Communication transmission rate adaptation and flow control method and system for underground power distribution room

Abstract

The invention discloses an underground power distribution room communication transmission rate adaptation and flow control method and system. The system comprises a multi-system communication module, a reinforcement learning relay module, a self-adaptive rate adaptation module, an intelligent flow control module and an energy efficiency optimization module. The output of the multi-system communication module is connected with the reinforcement learning relay module, the reinforcement learning relay module processes the data and transmits the data to the adaptive rate adaptation module, the data is returned to the multi-system communication module for transmission through the intelligent flow control module, and the energy efficiency optimization module provides power supply and energy consumption management support. According to the method, the communication stability and reliability of the underground power distribution room are remarkably enhanced, the transmission efficiency and the resource utilization rate are optimized, the intelligent adaptive capacity is improved, the bottleneck of green energy conservation and self-maintenance is broken through, the method is compatible with multi-system standards, the adaptability is high, and the method is suitable for popularization and application. And efficient, low-consumption and intelligent underlying communication support is provided for an intelligent power grid and a ubiquitous power internet of things.

Description

Technical Field

[0001] The present invention relates to the field of power communication technology, and in particular to a method for communication transmission rate adaptation and flow control in an underground distribution room. Background Art

[0002] In modern power systems, underground distribution rooms serve as the core hub for power distribution and management, and the stability and efficiency of their communication systems are crucial. Traditional single-mode communication technologies have significant limitations: medium-voltage carrier communications are susceptible to interference from power line harmonics and pulse noise, resulting in severe signal attenuation and difficulty ensuring data integrity over long distances. While LoRa technology offers the advantages of low power consumption and wide-area coverage, its transmission rate is only tens of kbps, failing to meet the high-speed demands of real-time equipment status monitoring and rapid fault diagnosis. While 5G-RedCap technology offers high-speed transmission, in closed and complex underground distribution rooms, obstacles such as walls and equipment exacerbate signal penetration losses, leading to numerous communication blind spots and making full coverage difficult.

[0003] At the same time, the unique operating environment of underground distribution rooms further exacerbates communication challenges. Strong electromagnetic interference generated by densely deployed equipment can severely impact wireless signal quality, causing signal distortion and increased bit error rates. The complex structure of cable trenches and metal pipes creates multipath effects, leading to signal reflection and refraction, causing transmission delays and packet disarray. The long-term effects of a humid, high-temperature environment accelerate the aging of communication equipment and reduce its reliability. Traditional communication systems lack dynamic adaptability and are unable to adjust communication strategies in a timely manner to complex and changing environments. This leads to unstable communication links, posing a serious threat to the safe operation of the power system.

[0004] With the advancement of smart grid construction, underground distribution rooms have placed higher demands on communication systems. Massive amounts of power monitoring data and equipment status information need to be transmitted quickly and accurately. The traditional fixed-rate transmission mode cannot adapt to the dynamic changes in network load, which can easily lead to bandwidth resource waste or congestion. At the same time, the key services of power control signals and fault alarm signals lack a priority guarantee mechanism, and real-time transmission cannot be ensured when the network is congested. In addition, traditional communication equipment has high energy consumption and low energy utilization. In the context of advocating green energy conservation, innovative technologies are urgently needed to achieve high-efficiency and low-power operation of communication systems. Therefore, the development of communication technologies that integrate multi-standard signals and have intelligent adaptation and flow control functions has become the key to solving the communication problems in underground distribution rooms. Summary of the Invention

[0005] The technical problem to be solved by the present invention is that, in order to address the problems of underground distribution room communication being susceptible to interference, low efficiency and high energy consumption, a method for communication transmission rate adaptation and flow control in underground distribution rooms is provided. By integrating medium-voltage carrier, LoRa and 5G-RedCap multi-standard modules to build a link matrix and intelligently switch signals, an improved reinforcement learning model is used to optimize the relay path, the transmission rate is dynamically adapted with the help of a generative adversarial network, a dual closed-loop fuzzy PID algorithm is used to control the flow, and energy sensing is combined to adjust the power and self-power supply, thereby reducing energy consumption and improving communication stability, efficiency and intelligence.

[0006] The present invention provides a method for communication transmission rate adaptation and flow control in an underground distribution room, comprising the following steps:

[0007] S1, dynamic fusion of multi-standard signals:

[0008] By integrating multi-standard communication modules including medium-voltage carrier, LoRa, and 5G-RedCap, a three-dimensional communication link matrix is constructed. The weight coefficients of medium-voltage carrier, LoRa, and 5G-RedCap are α1, α2, and α3, respectively. A dynamic weight allocation algorithm based on the weight coefficients α1 and α3 enables parallel transmission and intelligent switching of multi-standard signals.

[0009] S2, reinforcement learning relay path selection:

[0010] An improved reinforcement learning model is designed to construct a state transition matrix including a spatial distance attenuation factor of 0.78 / km, a time delay compensation coefficient of 0.96 / ms, and an interference intensity threshold of -90dBm. The model dynamically selects relay nodes and optimizes relay paths to reduce transmission delay to less than 50ms. The improved reinforcement learning model includes a spatiotemporal feature fusion module, a data encoding module, a spatiotemporal Bayesian network module, a dynamic channel switching module, and a multi-node interaction module.

[0011] S3, adaptive rate adaptation:

[0012] An improved generative adversarial network (GAN) was introduced to build a traffic prediction model. Combining real-time network load and service priority, the modulation and coding module was dynamically adjusted to achieve hierarchical adaptation of the transmission rate, with a rate adjustment step size of 1 Mbps.

[0013] S4, intelligent flow control command output:

[0014] A dual-closed-loop flow control architecture is designed. The outer loop dynamically adjusts the congestion window based on a fuzzy PID algorithm, while the inner loop reserves bandwidth for power control signals through priority queue management.

[0015] S5, energy efficiency optimization:

[0016] An energy-aware power control algorithm is designed. By combining the remaining power of the device and the location of the relay node, which triggers the low-power mode when the power threshold is ≤20%, the transmission power is dynamically adjusted. Energy harvesting technology is used to achieve self-powering of the communication equipment, reducing the system energy consumption by more than 30%.

[0017] Preferably, the steps of constructing the three-dimensional communication link matrix in step S1 are:

[0018] S1-1, through the quantum state parameter normalization mechanism, the transmission bandwidth, bit error rate and wall penetration loss of the three standards of medium voltage carrier, LoRa and 5G-RedCap are mapped to the quantum bit space. The quantum measurement operator is used to realize the non-deterministic normalization of multi-dimensional parameters, forming a time-space asymmetric tensor matrix of number of nodes × number of nodes × number of standards;

[0019] S1-2: A quantum Bayesian cognitive engine is introduced to estimate channel quality in real time. A Nash equilibrium solution space for multi-standard resource allocation is constructed based on quantum game theory. The initial weight coefficients are α1 = 0.38, α2 = 0.25, and α3 = 0.29. When the SNR is ≥ 20 dB, the 5G-RedCap weight α3 is increased to 0.45 through the quantum entanglement exchange mechanism. When the SNR is < 20 dB, the LoRa weight α2 is dynamically adjusted to 0.35.

[0020] S1-3 uses a non-Hermitian quantum evolution model to describe link state changes, and uses the PT symmetric phase transition detection mechanism to find the optimal switching time in real time and switch the signal to the optimal standard.

[0021] Preferably, the workflow of the spatiotemporal feature fusion module in step S2 includes:

[0022] S2-1 uses a quantum entanglement clock synchronization protocol to achieve ultra-high-precision synchronization of ≤100ns. The packet transmission delay and relay node processing delay are encoded as quantum phase differences. The time delay compensation coefficient of 0.99 / ms is obtained through the quantum state time evolution equation and measurement.

[0023] S2-2 uses the cyberspace gravitational model and non-Euclidean distance metrics, combined with the quantum wall-penetrating attenuation model, to map node energy and bandwidth to gravitational field strength. Through the curvature tensor and quantum tunneling effect, the node spatial distance is accurately calculated and spatial characteristics are formed.

[0024] S2-3, through the quantum entanglement exchange network and quantum Fourier transform, realizes the non-local correlation fusion of spatiotemporal characteristics in the quantum frequency domain, and uses quantum principal component analysis to adaptively extract key feature components; based on the quantum reinforcement learning framework and quantum perception hash algorithm, combined with the quantum neural network, predicts channel interference in advance. When the RSSI monitors that the interference intensity exceeds the threshold of -90dBm, the intelligent optimal switching of multi-standard signals can be completed within 5ms.

[0025] Preferably, the workflow of the data encoding module in step S2 includes:

[0026] S2-4 uses a quantum superposition state coding algorithm to map the medium voltage carrier voltage signal to the quantum Bloch sphere, uses quantum phase coding to encode LoRa's spreading factors SF7-SF12, and uses quantum amplitude modulation to process 5G-RedCap's modulation coding strategy MCS0-MCS16. Combined with ternary quantum entanglement coding, it forms a 128-dimensional quantum feature vector.

[0027] S2-5, using the space-time spinor representation and geometric algebra, the latitude and longitude of the sensor, altitude coordinates, and the temperature, voltage, and load rate of the equipment operating status are converted into geometric algebra elements with physical meaning for fusion;

[0028] S2-6 introduces quantum fractional Fourier transform, realizes flexible conversion of signal characteristics in time and frequency domain by adaptively adjusting the transformation order, and uses quantum entanglement exchange network to establish non-local correlation of multi-standard signals; at the same time, adopts quantum Lorentz mapping and quantum measurement entropy normalization to enhance feature discrimination based on chaotic system and quantify feature uncertainty.

[0029] Preferably, the workflow of the spatiotemporal Bayesian network module in step S2 includes:

[0030] S2-7, using quantum tensor product networks to embed spatial distance, time delay, and interference intensity into quantum Hilbert space, using quantum phase estimation algorithms to parallelly calculate three-dimensional state transition probabilities, and constructing non-local causal relationships between nodes based on quantum entanglement exchange mechanisms;

[0031] S2-8 uses a hidden Markov model in the form of a quantum superoperator, combined with quantum Monte Carlo sampling and error correction coding, to predict the delay distribution within the next five time windows in an exponential state space;

[0032] S2-9, in terms of path optimization, a space-time curvature routing algorithm is designed based on gravitational wave theory, which converts node energy and load into space-time curvature tensor. The shortest path is solved by the Dijkstra algorithm optimized by quantum annealing, and the quantum field theory path integral model is used for calculation.

[0033] Preferably, the workflow of the dynamic channel switching module in step S2 includes:

[0034] S2-10 uses a hybrid modulation architecture assisted by quantum stochastic resonance, integrating frequency hopping frequency spectrum (FHSS) and direct sequence spread spectrum (DSSS), and introduces quantum tunneling effect to enhance signal anti-interference capability;

[0035] S2-11, a dynamic decision-making model is constructed based on the quantum annealing reinforcement learning framework and the biological immune memory mechanism. When the SNR is less than 20dB, the quantum parallel search algorithm is triggered to globally scan the multidimensional channel space and simultaneously initiate the biological chemotaxis path optimization, shortening the channel switching cycle from 500ms to microseconds. Quantum entanglement is used to establish non-local correlation of channel states, enabling pre-evaluation and pre-configuration of backup channels.

[0036] S2-12 adopts the quantum genetic Q-learning algorithm, encodes multiple switching strategies through quantum bit superposition states, and simulates the dynamic optimization strategy of biological gene evolution process; the module has a built-in adaptive environmental perception system and uses the quantum perception hash algorithm to quickly match the optimal strategy.

[0037] Preferably, the multi-node interaction module in step S2 adopts the FedAvg algorithm framework, and its workflow includes:

[0038] S2-13, the multi-node interaction module builds a quantum entanglement federated-biological symbiotic network architecture, which innovates the traditional federated learning framework. Each relay node encodes local training parameters into quantum entangled states, and achieves secure parameter aggregation through quantum teleportation and remote preparation protocols, fundamentally protecting data privacy.

[0039] S2-14, uses the principle of biological symbiosis to build a defense mechanism, maps the parameter vector into an antigen-antibody reaction pattern, and resists malicious attacks through immune memory and clonal selection; innovates the quantum differential privacy protection mechanism, uses the unpredictability of quantum noise to enhance privacy protection, and reduces the risk of data leakage to 1 / 1000 of the traditional method under a privacy budget of ε=0.5; proposes a quantum game reinforcement learning framework, in which each node establishes non-local associations through quantum entanglement, achieves a Pareto optimal strategy without sharing private information, and designs a quantum state reward function to balance global optimization and local interests.

[0040] Preferably, the implementation method of the improved generative adversarial network GAN in step S3 is:

[0041] S3-1, Generator Construction and Implementation

[0042] The generator is constructed using a hybrid attention mechanism-enhanced LSTM network, embedding a spatiotemporal module and a self-attention module. The spatiotemporal module uses multi-head attention to capture the long-term spatiotemporal dependencies of power flow data, while the self-attention module focuses on key features. The number of neurons in the three hidden layers is 128, 64, and 32, respectively. The gating and attention mechanisms work together to process complex temporal relationships.

[0043] The system inputs one hour of multi-source heterogeneous traffic data, including power control, equipment monitoring, environmental sensing, weather warning, and grid topology dynamics data. It applies multimodal feature fusion encoding, builds semantic associations using knowledge graphs, maps the data into a unified feature space, and collaboratively optimizes the encoding model through federated learning, scaling it to the [0, 1] range.

[0044] The output of the generator is the traffic forecast value for the next 10 minutes. The improved composite loss function is used in the training process, and the formula is:

[0045] L gen =α(t)*L mes *1+λ spatial *S adj )+β(t)*L TV *(1+λ temporal *T corr )+γ(t)*L adv *(1+λ hybrid *H factor ) (1)

[0046] In formula (1), L mes is the mean square error, which measures the difference between the predicted value and the true value; L TV is the total variation loss, which is used to smooth the prediction results and prevent overfitting; L adv To combat the loss and encourage the generator to generate more realistic data, α(t), β(t), and γ(t) are dynamic weight coefficients, which are optimized through dynamic adjustment algorithms and adaptively changed according to the training stage; λ spatial ,λ temporal ,λ hybrid S is a newly added adjustment parameter that controls the influence of space, time and mixed factors on the loss function; adj L is a spatial correlation factor that considers the spatial correlation between the traffic data in the prediction area and the surrounding areas, explores the potential connection in the spatial dimension, and improves the rationality of the prediction; adv is the time correlation factor, which captures the dependency of traffic data in the time series and optimizes the prediction by using the time correlation between history and future. factor It is a mixed correlation factor that comprehensively considers the impact of various complex factors of time and space on traffic, so that the model can cope with more complex traffic change scenarios; the LSTM network parameters θ are optimized through the back propagation algorithm gen,Set the prediction error target to ≤5%. When the average prediction error on the validation set exceeds this threshold, the network structure and parameters are quickly adjusted in combination with the meta-learning mechanism;

[0047] S3-2, Discriminator Construction and Implementation

[0048] The discriminator uses a multi-scale feature pyramid convolutional neural network (MSF-PCNN), which includes four convolutional layer groups, three pyramid pooling layers, and two fully connected layers. Each convolutional layer group contains parallel convolutional layers with different convolution kernel sizes to extract multi-scale local features. The pyramid pooling layer aggregates features at different granularities to obtain a more comprehensive feature representation, and finally the fully connected layer makes classification decisions. The improved Wasserstein-f divergence metric is used to measure the distribution difference between the predicted traffic and the actual traffic. The formula is:

[0049]

[0050] In formula (2), f(.) is a customized nonlinear mapping function that enhances the sensitivity to data distribution differences; P data is the true data distribution, P gen is the generated data distribution, Π(P data ,P gen ) is the set of all possible joint distributions between the two, γ is a joint distribution sampled from the joint distribution set, and (x, y) is a pair of data sampled from γ; when training the discriminator, a quantum computing acceleration mechanism is introduced, and the discrimination results of multiple samples are calculated in parallel using quantum superposition states. The RMSProp optimizer with adaptive learning rate is used to update the discriminator parameters θ. dis , by maximizing W f (P data ,P gen ), so that the discriminator can distinguish the distribution differences between the two more accurately;

[0051] The discriminator outputs a three-category classification, including determining whether the input traffic data is "real data," "generated data," or "suspected anomaly." Combining a causal inference algorithm, the discriminator further analyzes the causal relationship between data identified as "suspected anomaly" and other related data to determine whether it is a true abnormal traffic pattern. A comprehensive recognition accuracy target of ≥95% is set, and the discriminator's performance is evaluated on a test set. If the target is not achieved, the network structure and parameters are automatically adjusted through reinforcement learning.

[0052] S3-3, Generative Adversarial Training

[0053] Initialize the network parameters of the generator and discriminator separately, using a quantum-inspired parameter initialization method that utilizes the superposition and entanglement properties of quantum states to generate better initial weights, and initialize the bias term to 0;

[0054] During the training process, a dynamic task priority alternating training strategy is adopted; the ratio of the number of training rounds of the generator and discriminator is dynamically adjusted according to the current training stage and data characteristics;

[0055] The generator introduces a conditional generative adversarial mechanism, and the loss function is:

[0056]

[0057] In formula (3), z is derived from the noise distribution p z The noise vector sampled from the conditional distribution p is c. c The conditional information sampled in , G(z,c) is the flow data generated by the generator based on the noise z and condition c, is a dynamic conditional fusion operation that fuses the conditional information after noise Q(c) processing. This function simulates the superposition characteristics of quantum states and enhances the interaction effect between conditional information and noise. D(·) is the discriminator, and the goal of the generator is to minimize L gen , generating realistic data that meets specific conditions; A(c) is an adaptive condition enhancement module, which dynamically adjusts the weights and dimensions of the generator input conditions according to the characteristics of the conditions, so that the generator focuses more on key condition information; M(c) is a condition information mapping module, which transforms the original conditions to generate a form that is more suitable for the discriminator input; L ms is a multi-scale feature penalty term, which calculates the feature differences of generated data at different scales, making the generated data closer to the real data in terms of details and overall structure, thereby improving the quality of generated data; qe It is a quantum entanglement weighting term that simulates the quantum entanglement phenomenon, assigns weight to the association between conditional information and generated data, and strengthens the constraint effect of conditions on the generation process; ms and λ qe are the weight coefficients of the multi-scale feature penalty term and the quantum entanglement weighted term, which are dynamically adjusted through reinforcement learning to balance the contribution of each part in the loss function;

[0058] The discriminator uses a multi-task learning loss function:

[0059]

[0060] In formula (4), D STQ It is a spatiotemporal quantization discriminator that processes input through the spatiotemporal feature extraction layer, quantum enhancement layer and attention fusion layer architecture; x is the distribution of real data p data The real traffic data sampled in L aux is the auxiliary task loss, λ(t) is the dynamic weight coefficient, and the discriminator’s goal is to maximize L dis , enhance the discriminative ability and understanding of data features through multi-task learning; Lqef is the quantum dining characteristic loss; L sa is the spatiotemporal attention loss, which constrains the discriminator to focus on the key spatiotemporal domain; γ and β are the weight coefficients of quantum feature loss and attention loss, respectively, which are dynamically optimized by the reinforcement learning agent. Both are continuously optimized during the confrontation process, which improves the data quality generated by the generator and enhances the recognition ability of the discriminator;

[0061] S3-4, Overall Operation and Evaluation

[0062] The trained generator and discriminator are deployed into a distributed system based on edge computing and blockchain. A lightweight edge computing unit is deployed at each data collection node to perform local data preprocessing and preliminary discrimination. Blockchain technology is used to record the data source, processing process, and model decision results to ensure data credibility and traceability. The generator is used to predict traffic flow for the next 10 minutes, and the discriminator monitors traffic data in real time to identify abnormal traffic patterns.

[0063] Construct a multi-dimensional dynamic evaluation system. In addition to calculating the prediction error of the generator and the recognition accuracy of the discriminator, it also introduces data diversity indicators, model robustness indicators, and model interpretability indicators. Regularly collect data from actual applications, and use the federated learning framework to summarize the evaluation results of each node without leaking privacy, and comprehensively evaluate the model performance. If the performance degrades, locate the problem through the automatic model diagnosis system, and use transfer learning and incremental learning techniques for optimization.

[0064] Preferably, the working method of the fuzzy PID algorithm in step S4 is:

[0065] S4-1, variable definition and quantification upgrade

[0066] Network load and delay fluctuation are defined as core input variables, with packet arrival rate and link jitter coefficient introduced as auxiliary references. Each variable is quantized to the domain of [-3, 3]. Network load is divided into levels based on bandwidth utilization, packet loss rate, and arrival rate. Delay fluctuation is comprehensively evaluated using the delay standard deviation and jitter coefficient. When the delay standard deviation is less than 5ms and the jitter coefficient is low, it is considered "small fluctuation." The output variable, congestion window adjustment step size, takes values of -2KB, 0, and +2KB, and is also mapped to the above domain.

[0067] S4-2, Fuzzy Rules and Reasoning Enhancement

[0068] A sophisticated fuzzy rule base was constructed, with rules that take into account link jitter and packet arrival rate. When network load is high, latency fluctuations are large, and link jitter is severe, the congestion window adjustment step size is set to -2KB. An improved Mamdani fuzzy inference method was used to dynamically adjust membership function parameters to adapt to network state changes. A weighted average method was used for defuzzification, with weights set based on rule importance to quickly determine the adjustment step size.

[0069] S4-3, dynamic response optimization

[0070] When bandwidth utilization exceeds 80% and latency fluctuation is greater than 15ms, the window reduction rule is triggered. PID parameters are automatically optimized through a reinforcement learning-based strategy. The adaptive learning rate adjustment formula is:

[0071]

[0072] In formula (5), α is the learning rate, β is the error factor, and ∈ is the minimum value to avoid the denominator being zero. This formula dynamically adjusts the learning rate based on historical error changes to ensure the stability and convergence speed of the system under different states. is to simulate the temperature decay during quantum annealing; γ is the control exploration amplitude; T is the control exploration period;

[0073] S4-4, Thermoelectric Power Generation Module Design

[0074] Using a nanocomposite device based on the Seebeck effect, the device activates when the surface temperature difference reaches 5°C or above. Based on the traditional Seebeck effect to calculate output power, efficiency enhancement factors related to nanostructures and materials are introduced. Through material screening, structural design, and intelligent temperature control and heat dissipation systems, the output power is guaranteed to be no less than 50μW. Energy management circuits are also used to efficiently store and distribute electrical energy.

[0075] S4-5, vibration power generation module performance improvement

[0076] Using piezoelectric ceramic materials with a gradient polarization structure, it works at a vibration frequency of 10-100Hz and an amplitude of 0.1mm or above; designing a vibration frequency adaptive matching circuit to improve power generation performance and automatically adjust the capacitance and inductance parameters to put the piezoelectric ceramics in the optimal resonant state; at the same time, adopting a multi-level energy capture structure to enhance the response capability to different vibrations; ultimately achieving an output power of no less than 30μW. Compared with traditional designs, the power generation power in a typical vibration environment is increased by 20%-30%, extending the self-powered battery life of communication equipment.

[0077] The present invention also discloses a communication transmission rate adaptation and flow control system for an underground power distribution room, comprising a multi-standard communication module, a reinforcement learning relay module, an adaptive rate adaptation module, an intelligent flow control module, and an energy efficiency control module, wherein:

[0078] The multi-standard communication module integrates three communication modes: medium-voltage carrier, LoRa, and 5G-RedCap. These three communication modes are respectively suitable for short-range stable transmission, long-range low-power transmission, and high-speed and low-latency scenarios. By normalizing quantum state parameters, a three-dimensional communication link matrix is constructed to achieve parallel transmission and intelligent switching of multi-standard signals.

[0079] The reinforcement learning relay module includes five submodules: the spatiotemporal feature fusion module achieves high-precision synchronization and feature fusion; the data encoding module uses quantum algorithms to process signals and information; the spatiotemporal Bayesian network module predicts delays and optimizes paths; the dynamic channel switching module achieves fast switching; and the multi-node interaction module ensures data privacy and optimizes strategies.

[0080] The adaptive rate adaptation module uses an improved generative adversarial network (GAN). The generator predicts traffic, the discriminator identifies data and detects anomalies, and dynamically adjusts the transmission rate in steps of 1 Mbps based on network load and service priority.

[0081] The intelligent flow control module adopts a dual closed-loop architecture. The outer loop uses a fuzzy PID algorithm to adjust the congestion window, and the inner loop reserves the power control signal bandwidth through priority queue management.

[0082] The energy efficiency optimization module adjusts the transmission power based on energy sensing, triggering low-power mode when the power level drops below 20%. It also integrates temperature difference and vibration power generation modules to achieve self-power supply, reducing energy consumption by more than 30%.

[0083] The multi-standard communication module processes the three standard signals through quantum state processing to construct a link matrix, and then outputs it to the reinforcement learning relay module;

[0084] The five submodules of the reinforcement learning relay module collaboratively plan the optimal path, and the processed signal is transmitted to the adaptive rate adaptation module;

[0085] The adaptive rate adaptation module predicts the flow rate based on GAN and adjusts the rate, and then transmits the signal to the intelligent flow control module;

[0086] The intelligent flow control module accurately controls the flow through a dual closed-loop architecture, and the processed signal is returned to the multi-standard communication module for transmission;

[0087] The energy efficiency optimization module monitors the power consumption and node location, adjusts the power and self-powers the modules to support the operation.

[0088] Beneficial effects of the present invention: The underground distribution room communication transmission rate adaptation and flow control method and system of the present invention achieves transmission delay ≤ 50ms through dynamic fusion and intelligent switching of multi-standard signals and reinforcement learning path optimization; dynamically adjusts the transmission rate through improved GAN adaptive rate adaptation, making the step size 1Mbps; ensures the bandwidth of key signals through dual closed-loop fuzzy PID flow control; reduces energy consumption by more than 30% through energy-aware power control combined with self-power supply. Through the implementation of the present invention, the stability and reliability of underground distribution room communication are significantly enhanced, the transmission efficiency and resource utilization are optimized, the intelligent adaptive capability is improved, and the bottleneck of green energy saving and self-sustaining is broken through. It is also compatible with multiple standards and has strong adaptability, providing efficient, low-consumption, and intelligent underlying communication support for smart grids and ubiquitous power Internet of Things. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is the overall workflow diagram of the present invention;

[0090] Figure 2 This is a workflow diagram of the improved generative adversarial network GAN in the present invention;

[0091] Figure 3 It is the working flow chart of the fuzzy PID algorithm in the present invention;

[0092] Figure 4 It is a system structure diagram of the present invention;

[0093] Figure 5 for Figure 4 Structural diagram of the reinforcement learning module. DETAILED DESCRIPTION

[0094] In order to make the technical solution of the present invention easier to understand, the technical solution of the present invention is now clearly and completely described in the form of specific embodiments in conjunction with the accompanying drawings.

[0095] Example 1:

[0096] like Figures 1 to 3 As shown, the underground distribution room communication transmission rate adaptation and flow control method of this embodiment includes the following steps:

[0097] S1, dynamic fusion of multi-standard signals:

[0098] By integrating multi-standard communication modules of medium-voltage carrier, LoRa and 5G-RedCap, a three-dimensional communication link matrix is constructed. The weight coefficients of medium-voltage carrier, LoRa and 5G-RedCap are α1, α2 and α3 respectively. A dynamic weight allocation algorithm based on the weight coefficients α1, α2 and α3 realizes parallel transmission and intelligent switching of multi-standard signals. The steps of constructing the three-dimensional communication link matrix are as follows:

[0099] S1-1, through the quantum state parameter normalization mechanism, the transmission bandwidth, bit error rate and wall penetration loss of the three standards of medium voltage carrier, LoRa and 5G-RedCap are mapped to the quantum bit space. The quantum measurement operator is used to realize the non-deterministic normalization of multi-dimensional parameters, forming a time-space asymmetric tensor matrix of number of nodes × number of nodes × number of standards;

[0100] S1-2: A quantum Bayesian cognitive engine is introduced to estimate channel quality in real time. A Nash equilibrium solution space for multi-standard resource allocation is constructed based on quantum game theory. The initial weight coefficients are α1 = 0.38, α2 = 0.25, and α3 = 0.29. When the SNR is ≥ 20 dB, the 5G-RedCap weight α3 is increased to 0.45 through the quantum entanglement exchange mechanism. When the SNR is < 20 dB, the LoRa weight α2 is dynamically adjusted to 0.35.

[0101] S1-3 uses a non-Hermitian quantum evolution model to describe link state changes, and uses the PT symmetric phase transition detection mechanism to find the optimal switching time in real time and switch the signal to the optimal standard.

[0102] S2, reinforcement learning relay path selection:

[0103] An improved reinforcement learning model is designed to construct a state transition matrix including a spatial distance attenuation factor of 0.78 / km, a time delay compensation coefficient of 0.96 / ms, and an interference intensity threshold of -90dBm. The model dynamically selects relay nodes and optimizes relay paths to reduce transmission delay to less than 50ms. The improved reinforcement learning model includes a spatiotemporal feature fusion module, a data encoding module, a spatiotemporal Bayesian network module, a dynamic channel switching module, and a multi-node interaction module.

[0104] The workflow of the spatiotemporal feature fusion module includes:

[0105] S2-1 uses a quantum entanglement clock synchronization protocol to achieve ultra-high-precision synchronization of ≤100ns. The packet transmission delay and relay node processing delay are encoded as quantum phase differences. The time delay compensation coefficient of 0.99 / ms is obtained through the quantum state time evolution equation and measurement.

[0106] S2-2 uses the cyberspace gravitational model and non-Euclidean distance metrics, combined with the quantum wall-penetrating attenuation model, to map node energy and bandwidth to gravitational field strength. Through the curvature tensor and quantum tunneling effect, the node spatial distance is accurately calculated and spatial characteristics are formed.

[0107] S2-3, through the quantum entanglement exchange network and quantum Fourier transform, realizes the non-local correlation fusion of spatiotemporal characteristics in the quantum frequency domain, and uses quantum principal component analysis to adaptively extract key feature components; based on the quantum reinforcement learning framework and quantum perception hash algorithm, combined with the quantum neural network, predicts channel interference in advance. When the RSSI monitors that the interference intensity exceeds the threshold of -90dBm, the intelligent optimal switching of multi-standard signals can be completed within 5ms.

[0108] The workflow of the data encoding module includes:

[0109] S2-4 uses a quantum superposition state coding algorithm to map the medium voltage carrier voltage signal to the quantum Bloch sphere, uses quantum phase coding to encode LoRa's spreading factors SF7-SF12, and uses quantum amplitude modulation to process 5G-RedCap's modulation coding strategy MCS0-MCS16. Combined with ternary quantum entanglement coding, it forms a 128-dimensional quantum feature vector.

[0110] S2-5, using the space-time spinor representation and geometric algebra, the latitude and longitude of the sensor, altitude coordinates, and the temperature, voltage, and load rate of the equipment operating status are converted into geometric algebra elements with physical meaning for fusion;

[0111] S2-6 introduces quantum fractional Fourier transform, realizes flexible conversion of signal characteristics in time and frequency domain by adaptively adjusting the transformation order, and uses quantum entanglement exchange network to establish non-local correlation of multi-standard signals; at the same time, adopts quantum Lorentz mapping and quantum measurement entropy normalization to enhance feature discrimination based on chaotic system and quantify feature uncertainty.

[0112] The workflow of the spatiotemporal Bayesian network module includes:

[0113] S2-7, using quantum tensor product networks to embed spatial distance, time delay, and interference intensity into quantum Hilbert space, using quantum phase estimation algorithms to parallelly calculate three-dimensional state transition probabilities, and constructing non-local causal relationships between nodes based on quantum entanglement exchange mechanisms;

[0114] S2-8 uses a hidden Markov model in the form of a quantum superoperator, combined with quantum Monte Carlo sampling and error correction coding, to achieve efficient prediction of delay distribution within the next five time windows in an exponential state space;

[0115] S2-9, in terms of path optimization, a space-time curvature routing algorithm is designed based on gravitational wave theory, which converts node energy and load into space-time curvature tensor. The shortest path is solved by the Dijkstra algorithm optimized by quantum annealing, and the quantum field theory path integral model is used to replace the traditional formula for calculation.

[0116] The workflow of the dynamic channel switching module includes:

[0117] S2-10 uses a hybrid modulation architecture assisted by quantum stochastic resonance, integrating frequency hopping frequency spectrum (FHSS) and direct sequence spread spectrum (DSSS), and introduces quantum tunneling effect to enhance signal anti-interference capability;

[0118] S2-11, a dynamic decision-making model is constructed based on the quantum annealing reinforcement learning framework and the biological immune memory mechanism. When the SNR is less than 20dB, the quantum parallel search algorithm is triggered to globally scan the multidimensional channel space and simultaneously initiate the biological chemotaxis path optimization, shortening the channel switching cycle from 500ms to microseconds. Quantum entanglement is used to establish non-local correlation of channel states, enabling pre-evaluation and pre-configuration of backup channels.

[0119] S2-12 adopts the quantum genetic Q-learning algorithm, encodes multiple switching strategies through quantum bit superposition states, and simulates the dynamic optimization strategy of biological gene evolution process; the module has a built-in adaptive environmental perception system and uses the quantum perception hash algorithm to quickly match the optimal strategy.

[0120] The multi-node interaction module adopts the FedAvg algorithm framework of federated reinforcement learning, and its workflow includes:

[0121] S2-13, the multi-node interaction module builds a quantum entanglement federated-biological symbiotic network architecture, which innovates the traditional federated learning framework. Each relay node encodes local training parameters into quantum entangled states, and achieves secure parameter aggregation through quantum teleportation and remote preparation protocols, fundamentally protecting data privacy.

[0122] S2-14, uses the principle of biological symbiosis to build a defense mechanism, maps the parameter vector into an antigen-antibody reaction pattern, and resists malicious attacks through immune memory and clonal selection; innovates the quantum differential privacy protection mechanism, uses the unpredictability of quantum noise to enhance privacy protection, and reduces the risk of data leakage to 1 / 1000 of the traditional method under a privacy budget of ε=0.5; proposes a quantum game reinforcement learning framework, in which each node establishes non-local associations through quantum entanglement, achieves a Pareto optimal strategy without sharing private information, and designs a quantum state reward function to balance global optimization and local interests.

[0123] S3, adaptive rate adaptation:

[0124] An improved generative adversarial network (GAN) was introduced to build a traffic prediction model. Combining real-time network load and service priority, the modulation and coding module was dynamically adjusted to achieve hierarchical adaptation of the transmission rate, with a rate adjustment step size of 1 Mbps.

[0125] The implementation method of the improved generative adversarial network GAN is as follows:

[0126] S3-1, Generator Construction and Implementation

[0127] The generator is constructed using a hybrid attention mechanism-enhanced LSTM network, embedding a spatiotemporal module and a self-attention module. The spatiotemporal module uses multi-head attention to capture the long-term spatiotemporal dependencies of power flow data, while the self-attention module focuses on key features. The number of neurons in the three hidden layers is 128, 64, and 32, respectively. The gating and attention mechanisms work together to process complex temporal relationships.

[0128] The system inputs one hour of multi-source heterogeneous traffic data, including power control, equipment monitoring, environmental sensing, weather warning, and grid topology dynamics data. It applies multimodal feature fusion encoding, builds semantic associations using knowledge graphs, maps the data into a unified feature space, and collaboratively optimizes the encoding model through federated learning, scaling it to the [0, 1] range.

[0129] The output of the generator is the traffic forecast value for the next 10 minutes. The improved composite loss function is used in the training process, and the formula is:

[0130] L gen =α(t)*L mes *(1+λ spatial *S adj )+β(t)*L TV *(1+λ temporal *T corr )+γ(t)*L adv *(1+λ hybrid *H factor ) (1)

[0131] In formula (1), L mes is the mean square error, which measures the difference between the predicted value and the true value; L TV is the total variation loss, which is used to smooth the prediction results and prevent overfitting; L adv To combat the loss and encourage the generator to generate more realistic data, α(t), β(t), and γ(t) are dynamic weight coefficients, which are optimized through dynamic adjustment algorithms and adaptively changed according to the training stage; λ spatial ,λ temporal ,λ hybrid S is a newly added adjustment parameter that controls the influence of space, time and mixed factors on the loss function; adjL is a spatial correlation factor that considers the spatial correlation between the traffic data in the prediction area and the surrounding areas, explores the potential connection in the spatial dimension, and improves the rationality of the prediction; adv is the time correlation factor, which captures the dependency of traffic data in the time series and optimizes the prediction by using the time correlation between history and future. factor It is a mixed correlation factor that comprehensively considers the impact of various complex factors of time and space on traffic, so that the model can cope with more complex traffic change scenarios; the LSTM network parameters θ are optimized through the back propagation algorithm gen ,Set the prediction error target to ≤5%. When the average prediction error on the validation set exceeds this threshold, the network structure and parameters are quickly adjusted in combination with the meta-learning mechanism;

[0132] S3-2, Discriminator Construction and Implementation

[0133] The discriminator uses a multi-scale feature pyramid convolutional neural network (MSF-PCNN), which includes four convolutional layer groups, three pyramid pooling layers, and two fully connected layers. Each convolutional layer group contains parallel convolutional layers with different convolution kernel sizes to extract multi-scale local features. The pyramid pooling layer aggregates features at different granularities to obtain a more comprehensive feature representation, and finally the fully connected layer makes classification decisions. The improved Wasserstein-f divergence metric is used to measure the distribution difference between the predicted traffic and the actual traffic. The formula is:

[0134]

[0135] In formula (2), f(.) is a customized nonlinear mapping function that enhances the sensitivity to data distribution differences; P data is the true data distribution, P gen is the generated data distribution, Π(P data ,P gen ) is the set of all possible joint distributions between the two, γ is a joint distribution sampled from the joint distribution set, and (x, y) is a pair of data sampled from γ; when training the discriminator, a quantum computing acceleration mechanism is introduced, and the discrimination results of multiple samples are calculated in parallel using quantum superposition states. The RMSProp optimizer with adaptive learning rate is used to update the discriminator parameters θ. dis , by maximizing W f (P data ,P gen ), so that the discriminator can distinguish the distribution differences between the two more accurately;

[0136] The discriminator outputs a three-category result, including determining whether the input traffic data is "real data," "generated data," or "suspected anomaly." Combined with a causal inference algorithm, the discriminator further analyzes the causal relationship between data identified as "suspected anomaly" and other related data to determine whether it is a true abnormal traffic pattern, achieving more intelligent anomaly identification. A comprehensive identification accuracy target of ≥95% is set, and the discriminator performance is evaluated on a test set. If the target is not achieved, the network structure and parameters are automatically adjusted through reinforcement learning.

[0137] S3-3, Generative Adversarial Training

[0138] Initialize the network parameters of the generator and discriminator separately, using a quantum-inspired parameter initialization method that utilizes the superposition and entanglement properties of quantum states to generate better initial weights, and initialize the bias term to 0;

[0139] During the training process, a dynamic task priority alternating training strategy is adopted; the ratio of the number of training rounds of the generator and discriminator is dynamically adjusted according to the current training stage and data characteristics;

[0140] The generator introduces a conditional generative adversarial mechanism, and the loss function is:

[0141]

[0142] In formula (3), z is derived from the noise distribution p z The noise vector sampled from the conditional distribution p is c. c The conditional information sampled in , G(z,c) is the flow data generated by the generator based on the noise z and condition c, is a dynamic conditional fusion operation that fuses the conditional information after noise Q(c) processing. This function simulates the superposition characteristics of quantum states and enhances the interaction effect between conditional information and noise. D(·) is the discriminator, and the goal of the generator is to minimize L gen , generating realistic data that meets specific conditions; A(c) is an adaptive condition enhancement module, which dynamically adjusts the weights and dimensions of the generator input conditions according to the characteristics of the conditions, so that the generator focuses more on key condition information; M(c) is a condition information mapping module, which transforms the original conditions to generate a form that is more suitable for the discriminator input; L ms is a multi-scale feature penalty term, which calculates the feature differences of generated data at different scales, making the generated data closer to the real data in terms of details and overall structure, thereby improving the quality of generated data; qe It is a quantum entanglement weighting term that simulates the quantum entanglement phenomenon, assigns weight to the association between conditional information and generated data, and strengthens the constraint effect of conditions on the generation process; ms and λ qeare the weight coefficients of the multi-scale feature penalty term and the quantum entanglement weighted term, which are dynamically adjusted through reinforcement learning to balance the contribution of each part in the loss function;

[0143] The discriminator uses a multi-task learning loss function:

[0144]

[0145] In formula (4), D STQ It is a spatiotemporal quantization discriminator that processes input through the spatiotemporal feature extraction layer, quantum enhancement layer and attention fusion layer architecture; x is the distribution of real data p data The real traffic data sampled in L aux is the auxiliary task loss, λ(t) is the dynamic weight coefficient, and the discriminator’s goal is to maximize L dis , enhance the discriminative ability and understanding of data features through multi-task learning; L qef is the quantum dining characteristic loss; L sa is the spatiotemporal attention loss, which constrains the discriminator to focus on the key spatiotemporal domain; γ and β are the weight coefficients of quantum feature loss and attention loss, respectively, which are dynamically optimized by the reinforcement learning agent. Both are continuously optimized during the confrontation process, which improves the data quality generated by the generator and enhances the recognition ability of the discriminator;

[0146] S3-4, Overall Operation and Evaluation

[0147] The trained generator and discriminator are deployed into a distributed system based on edge computing and blockchain. A lightweight edge computing unit is deployed at each data collection node to perform local data preprocessing and preliminary discrimination. Blockchain technology is used to record the data source, processing process, and model decision results to ensure data credibility and traceability. The generator is used to predict traffic flow for the next 10 minutes, and the discriminator monitors traffic data in real time to identify abnormal traffic patterns.

[0148] Construct a multi-dimensional dynamic evaluation system. In addition to calculating the prediction error of the generator and the recognition accuracy of the discriminator, it also introduces data diversity indicators, model robustness indicators, and model interpretability indicators. Regularly collect data from actual applications, and use the federated learning framework to summarize the evaluation results of each node without leaking privacy, and comprehensively evaluate the model performance. If the performance degrades, locate the problem through the automatic model diagnosis system, and use transfer learning and incremental learning techniques for optimization.

[0149] In a specific embodiment, the present invention reconstructs the GAN architecture based on quantum inspiration and spatiotemporal attention mechanism, constructs a generator through a hybrid attention LSTM network, captures the long-range dependency of power flow and focuses on key features, inputs multi-source heterogeneous data and generates a 10-minute flow forecast through knowledge graph encoding, uses a composite loss function to constrain the rationality of the prediction, and triggers meta-learning to adjust the network structure when the verification error is >5%; designs an MSF-PCNN discriminator, uses an improved Wasserstein-f divergence to measure distribution differences, introduces a quantum acceleration mechanism to parallelly calculate the discrimination results, outputs three-category results and performs causal reasoning analysis on "suspected abnormal" data, and automatically adjusts the structure through reinforcement learning if the test accuracy is <95%; based on quantum-inspired initialization parameters, dynamically adjusts the ratio of generator and discriminator training rounds, incorporates quantum entanglement weighting and multi-scale feature penalties into the generator, and adopts a spatiotemporal quantization architecture and multi-task loss into the discriminator; deploys the model to the edge computing node and combines the blockchain to record the data processing process, regularly summarizes the evaluation results through federated learning, and uses transfer learning to optimize the model. During the implementation, we achieved a traffic prediction error of ≤3.8%, anomaly recognition accuracy of ≥97.5%, discrimination processing delay reduced to 5ms, training convergence speed increased by 40%, and data diversity index increased to 0.08, significantly enhancing the model's prediction accuracy, real-time performance, robustness and system scalability.

[0150] Table 1 Performance comparison between improved GAN and traditional GAN

[0151] Evaluation Metrics Improved GAN Traditional GAN Improvement Forecast Error (MAE) ≤3.8％ 8.2％ ↓53.7％ Anomaly recognition accuracy ≥97.5％ 85.0％ ↑14.7％ Delay in processing 5ms 20ms ↓75.0％ Robustness to missing data Error increased by 1.5% Error increased by 5.8% ↓74.1％ Training convergence speed 40 rounds of convergence 67 rounds of convergence ↑40.3％ Data diversity (MMD) 0.08 0.15 ↓46.7％

[0152] Table 1 compares the performance of the improved GAN and traditional GAN in core indicators such as prediction error, anomaly recognition, processing delay, robustness, convergence speed, and data diversity. It shows that the improved GAN is significantly better than the traditional model in all indicators, especially in prediction accuracy and real-time performance.

[0153] S4, intelligent flow control command output:

[0154] A dual-closed-loop flow control architecture is designed. The outer loop dynamically adjusts the congestion window based on a fuzzy PID algorithm, while the inner loop reserves bandwidth for power control signals through priority queue management.

[0155] The working method of the fuzzy PID algorithm is as follows:

[0156] S4-1, variable definition and quantification upgrade

[0157] Network load and delay fluctuation are defined as core input variables, with packet arrival rate and link jitter coefficient introduced as auxiliary references. Each variable is quantized to the domain of [-3, 3]. Network load is divided into levels based on bandwidth utilization, packet loss rate, and arrival rate. Delay fluctuation is comprehensively evaluated using the delay standard deviation and jitter coefficient. When the delay standard deviation is less than 5ms and the jitter coefficient is low, it is considered "small fluctuation." The output variable, congestion window adjustment step size, takes values of -2KB, 0, and +2KB, and is also mapped to the above domain.

[0158] S4-2, Fuzzy Rules and Reasoning Enhancement

[0159] A sophisticated fuzzy rule base was constructed, with rules that take into account link jitter and packet arrival rate. When network load is high, latency fluctuations are large, and link jitter is severe, the congestion window adjustment step size is set to -2KB. An improved Mamdani fuzzy inference method was used to dynamically adjust membership function parameters to adapt to network state changes. A weighted average method was used for defuzzification, with weights set based on rule importance to quickly determine the adjustment step size.

[0160] S4-3, dynamic response optimization

[0161] When bandwidth utilization exceeds 80% and latency fluctuation is greater than 15ms, the window reduction rule is triggered. PID parameters are automatically optimized through a reinforcement learning-based strategy. The adaptive learning rate adjustment formula is:

[0162]

[0163] In formula (5), α is the learning rate, β is the error factor, and ∈ is the minimum value to avoid the denominator being zero. This formula dynamically adjusts the learning rate based on historical error changes to ensure the stability and convergence speed of the system under different states. is to simulate the temperature decay during quantum annealing; γ is the control exploration amplitude; T is the control exploration period;

[0164] S4-4, Thermoelectric Power Generation Module Design

[0165] Using a nanocomposite device based on the Seebeck effect, the device activates when the surface temperature difference reaches 5°C or above. Based on the traditional Seebeck effect to calculate output power, efficiency enhancement factors related to nanostructures and materials are introduced. Through material screening, structural design, and intelligent temperature control and heat dissipation systems, the output power is guaranteed to be no less than 50μW. Energy management circuits are also used to efficiently store and distribute electrical energy.

[0166] S4-5, vibration power generation module performance improvement

[0167] Using piezoelectric ceramic materials with a gradient polarization structure, it works at a vibration frequency of 10-100Hz and an amplitude of 0.1mm or above; designing a vibration frequency adaptive matching circuit to improve power generation performance, automatically adjust the capacitance and inductance parameters to put the piezoelectric ceramics in the optimal resonant state; at the same time, adopting a multi-level energy capture structure to enhance the response capability to different vibrations; ultimately achieving an output power of no less than 30μW. Compared with traditional designs, the power generation power in a typical vibration environment is increased by 20%-30%, significantly extending the self-powered battery life of communication equipment.

[0168] In a specific embodiment, this invention utilizes fuzzy logic and reinforcement learning as its core, combined with energy harvesting technology to achieve intelligent network traffic control and device self-powering. Network load and latency fluctuation variables are quantized to the [-3,3] domain, a fuzzy rule base incorporating link jitter and packet arrival rate is constructed, and improved Mamdani reasoning and weighted average defuzzification are used to precisely adjust the congestion window. PID parameters are dynamically optimized using reinforcement learning formulas, and the learning rate is adjusted using a simulated quantum annealing mechanism, improving system response speed and stability. Regarding self-powering, a nano-thermoelectric power generation module based on the Seebeck effect outputs ≥50μW at a temperature difference of ≥5°C, while a gradient-polarized piezoelectric ceramic vibration power generation module generates ≥30μW at 10-100Hz vibration, representing a 20%-30% improvement over traditional designs. This results in a 28% reduction in network latency, a 1.3% drop in packet loss rate, a 35% increase in throughput, a 2.5-fold increase in communication device battery life, and a 40% reduction in energy consumption, significantly enhancing the system's adaptability and reliability in complex environments.

[0169] Table 2 Fuzzy PID algorithm and self-powered system comprehensive performance comparison verification table

[0170]

[0171]

[0172] Table 2 compares the key performance indicators of traditional solutions and the present invention from the perspectives of flow control, self-powered system, and algorithm optimization. In terms of flow control, the present invention reduces network latency by 28.2%, packet loss by 75%, and throughput by 35%. In the self-powered system, the thermoelectric power generation module achieves a stable output of ≥50μW, the vibration power generation module increases power by 20%-30%, and the device's battery life is extended by 150%. In terms of algorithm optimization, the learning rate convergence speed is accelerated by 46.7%, and the system stability coefficient is improved by 31.9%, fully demonstrating the present invention's advantages in network control and energy management.

[0173] S5, energy efficiency optimization:

[0174] An energy-aware power control algorithm is designed. By combining the remaining power of the device and the location of the relay node, which triggers the low-power mode when the power threshold is ≤20%, the transmission power is dynamically adjusted. Energy harvesting technology is used to achieve self-powering of the communication equipment, reducing the system energy consumption by more than 30%.

[0175] Example 2:

[0176] like Figures 4 to 5 As shown, the underground distribution room communication transmission rate adaptation and flow control system of the present invention includes a multi-standard communication module, a reinforcement learning relay module, an adaptive rate adaptation module, an intelligent flow control module and an energy efficiency control module, wherein:

[0177] The multi-standard communication module integrates three communication modes: medium-voltage carrier, LoRa, and 5G-RedCap. These three communication modes are respectively suitable for short-range stable transmission, long-range low-power transmission, and high-speed and low-latency scenarios. By normalizing quantum state parameters, a three-dimensional communication link matrix is constructed to achieve parallel transmission and intelligent switching of multi-standard signals.

[0178] The reinforcement learning relay module includes five submodules: the spatiotemporal feature fusion module achieves high-precision synchronization and feature fusion; the data encoding module uses quantum algorithms to process signals and information; the spatiotemporal Bayesian network module predicts delays and optimizes paths; the dynamic channel switching module achieves fast switching; and the multi-node interaction module ensures data privacy and optimizes strategies.

[0179] The adaptive rate adaptation module uses an improved generative adversarial network (GAN). The generator predicts traffic, the discriminator identifies data and detects anomalies, and dynamically adjusts the transmission rate in steps of 1 Mbps based on network load and service priority.

[0180] The intelligent flow control module adopts a dual closed-loop architecture. The outer loop uses a fuzzy PID algorithm to adjust the congestion window, and the inner loop reserves the power control signal bandwidth through priority queue management.

[0181] The energy efficiency optimization module adjusts the transmission power based on energy sensing, triggering low-power mode when the power level drops below 20%. It also integrates temperature difference and vibration power generation modules to achieve self-power supply, reducing energy consumption by more than 30%.

[0182] The multi-standard communication module processes the three standard signals through quantum state processing to construct a link matrix, and then outputs it to the reinforcement learning relay module;

[0183] The five submodules of the reinforcement learning relay module collaboratively plan the optimal path, and the processed signal is transmitted to the adaptive rate adaptation module;

[0184] The adaptive rate adaptation module predicts the flow rate based on GAN and adjusts the rate, and then transmits the signal to the intelligent flow control module;

[0185] The intelligent flow control module accurately controls the flow through a dual closed-loop architecture, and the processed signal is returned to the multi-standard communication module for transmission;

[0186] The energy efficiency optimization module monitors the power consumption and node location, adjusts the power and self-powers the modules to support the operation.

[0187] It should be noted that the embodiments described herein are only some embodiments of the present invention, not all implementations of the present invention. The embodiments are merely illustrative and serve only to provide a more intuitive and clear way to understand the content of the present invention, rather than to limit the technical solutions described in the present invention. Without departing from the concept of the present invention, all other implementations that can be thought of by ordinary technicians in this field without making creative efforts, as well as other simple replacements and various variations of the technical solutions of the present invention, are within the scope of protection of the present invention.

Claims

1. A method for communication transmission rate adaptation and flow control in an underground distribution room, characterized in that: The following steps are involved: S1, dynamic fusion of multi-standard signals: By integrating multi-standard communication modules including medium-voltage carrier, LoRa, and 5G-RedCap, a three-dimensional communication link matrix is constructed. The weight coefficients of the medium-voltage carrier, LoRa, and 5G-RedCap are α1, α2, and α3, respectively. A dynamic weight allocation algorithm based on the weight coefficients α1, α2, and α3 enables parallel transmission and intelligent switching of multi-standard signals. S2, reinforcement learning relay path selection: An improved reinforcement learning model is designed to construct a state transition matrix including a spatial distance attenuation factor of 0.78 / km, a time delay compensation coefficient of 0.96 / ms, and an interference intensity threshold of -90dBm. The model dynamically selects relay nodes and optimizes relay paths to reduce transmission delay to less than 50ms. The improved reinforcement learning model includes a spatiotemporal feature fusion module, a data encoding module, a spatiotemporal Bayesian network module, a dynamic channel switching module, and a multi-node interaction module. S3, adaptive rate adaptation: An improved generative adversarial network (GAN) was introduced to build a traffic prediction model. Combining real-time network load and service priority, the modulation and coding module was dynamically adjusted to achieve hierarchical adaptation of the transmission rate, with a rate adjustment step size of 1 Mbps. S4, intelligent flow control command output: A dual-closed-loop flow control architecture is designed. The outer loop dynamically adjusts the congestion window based on a fuzzy PID algorithm, while the inner loop reserves bandwidth for power control signals through priority queue management. S5, energy efficiency optimization: An energy-aware power control algorithm is designed. By combining the remaining power of the device and the location of the relay node, which triggers the low-power mode when the power threshold is ≤20%, the transmission power is dynamically adjusted, and self-powered communication equipment is achieved through energy harvesting technology.

2. The underground distribution room communication transmission rate adaptation and flow control method according to claim 1, characterized in that: The steps for constructing the three-dimensional communication link matrix in step S1 are: S1-1, through the quantum state parameter normalization mechanism, the transmission bandwidth, bit error rate and wall penetration loss of the three standards of medium voltage carrier, LoRa and 5G-RedCap are mapped to the quantum bit space. The quantum measurement operator is used to realize the non-deterministic normalization of multi-dimensional parameters, forming a time-space asymmetric tensor matrix of number of nodes × number of nodes × number of standards; S1-2: A quantum Bayesian cognitive engine is introduced to estimate channel quality in real time. A Nash equilibrium solution space for multi-standard resource allocation is constructed based on quantum game theory. The initial weight coefficients are α1 = 0.38, α2 = 0.25, and α3 = 0.

29. When the SNR is ≥ 20 dB, the 5G-RedCap weight α3 is increased to 0.45 through the quantum entanglement exchange mechanism. When the SNR is < 20 dB, the LoRa weight α2 is dynamically adjusted to 0.

35. S1-3 uses a non-Hermitian quantum evolution model to describe link state changes, and uses the PT symmetric phase transition detection mechanism to find the optimal switching time in real time and switch the signal to the optimal standard.

3. The underground distribution room communication transmission rate adaptation and flow control method according to claim 1, characterized in that: The workflow of the spatiotemporal feature fusion module in step S2 includes: S2-1 uses a quantum entanglement clock synchronization protocol to achieve ultra-high-precision synchronization of ≤100ns. The packet transmission delay and relay node processing delay are encoded as quantum phase differences. The time delay compensation coefficient of 0.99 / ms is obtained through the quantum state time evolution equation and measurement. S2-2 uses the cyberspace gravitational model and non-Euclidean distance metrics, combined with the quantum wall-penetrating attenuation model, to map node energy and bandwidth to gravitational field strength. Through the curvature tensor and quantum tunneling effect, the node spatial distance is accurately calculated and spatial characteristics are formed. S2-3, through the quantum entanglement exchange network and quantum Fourier transform, realizes the non-local correlation fusion of spatiotemporal characteristics in the quantum frequency domain, and uses quantum principal component analysis to adaptively extract key feature components; based on the quantum reinforcement learning framework and quantum perception hash algorithm, combined with the quantum neural network, predicts channel interference in advance. When the RSSI monitors that the interference intensity exceeds the threshold of -90dBm, the intelligent optimal switching of multi-standard signals can be completed within 5ms.

4. The underground distribution room communication transmission rate adaptation and flow control method according to claim 1, characterized in that: The workflow of the data encoding module in step S2 includes: S2-4 uses a quantum superposition state coding algorithm to map the medium voltage carrier voltage signal to the quantum Bloch sphere, uses quantum phase coding to encode LoRa's spreading factors SF7-SF12, and uses quantum amplitude modulation to process 5G-RedCap's modulation coding strategy MCS0-MCS16. Combined with ternary quantum entanglement coding, it forms a 128-dimensional quantum feature vector. S2-5, using the space-time spinor representation and geometric algebra, the latitude and longitude of the sensor, altitude coordinates, and the temperature, voltage, and load rate of the equipment operating status are converted into geometric algebra elements with physical meaning for fusion; S2-6 introduces quantum fractional Fourier transform, realizes flexible conversion of signal characteristics in time and frequency domain by adaptively adjusting the transformation order, and uses quantum entanglement exchange network to establish non-local correlation of multi-standard signals; at the same time, adopts quantum Lorentz mapping and quantum measurement entropy normalization to enhance feature discrimination based on chaotic system and quantify feature uncertainty.

5. The underground distribution room communication transmission rate adaptation and flow control method according to claim 1, characterized in that: The workflow of the spatiotemporal Bayesian network module in step S2 includes: S2-7, using quantum tensor product networks to embed spatial distance, time delay, and interference intensity into quantum Hilbert space, using quantum phase estimation algorithms to parallelly calculate three-dimensional state transition probabilities, and constructing non-local causal relationships between nodes based on quantum entanglement exchange mechanisms; S2-8 uses a hidden Markov model in the form of a quantum superoperator, combined with quantum Monte Carlo sampling and error correction coding, to predict the delay distribution within the next five time windows in an exponential state space; S2-9, in terms of path optimization, a space-time curvature routing algorithm is designed based on gravitational wave theory, which converts node energy and load into space-time curvature tensor. The shortest path is solved by the Dijkstra algorithm optimized by quantum annealing, and the quantum field theory path integral model is used for calculation.

6. The underground distribution room communication transmission rate adaptation and flow control method according to claim 1, characterized in that: The workflow of the dynamic channel switching module in step S2 includes: S2-10 uses a hybrid modulation architecture assisted by quantum stochastic resonance, integrating frequency hopping frequency spectrum (FHSS) and direct sequence spread spectrum (DSSS), and introduces quantum tunneling effect to enhance signal anti-interference capability; S2-11, a dynamic decision-making model is constructed based on the quantum annealing reinforcement learning framework and the biological immune memory mechanism. When the SNR is less than 20dB, the quantum parallel search algorithm is triggered to globally scan the multidimensional channel space and simultaneously initiate the biological chemotaxis path optimization, shortening the channel switching cycle from 500ms to microseconds. Quantum entanglement is used to establish non-local correlation of channel states, enabling pre-evaluation and pre-configuration of backup channels. S2-12 adopts the quantum genetic Q-learning algorithm, encodes multiple switching strategies through quantum bit superposition states, and simulates the dynamic optimization strategy of biological gene evolution process; the module has a built-in adaptive environmental perception system and uses the quantum perception hash algorithm to quickly match the optimal strategy.

7. The underground distribution room communication transmission rate adaptation and flow control method according to claim 1, characterized in that: The multi-node interaction module in step S2 adopts the FedAvg algorithm framework of federated reinforcement learning, and its workflow includes: S2-13, the multi-node interaction module builds a quantum entanglement federated-biological symbiotic network architecture, which innovates the traditional federated learning framework. Each relay node encodes local training parameters into quantum entangled states, and achieves secure parameter aggregation through quantum teleportation and remote preparation protocols, fundamentally protecting data privacy. S2-14, uses the principle of biological symbiosis to build a defense mechanism, maps the parameter vector into an antigen-antibody reaction pattern, and resists malicious attacks through immune memory and clonal selection; innovates the quantum differential privacy protection mechanism, uses the unpredictability of quantum noise to enhance privacy protection, and reduces the risk of data leakage to 1 / 1000 of the traditional method under a privacy budget of ε=0.5; proposes a quantum game reinforcement learning framework, in which each node establishes non-local associations through quantum entanglement, achieves a Pareto optimal strategy without sharing private information, and designs a quantum state reward function to balance global optimization and local interests.

8. The underground distribution room communication transmission rate adaptation and flow control method according to claim 1, characterized in that: The implementation method of the improved generative adversarial network GAN in step S3 is: S3-1, Generator Construction and Implementation The generator is constructed using a hybrid attention mechanism-enhanced LSTM network, embedding a spatiotemporal module and a self-attention module. The spatiotemporal module uses multi-head attention to capture the long-term spatiotemporal dependencies of power flow data, while the self-attention module focuses on key features. The number of neurons in the three hidden layers is 128, 64, and 32, respectively. The gating and attention mechanisms work together to process complex temporal relationships. The system inputs one hour of multi-source heterogeneous traffic data, including power control, equipment monitoring, environmental sensing, weather warning, and grid topology dynamics data. It applies multimodal feature fusion encoding, builds semantic associations using knowledge graphs, maps the data into a unified feature space, and collaboratively optimizes the encoding model through federated learning, scaling it to the [0, 1] range. The output of the generator is the traffic forecast value for the next 10 minutes. The improved composite loss function is used in the training process, and the formula is: L gen =α(t)*L mes *(1+λ spatial *S adj )+β(t)*L TV *(1+λ temporal *T corr )+γ(t)*L adv *(1+λ hybrid *H factor ) (1) In formula (1), L mes is the mean square error, which measures the difference between the predicted value and the true value; L TV is the total variation loss, which is used to smooth the prediction results and prevent overfitting; L adv To combat the loss and encourage the generator to generate more realistic data, α(t), β(t), and γ(t) are dynamic weight coefficients that are optimized through a dynamic adjustment algorithm and adaptively changed according to the training stage. λ spatial ,λ temporal ,λ hybrid S is a newly added adjustment parameter that controls the influence of space, time and mixed factors on the loss function; adj L is a spatial correlation factor that considers the spatial correlation between the traffic data in the prediction area and the surrounding areas, explores the potential connection in the spatial dimension, and improves the rationality of the prediction; adv is the time correlation factor, which captures the dependency of traffic data in the time series and optimizes the prediction by using the time correlation between history and future. factor It is a mixed correlation factor that comprehensively considers the impact of various complex factors of time and space on traffic, so that the model can cope with more complex traffic change scenarios; the LSTM network parameters θ are optimized through the back propagation algorithm gen ,Set the prediction error target to ≤5%. When the average prediction error on the validation set exceeds this threshold, the network structure and parameters are quickly adjusted in combination with the meta-learning mechanism; S3-2, Discriminator Construction and Implementation The discriminator uses a multi-scale feature pyramid convolutional neural network (MSF-PCNN), which includes four convolutional layer groups, three pyramid pooling layers, and two fully connected layers. Each convolutional layer group contains parallel convolutional layers with different convolution kernel sizes to extract multi-scale local features. The pyramid pooling layer aggregates features at different granularities to obtain a more comprehensive feature representation, and finally the fully connected layer makes classification decisions. The improved Wasserstein-f divergence metric is used to measure the distribution difference between the predicted traffic and the actual traffic. The formula is: In formula (2), f(.) is a customized nonlinear mapping function that enhances the sensitivity to data distribution differences; P data is the true data distribution, P gen is the generated data distribution, Π(P data ,P gen ) is the set of all possible joint distributions between the two, γ is a joint distribution sampled from the joint distribution set, and (x, y) is a pair of data sampled from γ; when training the discriminator, a quantum computing acceleration mechanism is introduced, and the discrimination results of multiple samples are calculated in parallel using quantum superposition states. The RMSProp optimizer with adaptive learning rate is used to update the discriminator parameters θ. dis , by maximizing W f (P data ,P gen ), so that the discriminator can distinguish the distribution differences between the two more accurately; The discriminator outputs a three-category classification result, including determining whether the input traffic data is "real data," "generated data," or "suspected anomaly." Combining a causal inference algorithm, the discriminator further analyzes the causal relationship between the data identified as "suspected anomaly" and other related data to determine whether it is a true abnormal traffic pattern. Set the overall recognition accuracy target to ≥95% and evaluate the discriminator performance on the test set. If the target is not achieved, automatically adjust the network structure and parameters through reinforcement learning. S3-3, Generative Adversarial Training Initialize the network parameters of the generator and discriminator separately, using a quantum-inspired parameter initialization method that utilizes the superposition and entanglement properties of quantum states to generate better initial weights, and initialize the bias term to 0; During the training process, a dynamic task priority alternating training strategy is adopted; the ratio of the number of training rounds of the generator and discriminator is dynamically adjusted according to the current training stage and data characteristics; The generator introduces a conditional generative adversarial mechanism, and the loss function is: In formula (3), z is derived from the noise distribution p z The noise vector sampled from the conditional distribution p is c. c The conditional information sampled in , G(z,c) is the flow data generated by the generator based on the noise z and condition c, is a dynamic conditional fusion operation that fuses the conditional information after noise Q(c) processing. This function simulates the superposition characteristics of quantum states and enhances the interaction effect between conditional information and noise. D(·) is the discriminator, and the goal of the generator is to minimize L gen , generating realistic data that meets specific conditions; A(c) is an adaptive condition enhancement module, which dynamically adjusts the weights and dimensions of the generator input conditions according to the characteristics of the conditions, so that the generator focuses more on key condition information; M(c) is a condition information mapping module, which transforms the original conditions to generate a form that is more suitable for the discriminator input; L ms is a multi-scale feature penalty term, which calculates the feature differences of generated data at different scales, making the generated data closer to the real data in terms of details and overall structure, thereby improving the quality of generated data; qe It is a quantum entanglement weighting term that simulates the quantum entanglement phenomenon, assigns weight to the association between conditional information and generated data, and strengthens the constraint effect of conditions on the generation process; ms and λ qe are the weight coefficients of the multi-scale feature penalty term and the quantum entanglement weighted term, which are dynamically adjusted through reinforcement learning to balance the contribution of each part in the loss function; The discriminator uses a multi-task learning loss function: In formula (4), D STQ It is a spatiotemporal quantization discriminator that processes input through the spatiotemporal feature extraction layer, quantum enhancement layer and attention fusion layer architecture; x is the distribution of real data p data The real traffic data sampled in L aux is the auxiliary task loss, λ(t) is the dynamic weight coefficient, and the discriminator’s goal is to maximize L dis , enhance the discriminative ability and understanding of data features through multi-task learning; L qef It is the loss of quantum dining characteristics; L sa is the spatiotemporal attention loss, which constrains the discriminator to focus on the key spatiotemporal domain; γ and β are the weight coefficients of quantum feature loss and attention loss, respectively, which are dynamically optimized by the reinforcement learning agent; The two are continuously optimized during the confrontation process, which improves the quality of data generated by the generator and enhances the recognition ability of the discriminator. S3-4, Overall Operation and Evaluation Deploy the trained generator and discriminator to a distributed system based on edge computing and blockchain; Deploy lightweight edge computing units at each data collection node to achieve local preprocessing and preliminary judgment of data; use blockchain technology to record the source of data, processing process and model decision results to ensure data credibility and traceability; The generator is used to predict traffic flow in the next 10 minutes, and the discriminator monitors traffic data in real time and identifies abnormal traffic patterns; Construct a multi-dimensional dynamic evaluation system. In addition to calculating the prediction error of the generator and the recognition accuracy of the discriminator, it also introduces data diversity indicators, model robustness indicators, and model interpretability indicators. Regularly collect data from actual applications and use the federated learning framework to summarize the evaluation results of each node without leaking privacy, and comprehensively evaluate model performance. If performance degrades, the problem is located through the automatic model diagnosis system, and optimization is carried out using transfer learning and incremental learning techniques.

9. The underground distribution room communication transmission rate adaptation and flow control method according to claim 1, characterized in that: The working method of the fuzzy PID algorithm in step S4 is: S4-1, variable definition and quantification upgrade Network load and delay fluctuation are defined as core input variables, with packet arrival rate and link jitter coefficient introduced as auxiliary references. Each variable is quantized to the domain of [-3, 3]. Network load is classified into levels based on bandwidth utilization, packet loss rate, and arrival rate. Delay fluctuation is comprehensively evaluated using the delay standard deviation and jitter coefficient. When the delay standard deviation is less than 5ms and the jitter coefficient is low, it is considered "small fluctuation." The output variable, congestion window adjustment step size, takes values of -2KB, 0, and +2KB, and is also mapped to the above domain. S4-2, Fuzzy Rules and Reasoning Enhancement A sophisticated fuzzy rule base was constructed, with rules that take into account link jitter and packet arrival rate. When network load is high, latency fluctuations are large, and link jitter is severe, the congestion window adjustment step size is set to -2KB. An improved Mamdani fuzzy inference method was used to dynamically adjust membership function parameters to adapt to network state changes. A weighted average method was used for defuzzification, with weights assigned based on rule importance to quickly determine the adjustment step size. S4-3, dynamic response optimization When bandwidth utilization exceeds 80% and latency fluctuation is greater than 15ms, the window reduction rule is triggered. PID parameters are automatically optimized through a reinforcement learning-based strategy. The adaptive learning rate adjustment formula is: In formula (5), α is the learning rate, β is the error factor, and ∈ is the minimum value to avoid the denominator being zero. This formula dynamically adjusts the learning rate based on historical error changes to ensure the stability and convergence speed of the system under different states. is to simulate the temperature decay during quantum annealing; γ is the control exploration amplitude; T is the control exploration period; S4-4, Thermoelectric Power Generation Module Design Using a nanocomposite device based on the Seebeck effect, the device activates when the surface temperature difference reaches 5°C or above. Based on the traditional Seebeck effect to calculate output power, efficiency improvement factors related to nanostructures and materials are introduced. Through material screening, structural design, and intelligent temperature control and heat dissipation systems, the output power is guaranteed to be no less than 50μW. It is equipped with an energy management circuit to store and distribute electrical energy. S4-5, vibration power generation module performance improvement Using piezoelectric ceramic materials with a gradient polarization structure, it works at a vibration frequency of 10-100Hz and an amplitude of 0.1mm or above; designing a vibration frequency adaptive matching circuit to automatically adjust the capacitance and inductance parameters to put the piezoelectric ceramics in the optimal resonant state; at the same time, adopting a multi-level energy capture structure to enhance the response ability to different vibrations.

10. A communication transmission rate adaptation and flow control system for an underground power distribution room, characterized in that: It includes a multi-standard communication module, a reinforcement learning relay module, an adaptive rate adaptation module, an intelligent flow control module, and an energy efficiency control module, among which: The multi-standard communication module integrates three communication modes: medium-voltage carrier, LoRa, and 5G-RedCap. These three communication modes are respectively suitable for short-range stable transmission, long-range low-power transmission, and high-speed and low-latency scenarios. By normalizing quantum state parameters, a three-dimensional communication link matrix is constructed to achieve parallel transmission and intelligent switching of multi-standard signals. The reinforcement learning relay module includes five submodules: the spatiotemporal feature fusion module achieves high-precision synchronization and feature fusion; the data encoding module uses quantum algorithms to process signals and information; the spatiotemporal Bayesian network module predicts delays and optimizes paths; the dynamic channel switching module achieves fast switching; and the multi-node interaction module ensures data privacy and optimizes strategies. The adaptive rate adaptation module uses an improved generative adversarial network (GAN). The generator predicts traffic, the discriminator identifies data and detects anomalies, and dynamically adjusts the transmission rate in steps of 1 Mbps based on network load and service priority. The intelligent flow control module adopts a dual closed-loop architecture. The outer loop uses a fuzzy PID algorithm to adjust the congestion window, and the inner loop reserves the power control signal bandwidth through priority queue management. The energy efficiency optimization module adjusts the transmission power based on energy sensing, triggering low-power mode when the power level drops below 20%. It also integrates temperature difference and vibration power generation modules to achieve self-power supply, reducing energy consumption by more than 30%. The multi-standard communication module processes the three standard signals through quantum state processing to construct a link matrix, and then outputs it to the reinforcement learning relay module; The five submodules of the reinforcement learning relay module collaboratively plan the optimal path, and the processed signal is transmitted to the adaptive rate adaptation module; The adaptive rate adaptation module predicts the flow rate based on GAN and adjusts the rate, and then transmits the signal to the intelligent flow control module; The intelligent flow control module accurately controls the flow through a dual closed-loop architecture, and the processed signal is returned to the multi-standard communication module for transmission; The energy efficiency optimization module monitors the power consumption and node location, adjusts the power and self-powers the modules to support the operation.

Images