A satellite-to-ground link power control allocation method based on local recurrent neural network

Abstract

The present invention provides a satellite-to-ground link power control allocation method based on a local recurrent neural network, and the present invention relates to the field of satellite-to-ground link power control technology. The present invention establishes a data set for training a neural network based on quantized state data of satellites and users, and pre-processes the data set; selects characteristic parameters for each link, and determines the optimal transmission power of each link under different states; establishes a neural network, and selects parameters, and trains the neural network according to the selected parameters; uses an enumeration method to determine the output judgment value of the neural network, and determines the maximum power transmission or the minimum power transmission according to the data judgment value. The present invention proposes an algorithm that allows users to access the satellite network based on the shortest distance as the access principle, and performs power allocation through an Elman neural network with feedback after accessing the satellite network, thereby achieving a larger network maximum channel capacity with a lower algorithm complexity.

Description

Technical Field

[0001] The present invention relates to the technical field of satellite-to-ground link power control, and in particular to a satellite-to-ground link power control allocation method based on a local recurrent neural network. Background Art

[0002] With the rapid development of global informatization and networking, people are constantly pursuing all-weather, seamless, and high-quality communication methods. With the vigorous development of low-orbit satellite Internet, the wide-area coverage advantage of satellite communication can well make up for the limited coverage of ground cellular networks. Therefore, studying how ground users access low-orbit satellite Internet and how to allocate resources after access is an important part of the realization of the integrated space-ground network in the 5G era. Deep learning has excellent learning and data recognition capabilities. As a part of machine learning, it has shined in solving various resource scheduling optimization problems. From another perspective, the resource allocation problem between satellites and users can be regarded as a problem of finding a mapping function - each satellite obtains the distribution result in each allocation interval through a mapping function according to certain parameters, and neural networks are very suitable for finding mapping problems that are difficult to solve mathematically. Some scholars have used deep learning to solve the problem of power control and allocation of sensor transmitters in a space where a large number of sensors are deployed.

[0003] The research on improving the maximum channel capacity of low-orbit satellite networks can be divided into two aspects: on the one hand, the criteria for users to access satellite networks, that is, the parameters by which users determine the satellites they access, such as the shortest distance priority access algorithm, the longest coverage time priority access algorithm, the elevation angle weighted coverage time priority access algorithm, etc.; on the other hand, the algorithm for allocating the satellite transmission power after access is studied, such as heuristic algorithms and some mathematical methods that transform non-convex optimization problems into convex optimization problems. However, the above solutions all have certain shortcomings, and the main common defect is the problem of algorithm complexity. Considering too many parameters to select the satellite to access and overly complex allocation algorithms will bring delay problems, while relatively low-complexity methods will make the results after allocation less than excellent. Summary of the invention

[0004] The present invention realizes a larger network maximum channel capacity with a lower algorithm complexity by proposing an algorithm that allows users to access the satellite network based on the shortest distance as the access principle and performs power allocation through an Elman neural network with feedback after accessing the satellite network.

[0005] The present invention provides a satellite-to-ground link power control allocation method based on a local recurrent neural network, and the specific scheme is as follows:

[0006] A satellite-to-ground link power control allocation method based on a local recurrent neural network comprises the following steps:

[0007] Step 1: Based on the quantified state data of satellites and users, a data set for training the neural network is established and the data set is preprocessed;

[0008] Step 2: Select characteristic parameters for each link to determine the optimal transmit power of each link under different conditions;

[0009] Step 3: Establish a neural network, select parameters, and train the neural network according to the selected parameters;

[0010] The neural network structure is Elman neural network. The network has five layers, namely input layer, first hidden layer, second hidden layer, output layer, first receiving layer and second receiving layer. The input layer has eight nodes, corresponding to eight eigenvalues; both hidden layers have 50 nodes, and the output layer has only one node.

[0011] Step 4: Use the enumeration method to determine the output judgment value of the neural network, and determine the maximum power transmission or the minimum power transmission according to the data judgment value.

[0012] Preferably, the step 1 is specifically:

[0013] Based on the quantified state data of satellites and users, a data set for training the neural network is established. The data set is divided into three parts: training set, validation set and test set. The training set is used to train and update the network parameters; the validation set is used to test the generalization of the network and determine whether overfitting occurs; and the test set reflects the final performance of the neural network.

[0014] Determine the power allocation occasions; the access scenario is: each low-orbit satellite has multiple wide beams and the communication frequency is the same, then the coverage range of a certain beam of a certain low-orbit satellite is also the coverage range of the beams of other low-orbit satellites. There are a certain number of user terminals within this beam range; within each access interval, the user terminal uses the ephemeris broadcast by each satellite to calculate its distance from each satellite, and the terminal selects the nearest satellite for access;

[0015] In each access interval, all user terminals are connected to the satellite network. When only the downlink is considered, the transmission power of each satellite is reasonably allocated to maximize the total channel capacity of the entire satellite network. Each satellite has a minimum transmission power. For the channel condition between the satellite and the user terminal, a shadow Rice channel is used.

[0016] Preferably, the step 2 is specifically:

[0017] Eight characteristic parameters considered when performing power allocation, for each link, the convolution result of the horizontal coordinate of the satellite that affects it, for each link, the convolution result of the vertical coordinate of the satellite that affects it, for each link, the convolution result of the horizontal coordinate of the average position of the users that affect it, for each link, the convolution result of the vertical coordinate of the average position of the users that affect it, the channel fading of the current link, the minimum channel fading of all links in the current access interval, and the maximum channel fading of all links in the current access interval;

[0018] After obtaining the characteristic parameters, the optimal transmission power of each link under different conditions is determined, and the maximum network channel capacity is obtained after power allocation using the genetic algorithm.

[0019] Preferably, a roulette wheel selection algorithm is used for the selection operation, and individuals with good traits are selected to be passed on to the next generation according to their fitness in a certain way, and the proportion of each individual's fitness to the sum of the fitness of all individuals in this generation of the population is used as the probability of each individual being selected, that is, the probability of being selected is proportional to the fitness of each individual; the crossover operation is to exchange the components of the feasible solution between a pair of selected individuals with a certain probability, and the genetic algorithm used obtains the data set required for supervised training of the network.

[0020] Preferably, the hidden layer and output node of the neural network both require an activation function. The activation functions of the hidden layer and the receiving layer use the ReLU function, and the activation function of the output layer uses the Sigmoid function. When the input is large, the output tends to 1, and when the input is small, the output tends to 0.

[0021] Preferably, the neural network adopts a learning function of the gradient descent momentum function learngdm. On the basis of back propagation, a momentum is added when updating the weights. α is the step size of each weight change, and m and m - is the current momentum and the momentum of the previous moment, β is a constant less than 1. As the iteration proceeds, the learning function with momentum m is used, and the updated value Δω of the weight in the previous iteration will be added to each momentum.

[0022] Preferably, the training function used in the training in step 3 is the trainlm function, which refers to the training function using the LM algorithm, and the key parameter is the coefficient μ. When it is 0, the algorithm is the Newton method; when the coefficient μ is very large, the algorithm is the gradient descent method with a smaller step size; when the training just starts, a smaller coefficient μ is selected, and the Newton descent method will be used to update the weights. The Newton method has a second-order convergence speed and a faster training speed; as the training proceeds, the μ value will gradually increase, and the training will become the gradient descent method, the training speed will gradually slow down, the training results will gradually converge, and finally stabilize, and the training is completed; for medium-sized neural networks, the LM algorithm can obtain a very small mean square error.

[0023] Preferably, the loss function used is the mean square error. In each iteration, the square sum of the difference between the output value of the neural network and the target value is averaged. The output of the training function is the trained network and training records. During the training process, the training function continuously calls the learning function to correct the weights. The training is terminated by detecting the set number of training steps or the error calculated by the loss function is less than the set error.

[0024] Preferably, the stopping training condition is set as when the gradient is less than 1e-7, or the coefficient is less than 0.001 or greater than 1e10.

[0025] Preferably, the output result of the neural network is a result that fits as much as possible after the output of the activation function. The output judgment value of the neural network is found to be 0.505 by enumeration method. When it exceeds the judgment value, the network output is judged as "1", that is, maximum power transmission. When it is lower than the judgment value, it is judged as "0.1", that is, minimum power transmission.

[0026] Beneficial effects:

[0027] The present invention mainly proposes a method for increasing the maximum channel capacity of a satellite network formed by a ground mobile terminal and a low-orbit satellite. The method requires the ground mobile terminal to select an accessible low-orbit satellite based on the principle of proximity, and to control the satellite's transmission power through a neural network to increase the maximum channel capacity of the network.

[0028] The complexity comparison results of the present invention and the maximum signal-to-noise ratio access algorithm, the neural network optimization algorithm after the nearest access, and the greedy algorithm optimization algorithm after the nearest access. As the number of users increases, the operation time of the maximum signal-to-noise ratio access increases exponentially. When the number of users is large, the operation time is greater than the greedy algorithm and the neural network, which is consistent with the above theoretical analysis. The curve of the greedy algorithm is difficult to fit with a mathematical curve, but as the number of users increases, the time it takes also increases, which is consistent with the above theoretical analysis. Among the three algorithms, the simulation time used by the neural network is the shortest. It can be seen that the present invention has a lower complexity than other algorithms. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 It is a system model diagram of the present invention;

[0030] Figure 2 The comparison results of the maximum network channel capacity obtained after power allocation using the greedy algorithm, simulated annealing algorithm, simulated particle swarm algorithm, and genetic algorithm;

[0031] Figure 3 is the genetic algorithm process adopted;

[0032] Figure 4 It is a network structure diagram;

[0033] Figure 5 is the error curve during the neural network training process;

[0034] Figure 6 are the gradient change curve, coefficient μ change curve and verification times curve;

[0035] Figure 7 The comparison results of the present invention and other three algorithms obtained by simulating fifty times under different satellite conditions in the same simulation environment;

[0036] Figure 8 This is a comparison result of the complexity of the present invention, the maximum signal-to-noise ratio access algorithm, the nearest access followed by a neural network optimization algorithm, and the nearest access followed by a greedy algorithm optimization algorithm. DETAILED DESCRIPTION

[0037] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0038] Combination Figures 1 to 8 As shown, the present invention provides a satellite-to-ground link power control allocation method based on a local recurrent neural network, comprising the following steps:

[0039] A satellite-to-ground link power control allocation method based on a local recurrent neural network is characterized by comprising the following steps:

[0040] Step 1: Based on the quantified state data of satellites and users, a data set for training the neural network is established and the data set is preprocessed;

[0041] The step 1 is specifically as follows:

[0042] Based on the quantified state data of satellites and users, a data set for training the neural network is established. The data set is divided into three parts: training set, validation set and test set. The training set is used to train and update the network parameters; the validation set is used to test the generalization of the network and determine whether overfitting occurs; and the test set reflects the final performance of the neural network.

[0043] Determine the power allocation occasions; the access scenario is: each low-orbit satellite has multiple wide beams and the communication frequency is the same, then the coverage range of a certain beam of a certain low-orbit satellite is also the coverage range of the beams of other low-orbit satellites. There are a certain number of user terminals within this beam range; within each access interval, the user terminal uses the ephemeris broadcast by each satellite to calculate its distance from each satellite, and the terminal selects the nearest satellite for access;

[0044] In each access interval, all user terminals are connected to the satellite network. When only the downlink is considered, the transmission power of each satellite is reasonably allocated to maximize the total channel capacity of the entire satellite network. Each satellite has a minimum transmission power. For the channel condition between the satellite and the user terminal, a shadow Rice channel is used.

[0045] Step 2: Select characteristic parameters for each link to determine the optimal transmit power of each link under different conditions;

[0046] The step 2 is specifically as follows:

[0047] Eight characteristic parameters considered when performing power allocation, for each link, the convolution result of the horizontal coordinate of the satellite that affects it, for each link, the convolution result of the vertical coordinate of the satellite that affects it, for each link, the convolution result of the horizontal coordinate of the average position of the users that affect it, for each link, the convolution result of the vertical coordinate of the average position of the users that affect it, the channel fading of the current link, the minimum channel fading of all links in the current access interval, and the maximum channel fading of all links in the current access interval;

[0048] After obtaining the characteristic parameters, the optimal transmission power of each link under different conditions is determined, and the maximum network channel capacity is obtained after power allocation using the genetic algorithm.

[0049] Step 3: Establish a neural network, select parameters, and train the neural network according to the selected parameters;

[0050] The neural network structure is Elman neural network. The network has five layers, namely input layer, first hidden layer, second hidden layer, output layer, first receiving layer and second receiving layer. The input layer has eight nodes, corresponding to eight eigenvalues; both hidden layers have 50 nodes, and the output layer has only one node.

[0051] Step 4: Use the enumeration method to determine the output judgment value of the neural network, and determine the maximum power transmission or the minimum power transmission according to the data judgment value.

[0052] The roulette wheel selection algorithm is used for selection operation. According to the individual fitness, individuals with good traits are selected to the next generation in a certain way. The proportion of each individual's fitness to the sum of the fitness of all individuals in this generation is taken as the probability of each individual being selected, that is, the probability of being selected is proportional to the fitness of each individual. The crossover operation is to exchange the components of the feasible solution between a pair of selected individuals with a certain probability. The genetic algorithm used obtains the data set required for supervised training of the network.

[0053] The hidden layer and output node of the neural network both require an activation function. The activation function of the hidden layer and the receiving layer adopts the ReLU function, and the activation function of the output layer adopts the Sigmoid function. When the input is large, the output tends to 1, and when the input is small, the output tends to 0.

[0054] The neural network adopts the learning function of the gradient descent momentum function learngdm. On the basis of back propagation, a momentum is added when updating the weights. α is the step size of each weight change, and m and m - is the current momentum and the momentum of the previous moment, β is a constant less than 1. As the iteration proceeds, the learning function with momentum m is used, and the updated value Δω of the weight in the previous iteration will be added to each momentum.

[0055] The training function used in the step 3 is the trainlm function, which refers to the training function using the LM algorithm. The key parameter is the coefficient μ. When it is 0, the algorithm is the Newton method; when the coefficient μ is very large, the algorithm is the gradient descent method with a smaller step size; when the training just starts, a smaller coefficient μ is selected, and the Newton descent method will be used to update the weights. The Newton method has a second-order convergence speed and a faster training speed; as the training progresses, the μ value will gradually increase, and the training will become the gradient descent method, the training speed will gradually slow down, the training results will gradually converge, and finally stabilize, and the training is completed; for medium-sized neural networks, the LM algorithm can obtain a very small mean square error.

[0056] The loss function used is the mean square error. In each iteration, the square of the difference between the output value of the neural network and the target value is averaged. The output of the training function is the trained network and training records. During the training process, the training function continuously calls the learning function to correct the weights. The training ends by detecting the set number of training steps or the error calculated by the loss function is less than the set error.

[0057] The error curve during neural network training is as follows: Figure 5 As shown. The training set accounts for 70% of the data set, and its function is to train the neural network to minimize the error. The green part is the validation set, which accounts for 15% of the data set. Its function is to determine whether the trained neural network is suitable for this part of the input, that is, to check whether the neural network has overfitting or underfitting problems. If there are problems, retraining or further changing the feature values ​​are required. The red part is the test set, and its function is to test the performance of the neural network on the new set after verification and correction of the validation set, to check whether the above-mentioned overfitting or underfitting problems still exist. From Figure 5 It can be seen that the errors in the three sets are small, and there is no problem of overfitting or underfitting, so it is considered that the neural network training is successful.

[0058] The gradient change curve, coefficient μ change curve and verification times curve are as follows: Figure 6 As shown. It can be seen that the gradient is getting smaller and smaller, because as the iteration proceeds, the error becomes smaller and smaller, and the error function is closer to the minimum point, so its gradient is getting smaller and smaller and finally tends to 0; the coefficient μ curve first decreases, then the LM algorithm tends to the Newton method, and the training speed is faster, and then the curve rises, and the training tends to the gradient descent method, and the training becomes slower, but it is easier to converge; the verification curve shows that when the training is carried out for the fifth time, the condition for stopping training has been met. The condition for stopping training set by the present invention is when the gradient is less than 1e-7, or the coefficient is less than 0.001 or greater than 1e10.

[0059] The output result of the neural network is the result that is fitted as much as possible after the output of the activation function. The output judgment value of the neural network is found to be 0.505 through enumeration method. When it exceeds the judgment value, the network output is judged as "1", which is the maximum power transmission. When it is lower than the judgment value, it is judged as "0.1", which is the minimum power transmission.

[0060] With reference to the characteristic value selection method, network structure and parameter settings, after selecting the optimal value, the performance of the present invention is simulated and verified. The simulation includes two parts. The first part is the comparison of power allocation results. Under the same simulation conditions, after training a suitable neural network, the maximum channel capacity of the network after power allocation through the neural network, as well as through the maximum signal-to-noise ratio access algorithm, the greedy algorithm and the random allocation are compared; the second part is the comparison of complexity. The present invention compares the complexity of different algorithms according to the time they run in the same simulation environment. The longer the time it takes to run an algorithm, the higher its complexity.

[0061] Figure 7 The comparison results of the present invention and the other three algorithms obtained by simulating fifty times under different satellite conditions in the same simulation environment are shown. Taking the channel capacity after genetic algorithm optimization as the standard, the ratio of the channel capacity after the four algorithms are optimized to the channel capacity after genetic algorithm optimization is compared in each simulation. The larger the ratio, the better the result. The results output by the Elman neural network are better than those of the greedy algorithm and the maximum signal-to-noise ratio access algorithm. At the same time, it can be seen that the ratio of the network capacity after the Elman neural network optimization to the network capacity solved by the optimal genetic algorithm is more than 85%. It can be intuitively seen that the present invention has good power allocation capabilities and the stability of its allocation results.

[0062] Figure 8 This is the result of the comparison of the complexity of the present invention with the maximum signal-to-noise ratio access algorithm, the neural network optimization algorithm after the nearest access, and the greedy algorithm optimization algorithm after the nearest access. As the number of users increases, the operation time of the maximum signal-to-noise ratio access increases exponentially. When the number of users is large, the operation time is greater than that of the greedy algorithm and the neural network, which is consistent with the above theoretical analysis. The curve of the greedy algorithm is difficult to fit with a mathematical curve, but as the number of users increases, the time it takes also increases, which is consistent with the above theoretical analysis. Among the three algorithms, the simulation time used by the neural network is the shortest. It can be seen that the present invention has a lower complexity than other algorithms.

[0063] Finally, the complexity of the present invention is analyzed. The present invention mainly compares the complexity of three algorithms, namely the maximum signal-to-noise ratio access algorithm, the neural network optimization algorithm after the nearest access, and the greedy algorithm optimization algorithm after the nearest access. The reason for choosing to compare the neural network with these two algorithms is that they are the two most commonly used algorithms in current engineering practice. The most intuitive way to evaluate the complexity of different algorithms is to compare the number of operation symbols used by different algorithms. If the complexity of multiplication and division is considered to be the same, and the complexity of addition and subtraction is considered to be the same, then the most intuitive comparison method is to compare the number of multiplication and addition operations performed on the two. For the maximum signal-to-noise ratio access algorithm, the signal-to-noise ratio of each link needs to be calculated. The number of multiplications required for each calculation of the signal-to-noise ratio of a link is 3N times. Since there are N links in total, when the maximum signal-to-noise ratio access algorithm is used, the number of multiplications required for each user to calculate its signal-to-noise ratio to each satellite to be accessed is 3N. 2 times, thus selecting the link with the largest signal-to-noise ratio. Each user needs to go through this process, so the total number of multiplications required is 3N 3 Next, we need to calculate the channel capacity of each link and add them up. This step requires N additions, N multiplications, and N exponential operations. Therefore, the algorithm requires a total of (3N 3 +N) times of multiplication and N times of addition, and N times of exponential operation. For the nearest access method optimized by the greedy algorithm, in each iteration, the current signal-to-noise ratio of each link needs to be calculated, which requires 3N times of multiplication, and then the channel capacity of the entire network is calculated. Compared with the optimal result obtained in the previous iteration, the previous optimal result still requires 3N times of multiplication. This step requires N times of addition, N times of multiplication and N times of exponential operation. Assuming the number of iterations of the greedy algorithm is P, the algorithm requires a total of 7PN times of multiplication, PN times of addition and PN times of exponential operation. For the nearest access algorithm optimized by the neural network, the main amount of calculation is to calculate the weight of each link. Since the input layer of the Elman neural network has 8 nodes, the two hidden layers and the receiving layer each have 50 nodes and an output node, it takes 7950 times of multiplication and 7950 times of addition to calculate the weight of each link. Therefore, the algorithm needs to perform a total of 7951N times of multiplication, 7951N times of addition and N times of exponential operation. The main comparison is the number of multiplications required by the three algorithms. The number of multiplications required for the maximum signal-to-noise ratio access algorithm is (3N 3+N) times, the number of multiplications required for the greedy algorithm optimization after the nearest access is 7PN times, and the number of multiplications required for the neural network optimization algorithm is 7951PN times. It is worth noting that the P in the complexity of the greedy algorithm does not refer to the number of iterations in the fitness evolution curve, because the greedy algorithm only retains the solution with a greater fitness than this time in the next iteration, and those iterations that are not retained also take up a lot of time. According to multiple calculation experiences, when the set stop iteration number G is 1000, P is generally 8 to 10 times G. Therefore, in theory, the neural network method has the lowest complexity, and the maximum signal-to-noise ratio access has the highest complexity. Specific embodiment 2:

[0065] The present invention combines the traditional heuristic algorithm with a supervised trained neural network and proposes a new power allocation scheme, which mainly includes the generation and processing of eigenvalues ​​for neural network training and the selection of network structure and parameters. Figure 1 A system model of the present invention is presented.

[0066] The system model of the present invention can be divided into a training mode and a judgment mode. In the training mode, the system input is the quantized state data of the satellite and the user at this time. On the one hand, the data obtains the desired allocation result through a suitable allocation algorithm, and on the other hand, the processed feature parameters are obtained through a feature parameter extraction module. The output result of the network is obtained through an initialized neural network, and the error between the two is used as the basis for updating the neural network parameters. Through supervised training, the network parameters are updated with the minimum mean square error algorithm; after the network training is completed, it is switched to the judgment mode, the parameters of the neural network no longer change, and after the system inputs the processed feature parameters, the allocation result at this time is judged according to the output result of the neural network.

[0067] The present invention is used to generate and process a data set for training a neural network. In the present invention, the data set can be divided into three parts: a training set, a validation set, and a test set. The role of the training set is to use it to train and update network parameters; the role of the validation set is to test the generalization of the network and determine whether overfitting occurs; the test set is used to illustrate the final performance of the neural network. This part of the invention content can be divided into two parts: the selection of feature parameters and the calculation of expected results.

[0068] The first is how to select characteristic parameters. Before generating characteristic parameters, it is necessary to clarify the power allocation scenarios applicable to the present invention. The access scenario applicable to the present invention is: assuming that each low-orbit satellite has multiple wide beams and the communication frequency is the same, then a certain beam coverage range of a low-orbit satellite is also the beam coverage range of other low-orbit satellites. At this time, there are a certain number of user terminals within this beam range. Within each access interval, these user terminals can calculate their distance from each satellite using the ephemeris broadcast by each satellite, so that these terminals choose the satellite closest to them for access. User terminals access the satellite network in each access interval. Since this is a full-frequency multiplexing scenario, for a certain user, the communication signals between other users and their respective satellites will be regarded as noise. Therefore, when only considering the downlink, the transmission power of each satellite should be reasonably allocated to maximize the total channel capacity of the entire satellite network. At the same time, considering the principle of fairness, each satellite has a minimum transmission power. For the channel conditions between the satellite and the user terminal, the present invention adopts a shadow Rice channel.

[0069] After clarifying the application scenario, the next step is to select the feature parameters. Since the satellite channel attenuation is mainly related to the path transmission loss and shadow fading, and the path transmission loss is only related to the signal transmission distance, and the signal transmission distance is related to the coordinates of the satellite and the user. At the same time, the shadow fading of the channel is mainly related to the elevation angle between the user and the satellite, and the elevation angle is also related to the coordinates between the satellite and the user. Therefore, it can be considered that the final output of the neural network is related to the position coordinates of the user and the satellite. Based on this, the present invention proposes eight characteristic parameters to be considered when performing power allocation: for each link, the convolution result of the horizontal coordinate of the satellite that affects it, for each link, the convolution result of the vertical coordinate of the satellite that affects it, for each link, the convolution result of the horizontal coordinate of the average position of the users that affect it, for each link, the convolution result of the vertical coordinate of the users that affect it, the channel fading of the current link, the minimum channel fading among all links in the current access interval, and the maximum channel fading among all links in the current access interval.

[0070] After obtaining the characteristic parameters, the next step is to obtain the optimal transmission power of each link under different conditions. The present invention compares the maximum network channel capacity obtained after power allocation using the greedy algorithm, simulated annealing algorithm, simulated particle swarm algorithm, and genetic algorithm. The comparison results are as follows: Figure 2 As shown. Figure 2It can be seen that the greedy algorithm is the fastest algorithm, but it does not have the ability to obtain the global optimal solution. The simulated annealing algorithm requires a large number of iterations and a long operation time, but the final result is relatively excellent. The particle swarm algorithm has a strong global search capability at the beginning, but it will easily converge as the number of iterations increases, so it is easy to fall into a local optimal solution, but its convergence speed is very fast. Each generation of the genetic algorithm performs large-scale crossover and mutation operations, so although it can break the local optimal, it is easy to not converge. Figure 2 The fitness curve comparison of the four heuristic algorithms shows that the genetic algorithm has the best final result among the four algorithms and is the fastest to converge. The final result of the greedy algorithm is the worst because it does not have the ability of global search, but when the number of iterations increases, its final result also tends to converge. Although the particle swarm algorithm can converge quickly, the convergence point is not the global optimal solution, and as the number of iterations increases, it eventually converges to a local optimal point. The final output result of the simulated annealing algorithm is close to that of the genetic algorithm, but the genetic algorithm converges faster. Therefore, the genetic algorithm has the best optimization result and faster optimization speed, and is most suitable for the present invention. However, among the four algorithms, the greedy algorithm has the fastest solution speed, so the greedy algorithm is often used in various delay-sensitive solution occasions. Therefore, although the optimization result of the greedy algorithm is not excellent, it will be used as a comparison object for the next neural network optimization result. If the result of the neural network optimization is better than the greedy algorithm, its performance is considered to be up to standard.

[0071] For genetic algorithms, the principle is as follows: At the beginning, the first generation of population can be obtained. After applying a series of genetic operations such as selection, crossover and mutation to this generation of population, a new generation of population can be generated. According to a certain selection operator, the excellent individuals in this generation of population are selected and mated with other individuals. After the genes of the excellent individuals are selected, crossover and mutation are performed, a new generation of population will be generated and the excellent genes will be passed on. This process is repeated, and the best individuals in the last generation of population are considered to be the required individuals, that is, the optimal solution.

[0072] The present invention adopts the "roulette wheel" selection algorithm for selection operation. The so-called selection operation refers to selecting individuals with good traits to be passed on to the next generation according to the fitness of the individuals in a certain way. The ratio of the fitness of each individual to the sum of the fitness of all individuals in this generation of the population is taken as the probability of each individual being selected, that is, the probability of being selected is proportional to the fitness of each individual. The crossover operation is to exchange the components of the feasible solution between a pair of selected individuals with a certain probability, in order to improve the search ability of the genetic algorithm. The mutation operation refers to that for all individuals in the population, in each generation, according to the set mutation probability, the random gene mutation of each individual is an arbitrary random value within the domain of feasible solutions. The genetic algorithm process used in the present invention is as follows Figure 3 As shown. This can obtain the data set required for supervised training of the network.

[0073] The second part of the present invention is the selection and design of network structure and parameters. The neural network structure adopted by the present invention is Elman neural network, which is a typical local regression neural network. The network constructed by the present invention has five layers, namely input layer, hidden layer 1, hidden layer 2, output layer, receiving layer 1 and receiving layer 2. The input layer has eight nodes, corresponding to the eight eigenvalues ​​obtained in the first part. Both hidden layers have 50 nodes. The output layer has only one node, because in the scenario applied by the present invention, the power allocation result after optimization has only two possibilities of maximum power output or minimum power output, so only one output node is needed, and the output is "1" for maximum power output, and "0" for minimum power output. The number of nodes of the two receiving layers is also 50. The first receiving layer stores the output of the last round of hidden layer 1, which will be used as part of the input of this round of hidden layer 1. Similarly, the second receiving layer stores the output value of hidden layer 2 at the last moment, as part of the input of hidden layer 2 at this moment. The neural network uses the output of each hidden layer in this iteration as the input of the layer in the next iteration, realizing a local regression. Such a neural network with feedback is called a dynamic network. Compared with a static neural network without feedback, it can better learn a complex system and achieve better training results. Figure 4 shown.

[0074] Both the hidden layer and the output node require an activation function. The so-called activation function means that the value input to this node must be obtained after the activation function before it can be used as its output. The activation function of the hidden layer and the receiving layer uses the ReLU function, and the activation function of the output layer uses the Sigmoid function. The characteristic of this function is that when the input is large, the output tends to 1, and when the input is small, the output tends to 0. It is suitable for neural networks with only two output results.

[0075] To train a neural network, three functions need to be determined, namely, the learning function, the training function, and the loss function. The learning function used in the present invention is the gradient descent momentum function learngdm. The characteristic of this function is that a momentum is added when updating the weights based on back propagation. In this method, α is the step size of each weight change, which is generally a very small number. And m and m - is the current momentum and the momentum of the previous moment, and β is a constant less than 1. As the iteration proceeds, the learning function with momentum m is used, and the updated value Δω of the weight in the previous iteration will be added to each momentum. Because β is a value less than 1, the earlier the weight update occurs, the smaller the impact on the momentum will be, but its impact will always remain in the momentum, so the weight update will be affected by more than one update, which can improve the stability of learning. And as the weight is updated, the momentum will become larger and larger, and the learning speed will become faster and faster. The loss function is a function used to measure the error between the output of the neural network and the target value for supervised training neural networks.

[0076] The loss function adopted by the present invention is mean square error (MSE). MSE is the average of the sum of squares of the difference between the output value of the neural network and the target value at each iteration. The smaller the mean square error, the better the performance of the neural network and the closer it is to the target value. The difference between the training function and the learning function is that the purpose of the training function is to make the final loss function of the entire network as small as possible, which is a way to train the network as a whole, while the learning function is a function that adjusts the weight of each node in the network. The output of the training function is the trained network and the training record. During the training process, the training function continuously calls the learning function to correct the weights, and the training is terminated by detecting that the set number of training steps or the error calculated by the loss function is less than the set error.

[0077] The training function used in the present invention is the trainlm function. The trainlm function refers to a training function using the L (Levenberg-marquardt) algorithm. The LM algorithm is a nonlinear optimization method between the Newton method and the gradient descent method. It is insensitive to the over-parameterization problem, can effectively handle the redundant parameter problem, and greatly reduces the chance of the cost function falling into the local minimum. The key parameter in the LM algorithm is the coefficient μ. When it is 0, the algorithm is the Newton method; when the coefficient μ is very large, the algorithm is the gradient descent method with a small step size. When the training just starts, a smaller coefficient μ is selected, and the Newton descent method will be used to update the weight. The Newton method has a second-order convergence speed and a faster training speed; as the training proceeds, the μ value will gradually increase, and the training will become the gradient descent method, and the training speed will gradually slow down, but the training results will gradually converge, and eventually stabilize, and the training is completed. For medium-sized neural networks, a very small mean square error can be obtained by using the LM algorithm, so the present invention uses this training algorithm.

[0078] The error curve during neural network training is as follows: Figure 5 As shown. The training set accounts for 70% of the data set, and its function is to train the neural network to minimize the error. The green part is the validation set, which accounts for 15% of the data set. Its function is to determine whether the trained neural network is suitable for this part of the input, that is, to check whether the neural network has overfitting or underfitting problems. If there are problems, retraining or further changing the feature values ​​are required. The red part is the test set, and its function is to test the performance of the neural network on the new set after verification and correction of the validation set, to check whether the above-mentioned overfitting or underfitting problems still exist. From Figure 5 It can be seen that the errors in the three sets are small, and there is no problem of overfitting or underfitting, so it is considered that the neural network training is successful.

[0079] The gradient change curve, coefficient μ change curve and verification times curve are as follows: Figure 6 As shown. It can be seen that the gradient is getting smaller and smaller, because as the iteration proceeds, the error becomes smaller and smaller, and the error function is closer to the minimum point, so its gradient is getting smaller and smaller and finally tends to 0; the coefficient μ curve first decreases, then the LM algorithm tends to the Newton method, and the training speed is faster, and then the curve rises, and the training tends to the gradient descent method, and the training becomes slower, but it is easier to converge; the verification curve shows that when the training is carried out for the fifth time, the condition for stopping training has been met. The condition for stopping training set by the present invention is when the gradient is less than 1e-7, or the coefficient is less than 0.001 or greater than 1e10.

[0080] The output of the neural network is not "1" or "0.1", but the result of the best possible fit after the activation function output. This study found out the output judgment value of the neural network is 0.505 through enumeration. When it exceeds this value, the network output is judged as "1", which is the maximum power transmission. If it is lower than this value, it is judged as "0.1", which is the minimum power transmission.

[0081] The above is a detailed introduction to a satellite-to-ground link power control allocation method based on a local recurrent neural network provided by the present invention. The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only used to help understand the method of the present invention and its core idea. At the same time, for those skilled in the art, according to the idea of ​​the present invention, there will be changes in the specific implementation methods and application scopes. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A satellite-to-ground link power control allocation method based on local recurrent neural network, characterized in that: The following steps are involved: Step 1: Based on the quantified state data of satellites and users, a data set for training the neural network is established and preprocessed. Within each access interval, the user terminal calculates its distance from each satellite using the ephemeris broadcast by each satellite, and the terminal selects the nearest satellite for access. Step 2: Select characteristic parameters for each link to determine the optimal transmission power of each link under different conditions; eight characteristic parameters considered when performing power allocation, for each link, the convolution result of the horizontal coordinate of the satellite that affects it, for each link, the convolution result of the vertical coordinate of the satellite that affects it, for each link, the convolution result of the horizontal coordinate of the average position of the users that affect it, for each link, the convolution result of the vertical coordinate of the average position of the users that affect it, the channel fading of the current link, the minimum channel fading of all links in the current access interval, and the maximum channel fading of all links in the current access interval; Step 3: Establish a neural network, select parameters, and train the neural network according to the selected parameters; The neural network structure is Elman neural network. The network has five layers, namely input layer, first hidden layer, second hidden layer, output layer, first receiving layer and second receiving layer. The input layer has eight nodes, corresponding to eight eigenvalues. Both hidden layers have 50 nodes, and the output layer has only one node; Step 4: Determine the output judgment value of the neural network by enumeration method, and determine the maximum power transmission or the minimum power transmission according to the data judgment value; Through enumeration method, it is found that the output judgment value of the neural network is 0.

505. When it exceeds the judgment value, the network output is judged as "1", that is, the maximum power transmission. When it is lower than the judgment value, it is judged as "0.1", that is, the minimum power transmission.

2. The satellite-to-ground link power control allocation method based on local recurrent neural network according to claim 1, characterized in that: The step 1 is specifically as follows: Based on the quantified state data of satellites and users, a data set for training a neural network is established. The data set is divided into three parts: a training set, a validation set, and a test set. The training set is used to train and update network parameters. The validation set tests the generalization of the network and determines whether overfitting occurs; the test set reflects the final performance of the neural network. Determine power allocation scenarios; The access scenario is: each low-orbit satellite has multiple wide beams and the communication frequency is the same. Then the coverage of a beam of a certain low-orbit satellite is also the coverage of the beam of other low-orbit satellites. There are a certain number of user terminals within this beam range. In each access interval, all user terminals are connected to the satellite network. When only the downlink is considered, the transmission power of each satellite is reasonably allocated to maximize the total channel capacity of the entire satellite network. Each satellite has a minimum transmission power. For the channel condition between the satellite and the user terminal, a shadow Rice channel is used.

3. The satellite-to-ground link power control allocation method based on local recurrent neural network according to claim 2, characterized in that: The step 2 is specifically as follows: After obtaining the characteristic parameters, the optimal transmission power of each link under different conditions is determined, and the maximum network channel capacity is obtained after power allocation using the genetic algorithm.

4. The satellite-to-ground link power control allocation method based on local recurrent neural network according to claim 3, characterized in that: The roulette wheel selection algorithm is used for selection operation. According to the individual fitness, individuals with good traits are selected to be passed on to the next generation in a certain way. The proportion of each individual's fitness to the sum of the fitness of all individuals in this generation is taken as the probability of each individual being selected, that is, the probability of being selected is proportional to the fitness of each individual. The crossover operation is to exchange the components of the feasible solution between a pair of selected individuals with a certain probability. The genetic algorithm used obtains the data set required for supervised training of the network.

5. The satellite-to-ground link power control allocation method based on local recurrent neural network according to claim 1, characterized in that: The hidden layer and output node of the neural network both require an activation function. The activation functions of the hidden layer and the receiving layer use the ReLU function, and the activation function of the output layer uses the Sigmoid function.

6. The satellite-to-ground link power control allocation method based on local recurrent neural network according to claim 1, characterized in that: The neural network adopts the learning function of the gradient descent momentum function learngdm. On the basis of back propagation, a momentum is added when updating the weights. is the step size of each weight change, and m and is the current momentum and the momentum at the previous moment, is a constant less than 1. As the iteration proceeds, the learning function with momentum m is used, and the updated value of the weight in the previous iteration is will be added to each momentum.

7. The satellite-to-ground link power control allocation method based on local recurrent neural network according to claim 1, characterized in that: The training function used in step 3 is the trainlm function, which refers to the training function using the LM algorithm, and the key parameters are the coefficients , when it is 0, the algorithm is Newton's method; As the training progresses, The value will gradually increase, the training will become the gradient descent method, the training speed will gradually slow down, the training results will gradually converge, and finally stabilize, and the training will be completed.

8. The satellite-to-ground link power control allocation method based on local recurrent neural network according to claim 7, characterized in that: The loss function used is the mean square error. In each iteration, the square of the difference between the output value of the neural network and the target value is averaged. The output of the training function is the trained network and training records. During the training process, the training function continuously calls the learning function to correct the weights. The training ends by detecting the set number of training steps or the error calculated by the loss function is less than the set error.

9. The satellite-to-ground link power control allocation method based on local recurrent neural network according to claim 8, characterized in that: The training stop condition is set to when the gradient is less than 1e-7, or the coefficient is less than 0.001 or greater than 1e10.

Images