Optimization method, system and device for predicting uplink throughput based on rsrp and medium

Abstract

The application discloses an optimization method, system and device for predicting uplink throughput based on RSRP and a medium. The method comprises the following steps: constructing a data set comprising RSRP values and corresponding uplink throughputs; constructing an RSRP and uplink throughput mapping relationship to obtain a mapping model; performing multi-step prediction on the time sequence of RSRP based on an improved LSTM model, wherein the improved LSTM model comprises an encoder, an attention mechanism layer, a decoder and a full connection layer; the attention mechanism layer is used to calculate the correlation weight between the current time hidden state and all hidden states of the encoder at each step of the decoder prediction, and generate a weighted sum as the input of the decoder at the current time according to the correlation weight; and the value of the future time slot RSRP is predicted by using the improved LSTM model, and the uplink throughput is predicted in combination with the mapping model. The application predicts the uplink throughput by using the 5G network parameter RSRP based on the deep learning technology, so that the resources can be allocated in advance to meet the demand of the current system on the communication performance.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and in particular to an optimization method, system, apparatus and medium for predicting uplink throughput based on RSRP, which uses RSRP to predict the uplink throughput of future time slot terminals. Background Technology

[0002] With the continuous development of wireless communication technology, people's demands for network performance and user experience are also increasing. In the field of mobile communications, how to effectively predict and optimize the uplink throughput of terminal devices has become a key challenge.

[0003] In wireless communication networks, terminal devices transmit and receive data by communicating with base stations. During this process, RSRP (Reference Signal Received Power), as a crucial indicator of signal strength between the terminal device and the base station, is essential for assessing network coverage and connection quality. Uplink throughput, on the other hand, directly relates to the data upload rate of the terminal device and the network performance.

[0004] Traditionally, network operators and engineers assess network performance through real-time monitoring and statistical data, adjusting network parameters and resource allocation as needed. However, this approach often only reflects the current network state and lacks accurate predictions of future network performance.

[0005] To address this issue, Eyeceyurt E et al. proposed a method for predicting uplink throughput based on RSRP. However, while explaining the importance of RSRP in predicting uplink throughput, they did not specifically propose a method for predicting uplink throughput at future times (Eyceyurt E, Egi Y, Zec J. Machine-learning-based uplink throughput prediction from physical layer measurements[J]. Electronics, 2022, 11(8):1227). Chen Fangjiong et al. disclosed a "multi-user multi-task computing offloading method, device and medium with throughput prediction" in Chinese invention patent CN117177279A. This method analyzes historical RSRP data and uses data modeling and prediction algorithms to predict the RSRP value of terminal devices in future time slots, thereby further estimating the uplink throughput of terminal devices. However, the KNN+ARIMA model used in this scheme performed poorly, with low prediction accuracy. Summary of the Invention

[0006] To address the shortcomings of existing technologies to some extent, this invention proposes an optimization method, system, device, and medium for predicting uplink throughput based on RSRP, which can effectively improve the accuracy of prediction, optimize network resource allocation, and enhance network performance and user experience.

[0007] To achieve the objectives of this invention, the present invention provides an optimization method for predicting uplink throughput based on RSRP, comprising the following steps:

[0008] Data collection was conducted to construct a dataset including RSRP values ​​and corresponding uplink throughput;

[0009] The mapping relationship between RSRP and uplink throughput is constructed based on the decision tree regression algorithm, and the mapping model is obtained;

[0010] This paper proposes a multi-step prediction method for RSRP based on an improved LSTM model. The improved LSTM model includes an encoder, an attention mechanism layer, a decoder, and a fully connected layer. The encoder processes RSRP sequence data and generates corresponding hidden states. The attention mechanism layer calculates the correlation weights between the current hidden state and all hidden states of the encoder at each prediction step of the decoder, and generates a weighted sum as the input of the decoder at the current step. The decoder combines the weighted sum generated by the attention mechanism with the prediction value of the previous step to perform multi-step prediction. The fully connected layer performs a linear transformation on the output of the decoder and outputs the RSRP prediction value.

[0011] The improved LSTM model is used to predict the future slot RSRP value, and the uplink throughput is predicted in combination with the mapping model.

[0012] Furthermore, mobile terminals such as Android phones can be used to collect RSRP data and corresponding uplink throughput in historical time slots around the campus via the cellular-z software, and then preprocess the data.

[0013] Furthermore, a decision tree model is constructed using the CART algorithm for decision tree regression, fitted to the training data, and the input feature RSRP is used to output uplink throughput. Preferably, the model is evaluated using test data, and its performance is measured by calculating metrics such as error, precision, and recall. Then, the model parameters are adjusted as needed to optimize model performance.

[0014] RSRP mapping uplink throughput is a system with a single characteristic. Assuming that n sets of data are collected, the dataset is represented as D = {e1, e2, K, e...} n The corresponding response set is L={r1,r2,K,r}. n}, where D is an ordered set, e ir represents the value represented by the i-th group of RSRP. i It is e i The corresponding uplink throughput, i = 1...n; split the dataset D at the split point m, D1 = {e1, e2, K, e m} and D2={e m+1 ,e m+2 ,K,e n} represents the two subsets of data after splitting, L1={r1,r2,K,r m} and L2={r m+1 ,r m+2 ,K,r n} represent the response sets of data subsets D1 and D2 respectively. The weighted mean square error of the two response subsets L1 and L2 is calculated, and the split point m with the lowest weighted mean square error of the response subset is regarded as the optimal split point.

[0015] Furthermore, an improved LSTM model combining LSTM and an attention mechanism is trained using historical RSRP data from the cache. Through learning, a multi-step prediction model for RSRP is finally obtained. The prediction performance of the neural network is improved by adding an attention mechanism to the LSTM layer, and the model's generalization ability is enhanced by optimizing the model parameters using a genetic algorithm.

[0016] Furthermore, the selection operator of the genetic algorithm is a random acceptance-based method, which includes the following steps:

[0017] Calculate the sum of fitness scores;

[0018] Normalized fitness;

[0019] Accumulate the normalized fitness values ​​to construct a cumulative probability distribution;

[0020] Generate a random number between 0 and 1, use binary search to find the interval corresponding to the random number in the cumulative probability distribution, and determine the selected individual.

[0021] Furthermore, both the encoder and the decoder include multiple LSTM units.

[0022] Furthermore, the processing steps in the attention mechanism layer include:

[0023] The hidden state sequence output by the encoder is obtained. These hidden states capture the dynamic information and features of the input sequence through time step propagation and serve as the input to the attention mechanism layer, i.e., the initial sample {x}. 1 x 2 , ..., x N};

[0024] A fully connected layer is constructed using a linear transformation function and an activation function, and a new sample {e} is obtained through the fully connected layer. 1 e 2 , ..., e N};

[0025] A new sample {e} will be obtained 1 e 2 , ..., e N The weight vector W is obtained by inputting the time steps into the softmax() function. The number of column vector elements in the weight vector W is equal to the number of input time steps, and each element represents the weight of a time step.

[0026] The initial sample {x 1 x 2 , ..., x N The sample {x} is obtained by multiplying the weight vector W with the weight W. 1 W, x 2 W, ..., x N W};

[0027] The weighted sample {x 1 W, x 2 W, ..., x N W is summed according to the time step sequence to obtain the final output of the attention mechanism layer, which is a context vector.

[0028] In this layer, the query vector q in the attention mechanism layer is the corresponding dependent variable, i.e., RSRP. Therefore, the weight vector W is dynamically adjusted with q during the neural network training process.

[0029] Furthermore, the input to the decoder at each step includes a weighted sum of the predicted value from the previous step and the output of the attention mechanism layer. This weighted sum forms a context vector, which, along with the context vector, constitutes the input vector. The decoder generates the output predicted value y for the next time step based on the current hidden state, the input vector, and the memory unit. t With the new hidden state, the decoder updates the hidden state and memory unit, and sets the current output prediction value y. t Input for the next time step.

[0030] The present invention also provides an optimization system for predicting uplink throughput based on RSRP, the system being used to implement the aforementioned method, the system comprising the following modules:

[0031] The dataset building module is used for data collection and to build a dataset that includes RSRP values ​​and corresponding uplink throughput.

[0032] The mapping model building module is used to construct the mapping relationship between RSRP and uplink throughput based on the decision tree regression algorithm, and obtain the mapping model.

[0033] The LSTM model building module is used to perform multi-step prediction of RSRP on a time series based on an improved LSTM model. The improved LSTM model includes an encoder, an attention mechanism layer, a decoder, and a fully connected layer. The encoder is used to process RSRP sequence data and generate corresponding hidden states. The attention mechanism layer is used to calculate the correlation weight between the current hidden state and all hidden states of the encoder at each step of the decoder prediction, and generate a weighted sum based on this, which serves as the input of the decoder at the current step. The decoder is used to combine the weighted sum generated by the attention mechanism and the prediction value of the previous step to perform multi-step prediction. The fully connected layer is used to perform a linear transformation on the output of the decoder and output the RSRP prediction value.

[0034] The prediction module is used to predict the value of future slot RSRP using the improved LSTM model and to predict uplink throughput using the mapping model.

[0035] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the aforementioned method.

[0036] The present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, wherein the computer program or instructions, when executed by a processor, implement the steps of the aforementioned method.

[0037] Compared with the prior art, the present invention can achieve at least the following beneficial effects:

[0038] (1) This invention is based on deep learning technology and uses the 5G network parameter RSRP to predict uplink throughput, thereby allocating resources in advance to meet the current system's communication performance requirements.

[0039] (2) This invention employs decision tree and LSTM (Long Short-Term Memory) techniques for prediction. Compared to the KNN+ARIMA model, the decision tree model can more accurately capture the complex relationship between RSRP and uplink throughput. Using the LSTM network to handle long-term dependencies in time-series data allows for more effective utilization of historical data to predict future RSRP and uplink throughput. This improvement not only enhances prediction accuracy but also effectively optimizes network resource allocation, resulting in significant improvements in network performance and user experience. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating RSRP prediction of uplink throughput in an embodiment of the present invention;

[0041] Figure 2 This is a schematic diagram of the decision tree regression process for a single feature dataset in an embodiment of the present invention;

[0042] Figure 3 This is a schematic diagram of the recurrent neural network structure in an embodiment of the present invention;

[0043] Figure 4 This is a schematic diagram of the recurrent neural network unfolding structure in an embodiment of the present invention;

[0044] Figure 5 This is a diagram of the long short-term memory network structure in an embodiment of the present invention;

[0045] Figure 6 This is a flowchart of the genetic algorithm in an embodiment of the present invention;

[0046] Figure 7 This is a schematic diagram of a multi-step prediction model for RSRP based on an improved LSTM in an embodiment of the present invention.

[0047] Figure 8 This is a schematic diagram of the average root mean square error of SS-RSRP prediction in a low-speed moving scenario in an embodiment of the present invention.

[0048] Figure 9 This is a schematic diagram of the average root mean square error of SS-RSRP prediction in a high-speed moving scene according to an embodiment of the present invention. Detailed Implementation

[0049] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the technical solutions in the embodiments of the present invention will be described in detail and clearly below with reference to the accompanying drawings. It should be noted that the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] Please see Figure 1 The present invention provides an optimization method for predicting uplink throughput based on RSRP, comprising the following steps:

[0051] Step 1: Data acquisition and preprocessing.

[0052] Step 1 specifically includes: collecting data through the terminal during the initial process to form a 5G dataset, including RSRP values ​​and their corresponding uplink throughput.

[0053] In some embodiments of the present invention, the collected 5G dataset may contain non-numerical and duplicate values ​​due to noise and environmental influences, which affects the deep learning analysis in the feature extraction and normalization stages. Therefore, erroneous values ​​should be removed from the data before feature extraction.

[0054] In some embodiments of the present invention, a mobile user terminal, such as an Android phone, is used to collect RSRP data and corresponding uplink throughput in historical time slots around a collection location, such as a campus, using software cellular-z, and preprocessing is performed to remove non-numerical and duplicate values.

[0055] Step 2: Construct the mapping relationship between RSRP and uplink throughput based on the decision tree regression algorithm to obtain the mapping model.

[0056] Step 2 specifically includes: constructing a regression model using the CART algorithm for decision tree regression, fitting the model to the training data, with the RSRP collected in Step 1 as the input and the uplink throughput as the output. The regression model is then evaluated using test data, and its performance is measured by calculating metrics such as error.

[0057] Next, we will introduce the principle of decision tree regression and provide an example of the mapping model between RSRP and uplink throughput.

[0058] Decision Tree Regression (DTR): A decision tree is a hierarchical model for partitioning a dataset. Its goal is to divide the dataset into clean subsets, where data within each subset has highly similar response values. The smaller the mean squared error of the response values ​​in the partitioned subsets, the more concentrated the response values ​​are around the average response value of that subset. This indicates that the average response value of that subset can be considered a reasonable prediction of the response value of any data point within it. Therefore, minimizing the weighted mean squared error of the subset response sets becomes a metric when evaluating potential split points in a dataset, ensuring a balance between predictive accuracy and model complexity.

[0059] In this invention, RSRP mapping uplink throughput is a system with a single characteristic. Assuming step 1 collects n sets of data, the resulting dataset is represented as D = {e1, e2, K, e...}. n}(D is an ordered set, which is the set of RSRPs collected in step 1, e i The value represented by the i-th RSRP is given by the response set L = {r1, r2, K, r...}. n}, where L is the set of uplink throughput corresponding to RSRP, and r i It is a continuous numerical value, representing e. iThe corresponding uplink throughput, i = 1...n. Split the dataset D at the split point m, D1 = {e1, e2, K, e...} m} and D2={e m+1 ,e m+2 ,K,e n} represents the two subsets after splitting, L1={r1,r2,K,r m} and L2={r m+1 ,r m+2 ,K,r n} represent the response sets of data subsets D1 and D2 respectively. The weighted mean square error of the two response subsets L1 and L2 is calculated, and the split point m with the lowest weighted mean square error of the response subset is regarded as the optimal split point.

[0060] In some embodiments of the present invention, an example of decision tree regression is given. Assume there exists a dataset X1 = {1, 2, 3, 8, 9, 10}, and a corresponding response set L = {1.1, 1.2, 0.5, 1.3, 1.4, 0.8}. The root node of the decision tree is feature X1, and the decision tree segmentation process is as follows: Figure 2 As shown, for simplicity, this example uses the dataset mean as the splitting parameter. Data sets with a mean less than or equal to the mean are assigned to the left subset, and data sets with a mean greater than the mean are assigned to the right subset. Once the decision tree reaches a preset depth, it stops splitting the dataset to control the tree's complexity.

[0061] The advantages of decision tree regression include the model's intuitiveness and interpretability. Because it clearly maps the feature space to the response space, it is easy to understand and interpret. The non-parametric nature of decision trees allows the model to capture complex non-linear patterns in the training data without relying on prior assumptions about the data's distribution.

[0062] Step 3: Perform multi-step prediction of RSRP on the time series based on the improved LSTM model.

[0063] In a 5G module, this invention can periodically obtain the RSRP value, resulting in a time series with multiple RSRP values. Then, it performs multi-step predictions and substitutes the data into the mapping model obtained in step 2 to derive the uplink throughput.

[0064] 1. Relevant theories and models

[0065] a. LSTM Neural Network

[0066] To address the problem of poor time series prediction performance of traditional Long Short Term Memory (LSTM) neural networks, this invention proposes an improved LSTM model. By adding an attention mechanism to the LSTM layer, the prediction performance of the neural network is improved, and the model generalization ability is enhanced by optimizing the model parameters through a genetic algorithm.

[0067] LSTM is a variant of Recurrent Neural Network (RNN), and the structure of a recurrent neural network is as follows: Figure 3 As shown, the main characteristic of recurrent neural networks (RNNs) is the introduction of recurrent connections between the hidden layers of the neural network, allowing information to be passed within the network. The output at each time step depends not only on the current input but also on the outputs of previous time steps. Therefore, RNNs can capture the temporal dependencies in a sequence.

[0068] The internal structure of the hidden layers of a recurrent neural network (RNN) is expanded and refined, such as... Figure 4 As shown, u represents the weight matrix from the input layer to the hidden layer, v represents the weight matrix from the hidden layer to the output layer, and w is the weight of the hidden layer's previous value used as the input at the current time step. The hidden state at each time step is propagated to the next time step; therefore, the RNN uses a chain rule in its computation. This leads to the vanishing gradient problem during backpropagation (BP), making it unable to handle long-term dependencies.

[0069] Compared to traditional recurrent neural networks, LSTM models introduce memory units, with the core being the cell state, responsible for storing and transmitting information. LSTM models control the input, output, and forgetting of cell states through three gates: the input gate, the output gate, and the forget gate. Figure 5 As shown:

[0070] The state update formulas for the three types of gates are expressed as follows:

[0071] f t =σ(W f ·[h t-1 ,x t ]+b f (1)

[0072] i t =σ(W i ·[h t-1 ,x t ]+b i (2)

[0073] ot =σ(W o ·[h t-1 ,x t ]+b o (3)

[0074] Where, the subscript t represents the position in the current time series being processed, f represents the forget gate, i represents the input gate, o represents the output gate, C represents the cell state, and h represents the hidden state. The forget gate determines how much information to delete from the memory unit, the input gate determines the degree to which the memory unit is updated, and the output gate determines how much content of the memory unit is output. W and b represent the weight matrix and bias vector, respectively, and σ is the sigmoid function, which is used as an activation function with an activation value ranging from 0 to 1. In equation (1), σ = 0 indicates complete forgetting, and 1 indicates complete retention. In equation (2), σ determines the updated part of the data, and is related to the new candidate value vector C created by the tanh layer. t Merging is used to update the cell state. C t The updated formula is as follows:

[0075] C t '=tanh(W C ·[h t-1 ,x t ]+b C (4)

[0076] New cell state C t It will depend on the information forgotten and the information newly remembered:

[0077] C t =f t ·C t-1 +i t ·C t (5)

[0078] Finally, the output value h of the hidden node t The value of the output gate will be determined by o. t and cell state C t constitute:

[0079] h t =o t ·tanh(C t (6)

[0080] This type of gating mechanism allows LSTM models to adaptively retain and forget information, thus better handling long-term dependencies. The state is calculated by accumulation, which solves the gradient vanishing problem that often occurs in traditional recurrent neural networks when dealing with long-term series. RSRP time series often contain long-term temporal dependencies, so LSTM models are more likely to capture these dependencies than RNN models and can more accurately predict the future values ​​of RSRP sequences.

[0081] b. Attention mechanism

[0082] The attention mechanism is inspired by attention in cognitive neuroscience. When processing complex external information, the human brain consciously or unconsciously selects a small portion of the input information for focused processing while ignoring other information; this ability is called attention. Neural networks can also borrow this attention mechanism when processing large amounts of input information, selecting only key information for processing. If the selected information is the expectation of all input vectors under the attention distribution, it is called a soft attention mechanism; conversely, if only one input vector is focused on, it is called a hard attention mechanism.

[0083] The computation of soft attention mechanisms generally involves two steps: first, calculating the attention distribution, or weights, across all input information; and second, calculating a weighted average of the input information based on the attention distribution. Different weights are assigned to reflect the importance of various information types.

[0084] This invention proposes an improved LSTM model that combines an attention mechanism with a Long Short-Term Memory (LSTM) network for multi-step prediction of RSRP (Reference Signal Received Power). In the original LSTM model, RSRP data is input into an encoder for encoding. The encoder learns the final cell state by progressively processing the input sequence data. This cell state is then fed into the decoder to generate multi-step RSRP predictions. In this invention, an attention mechanism is introduced to enhance the model's predictive performance. Specifically, the attention mechanism allows the decoder to reference the hidden states of all time steps output by the encoder when generating each prediction, rather than relying solely on the final cell state. This allows the decoder to dynamically focus on different parts of the input sequence, thereby capturing richer temporal information and improving prediction accuracy.

[0085] The specific implementation steps of this invention are as follows:

[0086] Data preprocessing: The original RSRP data is normalized to ensure that the data is within a reasonable numerical range, preventing the values ​​from being too large or too small from affecting the stability of model training.

[0087] Model structure design:

[0088] • Encoder: Includes multiple LSTM units for processing the input RSRP sequence data and generating the corresponding hidden states and cell states.

[0089] • Attention mechanism: Used to calculate the correlation weights between the current hidden state and all hidden states of the encoder at each prediction step of the decoder, and generate a weighted sum (context vector) based on this, which serves as the input of the decoder at the current step.

[0090] • Decoder: Includes multiple LSTM units, which combine the weighted sum generated by the attention mechanism with the previous predictions to perform multi-step prediction.

[0091] Training process: The improved LSTM model is trained end-to-end using training data. A genetic algorithm is employed to optimize the model parameters by minimizing the loss function between the predicted and true values. In some embodiments of this invention, the loss function used is the mean squared error (MSE).

[0092] Prediction process: After the model training is completed, the trained improved LSTM model is used to perform multi-step predictions on the new RSRP data. First, the new data is input into the encoder to generate hidden states. Then, at each prediction step, the decoder, combined with an attention mechanism, refers to the hidden states of the encoder to generate the RSRP prediction value for the next time step.

[0093] Post-processing: The predicted RSRP values ​​are denormalized to restore them to their original range so that they can be compared and evaluated with the actual RSRP values.

[0094] c. Genetic Algorithm

[0095] Genetic Algorithms (GA) are stochastic global search and optimization methods developed based on biological evolution mechanisms such as natural selection and genetic variation, mimicking these mechanisms. Essentially, they are efficient, parallel, and global search methods that automatically acquire and accumulate knowledge about the search space during the search process and adaptively control the search to find the optimal solution. For example... Figure 6 As shown, the implementation process of a genetic algorithm can be roughly divided into the following steps:

[0096] Population initialization: First, represent the feasible solutions to the problem as chromosomes or individuals in the genetic space through encoding. Let the population size be M.

[0097] Fitness: Fitness is the degree of an individual's advantage in surviving within a population, and is measured using a fitness function or evaluation function.

[0098] Selection: Defines the solution to be retained for further replication. Examples include roulette wheel selection and tournament selection.

[0099] Crossover: describes how to create new solutions from existing solutions.

[0100] Mutation: Its purpose is to introduce diversity and novelty into the solution pool by randomly swapping or closing solutions. For example, binary mutation.

[0101] Since this invention uses a recurrent neural network (RNN) model, the recurrent structure of RNNs makes them difficult to parallelize, thus training RNNs requires a significant amount of time. Using a genetic algorithm to fine-tune the parameters of the recurrent random network further increases the overall model tuning time. Therefore, enabling the genetic algorithm to converge more quickly is crucial.

[0102] As can be seen from the above process, the crossover operator is responsible for combining and optimizing genes, the mutation operator is responsible for extensively searching the entire space, and the selection operator is responsible for selecting solutions with high fitness values. Both crossover and mutation are unguided operations, so the selection operator must ensure that the population evolves towards the desired goal. Through the selection operator, individuals with high fitness values ​​will have a greater chance of being selected as parents for the next generation. Based on this, this invention improves the selection operator part of the genetic algorithm to accelerate the convergence speed of the genetic algorithm.

[0103] The selection operator commonly used in genetic algorithms is the roulette wheel selection method. The traditional roulette wheel selection method involves the following steps:

[0104] • Calculate the sum of fitness values: sum the fitness values ​​of all individuals in the population.

[0105] • Normalized fitness: Normalize the fitness value of each individual and calculate its proportion in the total fitness.

[0106] • Select an individual: Generate a random number between 0 and 1, and select the corresponding individual based on which segment of the normalized fitness interval the random number falls into.

[0107] This method has a time complexity of O(M) because it requires iterating through all fitness values ​​each time an individual is selected.

[0108] The selection operator used in this invention is based on a random acceptance method.

[0109] To reduce the time complexity of the selection process, this invention introduces a random acceptance-based method. This method reduces the time complexity of each individual selection by preprocessing, constructing a cumulative probability distribution, and performing a binary search. The specific steps are as follows:

[0110] • Calculate the sum of fitness values: sum the fitness values ​​of all individuals in the population.

[0111] • Normalized fitness: Normalize the fitness value of each individual and calculate the proportion of each individual's fitness value in the total fitness.

[0112] • Construct a cumulative probability distribution (CDF): Accumulate the normalized fitness values ​​to construct a cumulative probability distribution.

[0113] • Select an individual:

[0114] a. Generate a random number between 0 and 1.

[0115] b. Use binary search to find the interval corresponding to the random number in the cumulative probability distribution, and determine the selected individual.

[0116] The improvement of this method is that it uses binary search instead of traversal search, which reduces the time complexity of the selection process from O(M) to O(log(M)).

[0117] 2. Model Building

[0118] Model construction includes: data processing, inputting data into the encoder's LSTM unit, an attention mechanism layer, the decoder's LSTM unit, a fully connected layer, and genetic algorithm tuning. The overall model flow is as follows: Figure 7 As shown. Data processing ensures the quality and applicability of the input data. LSTM units are used to process dependencies in the RSRP dataset collected in step 1, extracting important temporal features. The attention mechanism layer emphasizes important time steps, improving the model's prediction accuracy and interpretability. Fully connected layers are used to transform high-level features into the final prediction results. Genetic algorithms are used to tune and optimize the model's hyperparameters, improving overall performance.

[0119] a. Data processing

[0120] First, data cleaning is performed to handle missing values, outliers, etc. Second, the total sample is divided into training, validation, and test sets according to a certain ratio. Then, the data is normalized: the data is scaled to a certain range (e.g., between 0 and 1) to improve model stability and convergence speed.

[0121] Let 'samples' represent the number of 5G datasets collected in step 1, indicating the number of samples the neural network can process during each training iteration. Let 'timesteps' represent the number of past steps of data used to predict the future. Let 'features' represent the number of features, which is 1 here, i.e., the RSRP value. Finally, the sample data is organized into a [samples, timesteps, features] shape according to the input format of the LSTM neural network layers.

[0122] b. LSTM unit

[0123] The relevant parameters that need to be set for the LSTM unit are as follows. In the PyTorch package, you can set the number of hidden neurons in each layer through hidden_size, set the number of layers through n_layers, and set whether the LSTM is bidirectional through the bidirectional parameter.

[0124] In the case of unidirectional LSTM, the dimension of the input data of the LSTM unit is [samples, timesteps, features], and the dimension of the output data is [samples, timesteps, hidden_size].

[0125] c. Attention mechanism layer

[0126] Currently, there are many attention mechanisms available. Since RSRP is a mobile communication time series, the RSRP at each moment is influenced by historical values ​​to varying degrees. To better reflect the different degrees of influence of historical values ​​on the current value, this invention selects the timestep parameters in the output data of the LSTM unit as the attention mechanism, that is, assigning different attention to data at each lagging time step. The timestep weight parameter W is obtained through deep learning. The specific algorithm flow is as follows:

[0127] 1) Obtain the hidden state sequence output by the LSTM unit of the encoder. The hidden state sequence refers to the hidden state of the LSTM unit at each time step. These hidden states capture the dynamic information and features of the input sequence through time step propagation, serving as the input to the attention mechanism layer, i.e., the initial sample {x}. 1 x 2 , ..., x N};

[0128] 2) Construct a fully connected layer using the Linear Transform function and the ReLU activation function, and obtain new samples {e} through the fully connected layer. 1 e 2 , ..., e N};

[0129] 3) Take the new samples obtained in 2) and input them into the softmax() function according to the timesteps. The softmax() function maps the output to values ​​in the range (0, 1) and performs normalization. The sum of all elements equals 1. After inputting into the softmax() function, the weight vector W is obtained. The number of column vector elements in the weight vector W is equal to the number of input timesteps, and each element represents the weight of a timestep.

[0130] 4) The initial sample {x} from step 1) 1 x 2 , ..., x N The sample {x} is multiplied by the weight vector W to obtain the sample {x} with weight W. 1 W,x 2 W, ..., x N W}, such as Figure 7 As shown;

[0131] 5) Sum the samples assigned weight W along the timesteps column to obtain the final output of the attention mechanism layer, which is a context vector. This context vector is obtained by weighted averaging of the hidden state sequence output by the encoder's LSTM unit. Each hidden state at each time step is assigned an attention weight, which represents the relative importance of that time step to the final output.

[0132] In the attention mechanism used in this invention, the query vector q is the corresponding dependent variable, i.e., RSRP. Therefore, during the neural network training process, the weight vector W is dynamically adjusted along with q.

[0133] This invention employs a soft attention mechanism and calculates relevance scores based on features and query vectors after linear transformation, generating a bilinear model scoring function for attention weights. The bilinear scoring function is used as the probability calculation method for selecting a specific time step, implemented in a fully connected layer during network design. In terms of feature dimensions, time weights are shared; that is, the weights corresponding to each feature at the same time step are the same. The weights of each time step are assigned to each feature and summed to obtain the final time-weighted feature vector for each feature, which serves as the input to the next layer.

[0134] The input dimension of the entire attention mechanism layer is [batch_size, timesteps, hidden_size], and the output dimension is [batch_size, hidden_size].

[0135] In the decoder, the process of multi-step prediction, combining the weighted sum (context vector) generated by the attention mechanism layer with the previous predictions, is as follows:

[0136] 1. Output of the Attention Mechanism Layer (Context Vector): The output of the attention mechanism layer is a context vector, which is obtained by weighted averaging of the hidden state sequence output by the encoder's LSTM unit. Each time step's hidden state is assigned an attention weight, which represents the relative importance of that time step to the final output.

[0137] 2. Decoder input: During the decoding process, the input of the LSTM unit of the decoder at each step includes not only the predicted value of the previous step, but also the context vector generated by the attention mechanism layer.

[0138] 3. Decoder workflow:

[0139] • Initial Input: At the first time step of the predicted sequence, the decoder typically uses an initial input (e.g., a special start marker or zero vector) and a context vector.

[0140] • Step-by-step prediction:

[0141] a) At time step t, the input received by the decoder is the predicted value y from the output of the previous time step. t-1 The input vector, composed of the context vector, is specifically represented as: Input t =[y t-1 [context];

[0142] b) The input vector is passed to the LSTM unit of the decoder, which generates the output prediction y for the next time step based on the current hidden state, the input vector, and the internal state (memory unit) of the LSTM. t And a new hidden state.

[0143] • Update State: The decoder's LSTM unit updates its hidden state and memory unit, and sets the current output prediction value y. t Input for the next time step.

[0144] 4. Multi-step prediction:

[0145] This input method ensures that the prediction at each time step depends not only on past predictions (stepwise recursive input) but also on global contextual information (context vector).

[0146] By continuously combining the context vector and the prediction value from the previous time step into the decoder input, the decoder is able to generate the prediction value for subsequent time steps.

[0147] d. Fully connected layer

[0148] The fully connected layer has 1 neuron, an input dimension of [batch_size, hidden_size], and an output dimension of [batch_size, 1]. The hidden_size is equal to the hidden_size set in the previous LSTM unit. The fully connected layer maps to the target value: the output of the decoder's LSTM unit is normalized data, so it undergoes a linear transformation through the fully connected layer to obtain the final predicted value (RSRP value).

[0149] e. Genetic Algorithm Tuning

[0150] In this invention, a genetic algorithm is used to optimize the hyperparameters of an improved LSTM model with an added attention mechanism layer to improve the accuracy of RSRP in predicting uplink throughput. Specifically, the optimized hyperparameters include the number of timesteps and the hidden layer size.

[0151] Specific steps of genetic algorithm:

[0152] 1) Initialize the population:

[0153] Generate an initial set of individuals, each representing a combination of hyperparameters. For example, an individual can be represented as (timesteps=10, hidden_size=64).

[0154] 2) Fitness calculation:

[0155] The root mean square error (RMSE) of the model on the validation set is used as the fitness function. The fitness value of each individual is calculated, and it is determined whether it meets the optimization criteria. Optimization criteria include reaching the maximum number of iterations or satisfying the convergence condition.

[0156] 3) Selection:

[0157] Individuals are selected based on their fitness values, with individuals having a higher fitness level being more likely to be selected. The roulette wheel selection method is improved to a random acceptance method to balance exploration and exploitation.

[0158] 4) Crossover and mutation:

[0159] Set the crossover probability to generate new offspring from the selected individuals. Set the mutation probability to randomly mutate some genes of the offspring individuals. Generate offspring individuals. Iteratively calculate the fitness, generating a new generation population through crossover and mutation, defining the number of iterations, and gradually finding a better combination of hyperparameters. This optimal scheme will be used for the final evaluation of the model on the test set to verify its effect on improving the uplink throughput of RSRP prediction.

[0160] Step 4: Use the improved LSTM model from Step 3 to predict the future slot RSRP value, and combine it with the mapping model constructed in Step 2 to predict the uplink throughput.

[0161] In some embodiments of the present invention, RSRP time series prediction experiments were conducted for low-speed and high-speed movement scenarios, respectively. The experimental results include the average root mean square error (RMSE) of the three prediction models under different RSRP training set sizes. Figure 8 and Figure 9 The data shows that as the size of the RSRP training set increases, the mean root mean square error decreases rapidly, and the model's prediction performance improves accordingly. For low-speed moving scenarios, the mean root mean square error gradually stabilizes after the number of training points reaches approximately 1000, indicating convergence. For high-speed moving scenarios, the mean root mean square error converges after the number of training points reaches approximately 800. Furthermore, it can be observed that the improved LSTM model proposed in this invention yields the lowest mean root mean square error between the predicted results and the observed values, followed by the ARIMA model, while the quadratic exponential smoothing model yields the highest, meaning that the LSTM model has the best prediction performance.

[0162] Compared to the KNN+ARIMA model, this invention, by using decision trees and LSTM techniques, significantly improves prediction accuracy and reduces the error rate by 3%. This improvement not only enhances prediction accuracy but also effectively optimizes network resource allocation, resulting in a significant improvement in network performance and user experience.

[0163] In some embodiments of the present invention, an optimization system based on RSRP for predicting uplink throughput is also provided. This system implements the method provided in the foregoing embodiments and includes the following modules:

[0164] The dataset building module is used for data collection and to build a dataset that includes RSRP values ​​and corresponding uplink throughput.

[0165] The mapping model building module is used to construct the mapping relationship between RSRP and uplink throughput based on the decision tree regression algorithm, and obtain the mapping model.

[0166] The LSTM model building module is used to perform multi-step prediction of RSRP on a time series based on an improved LSTM model. The improved LSTM model includes an encoder, an attention mechanism layer, a decoder, and a fully connected layer. The encoder is used to process RSRP sequence data and generate corresponding hidden states. The attention mechanism layer is used to calculate the correlation weight between the current hidden state and all hidden states of the encoder at each step of the decoder prediction, and generate a weighted sum based on this, which serves as the input of the decoder at the current step. The decoder is used to combine the weighted sum generated by the attention mechanism and the prediction value of the previous step to perform multi-step prediction. The fully connected layer is used to perform a linear transformation on the output of the decoder and output the RSRP prediction value.

[0167] The prediction module is used to predict the value of future slot RSRP using the improved LSTM model and to predict uplink throughput using the mapping model.

[0168] In some embodiments of the present invention, a computer device is also provided, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the method provided in the foregoing embodiments.

[0169] In some embodiments of the present invention, a computer-readable storage medium is also provided, on which a computer program or instructions are stored, which, when executed by a processor, implement the steps of the method provided in the foregoing embodiments.

[0170] As for the RSRP-based uplink throughput prediction optimization system disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and relevant details can be found in the method section.

[0171] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. An optimization method for predicting uplink throughput based on RSRP, characterized in that, Includes the following steps: Data collection was conducted to construct a dataset including RSRP values ​​and corresponding uplink throughput; The mapping relationship between RSRP and uplink throughput is constructed based on the decision tree regression algorithm, and the mapping model is obtained; This paper proposes a multi-step prediction method for RSRP based on an improved LSTM model. The improved LSTM model includes an encoder, an attention mechanism layer, a decoder, and a fully connected layer. The encoder processes RSRP sequence data and generates corresponding hidden states. The attention mechanism layer calculates the correlation weights between the current hidden state and all hidden states of the encoder at each prediction step of the decoder, and generates a weighted sum as the input of the decoder at the current step. The decoder combines the weighted sum generated by the attention mechanism with the prediction value of the previous step to perform multi-step prediction. The fully connected layer performs a linear transformation on the output of the decoder and outputs the RSRP prediction value. The improved LSTM model is used to predict the value of future slot RSRP, and the uplink throughput is predicted in combination with the mapping model. The CART algorithm for decision tree regression is used to construct a regression model, which is then fitted to the training data. The input is the collected RSRP value, and the output is the uplink throughput. RSRP mapping uplink throughput is a system with a single characteristic. Assuming n sets of data are collected, the resulting dataset is represented as... The corresponding response set is ,in, D For an ordered set, This represents the value represented by the i-th group of RSRP. yes The corresponding uplink throughput, i = 1...n; at the split point m Data set D Split into pieces. and This represents the two subsets of data after splitting. and These represent data subsets respectively. and Given a set of responses, compute two subsets of responses. and The weighted mean square error is used to determine the split point with the lowest weighted mean square error in the response subset. m Considered the optimal dividing point; When training the improved LSTM model, the improved LSTM model is trained end-to-end using training data. The model parameters are optimized by minimizing the loss function between the predicted and the true values ​​using a genetic algorithm. The selection operator of the genetic algorithm is a random acceptance method, which includes the steps of: calculating the sum of fitness; Normalized fitness; Accumulate the normalized fitness values ​​to construct a cumulative probability distribution; Generate a random number between 0 and 1, use binary search to find the interval corresponding to the random number in the cumulative probability distribution, and determine the selected individual; The processing steps in the attention mechanism layer include: The hidden state sequence output by the encoder is obtained. These hidden states capture the dynamic information and features of the input sequence through time step propagation and serve as the input to the attention mechanism layer, i.e., the initial samples. ; A fully connected layer is constructed using a linear transformation function and an activation function, and new samples are obtained through the fully connected layer. ; A new sample will be obtained. The weight vector is obtained by inputting the data into the softmax() function according to the time steps. Weight vector The number of elements in the column vector is equal to the number of input time steps, and each element represents the weight of a time step; Initial sample and weight vector Multiplication yields weighted results Subsequent samples ; The weighted samples The summation is performed according to the time step sequence to obtain the final output of the attention mechanism layer, which is a context vector. Among them, the query vector in the attention mechanism layer For the corresponding dependent variable RSRP, the weight vector during neural network training. With query vector It is dynamically adjusted.

2. The optimization method for predicting uplink throughput based on RSRP according to claim 1, characterized in that, Both the encoder and decoder include multiple LSTM units.

3. An optimization method for predicting uplink throughput based on RSRP according to any one of claims 1-2, characterized in that, The input to the decoder at each step includes a weighted sum of the predicted value from the previous step and the output of the attention mechanism layer. This weighted sum forms a context vector. The predicted value from the previous step and the context vector together form the input vector. The decoder generates the output predicted value for the next time step based on the current hidden state, the input vector, and the memory unit. With the new hidden state, the decoder updates the hidden state and memory unit, and sets the current output prediction value. Input for the next time step.

4. An optimization system based on RSRP for predicting uplink throughput, characterized in that, The system is used to implement the method according to any one of claims 1-3, and the system includes the following modules: The dataset building module is used for data collection and to build a dataset that includes RSRP values ​​and corresponding uplink throughput. The mapping model building module is used to construct the mapping relationship between RSRP and uplink throughput based on the decision tree regression algorithm, and obtain the mapping model. The LSTM model building module is used to perform multi-step prediction of RSRP on a time series based on an improved LSTM model. The improved LSTM model includes an encoder, an attention mechanism layer, a decoder, and a fully connected layer. The encoder is used to process RSRP sequence data and generate corresponding hidden states. The attention mechanism layer is used to calculate the correlation weight between the current hidden state and all hidden states of the encoder at each step of the decoder prediction, and generate a weighted sum based on this, which serves as the input of the decoder at the current step. The decoder is used to combine the weighted sum generated by the attention mechanism and the prediction value of the previous step to perform multi-step prediction. The fully connected layer is used to perform a linear transformation on the output of the decoder and output the RSRP prediction value. The prediction module is used to predict the value of future slot RSRP using the improved LSTM model and to predict uplink throughput using the mapping model.

5. A computer device comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the method according to any one of claims 1-3.

6. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the steps of the method according to any one of claims 1-3.

Images