[0051] Based on the initial hidden representation vector, a nonlinear activation function processing unit is called, and the hyperbolic tangent function is used to compress the feature amplitude and reverse the direction, so as to form an intermediate feature representation with nonlinear discrimination capability.

[0052] Multi-level fully connected mapping operations are performed on the intermediate feature representations. Each level of mapping is batch normalized by combining the normalization parameters in the structure-sensitive weight mapper, and a rectified linear unit is added before the output to enhance the sparse response characteristics.

[0053] For the high-dimensional feature vectors output by multi-level mapping, the feature reconstruction module is called to merge similar feature channels and redistribute weights according to the topological change sensitivity coefficient to generate a sequence of topological change latent variables that meet the requirements of semantic expression.

[0054] Through the above chain mapping and activation processing, the topological evolution feature vector is transformed into a high-dimensional latent space encoding that can be parsed by the subsequent channel attention mechanism, thereby realizing the semantic representation of the topological change pattern.

[0055] The structure-sensitive weight mapper architecture includes: a multi-layer fully connected neural network, activation functions, a batch normalization module, a feature reconstruction module, and an output module.

[0056] The activation function uses the hyperbolic tangent function for nonlinear projection.

[0057] Batch normalization: Batch normalization is performed after each fully connected mapping (initial mean 0, variance 1).

[0058] Feature reconstruction module: Merge similar feature channels and redistribute weights according to the topology change sensitivity coefficient (0.1~1.0).

[0059] Output module: Generates a 128-dimensional sequence of latent variables for topological changes.

[0060] The initial configuration of the structure-sensitive weight mapper includes: an input feature vector dimension of 64, a first-layer fully connected neural network matrix size of 64×128, a second-layer matrix of 128×256, and a third-layer matrix of 256×128. Batch normalization parameters are set to a mean of 0 and a variance of 1. The hyperbolic tangent function is used as the nonlinear activation function. When processing the 64-dimensional topological evolution feature vector from the output of step S2, a linear transformation is first performed using a 64×128 weight array, followed by a hyperbolic tangent mapping to form a 128-dimensional intermediate representation. This is then processed by a 128×256 fully connected mapping and batch normalization to obtain 256-dimensional features. Finally, a 256×128 mapping is performed to form the final 128-dimensional high-dimensional latent space encoding, where the weights of each channel are redistributed within the range of 0.1 / 1.0 according to the topological change sensitivity coefficient. The final output 128-dimensional topological change latent variable sequence significantly improves the accuracy of first-hop aggregation layer weight adjustment under different topological change modes in channel attention mechanism input tests.

[0061] S3.2: Receive the topology change latent variable sequence and the original weight matrix of the first hop aggregation layer in the energy consumption balance control strategy model, calculate the contribution score of the topology change latent variable sequence to each weight channel based on the channel attention mechanism, so as to generate a channel importance mask matrix that characterizes the influence weight of different topology change modes on the aggregation operation.

[0062] The energy consumption balance control strategy model consists of four modules: input encoding module, first-hop aggregation layer, channel attention weight layer, and energy consumption prediction and regulation output layer.

[0063] 1. Module composition and network layers Input encoding module Number of layers: 2 fully connected neural networks.

[0064] Structure: The first fully connected layer maps the original node features (such as device power and load rate) into 128-dimensional latent vectors; the second fully connected layer maps the latent vectors into 64-dimensional encodings that match the dimension of the topological evolution feature vectors.

[0065] Activation functions: The first layer uses the hyperbolic tangent function, and the second layer uses the rectified linear unit.

[0066] Batch normalization: A batch normalization layer is added after each fully connected layer, with initial parameters set to mean 0 and variance 1.

[0067] First Jump Aggregation Layer Type: Graph convolutional aggregation, realizing first-order neighbor information aggregation.

[0068] Number of layers: Single layer, but includes a learnable weight matrix with a size of 64×64.

[0069] Aggregation method: Summation-normalized aggregation, outputting a 64-dimensional node feature matrix.

[0070] Correctable parameters: This weight matrix is ​​one of the targets of the local parameter correction coefficients in S3, and can be directly corrected by the channel importance mask matrix generated by S3.2.

[0071] Channel attention weight layer Structure: A multi-head self-attention mechanism based on a query-key-value structure, with the number of heads set to 4.

[0072] Number of layers: 1 Transformer encoder layer (excluding position encoding).

[0073] Internal parameters include query weight matrix, key weight matrix, value weight matrix (all 64×64), and learnable scaling factor and bias term.

[0074] Output: Attention-weighted feature matrix with the same dimensions as the input (64 × number of nodes).

[0075] Correctable parameters: The scaling factor and bias term in the attention weight layer, as well as some channels of the query, key, and value weight matrices, can all receive the adaptive attention adjustment coefficient vector generated by S3.3 and the local parameter correction coefficients generated by S3.5.

[0076] Energy consumption prediction and regulation output layer Number of layers: 3 fully connected neural networks.

[0077] Activation function: The intermediate layer uses rectified linear units, and the output layer has no activation function (linear output).

[0078] Output: Energy consumption regulation coefficient (or power allocation value) for each node, used for actual equipment control.

[0079] 2. Connection method Main connection: Input encoding module → first hop aggregation layer → channel attention weight layer → output layer, in sequence.

[0080] Residual connection: Add a residual connection between the first-hop aggregation layer and the channel attention weight layer. That is, first sum the output of the first-hop aggregation layer with the input of the channel attention weight layer, and then feed it into the attention layer after layer normalization.

[0081] Feedforward network supplement: A two-layer fully connected network (hidden units 64→256→64) is followed by the attention layer. The activation function is rectified linear unit, and a dropout ratio of 0.1 is added to enhance non-linear expressive power.

[0082] 3. Weight initialization method Fully connected layer: It adopts uniform initialization with Xavier and the gain coefficient is automatically calculated according to the input and output dimensions.

[0083] The first hop aggregation layer weight matrix is ​​orthogonally initialized to ensure the stability of the initial aggregation gradient.

[0084] Attention weight matrix (query, key, value): all initialized with a normal distribution, mean 0, standard deviation 0.02.

[0085] Scaling factor: Instead of being set as a learnable parameter, it is fixed as a value that depends on the dimension of the key vector; the bias term is initialized to 0.

[0086] Batch normalization layer: The scaling parameter is initialized to 1, and the offset parameter is initialized to 0.

[0087] 4. Training Process Training data: Extract topological evolution sequences and corresponding optimal energy consumption allocation labels from historical energy consumption operation data (labels are obtained through simulation or expert rules).

[0088] The loss function consists of three terms: the mean square error between the predicted energy consumption adjustment coefficient and the actual optimal coefficient, the energy consumption balancing loss (i.e., the variance of the load at each node), and a smoothing regularization term for parameter changes in adjacent time steps. These three terms are weighted and summed using a preset balancing coefficient.

[0089] Optimizer: The AdamW optimizer is used, with an initial learning rate of 0.001 and a weight decay coefficient of 0.0001.

[0090] Learning rate scheduling: A cosine annealing decay strategy is adopted, with a decay period of 50 training rounds.

[0091] Batch size: Each batch contains 32 topology snapshots.

[0092] Training rounds: A total of 200 rounds of training were conducted, and an early stop strategy was adopted (training was stopped if the loss did not decrease for 20 consecutive rounds).

[0093] Fine-tuning phase: After S3 generates local parameter correction coefficients, it no longer performs full training. Instead, it directly adjusts the corresponding parameters of the first-hop aggregation layer and attention weight layer (such as some elements of the weight matrix, scaling factors and bias terms) based on the correction coefficients to achieve fast adaptation.

[0094] The topology change latent variable sequence output from the preceding step S3.1 and the original weight matrix of the first hop aggregation layer of the backbone graph neural network are received as input objects. A weight channel mapping table is established to uniquely bind the elements of the latent variable sequence to the weight channels of the aggregation layer.

[0095] The core computational unit of the channel attention mechanism is invoked to perform linear transformation and normalization on the latent variable input of each channel, thereby obtaining the latent variable intensity vector of each channel, which serves as the preprocessing result for calculating the channel contribution.

[0096] During the contribution calculation phase, a scoring function based on Softmax normalization is set, and the contribution score corresponding to each weight channel is calculated using the following formula: Among them, a i S is the contribution score for the i-th channel. i Let be the latent variable intensity value of the i-th channel, and the denominator is the sum of the latent variable intensity indices of all channels.

[0097] The contribution score vector and the channel index matrix are fused element-wise to generate a preliminary importance matrix containing the weights of each channel's influence on the first-hop aggregation operation.

[0098] Threshold truncation and sparsification are performed on the initial importance matrix to remove channels with contributions below a preset threshold, thus avoiding the introduction of irrelevant or noise effects in subsequent weight mapping transformations.

[0099] After sparsification, the importance matrix index is rearranged and output as a channel importance mask matrix. This matrix numerically preserves the weight correction ratio of channels that make significant contributions and structurally corresponds to the weight channel arrangement encoding of the first hop aggregation layer.

[0100] By using a channel attention mechanism and a sparsity processing chain, the sequence of latent variables of topological changes is transformed into a channel importance mask matrix that can intuitively represent the influence weight of different topological change patterns on the aggregation operation, thereby enabling the backbone model to respond sensitively to local topological changes.

[0101] S3.3: Based on the query key-value pair structure of the attention weight layer in the channel importance mask matrix and energy consumption balance control strategy model, perform affine transformation operation to dynamically adjust the scaling factor and bias term in the attention mechanism to generate an adaptive attention adjustment coefficient vector that can respond to local connectivity shift.

[0102] Based on the query key-value pair structure of the attention weight layer of the channel importance mask matrix and energy consumption balance control strategy model, the mask matrix is ​​loaded as a reference signal for weight adjustment. The structural features of each query and key in the attention weight layer are analyzed and a parameter mapping index table is established to realize a one-to-one correspondence between channel contribution and attention parameter position.

[0103] The affine transformation operator is invoked to perform a linear combination operation on the scaling factor of each target site in the mapping index table. The mask matrix elements are multiplied by the query-key correlation coefficients and then appended with a reference bias value to form a preliminary adjusted signal matrix. The combination of the scaling factor and the bias term follows the affine transformation formula: Where A′ is the adjusted attention scaling factor matrix, A is the original attention scaling factor matrix, α is the dynamic scaling ratio matrix calculated based on the mask matrix, and β is the baseline bias term matrix.

[0104] The initial adjustment signal matrix is ​​normalized using a min-max normalization method to map all adjustment values ​​to a preset valid range.

[0105] The normalized adjustment coefficient vector is subjected to gating filtering. Low contribution coefficients are removed based on the local connectivity offset threshold. The set of remaining high contribution coefficients is transmitted to the corresponding query key-value pair position of the attention weight layer to replace the original scaling factor and bias term value to complete the dynamic adjustment.

[0106] By using affine transformation and normalization based on the channel importance mask matrix, the contribution of the topological change latent variable generated in the previous step is transformed into an adaptive adjustment coefficient vector that can be directly applied to the attention weight layer, thereby achieving the model's accurate attention response capability when local connectivity shifts.

[0107] S3.4: Perform weighted fusion processing on the channel importance mask matrix and the adaptive attention adjustment coefficient vector to eliminate parameter conflicts caused by single-dimensional correction and generate a unified-dimensional comprehensive parameter correction guidance tensor.

[0108] S3.5: Based on the comprehensive parameter correction guide tensor and the preset parameter update magnitude constraint threshold, perform nonlinear bounded mapping and scaling normalization to generate the local parameter correction coefficients. Specifically, the comprehensive parameter correction guide tensor is input to a nonlinear activation function for nonlinear mapping, and the mapping result is numerically normalized using the parameter update magnitude constraint threshold to convert the comprehensive correction signal into local parameter correction coefficients that conform to probability distribution characteristics, completing the end-to-end mapping from topological evolution features to model weight fine-tuning factors. In this embodiment, the nonlinear activation function is the sigmoid function. Of course, in other embodiments, other nonlinear activation functions such as the ReLU function can also be used.

[0109] The system receives the comprehensive parameter correction guide tensor and the parameter update amplitude constraint threshold as input objects. It then calls the nonlinear mapping module to perform sigmoid activation processing on each channel dimension, mapping the input tensor elements to the interval [0,1] to form a preliminary representation of the probability distribution characteristics. Based on the preliminary representation results, an element-by-element comparison is performed with the parameter update amplitude constraint threshold. A normalization operator is used to compress the differential components exceeding the range according to the threshold scaling factor, ensuring they meet the preset stability standard. Numerical normalization is performed on the threshold-constrained tensor. Max-min normalization or mean-variance normalization is used to eliminate the uneven response amplitude between different channels on a global scale, ensuring that each correction coefficient has a consistent statistical distribution. The normalized tensor is used as the probability weight coefficient, combined with the channel importance mask determined in the preceding structure-sensitive mapping process, to generate a local parameter correction coefficient matrix for the corresponding level of weight parameters. Through this chained processing, the comprehensive correction signal is transformed into a local weight fine-tuning factor that conforms to the probability distribution characteristics and satisfies the amplitude constraint, achieving an end-to-end mapping from topological evolution characteristics to factors that can be directly applied to model weights.

[0110] For example, in a smart city street light group dynamic topology adaptation scenario, the comprehensive parameter correction guide tensor has a value range of [-2.5, 3.8], and the preset parameter update amplitude constraint threshold is 0.6. When performing sigmoid mapping, the formula is: Where x is the guide tensor element value and y is the mapped output value. After mapping, the channel element values ​​fall into [0,1]. For elements with mapped values ​​greater than the threshold of 0.6, they are scaled by a scaling factor of 0.6 / y to obtain a matrix that satisfies the constraints. Max-min normalization is selected in the normalization stage. After normalization, the response amplitudes of each channel are consistent. The final generated local parameter correction coefficient matrix directly applies to the corresponding parameters of the first hop aggregation layer and attention weight layer of the backbone graph neural network, realizing smooth fine-tuning of model weights under dynamic topology changes. In the demonstration area test, this processing chain significantly improved the stability of energy consumption balance calculation under high-frequency topology fluctuations, and the power distribution error of the forward inference output remained within a low amplitude fluctuation range in continuous multiple updates.

[0111] like Figure 3 As shown, step S4: The corresponding level parameters of the energy consumption balance control strategy model are dynamically fine-tuned using the local parameter correction coefficients to construct an online adaptive energy consumption balance graph neural network model with real-time adaptability. Specifically, this includes: S4.1: Obtain the original weight matrix of the first hop aggregation layer and the attention weight layer in the energy consumption balance control strategy model, and extract the set of target parameters to be updated based on the original weight matrix to establish the scope of the operation object for dynamic fine-tuning.

[0112] Based on the application requirements of the local parameter correction coefficients generated in the preceding steps, the input object is the memory-resident weight structure and its hierarchical index information of the energy consumption balance control strategy model, and the execution object is the original weight matrix of the first-hop aggregation layer and the attention weight layer. The model structure description file is parsed to extract hierarchical metadata containing the weight matrix, generating a hierarchical parameter index table. The weight matrix reading interface is called to load the numerical matrices of the current version of the first-hop aggregation layer and the attention weight layer from the model storage unit according to the index, and precision unification processing is performed during the loading process to ensure consistent floating-point precision. In the weight matrix of the first-hop aggregation layer, the core weight components participating in the neighbor feature aggregation calculation are located through the mapping rules between the matrix dimension and the adjacency matrix structure, and this set is marked as the aggregation layer update target. In the attention weight layer, according to the parameter mapping relationship between the query key-value pair structure and the attention allocation formula, the scaling factor matrix and bias term vector participating in the weight allocation calculation are located and extracted, and they are marked as the attention layer update target. The two types of update targets are stored in the target parameter set cache, and a parameter component index mapping table is established for subsequent element-wise multiplication processing to support accurate matching between the input local parameter correction coefficients and the corresponding weight components. By using the above processing method, the application scope of the local parameter correction coefficients in the previous step is limited to the set of objects to be dynamically fine-tuned, so as to realize the update preparation only for the topology evolution sensitive weights.

[0113] For example, in a smart city street light group collaborative control system, the first-hop aggregation layer weight matrix of the energy consumption balance control strategy model has a dimension of 256×256, and the floating-point precision is unified to 32-bit single precision (FP32). The attention weight layer contains a scaling factor matrix of size 128×128 and a bias term vector of length 128. During this step, the model structure description file is parsed to generate a hierarchical parameter index table, where index [LayerID=1] corresponds to the first-hop aggregation layer and [LayerID=3] corresponds to the attention weight layer. The loading interface is called to obtain the two sets of weight matrices and perform FP32 precision conversion. In the first-hop aggregation layer, weight components involving newly added edge nodes are selected according to the adjacency matrix mapping relationship, totaling 512 elements as the aggregation layer update target; in the attention weight layer, the scaling factor matrix and each element of the bias term are located and extracted as the update target. An index mapping is established so that the local parameter correction coefficients can perform precise matching for each weight component in the target parameter set. This process ensures that dynamic fine-tuning only affects the weight components impacted by the newly added topology snapshot, thereby improving the computational efficiency and model stability of incremental adaptation.

[0114] S4.2: Apply the local parameter correction coefficients to the corresponding weight components in the target parameter set using element-wise multiplication to generate an intermediate correction weight matrix containing topological evolution information.

[0115] The system receives the local parameter correction coefficients generated in the previous steps as input data and maps them by index to the corresponding weight components of the first-hop aggregation layer and attention weight layer in the target parameter set. Element-wise multiplication is performed on the mapped correction coefficients and target weight components to obtain a weight change matrix reflecting the impact of topology evolution through element-wise numerical multiplication. The resulting weight change matrix is ​​numerically accumulated with the original weight matrix to embed the adjustment offset brought by the correction coefficients while maintaining the original baseline. The accumulated result is rearranged according to parameter position and index relationship to generate an intermediate corrected weight matrix with consistent structure and containing topology evolution information. Through element-wise multiplication accumulation and numerical rearrangement, the local parameter correction coefficients are effectively applied to the target weight components, transforming them into an intermediate corrected weight matrix that can be used for subsequent gradient clipping, achieving a direct mapping from topology change patterns to the parameter matrix.

[0116] For example, in a smart city street light group collaborative control application, the first-hop aggregation layer weight matrix is ​​set to a dimension of 128×256, the attention weight layer matrix is ​​set to a dimension of 256×256, and the dimension of the local parameter correction coefficient matrix perfectly matches the corresponding weight matrix. Elements in the correction coefficient matrix are multiplied element-wise by the target weight matrix according to their index to obtain the change matrix. Taking the weight value of 4.75 at position (12,45) in the aggregation layer as an example, the corresponding correction coefficient is 0.92, and the product result is 4.37. This change value replaces the original position value to form an accumulation matrix. An index mapping arrangement is performed on the accumulation matrix to generate a complete intermediate correction weight matrix, which serves as the input for subsequent gradient pruning. Verification shows that this matrix can significantly improve the response accuracy of energy consumption balance adjustment under dynamic topology in real-time inference of the control system.

[0117] S4.3: Perform gradient pruning on the intermediate corrected weight matrix, remove abnormal gradient components according to the preset stability constraint threshold, and output a smoothed stable corrected weight matrix to prevent parameter oscillations in the model during frequent topology changes.

[0118] In this embodiment, the gradient pruning process performed on the intermediate corrected weight matrix aims to address the potential instability in model training caused by frequent changes in the street light network topology, such as node additions / deletions and dynamic adjustments to connection relationships. Specifically, it involves checking the absolute value of each gradient component in the intermediate corrected weight matrix based on a preset stability constraint threshold, typically a small positive number, such as 1.0, or determined through a validation set. If the absolute value of a component exceeds the threshold, its sign is preserved, but its amplitude is truncated to the threshold value; otherwise, the original value is retained. This operation is equivalent to projecting the gradient vector into a sphere with a radius equal to the threshold, thus mathematically limiting the maximum step size of a single parameter update. The stable corrected weight matrix output after this processing retains the update direction, but the update amplitude is subject to smoothing constraints, thereby fundamentally suppressing the violent oscillations of internal model parameters caused by topological abrupt changes (i.e., gradient explosion), ensuring the numerical stability and convergence reliability of the incremental learning process.

[0119] S4.4: Based on the stable correction weight matrix, replace the corresponding original weight components in the energy consumption balance control strategy model, complete the in-situ update operation of the internal parameters of the energy consumption balance control strategy model, and construct an online adaptive energy consumption balance graph neural network model that reflects the current road network topology characteristics.

[0120] Using the stabilized and corrected weight matrix after gradient pruning as direct input, the model parameter management interface is called to locate the corresponding original weight component index positions of the first-hop aggregation layer and attention weight layer in the energy consumption balance control strategy model, establishing an operation mapping table for weight replacement. Utilizing memory pointer binding, the values ​​in the stabilized and corrected weight matrix are matched element-wise without creating temporary copies, ensuring that the replacement operation only applies to the parameter positions defined in the mapping table. An in-situ update mechanism is executed, directly overwriting the matching elements in the original weight matrix with the corresponding elements of the stabilized and corrected weight matrix, and a numerical verification module is called to perform layer-by-layer consistency checks on the updated parameter set to confirm the integrity of the weight data structure and the consistency of the matrix size. The updated parameter set, after consistency verification, is loaded, and the model's internal weight cache state is refreshed to ensure that subsequent inference calls are calculated based on the updated parameter values. Combining the structural information of the current topology snapshot, the topology feature identifier field in the model metadata is updated, enabling the online adaptive energy consumption balance graph neural network model to reflect the current road network topology characteristics. Through the above processing methods, the stabilized and corrected weight matrix from the previous step is transformed into an internal parameter set embedding real-time topology features, achieving an immediate adaptation effect to dynamic topology changes.

[0121] For example, in a city road network, the first-hop aggregation layer weight matrix is ​​set to 256×256, and the attention weight layer query matrix is ​​128×128. After gradient clipping, the stable correction weight matrix has elements in the range [-0.08, 0.12]. The replacement operation mapping table covers 64 channel parameters in the 256×256 matrix and 32 sets of query-key-value pair parameters in the attention weight matrix. During replacement, the positions of these parameters are locked through the mapping index, and the corresponding values ​​are directly updated using an in-situ overwrite method, avoiding additional memory consumption. After replacement, the consistency check module is called to perform the following parameter integrity judgment formula on each matrix: Where, p i For the updated parameter value, μ ref Here, N represents the historical reference value for the corresponding location, and C represents the total number of parameters in the matrix. valid This serves as the integrity assessment metric. If the calculated result of the integrity assessment metric is within the allowable threshold range, the parameter structure is confirmed to be correct. Through the execution of this embodiment, the updated online adaptive energy consumption balance graph neural network model, when loading the latest topology snapshot data for inference, can significantly improve the response stability and energy efficiency balance capability under conditions of frequent road construction and temporary light pole installations.

[0122] S4.5: Call the online adaptive energy consumption balance graph neural network model to perform forward reasoning on the current street light node status data, generate a preliminary energy consumption balance control strategy, and verify the model's real-time response capability and decision effectiveness after dynamic fine-tuning.

[0123] The system receives the online adaptive energy consumption balance graph neural network model updated in step S4.4 and the current street light node status data. Based on the street light node status data, a comprehensive feature tensor containing real-time power values, node location coordinates, and adjacency indexes is constructed, ensuring that the input dimension is consistent with the model's expected format. The comprehensive feature tensor is input to the model's input layer, and the first-hop aggregation layer is called to perform neighbor feature aggregation operations based on the fine-tuned weight matrix to generate primary node embedding vectors covering locally connected change regions. These primary node embedding vectors are input to the attention weight layer, where the weight distribution of each neighbor node is dynamically adjusted using adaptive adjustment coefficients, outputting a refined weighted node embedding representation. This weighted node embedding representation is then input to subsequent multi-layer graph convolution or message propagation modules to perform feature updates and nonlinear mapping operations across the entire graph or a specific region, obtaining terminal node feature vectors reflecting the current energy consumption distribution characteristics. The energy consumption balance calculation module is then called to calculate a preliminary energy consumption balance control strategy based on the terminal node feature vectors and a preset energy consumption optimization objective function, where the objective function can be expressed as: , Where p i Let p be the power allocation value for the i-th node in strategy P. ideal Let N be the ideal balanced power value for this node, N be the total number of nodes covered by the current topology snapshot, and P be the set of power allocation strategies to be solved. Through the above processing method, the parameter fine-tuning results of the previous step are transformed into energy consumption balance control strategies with verifiable real-time response capabilities, enabling the evaluation of the decision effectiveness of the dynamically fine-tuned model.

[0124] The specific implementation logic of the energy consumption balance control strategy closely revolves around the aforementioned objective function, aiming to reduce the power allocation value p of each node through dynamic power allocation. i With ideal equilibrium power p ideal The deviation between them. At the implementation level, this invention first calculates the real-time allocated power value p for each street light node. i The difference between the power output and the average power output of the entire network is used to identify abnormal nodes that are under high or low load. Subsequently, power reduction commands (such as lowering brightness levels or adjusting on / off times) are sent to high-load nodes, while power increase commands are sent to low-load nodes, thus guiding the overall network power distribution back to the mean. This process is not a simple linear increase or decrease, but rather combines the physical location coordinates and adjacency relationships of nodes, prioritizing power complementarity within local topology regions to reduce additional losses caused by cross-regional transmission.

[0125] Furthermore, to verify the immediate response capability after dynamic fine-tuning, the energy consumption balance control strategy also includes a closed-loop feedback mechanism. This application monitors in real time whether the adjusted power distribution meets the optimization conditions of the above objective function. If a significant deviation is detected, the online adaptive energy consumption balance graph neural network model will use the current node state data to perform forward inference again, generating a revised, refined control instruction. For example, for nodes that deviate from the ideal value due to a surge in lighting demand caused by sudden environmental changes such as a sudden downpour, the energy consumption balance control strategy will dynamically adjust their weights, allowing them to temporarily exceed the conventional power limit to ensure basic lighting, while further reducing the power in surrounding non-critical areas, thereby achieving overall energy consumption balance while ensuring user experience.

[0126] For example, during the evening rush hour, status data of 120 streetlight nodes within the current topological snapshot coverage area of ​​a city's main road are collected. Each node's features include real-time power value (in W), location coordinates (in meters), and neighbor node indices. The feature tensor is input to a finely tuned online adaptive energy balance graph neural network model. The first-hop aggregation layer, under the adjusted weights, aggregates only the neighbor features of nodes involved in topological changes, generating a 120-dimensional primary node embedding vector. The attention weight layer changes the neighbor node weight distribution through adaptive adjustment coefficients; for example, the weight of newly added edges is increased by 20% compared to the original energy balance control strategy model, while the weight of deleted edges is reduced by 15%, outputting a weighted node embedding representation. Subsequent graph convolutional layers perform message propagation and nonlinear mapping, transforming the embedding representation into terminal node feature vectors. The energy balance calculation module calculates the squared deviation between each node's power value and the ideal balanced power value according to the above formula, and averages the deviation under N=120 conditions, obtaining a regional energy balance error scalar of 35.6 (in W). 2 The initial control strategy for this region suggests reducing the power of some nodes by 2.3W and increasing it by 1.8W to improve overall balance. This strategy was used for consistency verification in the subsequent S5 step, verifying that the model's real-time response capability and decision-making effectiveness were within acceptable engineering ranges.

[0127] Step S5: Calculate the output deviation value of the historical stable topology sample data input into the online adaptive energy consumption balance graph neural network model and the original energy consumption balance control strategy model for inference verification. Specifically, this includes: S5.1: Obtain pre-stored historical stable topology sample data, and construct a standard adjacency matrix and node feature tensor based on the node connection relationships in the historical stable topology sample data to form a benchmark input data stream for model consistency verification.

[0128] In this embodiment, the acquisition of pre-stored historical stable topology sample data aims to construct a noise-free benchmark set for horizontal comparison of model performance. The specific process is as follows: Topology snapshot data marked as "stable" during long-term system operation is read from persistent storage. This topology snapshot data characterizes the inherent connection patterns of the network under normal conditions. Subsequently, based on the node connection relationships in the topology snapshot data, a standard adjacency matrix with fixed dimensions is constructed to accurately depict the physical or logical adjacency relationships between nodes in a binarized form. Simultaneously, multi-dimensional state features of each node at the corresponding stable moment, such as power, historical load, and device type encoding, are extracted and organized into a structured node feature tensor. Finally, the standard adjacency matrix and node feature tensor are aligned and packaged according to samples to form a benchmark input data stream.

[0129] S5.2: Input the reference input data stream into the energy consumption balance control strategy model without parameter correction and the online adaptive energy consumption balance graph neural network model after dynamic fine-tuning, respectively. Use the forward propagation algorithm to perform energy consumption balance prediction inference in parallel to generate historical reference energy consumption distribution vector and current adapted energy consumption distribution vector, respectively.

[0130] The baseline input data stream includes a standard adjacency matrix and node feature tensors constructed from historical stable topology sample data. The execution targets are an energy consumption balance control strategy model without parameter correction and an online adaptive energy consumption balance graph neural network model with dynamic fine-tuning.

[0131] The forward propagation operator of the energy consumption balance control strategy model is called on the baseline input data stream to sequentially complete the matrix multiplication of the adjacency matrix and the node feature tensor, normalization processing, and first-hop aggregation layer feature update, resulting in a set of historical baseline node latent vectors that do not include topology adaptation adjustments.

[0132] The set of historical baseline node latent vectors is input into the attention weight layer. The weight distribution between nodes is calculated according to the query key-value pair mechanism and a weighted aggregation operation is performed to generate a historical baseline energy consumption distribution vector that represents the energy consumption balance prediction result under stable topology conditions.

[0133] The forward propagation operator of the online adaptive energy-balancing graph neural network model is called on the same baseline input data stream to perform the same matrix operation and aggregation process as the backbone model. However, the updated local parameters are applied at the first hop aggregation layer and attention weight layer to obtain the current node latent vector set containing the effects of topology adaptation fine-tuning.

[0134] The current node latent vector set is processed by an attention weight layer to generate a current adaptive energy consumption distribution vector that reflects the inference results under stable topology conditions after dynamic fine-tuning.

[0135] By using the parallel inference processing method described above, the baseline input data stream from the previous step is transformed into a historical baseline energy consumption distribution vector and a currently adapted energy consumption distribution vector, thereby achieving direct comparability of prediction results between the old and new models in a stable topology scenario.

[0136] For example, in a smart city street light cluster system, a baseline input data stream is constructed using historical stable topology sample data containing 1000 nodes with an average degree of 4, where the node feature tensor dimension is 128. An energy consumption balance control strategy model is adopted, with the first-hop aggregation layer weight matrix having a size of 128×256 and the attention weight layer key-value pair dimension of 256×64. Forward propagation is performed on the baseline input data stream to obtain a historical baseline energy consumption distribution vector with a length of 1000. An online adaptive energy consumption balance graph neural network model is used, whose local parameter correction coefficients have been updated for topology changes. The magnitude of the change in the first-hop aggregation layer weight matrix follows the following formula: Where, Δ 2 Let w be the mean squared error of the local weight variation, M be the total number of elements in the weight matrix, and w be the mean squared error of the local weight variation. i and w i ′ These are the weights of the backbone model and the adapted model at the i-th position, respectively. In actual calculations, Δ 2 The squared value is significantly lower than the preset stability threshold. The baseline input data stream is input into the adaptation model, resulting in a current adaptation energy consumption distribution vector with the same length of 1000. The two models correspond one-to-one by node index, providing the necessary data foundation for subsequent deviation measurement. In this embodiment, under batch inference conditions, the average forward computation latency difference between the backbone model and the adaptation model is less than 5ms, which can significantly improve the accuracy of comparative analysis of inference efficiency under stable topology conditions.

[0137] S5.3: Perform element-wise difference operation on the historical baseline energy consumption distribution vector and the current adapted energy consumption distribution vector, and process the difference results using the absolute value norm calculation rule to generate an original deviation metric sequence that characterizes the prediction difference between the old and new models in a stable topology scenario.

[0138] Using the historical baseline energy consumption distribution vector and the currently adapted energy consumption distribution vector as inputs, element-wise difference operations are performed based on the aligned node index positions. The predicted energy consumption values ​​of each corresponding node are subtracted to obtain a node-level difference sequence. An absolute value mapping function is then called on this node-level difference sequence to remove the sign information of the differences, retaining only the magnitude values ​​to ensure that the deviation calculation is not affected by positive or negative directions. The node deviation sequence after absolute value processing is input to the norm calculation module, where a global scaling reduction operation is performed according to a preset norm type. The norm is selected as L1 or L2 to adapt to different model stability requirements. When using the L2 norm, square accumulation and square root operations are performed, with the mathematical expression as follows: Among them, y b,i y represents the predicted value of the i-th node in the historical baseline energy consumption distribution vector.a,i This represents the predicted value of the i-th node in the current adapted energy consumption distribution vector, where D is the original deviation metric and N is the total number of nodes. After the norm calculation is completed, the original deviation metric sequence representing the difference in predictions between the old and new models in a stable topology scenario is output. Through the above-mentioned differencing, absolute value, and norm processing methods, the result of the previous step is transformed into a quantitative deviation index that can be compared on a uniform scale, realizing a direct measurement of model consistency deviation under a stable topology.

[0139] S5.4: Based on the original deviation measurement sequence, smooth the local noise interference using moving average filtering, and extract the global convergence deviation index that reflects the overall drift trend of the model, so as to eliminate the misjudgment of stability assessment caused by single sampling fluctuations.

[0140] S5.5: Based on the global convergence deviation index, a scalar mapping transformation is performed in conjunction with the preset weight normalization coefficients to calculate the output deviation value between the final energy consumption balance control strategy model without parameter correction and the online adaptive energy consumption balance graph neural network model after dynamic fine-tuning, which serves as a direct quantitative basis for judging whether the topology consistency constraint threshold is exceeded.

[0141] Based on the global convergence deviation index input, the scalar mapping module is called to load the preset weight normalization coefficients to construct the mapping factor matrix, ensuring that the dimensions of different deviation indices are consistent and comparable.

[0142] The global convergence deviation index and the mapping factor matrix are multiplied element-wise to generate a normalized deviation quantization candidate sequence for subsequent numerical mapping calculations.

[0143] The normalized deviation quantization candidate sequence is processed by calling the nonlinear mapping function module to perform piecewise linear interpolation and smoothing coefficient compensation to improve the sensitivity of the deviation value when it is close to the constraint threshold.

[0144] Based on the processed mapping results, the cumulative weighted summation unit is called to calculate the overall deviation intensity, forming a new and old model output deviation metric value with a unified scalar representation.

[0145] The output deviation between the final energy consumption balance control strategy model without parameter correction and the dynamically fine-tuned online adaptive energy consumption balance graph neural network model is encapsulated into a direct quantization basis data package and transmitted to the topology consistency constraint judgment module for subsequent comparison calculations to determine whether the deviation exceeds the preset consistency constraint threshold.

[0146] Through the above multi-stage mapping and normalization process, the global convergence deviation index of the previous step is transformed into the output deviation value between the energy consumption balance control strategy model without parameter correction and the online adaptive energy consumption balance graph neural network model after dynamic fine-tuning. This enables quantifiable monitoring of changes in the model's generalization ability.

[0147] For example, in a smart city street light energy consumption balance control scenario, the system sets the global convergence deviation index to 2.7, the weight normalization coefficient to 0.85, and the mapping factor matrix to a single-element matrix of {0.85}. Element-wise multiplication yields a preliminary normalization result of 2.295. Piecewise linear interpolation is used to map the result to the sensitivity interval, and a smoothing coefficient of 0.05 is used to compensate, resulting in a corrected value of 2.345. The overall deviation intensity is calculated using a cumulative weighted summation unit, and the output scalar value of 2.345 is used as the output deviation between the uncorrected energy consumption balance control strategy model and the dynamically fine-tuned online adaptive energy consumption balance graph neural network model. When applied to the topology consistency constraint judgment module, this deviation value is lower than the constraint threshold of 3.0. The system confirms that the current online incremental update has not damaged the original model's generalization ability and proceeds to the subsequent lightweight supervision signal construction stage, verifying that the stability of the model's energy consumption control strategy is significantly improved when the daily topology change frequency is 50 times.

[0148] Step S6: If the output deviation value exceeds the preset topology consistency constraint threshold, a local rollback mechanism is triggered to restore the original historical parameter state of the energy consumption balance control strategy model and output control commands; if the output deviation value does not exceed the topology consistency constraint threshold, the online adaptive energy consumption balance graph neural network model is maintained and control commands are output. Specifically, this includes: S6.1: Obtain the pre-stored historical stable topology sample data set, and based on the node connection relationships and attribute features in the historical stable topology sample data set, construct a standard input tensor for verifying model consistency, so as to generate a benchmark dataset with spatiotemporal invariance.

[0149] Using historical stable topology sample data already recorded by the system as input, the data retrieval module is invoked to locate the corresponding data file by unique topology snapshot number and load it into processing memory, ensuring that the selected samples cover multiple time periods and that the stability of the topology structure meets preset judgment criteria. An adjacency matrix construction operation is performed on the node connection relationships in the historical stable topology sample data set. A sparse matrix storage method is used to compress empty connection redundancy, forming a global adjacency matrix baseline version with standard row and column indexes. Feature vector extraction and normalization are performed on the node attribute features in the historical stable topology sample data set. Node feature tensors are generated according to a preset attribute dimension order to ensure consistency in feature distribution among different topology samples. A matrix tensor concatenation method is used to combine the global adjacency matrix baseline version and the node feature tensors into a unified input structure, generating a multi-channel data tensor containing topology connection information and node state information. Spatiotemporal invariant feature encoding is performed, and the multi-channel data tensor is input to the time-series freezing module. Average pooling is performed on the time dimension, and dynamic attribute components are removed to ensure that the output benchmark dataset is invariant in both temporal and spatial structure. The above processing method transforms the results of the previous step into standard input tensors that can be directly used for model consistency verification, thereby achieving structured and standardized encapsulation of historical stable topological sample data.

[0150] S6.2: Input the benchmark test dataset into the energy consumption balance control strategy model and the online adaptive energy consumption balance graph neural network model respectively, and use the forward inference algorithm to perform energy consumption balance prediction calculation in parallel to obtain the historical benchmark energy consumption balance vector and the real-time incremental energy consumption balance vector respectively.

[0151] The benchmark dataset serves as a unified input object for parallel computing, containing a standard adjacency matrix and node attribute feature tensors constructed from historical stable topology sample data.

[0152] The energy consumption balance control strategy model inference engine is invoked to map the benchmark dataset to key computing units such as the first-hop aggregation layer and attention weight layer according to the internal data format, and an end-to-end forward propagation calculation is performed to obtain the historical benchmark energy consumption balance prediction result vector.

[0153] The online adaptive energy consumption balance graph neural network model inference engine is invoked to map the same benchmark dataset to the corresponding dynamically fine-tuned layer structure, perform forward propagation calculation, and obtain the real-time incremental energy consumption balance prediction result vector.

[0154] To ensure the synchronization of parallel computing, a unified batch indexing mechanism is adopted to start the inference process simultaneously within the two models, and the node feature tensors are subjected to the same normalization process to eliminate prediction bias caused by input differences.

[0155] During the inference process, the same energy consumption distribution reconstruction rule is used for the output layer of each model, and the obtained energy consumption prediction values ​​of each node are mapped to the energy consumption balance vector in the balance metric space so that subsequent difference calculations can be directly based on the energy consumption balance metric.

[0156] By using parallel inference and consistent preprocessing, the benchmark dataset is transformed into a directly comparable historical benchmark energy balance vector and a real-time incremental energy balance vector in two different models, achieving input consistency and result comparability for the model update effect.

[0157] For example, a smart city demonstration area selected a historical stable topology sample data set containing 400 street light nodes during low-traffic periods at night. The adjacency matrix was a sparse matrix of 400×400, and the node attribute feature tensor contained the average power and status identifier of each node over the past 30 minutes. After constructing this set into a benchmark dataset, it was input into an energy consumption balance control strategy model and an online adaptive energy consumption balance graph neural network model, respectively, and forward inference with a batch size of 64 was performed. Within both models, the node power data was min-max normalized according to the interval [0,300] watts to ensure that different power ranges were uniformly included in the aggregation calculation. The energy consumption balance vector output by the backbone model had a dimension of 400, and the values ​​were processed by balance degree mapping to form a historical benchmark energy consumption balance vector; the adaptation model output a real-time incremental energy consumption balance vector under the same input conditions. Both models used the same energy consumption distribution reconstruction rule: Among them, P i,t P is the normalized predicted power value of node i at time t. i,des Let be the normalized ideal equilibrium power value of node i, N be the total number of nodes, and E be the energy consumption equilibrium error value. After complete calculation, the energy consumption equilibrium vectors of both models are mapped to the same metric space to ensure the comparability of difference assessments. This synchronous inference result shows that, under the current topology, the adaptive model can quickly respond to changes and output an equilibrium distribution close to the historical benchmark, providing a clear quantitative basis for subsequent deviation calculation and consistency constraint determination.

[0158] S6.3: Based on the historical baseline energy consumption balance vector and the real-time incremental energy consumption balance vector, perform element-wise difference operation and norm normalization to calculate the topological consistency deviation scalar that characterizes the difference in output distribution between the old and new models.

[0159] S6.4: Obtain the preset topology consistency constraint threshold parameter, and perform a numerical comparison logic judgment between the topology consistency deviation scalar and the topology consistency constraint threshold parameter to generate a binary state judgment signal indicating the validity or failure of the model update.

[0160] Using the topology consistency deviation scalar output from the preceding sub-step S6.3 as the input object, the system configuration interface is called to read the preset topology consistency constraint threshold parameters and load them into the running memory, ensuring that the threshold data and the deviation scalar are in the same numerical precision and unit system for comparison.

[0161] The topology consistency deviation scalar and the topology consistency constraint threshold parameter are input into the numerical comparison and calculation unit. The difference between the two is calculated to obtain the basis for determining the deviation exceeding the limit. The difference calculation formula is as follows: Where P is the topology consistency deviation scalar, T is the topology consistency constraint threshold parameter, and Δ is the deviation margin value.

[0162] The deviation margin value is sent to the sign determination module, and the sign bit extraction and logical mapping processing are performed. When the sign bit is positive and the absolute value of the margin value exceeds zero, a failure indication status flag is generated. When the sign bit is negative or the absolute value of the margin value is equal to zero, a valid indication status flag is generated.

[0163] The aforementioned flag bits are encapsulated into a binary state determination signal, and a one-bit Boolean code is used to represent the validity of the model update. The encoding rules are given by a predefined mapping table, for example, a failure state is encoded as 1 and a valid state is encoded as 0, so that the downstream rollback mechanism or gradient update module can read it directly.

[0164] By using numerical comparison logic and encoding encapsulation processing, the deviation scalar result of the previous step is transformed into a binary state judgment signal that can directly trigger model rollback or continue to adapt and update, thereby realizing real-time and programmable model stability decision-making under dynamic topology changes.

[0165] S6.5: Based on the failure indication state in the binary state determination signal, call the parameter version management module to perform a historical parameter snapshot reading operation, and replace the key layer weight parameters of the energy consumption balance control strategy model with the historical parameter state before the update, so as to complete the local rollback recovery after abnormal topology adaptation.

[0166] Based on the failure indication state in the binary state determination signal, the snapshot index retrieval function of the parameter version management module is invoked to perform a key layer weight parameter version number matching operation on the stored historical parameter snapshot directory, locking the historical snapshot index corresponding to the update time of the current energy consumption balance control strategy model. A snapshot data read instruction is executed on the locked historical snapshot index, parsing all element values ​​of the first-hop aggregation layer and attention weight layer in the snapshot file and loading them into a temporary buffer in matrix form to ensure the integrity and losslessness of parameter values ​​during data reading. The weight matrices of the aggregation layer and attention weight layer in the temporary buffer are compared with the corresponding weight components in the current online adaptive energy consumption balance graph neural network model using a difference mapping calculation to generate a difference matrix containing numerical update magnitude information, providing a reference for parameter consistency verification during subsequent replacement. The parameter replacement interface is invoked to replace the weight components at the corresponding positions in the current model with the weight values ​​from the historical snapshot, ensuring no out-of-bounds errors in the difference matrix verification, and performing an in-situ update in memory to ensure that the key layer parameters of the replaced model are completely restored to their state before failure. The model consistency reload process is triggered, and the replaced energy consumption balance control strategy model is submitted to the inference engine for a global topology consistency verification to confirm that the rolled-back model state meets the preset stability and performance constraints. Through the above processing method, the binary state judgment result generated in the previous step is transformed into a data operation that completely rolls back the key layer weights, realizing the immediate guarantee of model stability recovery and energy consumption balance calculation accuracy after topology anomaly adaptation.

[0167] For example, in a smart city pilot area, the energy consumption balance control strategy model is deemed ineffective when the topology consistency deviation scalar is 0.085 and the preset threshold is 0.08. The parameter version management module stores key layer weight snapshot files within 48 hours. The first-hop aggregation layer weight matrix has a size of 128×128, and the attention weight layer query key-value matrix has a size of 128×64. When reading historical snapshot data, a snapshot file with version number 21045 generated 3 minutes before the update is retrieved. The average update magnitude of the difference matrix generated through difference mapping is 0.012. During the replacement process, all elements of the difference matrix satisfy the stability constraint that the magnitude is less than 0.05. After the replacement is completed, the global topology consistency verification module performs an energy consumption balance prediction deviation calculation based on the mean square error. Where y i Let y be the power value of the i-th node in the historical baseline energy consumption distribution vector. i ′This represents the power value predicted by the model for the i-th node after rollback, where n is the total number of nodes, and the summation sign covers all nodes. The mean square error of the verification results is significantly lower than before the failure determination, indicating that the rolled-back model has restored its high-fidelity energy balance calculation capability under unstable topology conditions and once again meets the real-time control accuracy requirements of field operation.

[0168] In one embodiment, a smart city street light group coordination control system is provided, comprising: a topology difference matrix acquisition module, a topology evolution feature vector acquisition module, a local parameter correction coefficient acquisition module, an online adaptive energy consumption balance graph neural network model construction module, an output deviation value acquisition module, and a model adjustment module, wherein: Topology difference matrix acquisition module: used to acquire node connection status change events of smart city street light groups within a continuous time window, and generate a topology difference matrix based on the node connection status change events; Topology evolution feature vector acquisition module: used in the pre-built topology evolution encoder to perform feature compression processing on the topology difference matrix using multi-layer sparse convolution and symbol-aware gating unit, and output topology evolution feature vector; Local parameter correction coefficient acquisition module: Based on the topology evolution feature vector and the key layer structure parameters of the preset energy consumption balance control strategy model, it performs a weight mapping transformation operation to generate local parameter correction coefficients for correcting the original energy consumption balance control strategy model; Online adaptive energy consumption balance graph neural network model construction module: used to dynamically adjust the energy consumption balance control strategy model using the local parameter correction coefficients, and construct an online adaptive energy consumption balance graph neural network model; Output deviation value acquisition module: used to calculate the output deviation value when historical stable topology sample data is input into the online adaptive energy consumption balance graph neural network model and energy consumption balance control strategy model for inference verification; Model adjustment module: If the output deviation value exceeds the preset topology consistency constraint threshold, a local rollback mechanism is triggered to restore the historical parameter state of the energy consumption balance control strategy model; if the output deviation value does not exceed the topology consistency constraint threshold, the online adaptive energy consumption balance graph neural network model is maintained.

[0169] This application provides a smart city street light group coordination control method, which can be applied to, for example... Figure 4In the application environment shown, terminal 102 communicates with server 104 via a network. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices, and server 104 can be a standalone server or a server cluster consisting of multiple servers.

[0170] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 5 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores data generated during the implementation of a smart city street light group coordinated control method. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a smart city street light group coordinated control method.

[0171] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor: S1: Obtain node connection status change events of smart city street light groups within a continuous time window, and generate a topology difference matrix based on the node connection status change events; S2: In the pre-built topology evolution encoder, multi-layer sparse convolution and symbol-aware gating units are used to perform feature compression processing on the topology difference matrix and output the topology evolution feature vector. S3: Based on the topology evolution feature vector and the key layer structure parameters of the preset energy consumption balance control strategy model, perform a weight mapping transformation operation to generate local parameter correction coefficients for correcting the original energy consumption balance control strategy model. S4: The energy consumption balance control strategy model is dynamically fine-tuned using the local parameter correction coefficients to construct an online adaptive energy consumption balance graph neural network model. S5: Calculate the output deviation value of inputting historical stable topology sample data into the online adaptive energy consumption balance graph neural network model and energy consumption balance control strategy model for inference verification; S6: If the output deviation value exceeds the preset topology consistency constraint threshold, a local rollback mechanism is triggered to restore the original historical parameter state of the energy consumption balance control strategy model and output control commands; if the output deviation value does not exceed the topology consistency constraint threshold, the online adaptive energy consumption balance graph neural network model is maintained and control commands are output.

Claims

1. A method for coordinated control of street light groups in smart cities, characterized in that, Specifically, it includes: S1: Obtain node connection status change events of urban street light groups within a continuous time window, and generate a topology difference matrix based on the node connection status change events; S2: Input the topological difference matrix into the pre-constructed topological evolution encoder, and use multi-layer sparse convolution and symbolic gate unit to perform feature compression processing on the topological difference matrix, and output the topological evolution feature vector. S3: Based on the topology evolution feature vector and the key layer structure parameters of the preset energy consumption balance control strategy model, perform a weight mapping transformation operation to generate local parameter correction coefficients for correcting the original energy consumption balance control strategy model. S4: The energy consumption balance control strategy model is dynamically fine-tuned using the local parameter correction coefficients to construct an online adaptive energy consumption balance graph neural network model. S5: Calculate the deviation value of the historical stable topology sample data input into the online adaptive energy consumption balance graph neural network model and the original energy consumption balance control strategy model for inference verification. S6: If the output deviation value exceeds the preset topology consistency constraint threshold, a local rollback mechanism is triggered to restore the original historical parameter state of the energy consumption balance control strategy model and output control commands; if the output deviation value does not exceed the topology consistency constraint threshold, the online adaptive energy consumption balance graph neural network model is maintained and executed and output control commands.

2. The smart city street light group coordinated control method according to claim 1, characterized in that, Step S3 specifically includes: Obtain the initialization configuration parameters of the predefined structure-sensitive weight mapper and the topology evolution feature vector, and use a fully connected neural network layer to perform nonlinear projection transformation on the topology evolution feature vector to output a sequence of latent variables of topology change; Receive the topology change latent variable sequence and the original weight matrix of the first hop aggregation layer in the energy consumption balance control strategy model, calculate the contribution score of the topology change latent variable sequence to each weight channel based on the channel attention mechanism, and generate a channel importance mask matrix. Based on the channel importance mask matrix and the query key-value pair structure of the attention weight layer in the energy consumption balance control strategy model, an affine transformation operation is performed to generate an adaptive attention adjustment coefficient vector. The channel importance mask matrix and the adaptive attention adjustment coefficient vector are weighted and fused to generate a comprehensive parameter correction guidance tensor; Based on the comprehensive parameter correction guide tensor and the preset parameter update magnitude constraint threshold, nonlinear bounded mapping and scaling normalization are performed to generate the local parameter correction coefficients.

3. The smart city street light group coordinated control method according to claim 2, characterized in that, The process of performing nonlinear bounded mapping and scaling normalization includes: The comprehensive parameter correction guide tensor is input into a nonlinear activation function for nonlinear mapping, and the mapping result is numerically normalized using the parameter update magnitude constraint threshold to generate the local parameter correction coefficients.

4. The smart city street light group coordinated control method according to claim 1, characterized in that, The energy consumption balance control strategy model in S3 includes four modules: input encoding module, first hop aggregation layer, channel attention weight layer, and energy consumption prediction and adjustment output layer.

5. The smart city street light group coordinated control method according to claim 1, characterized in that, Step S4 specifically includes: Obtain the original weight matrix of the first hop aggregation layer and the attention weight layer in the energy consumption balance control strategy model, and extract the target parameter set to be updated based on the original weight matrix; The local parameter correction coefficients are applied to the corresponding weight components in the target parameter set using element-wise multiplication to generate an intermediate correction weight matrix. Gradient clipping is performed on the intermediate corrected weight matrix to remove abnormal gradient components according to a preset stability constraint threshold, and a stable corrected weight matrix is ​​output. Based on the stable correction weight matrix, the corresponding original weight components in the energy consumption balance control strategy model are replaced, and the in-situ update operation of the internal parameters of the energy consumption balance control strategy model is completed, thus constructing an online adaptive energy consumption balance graph neural network model. The online adaptive energy consumption balance graph neural network model is invoked to perform forward reasoning on the current street light node status data to generate a preliminary energy consumption balance control strategy, so as to verify the model's real-time response capability and decision effectiveness after dynamic fine-tuning.

6. The smart city street light group coordinated control method according to claim 1, characterized in that, Step S5 specifically includes: Obtain pre-stored historical stable topology sample data, and construct a standard adjacency matrix and node feature tensor based on the node connection relationships in the historical stable topology sample data to form a benchmark input data stream; The reference input data stream is input into the energy consumption balance control strategy model without parameter correction and the online adaptive energy consumption balance graph neural network model after dynamic fine-tuning, respectively. The forward propagation algorithm is used to perform energy consumption balance prediction inference in parallel to generate historical reference energy consumption distribution vector and current adapted energy consumption distribution vector, respectively. Perform element-wise difference operations on the historical baseline energy consumption distribution vector and the current adapted energy consumption distribution vector, process the difference results using the absolute value norm calculation rules, and generate the original deviation measurement sequence. Based on the original deviation measurement sequence, local noise interference is smoothed by moving average filtering, and the global convergence deviation index is extracted. Based on the global convergence deviation index, a scalar mapping transformation is performed in conjunction with the preset weight normalization coefficients to calculate the output deviation between the final energy consumption balance control strategy model without parameter correction and the online adaptive energy consumption balance graph neural network model after dynamic fine-tuning.

7. The smart city street light group coordinated control method according to claim 1, characterized in that, The topology difference matrix includes: edge addition / deletion identifiers and node activity changes.

8. A smart city street light group coordination and control system, comprising: Topology difference matrix acquisition module: used to acquire node connection status change events of smart city street light groups within a continuous time window, and generate a topology difference matrix based on the node connection status change events; Topology evolution feature vector acquisition module: used in the pre-built topology evolution encoder to perform feature compression processing on the topology difference matrix using multi-layer sparse convolution and symbol-aware gating unit, and output topology evolution feature vector; Local parameter correction coefficient acquisition module: Based on the topology evolution feature vector and the key layer structure parameters of the preset energy consumption balance control strategy model, it performs a weight mapping transformation operation to generate local parameter correction coefficients for correcting the original energy consumption balance control strategy model; Online adaptive energy consumption balance graph neural network model construction module: used to dynamically adjust the energy consumption balance control strategy model using the local parameter correction coefficients, and construct an online adaptive energy consumption balance graph neural network model; Output deviation value acquisition module: used to calculate the output deviation value when historical stable topology sample data is input into the online adaptive energy consumption balance graph neural network model and energy consumption balance control strategy model for inference verification; Model adjustment module: If the output deviation value exceeds the preset topology consistency constraint threshold, it triggers a local rollback mechanism to restore the original historical parameter state of the energy consumption balance control strategy model and outputs control commands; if the output deviation value does not exceed the topology consistency constraint threshold, it maintains the use of the online adaptive energy consumption balance graph neural network model and outputs control commands.

9. A computer device, characterized in that, include: A processor, a memory, and a computer program stored in the memory and executable on the processor; the computer program, when executed by the processor, implements the steps of a smart city street light group coordinated control method as described in any one of claims 1 to 7.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when executed by a processor, implements the steps of a smart city street light group coordinated control method as described in any one of claims 1 to 7.

Images