A power line carrier robust dynamic networking method combined with deep learning

Abstract

This invention discloses a robust dynamic networking method for power line carrier communication that incorporates deep learning. Sites extract channel features using lightweight convolutional neural networks, while the central coordinator employs an attention-enhanced long short-term memory network to predict channel states. Optimal routes and dual-mode switching strategies are dynamically generated based on deep reinforcement learning. The central coordinator performs millisecond-level topology reconstruction, achieving network-wide synchronization through special synchronization beacon frames. The system automatically and incrementally updates model parameters daily, continuously optimizing network performance. Experimental results show that this method significantly improves the robustness of power line communication networks in complex noise environments, drastically reduces burst noise packet loss rate, and significantly shortens topology response time, providing highly reliable communication assurance for smart grids.

Description

Technical Field

[0001] This invention belongs to the field of power line carrier communication technology, specifically relating to a dynamic networking method integrating a deep learning model to improve the robustness and adaptability of low-voltage power line dual-mode communication networks in complex noise environments. This method is particularly suitable for high-reliability communication scenarios such as advanced measurement systems for smart grids and distributed energy monitoring, solving the communication quality degradation problem caused by the time-varying characteristics of power line channels. Background Technology

[0002] Power line carrier communication faces severe challenges in smart grid applications, mainly due to the inherent high dynamic characteristics of power line channels (including signal-to-noise ratio fluctuations caused by time-varying impulse noise, impedance changes caused by load switching, and inter-symbol interference caused by multipath effects). Traditional networking strategies have limitations such as long static route update cycles, inability to adapt to millisecond-level topology changes, large false handover rates due to fixed signal-to-noise ratio threshold mechanisms, and large network reconstruction delays due to single-point failures in centralized decision-making architectures. Existing improvement schemes, such as game theory-based route optimization, have high computational complexity and rely on ideal models. Dynamic time-division multiple access technology does not solve the timing loss caused by sudden noise, and Kalman filter-based channel prediction models have large errors. At the same time, standardized periodic route evaluation mechanisms are difficult to respond to sudden failures in a timely manner, and fixed proxy coordinator (PCO) selection strategies ignore real-time link quality, leading to network bottlenecks. Summary of the Invention

[0003] This invention proposes a robust dynamic networking method for power line carriers that combines deep learning. The architecture first collects power line channel data in real time through stations (STAs), performs preliminary feature extraction locally, and then uploads it to the Central Coordinator (CCO) to complete the channel awareness phase. Subsequently, in the decision execution phase, the Central Coordinator uses an advanced deep learning model to analyze the processed information and generate the optimal decision instructions. Finally, the Central Coordinator accurately executes network topology reconstruction.

[0004] The core innovations of its technical solution are mainly reflected in two aspects: In channel prediction, it innovatively integrates Convolutional Neural Networks (CNN) and Long Short-Term Memory Networks (LSTM) to construct a hybrid model architecture, introducing three modes as input features: time-frequency domain noise spectrum, impedance gradient value, and historical packet error rate. Simultaneously, it employs an attention mechanism to dynamically focus on key burst noise features, with strict control over weight allocation bias, significantly improving prediction accuracy. In the decision-making mechanism, it constructs a multi-objective optimization strategy based on deep reinforcement learning, specifically designs a priority experience replay mechanism to accelerate the network's learning efficiency for key topology change events, and establishes a unique dual-mode switching decision tree structure, effectively reducing the risk of mode oscillations caused by sudden environmental changes during network operation. This architecture achieves high-precision real-time perception and intelligent, adaptive dynamic networking decision-making for complex power line channel environments through deep learning methods, providing a systematic solution to the dynamic challenges faced by traditional power line communication networks. Attached Figure Description

[0005] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are merely embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In the drawings:

[0006] Figure 1 This is a flowchart of the method of the present invention.

[0007] Figure 2 This is a flowchart of channel state prediction according to an embodiment of the method of the present invention. Detailed Implementation

[0008] Step 1: Hardware System Initialization. The site configuration utilizes a high-speed processor in conjunction with a power line carrier communication module, and integrates an impedance detection unit to achieve real-time acquisition of channel parameters. The central coordinator employs a high-performance computing platform with multi-core processing units, and is configured with ample high-speed storage space to meet the computational needs of deep learning models and the storage strategy mapping table.

[0009] Step 2: Channel Data Acquisition and Preprocessing. Synchronous sampling begins 100 microseconds after the power frequency zero-crossing point to avoid zero-crossing noise interference. Time-domain voltage waveform acquisition uses a 100 MHz sampling rate to obtain 1024 data points, with 14-bit quantization precision to ensure signal integrity. The local preprocessing process includes: performing a Fast Fourier Transform on the original time-domain waveform to generate spectral data, calculating the power spectral density, extracting impedance change gradient features, and finally constructing a multi-dimensional feature vector.

[0010] Step 3: Feature Extraction Processing. A lightweight convolutional neural network model is deployed on the site, employing a three-layer one-dimensional convolutional structure: the first layer contains 16 5×5 convolutional kernels, the second layer has 32 3×3 convolutional kernels, and the third layer has 64 3×3 convolutional kernels. The model undergoes a 32-bit floating-point to 8-bit integer quantization conversion, with precision loss controlled within 0.5%. The feature extraction process is completed on the main frequency processor, outputting a 128-dimensional compressed feature vector.

[0011] Step 4: Channel State Prediction. The central coordinator runs a Long Short-Term Memory (LSTM) network model with an attention mechanism to process the feature data. The model first extracts temporal features through a bidirectional LSM layer, then the attention layer calculates the feature weights for each time step and performs weighted fusion. Finally, a channel state score ranging from 0 to 1 is output through a fully connected layer. The attention weight calculation uses a combination of hyperbolic tangent activation function and soft maximum normalization to ensure that burst noise features receive higher weight allocation.

[0012] Step 5: Constructing the Decision State Space. A four-dimensional state space matrix is ​​constructed: the first dimension is the inter-node predicted signal-to-noise ratio matrix, converting channel state scores into decibel values; the second dimension is the queue load rate vector, reflecting the buffer occupancy of each node; the third dimension is the link stability matrix, using an exponential decay factor attenuation function; and the fourth dimension is the topology connectivity matrix, with binary values ​​representing the connectivity status between nodes. The state space is updated every 5 milliseconds to ensure real-time reflection of network state changes.

[0013] Step 6: Reinforcement Learning Decision Making. A dual-deep network algorithm is used to construct the decision engine, with the state space dimension being four times the square of the number of nodes. The action space is designed with 256 discrete actions, encoded using 8-bit binary codes: the high 4 bits represent relay node selection, the middle 2 bits control communication mode switching, and the low 2 bits determine the time slot allocation ratio. The reward function design incorporates three objectives: maximizing throughput, minimizing hop count, and a link stability penalty term. Action selection adopts an ε-greedy strategy, with a 90% probability of selecting the action with the maximum Q value and a 10% probability of exploratory random selection.

[0014] Step 7: Prioritized Experience Replay. A dual-buffer system is established to store experience data, prioritizing it based on the temporal difference error value. A high-priority buffer stores key topology change events, while a regular buffer stores routine communication data. The sampling probability is proportional to the 0.6 power of the temporal difference error, ensuring that important experiences receive higher training frequencies. The training process uses mini-batch gradient descent to update the policy network parameters, with the target network updated synchronously every 1000 iterations.

[0015] Step 8: Network Topology Switching Prediction. Before the central coordinator issues the reconfiguration command, it performs a topology switching simulation based on the output of a deep reinforcement learning policy network: First, it simulates the stability of communication links within 500ms under the new topology, calculating key performance indicators (including predicted signal-to-noise ratio margin, end-to-end delay distribution, and the proportion of potential conflict time slots); second, it uses a pre-trained graph neural network model to dynamically evaluate the topology switching success rate (real-time calculation of node synchronization failure probability, automatically triggering a decision rollback mechanism when the probability threshold exceeds 5%); finally, it generates a three-dimensional prediction matrix containing the switching time window (accurate to ±2ms), mode switching sequence, and fault avoidance contingency plan, ensuring that the reconfiguration operation has forward-looking risk control capabilities.

[0016] Step 9: Topology Reconfiguration Execution. After receiving the policy instructions, the central coordinator performs a three-stage operation: First, it updates the routing table, recording the relationship between the target node and the next hop and the communication mode; second, it broadcasts a beacon frame carrying a topology change flag; and third, it uses a special synchronization word to ensure that nodes accurately identify the frame's start position. Clock synchronization between nodes uses a precise time protocol to achieve microsecond-level synchronization accuracy.

[0017] Step 10: Online Model Optimization. The daily model maintenance process begins with: first, collecting 24 hours of runtime data to build an incremental dataset; then, freezing the convolutional neural network parameters and fine-tuning the long short-term memory network and reinforcement learning policy network with a learning rate of 10 to the power of -4; next, using cross-validation to evaluate model performance, requiring an accuracy of at least 85%; and finally, deploying the updated model to the central coordinator via a secure channel.

[0018] This invention provides a robust dynamic networking method for power line carrier communication that combines deep learning. It significantly outperforms traditional methods in prediction accuracy, keeping channel state scoring errors to a low level. Real-time decision-making is effectively guaranteed through a deep reinforcement learning model, fully meeting the rapid response requirements of communication systems. Resource utilization efficiency is greatly improved by the introduction of a dual-mode intelligent switching mechanism, effectively compressing redundant data transmission. Simultaneously, the system design possesses excellent industry compatibility, fully supporting the current State Grid power communication industry standards, ensuring seamless deployment of the technical solution in practical application scenarios. Overall, it achieves a comprehensive performance improvement in prediction capability, response speed, resource optimization, and standard compatibility.

[0019] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1.A power line carrier robust dynamic networking method combined with deep learning, characterized in that Includes the following steps: Step 1: The site collects the time-domain waveform, impedance value, and historical packet error rate parameters of the power line channel in real time; Step 2: The station performs convolution operations on the collected time-domain waveforms, impedance values ​​and historical packet error rate parameters through a three-layer one-dimensional convolutional neural network architecture, extracts local features, and outputs a 128-dimensional compressed feature vector to the central coordinator. Step 3: The central coordinator receives the compressed feature vector and runs a long short-term memory network model with attention mechanism to predict the channel state and obtain the channel state score as the predicted signal-to-noise ratio; the model extracts temporal features through bidirectional long short-term memory layers, and the attention mechanism uses a combination of hyperbolic tangent activation function and soft maximum normalization to calculate the feature weights at each time step, so that burst noise features receive higher weight allocation. Step 4: Construct a four-dimensional state space matrix, where the first dimension is the inter-node predicted signal-to-noise ratio matrix obtained by converting the channel state score, the second dimension is the queue load rate vector reflecting the buffer occupancy of each node, the third dimension is the link stability matrix using an exponential decay factor, and the fourth dimension is the topology connection matrix representing the connectivity state between nodes. Step 5: Using the four-dimensional state space matrix as the state input, a deep reinforcement learning policy network is used to generate routing selection, communication mode switching, and time slot allocation decisions. The output is a discrete action encoded in 8 bits, where the high 4 bits represent relay node selection, the middle 2 bits control communication mode switching, and the low 2 bits determine the time slot allocation ratio. Step 6: The central coordinator parses the discrete actions and performs a topology reconstruction operation, which includes: first updating the routing table, and then broadcasting a beacon frame carrying a topology change flag. Step 7: The node synchronously switches the communication mode according to the communication mode switching type and time slot allocation ratio contained in the beacon frame; Step 8: Store system experience data into a dual-priority experience buffer, where the high-priority buffer stores experience data on key topology change events; Step 9: Update the deep learning model parameters incrementally every day. During the update, the experience in the dual-priority experience buffer is sampled and replayed for training based on the sampling probability that is proportional to the power of 0 or 6 of the time-series difference error.

Images