Lightweight edge calculation time sequence prediction method for power transmission line monitoring

Abstract

The invention is suitable for the technical field of time sequence prediction, and provides a lightweight edge calculation time sequence prediction method for power transmission line monitoring, and the method comprises the steps: carrying out the preprocessing of obtained power transmission line monitoring data, and generating an enhanced time sequence data set with a spatio-temporal context; performing feature extraction on the enhanced time sequence data set to generate a wavelet coefficient matrix and a graph embedding vector; fusing the wavelet coefficient matrix, the static environment data and the graph embedding vector through an attention mechanism to generate a cross-modal time sequence feature vector; inputting the cross-modal time sequence feature vector into a preset hybrid prediction model for preliminary training, and performing random forest compensation and dynamic cutting to obtain a target hybrid prediction model; dividing the power transmission line topological graph into sub-graph areas, and performing local training on each node based on the target hybrid prediction model to generate a global optimization model; and a resource optimization instruction is generated according to a time sequence prediction value output by the global optimization model, so that the prediction precision and scene generalization are improved.

Description

Technical Field

[0001] The present invention relates to the field of time series prediction technology, and in particular to a lightweight edge computing time series prediction method for transmission line monitoring. Background Art

[0002] With the rapid development of smart grid and IoT technologies, power transmission line monitoring systems must process massive amounts of time-series data in real time to achieve fault warnings and load balancing. Traditional centralized cloud computing models struggle to meet real-time requirements due to high transmission latency and bandwidth constraints. Edge computing significantly reduces latency by processing data locally. However, power transmission line edge devices commonly face challenges such as limited computing resources and dynamically changing network environments.

[0003] Among current time series forecasting technologies, deep learning models based on LSTM and Transformer offer excellent accuracy, but their large number of parameters and high training energy consumption make them difficult to deploy directly on resource-constrained edge nodes. Traditional statistical models, while computationally lightweight, lack adaptability to non-stationary time series signals. Furthermore, existing methods often employ fixed model structures or centralized training, lacking adaptability to dynamic network topologies. This results in low cross-node coordination efficiency and a failure to incorporate the spatiotemporal correlation characteristics of transmission lines, resulting in low prediction accuracy and poor generalization across scenarios. Summary of the Invention

[0004] This application provides a lightweight edge computing time series prediction method for transmission line monitoring, which is used to solve the problems of low prediction accuracy and poor generalization with scenarios.

[0005] This application provides a lightweight edge computing time series prediction method for transmission line monitoring, including:

[0006] Acquire transmission line monitoring data, including dynamic time series data, static environmental data, and spatial topology data, and preprocess the monitoring data to generate an enhanced time series dataset with spatiotemporal context;

[0007] Performing feature extraction on the enhanced time series data set; performing dynamic wavelet transform on the dynamic time series data to generate a wavelet coefficient matrix; performing graph convolutional network processing on the spatial topology data to generate a graph embedding vector; fusing the wavelet coefficient matrix, static environment data, and graph embedding vector through an attention mechanism to generate a cross-modal time series feature vector;

[0008] Inputting the cross-modal time series feature vector into a preset hybrid prediction model for preliminary training, and performing random forest compensation and dynamic pruning to obtain a target hybrid prediction model;

[0009] The transmission line topology map is divided into sub-graph areas, and sub-graph training tasks are assigned to each edge node. Each node performs local training based on the target hybrid prediction model to generate a global optimization model.

[0010] A resource optimization instruction is generated according to the timing prediction value output by the global optimization model, which is used to drive the edge device to perform node wake-up and early warning upload operations.

[0011] Furthermore, performing dynamic wavelet transform on the dynamic time series data to generate a wavelet coefficient matrix includes:

[0012] Dynamically select the wavelet basis function according to the current fluctuation state of the transmission line;

[0013] The dynamic time series data is subjected to time-frequency decomposition based on the selected wavelet basis function, the time window length is set to the inverse of the current sampling frequency, the number of decomposition layers is set to at least three layers, and a corresponding wavelet coefficient matrix is ​​generated.

[0014] Furthermore, performing graph convolutional network processing on the spatial topology data to generate a graph embedding vector includes:

[0015] Construct an adjacency matrix of the transmission line topology graph, defining nodes as tower coordinates and edges as electrical connections between adjacent towers;

[0016] A two-layer cascaded graph convolutional network layer is adopted. The first layer aggregates the first-order neighbor node features, and the second layer aggregates the second-order neighbor node features. Each layer of the graph convolutional network uses the ReLU activation function and adopts mean pooling to generate the node-level graph embedding vector.

[0017] Furthermore, the wavelet coefficient matrix, static environment data and graph embedding vector are fused through the attention mechanism to generate a cross-modal time series feature vector, including:

[0018] The wavelet coefficient matrix is ​​flattened into a time-frequency feature vector according to the time window, the static environment data is encoded into the environment feature vector, and the graph embedding vector retains the topological correlation characteristics;

[0019] Generate query vector, key vector and value vector of time-frequency feature vector, environment feature vector and graph embedding vector based on the fully connected layer;

[0020] Calculating the dot product attention weight of the time-frequency feature vector and the environment feature vector, superimposing the topological association weight of the graph embedding vector, and normalizing by the softmax function to generate a fusion weight;

[0021] The time-frequency feature vector, the environment feature vector and the graph embedding vector are weightedly summed according to the fusion weight to generate a cross-modal time series feature vector.

[0022] Furthermore, the cross-modal time series feature vector is input into a preset hybrid prediction model for preliminary training, and random forest compensation and dynamic pruning are performed to obtain a target hybrid prediction model, including:

[0023] A spatiotemporal wavelet LSTM unit of a hybrid prediction model is constructed, the cross-modal time series feature vector is input, it is iteratively trained through a sliding time window based on a historical enhanced time series dataset, and a preliminary prediction sequence is output.

[0024] Furthermore, the step of inputting the cross-modal time series feature vector into a preset hybrid prediction model for preliminary training, and performing random forest compensation and dynamic pruning to obtain a target hybrid prediction model further includes:

[0025] Inputting the preliminary prediction sequence and static environment data into a preset random forest model to generate a residual prediction result;

[0026] The preliminary prediction sequence is modified based on the residual prediction result to generate a final time series prediction value, and a fault risk score is calculated.

[0027] Furthermore, the step of inputting the cross-modal time series feature vector into a preset hybrid prediction model for preliminary training, and performing random forest compensation and dynamic pruning to obtain a target hybrid prediction model further includes:

[0028] According to the energy distribution of the wavelet coefficient matrix, the neuron branches of the low-frequency components in the spatiotemporal wavelet LSTM unit are pruned;

[0029] Sparsifying the model parameter transmission links between edge nodes based on the topological distance of transmission lines;

[0030] The pruned hybrid prediction model is retrained to obtain the target hybrid prediction model.

[0031] Furthermore, the transmission line topology map is divided into sub-graph areas, a sub-graph training task is assigned to each edge node, and each node performs local training based on the target hybrid prediction model to generate a global optimization model, including:

[0032] Based on the electrical connection relationship and geographical distance of the transmission lines, the tower nodes in the topological map are clustered. The electrical connection weight between adjacent towers is inversely proportional to the geographical distance, and the sub-map area boundary is set as the breakpoint where the electrical connection weight is lower than the preset threshold.

[0033] Furthermore, the transmission line topology graph is divided into subgraph regions, a subgraph training task is assigned to each edge node, each node performs local training based on the target hybrid prediction model to generate a global optimization model, and further includes:

[0034] Each edge node exchanges the locally trained model parameters with adjacent nodes;

[0035] Based on the connectivity of the transmission line topology graph, the parameter aggregation weights between nodes are calculated through the graph attention mechanism, and the model parameters are weighted averaged to generate a global optimization model.

[0036] Furthermore, the resource optimization instruction is generated according to the timing prediction value output by the global optimization model, and is used to drive the edge device to perform node wake-up and early warning upload operations, including:

[0037] Generate node wake-up instructions, warning signal upload instructions, and sampling frequency adjustment instructions based on the timing prediction value and the fault risk score;

[0038] When the timing prediction value exceeds the preset load threshold, the node wake-up instruction is triggered to activate the dormant edge device; when the fault risk score reaches the preset risk level, the early warning signal upload instruction is triggered and the sampling frequency of the edge node in the corresponding area is adjusted.

[0039] It can be seen from the above technical solutions that this application has the following advantages:

[0040] This application uses dynamic wavelet transform to extract the time-frequency characteristics of non-stationary time series signals of transmission lines, combines graph convolutional networks to extract spatial topological associations, and uses cross-modal attention mechanisms to fuse multi-source data, thereby improving the prediction accuracy in complex scenarios; through topology-aware federated learning and dynamic clipping mechanisms, while ensuring the lightweight model, efficient collaborative training of multiple edge nodes is achieved, reducing computing and communication overhead; and finally outputs multi-dimensional prediction results and resource optimization instructions, effectively adapting to the dynamic edge environment of transmission line monitoring, and providing highly reliable and low-latency monitoring technology support for the power Internet of Things. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The present invention is a flowchart of an embodiment of a lightweight edge computing timing prediction method for transmission line monitoring. DETAILED DESCRIPTION

[0042] The terms "first," "second," "third," "fourth," etc. (if any) in the specification and claims of the present application and in the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequential sequence. It should be understood that the numbers used in this way are interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "corresponding to," and any variations thereof, are intended to cover non-exclusive inclusions, for example, a process, method, system, product, or apparatus comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products, or apparatus.

[0043] Example 1

[0044] The implementation method in this embodiment can be implemented in the system, can be implemented in the server, and can also be implemented in the terminal, without specific limitation. The following will introduce the lightweight edge computing time series prediction method for transmission line monitoring in this application from the perspective of system implementation. Figure 1 , the method provided in the embodiment of the present application includes the following steps:

[0045] S11. Acquire transmission line monitoring data, including dynamic time series data, static environmental data, and spatial topology data, and preprocess the monitoring data to generate an enhanced time series dataset with spatiotemporal context;

[0046] In this embodiment, during the data collection phase, edge sensors deployed on transmission line towers are used to acquire dynamic time series data of current, voltage, and temperature in real time. Conductor material, meteorological parameters (humidity, wind speed), and equipment aging coefficients are simultaneously collected as static environmental data. A spatial topology map with towers as nodes and electrical connections as edges is constructed based on the line design drawings. During preprocessing, the dynamic time series data is timestamp aligned and segmented with sliding windows. The spatial topology data is converted into an adjacency matrix and node coordinate vectors, and the static environmental data is normalized into a fixed-dimensional vector. By associating timestamps with tower location labels, the dynamic time series data is bound to the static environmental data and spatial topology data in the time-space dimension, generating an enhanced time series dataset with spatiotemporal labels for subsequent feature extraction and model training.

[0047] S12. Perform feature extraction on the enhanced time series dataset; perform dynamic wavelet transform on the dynamic time series data to generate a wavelet coefficient matrix; perform graph convolutional network processing on the spatial topology data to generate a graph embedding vector; and fuse the wavelet coefficient matrix, static environment data, and graph embedding vector through an attention mechanism to generate a cross-modal time series feature vector.

[0048] S121. Perform dynamic wavelet transform on dynamic time series data to generate a wavelet coefficient matrix, including the following:

[0049] 1. Dynamically select the wavelet basis function according to the current fluctuation state of the transmission line;

[0050] 2. Perform time-frequency decomposition on the dynamic time series data based on the selected wavelet basis function, set the time window length to the inverse of the current sampling frequency, set the number of decomposition layers to at least three, and generate the corresponding wavelet coefficient matrix.

[0051] When performing dynamic wavelet transform on dynamic time series data, the wavelet basis function is adaptively selected based on the current fluctuation state of the transmission line. When a significant increase in the current fluctuation amplitude is detected, such as from a lightning strike or a sudden load change, the Haar wavelet basis is used to capture the instantaneous high-frequency characteristics. When the current is stable, the DB4 wavelet basis is used to extract the periodic fluctuation pattern. The time window length is dynamically set based on the current data sampling frequency. For example, the window length is 1 second when the sampling frequency is 1 Hz, and the number of wavelet decomposition layers is set to 3 to 5. The layers are adjusted according to the signal complexity to ensure that the low-frequency components cover the line load trend and the high-frequency components capture the sudden noise. The wavelet coefficient matrix of the time-frequency energy distribution is generated. Through dynamic basis function selection and hierarchical decomposition, the time-frequency characteristics of the non-stationary transmission line signal are adapted, avoiding the spectrum leakage problem of the traditional Fourier transform.

[0052] S122. Perform graph convolutional network processing on the spatial topology data to generate a graph embedding vector, including the following:

[0053] 1. Construct an adjacency matrix for the transmission line topology, defining nodes as tower coordinates and edges as electrical connections between adjacent towers;

[0054] 2. A two-layer cascaded graph convolutional network is used. The first layer aggregates first-order neighbor node features, and the second layer aggregates second-order neighbor node features. Each layer of the graph convolutional network uses the ReLU activation function and adopts mean pooling to generate node-level graph embedding vectors.

[0055] When constructing the adjacency matrix of the transmission line topology, a weighted undirected graph is constructed using tower coordinates as nodes and electrical connection relationships, such as conductor impedance and physical connectivity, as edge weights. A two-layer cascaded graph convolutional network (GCN) layer is used: the first layer aggregates the features of each tower's first-order neighbors (directly connected towers) to extract local line status; the second layer aggregates the features of second-order neighbors (indirectly connected towers) to capture long-distance fault propagation correlations. Each GCN layer uses the ReLU activation function to enhance nonlinear expression capabilities and generates node-level graph embedding vectors through mean pooling. For example, each tower outputs a 128-dimensional vector, preserving topological spatial relationship characteristics. This step effectively models the fault propagation path of the transmission line and improves the accuracy of spatial correlation prediction.

[0056] S123. The wavelet coefficient matrix, static environment data, and graph embedding vector are fused through the attention mechanism to generate a cross-modal temporal feature vector, including the following:

[0057] 1. Flatten the wavelet coefficient matrix into a time-frequency feature vector according to the time window, encode the static environment data into the environment feature vector, and embed the graph vector to preserve the topological association characteristics;

[0058] 2. Generate query vector, key vector, and value vector of time-frequency feature vector, environment feature vector, and graph embedding vector based on the fully connected layer;

[0059] 3. Calculate the dot product attention weight of the time-frequency feature vector and the environment feature vector, superimpose the topological association weight of the graph embedding vector, and generate the fusion weight by normalizing it through the softmax function;

[0060] 4. Perform weighted summation of the time-frequency feature vector, environmental feature vector, and graph embedding vector based on the fusion weight to generate a cross-modal time series feature vector.

[0061] First, the wavelet coefficient matrix is ​​flattened into a time-frequency feature vector according to the time window (for example, a 10-second window is flattened into a 100-dimensional vector), and the static environmental data (such as material and humidity) is encoded as an environmental feature vector. The graph embedding vector retains the original topological dimension. The three types of features are mapped to the same semantic space through a fully connected layer to generate a query vector (Query), a key vector (Key), and a value vector (Value). The attention weights of the time-frequency features and the environmental features are calculated, and the topological association weights of the graph embedding vector are superimposed (for example, the weights of adjacent towers are doubled). The fusion weights are generated after softmax normalization. Finally, the three types of feature vectors are weighted and summed to generate a cross-modal time series feature vector (for example, 256 dimensions), realizing the joint modeling of time-frequency mutation signals, environmental physical constraints, and spatial topological dependencies.

[0062] Taking a sudden surge in current caused by a lightning strike on a transmission line as an example, step S12 first performs a dynamic wavelet transform on the current time series data. When a sudden change in current is detected (e.g., a momentary increase of 100A), the system automatically selects a Haar wavelet basis to capture the high-frequency pulse signal and decomposes it into a three-layer wavelet coefficient matrix, reflecting the instantaneous impact, short-term fluctuations, and long-term trends, respectively. Simultaneously, a topological graph is constructed based on the tower coordinates and conductor connections. A graph convolutional network is used to extract the electrical correlation features of adjacent towers. For example, a failure in tower A may affect downstream towers B and C, generating a graph embedding vector that represents the spatial propagation pattern. Finally, an attention mechanism is used to dynamically fuse the current time-frequency features, thunderstorm weather parameters (static environment), and topological features. For example, the weight of high-frequency wavelet coefficients during the lightning strike period is increased, and the weight of the topological features of towers near the fault point is enhanced. This generates a cross-modal feature vector that integrates temporal, spatial, and environmental information, providing accurate input for subsequent prediction models.

[0063] S13. Input the cross-modal time series feature vector into a preset hybrid prediction model for preliminary training, and perform random forest compensation and dynamic pruning to obtain a target hybrid prediction model;

[0064] In this embodiment, the initial training of the hybrid prediction model includes the following:

[0065] 1. Build a spatiotemporal wavelet LSTM unit for the hybrid prediction model, input a cross-modal time series feature vector, perform iterative training using a sliding time window based on a historical enhanced time series dataset, and output a preliminary prediction sequence.

[0066] Specifically, when constructing a spatiotemporal wavelet LSTM unit, a cross-modal time series feature vector (e.g., 256 dimensions) is input, and a sliding time window is used to train the historical enhanced time series dataset. The window length is dynamically adjusted based on the periodicity of the line load. For example, when the load fluctuation period is 24 hours, the window is set to 8 hours, and the data in each window is input into the LSTM network in chronological order. The LSTM unit is designed as a multi-layer cascade structure, with the number of neurons in each layer adapted to the computing power of the edge device. The long-term dependency between time-frequency features and spatial topology is learned through a gating mechanism, and a preliminary prediction sequence of current and temperature for future time periods (e.g., the next hour) is output. During training, an adaptive optimizer is used to dynamically adjust the learning rate to avoid gradient explosion or disappearance and ensure model convergence stability.

[0067] In this embodiment, random forest compensation includes the following:

[0068] 1. Input the preliminary prediction sequence and static environment data into the preset random forest model to generate residual prediction results;

[0069] 2. Based on the residual prediction results, the preliminary prediction sequence is modified to generate the final time series prediction value and calculate the fault risk score.

[0070] The initial prediction sequence and static environmental data (such as conductor material and meteorological parameters) are input into a random forest model to generate residual prediction results. The input features of the random forest include the predicted value, environmental parameters, and historical residual statistics (such as mean and variance). Multiple decision trees vote to output a residual correction. The initial prediction value is compensated based on the correction. For example, if the initial predicted current is 500A and the residual prediction is +20A, it is corrected to 520A. A fault risk score is then calculated. Based on the residual distribution and static environmental thresholds (such as the maximum current carrying capacity of the conductor and the critical humidity for icing), a probability density function is used to estimate the joint probability of faults such as overload and disconnection, and a risk score is output within a range of 0 to 1. For example, if the residual is continuously higher than twice the historical mean and the humidity exceeds the threshold, the risk score increases to above 0.9.

[0071] After the initial training of the hybrid prediction model, dynamic pruning is performed to adapt to edge computing resource constraints:

[0072] 1. Prune the neuron branches of low-frequency components in the spatiotemporal wavelet LSTM unit according to the energy distribution of the wavelet coefficient matrix;

[0073] Based on the energy distribution of the wavelet coefficient matrix, the energy contribution of each frequency band is calculated (low-frequency component energy contribution = sum of squares of low-frequency wavelet coefficients / total energy). If the low-frequency energy contribution exceeds a preset threshold, for example, if the low-frequency component covers the long-term trend of line load but has high information redundancy, the neuron branches processing the low-frequency component in the spatiotemporal wavelet LSTM unit are pruned. The pruning strategy uses gradient importance sorting to remove neuron connections with low prediction contribution. For example, the 20% of connections with the smallest absolute weights are pruned, while retaining the high-frequency feature processing branches to capture sudden fluctuations.

[0074] 2. Sparsifying the model parameter transmission links between edge nodes based on the transmission line topology distance;

[0075] Based on the physical distance between towers in the transmission line topology, if the distance between adjacent towers is less than a preset distance threshold, only adjacent edge nodes are allowed to exchange key model parameters, such as the output gate weights of LSTM units and important decision tree nodes in random forests. Parameter transmission links between non-adjacent nodes are directly blocked. Sparsification rules are dynamically generated from the adjacency matrix of the topology to ensure that the communication load matches the spatial correlation of the lines.

[0076] 3. Retrain the pruned hybrid prediction model to obtain the target hybrid prediction model.

[0077] The pruned model is retrained, using knowledge distillation to use the original model's output distribution as soft labels to constrain the lightweight model and minimize prediction error. During retraining, the pruned neuron branches are frozen, and only the parameters of the retained components are updated to ensure that the model complexity is compatible with the computing power of the edge device.

[0078] Taking a winter ice-covered transmission line scenario in a mountainous area as an example, a spatiotemporal wavelet LSTM system used cross-modal features to predict the line load for the next three hours, initially predicting 600A. A random forest algorithm, combined with the current temperature (-5°C), humidity (90%), and ice thickness, predicted a residual error of +50A (due to increased resistance caused by ice). After correction, the final prediction was 650A, and an ice-covered fracture risk score of 0.85 was calculated. The system detected that the low-frequency energy in the wavelet accounted for 70%, pruning 30% of the low-frequency neurons in the LSTM reduced computational effort by 40%. Furthermore, by exchanging parameters with only three adjacent towers based on topological distance, communication overhead was reduced by 50%. Ultimately, the system triggered a high-risk alert and activated collaborative monitoring among surrounding nodes, detecting potential line outages two hours in advance.

[0079] S14. Divide the transmission line topology into subgraph regions, assign subgraph training tasks to each edge node, and perform local training on each node based on the target hybrid prediction model to generate a global optimization model;

[0080] In this embodiment, subgraph area division and task allocation include the following:

[0081] Based on the electrical connection relationship and geographical distance of the transmission lines, the tower nodes in the topological map are clustered. The electrical connection weight between adjacent towers is inversely proportional to the geographical distance, and the sub-map area boundary is set as the breakpoint where the electrical connection weight is lower than the preset threshold.

[0082] Specifically, the connection weights between tower nodes are constructed based on the electrical connection relationship of the transmission line (such as conductor impedance, current transmission capacity) and the geographical distance. ij The calculation formula is:

[0083]

[0084] Where: D ij is the geographical distance between towers i and j, Z ij is the conductor impedance, and α and β are balance coefficients (set according to the line design specifications). Connections with shorter geographical distances and lower impedances have higher weights.

[0085] A spectral clustering algorithm is used to partition the tower nodes in the topology graph. A Laplacian matrix is ​​constructed, and node similarity is calculated based on connection weights. Subgraph region boundaries are set to require connection weights below a preset threshold. For example, edges with weights < 0.1 are considered weak links and serve as subgraph segmentation points. Multiple subgraph regions are output to ensure high electrical connectivity between nodes within each subgraph (e.g., clusters of adjacent towers). Each edge node is assigned a training task for a subgraph region, which matches the edge node's physical location. For example, a subgraph region represents a tower subgraph within the power supply radius of a substation.

[0086] In this embodiment, federated collaborative training and global optimization include the following:

[0087] 1. Each edge node exchanges the locally trained model parameters with adjacent nodes;

[0088] Each edge node performs local training based on the target hybrid prediction model (from step S13), iteratively updating model parameters using the augmented time series dataset within the subgraph region. After training, the following parameters are extracted: the hidden layer weight matrix of the spatiotemporal wavelet LSTM unit; and the key decision tree splitting node of the random forest compensator. These parameters are only allowed to be exchanged between edge nodes in adjacent subgraph regions to reduce communication redundancy.

[0089] 2. Based on the connectivity of the transmission line topology graph, the graph attention mechanism is used to calculate the parameter aggregation weights between nodes, and the model parameters are weighted averaged to generate a global optimization model.

[0090] The parameter aggregation weight between nodes is calculated through the graph attention mechanism (GAT), where the attention coefficient is calculated based on the difference in model parameters between adjacent nodes i and j:

[0091] e ij =LeakyReLU(a T [Wh i ||Wh j ])

[0092] Where: h i and h j is the model parameter vector of nodes i and j, W is the learnable weight matrix, and a is the attention vector.

[0093] Attention coefficients are normalized using softmax to obtain aggregate weights. Globally optimized model parameters are generated by weighted summing of adjacent node parameters. These aggregated parameters are synchronized to all edge nodes, replacing local model parameters, completing a round of federated training. Iterations are repeated until the model converges, as indicated by a loss function change rate below 1e-5.

[0094] Taking a provincial-level main transmission network as an example: 1. Subgraph partitioning: Calculate the connection weights of 500 tower nodes, with a geographical distance range of 1 to 50 kilometers and an impedance range of 0.1 to 5Ω; set a weight threshold of 0.1 to divide the network into 20 subgraph regions (each containing 25 highly correlated towers); assign subgraph training tasks to 20 edge nodes (for example, node A is responsible for the eastern coastal subgraph). 2. Parameter exchange and aggregation: After node A is trained, it exchanges LSTM hidden layer weights with its neighboring node B (responsible for the adjacent inland subgraph); calculate the parameter aggregation weights of nodes A and B using the graph attention mechanism (for example, A's weight is 0.7, B's weight is 0.3); generate global model parameters and synchronize them to all nodes, improving the ability to predict cross-regional load fluctuations. 3. Effectiveness verification: During the typhoon, the global optimization model reduced the load forecast error in the eastern coastal and inland areas from 12% to 6%, reducing communication overhead by 40%.

[0095] This step achieves distributed and efficient training of the transmission line monitoring model through electrical-geographic joint weighted sub-graph partitioning and topology-aware federated aggregation, retains the learning ability of local features, improves the accuracy of cross-regional collaborative prediction through global optimization, and significantly reduces the communication and computing power pressure of edge computing nodes.

[0096] S15. Generate resource optimization instructions based on the timing prediction value output by the global optimization model to drive the edge device to perform node wake-up and warning upload operations.

[0097] In this embodiment, generating resource optimization instructions includes the following:

[0098] 1. Generate node wake-up instructions, warning signal upload instructions, and sampling frequency adjustment instructions based on the timing prediction value and fault risk score;

[0099] 2. When the time series prediction value exceeds the preset load threshold, the node wake-up instruction is triggered to activate the dormant edge device; when the fault risk score reaches the preset risk level, the early warning signal upload instruction is triggered and the sampling frequency of the edge node in the corresponding area is adjusted.

[0100] Specifically, the system dynamically generates an instruction set based on the time series prediction values ​​(such as current, temperature) and fault risk scores output by the global optimization model: when the prediction value exceeds the preset load threshold, which is set according to the line's rated current carrying capacity and historical peak value, the system triggers the node wake-up instruction to activate the dormant edge device to enhance the monitoring density. When the fault risk score reaches the preset risk level, which is divided into levels such as low, medium, and high based on historical fault data and expert experience, the system triggers the early warning signal to be uploaded to the central server and increases the sampling frequency of the edge nodes in the corresponding area (such as from 1Hz to 5Hz) to capture high-frequency abnormal signals. For example, if the predicted current of a line exceeds the threshold by 10% and the icing risk score is "high risk", the system wakes up the backup node within 10 kilometers and increases the sampling rate to 10Hz. The real-time monitoring data is transmitted back through the optimized communication link, reducing the fault warning response time by 40%.

[0101] The above embodiment significantly improves the prediction accuracy and real-time performance of transmission line monitoring by integrating time-frequency features, spatial topology and multimodal data, and combining lightweight models with dynamic resource optimization, while reducing the energy consumption and communication overhead of edge computing nodes.

[0102] It is understandable that those skilled in the art can, under the guidance of the above embodiments, combine various implementation methods in the above embodiments to obtain technical solutions of multiple implementation methods.

[0103] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A lightweight edge computing time series prediction method for transmission line monitoring, characterized in that: include: Acquire transmission line monitoring data, including dynamic time series data, static environmental data, and spatial topology data, and preprocess the monitoring data to generate an enhanced time series dataset with spatiotemporal context; Performing feature extraction on the enhanced time series data set; performing dynamic wavelet transform on the dynamic time series data to generate a wavelet coefficient matrix; performing graph convolutional network processing on the spatial topology data to generate a graph embedding vector; fusing the wavelet coefficient matrix, static environment data, and graph embedding vector through an attention mechanism to generate a cross-modal time series feature vector; Inputting the cross-modal time series feature vector into a preset hybrid prediction model for preliminary training, and performing random forest compensation and dynamic pruning to obtain a target hybrid prediction model; The transmission line topology map is divided into sub-graph areas, and sub-graph training tasks are assigned to each edge node. Each node performs local training based on the target hybrid prediction model to generate a global optimization model. A resource optimization instruction is generated according to the timing prediction value output by the global optimization model, which is used to drive the edge device to perform node wake-up and early warning upload operations.

2. The lightweight edge computing time series prediction method for transmission line monitoring according to claim 1 is characterized in that: The performing of dynamic wavelet transform on the dynamic time series data to generate a wavelet coefficient matrix includes: Dynamically select the wavelet basis function according to the current fluctuation state of the transmission line; The dynamic time series data is subjected to time-frequency decomposition based on the selected wavelet basis function, the time window length is set to the inverse of the current sampling frequency, the number of decomposition layers is set to at least three layers, and a corresponding wavelet coefficient matrix is ​​generated.

3. The lightweight edge computing time series prediction method for transmission line monitoring according to claim 1 is characterized in that: The performing graph convolutional network processing on the spatial topology data to generate a graph embedding vector includes: Construct an adjacency matrix of the transmission line topology graph, defining nodes as tower coordinates and edges as electrical connections between adjacent towers; A two-layer cascaded graph convolutional network layer is adopted. The first layer aggregates the first-order neighbor node features, and the second layer aggregates the second-order neighbor node features. Each layer of the graph convolutional network uses the ReLU activation function and adopts mean pooling to generate the node-level graph embedding vector.

4. The lightweight edge computing time series prediction method for transmission line monitoring according to any one of claims 1 to 3, characterized in that: The method of fusing the wavelet coefficient matrix, static environment data, and graph embedding vector through the attention mechanism to generate a cross-modal temporal feature vector includes: The wavelet coefficient matrix is ​​flattened into a time-frequency feature vector according to the time window, the static environment data is encoded into the environment feature vector, and the graph embedding vector retains the topological correlation characteristics; Generate query vector, key vector and value vector of time-frequency feature vector, environment feature vector and graph embedding vector based on the fully connected layer; Calculating the dot product attention weight of the time-frequency feature vector and the environment feature vector, superimposing the topological association weight of the graph embedding vector, and normalizing by the softmax function to generate a fusion weight; The time-frequency feature vector, the environment feature vector and the graph embedding vector are weightedly summed according to the fusion weight to generate a cross-modal time series feature vector.

5. The lightweight edge computing time series prediction method for transmission line monitoring according to claim 4 is characterized in that: The cross-modal time series feature vector is input into a preset hybrid prediction model for preliminary training, and random forest compensation and dynamic pruning are performed to obtain a target hybrid prediction model, including: A spatiotemporal wavelet LSTM unit of a hybrid prediction model is constructed, the cross-modal time series feature vector is input, it is iteratively trained through a sliding time window based on a historical enhanced time series dataset, and a preliminary prediction sequence is output.

6. The lightweight edge computing time series prediction method for transmission line monitoring according to claim 5 is characterized in that: The step of inputting the cross-modal time series feature vector into a preset hybrid prediction model for preliminary training, performing random forest compensation and dynamic pruning to obtain a target hybrid prediction model further includes: Inputting the preliminary prediction sequence and static environment data into a preset random forest model to generate a residual prediction result; The preliminary prediction sequence is modified based on the residual prediction result to generate a final time series prediction value, and a fault risk score is calculated.

7. The lightweight edge computing time series prediction method for transmission line monitoring according to claim 6 is characterized in that: The step of inputting the cross-modal time series feature vector into a preset hybrid prediction model for preliminary training, and performing random forest compensation and dynamic pruning to obtain a target hybrid prediction model further includes: According to the energy distribution of the wavelet coefficient matrix, the neuron branches of the low-frequency components in the spatiotemporal wavelet LSTM unit are pruned; Sparsifying the model parameter transmission links between edge nodes based on the topological distance of transmission lines; The pruned hybrid prediction model is retrained to obtain the target hybrid prediction model.

8. The lightweight edge computing time series prediction method for transmission line monitoring according to claim 1 is characterized in that: The transmission line topology graph is divided into subgraph areas, a subgraph training task is assigned to each edge node, each node performs local training based on the target hybrid prediction model, and generates a global optimization model, including: Based on the electrical connection relationship and geographical distance of the transmission lines, the tower nodes in the topological map are clustered. The electrical connection weight between adjacent towers is inversely proportional to the geographical distance, and the sub-map area boundary is set as the breakpoint where the electrical connection weight is lower than the preset threshold.

9. The lightweight edge computing time series prediction method for transmission line monitoring according to claim 8, characterized in that: The method further includes dividing the transmission line topology graph into subgraph areas, assigning subgraph training tasks to each edge node, and each node performing local training based on the target hybrid prediction model to generate a global optimization model. Each edge node exchanges the locally trained model parameters with adjacent nodes; Based on the connectivity of the transmission line topology graph, the parameter aggregation weights between nodes are calculated through the graph attention mechanism, and the model parameters are weighted averaged to generate a global optimization model.

10. The lightweight edge computing time series prediction method for transmission line monitoring according to claim 1, characterized in that: Generating a resource optimization instruction based on the timing prediction value output by the global optimization model, which is used to drive the edge device to perform node wake-up and early warning upload operations, includes: Generate node wake-up instructions, warning signal upload instructions, and sampling frequency adjustment instructions based on the timing prediction value and the fault risk score; When the timing prediction value exceeds the preset load threshold, the node wake-up instruction is triggered to activate the dormant edge device; when the fault risk score reaches the preset risk level, the early warning signal upload instruction is triggered and the sampling frequency of the edge node in the corresponding area is adjusted.

Images