An adaptive cellular traffic prediction method based on asymmetric fast-slow feature decoupling

Abstract

The application discloses a kind of self-adapting cellular traffic prediction methods based on asymmetric fast and slow feature decoupling, belong to mobile communication network traffic prediction technical field.The application is based on freezing pre-training model main parameter, by constructing fast adaptive module and asynchronous adaptive module, the continuous online adaptation of target domain cellular traffic data is realized.The fast adaptive module utilizes the spatiotemporal structure information in the current observation window for zero-delay update, the asynchronous adaptive module utilizes the long-term global view obtained by the delayed arrival of real values, and the collaborative optimization of the two types of adaptive capabilities is achieved through a dynamic fusion mechanism, thereby improving the prediction accuracy, robustness and generalization ability of the model in complex environments.

Description

Technical Field

[0001] This invention relates to the field of mobile communication network traffic prediction technology, and in particular to an adaptive cellular traffic prediction method based on asymmetric fast and slow feature decoupling. Background Technology

[0002] With the rapid development of 5G and 6G mobile communication networks and the continuous growth of mobile internet services, cellular network traffic exhibits significant spatiotemporal correlation, dynamism, and non-stationarity. Accurately predicting future cellular traffic trends is crucial for network resource scheduling, load balancing, edge caching deployment, and network operation and maintenance optimization. Therefore, cellular traffic prediction has become an important research direction in the field of mobile communications.

[0003] Currently, researchers have proposed cellular traffic prediction methods based on statistical models, machine learning models, and deep learning models. Among them, deep learning methods based on Long Short-Term Memory (LSTM), Graph Neural Networks (GNN), and Transformers can effectively uncover complex spatiotemporal dependencies in cellular traffic and have achieved good results in prediction accuracy. However, when pre-trained models are deployed to new network environments, changes in user behavior patterns, network topology, and traffic load distribution can easily lead to a shift between the target domain data distribution and the source domain data distribution, resulting in a significant decrease in model prediction performance.

[0004] To address these issues, existing research has proposed methods such as online domain adaptation and test-time domain adaptation, which continuously update model parameters to improve the model's adaptability to new environments. However, existing methods generally suffer from the following problems: First, most methods require waiting for the arrival of true labels before updating the model, resulting in an unavoidable adaptation delay and difficulty in quickly responding to sudden changes in traffic. Second, existing methods typically fit both current observation data and historical experience data on the same set of parameters, which can easily lead to conflicting optimization directions and parameter oscillations, thus affecting the stability and predictive performance of the model during long-term adaptation.

[0005] Therefore, there is an urgent need to propose an adaptive cellular traffic prediction method that can achieve zero-latency rapid response, take into account long-term global adaptability, and effectively avoid mutual interference between different optimization objectives. Summary of the Invention

[0006] This invention provides an adaptive cellular traffic prediction method based on asymmetric fast and slow feature decoupling, aiming to solve the problem of decreased prediction performance of pre-trained cellular traffic prediction models in complex dynamic network environments due to continuous changes in data distribution.

[0007] The technical solution adopted in this invention is: an adaptive cellular traffic prediction method based on asymmetric fast-slow feature decoupling, which includes the following steps:

[0008] Step 1: Obtain the cellular network traffic dataset and preprocess it to build a training dataset for traffic prediction;

[0009] Step 2: Set up a cellular traffic prediction model. Its input is historical cellular traffic data for a specified time period, and its prediction target is cellular traffic for multiple future time steps. Train the cellular traffic prediction model using the training dataset obtained in Step 1 to obtain a pre-trained cellular traffic prediction model. ;

[0010] Step 3: Construct a fast adaptive module and an asynchronous adaptive module. The fast adaptive module includes a first input adapter and a first output adapter; the asynchronous adaptive module includes a second input adapter and a second output adapter. The first and second input adapters adopt a low-rank spatiotemporal decomposition linear mapping structure. First, the temporal evolution mapping representation is extracted along the time dimension through a mixed weight matrix, and then a projection mapping is performed along the spatial dimension to obtain the reconstructed input, which is then fed into the model. The first and second output adapters employ a low-rank feedforward bottleneck structure with residual connections, based on the model. The output is used as the input, and the residual calibration value is obtained through a low-rank feedforward bottleneck structure, and then compared with the model. The outputs are summed to output the calibrated prediction result;

[0011] Step 4: Input the target domain cellular traffic data into the fast adaptive module, and update the fast adaptive module through the mask self-supervised reconstruction task and the spatiotemporal consistency constraint task;

[0012] Step 5: Construct a historical experience replay queue, update the asynchronous adaptive module using delayed real traffic data, and achieve global vision calibration through a similarity historical experience sampling mechanism;

[0013] Step 6: Dynamically calculate the fusion weight based on the adaptation maturity of the asynchronous adaptive module and the mutation rate of the current traffic data, and fuse the prediction results output by the fast adaptive module and the asynchronous adaptive module to obtain the final cellular traffic prediction result.

[0014] Furthermore, in step 1, the preprocessing includes:

[0015] The raw cellular traffic data in the cellular network traffic dataset is arranged in chronological order, and a continuous time series is constructed.

[0016] After normalizing the raw cellular traffic data, supervised learning samples are constructed using a sliding time window approach. The time length of each sample is L, and the cellular traffic data at the H time steps following the current window are used as the predicted labels for the samples. The sample data is then combined... spacetime matrix That is, the historical cellular traffic matrix, in which, This represents the number of spatial nodes.

[0017] Furthermore, in step 2, the pre-trained cellular traffic prediction model is a pre-trained spatiotemporal prediction model, whose input is the historical cellular traffic matrix. The output is a future cellular traffic prediction matrix. ,in, Indicates the length of the historical observation window. Indicates the number of spatial nodes. This indicates the prediction step size. In the subsequent target domain adaptation phase, the main parameters of the pre-trained cellular traffic prediction model remain frozen, and target domain calibration is completed only through the adaptive modules at the input and output ends.

[0018] Furthermore, in step 2, when training the cellular traffic prediction model using the training dataset obtained in step 1, the future cellular traffic prediction matrix output by the cellular traffic prediction model is used. The error between the predicted label and the actual value is used as the optimization objective to optimize the model parameters of the cellular traffic prediction model. Iterative updates will be performed.

[0019] Furthermore, step 4 specifically includes:

[0020] Obtain the historical cellular traffic matrix of the target domain Its dimensions are consistent with the input data of the cellular traffic prediction model;

[0021] Based on a preset mask ratio, generate and Binary spatiotemporal mask matrix of the same dimension ;

[0022] Based on spatiotemporal mask matrix Generate damaged input ,in, Indicate the Hadamard accumulation; and input the damaged data. Input the first input adapter to obtain the reconstructed input. ;

[0023] right Apply random perturbation to obtain enhanced input ,in Represents the perturbation matrix. It is a matrix of all 1s;

[0024] Reconstruct the input Input pre-trained cellular traffic prediction model The first baseline prediction result is obtained, and then the first baseline prediction result is input into the first output adapter to obtain the first residual calibration value. Finally, the first baseline prediction result is added to the first residual calibration value to obtain the first prediction result. ;

[0025] Enhance input Input pre-trained cellular traffic prediction model The second benchmark prediction result is obtained, and then input into the first output adapter to obtain the second residual calibration value. The second benchmark prediction result is then added to the second residual calibration value to obtain the second prediction result. ;

[0026] Calculate the self-supervised reconstruction loss: ;

[0027] Calculate the loss due to spatiotemporal consistency constraints: ;in, This indicates that the gradient operation has been stopped.

[0028] Based on self-supervised reconstruction loss Loss due to spatiotemporal consistency constraints The weighted fusion loss is used to update the parameters of the fast adaptive module;

[0029] When the preset update convergence condition is met, the trained fast adaptive module is obtained; and based on the sum of the first benchmark prediction result and the residual calibration amount generated by the first output adapter, the fast prediction result is obtained. (i.e., the prediction results of the adaptive module).

[0030] Furthermore, in step 4, the spatiotemporal mask matrix is ​​determined using either random independent masking or continuous time block masking. The position of the element that is being obscured.

[0031] Furthermore, in step 4, the self-supervised reconstruction loss... The weighting coefficient is Spatiotemporal consistency constraint loss The weighting coefficient is ,and .

[0032] Furthermore, in step 4, the perturbation matrix Each element is independently and identically distributed from a uniform distribution Mid-sampling, This is a hyperparameter.

[0033] Furthermore, step 5 specifically includes:

[0034] Maintain a fixed capacity First-in-first-out experience replay queue Used to store historical observation window input and their corresponding prediction labels (True value), where, and Composing historical moments spacetime experience tuple ;

[0035] When a historical moment of When the delay occurs, match it with the corresponding input. Encapsulated as a spatiotemporal empirical tuple And store it in the experience replay queue. middle;

[0036] Historical cellular traffic matrix of the current target domain Using the model The output calculation is compared with the experience replay queue. Input in the historical observation window Cosine similarity between : ;

[0037] Similarity After normalization using the Softmax function, the sampling weights of each spatiotemporal empirical tuple are obtained. ;

[0038] Calculate each Reliability penalty factor ;

[0039] Experience replay queue Each of them The corresponding reconstructed input is obtained through the second input adapter and then processed by the model. The second output adapter obtains the calibrated prediction results. ;

[0040] Based on sampling weight and reliability penalty factor Calculate the weighted mean square error loss :

[0041]

[0042] in, For the corresponding The predicted labels, where K is the experience replay queue. The number of stored spatiotemporal experience tuples;

[0043] Based on weighted mean square error loss The parameters of the asynchronous adaptive module are updated to enable the model to adapt to long-term global patterns while avoiding overfitting to local variations. When the preset update convergence condition is met, the slow prediction result is obtained based on the output of the second output adapter of the asynchronous adaptive module. (i.e., the prediction results of the asynchronous adaptive module) to provide global perspective information for subsequent dynamic fusion.

[0044] Furthermore, the adaptation maturity in step 6 is:

[0045]

[0046] in, The cumulative number of asynchronous updates completed when updating the asynchronous adaptive module.

[0047] Furthermore, in step 6, the fusion weights are set as follows:

[0048]

[0049] in, Indicates the degree of adaptation maturity. This represents the rate of variation.

[0050] Furthermore, the rate of variation Set to:

[0051] Calculate the absolute first-order difference between adjacent time steps:

[0052]

[0053] in, Indicates the current observation window The Middle Each time step corresponds to Cellular flow vector of each spatial node Indicates the current observation window The Middle Each time step corresponds to Cellular flow vector of each spatial node This represents element-wise absolute value operation;

[0054] Based on the average vector of flow data within the current observation window Calculate the rate of variation :

[0055]

[0056] in, This is a preset constant used to prevent the denominator from being zero.

[0057] Furthermore, in step 6, the final cellular traffic prediction result is:

[0058]

[0059] in, It represents the Hadamardi (or Hadama) stack.

[0060] The technical solution provided by this invention brings at least the following beneficial effects:

[0061] This invention, based on freezing the main parameters of the pre-trained model, achieves continuous online adaptation to cellular traffic data in the target domain by constructing a fast adaptive module and an asynchronous adaptive module. The fast adaptive module utilizes the spatiotemporal structure information within the current observation window for zero-latency updates, while the asynchronous adaptive module uses delayed-arriving ground truth values ​​to obtain a long-term global perspective. A dynamic fusion mechanism is used to synergistically optimize these two types of adaptation capabilities, thereby improving the model's prediction accuracy, robustness, and generalization ability in complex environments. Attached Figure Description

[0062] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] Figure 1 This is a schematic diagram of the adaptive cellular traffic prediction method based on asymmetric fast and slow feature decoupling, which is the subject of this invention.

[0064] Figure 2 This is a schematic diagram of the pre-trained cellular traffic prediction model structure in an optional embodiment of the present invention.

[0065] Figure 3 This is a schematic diagram of the input adapter structure involved in the present invention.

[0066] Figure 4 This is a schematic diagram of the output adapter involved in the present invention. Detailed Implementation

[0067] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be described in detail and completely below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Generally, the components of the embodiments of the present invention described and shown in the accompanying drawings can be arranged and designed using different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of the present invention.

[0068] This invention provides an adaptive cellular traffic prediction method based on asymmetric fast-slow feature decoupling. By freezing the main parameters of the pre-trained model, a fast adaptive module and an asynchronous adaptive module are constructed to achieve continuous online adaptation to cellular traffic data in the target domain. The final prediction result is obtained through a dynamic fusion mechanism, thereby improving the prediction accuracy and robustness of the model in complex environments. In this invention, asymmetry is reflected in the following: the fast adaptive module uses an unsupervised self-supervised task to update immediately step-by-step, while the asynchronous adaptive module uses a supervised task to update periodically in the background, relying on delayed real values. The parameters of the two modules are independent of each other, thus achieving asymmetric decoupling in terms of time scale and optimization objective.

[0069] In one embodiment, the adaptive cellular traffic prediction method based on asymmetric fast-slow feature decoupling provided in this embodiment includes the following steps:

[0070] Step 1: Obtain the cellular network traffic dataset and preprocess it to build a training dataset for traffic prediction;

[0071] Step 2: Obtain and freeze the pre-trained cellular traffic prediction model, and use the training dataset obtained in Step 1 to train the model and obtain the pre-trained model.

[0072] Step 3: Based on the frozen model parameters of the pre-trained model, construct the fast adaptive module and the asynchronous adaptive module respectively, and establish a dynamic fusion mechanism;

[0073] Step 4: Input the target domain cellular traffic data into the fast adaptive module, and achieve rapid online adaptation of the model through the mask self-supervised reconstruction task and the spatiotemporal consistency constraint task;

[0074] Step 5: Construct a historical experience replay queue, update the asynchronous adaptive module using delayed real traffic data, and achieve global vision calibration through a similarity historical experience sampling mechanism;

[0075] Step 6: Dynamically calculate the fusion weight based on the adaptation maturity of the asynchronous adaptive module and the variation rate of the current traffic data, and fuse the prediction results of the fast adaptive module and the asynchronous adaptive module to obtain the final cellular traffic prediction result.

[0076] Step 1 includes the following steps:

[0077] Step 1-1: Obtain the Milan dataset and C2TM dataset of cellular network traffic, that is, obtain the raw cellular traffic data; in this embodiment, two datasets are selected to ensure the diversity of sample data, thereby improving the reliability of the results;

[0078] Steps 1-2: Arrange the raw cellular traffic data in chronological order and construct a continuous time series;

[0079] Steps 1-3: Supervised learning samples are constructed using a sliding time window approach, where a historical observation window of length L is used as the model input, and cellular traffic over the next H time steps is used as the prediction target; where the time length L and time steps H are preset values ​​that can be set based on the actual application scenario.

[0080] Steps 1-4 involve normalizing the cellular traffic data to map traffic data from different base stations to a unified numerical range; that is, extracting the specified service traffic fields from the raw cellular traffic data and then organizing them by hour and by spatial node into a single data structure. The spatiotemporal matrix (i.e., the historical cellular traffic matrix) is then processed, missing and outlier data are filled in, and finally, Min-Max normalization is applied to 0 to 1. The dimensions of the spatiotemporal matrix are preset values;

[0081] Steps 1-5 divide the processed data into training and test sets according to a preset ratio.

[0082] Step 2 includes the following steps:

[0083] Step 2-1: Construct a pre-trained cellular traffic prediction model .

[0084] The pre-trained cellular traffic prediction model This tool is used to predict cellular traffic over a future time period based on historical cellular traffic observation windows. Its input is a historical cellular traffic matrix. The output is a future cellular traffic prediction matrix. .in, Indicates the length of the historical observation window. Indicates the number of spatial nodes. Indicates the prediction step size. This represents all the original learnable parameters in the pre-trained cellular traffic prediction model.

[0085] In one alternative embodiment, such as Figure 2 As shown, the pre-trained cellular traffic prediction model A spatiotemporal prediction structure based on dual-stream divide-and-conquer can be adopted, which includes a high- and low-frequency information separation module, a dynamic spatial correlation modeling module, a dual-stream spatiotemporal feature aggregation module, a prediction module, and a prediction result fusion module. The high- and low-frequency information separation module decomposes the input cellular traffic observation window into a low-frequency stationary component and a high-frequency fluctuating component; the dynamic spatial correlation modeling module characterizes the correlation between multiple spatial nodes under different frequency components; the dual-stream spatiotemporal feature aggregation module extracts spatiotemporal features from the low-frequency / high-frequency components; and the prediction module, based on these spatiotemporal features as input, outputs low-frequency / high-frequency prediction results; i.e., a pre-trained cellular traffic prediction model. The system comprises two branches: a high-frequency branch and a low-frequency branch. The low-frequency stationary component, decomposed by the high-low frequency information separation module, is fed into the low-frequency branch, while the high-frequency fluctuating component is fed into the high-frequency branch. The high-frequency and low-frequency branches have identical structures, sequentially including: a dynamic spatial correlation modeling module, a dual-stream spatiotemporal feature aggregation module, and a prediction module. The dynamic spatial correlation modeling module constructs a dynamic hypergraph based on an attention mechanism, generates a corresponding hypergraph correlation matrix, and performs Laplacian normalization on the hypergraph correlation matrix to obtain a normalized hypergraph correlation matrix. The dual-stream spatiotemporal feature aggregation module extracts spatiotemporal features from the normalized hypergraph correlation matrix and feeds them into the prediction module. The prediction module outputs corresponding sub-prediction results to obtain high / low-frequency sub-prediction results. The prediction result fusion module performs weighted fusion of the prediction results from different branches to obtain a baseline prediction result.

[0086] It should be noted that this dual-stream divide-and-conquer structure is only an optional implementation of the pre-trained cellular traffic prediction model and does not constitute a limitation on the scope of protection of the adaptive method of this invention. This invention is also applicable to other pre-trained cellular traffic spatiotemporal prediction models.

[0087] Step 2-2: Input the training set obtained in Step 1 into the pre-trained cellular traffic prediction model. Conduct training.

[0088] During training, historical cellular traffic observation windows were used. As input to the model, corresponding to real cellular traffic in future time periods. As a supervisory signal, it enables the pre-trained cellular traffic prediction model to learn the spatiotemporal evolution patterns in the source domain cellular traffic data.

[0089] Steps 2-3 involve using the error between the predicted results and the true values ​​as the optimization objective to adjust the model parameters. Iterative updates will be performed.

[0090] Specifically, based on the prediction results output by the pre-trained cellular traffic prediction model Compared with the true value The difference between them is used to construct the prediction loss function, and the model parameters are updated through the backpropagation algorithm. This continues until the model can reliably predict future cellular traffic.

[0091] Steps 2-4: Once the model's performance on the validation set reaches the preset convergence condition, save the trained model parameters to obtain the pre-trained cellular traffic prediction model. .

[0092] In this step, the preset convergence conditions include the validation set prediction error no longer decreasing significantly, reaching the preset number of training rounds, or the model's evaluation metric on the validation set reaching the preset threshold.

[0093] Steps 2-5: During subsequent time-domain adaptation testing, freeze the pre-trained cellular traffic prediction model. All raw parameters .

[0094] In all subsequent steps of this invention, a pre-trained cellular traffic prediction model is used. parameters It remains frozen and does not participate in any gradient updates. Subsequent target domain adaptation processes only update the lightweight adapter parameters attached to the input and output of the pre-trained cellular traffic prediction model.

[0095] Step 3 includes the following steps:

[0096] Step 3-1: Construct an input adapter at the input end of the frozen pre-trained model. ;

[0097] Among them, the input adapter Employing a low-rank spatiotemporal decomposition linear mapping structure, such as Figure 3 As shown, the temporal evolution trend is first extracted along the time dimension using a hybrid weight matrix, and then node feature interaction is achieved along the spatial dimension using a low-rank projection matrix.

[0098] Wherein, the hybrid weight matrix is ​​the input adapter. The learnable parameter matrix in the matrix is ​​denoted as . This matrix is ​​used to linearly weight and combine cellular traffic characteristics at different time steps within a historical observation window, thereby extracting the temporal evolution trend within the current observation window. Specifically, the learnable parameter matrix can be used... The initial value can be set to The identity matrix.

[0099] For input When realizing node feature interaction along the spatial dimension through a low-rank projection matrix, the first step is to use a time-mixed weight matrix. Extract the temporal evolution trend within the observation window along the time dimension; then, based on the spatial low-rank dimensionality reduction projection matrix... Perform dimensionality reduction projection to obtain a dimensionless projection. The low-dimensional space representation, in which, It is the preset value, and Then, the low-dimensional spatial representation is projected onto a low-rank spatially upgraded matrix. Perform up-dimensional projection and feature interaction along the spatial node dimensions:

[0100]

[0101] in, .because It is equivalent to a low-rank spatial node mixing matrix. The representation of each spatial node in the output is determined by the flow characteristics of multiple input nodes. Therefore, it can realize information interaction between spatial nodes with a low number of parameters.

[0102] Step 3-2: Build an output adapter at the output of the frozen pre-trained model. ;

[0103] Among them, the output adapter Employing a low-rank feedforward bottleneck structure with residual connections, such as... Figure 4 As shown, it consists of a dimension-reducing linear layer, a nonlinear activation function, and a dimension-upgrading linear layer in sequence.

[0104] The input after processing by the input adapter Feed into the frozen pre-trained cellular traffic prediction model The baseline prediction results are obtained as follows:

[0105]

[0106] in, Indicates the prediction step size. Indicates the number of nodes in the space.

[0107] The output adapter first uses a dimensionality reduction linear layer to convert the values ​​at each prediction time step. 3D space node prediction vector mapping to Bottleneck space:

[0108]

[0109] in, , represents the weight matrix of the linear reduction layer of the output adapter, and , It is a non-linear activation function. , representing the bias vector of the dimension reduction linear layer. Subsequently, the original prediction dimension is restored through the dimension increase linear layer to obtain the residual calibration value:

[0110]

[0111] in, . , represents the weight matrix of the linear layer of the output adapter. , representing the bias vector of the upgraded linear layer. Finally, the baseline prediction result is added to the residual calibration value to obtain the calibrated prediction result:

[0112]

[0113] in, .

[0114] Among them, the input adapter With output adapter This is the fast adaptive module of the present invention.

[0115] Step 3-3: Construct an asynchronous adaptive module independent of the fast adaptive module parameters. The asynchronous adaptive module includes a slow input adapter. and slow output adapter The parameters of the two modules are independent to isolate the optimization objectives of the fast and slow modules and avoid gradient conflicts. The input adapter... With output adapter This is the asynchronous adaptive module of the present invention.

[0116] It should be noted that the input and output adapters in both the fast and asynchronous adaptive modules can employ the same lightweight network structure, but their parameter sets are independent and do not share parameters. The fast adaptive module constructs a masked self-supervised reconstruction loss and a spatiotemporal consistency loss based on the observation window data directly available at the current time step, and updates the fast adapter parameters in real-time at the current time step. The asynchronous adaptive module, on the other hand, updates the asynchronous adapter parameters asynchronously in the background after the historical ground truth values ​​arrive with a delay, based on a historical experience queue, similarity addressing, and a reliability penalty mechanism. Therefore, the difference in speed between the two is not due to differences in the adapter network structure, but rather to a combination of differences in data timeliness, the source of the supervision signal, the timing of parameter updates, and the execution method.

[0117] During the target domain testing time adaptation process, only the historical cellular traffic matrix can be obtained at the current moment. The corresponding future real cellular traffic at that moment This information can only be obtained after the actual occurrence of a future time step. Therefore, in this embodiment, historical samples that have already completed predictions and subsequently obtained actual results are denoted as... ,in This indicates the delay in reaching the actual value corresponding to the historical prediction window. The fast adaptive module does not depend on... Based on the current It performs self-supervised updates immediately, making it fast. Asynchronous adaptive modules, on the other hand, must update the actual value immediately. Upon arrival, it undergoes supervised calibration in the background based on historical experience samples, which is why it is slow.

[0118] Steps 3-4: Construct a first-in-first-out experience replay queue with a fixed capacity of m. When the queue is full, the oldest sample is automatically discarded.

[0119] Steps 3-5: Construct a dynamic fusion mechanism based on adaptation maturity and mutation rate.

[0120] Step 4 includes the following steps:

[0121] Step 4-1: Obtain the current observation window input at time step t. That is, the historical cellular traffic matrix;

[0122] Step 4-2, generate and Binary spatiotemporal mask matrix of the same dimension The mask ratio can be preset to γ ​​(e.g., 0.3), and the position of the occluded element can be determined by random independent masking or continuous time block masking.

[0123] Step 4-3: Apply the spatiotemporal mask matrix to the original input to obtain the damaged input: ,in, It represents the Hadamardi (or Hadama) stack;

[0124] Step 4-4: The damaged input is sent to the input adapter for reconstruction to obtain the reconstructed input. ;

[0125] Steps 4-5: Calculate the self-supervised reconstruction loss only for the masked region.

[0126]

[0127] in, The total number of mask elements is used to normalize the loss.

[0128] Steps 4-6, for By applying random perturbations, we obtain enhanced inputs: ,in This represents a perturbation matrix, where each element is independently and identically distributed from a uniform distribution. Mid-sampling, This is a hyperparameter that defines the maximum permissible relative fluctuation tolerance. It is a matrix of all 1s.

[0129] Steps 4-7 involve feeding the reconstructed input and augmented input into the pre-trained model and output adapter, respectively, to obtain the prediction results. and :

[0130]

[0131]

[0132] Steps 4-8: Calculate the spatiotemporal consistency constraint loss:

[0133]

[0134] in, This indicates that gradient operations are stopped. The purpose is to prevent the network from degenerating into an identity mapping and to avoid all adapter parameters being zero.

[0135] Steps 4-9: Jointly optimize the self-supervised reconstruction loss and the spatiotemporal consistency constraint loss:

[0136]

[0137] in, The hyperparameters that control the contribution of the two loss functions are usually set to [0.3, 0.7].

[0138] Steps 4-10: Update the parameters of the fast adaptive module according to the loss function, and... As a rapid prediction result .

[0139] Step 5 includes the following steps:

[0140] Step 5-1, maintain a fixed capacity of First-in-first-out experience replay queue It is used to store historical observation window inputs and their corresponding true values, forming spatiotemporal empirical tuples. And store it in the memory bank:

[0141]

[0142] The asynchronous adaptive module checks for newly arrived real values ​​after each time step. If so, it performs a background update; otherwise, it skips the update.

[0143] Step 5-2, when the historical moment The true value When the arrival is delayed, compare it with the corresponding historical input. Encapsulated as a spatiotemporal empirical tuple And store it in the experience replay queue. If the queue is full, the oldest sample is replaced using a first-in, first-out (FIFO) strategy.

[0144] Step 5-3, input the current observation window With experience replay queue Historical input in China Deep semantic feature extraction is performed, and cosine similarity is calculated to quantify the reference value of each historical sample to the current moment:

[0145]

[0146] in, This is a pre-trained model.

[0147] Step 5-4, similarity After normalization using the Softmax function, the sampling weights of each historical empirical tuple are obtained. :

[0148]

[0149] in, The smoothing hyperparameter is used to adjust the smoothness of the sampling weight distribution, where K is the empirical replay queue. Historical input in China The quantity.

[0150] Step 5-5: Calculate historical samples Overall variance across all time steps and all base stations Define the reliability penalty factor :

[0151]

[0152] The larger the variance, the higher the degree of anomaly in the historical samples, and the lower their contribution to asynchronous adaptive updates.

[0153] Steps 5-6: For each historical sample in the experience replay queue First, use the slow input adapter. Perform feature mapping, and then use a pre-trained model. and slow output adapter The predicted output is obtained:

[0154]

[0155] Steps 5-7, based on sampling weights and reliability factor Calculate the weighted mean square error loss :

[0156]

[0157] Steps 5-8, minimize For slow input adapters and slow output adapter The parameters are updated to adapt the model to long-term global patterns while avoiding overfitting to local changes.

[0158] Steps 5-9: The asynchronous adaptive module outputs the slow prediction result. (Right now The final output provides global perspective information for subsequent dynamic fusion.

[0159] Step 6 includes the following steps:

[0160] Step 6-1, Define Adaptation Maturity Indicators This is used to measure the degree to which the slow adaptive module has learned the target domain environment. Let the current time be... The slow adaptive module has completed an average of [number] asynchronous updates. Then, the maturity of adaptation is defined as:

[0161]

[0162] in, This value gradually increases with the number of asynchronous updates. When the model is first deployed to the target environment... Approaching 0; as the model continuously absorbs delayed true values ​​and completes updates, It gradually approaches 1.

[0163] Step 6-2: Calculate the variation rate index of the data in the current observation window. This is used to describe the degree of traffic variation at each base station. First, the absolute first-order difference between adjacent time steps is calculated:

[0164]

[0165] in, Indicates the current observation window The Middle Each time step corresponds to Cellular flow vector of each spatial node Indicates the current observation window The Middle Each time step corresponds to Cellular flow vector of each spatial node This represents element-wise absolute value operation;

[0166] Then, the mean of the data within the observation window is used for normalization to obtain the variation rate:

[0167]

[0168] in, This represents the average value vector of the flow data within the current observation window. To prevent extremely small constants with a denominator of zero.

[0169] Step 6-3, based on adaptation maturity With the rate of variation Calculate the dynamic fusion weights of the slow adaptive module :

[0170]

[0171] Both exponentiation and multiplication are element-wise operations, giving each base station independent fusion weights. When the target environment is relatively stable... Smaller, at this time When traffic is relatively large, the model tends to use the global perspective information provided by the slow adaptive module; when sudden changes in traffic occur, It grows rapidly, at this time As the number of cells decreases, the model tends to use the predictions from the fast adaptive module.

[0172] Step 6-4: Utilize dynamic fusion weights to evaluate the rapid prediction results. and slow prediction results By performing weighted fusion, we obtain the final prediction result for the current moment:

[0173]

[0174] in, It represents the Hadamardi (or Hadama) stack.

[0175] In step 6-5, when the model is in the early stage of target domain adaptation, the slow adaptive module has not yet accumulated enough historical experience and its adaptation maturity is low. Therefore, the final prediction result mainly depends on the output of the fast adaptive module. As the model continues to run, the slow adaptive module gradually obtains more real value supervision information, and its prediction result gradually increases in proportion to the final decision.

[0176] Step 6-6: When a sudden event occurs in the cellular network, user behavior changes drastically, or the network structure is adjusted, the variation rate of the current observation window increases, and the dynamic fusion weight automatically decreases, thereby increasing the decision-making proportion of the fast adaptive module and enabling the model to respond quickly to new data distribution changes.

[0177] Steps 6-7: Through the above dynamic fusion mechanism, the fast adaptive module and the slow adaptive module work together. While ensuring the model's real-time response capability, the model introduces guidance from long-term global evolution laws to improve the prediction accuracy and robustness of the model in dynamic network environments.

[0178] Steps 6-8: Use the final prediction result obtained in step 6-4 as the future... The predicted cellular traffic at each time step is output, and the data processing flow for the next time step is entered. Steps 4 to 6 are executed continuously to achieve online adaptive updating of the model in the target domain environment.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the foregoing embodiments have described the present invention in detail, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the embodiments of the present invention.

[0180] The above descriptions are merely some embodiments of the present invention. Those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. An adaptive cellular traffic prediction method based on asymmetric fast-slow feature decoupling, characterized in that, Includes the following steps: Step 1: Obtain the cellular network traffic dataset and preprocess it to build a training dataset for traffic prediction; Step 2: Set up a cellular traffic prediction model. Its input is historical cellular traffic data for a specified time period, and the prediction target is cellular traffic for multiple future time steps. The cellular traffic prediction model is then trained using the training dataset obtained in step 1 to obtain a pre-trained cellular traffic prediction model. ; Step 3: Construct a fast adaptive module and an asynchronous adaptive module. The fast adaptive module includes a first input adapter and a first output adapter; the asynchronous adaptive module includes a second input adapter and a second output adapter. The first and second input adapters adopt a low-rank spatiotemporal decomposition linear mapping structure. First, the temporal evolution mapping representation is extracted along the time dimension through a mixed weight matrix, and then a projection mapping is performed along the spatial dimension to obtain the reconstructed input, which is then fed into the model. The first and second output adapters employ a low-rank feedforward bottleneck structure with residual connections, based on the model. The output is used as the input, and the residual calibration value is obtained through a low-rank feedforward bottleneck structure, and then compared with the model. The outputs are summed to output the calibrated prediction result; Step 4: Input the target domain cellular traffic data into the fast adaptive module, and update the fast adaptive module through the mask self-supervised reconstruction task and the spatiotemporal consistency constraint task; Step 5: Build a historical experience replay queue and update the asynchronous adaptive module using delayed real traffic data; Step 6: Dynamically calculate the fusion weight based on the adaptation maturity of the asynchronous adaptive module and the mutation rate of the current traffic data, and fuse the prediction results output by the fast adaptive module and the asynchronous adaptive module to obtain the final cellular traffic prediction result.

2. The method as described in claim 1, characterized in that, In step 1, preprocessing includes: The raw cellular traffic data in the cellular network traffic dataset is arranged in chronological order, and a continuous time series is constructed. After normalizing the raw cellular traffic data, supervised learning samples are constructed using a sliding time window approach. The time length of each sample is L, and the cellular traffic data at the H time steps following the current window are used as the predicted labels for the samples. The sample data is then combined... Historical cellular traffic matrix ,in, This represents the number of spatial nodes.

3. The method as described in claim 1, characterized in that, In step 2, the pre-trained cellular traffic prediction model The input is the historical cellular traffic matrix. The output is a future cellular traffic prediction matrix. ,in, Indicates the length of the historical observation window. Indicates the number of spatial nodes. Indicates the prediction step size; In step 4, the pre-trained cellular traffic prediction model is frozen. Target domain calibration is performed only through the input adapter at its input and the output adapter at its output.

4. The method as described in claim 1, characterized in that, Step 4 specifically includes: Obtain the historical cellular traffic matrix of the target domain Its dimensions are consistent with the input data of the cellular traffic prediction model; Based on a preset mask ratio, generate and Binary spatiotemporal mask matrix of the same dimension ; Based on spatiotemporal mask matrix Generate damaged input ,in, Indicate the Hadamard accumulation; and input the damaged data. Input the first input adapter to obtain the reconstructed input. ; right Apply random perturbation to obtain enhanced input ,in Represents the perturbation matrix. It is a matrix of all 1s; Reconstruct the input Input pre-trained cellular traffic prediction model The first baseline prediction result is obtained, and then the first baseline prediction result is input into the first output adapter to obtain the first residual calibration value. Finally, the first baseline prediction result is added to the first residual calibration value to obtain the first prediction result. ; Enhance input Input pre-trained cellular traffic prediction model The second benchmark prediction result is obtained, and then input into the first output adapter to obtain the second residual calibration value. The second benchmark prediction result is then added to the second residual calibration value to obtain the second prediction result. ; Calculate the self-supervised reconstruction loss: ; Calculate the loss due to spatiotemporal consistency constraints: ;in, This indicates that the gradient operation has been stopped. Based on self-supervised reconstruction loss Loss due to spatiotemporal consistency constraints The weighted fusion loss is used to update the parameters of the fast adaptive module; When the preset update convergence condition is met, the trained fast adaptive module is obtained; and based on the sum of the first benchmark prediction result and the residual calibration amount generated by the first output adapter, the fast prediction result is obtained. .

5. The method as described in claim 4, characterized in that, Self-supervised reconstruction loss The weighting coefficient is Spatiotemporal consistency constraint loss The weighting coefficient is .

6. The method as described in claim 4, characterized in that, In step 4, the perturbation matrix Each element is independently and identically distributed from a uniform distribution Mid-sampling, This is a hyperparameter.

7. The method as described in claim 1, characterized in that, Step 5 specifically includes: Maintain a fixed capacity First-in-first-out experience replay queue Used to store historical observation window input and their corresponding prediction labels (True value), where, and Composing historical moments spacetime experience tuple ; When a historical moment of When the delay occurs, match it with the corresponding input. Encapsulated as a spatiotemporal empirical tuple And store it in the experience replay queue. middle; Historical cellular traffic matrix of the current target domain Using the model The output calculation is compared with the experience replay queue. Input in the historical observation window Cosine similarity between : ; Similarity After normalization using the Softmax function, the sampling weights of each spatiotemporal empirical tuple are obtained. ; Calculate each Reliability penalty factor ; For experience replay queue Each of them The corresponding reconstructed input is obtained through the second input adapter and then processed by the model. The second output adapter obtains the calibrated prediction results. ; Based on sampling weight and reliability penalty factor Calculate the weighted mean square error loss : in, For the corresponding The predicted labels, where K is the experience replay queue. The number of stored spatiotemporal experience tuples; Based on weighted mean square error loss The parameters of the asynchronous adaptive module are updated, and the pre-trained cellular traffic prediction model is activated when the preset update convergence condition is met. The slow prediction result is obtained by summing the baseline prediction result with the residual calibration value generated by the second output adapter based on the asynchronous adaptive module. .

8. The method as described in claim 1, characterized in that, In step 6, the adaptation maturity level is: in, The cumulative number of asynchronous updates completed when updating the asynchronous adaptive module.

9. The method as described in claim 1, characterized in that, In step 6, the fusion weights are set as follows: in, Indicates the degree of adaptation maturity. Indicates the rate of variation; Variation rate Set to: Calculate the absolute first-order difference between adjacent time steps: in, Indicates the current observation window The Middle Each time step corresponds to Cellular flow vector of each spatial node Indicates the current observation window The Middle Each time step corresponds to Cellular flow vector of each spatial node This represents element-wise absolute value operation; Based on the average vector of flow data within the current observation window Calculate the rate of variation : in, This is a preset constant used to prevent the denominator from being zero.

10. The method as described in claim 9, characterized in that, In step 6, the final cellular traffic prediction result is: in, , These are the prediction results output by the fast adaptive module and the asynchronous adaptive module, respectively.

Images