A 5G base station traffic prediction method based on a decomposable periodic pattern network

Abstract

The application discloses a 5G base station traffic prediction method based on a decomposable periodic pattern network, and relates to the technical field of 5G base station traffic prediction. The method comprises the following steps: an intelligent network management center collects 5G transmission related feature data in a 5G base station through a data acquisition module, pre-processes the collected feature data, and divides samples into a training set, a verification set and a test set in time sequence; the training set and the verification set are input into a 5G base station traffic prediction model based on a decomposable periodic pattern network for training, the 5G base station traffic prediction model based on the decomposable periodic pattern network comprises a periodic pattern modeling module, a component decomposition module and an xLSTM module; after multiple training iterations, the weight parameters that perform best on the test set are finally selected for prediction. The application fully utilizes the periodic time characteristics of the traffic data of each base station, improves the accuracy of 5G base station traffic prediction, provides data reference for base station management and network optimization, and reduces the labor and time cost.

Description

Technical Field

[0001] This invention relates to the field of 5G base station traffic prediction technology, and specifically to a 5G base station traffic prediction method based on a decomposable periodic pattern network. Background Technology

[0002] With the rapid popularization and large-scale deployment of 5G (fifth-generation mobile communication technology) networks, the number and types of 5G base stations are also constantly increasing. 5G technology, with its significant advantages such as ultra-high transmission speed, large capacity, and ultra-low latency, has become an important component of modern communication networks, widely used in many fields such as smart manufacturing, smart cities, autonomous driving, and telemedicine. However, as the coverage and application scenarios of 5G networks continue to expand, the energy consumption of 5G base stations is becoming increasingly prominent. Compared to 4G networks, 5G base stations consume significantly more energy in data processing and transmission, with overall energy consumption more than three times that of 4G base stations. This high energy consumption not only increases the operating costs of operators but also poses a significant challenge to global carbon emission control. According to the latest standards of the International Telecommunication Union (ITU), in order to comply with the Paris Agreement, the information and communication technology industry needs to reduce greenhouse gas emissions by 45% between 2020 and 2030. Therefore, reducing the energy consumption of 5G base stations has become a critical issue that the global telecommunications industry urgently needs to address.

[0003] In the architecture of 5G networks, base stations typically cover multiple logical cells, and the traffic load of each logical cell directly affects the base station's energy consumption performance. However, due to the complexity and high dynamism of 5G networks, the traffic load of each logical cell exhibits strong temporal and spatial dependencies. To achieve more efficient energy management, base stations need to be able to accurately predict the future traffic load of each logical cell. By predicting future traffic loads, base stations can dynamically adjust power output during off-peak hours, and even shut down some equipment without affecting service quality, thereby significantly reducing energy consumption.

[0004] Current research primarily employs traditional time series forecasting models such as autoregressive (AR) models, moving average (MA) models, and long short-term memory (LSTM) networks for 5G base station traffic load prediction. However, these methods have limitations in handling the complex time dependencies in 5G networks, leading to insufficient prediction accuracy. Furthermore, as 5G networks evolve towards edge computing and distributed architectures, higher demands are placed on the computational complexity of prediction models. High-computational-complexity models such as Transformer and GNN are difficult to deploy efficiently on resource-constrained edge devices, further limiting their feasibility in practical applications. Summary of the Invention

[0005] The present invention aims to provide a 5G base station traffic prediction method based on decomposable periodic pattern networks, in order to solve the limitations of existing 5G base station traffic prediction methods in handling complex time dependencies, which leads to insufficient prediction accuracy, excessive computational complexity, and unfavorable deployment of edge devices.

[0006] This invention adopts the following technical solution: a 5G base station traffic prediction method based on a decomposable periodic pattern network, comprising the following steps:

[0007] Step S1: The intelligent network management center collects characteristic data related to 5G transmission in 5G base stations through the data acquisition module and sets the coverage area of ​​the intelligent network management center;

[0008] Step S2: Organize the collected feature data according to the time level, and resample the feature data at hourly and daily collection intervals to obtain the resampled dataset;

[0009] Step S3: Clean the resampled dataset by filling all empty values ​​with zero values. Strictly set the data for each day to 24 time nodes in hours. If a time node is missing during the collection process, fill it with the average of the historical data of that base station. Divide the resampled dataset into training set, validation set and test set according to time order.

[0010] Step S4: Design a 5G base station traffic prediction model based on a decomposable periodic pattern network; the 5G base station traffic prediction model based on a decomposable periodic pattern network includes a periodic pattern modeling module, a component decomposition module, and an xLSTM module; the periodic pattern modeling module includes removing periodic pattern units and restoring periodic pattern units, the periodic pattern modeling module is used to extract periodic features from the data, the component decomposition module decomposes the data into trend components and seasonal components, and the xLSTM module is used to capture long-term dependencies;

[0011] Step S5: Input the training set data into the 5G base station traffic prediction model of the decomposable periodic pattern network for training. After iterative optimization, select the model parameters with the best performance on the test set for 5G base station traffic prediction. Deploy the trained 5G base station traffic prediction model of the decomposable periodic pattern network to the intelligent network management center. In subsequent operation, input the historical 96 time steps of the real-time or prediction period feature data of the corresponding base station to obtain the traffic prediction results of each base station for 96 time steps in the future time period.

[0012] Furthermore, step S1 includes the following steps:

[0013] Step S11: Set the coverage area of ​​the intelligent network management center. The coverage area of ​​the 5G base station is the selected base station area.

[0014] Step S12: Data acquisition modules are installed in both the BBU and RRU / AAU devices in the 5G base station. The intelligent network management center uses the data acquisition modules to monitor and collect the device current in real time.

[0015] Step S13: The intelligent network management center collects and summarizes 5G transmission-related feature data at a granularity of 5 minutes. The feature data includes the current PRB utilization rate of the base station, the number of successful RRC establishments, the RRC establishment success rate, the number of RRC connected users, user uplink and downlink traffic, RRU model, and base station model.

[0016] Furthermore, step S3 includes the following steps:

[0017] Step S31: Fill all null values ​​in the resampled dataset with zeros;

[0018] Step S32: Strictly divide the data of each day into 24-hour nodes. If data is missing at certain time nodes during the data collection process, fill it with the historical average value of the base station.

[0019] Step S33: Divide the resampled dataset into training, validation and test sets in a 6:2:2 ratio according to time order to ensure the continuity and integrity of the training data.

[0020] Furthermore, the function implemented in step S4 based on the 5G base station traffic prediction model of the decomposable periodic pattern network includes the following steps:

[0021] Step S41: Standardize the input sequence to obtain a standardized sequence;

[0022] Step S42: Standardize the sequence by length... Using a sliding window approach, multiple standardized subsequences are generated in chronological order.

[0023] Step S43: For each standardized subsequence, the periodic pattern removal unit of the periodic pattern modeling module removes the periodicity from the standardized subsequence to obtain the residual sequence;

[0024] Step S44: The component decomposition module further divides the residual sequence into a trend component with smooth changes and a seasonal component with short-term fluctuations.

[0025] Step S45: Apply linear mapping to the trend component and the seasonal component respectively, sum them, perform batch normalization to obtain the normalized fusion sequence, and input it into the xLSTM module;

[0026] Step S46: The xLSTM module is composed of multiple sLSTM modules and multiple mLSTM modules in a linear combination. The xLSTM module receives the normalized fusion sequence and uses the module structure that combines the gated memory characteristics of sLSTM and the matrix memory characteristics of mLSTM to model the time dependency in the normalized fusion sequence and the spatial dependency between different base stations, thereby obtaining the feature information of the base station data.

[0027] Step S47: The feature information of the base station data obtained by the xLSTM module is mapped through the third linear layer to obtain the future... The residual prediction sequence at each time step;

[0028] Step S48: The periodic pattern modeling module restores the periodic characteristics in the standardized subsequence by recovering the periodic pattern unit, and obtains the predicted sequence by combining it with the residual prediction sequence;

[0029] Step S49: Perform denormalization on the predicted sequence to obtain the final prediction result.

[0030] Further, step S43 includes the following steps:

[0031] Step S431: Remove periodic pattern units to construct a periodic memory matrix ,in The periodic memory matrix represents the length of the periodic window. These are trainable parameters;

[0032] Step S432: At time step At that time, for the periodic memory matrix Perform a circular left shift Bitwise operations to obtain the current time step Synchronization period template ;

[0033] Step S433: Synchronization period template Repeated splicing Next, and add to its preceding bit sequence This yields a historical periodic segment aligned with the length of the sliding window. ;

[0034] Step S434: Subtract the corresponding elements from the standardized subsequence and the historical periodic segment generated by removing the periodic pattern unit to obtain the residual sequence.

[0035] Further, step S44 includes the following steps:

[0036] Step S441: Further decompose the residual series into trend and seasonal components. First, use local average pooling to extract the trend component.

[0037] Step S442: Subtract the residual sequence from its trend component to obtain the seasonal component.

[0038] Furthermore, step S5 includes the following steps:

[0039] Step S51: Use the mean absolute error as the loss function of the model and compare the predicted value output by the model with the actual flow rate.

[0040] Step S52: Input the training set data into the 5G base station traffic prediction model of the constructed decomposable periodic pattern network for training;

[0041] Step S53: Use the Adam optimizer to perform backpropagation optimization on the network parameters to minimize the above loss function, and save the training weight parameters of each round during the training process; finally, select the parameter with the smallest loss value on the test set as the optimal model parameter corresponding to the base station.

[0042] Step S54: Deploy the optimal model parameters obtained from training all base stations to the intelligent network management center, and input the historical 96 time steps of the real-time or predicted period feature data of the corresponding base stations in subsequent operations, thereby obtaining the traffic prediction results of each base station for 96 time steps in the future time period. The beneficial effects of this invention are:

[0043] (1) The Decomposable Periodic Pattern Network (DPPNet) proposed in this invention can accurately extract the long-term periodic structure in 5G base station traffic data by explicitly modeling the periodic memory matrix and combining synchronous reconstruction and periodic segment splicing mechanism. This effectively improves the model's ability to capture complex periodic patterns and long-term dependent information, thereby achieving high-precision prediction of future traffic.

[0044] (2) The present invention further divides the residual sequence into trend components and seasonal components through the component decomposition module, which significantly improves the model’s adaptability to non-stationary noise and local disturbances; at the same time, the xLSTM module is introduced, which integrates scalar memory and matrix memory mechanisms, while maintaining high modeling ability, significantly suppressing gradient vanishing and parameter inflation problems, so that the prediction system has good stability and generalization ability.

[0045] (3) The model of the present invention has a lightweight overall structure, low parameter quantity, high operating efficiency, and is easy to deploy in edge computing environment or intelligent network management center. While ensuring prediction accuracy, it reduces runtime resource consumption, improves the feasibility of long-term online deployment and rapid response of the model in actual 5G network, and further provides a reliable basis for base station energy saving strategy optimization and operation and maintenance scheduling. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a schematic diagram of the process of the present invention;

[0048] Figure 2 This is a schematic diagram of the overall network structure of the predictive model based on decomposable periodic patterns of the present invention.

[0049] Figure 3 This is a schematic diagram of the periodic pattern modeling module based on the decomposable periodic pattern network prediction model of the present invention.

[0050] Figure 4 This is a schematic diagram of the sLSTM and mLSTM modules included in the xLSTM module of the decomposable periodic pattern network prediction model of the present invention.

[0051] Figure 5 This is a schematic diagram of the base station prediction results based on the decomposable periodic pattern network of the present invention. Detailed Implementation

[0052] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, 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.

[0053] Deploy a 5G base station traffic prediction model based on a decomposable periodic pattern network with hourly time granularity in the intelligent network management center.

[0054] like Figure 1 As shown, this invention provides a 5G base station traffic prediction method based on a decomposable periodic pattern network, the method comprising the following steps:

[0055] Step S1: The intelligent network management center collects characteristic data related to 5G transmission in 5G base stations through the data acquisition module and sets the coverage area of ​​the intelligent network management center.

[0056] Step S11: Set the coverage area of ​​the intelligent network management center. The coverage area of ​​the 5G base station is the selected base station area.

[0057] Step S12: Data acquisition modules are installed in the BBU (Baseband Unit) and RRU (Remote Radio Unit) / AAU (Active Antenna Unit) devices in the 5G base station. The intelligent network management center uses the data acquisition modules to monitor and collect the device current in real time.

[0058] Step S13: The intelligent network management center collects and summarizes 5G transmission-related feature data at a granularity of 5 minutes. This feature data includes the base station's current PRB (Physical Resource Block) utilization rate, RRC (Radio Resource Control) establishment success count, RRC establishment success rate, number of RRC connected users, user uplink and downlink traffic, RRU model, and base station model; therefore, each 5G base station has 7 feature dimensions. Then, the energy consumption generated by the device per hour is calculated using the following formula:

[0059] Power = Voltage × Current; Energy consumption = Power × Time.

[0060] Step S2: Organize the collected feature data according to time level, and resample the feature data at hourly and daily collection intervals to obtain a resampled dataset.

[0061] Specifically, since it is necessary to deploy a 5G base station traffic prediction model based on a decomposable periodic pattern network with an hourly time granularity in the intelligent network management center, the data acquisition module resamples the sample data collected with a 5-minute acquisition granularity at hourly acquisition intervals.

[0062] Step S3: Clean the resampled dataset by filling all empty values ​​with zero values. Strictly set the data for each day to 24 time nodes in hours. If a time node is missing during the collection process, fill it with the average of the historical data of that base station. Divide the resampled dataset into training set, validation set and test set according to time order.

[0063] Step S31: Due to equipment failure and loss of collected information during the data acquisition process, the data acquisition module fills all null values ​​in the resampled dataset with zero values.

[0064] Step S32: To address the issue of missing 24-hour data entries for some dates due to data loss during collection, the data for each day is strictly divided into 24-hour nodes to ensure that each date contains 24 hourly data entries. For newly added hourly data, the historical average value of this base station is used to fill in the gaps.

[0065] Step S33: Divide the resampled dataset into training, validation, and test sets in a 6:2:2 ratio according to time sequence to ensure the continuity and integrity of the training data. In this embodiment, the resampled dataset contains 8808 hourly data entries, which are ultimately divided into 5282 training data entries, 1762 validation data entries, and 1760 test data entries.

[0066] Step S4: Design a 5G base station traffic prediction model based on a decomposable periodic pattern network, with the specific structure as follows: Figure 2 As shown, the 5G base station traffic prediction model based on decomposable periodic pattern networks includes a periodic pattern modeling module, a component decomposition module, and an xLSTM module. The periodic pattern modeling module includes a periodic pattern removal unit and a periodic pattern recovery unit. The periodic pattern modeling module is used to extract periodic features from the data. The component decomposition module decomposes the data into trend components and seasonal components. The xLSTM module is used to capture long-term dependencies.

[0067] The 5G base station traffic prediction model based on decomposable periodic pattern networks receives data in a sliding window format with a length of [missing information]. Standardized historical observation data Output the future Predicted sequence at time steps Then, inverse standardization is used to obtain the final prediction result. Considering that the time-varying distribution of actual 5G base station traffic data may lead to biases in model feature extraction, this model first standardizes the input historical observation data to reduce the interference of distribution variations and outliers on subsequent modeling. Then, the periodic pattern removal unit of the periodic modeling module learns cyclical patterns from the historical observation data. (Length is determined by hyperparameters) (Control), and perform phase alignment on the historical window to obtain standardized historical observation data. Periodic components of the same dimension This periodic component, as an interpretable global prior, is removed from the standardized historical observation data, resulting in a residual sequence that is easier to model. This explicitly decouples periodic and non-periodic information, facilitating subsequent component decomposition and feature extraction.

[0068] Based on this, the component decomposition module further decomposes the residual sequence Divided into trend components with smooth changes Seasonal component of short-term fluctuations The two decomposed signals are linearly projected to a higher dimension and then feature fusion is performed to provide more stable and modelable input data for the subsequent xLSTM module.

[0069] To address spatiotemporal dependencies, this invention introduces an xLSTM module as the backbone network of the prediction model. While maintaining causality, the xLSTM module receives input data processed by the component decomposition module and utilizes a module structure combining gated memory (sLSTM) and matrix memory (mLSTM) to model the temporal dependencies in the time-series data and the spatial dependencies between different base stations, thereby outputting the future... Residual prediction sequence at time step Meanwhile, the cyclical pattern With residual prediction sequence Then align them to obtain the periodic components. It is used to recover the periodic information of the residual sequence to obtain the predicted sequence. Finally, through destandardization, the final predicted sequence is obtained. .

[0070] The functions implemented by the 5G base station traffic prediction model based on decomposable periodic pattern networks in step S4 include the following steps:

[0071] Step S41: For the input sequence Standardization is performed to obtain the standardized sequence. To reduce the interference of distribution variations and outliers on subsequent modeling, the standardization process is as follows:

[0072]

[0073] in, This represents the mean of the input sequence. Indicates standard deviation, , Represents the set of real numbers. Indicates the number of samples in the batch. This represents the time step length for each sample. This represents the number of feature channels at each time step.

[0074] Step S42: Standardize the sequence With length as Using a sliding window method to generate multiple time-sequentially consecutive normalized subsequences .

[0075] Step S43: For each normalized subsequence The periodic pattern modeling module removes periodic pattern units and removes normalized subsequences. The periodicity of the residual sequence is used to obtain the residual sequence. .

[0076] Step S431: Remove periodic pattern units to construct a periodic memory matrix ,in This represents the length of the periodic window. The periodic memory matrix... These are trainable parameters, initially set to zero, and jointly optimized with the xLSTM backbone network during model training, such as... Figure 3 As shown.

[0077] Step S432: At time step At that time, for the periodic memory matrix Perform a circular left shift Bitwise operations to obtain the current time step Synchronization period template This is to achieve periodic alignment operations.

[0078] Step S433: Synchronization period template Repeated splicing Next, and add to its preceding bit sequence This yields a historical periodic segment aligned with the length of the sliding window. , represented as:

[0079]

[0080] Step S434: Standardize the subsequence Subtracting the corresponding elements from the historical periodic segments generated by removing the periodic pattern units yields the residual sequence. , represented as:

[0081]

[0082] The above steps utilize trainable periodic templates. The synchronization and reconstruction mechanism explicitly learns the periodic structure in the sequence, thereby providing stable periodic prior knowledge in the historical input and future prediction stages, enhancing the model's ability to capture long-term temporal dependencies, and improving prediction accuracy and model interpretability.

[0083] Step S44: The component decomposition module further decomposes the residual sequence Trend components classified as smooth changes Seasonal components of short-term fluctuations .

[0084] Step S441: Process the residual sequence To further decompose the trend and seasonal components, the trend component is first extracted using local average pooling. The calculation formula is as follows:

[0085]

[0086] in, Indicates trend components, This indicates a local average pooling operation, used to smooth short-term disturbances and highlight long-term trends.

[0087] Step S442: Subtract the residual series from its trend component to obtain the seasonal component. The formula is:

[0088]

[0089] in, This indicates short-term cyclical changes, i.e., seasonal components.

[0090] Step S45: Apply linear mapping to the trend component and the seasonal component respectively, sum them, and then perform batch normalization to obtain a normalized fusion sequence, which is then input into the xLSTM module.

[0091] Step S451: Perform a first linear layer mapping on the trend component and a second linear layer mapping on the seasonal component to obtain the same and higher-dimensional representation; in this embodiment, the feature dimension after mapping is set to 256. Add the two elements together to obtain a fused sequence that simultaneously contains trend and seasonal information.

[0092] Step S452: Perform batch normalization (BN) on the fused sequence to obtain a normalized fused sequence.

[0093] By introducing batch normalization after the fusion sequence, feature distribution drift can be effectively alleviated and numerical stability can be improved, thereby accelerating model convergence, reducing sensitivity to parameter initialization, and obtaining a certain regularization effect without introducing additional supervision, thus improving the training efficiency and generalization performance of the xLSTM module as a whole.

[0094] Step S46: The xLSTM module is composed of multiple sLSTM modules and multiple mLSTM modules in a linear combination. The xLSTM module receives the normalized fusion sequence and uses the module structure that combines the gated memory characteristics of sLSTM and the matrix memory characteristics of mLSTM to model the time dependency in the normalized fusion sequence and the spatial dependency between different base stations, thereby obtaining the feature information of the base station data.

[0095] Step S461: In the first sLSTM module, the normalized fusion sequence is processed by the first layer of normalization (LN). The first normalized fusion sequence is as follows: Figure 4 As shown in (a) of the diagram.

[0096] Step S462: The first normalized fusion sequence is subjected to a causal convolution (casual convolution, kernel=4) with a kernel size of 4. The unidirectional convolution maintains the temporal order constraint and aggregates local context information. Then, it is subjected to the Swish activation function to enhance the nonlinear expression. The final result after convolution is fed into a block-diagonal projection with 4 parallel heads to generate the parameters of the input gate and forget gate, respectively.

[0097] Step S463: The first normalized fusion sequence is fed into the block diagonal linear mapping to obtain the intermediate candidate memory state and the parameters of the output gate.

[0098] The block diagonal design limits the mutual projection between different heads, thereby reducing parameters and computational load while ensuring parallelism, and avoiding noise amplification caused by the mixing of irrelevant channels.

[0099] Step S464: Input gate, forget gate, output gate and intermediate candidate memory state are input into the sLSTM unit, and the update is completed in a single step to obtain the sLSTM hidden state output.

[0100] In the sLSTM cell, define the time step. Scalar memory state and normalized state The calculations are performed recursively using the following formulas:

[0101]

[0102] in, , Representing time steps Scalar memory state and normalized state under the given conditions; , These are the input gate and the forget gate, respectively. Input for intermediate candidate memories, , , Both are derived from the input vector and the previous hidden state through exponentiation. The gating function is calculated and represented as follows:

[0103]

[0104]

[0105]

[0106] Final hidden state The hidden state output of the sLSTM is obtained by normalizing the state and output gates. The calculation yielded:

[0107]

[0108] in, express Activation function Indicates candidate memory input Nonlinear activation functions are used to enhance expressive power, and are typically selected... or Activation function; Represents an exponential function; Represents the input sequence; Indicates the previous hidden state; variable , , , These represent the linear weights of the intermediate candidate memory input, input gate, forget gate, and output gate, respectively. , , , These represent the recurrent connection weights of the intermediate candidate memory input, input gate, forget gate, and output gate, respectively. , , , These represent the bias terms for intermediate candidate memory input, input gate, forget gate, and output gate, respectively. Indicates transpose; , , , These represent the intermediate states of the candidate memory input, input gate, forget gate, and output gate after linear operations, respectively. Indicates a hidden state; This indicates the output gate.

[0109] Step S464: The sLSTM hidden state output is integrated and scaled by the multi-head output through the first group normalization (GN). Then, the result of the first group normalization is added to the normalized fusion sequence through a short-circuit connection to form the first residual.

[0110] The first set of normalizations can reduce the impact of inconsistent statistical distributions between different heads on subsequent calculations. By adding the short-circuit connections, the stable features of the lower layers can be preserved and the difficulty of deep optimization can be reduced.

[0111] Step S465: After the first residual is normalized by the second layer, it passes through the feedforward projection branch consisting of "upward projection - gating - downward projection" to obtain the output of the sLSTM module.

[0112] The channel dimension of the feedforward projection branch is first increased by the projection factor (PF) of 4 / 3. After upprojection, it is split into two parallel branches, one of which is a direct linear mapping, and the other is smoothed by the GeLU activation function. Then, the two branches are multiplied element-wise to form a lightweight gated feedforward structure to suppress ineffective activations. After gating, downprojection is performed by PF of 3 / 4 to reduce the channel dimension to a scale consistent with the normalized fusion sequence. Finally, it is added to the first residual to obtain the final result of this layer, which is the output of the sLSTM module.

[0113] Step S466: The output of the sLSTM module is used as the input of the next sLSTM module until the last sLSTM module; the output of the last sLSTM module is used as the input of the first mLSTM module.

[0114] Step S467: The first mLSTM module uses the output of the last sLSTM module as input for the third layer of normalization, completing scale alignment and numerical stabilization, resulting in a normalized input sequence, such as... Figure 4 As shown in (b) of the diagram.

[0115] Step S468: Upscale the normalized input sequence by projection factor PF=2 to expand the intermediate channel dimension and obtain two projection sequences.

[0116] Step S469: One projection sequence is used to model local temporal variations. It is first processed by a causal convolution with a kernel size of 4 to extract the nearest neighbor temporal dependencies, and then connected to the Swish activation function to enhance the nonlinear response, resulting in an enhanced sequence. The enhanced sequence is fed into a block diagonal linear mapping with a block size of 4, and another projection sequence is directly fed into the block diagonal linear mapping. The block diagonal linear mapping performs matrix memory modeling, forming a multi-head query (Q), key (K), and value (V) representation under the constraints of structured projection, and input into the mLSTM unit.

[0117] Step S4610: The mLSTM unit receives the query (Q), key (K), and value (V) representations, controls memory writing and decay through the input gate and forget gate, and realizes the association retrieval and update of cross-time interaction by relying on matrix storage, and obtains the mLSTM hidden state output.

[0118] In the mLSTM unit, a matrix-based memory mechanism further enhances the network's ability to model complex temporal structures. The mLSTM unit updates the memory matrix in key-value pairs, employing a covariance-driven update rule and introducing a threshold normalization strategy to improve the stability of long-term information retrieval. Its forward computation formula is as follows:

[0119]

[0120]

[0121]

[0122]

[0123]

[0124]

[0125]

[0126]

[0127]

[0128] in, Indicates time step The associated memory matrix below, Indicates time step The associated memory matrix below; , , These are vectors representing the query, key, and value, respectively. , , These are the input gate, forget gate, and output gate of mLSTM, respectively. Represents the input sequence; Indicates time step The normalized state below; express Element-wise multiplication operation; This represents an intermediate state of the hidden state after a non-linear operation. This represents the final hidden state, which is the output of the mLSTM hidden state. , , , These represent the linear transformation weight matrices for the query, key, value, and output gates, respectively. , These represent the linear weights of the input gate and the forget gate, respectively. , , , , , These represent the biases for query, key, value, input gate, forget gate, and output gate, respectively. Indicates the dimension of a vector; It is a normalization factor used to avoid numerical instability issues such as division by zero.

[0129] Step S4611: The mLSTM hidden state output, after being normalized by the second set, is fused with the augmented sequence through the learnable skip connection result to obtain the second residual.

[0130] The mLSTM hidden state output, after a second set of normalizations, reduces training instability caused by differences in statistics across different channels. Residual fusion then transforms the second residual into an intermediate representation that combines long-range dependencies with local details.

[0131] Step S4612: After the normalized input sequence is increased in dimension by projection factor PF=2, the result processed by the Swish activation function is multiplied by the second residual row channel by channel, then reduced in dimension by downprojection with projection factor PF=1 / 2, and finally residually concatenated with the input sequence of the mLSTM module to obtain the output of the mLSTM module.

[0132] Adaptive modulation and gating of content are achieved through channel-wise dot product. Down-projection dimensionality reduction ensures that the output of the mLSTM module maintains the same channel size as its input, and the feature dimension of the mLSTM module's output is consistent with the feature dimension of the normalized fused sequence input to the xLSTM module; in this embodiment, this dimension is set to 256. Residual connections ensure gradient propagation and information fidelity during deep training.

[0133] Step S4613: The output of the mLSTM module is used as the input of the next mLSTM module until the last mLSTM module; the output of the last mLSTM module is used as the final output of the xLSTM module, which is the feature information of the base station data.

[0134] The xLSTM module combines the advantages of scalar and matrix memory mechanisms, significantly enhancing the model's stability and expressive power in long sequence modeling. It is suitable for effectively extracting potential long-term dependency patterns in 5G base station traffic data, thereby improving prediction accuracy.

[0135] Step S47: The feature information of the base station data obtained by the xLSTM module is mapped through the third linear layer to obtain the future... Residual prediction sequence at each time step .

[0136] To enable the xLSTM module to output a length of The residual prediction sequence requires a linear layer mapping of the xLSTM module to transform its output base station data and align it with the length of the residual prediction sequence in the time dimension. The residual prediction sequence obtained through this mapping provides stable and usable input data for subsequent processes.

[0137] Step S48: Recovery of normalized subsequences from the recovery periodic pattern unit of the periodic pattern modeling module. The periodic characteristics in the residual prediction sequence Obtain the predicted sequence .

[0138] Step S481: Restore the periodic pattern unit to construct the periodic memory matrix ,in This represents the length of the periodic window. The periodic memory matrix... These are trainable parameters, initially set to zero, and jointly optimized with the xLSTM backbone network during model training.

[0139] Step S482: At time step At that time, for the periodic memory matrix Perform a circular right shift Bitwise operations to obtain the current time step Synchronization period template This is to achieve periodic alignment operations.

[0140] Step S483: Synchronization period template Repeated splicing Next, and add to its preceding bit sequence This yields future periodic segments aligned with the length of the sliding window. , represented as:

[0141]

[0142] in, Indicates the current time step, symbol Indicates rounding down. Indicates a modulo operation;

[0143] Step S484: Add the residual prediction sequence to the future periodic segments generated by the recovery periodic mode unit at corresponding positions to obtain the prediction sequence. , represented as:

[0144]

[0145] Step S49: The predicted sequence is destandardized to obtain the final prediction result. The destandardization process is as follows:

[0146]

[0147] in, This represents the mean of the input sequence. Indicates standard deviation, This indicates the final prediction result.

[0148] Step S5: Input the training set data into the 5G base station traffic prediction model of the decomposable periodic pattern network for training. After iterative optimization, select the model parameters with the best performance on the test set for 5G base station traffic prediction. Deploy the trained 5G base station traffic prediction model of the decomposable periodic pattern network to the intelligent network management center. In subsequent operation, input the historical 96 time steps of the real-time or prediction period feature data of the corresponding base station to obtain the traffic prediction results of each base station for 96 time steps in the future time period. The specific steps are as follows:

[0149] Step S51: Using the mean absolute error (MAE) as the model's loss function, compare the predicted values ​​output by the model with the actual flow rates. The specific calculation formula is as follows:

[0150]

[0151] in, Represents the loss function; This represents the flow rate predicted by the model. This represents the actual observed flow rate at the corresponding time step. This represents the total number of prediction time steps.

[0152] Step S52: Input the training set data into the constructed 5G base station traffic prediction model of the decomposable periodic pattern network for training, setting the number of training epochs to 80 and the initial learning rate to 0.002. After each training epoch, evaluate the model performance using the validation set and calculate the loss value. If the validation performance does not improve further, halve the learning rate to continue optimizing the network parameters.

[0153] Step S53: Use the Adam optimizer to perform backpropagation optimization on the network parameters to minimize the above loss function, and save the training weight parameters of each round during training; finally, select the parameter with the smallest loss value on the test set as the optimal model parameter corresponding to the base station.

[0154] Step S54: Deploy the optimal model parameters obtained from training all base stations to the intelligent network management center, and input the historical 96 time steps of the real-time or predicted period feature data of the corresponding base stations in subsequent operations, thereby obtaining the traffic prediction results of each base station for 96 time steps in the future period, such as... Figure 5 As shown.

[0155] Furthermore, the intelligent network management center can intelligently adjust the power output of each base station or shut down some equipment during off-peak hours based on the traffic prediction results of each base station, thereby significantly reducing the overall energy consumption of the base station and reducing manpower and time costs.

[0156] It should be understood that the input historical time steps are not limited to the 96 time steps mentioned above, the future time steps are not limited to the 96 time steps mentioned above, and the training parameter settings are not limited to the specific settings mentioned above. No restrictions are imposed here.

[0157] The above description is only a typical embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1.A 5G base station traffic prediction method based on a decomposable periodic pattern network, characterized by, The method comprises the following steps: Step S1: The intelligent network management center collects 5G transmission related feature data in the 5G base station through the data acquisition module, and sets the coverage area of the intelligent network management center; Step S2: The collected feature data is sorted by time level, and the feature data is resampled at an hourly and daily collection interval to obtain a resampled data set; Step S3: The resampled data set is data cleaned, all null values are filled with zero values, and the data of each day is strictly set to 24 time nodes, if a certain time node is missing in the collection process, the historical data of the base station is filled with the average value, and the resampled data set is divided into a training set, a validation set and a test set in time sequence; Step S4: A 5G base station traffic prediction model based on a decomposable periodic pattern network is designed; the 5G base station traffic prediction model based on the decomposable periodic pattern network comprises a periodic pattern modeling module, a component decomposition module and an xLSTM module; the periodic pattern modeling module comprises a remove periodic pattern unit and a restore periodic pattern unit, the periodic pattern modeling module is used to extract the periodic characteristics in the data, the component decomposition module decomposes the data into a trend component and a seasonal component, and the xLSTM module is used to capture long-time dependence; Step S5: The training set data is input into the 5G base station traffic prediction model based on the decomposable periodic pattern network for training, and after iterative optimization, the model parameters with the best performance on the test set are selected for 5G base station traffic prediction, and the trained 5G base station traffic prediction model based on the decomposable periodic pattern network is deployed to the intelligent network management center, and in subsequent operation, the historical 96 time steps of the real-time or predicted period feature data of the corresponding base station are input, so as to obtain the traffic prediction result of each base station in the future time period; The step S4 comprises the following steps: Step S41: The input sequence is standardized to obtain a standardized sequence; Step S42: standardizing the sequence to length In the sliding window manner, a plurality of time-sequentially continuous standardized sub-sequences are generated. Step S43: For each standardized subsequence, the remove periodic pattern unit of the periodic pattern modeling module removes the periodic characteristics in the standardized subsequence to obtain a residual sequence; Step S44: The component decomposition module further divides the residual sequence into a smooth trend component and a short-term fluctuation seasonal component; Step S45: After linear mapping and summation are performed on the trend component and the seasonal component respectively, batch normalization processing is performed to obtain a normalized fusion sequence, which is input into the xLSTM module; Step S46: The xLSTM module is linearly combined by a plurality of sLSTM modules and a plurality of mLSTM modules, the xLSTM module receives the normalized fusion sequence, and models the time dependence in the normalized fusion sequence and the spatial dependence between different base stations by using the module structure combining the gating memory characteristics of the sLSTM and the matrix memory characteristics of the mLSTM, to obtain feature information of the base station data; Step S47: The feature information of the base station data obtained by the xLSTM module is mapped by a third linear layer to obtain a residual prediction sequence of future time steps; Step S48: The restore periodic pattern unit of the periodic pattern modeling module restores the periodic characteristics in the standardized subsequence to obtain a prediction sequence in combination with the residual prediction sequence; Step S49: The prediction sequence is subjected to inverse standardization processing to obtain a final prediction result; The step S43 comprises the following steps: Step S431: remove period pattern unit to build period memory matrix wherein denotes the period window length; the period memory matrix is a trainable parameter; Step S432: at the time step , the cycle memory matrix is cyclically left-shifted by one bit to obtain the synchronization cycle template of the current time step ; Step S433: repeating splicing on the synchronization period template ;​​​​ Step S434: subtract the corresponding position elements of the standardized sub-sequence and the historical period segment generated by the period pattern removal unit to obtain a residual sequence; The step S44 comprises the following steps: Step S441: further decompose the residual sequence into trend and seasonal components, and first extract the trend component using a local average pooling method; Step S442: subtract the trend component from the residual sequence to obtain the seasonal component. 2.The 5G base station traffic prediction method based on the decomposable periodic pattern network according to claim 1, wherein, The step S1 comprises the following steps: Step S11: set the range of the intelligent network management center coverage area, and the range of the 5G base station coverage area is the selected base station area; Step S12: the BBU and RRU / AAU devices in the 5G base station are both installed with a data acquisition module, and the intelligent network management center performs real-time monitoring and acquisition of device current through the data acquisition module; Step S13: the intelligent network management center collects and summarizes the feature data related to 5G transmission with a collection granularity of 5 minutes, and the feature data includes the current PRB utilization rate of the base station, the number of successful RRC establishment, the success rate of RRC establishment, the number of RRC connection users, the uplink and downlink traffic of users, the RRU model, and the base station model. 3.The 5G base station traffic prediction method based on the decomposable periodic pattern network according to claim 1, wherein, The step S3 comprises the following steps: Step S31: zero value filling is performed on all null values in the resampled data set; Step S32: strictly divide the data of each day into 24-hour nodes, and if the data of some time nodes is missing in the data acquisition process, the historical average value of the base station is used for filling; Step S33: divide the resampled data set into a training set, a validation set and a test set in the order of time with a ratio of 6:2:2, to ensure the continuity and integrity of the training data. 4.The 5G base station traffic prediction method based on the decomposable periodic pattern network according to claim 1, wherein, The step S5 comprises the following steps: Step S51: use the mean absolute error as the loss function of the model, and compare the predicted value output by the model with the actual traffic value; Step S52: input the training set data into the 5G base station traffic prediction model of the decomposable period pattern network constructed to perform training; Step S53: use the Adam optimizer to perform back propagation optimization on the network parameters to minimize the loss function, and save the training weight parameters in each round during the training process; finally, select the parameter with the minimum loss value on the test set as the optimal model parameter corresponding to the base station; Step S54: deploy the optimal model parameters obtained by training all base stations to the intelligent network management center, and input the historical 96 time steps of real-time or predicted period feature data of the corresponding base station in subsequent operation, to obtain the traffic prediction result of each base station in the future time period.

Images