Downlink NOMA system power distribution method based on deep learning
By generating a training dataset through an improved particle swarm optimization algorithm and combining it with a cascaded deep neural network, the real-time and optimality challenges of the NOMA power allocation method are solved, enabling fast and accurate power allocation in dynamic wireless environments and meeting the requirements of low-latency communication.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- JILIN UNIVERSITY
- Filing Date
- 2026-02-05
- Publication Date
- 2026-05-12
AI Technical Summary
Existing NOMA power allocation methods lack real-time performance in dynamic wireless environments, making it difficult to meet the requirements of low-latency communication. Furthermore, traditional optimization methods have high computational complexity and are difficult to deploy online, while data-driven methods are limited by the quality of training data and cannot achieve rapid decision-making.
An improved particle swarm optimization algorithm is used to generate a high-quality training dataset, and a cascaded deep neural network model is used to achieve a fast mapping from channel state to optimal power allocation. Combined with effective capacity theory, a two-stage decoupled architecture of offline data generation and online deep learning fast inference is constructed.
It achieves real-time power allocation within millisecond latency, meets the requirements of low-latency communication, reduces data acquisition and storage costs, improves the practicality and deployability of the solution, ensures statistical latency QoS guarantee, and adapts to dynamic wireless environments.
Smart Images
Figure CN122028181A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of wireless communication technology, specifically relating to a power allocation method for a downlink NOMA system based on deep learning. Background Technology
[0002] To meet the stringent requirements of 5G and future mobile communication systems for massive connectivity, ultra-high capacity, and low latency, Non-Orthogonal Multiple Access (NOMA) technology has become a research hotspot due to its superior spectral efficiency. Unlike traditional orthogonal multiple access, NOMA achieves simultaneous service to multiple users on the same time-frequency resource block by performing non-orthogonal superposition of power domain or code domain data for multiple users at the transmitter and decoding using continuous interference cancellation technology at the receiver. Downlink power domain NOMA is a simple and efficient solution; its core lies in how to allocate appropriate power to coexisting users to balance fairness among users, total system throughput, and individual quality of service.
[0003] While NOMA holds immense potential, its performance advantages are highly dependent on precise, adaptive power allocation. Existing power allocation schemes face the following significant challenges when dealing with dynamic wireless environments and stringent QoS guarantee requirements: (1) Traditional optimization methods lack real-time performance and are difficult to adapt to dynamic environments: Most existing studies use traditional mathematical tools such as convex optimization, fractional programming, or the Lagrange multiplier method to solve power allocation problems. Although these methods can theoretically find (sub)optimal solutions, they usually involve complex iterative calculations and a large number of floating-point operations. In real-world scenarios where channel state information changes rapidly (such as high-speed movement or complex fading), the real-time computational complexity of these algorithms is extremely high, making it difficult to complete the solution and allocation within the channel coherence time. This results in a delayed system response, which cannot meet the millisecond or even microsecond-level decision-making requirements of scenarios such as ultra-reliable low-latency communication.
[0004] (2) Difficulty in solving the effective capacity optimization problem for statistical QoS guarantee: With the emergence of new services such as URLLC, traditional indicators such as "maximizing the sum rate" are no longer sufficient to characterize the system's latency performance. Effective capacity theory, as a tool for characterizing the maximum sustainable arrival rate under given latency constraints, has become a key indicator for evaluating the performance of statistical QoS guarantee. However, the NOMA power allocation problem, which aims to maximize the total effective capacity of the system, has more significant non-convex and nonlinear characteristics in its objective function, and usually does not have a closed-form solution. Existing methods rely heavily on metaheuristic algorithms such as genetic algorithms and particle swarm optimization for search, which have slow convergence speed and huge computational overhead, and are only suitable for offline simulation research, and cannot be deployed in online systems that require fast response.
[0005] (3) Data-driven intelligent methods face a bottleneck of high-quality training data: In recent years, machine learning and deep learning methods have been introduced to solve the problem of poor real-time performance of traditional optimization. These methods train a neural network to learn the mapping relationship from channel state to power allocation coefficient, thereby achieving millisecond-level fast decision-making in the inference stage. However, the performance of these methods is severely limited by the quality of training data. The "label dilemma" they face is that if traditional suboptimal algorithms (such as methods based on fixed ratios) are used to generate training labels, the upper limit of the performance of the trained model is locked, and it is difficult to surpass its "teacher" algorithm and obtain breakthrough gains; if metaheuristic algorithms that can approximate the global optimum are used to generate training labels, although high-quality data can be obtained, the data generation process itself (such as multiple iterative evaluations of PSO) is extremely time-consuming and computationally expensive, and the fundamental contradiction of "slow online computing" is still not solved. This makes the construction of a large-scale, high-quality training dataset itself an efficiency bottleneck. Summary of the Invention
[0006] To overcome the aforementioned problems, this invention provides a power allocation method for downlink NOMA systems based on deep learning, which is an intelligent NOMA power allocation method based on maximizing effective capacity. This invention employs a two-stage decoupled architecture: offline high-quality data generation + online deep learning fast inference. In the offline stage, an improved particle swarm optimization algorithm is used to generate a globally optimal or near-optimal power allocation dataset with the goal of maximizing the total effective capacity of NOMA users. In the online stage, this dataset is used to train a deep neural network model to achieve a fast mapping from channel states to optimal power allocation coefficients, thereby achieving millisecond-level real-time power allocation while ensuring statistical service quality.
[0007] A power allocation method for a downlink NOMA system based on deep learning includes the following steps: Step 1, obtain The data, specifically: Step 1.1: Generate a channel gain sample pair for the user at time slot t. , ),in Let be the channel gain of user 1 in time slot t. Let t be the channel gain of user 2; denote the channel gain sample pairs corresponding to M time slots as channel gain samples, and extract the original channel components of the channel gain samples: the complex channel gain from the base station to user 1 in each time slot. and the multiple channel gain from base station to user 2 Where t = 1, 2, 3...M; Step 1.2, the channel gain samples are processed as follows to obtain the channel gain samples. data: The sum of squared complex channel gains of all channel gain samples is concatenated to form the input feature vector of the channel gain samples. : , The first M dimensions represent the sum of squared channel gains sequence for user 1, and the last M dimensions represent the sum of squared channel gains sequence for user 2. Input feature vector of channel gain sample Perform Z-score standardization: ; in , These are the statistical feature vectors of channel state information for the channel gain samples. The mean and standard deviation; Input feature vector Z-score standardized results; Step 2: Construct a cascaded deep neural network model, which is formed by sequentially connecting the trained first-level neural network model and the second-level neural network model. The structure of the first-level neural network model is as follows: The input layer consists of 200 neurons; A fully connected layer; The output layer consists of two neurons; The second-level neural network model has the following structure: The input layer consists of two neurons; Three hidden layers, each using the ReLU activation function and applying Dropout regularization, with a dropout rate of 0.2; The output layer consists of one neuron and uses linear activation; Step 3, sample the channel gain. Data is input into a trained cascaded deep neural network model; the model outputs the optimal power allocation coefficients predicted from the channel gain samples. ; Step four, the base station according to Calculate the actual power allocated to user 1 within the M time slots. The actual power allocated to user 2 .
[0008] In step 1.1, the 4-dimensional original components of the channel gain sample are... , , , Reconstruct the complex channel gain summation for each time slot in the channel gain sample: ;in , These are the complex channel gains from the base station to user k in the corresponding time slot. The real and imaginary parts of .
[0009] The training process of the cascaded deep neural network model is as follows: Step 1, data preparation, specifically includes the following: S1.1: Define the downlink NOMA system model Received signal model: The received signal of user k in time slot t. Represented as: , in: This is the encoded signal sent by the base station to user k, where k = 1 or 2, representing the kth user; Let be the complex channel gain from the base station to user k, which follows independent Rayleigh fading; Let M be the transmit power that the base station allocates to user k within M time slots. Let M be the total transmit power allocated by the base station to all users over M time slots, which satisfies the following constraints: ; It is additive white Gaussian noise; Channel State Information: Definition According to the Rayleigh fading model, Follow the mean The exponential distribution of has a probability density function as: , Assuming the receiver noise power of the user is normalized, and strong and weak users are distinguished by the strength of the channel gain between the base station and the user, then the instantaneous service rate of user k is... for: Strong user refers to the instantaneous service rate of user 1. for: , The instantaneous service rate of weak users, i.e., user 2. for: , in: Let be the channel gain of user 1 in time slot t. Let be the channel gain for user 2 in time slot t; S1.2: Define statistical QoS metrics and optimization objectives based on effective capacity. Effective capacity formula: For user k with independent, co-distributed block fading channels and a specific transmission mode, its effective capacity is... for: , in , where is the latency QoS index for user k; The probability of successfully transmitting information to user k; where: Probability of User 1 successfully transmitting information : , Probability of User 2 successfully transmitting information : , Where the threshold The joint channel requirements that User 1 must meet when using SIC decoding; The preset transmission rate for user 1, The preset transmission rate for user 2; It is the probability density function; System optimization goals: Will and Substituting the expressions into the effective capacity respectively The calculation formula yields the total effective capacity of the system. : , S1.3: Implement interval-based particle swarm optimization to generate optimal labels. S1.31: Optimize the interval [0, The interval is divided into four sub-intervals, where: , , , S1.32: Batch generation and feature extraction of channel samples: Generate a channel gain sample pair for the user at time slot t. , Each M time slot corresponds to a channel gain sample pair, which is denoted as a group of channel gain samples. The mean squared channel gain of this group of channel gain samples is calculated. and As the statistical feature vector of the channel state information of this group of samples ; S1.34: Perform the following operations to find the optimal solution. : S1.341: First interval Optimization: Initialize a PSO particle swarm within this interval, so as to As the fitness function, PSO iteration is performed to record the optimal solution found within the interval. and ,in It is the optimal power allocation value within this interval, i.e., the optimal value within this interval. ,and- It is the maximum total effective system capacity value in this interval; Second interval Optimization: Initialize a PSO particle swarm within this interval, so as to Using the fitness function, perform PSO iteration to obtain the optimal solution. and ,in It is the optimal power allocation value for this interval, i.e., the optimal value for this interval. ,and- It is the maximum total effective system capacity value in this interval; Third interval Optimization: Initialize a PSO particle swarm within this interval, so as to Using the fitness function, perform PSO iteration to obtain the optimal solution. and ,in It is the optimal power allocation value for this interval, i.e., the optimal value for this interval. ,and- It is the maximum total effective system capacity value in this interval; S1.342: Perform a physical feasibility check: like ,think invalid; like ,think invalid; Remove invalid data and the maximum total effective system capacity corresponding to the invalid data from the three sub-intervals. Select the maximum total effective system capacity corresponding to the largest valid data. The value is the globally optimal value under the given set of channel gain sample states. ,Right now ; S1.343: Calculate the optimal power allocation coefficients for this set of channel gain samples: ; S1.4: Constructing the final dataset The corresponding channel gain samples Statistical feature vector of channel state information and optimal power allocation coefficient Pairing together to form a data sample ; obtain Groups of data samples form a large-scale dataset D; S1.5: Tag Data Preparation: Read the statistical feature vector of channel state information corresponding to each group of channel gain samples from dataset D. and optimal power allocation coefficient , take each set of data and Together they serve as tag data; S1.6: Calculate the corresponding channel gain samples for each group according to step 1.2. data; S1.7: Training process: Training the first-level neural network model: Will The z-data corresponding to each group of channel gain samples is used as the input information of the model. The statistical feature vector of channel state information in the label data corresponding to each group of channel gain samples is used. The output data of the model are input together into the first-level neural network model for training until the loss function of the first-level neural network model converges, and the training ends. The first-level neural network model uses the Adam optimizer for parameter updates, where the initial learning rate is... The learning rate uses exponential decay, decreasing to 0.95 times its original value every 50 training steps. The first-level neural network model uses the mean absolute error as the loss function: , in This represents the loss value of the first-level neural network model. The number of groups of channel gain samples used for training, For the first The target set value of the statistical feature vector of the base station for the channel state information of user 1 corresponding to the group of channel gain samples. For the first The target set value of the statistical feature vector of the base station for the channel state information of user 2 corresponding to the group of channel gain samples; Training the second-level neural network model: Will Group channel gain samples corresponding As input information to the model, the optimal power allocation coefficients from the tag data corresponding to each group of channel gain samples are used. The output data of the model are fed into the second-level neural network model for training until the loss function of the second-level neural network model converges, at which point the training ends. The second-level neural network model uses the Adam optimizer for parameter updates, where the initial learning rate is... The learning rate is a fixed learning rate; The second-level neural network model uses standard mean squared error as the loss function: , in This represents the loss value of the second-level neural network model. The number of groups of channel gain samples used for training. The target setpoint for the optimal power allocation coefficient; The output of the second-level neural network model Predicted power allocation coefficients for group channel gain samples; Training cascaded networks: The original training material for the first-level neural network model Optimal power allocation coefficients in the data and labeled data used to train the second-level neural network model The input and output information of the cascaded network are respectively used as input and output information to train the cascaded network until the loss function of the cascaded network converges, and the training ends. The cascaded network uses the Adam optimizer for parameter updates, where the initial learning rate... The learning rate uses exponential decay, decreasing to 0.95 times its original value every 30 training steps. The cascaded network employs a hybrid loss function. : , in: , where is the mean square error; For Huber's losses; This is an L2 regularization term.
[0010] The beneficial effects of this invention are: This invention effectively solves the core challenge of balancing real-time performance and optimality in power allocation within NOMA systems through an innovative data-driven and model-driven deep fusion architecture. First, it creatively decouples global optimization from rapid inference: an improved partitioned particle swarm optimization (PSO) algorithm generates high-quality training data approaching the global optimum, providing a solid performance ceiling for the model. Then, a cascaded deep neural network model is used to achieve real-time decision-making with microsecond-level latency in forward propagation, completely overcoming the fundamental contradiction of traditional iterative algorithms' long computation time and inability to meet low-latency communication requirements. Second, this method achieves data efficiency—by deeply fusing the effective capacity theoretical model (model-driven) with deep learning (data-driven), the network can learn from deep physical laws, thus requiring only a small number of data samples to train a high-performance model, significantly reducing data acquisition and storage costs and improving the practicality and deployability of the solution. Third, the system directly uses effective capacity as the optimization target, ensuring that the power allocation strategy strictly meets the statistical latency QoS guarantees for services such as URLLC, enhancing the advanced nature and practicality of the solution. Furthermore, the designed cascaded neural network structure is unique. The first-level network adopts a minimalist linear design without hidden layers, while the second-level network performs deep nonlinear mapping. The two are trained collaboratively, ensuring high accuracy while possessing excellent generalization ability and robustness, enabling adaptation to dynamically changing wireless environments. Finally, the entire solution is engineering-friendly, with a small number of trained model parameters and high computational efficiency, allowing for easy integration into existing base station platforms. This provides an efficient and reliable technical path for intelligent resource management in 5G-Advanced and future mobile communication networks. Attached Figure Description
[0011] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the content of the embodiments of the present invention and these drawings without creative effort.
[0012] Figure 1 This is a schematic diagram of the overall system architecture and process of the method of the present invention; Figure 2 This is a schematic diagram of the downlink transmission model of the NOMA system of the present invention; Figure 3 This is a schematic diagram of the structure of the cascaded deep neural network model of the present invention; Figure 4 This is a flowchart illustrating the workflow of the online reasoning and execution module of this invention. Figure 5 This is the convergence curve for training the deep neural network of this invention; Figure 6This is a scatter plot showing the prediction performance of the deep neural network in this invention. Figure 7 This is a box plot comparing the single computation time of PSO and deep learning in this invention. Detailed Implementation
[0013] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.
[0014] A power allocation method for downlink NOMA systems based on deep learning, such as Figure 1 As shown, it includes the following: Step S1: Generate the optimal power allocation dataset based on effective capacity theory and improved particle swarm optimization, specifically including the following: S1.1: Define the downlink NOMA system model Define a downlink power domain NOMA system scenario, which includes one base station and two users, with the base station and all users randomly distributed in a specific geographical area; where the average channel gain between user 1 and the base station is greater than that between user 2 and the base station, user 1 is defined as a strong user; and the average channel gain between user 2 and the base station is less than that between user 1 and the base station, user 2 is defined as a weak user. Received signal model: The received signal of user k in time slot t. Represented as: , in: This is the encoded signal sent by the base station to user k, where k = 1 or 2, representing the kth user; Let be the complex channel gain from the base station to user k, which follows independent Rayleigh fading; Let M be the transmit power that the base station allocates to user k within M time slots. Let M be the total transmit power allocated by the base station to all users over M time slots, which satisfies the following constraints: ; It is additive white Gaussian noise; Channel State Information: Definition Let be the sum of squares of the complex channel gain of user k in time slot t, according to the Rayleigh fading model. Follow the mean The exponential distribution of has a probability density function as: , in Represents the mathematical expectation; Assuming the receiver noise power of the user is normalized, and strong and weak users are distinguished by the strength of the channel gain between the base station and the user, then the instantaneous service rate of user k is... for: Strong user refers to the instantaneous service rate of user 1. for: , The instantaneous service rate of weak users, i.e., user 2. for: , in: Let be the channel gain of user 1 in time slot t. Let be the channel gain for user 2 in time slot t; S1.2: Define statistical QoS metrics and optimization objectives based on effective capacity. Effective capacity formula: For channels with independent and co-distributed block fading, using a specific transmission mode (the specific transmission mode is On-Off transmission mode, i.e., the transmission rate is [value missing] if the user can successfully decode), the transmission rate is [value missing]. For user k (whose effective capacity is 0 if the transmission rate fails), the effective capacity is... for: , in , where is the latency QoS index for user k; For the instantaneous service rate of user k, The probability of successfully transmitting information for user k; that is, the probability that user k's instantaneous service rate meets its target instantaneous service rate requirement under the current channel and power allocation; this is a key probability related to channel fading distribution, power allocation, and target instantaneous service rate.
[0015] The probability that user k successfully transmits information. The calculation, where: Probability of User 1 successfully transmitting information : , Probability of User 2 successfully transmitting information : , Where the threshold The joint channel requirements that User 1 must meet when using SIC decoding; The preset transmission rate for user 1, The preset transmission rate for user 2; Let be the probability density function. for The probability density function, for The probability density function; System optimization goals: Let the optimal power allocation factor for user 1 be... ,but ,Will and Substituting the expressions into the effective capacity respectively The calculation formula is used, and the results are summed to obtain the total effective capacity of the system. : , S1.3: Implement interval-based particle swarm optimization to generate optimal labels. S1.31: Based on the NOMA decoding conditions, , , and The probability of successfully transmitting information is converted into three key power thresholds. , and The optimized interval [0, The interval is divided into four sub-intervals, where: , , , By performing PSO iterations across all sub-intervals, further comparisons can be made to arrive at the optimal result.
[0016] S1.32: Batch generation and feature extraction of channel samples: Generate a channel gain sample pair for the user at time slot t. , To reflect real-world scenarios with both strong and weak users, the average channel gain for user 1 is... The average channel gain of user 2 ; S1.33: Group processing is adopted: the channel gain sample pairs corresponding to every 100 time slots t are recorded as a group of channel gain samples. The samples within the group share the same statistical characteristics, and the mean squared channel gain of a group of channel gain samples is calculated. and As the statistical feature vector of the channel state information of this group of samples ; S1.34: Interval Parallel PSO Optimization: Statistical feature vector of channel state information for each group of channel gain samples By maximizing To obtain the optimal power allocation strategy; perform the following operations to find the optimal : First interval Optimization: Initialize a PSO particle swarm within this interval, so as to As the fitness function, PSO iteration is performed to record the optimal solution found within the interval. and ,in It is the optimal power allocation value within this interval, i.e., the optimal value within this interval. ,and- It is the maximum total effective system capacity value in this interval; Second interval Optimization: Initialize a PSO particle swarm within this interval, so as to Using the fitness function, perform PSO iteration to obtain the optimal solution. and ,in It is the optimal power allocation value for this interval, i.e., the optimal value for this interval. ,and- It is the maximum total effective system capacity value in this interval; Third interval Optimization: Initialize a PSO particle swarm within this interval, so as to Using the fitness function, perform PSO iteration to obtain the optimal solution. and ,in It is the optimal power allocation value for this interval, i.e., the optimal value for this interval. ,and- It is the maximum total effective system capacity value in this interval; Because of the power allocation requirements of the NOMA system for strong and weak users, the fourth interval We can simply ignore it; Perform a physical feasibility check: like This indicates that the allocated power for user 1 is too low, therefore it is considered... invalid; Since the second interval does not contain 0 or Such power allocation boundary values, therefore There is no need to consider the problem of insufficient power allocation; like This indicates that the allocated power for user 2 is too low, therefore it is considered... invalid; Remove invalid data and the maximum total effective system capacity corresponding to the invalid data from the three sub-intervals. Select the maximum total effective system capacity corresponding to the largest valid data. The value is the globally optimal value under the given set of channel gain sample states. ,Right now ; Calculate the optimal power allocation coefficients for this set of channel gain samples: ; S1.4: Constructing the final dataset The channel gain samples in this group are in each time slot Statistical feature vector of channel state information and optimal power allocation coefficient Pairing together to form a data sample The data samples consisting of 100 sets of channel gain samples form a large-scale dataset D, which is used in step S2. Step S2: Power Allocation Model Based on Deep Neural Network like Figure 3 As shown, this step utilizes the dataset D generated in step S1 to construct and train a cascaded deep neural network model to learn a complex nonlinear mapping from raw channel information to optimal power allocation coefficients; specifically, it includes the following: S2.11: Extract the original channel components of each group of channel gain samples in dataset D: complex channel gain from base station to user 1 in each time slot. and the multiple channel gain from base station to user 2 in each time slot ; S2.12: Extract the 4-dimensional original components of each group of channel gain samples [ , , , Reconstruct the data and calculate the sum of squared channel gains for each time slot in each group of channel gain samples: ;in , These are the complex channel gains from the base station to user k in the corresponding time slot. The real and imaginary parts of .
[0017] S2.13: Concatenate the sum of squared channel gains in each time slot of each group of channel gain samples to form the input feature vector of that group of channel gain samples. ,in: , The first 100 dimensions represent the sum of squared channel gains sequence for user 1, and the last 100 dimensions represent the sum of squared channel gains sequence for user 2. S2.14: Tag Data Preparation: Read the statistical feature vector of channel state information corresponding to each group of channel gain samples from dataset D. and optimal power allocation coefficient Together, they serve as tag data; S2.15: Data Standardization: For the Input feature vector of group channel gain samples Perform Z-score standardization: ; in , The first Statistical feature vector of channel state information of group channel gain samples The mean and standard deviation; Input feature vector Z-score standardized results; S2.2: Construct a cascaded deep neural network model, which is formed by sequentially connecting the first-level neural network model and the second-level neural network model; The first-level neural network model has the following input information: Data, output information is the first The predicted value of the statistical feature vector of the base station for the channel state information of user k in the group channel gain samples. ; Its structure is as follows: The input layer consists of 200 neurons; A fully connected layer; The output layer consists of two neurons; The first-level neural network model uses the Adam optimizer for parameter updates, where the initial learning rate is... The learning rate uses exponential decay, decreasing to 0.95 times its original value every 50 training steps. The first-level neural network model uses the mean absolute error as the loss function: , in The loss value is for the first-level neural network model Net1. The number of groups of channel gain samples used for training, For the first The target set value of the statistical feature vector of the base station for the channel state information of user 1 corresponding to the group of channel gain samples. For the first The target set value of the statistical feature vector of the base station for the channel state information of user 2 corresponding to the group of channel gain samples; The second-level neural network model has the following input information: Data, output information is the first Predicted power allocation coefficients in group channel gain samples ; Its structure is as follows: The input layer consists of two neurons; Three hidden layers, each using the ReLU activation function and applying Dropout regularization, with a dropout rate of 0.2; The output layer consists of one neuron and uses linear activation; The second-level neural network model uses the Adam optimizer for parameter updates, where the initial learning rate... The learning rate is a fixed learning rate; The second-level neural network model uses standard mean squared error as the loss function: , in This represents the loss value of the second-level neural network model, Net2. The number of groups of channel gain samples used for training. The target setpoint for the optimal power allocation coefficient; After the first-level neural network model and the second-level neural network model are trained separately, they are sequentially connected to form a cascaded network. The input information of the cascaded network is... The final output information of the cascaded network is the standardized predicted value of the power allocation coefficient. The cascaded network uses the Adam optimizer for parameter updates, where the initial learning rate... The learning rate uses exponential decay, decreasing to 0.95 times its original value every 30 training steps. The cascaded network employs a hybrid loss function to balance accuracy and robustness: , in: It is a mixed loss function; , where is the mean square error; To enhance robustness to outliers using Huber loss; This is an L2 regularization term to prevent overfitting; S2.4: Perform phased model training Phase 1: Pre-training the first-level neural network model: The statistical feature vectors of channel state information from the z-data corresponding to 100 sets of channel gain samples and the tag data corresponding to 100 sets of channel gain samples are used to... The first-level neural network model is trained by inputting the loss function of the first-level neural network model as both input and output information, until the loss function of the first-level neural network model converges, at which point the training ends; this enables the first-level network to accurately learn a robust mapping from the channel sum of squares sequence to its statistical mean.
[0018] Phase Two: Training the Second-Level Neural Network Model The 100 sets of channel gain samples corresponding to each time slot The optimal power allocation coefficients in the tag data corresponding to 100 sets of channel gain samples The information is used as both input and output to train the second-level neural network model until the loss function of the second-level neural network model converges, at which point the training ends. This allows the second-level network to learn a nonlinear decision function that transforms channel statistical characteristics into the optimal power allocation strategy.
[0019] Phase 3: Training the cascaded network: Unlock and connect all parameters of the first and second level networks; The original training material for the first-level neural network model Optimal power allocation coefficients in the data and labeled data used to train the second-level neural network model The input and output information of the two levels of the cascaded network are used as input and output information respectively for training until the loss function of the cascaded network converges, at which point training ends. By backpropagating the gradient throughout the network, the parameters of the two levels of the network are jointly optimized and fine-tuned, improving the overall mapping accuracy.
[0020] This completes the training of a high-performance, low-complexity intelligent power allocation deep learning model. This model can approximate the computationally intensive PSO optimization effect of step S1 with extremely low latency in forward propagation, providing a core algorithm module for subsequent online real-time deployment.
[0021] Step S3: Online real-time power allocation and execution based on the trained cascaded deep neural network model: like Figure 4 As shown, the trained and satisfactory cascaded deep neural network model is deployed in the computing unit on the base station side.
[0022] The online workflow is as follows: S3.1: Channel Estimation The base station estimates the complex channel gain from the base station to user k in real time for each time slot using uplink pilot signals. ; S3.2: Feature Extraction The required feature vector for calculating the current channel state Specifically, it refers to the z-data of a set of channel gain samples within M time slots.
[0023] S3.3: Input the z data into the trained cascaded deep neural network model; the model outputs the predicted optimal power allocation coefficients for the M time slots within microseconds after one forward propagation. ; S3.4: Base station according to Calculate the actual power allocated to user 1 within the M time slots. The actual power allocated to user 2 Subsequently, the data from the two users were encoded using power domain superposition according to the NOMA protocol and then transmitted.
[0024] The simulation environment parameters of the NOMA system and the training hyperparameters of the cascaded deep neural network of this invention were set according to Table 1. To verify the performance of the method of this invention, the PSO algorithm was selected as the benchmark scheme. Under the premise of keeping the communication environment parameters consistent, each algorithm was trained separately, and their respective optimal policies were obtained. During the policy execution phase, the performance of the optimal policy network of each algorithm under the same test environment was observed, comparing the performance of the cascaded neural network itself with that of the PSO algorithm. The simulation results are as follows: Figure 6 , 7 As shown.
[0025] Depend on Figure 5 As can be seen, the training loss curve of this cascaded neural network exhibits excellent convergence characteristics and generalization ability. Both the training loss (solid line) and the test loss (dashed line) decrease rapidly within the first 20 epochs, indicating that the model quickly learns effective features. The two curves remain highly similar and decrease synchronously throughout the training process, without significant separation, proving that the model does not overfit and possesses good generalization performance. The final test loss stabilizes at a low level of approximately 0.0015, indicating that the cascaded network successfully learned a high-precision end-to-end mapping from the original channel features to the optimal power allocation, verifying the effectiveness of the "offline PSO data generation + online DNN fast inference" architecture of this invention.
[0026] Depend on Figure 6 As can be seen, the scatter plot of prediction performance visually demonstrates the superior power allocation accuracy of the cascaded neural network. The vast majority of prediction points are closely distributed near the ideal prediction diagonal (red dashed line), indicating a high degree of agreement between the network output and the globally optimal label generated by PSO. The concentrated distribution of data points without significant deviation verifies that the model maintains stable and accurate prediction capabilities across the entire range of power allocation coefficients. The minimal number of discrete points indicates good robustness to outliers. This result fully demonstrates that, through the "interval-based PSO data generation" and "cascaded network collaborative training" methods of this invention, deep learning models can accurately approximate the globally optimal solution of traditional optimization algorithms with extremely low computational cost in forward propagation.
[0027] A comparison chart of the allocation rate calculated using the method of this invention and using PSO alone is shown below. Figure 7 As shown, by Figure 7 As can be seen, the method of this invention is significantly faster than the PSO method. The median PSO time is approximately 0.135 seconds, concentrated in the range of 0.133–0.137 seconds, stable but relatively slow. The median deep learning method is approximately 0.001–0.002 seconds, mostly in the millisecond range, representing a speed improvement of approximately 67–135 times; a small number of outliers exist (the highest being approximately 0.022 seconds), possibly caused by initialization or preprocessing. These results fully demonstrate that the deep learning method is suitable for real-time applications, while the PSO method is suitable for offline optimization. It is recommended to prioritize the use of deep learning methods in real-time scenarios.
[0028] Table 1: Parameter Settings for NOMA System Training Set and Cascaded Deep Neural Network Training Stage hyperparameters describe Simulation value settings bandwidth bandwidth 1 (MHz) ue_num Number of users 2 bs_num Number of base stations 1 power_total Base station power limit 10(W) channel_user1 User 1 Channel Mean 2.1-4 channel_user2 User 2 Channel Mean 1-1.9 code_rate Normalized bit rate 1 (bit / MHz) delay_QoS_exponent Latency QoS Index 1 population_size Total number of PSO particles 1000 iteration_times Maximum number of PSO iterations 100 inertia_weight PSO Inertia Factor 0.9 learning_factor1 PSO cognitive factors 0.5 learning_factor2 PSO social factors 0.5 sample_size Group sample size 100 groups_train Number of training set samples 500 groups_test Number of test set samples 100 learning_rate Network learning rate 1e-3 number of layers for network1 Network Layer 1 0 Number of layers for network2 Network Layer 2 3 Number of nodes (network2) Network 2-node count (64, 32, 16) decay_rate Attenuation rate 0.95 optimizer Optimizer Adam activation function Activation function Relu This invention employs a decoupled intelligent computing architecture, comprising a server for offline data processing, an online-deployed base station, and a power allocation controller with deep learning inference capabilities. In the offline phase, the server generates a globally optimal power allocation dataset with the objective of maximizing the total effective capacity of two users through large-scale channel simulation using an improved particle swarm optimization (PSO) algorithm. In the online phase, a trained lightweight cascaded deep learning model is deployed at the base station, receiving channel state information (CSI) from user terminals in real time and outputting the optimal power allocation coefficients with a millisecond delay before forward propagation. The base station then performs NOMA overlay coding and signal transmission accordingly. This system avoids the excessive computational complexity and inability to meet real-time requirements of traditional online optimization algorithms, while enabling the base station to consistently execute a power allocation strategy approaching the global optimum in dynamically changing channel environments, thereby ensuring maximum effective system capacity. This not only significantly improves spectral efficiency and Quality of Service (QoS) but also resolves the core contradiction between the poor real-time performance of traditional model-driven methods and the low performance ceiling of purely data-driven methods. Meanwhile, the inter-regional PSO data generation and deep cascaded network collaborative design proposed in this invention combine the theoretical guarantee of the model with the efficient learning capability of the data, providing an efficient, reliable and easy-to-deploy system-level solution for intelligent resource management of 5G-Advanced and future mobile communication networks.
[0029] The preferred embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the scope of protection of the present invention is not limited to the specific details in the above embodiments. Within the scope of the technical concept of the present invention, any person skilled in the art can make equivalent substitutions or changes based on the technical solution and inventive concept of the present invention within the scope of the technology disclosed in the present invention. These simple modifications are all within the scope of protection of the present invention.
Claims
1. A power allocation method for a downlink NOMA system based on deep learning, characterized in that, Includes the following: Step 1, obtain The data, specifically: Step 1.1: Generate a channel gain sample pair for the user at time slot t. , ),in Let be the channel gain of user 1 in time slot t. Let t be the channel gain of user 2; denote the channel gain sample pairs corresponding to M time slots as channel gain samples, and extract the original channel components of the channel gain samples: the complex channel gain from the base station to user 1 in each time slot. and the multiple channel gain from base station to user 2 Where t = 1, 2, 3...M; Step 1.2, the channel gain samples are processed as follows to obtain the channel gain samples. data: The sum of squared complex channel gains of all channel gain samples is concatenated to form the input feature vector of the channel gain samples. : , The first M dimensions represent the sum of squared channel gains sequence for user 1, and the last M dimensions represent the sum of squared channel gains sequence for user 2. Input feature vector of channel gain sample Perform Z-score standardization: ， in , These are the statistical feature vectors of channel state information for the channel gain samples. The mean and standard deviation; Input feature vector Z-score standardized results; Step 2: Construct a cascaded deep neural network model, which is formed by sequentially connecting the trained first-level neural network model and the second-level neural network model. The structure of the first-level neural network model is as follows: The input layer consists of 200 neurons; A fully connected layer; The output layer consists of two neurons; The second-level neural network model has the following structure: The input layer consists of two neurons; Three hidden layers, each using the ReLU activation function and applying Dropout regularization, with a dropout rate of 0.2; The output layer consists of one neuron and uses linear activation; Step 3, sample the channel gain. Data is input into a trained cascaded deep neural network model; the model outputs the optimal power allocation coefficients predicted from the channel gain samples. ; Step four, the base station according to Calculate the actual power allocated to user 1 within the M time slots. The actual power allocated to user 2 .
2. The power allocation method for a downlink NOMA system based on deep learning according to claim 1, characterized in that, In step 1.1, the 4-dimensional original components of the channel gain sample are... , , , Reconstruct the complex channel gain summation for each time slot in the channel gain sample: ;in , These are the complex channel gains from the base station to user k in the corresponding time slot. The real and imaginary parts of .
3. The power allocation method for a downlink NOMA system based on deep learning according to claim 1, characterized in that, The training process of the cascaded deep neural network model is as follows: Step 1, data preparation, specifically includes the following: S1.1: Define the downlink NOMA system model Received signal model: The received signal of user k in time slot t. Represented as: ， in: This is the encoded signal sent by the base station to user k, where k = 1 or 2, representing the kth user; Let be the complex channel gain from the base station to user k, which follows independent Rayleigh fading; Let M be the transmit power that the base station allocates to user k within M time slots. Let M be the total transmit power allocated by the base station to all users over M time slots, which satisfies the following constraints: ; It is additive white Gaussian noise; Channel State Information: Definition According to the Rayleigh fading model, Follow the mean The exponential distribution of has a probability density function as: ， Assuming the user's receiver noise power is normalized, and strong and weak users are distinguished by the strength of the channel gain between the base station and the user, then the instantaneous service rate of user k is... for: Strong user refers to the instantaneous service rate of user 1. for: ， The instantaneous service rate of weak users, i.e., user 2. for: ， in: Let be the channel gain of user 1 in time slot t. Let be the channel gain for user 2 in time slot t; S1.2: Define statistical QoS metrics and optimization objectives based on effective capacity. Effective capacity formula: For user k with independent, co-distributed block fading channels and a specific transmission mode, its effective capacity is... for: ， in , where is the latency QoS index for user k; The probability of successfully transmitting information to user k; where: Probability of User 1 successfully transmitting information : ， Probability of User 2 successfully transmitting information : ， Where the threshold The joint channel requirements that User 1 must meet when using SIC decoding; The preset transmission rate for user 1, The preset transmission rate for user 2; It is the probability density function; System optimization goals: Will and Substituting the expressions into the effective capacity respectively The calculation formula yields the total effective capacity of the system. : ， S1.3: Implement interval-based particle swarm optimization to generate optimal labels. S1.31: Optimize the interval [0, The interval is divided into four sub-intervals, where: ， ， ， S1.32: Channel Sample Batch Generation and Feature Extraction: Generate a channel gain sample pair for the user at time slot t. , Each M time slot corresponds to a channel gain sample pair, which is denoted as a group of channel gain samples. The mean squared channel gain of this group of channel gain samples is calculated. and As the statistical feature vector of the channel state information of this group of samples ; S1.34: Perform the following operations to find the optimal solution. : S1.341: First interval Optimization: Initialize a PSO particle swarm within this interval, so as to As the fitness function, PSO iteration is performed to record the optimal solution found within the interval. and ,in It is the optimal power allocation value within this interval, i.e., the optimal value within this interval. ,and- It is the maximum total effective system capacity value in this interval; Second interval Optimization: Initialize a PSO particle swarm within this interval, so as to Using the fitness function, perform PSO iteration to obtain the optimal solution. and ,in It is the optimal power allocation value for this interval, i.e., the optimal value for this interval. ,and- It is the maximum total effective system capacity value in this interval; Third interval Optimization: Initialize a PSO particle swarm within this interval, so as to Using the fitness function, perform PSO iteration to obtain the optimal solution. and ,in It is the optimal power allocation value for this interval, i.e., the optimal value for this interval. ,and- It is the maximum total effective system capacity value in this interval; S1.342: Perform a physical feasibility check: like ,think invalid; like ,think invalid; Remove invalid data and the maximum total effective system capacity corresponding to the invalid data from the three sub-intervals. Select the maximum total effective system capacity corresponding to the largest valid data. The value is the globally optimal value under the given set of channel gain sample states. ,Right now ; S1.343: Calculate the optimal power allocation coefficients for this set of channel gain samples: ; S1.4: Constructing the final dataset The corresponding channel gain samples Statistical feature vector of channel state information and optimal power allocation coefficient Pairing together to form a data sample ; obtain Groups of data samples form a large-scale dataset D; S1.5: Tag Data Preparation: Read the statistical feature vector of channel state information corresponding to each group of channel gain samples from dataset D. and optimal power allocation coefficient , take each set of data and Together they serve as tag data; S1.6: Calculate the corresponding channel gain samples for each group according to step 1.
2. data; S1.7: Training process: Training the first-level neural network model: Will The z-data corresponding to each group of channel gain samples is used as the input information of the model. The statistical feature vector of channel state information in the label data corresponding to each group of channel gain samples is used. The output data of the model are input together into the first-level neural network model for training until the loss function of the first-level neural network model converges, and the training ends. The first-level neural network model uses the Adam optimizer for parameter updates, where the initial learning rate is... The learning rate uses exponential decay, decreasing to 0.95 times its original value every 50 training steps. The first-level neural network model uses the mean absolute error as the loss function: ， in This represents the loss value of the first-level neural network model. The number of groups of channel gain samples used for training, For the first The target set value of the statistical feature vector of the base station for the channel state information of user 1 corresponding to the group of channel gain samples. For the first The target set value of the statistical feature vector of the base station for the channel state information of user 2 corresponding to the group of channel gain samples; Training the second-level neural network model: Will Group channel gain samples corresponding As input information to the model, the optimal power allocation coefficients from the tag data corresponding to each group of channel gain samples are used. The output data of the model are fed into the second-level neural network model for training until the loss function of the second-level neural network model converges, at which point the training ends. The second-level neural network model uses the Adam optimizer for parameter updates, where the initial learning rate is... The learning rate is a fixed learning rate; The second-level neural network model uses standard mean squared error as the loss function: ， in This represents the loss value of the second-level neural network model. The number of groups of channel gain samples used for training. The target setpoint for the optimal power allocation coefficient; The output of the second-level neural network model Predicted power allocation coefficients for group channel gain samples; Training cascaded networks: The original training material for the first-level neural network model Optimal power allocation coefficients in the data and labeled data used to train the second-level neural network model The input and output information of the cascaded network are respectively used as input and output information to train the cascaded network until the loss function of the cascaded network converges, and the training ends. The cascaded network uses the Adam optimizer for parameter updates, where the initial learning rate... The learning rate uses exponential decay, decreasing to 0.95 times its original value every 30 training steps. The cascaded network employs a hybrid loss function. : ， in: , where is the mean square error; For Huber's losses; This is an L2 regularization term.