Method for determining load of regional communication base station and device thereof, and electronic device

Abstract

The application discloses a kind of regional communication base station load determination method and device and electronic equipment, it is related to artificial intelligence technical field, wherein the determination method includes: the historical load data sequence of target communication base station in the target area in pre-set historical time window is collected, and historical load data sequence is parsed, obtain multiple target periods of pre-set historical time window;Based on multiple target periods, data extraction is carried out to historical load data sequence, and multiple historical load data subsequences are obtained;Historical load data sequence and multiple historical load data subsequences are input to target load prediction model, and the load prediction value in pre-set time period is obtained;Standardization processing is carried out to historical load data sequence, and mean and standard deviation are obtained;Based on mean, standard deviation and load prediction value, the target load value of target communication base station is determined.The application solves the technical problem that the precision of base station load prediction in related art is low.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method, apparatus, and electronic device for determining the load of a regional communication base station. Background Technology

[0002] Base station energy management, network operation and maintenance, and rational resource allocation all rely on accurate prediction of regional communication base station load. Currently, base station load prediction mostly adopts time series analysis models and basic machine learning methods, with only a few solutions incorporating shallow network technology. These methods can predict accurately in scenarios with smooth load fluctuations, but with the upgrading of communication networks, these methods suffer from low prediction accuracy and poor stability, making it difficult to meet the needs of accurate day-ahead and short-term load prediction and refined energy management for base stations.

[0003] There is currently no effective solution to the above problems. Summary of the Invention

[0004] This invention provides a method, apparatus, and electronic device for determining the load of a regional communication base station, in order to at least solve the technical problem of low accuracy in predicting base station load in related technologies.

[0005] According to one aspect of the embodiments of this application, a method for determining the load of a regional communication base station is provided, comprising: collecting historical load data sequences of a target communication base station in a target area within a preset historical time window, and parsing the historical load data sequences to obtain multiple target time periods within the preset historical time window; extracting data from the historical load data sequences based on the multiple target time periods to obtain multiple historical load data sub-sequences; inputting the historical load data sequences and the multiple historical load data sub-sequences into a target load prediction model to obtain load prediction values ​​within the preset time period, wherein the target load prediction model is used to extract a first historical multi-scale interactive enhancement feature of the historical load data sequence, extract a second historical multi-scale interactive enhancement feature of each historical load data sub-sequence, and extract a time-series dependency feature of the historical load data sequence, and generate a fused feature vector based on the first historical multi-scale interactive enhancement feature, all second historical multi-scale interactive enhancement features, and the time-series dependency feature, and mapping the fused feature vector to a load prediction value; standardizing the historical load data sequence to obtain the mean and standard deviation; and determining the target load value of the target communication base station based on the mean, standard deviation, and load prediction value.

[0006] Furthermore, the network structure of the target load prediction model includes: a temporal convolutional network, an extended long short-term memory network, a feature fusion layer, and an output layer. The temporal convolutional network includes at least a first branch and a second branch. The first branch extracts the first historical multi-scale interactive enhancement features of the historical load data sequence, and the second branch extracts the second historical multi-scale interactive enhancement features of each historical load data subsequence. The extended long short-term memory network extracts the temporal dependency features of the historical load data sequence. The feature fusion layer includes at least an attention weight allocation module and a fully connected layer. The attention weight allocation module weights the first historical multi-scale interactive enhancement features, the second historical multi-scale interactive enhancement features, and the temporal dependency features. The fully connected layer fuses and compresses the weighted features to generate a fused feature vector. The output layer maps the fused feature vector to a one-dimensional load prediction value.

[0007] Further, the steps of parsing the historical load data sequence to obtain multiple target time periods within a preset historical time window include: traversing the historical load data sequence and performing time-series gradient analysis on each historical load data point to obtain load data gradient values ​​at multiple time points; comparing each load data gradient value with a preset time-series gradient threshold to obtain all time points where the load data gradient value is greater than or equal to the preset time-series gradient threshold; constructing a set of filtered time periods based on all time points; assigning values ​​to the time points within the preset historical time window based on the set of filtered time periods to construct a two-dimensional assignment coordinate system, where the horizontal axis of the two-dimensional assignment coordinate system is the time axis and the vertical axis is the number of assignments, and the two-dimensional assignment coordinate system includes at least multiple assignment points; and performing time period diffusion analysis on the multiple assignment points to determine multiple target time periods.

[0008] Furthermore, the steps of performing time-period diffusion analysis on multiple assignment points to determine multiple target time periods include: filtering multiple assignment points based on the maximum value of the number of assignments to determine multiple initial diffusion assignment points; taking each initial diffusion assignment point as a starting point and performing time-period diffusion on the assignment points based on each starting point to obtain initial diffusion time periods; and filtering all initial diffusion time periods based on a preset time period length threshold to determine multiple target time periods.

[0009] Further, the step of performing time-period diffusion on the assigned points based on each starting point to obtain the initial diffusion period includes: for each starting point, diffusion is performed on the starting point based on a preset diffusion bandwidth to obtain an initial diffusion neighborhood; based on the preset diffusion bandwidth, neighborhood edge diffusion is performed on the initial diffusion neighborhood to obtain a stage diffusion neighborhood; based on a preset time-period diffusion stopping strategy, the initial diffusion neighborhood and the stage diffusion neighborhood are analyzed to obtain the analysis result; if the analysis result indicates that diffusion has stopped, the time period corresponding to the initial diffusion neighborhood is determined as the initial diffusion period; if the analysis result indicates that diffusion should continue, diffusion is continued on the stage diffusion neighborhood based on the preset diffusion bandwidth until the analysis result indicates that diffusion has stopped, and the time period corresponding to the stage diffusion neighborhood is determined as the initial diffusion period.

[0010] Further, the steps of extracting the first historical multi-scale interactive enhancement features of the historical load data sequence through the first branch and extracting the second historical multi-scale interactive enhancement features of each historical load data subsequence through the second branch include: extracting features from the historical load data sequence through the first branch to obtain a historical multi-scale feature set; performing interactive enhancement on the historical multi-scale feature set to obtain the first historical multi-scale interactive enhancement features; extracting features from multiple historical load data subsequences through the second branch to obtain multiple historical multi-scale feature sets for target time periods; and performing interactive enhancement on the historical multi-scale feature set for each target time period to obtain the second historical multi-scale interactive enhancement features.

[0011] Further, the step of performing interactive enhancement on the historical multi-scale feature set to obtain the first historical multi-scale interactive enhanced feature includes: combining every two historical multi-scale features in the historical multi-scale feature set to generate multiple historical multi-scale feature combinations to be interacted, wherein the historical multi-scale feature combinations to be interacted correspond to a generation order; performing interactive enhancement on each historical multi-scale feature combination to be interacted according to the generation order to obtain the first-stage historical multi-scale interactive enhanced feature, and constructing the first-stage historical multi-scale interactive enhanced feature set based on all the first-stage historical multi-scale interactive enhanced features; calculating the mean of the first-stage historical multi-scale interactive enhanced feature set to obtain the first historical multi-scale interactive enhanced feature.

[0012] Further, following the generation order, the steps of performing interactive enhancement on each historical multi-scale feature combination to be interacted with to obtain the first-stage historical multi-scale interactive enhancement features include: calculating the similarity of features belonging to the same feature type in the historical multi-scale feature combination to be interacted with, obtaining an interactive enhancement similarity set, wherein the feature type is the category to which each historical multi-scale feature in the historical multi-scale feature set belongs, and the category is determined by the value of the convolution kernel in the first branch; normalizing each interactive enhancement similarity in the interactive enhancement similarity set, and filling each normalized interactive enhancement similarity into a preset initial matrix to obtain an interactive enhancement matrix; performing convolution operation on each historical multi-scale feature combination to be interacted with based on the interactive enhancement matrix to obtain an interactive historical multi-scale feature combination; calculating multiple means of the interactive historical multi-scale feature combination based on the feature dimension to obtain the mean feature, and determining the mean feature as the first-stage historical multi-scale interactive enhancement feature.

[0013] According to another aspect of the embodiments of this application, a device for determining the load of a regional communication base station is also provided, comprising: a parsing unit, used to collect historical load data sequences of a target communication base station in a target area within a preset historical time window, and to parse the historical load data sequences to obtain multiple target time periods within the preset historical time window; an extraction unit, used to extract data from the historical load data sequences based on the multiple target time periods to obtain multiple historical load data sub-sequences; an input unit, used to input the historical load data sequences and multiple historical load data sub-sequences into a target load prediction model to obtain load prediction values ​​within the preset time period, wherein the target load prediction model is used to extract a first historical multi-scale interactive enhancement feature of the historical load data sequence, extract a second historical multi-scale interactive enhancement feature of each historical load data sub-sequence, and extract a time-series dependency feature of the historical load data sequence, and generate a fused feature vector based on the first historical multi-scale interactive enhancement feature, all second historical multi-scale interactive enhancement features, and the time-series dependency feature, and map the fused feature vector to a load prediction value; a processing unit, used to standardize the historical load data sequence to obtain the mean and standard deviation; and a determination unit, used to determine the target load value of the target communication base station based on the mean, standard deviation, and load prediction value.

[0014] Furthermore, the network structure of the target load prediction model includes: a temporal convolutional network, an extended long short-term memory network, a feature fusion layer, and an output layer. The temporal convolutional network includes at least a first branch and a second branch. The first branch extracts the first historical multi-scale interactive enhancement features of the historical load data sequence, and the second branch extracts the second historical multi-scale interactive enhancement features of each historical load data subsequence. The extended long short-term memory network extracts the temporal dependency features of the historical load data sequence. The feature fusion layer includes at least an attention weight allocation module and a fully connected layer. The attention weight allocation module weights the first historical multi-scale interactive enhancement features, the second historical multi-scale interactive enhancement features, and the temporal dependency features. The fully connected layer fuses and compresses the weighted features to generate a fused feature vector. The output layer maps the fused feature vector to a one-dimensional load prediction value.

[0015] Furthermore, the parsing unit includes: a first analysis module, used to traverse the historical load data sequence and perform time-series gradient analysis on each historical load data in the historical load data sequence to obtain load data gradient values ​​at multiple time points; a first comparison module, used to compare each load data gradient value with a preset time-series gradient threshold to obtain all time points where the load data gradient value is greater than or equal to the preset time-series gradient threshold; a first construction module, used to construct a set of filtered time periods based on all time points; a second construction module, used to assign values ​​to time points in a preset historical time window based on the set of filtered time periods to construct a two-dimensional assignment coordinate system, wherein the horizontal axis of the two-dimensional assignment coordinate system is the time axis and the vertical axis is the number of assignments, and the two-dimensional assignment coordinate system includes at least multiple assignment points; and a first determination module, used to perform time period diffusion parsing on the multiple assignment points to determine multiple target time periods.

[0016] Furthermore, the first determining module includes: a first filtering submodule, used to filter multiple assignment points based on the maximum value of the assignment number to determine multiple initial diffusion assignment points; a first diffusion submodule, used to take each initial diffusion assignment point as a starting point and perform time-period diffusion on the assignment points based on each starting point to obtain an initial diffusion time period; and a second filtering submodule, used to filter all initial diffusion time periods based on a preset time period length threshold to determine multiple target time periods.

[0017] Further, the first diffusion submodule includes: a first diffusion subunit, used to diffuse each starting point based on a preset diffusion bandwidth to obtain an initial diffusion neighborhood; a second diffusion subunit, used to perform neighborhood edge diffusion on the initial diffusion neighborhood based on the preset diffusion bandwidth to obtain a stage diffusion neighborhood; a first analysis subunit, used to analyze the initial diffusion neighborhood and the stage diffusion neighborhood based on a preset time-period diffusion stopping strategy to obtain analysis results; a first determination subunit, used to determine the time period corresponding to the initial diffusion neighborhood as the initial diffusion time period when the analysis result indicates that diffusion should stop; and a second determination subunit, used to continue diffusion on the stage diffusion neighborhood based on the preset diffusion bandwidth when the analysis result indicates that diffusion should continue, until the analysis result indicates that diffusion should stop, and to determine the time period corresponding to the stage diffusion neighborhood as the initial diffusion time period.

[0018] Furthermore, the device for determining the load of regional communication base stations also includes: a first extraction module, used to extract features from historical load data sequences through a first branch to obtain a historical multi-scale feature set; a first enhancement module, used to perform interactive enhancement on the historical multi-scale feature set to obtain a first historical multi-scale interactive enhancement feature; a second extraction module, used to extract features from multiple historical load data sub-sequences through a second branch to obtain multiple historical multi-scale feature sets for target time periods; and a second enhancement module, used to perform interactive enhancement on the historical multi-scale feature set for each target time period to obtain a second historical multi-scale interactive enhancement feature.

[0019] Furthermore, the first enhancement module includes: a first combination submodule, used to combine every two historical multi-scale features in the historical multi-scale feature set to generate multiple historical multi-scale feature combinations to be interacted with, wherein the historical multi-scale feature combinations to be interacted with correspond to a generation order; a first enhancement submodule, used to perform interactive enhancement on each historical multi-scale feature combination to be interacted with according to the generation order to obtain the first-stage historical multi-scale interactive enhancement features, and to construct the first-stage historical multi-scale interactive enhancement feature set based on all the first-stage historical multi-scale interactive enhancement features; and a first calculation submodule, used to calculate the mean of the first-stage historical multi-scale interactive enhancement feature set to obtain the first historical multi-scale interactive enhancement features.

[0020] Further, the first enhancement submodule includes: a first calculation subunit, used to calculate the similarity of features belonging to the same feature type in the historical multi-scale feature combination to be interacted, to obtain an interaction enhancement similarity set, wherein the feature type is the category to which each historical multi-scale feature in the historical multi-scale feature set belongs, and the category is determined by the value of the convolution kernel in the first branch; a first processing subunit, used to normalize each interaction enhancement similarity in the interaction enhancement similarity set, and fill each normalized interaction enhancement similarity into a preset initial matrix to obtain an interaction enhancement matrix; a first operation subunit, used to perform convolution operation on each historical multi-scale feature combination to be interacted based on the interaction enhancement matrix to obtain an interaction historical multi-scale feature combination; and a second calculation subunit, used to calculate multiple means of the interaction historical multi-scale feature combination based on the feature dimension to obtain the mean feature, and determine the mean feature as the first-stage historical multi-scale interaction enhancement feature.

[0021] According to another aspect of the embodiments of this application, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements any of the above-described methods for determining the load of a regional communication base station.

[0022] According to another aspect of the embodiments of this application, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any of the above-described methods for determining the load of a regional communication base station.

[0023] In this invention, historical load data sequences of target communication base stations in a target area are collected within a preset historical time window. These historical load data sequences are then parsed to obtain multiple target time periods within the preset historical time window. Based on these multiple target time periods, data extraction is performed on the historical load data sequences to obtain multiple historical load data sub-sequences. The historical load data sequences and multiple historical load data sub-sequences are input into a target load prediction model to obtain load prediction values ​​within the preset time period. The target load prediction model extracts first historical multi-scale interactive enhancement features from the historical load data sequences, second historical multi-scale interactive enhancement features from each historical load data sub-sequence, and time-series dependency features from the historical load data sequences. Based on the first historical multi-scale interactive enhancement features, all second historical multi-scale interactive enhancement features, and the time-series dependency features, a fused feature vector is generated and mapped to load prediction values. The historical load data sequences are standardized to obtain the mean and standard deviation. Based on the mean, standard deviation, and load prediction values, the target load value of the target communication base station is determined, thus solving the technical problem of low accuracy in predicting base station load in related technologies.

[0024] In this invention, historical load data of communication base stations in a target area are collected within a preset historical time window. The target time period load subsequence is obtained through time-series parsing and other processing. Multi-scale feature extraction and interaction enhancement, as well as long-term time-series dependency feature analysis, are performed through multiple branches of the target load prediction model to calculate various load characteristic information. The results of the multi-branch feature processing are then fused to accurately predict the load situation of regional communication base stations, making the load prediction results more accurate and stable. By standardizing the historical load data sequence, the mean and standard deviation are obtained. Then, based on the mean, standard deviation, and load prediction values, the target load value of the target communication base station is calculated, improving the prediction accuracy and stability of the regional communication base station load and adapting to actual load prediction and scheduling needs. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0026] Figure 1 A hardware structure block diagram of a computer terminal (or mobile device) for implementing a method for determining the load of a regional communication base station is shown.

[0027] Figure 2 This is a flowchart of a method for determining the load of a regional communication base station according to Embodiment 1 of this application;

[0028] Figure 3This is a flowchart of a regional communication base station load prediction method that integrates multi-scale temporal features according to an embodiment of this application;

[0029] Figure 4 This is a structural diagram of a regional communication base station load prediction system that integrates multi-scale temporal features according to an embodiment of this application;

[0030] Figure 5 This is a schematic diagram of a device for determining the load of a regional communication base station according to an embodiment of this application;

[0031] Figure 6 This is a structural block diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0032] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0033] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0034] It should be noted that all relevant information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) collected and involved in this invention are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data comply with the relevant laws, regulations, and standards of the relevant regions, do not violate public order and good morals, and provide corresponding operation entry points for users to choose to authorize or refuse. For example, this system has an interface with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface. After receiving consent from the aforementioned user or organization, the relevant information is obtained. If the user chooses to refuse, the process enters the expert decision-making process.

[0035] The present invention will now be described in detail with reference to various embodiments.

[0036] Example 1

[0037] According to an embodiment of this application, an embodiment of a method for determining the load of a regional communication base station is also provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0038] The method embodiment provided in Embodiment 1 of this application can be executed on a mobile terminal, computer terminal, or similar computing device. Figure 1 A hardware block diagram of a computer terminal (or mobile device) for implementing a method for determining the load of a regional communication base station is shown. Figure 1 As shown, computer terminal 10 (or mobile device) may include one or more ( Figure 1 The processor 102 (which may include, but is not limited to, a microprocessor MCU or a programmable logic device FPGA, etc.), a memory 104 for storing data, and a transmission device 106 for communication functions may also be included. In addition, it may include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of a BUS bus), a network interface, a power supply, and / or a camera, wherein the network interface can be connected to wired and / or wireless networks. Those skilled in the art will understand that... Figure 1 The structure shown is for illustrative purposes only and does not limit the structure of the aforementioned electronic device. For example, computer terminal 10 may also include... Figure 1The more or fewer components shown, or having the same Figure 1 The different configurations shown.

[0039] It should be noted that the aforementioned one or more processors 102 and / or other data processing circuits are generally referred to herein as "data processing circuits". These data processing circuits may be embodied, in whole or in part, in software, hardware, firmware, or any other combination thereof. Furthermore, the data processing circuits may be a single, independent processing module, or may be integrated, in whole or in part, into any other element within the computer terminal 10 (or mobile device). As involved in the embodiments of this application, the data processing circuits serve as a processor control mechanism (e.g., selection of a variable resistor termination path connected to an interface).

[0040] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the method for determining the load of a regional communication base station in this embodiment of the application. The processor 102 executes various functional applications and data processing by running the software programs and modules stored in the memory 104, thereby realizing the aforementioned method for determining the load of a regional communication base station. The memory 104 may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include memory remotely located relative to the processor 102, and these remote memories can be connected to the computer terminal 10 via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0041] The transmission device 106 is used to receive or send data via a network. Specific examples of the network described above may include a wireless network provided by the communication provider of the computer terminal 10. In one example, the transmission device 106 includes a Network Interface Controller (NIC), which can connect to other network devices via a base station to communicate with the Internet. In another example, the transmission device 106 may be a Radio Frequency (RF) module, used for wireless communication with the Internet.

[0042] The display may be, for example, a touchscreen liquid crystal display (LCD) that allows the user to interact with the user interface of the computer terminal 10 (or mobile device).

[0043] Under the aforementioned operating environment, this application provides the following: Figure 2 The method for determining the load of regional communication base stations is shown. Figure 2This is a flowchart of a method for determining the load of a regional communication base station according to Embodiment 1 of this application, as follows: Figure 2 As shown, the method includes the following steps:

[0044] Step S201: Collect the historical load data sequence of the target communication base station in the target area within a preset historical time window, and parse the historical load data sequence to obtain multiple target time periods within the preset historical time window.

[0045] Optionally, the target communication base station in the target area is a mobile communication infrastructure deployed in a specific geographical area. It provides wireless communication signal transmission and reception, data transmission and other services to mobile terminals in the area through various communication hardware devices, and is a basic node of the mobile communication network.

[0046] The setting of the preset historical time window can be combined with the actual engineering needs of the daily and short-term load forecast of regional communication base stations. For example, based on engineering practice experience, the length of the historical time window can be initially set to 90 days. Then, the periodic characteristics of the historical load data of the base station can be analyzed to verify that the 90-day time length can fully cover the daily, weekly and monthly periodic characteristics of the base station load. Finally, the preset historical time window is determined to be a continuous 90-day time range. At the same time, the time axis within the window is equally divided into hourly granularities to ensure that the granularity of the time axis division is consistent with the granularity of the load data collection, providing a unified time dimension benchmark for subsequent two-dimensional time point assignment.

[0047] In this embodiment of the invention, firstly, historical load data sequences of target communication base stations in the target area are collected within a preset historical time window. Temporal gradient analysis is performed on the historical load data sequences, and time period filtering is completed based on a preset temporal gradient threshold to determine a set of filtered time periods. Then, based on this set of filtered time periods, two-dimensional time point assignments are applied to the preset historical time window, constructing a two-dimensional assignment coordinate system with the time axis as the horizontal axis and the assignment magnitude as the vertical axis, containing multiple assignment points. Finally, time period diffusion analysis is performed on the multiple assignment points in the coordinate system to determine multiple core time periods (i.e., target time periods) of the regional communication base station load.

[0048] Step S202: Based on multiple target time periods, extract data from the historical load data sequence to obtain multiple historical load data subsequences.

[0049] In this embodiment of the invention, the time intervals of the identified core time periods are first standardized. That is, the start and end times corresponding to each core time period on the horizontal time axis of the two-dimensional coordinate system are converted into timestamp formats consistent with the historical load data sequence. The timestamps of the historical load data sequence are continuous time identifiers within a preset historical time window, corresponding one-to-one with the load data. After completing the timestamp interval conversion of all core time periods, a unique start and end timestamp is formed for each core time period, clarifying the time range definition standard of each core time period and providing a unified time matching basis for subsequent data extraction.

[0050] Then, a full traversal of the historical load data sequence within the preset historical time window is performed. The timestamp and corresponding load monitoring data of each data point in the historical load data sequence are read sequentially according to time. The timestamp of each data point is compared with the start and end timestamps of each core period to determine whether the timestamp is within the timestamp interval of a certain core period. If the timestamp is within the timestamp interval of a certain core period, the load monitoring data of the data point is marked as the data to be extracted for the corresponding core period. If it is not within the timestamp interval of any core period, it is not marked. After the full traversal is completed, the data set to be extracted corresponding to each core period is obtained.

[0051] Next, the data sets to be extracted for each core time period are processed for time series processing. That is, the load monitoring data in each data set to be extracted are reordered in ascending order of timestamps to ensure the temporal continuity of the data in each set and to maintain consistency with the time granularity of the historical load data sequence. After sorting, the ordered load monitoring data sequence corresponding to each core time period is used as the historical load data subsequence of that core time period. Each historical load data subsequence is assigned a unique identifier that corresponds one-to-one with the core time period to distinguish the load data characteristics of different core time periods. Finally, the historical load data subsequence of each core time period is obtained, matching the number of core time periods.

[0052] Step S203: Input the historical load data sequence and multiple historical load data subsequences into the target load prediction model to obtain the load prediction values ​​within a preset time period. The target load prediction model is used to extract the first historical multi-scale interactive enhancement features of the historical load data sequence, extract the second historical multi-scale interactive enhancement features of each historical load data subsequence, and extract the time-series dependency features of the historical load data sequence. Based on the first historical multi-scale interactive enhancement features, all second historical multi-scale interactive enhancement features, and time-series dependency features, a fused feature vector is generated, and the fused feature vector is mapped to the load prediction values.

[0053] In this embodiment of the invention, the first historical multi-scale interactive enhancement feature of the historical load data sequence and the second historical multi-scale interactive enhancement feature of each historical load data subsequence can be extracted through the TCN (Temporal Convolutional Network) in the target load prediction model. The temporal dependency feature of the historical load data sequence can be extracted through the xLSTM (Extended Long Short-Term Memory) in the target load prediction model. Different features can be weighted through CA (Convolutional Attention) in the target load prediction model. Based on the first historical multi-scale interactive enhancement feature, all second historical multi-scale interactive enhancement features, and temporal dependency features, a fused feature vector is generated. Then, the fused feature vector is mapped to the load prediction value within a preset time period (such as a certain hour, a certain day, or a certain month) through the output layer.

[0054] The TCN branch is a multi-scale feature extraction branch that relies on multi-scale convolutional structures to extract base station load change features at different time scales, in order to characterize the intraday fluctuations, intraweekly periodic changes, and long-term trend characteristics of the load. The xLSTM branch uses a dual-mechanism approach—scalar gating of sLSTM (Scalar Long Short-Term Memory, an improved version of LSTM that enhances stability and performance by introducing exponential gating and normalization techniques) and matrix parallelism of mLSTM (Matrix Long Short-Term Memory, an extension and improvement of LSTM that enhances the model's memory and expressive capabilities by using matrix memory)—to model long-term time-series dependencies in load data. This branch can capture long-term load evolution patterns and alleviate the gradient vanishing problem in deep modeling. Specifically, the xLSTM branch first uses a preprocessing network layer to convert historical load data sequences into multi-dimensional time feature vector sequences. Then, sLSTM performs time-step scalar gating on this vector sequence to generate the first-stage time-series hidden state. Simultaneously, mLSTM performs matrix-based state mapping and parallel update processing on this hidden state, ultimately obtaining long-term time-series dependency features. The CA branch specifically focuses on the key periods and core features that influence load forecasting. By allocating causal attention weights, it strengthens the contribution of key period features such as peaks and troughs to the forecast results and weakens the feature focus on irrelevant noise interference.

[0055] Before inputting historical load data sequences into the xLSTM branch and performing dual-mechanism collaborative feature analysis using the sLSTM and mLSTM of the xLSTM branch to obtain long-term time-series dependency features, the overall construction of the xLSTM branch needs to be completed. This branch adopts a dual-mechanism collaborative architecture in which a preprocessing network layer, an sLSTM scalar gating module, and an mLSTM matrix parallel module are sequentially connected. Data dimension matching and unidirectional ordered transmission are achieved between the modules through a fully connected layer. The overall architecture is adapted to the long-term time-series characteristics of regional communication base station load data, taking into account both fine-grained feature capture and long-period regularity memory. The preprocessing network consists of two fully connected layers and two batch normalization layers. The first fully connected layer has 128 neurons, and the second layer has 64 neurons. ReLU is used as the activation function for both layers, and the Dropout rate is set to 0.2. This transforms the one-dimensional load data sequence into a multi-dimensional feature vector to meet the input dimension requirements of subsequent modules. The sLSTM scalar gating module consists of three sLSTM layers, each with 64 hidden units. The sigmoid activation function is used for gating, and tanh is used for output activation. The Dropout rate is set to 0.2. This scalar gating mechanism enables fine-grained time-step filtering of temporal features. The mLSTM matrix parallel module consists of two mLSTM layers, each with 128 hidden units. Softmax is used for matrix mapping activation, and tanh is used for state update activation. The Dropout rate is set to 0.2. This matrix parallel operation architecture enables synchronous multi-dimensional feature updates. At the module junction, a single-output fully connected layer is set up to map the 64-dimensional feature vector output by the sLSTM scalar gate module to 128 dimensions, achieving precise dimensional matching with the mLSTM matrix parallel module and ensuring the effectiveness of feature transfer.

[0056] If the entire model (i.e., the load prediction model including the temporal convolutional network, extended long short-term memory network, feature fusion layer, and output layer) is trained directly from scratch, the initial output will be extremely inaccurate due to the random parameters, and the gradient may be huge or extremely small, causing the model to learn nothing or converge very slowly. Instead, each module in the model can be pre-trained separately to learn how to extract effective features from the original data. Then, the final target load prediction model can be obtained through end-to-end joint training of the entire model.

[0057] Optionally, the training process for the xLSTM branch can be performed as follows: First, a training dataset is constructed, selecting continuous historical load data sequences of communication base stations in the target area over the past twelve months as the original data. Outliers in the data are removed using the 3σ principle, where σ is the standard deviation of the data. Then, the data is preprocessed using the Z-score standardization method to eliminate dimensional differences. Subsequently, the preprocessed dataset is divided into training, validation, and test sets in a 7:2:1 ratio. The training set is sampled using a time-series sliding window sampling method, with the sampling window length consistent with the preset historical time window and a sliding window step size of 1 to ensure the temporal continuity of the training samples. The training objective is to minimize the feature deviation between the long-term temporal dependency features output by the model and the true long-term temporal features of the load. True long-term temporal features such as long-term load trends, periodic evolution, and abrupt change correlations extracted by professional time-series feature extraction tools are used as labeled values.

[0058] Then, the core training parameters were set. The Adam optimizer was selected, with an initial learning rate of 0.001. A cosine annealing decay strategy was used to dynamically adjust the learning rate, with T_max set to 100, the lower bound of the learning rate set to 0.0001, the batch size set to 32, and the training period set to 200 epochs. The mean squared error loss function was selected, and L2 regularization was added with a weight decay coefficient of 0.0005 to avoid overfitting. Gradient optimization employed a gradient pruning strategy with a pruning threshold of 1.0 to prevent gradient explosion during training. During training, the backpropagation chain rule is used to pass the loss value from the mLSTM matrix parallel module to the sLSTM scalar gating module, and then to the preprocessing network layer. The weights and bias parameters of each layer are updated sequentially. The validation set is used to monitor the model training effect in real time. When the loss function value of the validation set no longer decreases for 20 consecutive epochs, or when the training cycle reaches the preset 200 epochs (rounds), the training is terminated immediately, the optimal parameters of the model at this time are saved, and the training of the entire xLSTM branch is completed.

[0059] By inputting the historical load data sequence of regional communication base stations within a preset historical time window into the xLSTM branch preprocessing network layer, the transformation from a one-dimensional data sequence to a multi-dimensional time feature vector sequence is completed. This historical load data sequence is one-dimensional time-series data. It is first input into the first fully connected layer of the preprocessing network layer, mapping the one-dimensional data into a 128-dimensional feature vector. After nonlinear transformation using the ReLU activation function, 20% of the neurons are randomly discarded through a Dropout layer. Then, a batch normalization layer normalizes the feature vector, eliminating numerical differences between different dimensions. Next, the processed 128-dimensional feature vector is input into the second fully connected layer, mapping it into a 64-dimensional feature vector. The ReLU activation, Dropout, and batch normalization operations are repeated, ultimately converting the one-dimensional historical load data sequence into a 64-dimensional multi-dimensional time feature vector sequence. The time steps of this vector sequence are completely consistent with the original historical load data sequence, providing suitable input data for subsequent scalar gating processing. Next, the multi-dimensional temporal feature vector sequence output from the preprocessed network layer is input into the sLSTM scalar gating module to perform time-step scalar gating processing, generating the first-stage time series hidden state. Following the time-step sequence, the 64-dimensional feature vector of each time step is sequentially input into three sLSTM layers. Each sLSTM layer uses a scalar gating mechanism consisting of an input gate, a forget gate, and an output gate to filter and memorize the input features of each time step. Specifically, the forget gate filters and discards historical feature information unrelated to long-term temporal patterns, the input gate selects and retains the effective load feature information of the current time step, and the output gate integrates the feature information from the memory unit and outputs the hidden state of the current time step. After time-step scalar gating processing by three sLSTM layers, the output of the last sLSTM layer is integrated according to the time-step sequence to generate the first-stage time series hidden state with a dimension of 64. This hidden state fully preserves the fine-grained temporal variation features of the regional communication base station load data, while mitigating the gradient vanishing problem in deep modeling through the scalar gating mechanism. Finally, the first-stage time-series hidden state output from the sLSTM scalar-gated module is input into the dimension-matching fully connected layer. This first performs feature dimension mapping and transformation, and then inputs it into the mLSTM matrix parallel module to perform matrix-based state mapping and parallel update processing, obtaining long-term temporal dependency features. The dimension-matching fully connected layer maps the 64-dimensional scalar hidden state into a 128-dimensional feature vector. After a non-linear transformation using the ReLU activation function, this vector is input into the mLSTM matrix parallel module. This module converts the scalar feature vector into matrix-based state features and uses matrix parallel computation to synchronously update the matrix state at each time step in multiple dimensions. Matrix mapping enhances the ability to capture long-term temporal dependencies.Two mLSTM layers sequentially update the matrix state in parallel. Each layer filters the effective load features in the long time series through a matrix gating mechanism and discards noise features. Finally, the output of the last mLSTM layer is integrated in time step order and transformed nonlinearly by the tanh activation function to obtain the long time series dependency features of the target communication base station load data. These features cover the long-term evolution trend, periodic changes, cross-time period correlations and other long time series information of the load data.

[0060] By constructing a preprocessing network layer, a dual-mechanism collaborative xLSTM branch, and completing the system training of the model, and then using fully connected layer mapping, scalar gating time-step filtering, and matrix parallel multi-dimensional updates to achieve feature transformation and analysis of historical load data sequences, the long-term temporal dependencies of regional communication base station load data are effectively captured, the model's ability to stably remember long-term load evolution patterns is strengthened, and the gradient vanishing problem in the deep modeling process is alleviated.

[0061] Step S204: Standardize the historical load data series to obtain the mean and standard deviation.

[0062] In this embodiment of the invention, the standardization process can be performed using the Z-score standardization method, which can calculate the mean and standard deviation of all data points in the historical load data sequence.

[0063] Step S205: Determine the target load value of the target communication base station based on the mean, standard deviation, and load prediction values.

[0064] In this embodiment of the invention, the target load value can be calculated by the following formula: target load value = load prediction value × standard deviation + mean. This value is the power load of the target communication base station in the target area during the prediction period (e.g., day-ahead or short-term hourly level), and the unit of measurement is kilowatts, corresponding to the actual power load value of the base station equipment operation.

[0065] In summary, by collecting historical load data sequences of target communication base stations within the target area, and processing them through core time period analysis and multi-scale feature extraction to obtain various load characteristics, and analyzing long-term load dependencies, prediction is performed by combining multi-branch output feature fusion, thereby accurately obtaining base station load prediction results. Furthermore, the historical load data sequences are standardized to obtain the mean and standard deviation. Based on the mean, standard deviation, and load prediction values, the target load values ​​of the target communication base stations are calculated, making the regional communication base station load prediction results more accurate and reliable, improving the accuracy and stability of communication base station load prediction, and thus solving the technical problem of low accuracy in base station load prediction in related technologies.

[0066] To improve the accuracy of the target load prediction model, in the method for determining the regional communication base station load provided in Embodiment 1 of this application, the network structure of the target load prediction model includes: a temporal convolutional network, an extended long short-term memory network, a feature fusion layer, and an output layer. The temporal convolutional network includes at least a first branch and a second branch. The first branch extracts the first historical multi-scale interactive enhancement features of the historical load data sequence, and the second branch extracts the second historical multi-scale interactive enhancement features of each historical load data subsequence. The extended long short-term memory network extracts the temporal dependency features of the historical load data sequence. The feature fusion layer includes at least an attention weight allocation module and a fully connected layer. The attention weight allocation module weights the first historical multi-scale interactive enhancement features, the second historical multi-scale interactive enhancement features, and the temporal dependency features. The fully connected layer fuses and compresses the weighted features to generate a fused feature vector. The output layer maps the fused feature vector to a one-dimensional load prediction value.

[0067] In this embodiment of the invention, the network structure of the target load prediction model includes: a temporal convolutional network, an extended long short-term memory (LSTM) network, a feature fusion layer, and an output layer. The LSTM network includes at least a first branch and a second branch. The first branch can extract the first historical multi-scale interaction enhancement features of the historical load data sequence, and the second branch can extract the second historical multi-scale interaction enhancement features of each historical load data subsequence. The historical load data sequence is then input into the LSTM network, and long-term temporal dependency features are obtained through dual-mechanism collaborative feature analysis using the LSTM and mLSTM of the LSTM network.

[0068] Dimensional normalization can be performed on the first historical multi-scale interaction enhancement features, multiple second historical multi-scale interaction enhancement features, and temporal dependent features. For the original dimensional differences among the three types of features, independent dimension-mapping fully connected layers are constructed, each with 64 neurons, using ReLU activation and a dropout rate of 0.2. The three types of features are input into their respective dimension-mapping fully connected layers. Through linear transformation and non-linear activation operations, all features are uniformly mapped to a 64-dimensional feature space, achieving consistency calibration of feature dimensions. This ensures effective information exchange among the various features during the fusion process, laying the foundation for subsequent feature fusion.

[0069] The feature fusion layer employs a structure design that combines attention-weighted features with a fully connected layer. First, the three types of features, after dimensionality standardization, are sequentially concatenated to generate a 192-dimensional concatenated feature vector. This vector is then input into the attention weight allocation module of the feature fusion layer. This module redistributes the weights of feature components at different time periods and scales within the concatenated feature vector, strengthening the weight ratio of core time-period interaction enhancement features and long-term dependent features, while weakening the influence of irrelevant feature components. The attention-weighted feature vector is then input into the fully connected layer of the feature fusion layer. This layer has 128 neuron units and uses ReLU as the activation function. Nonlinear fusion and information compression are performed on the weighted features to generate a 64-dimensional fused feature vector. This vector integrates the core information of the three types of features, comprehensively representing the multi-scale changes and long-term evolution patterns of regional communication base station load.

[0070] The 64-dimensional fused feature vector output from the feature fusion layer is then input into the load prediction layer (i.e., the output layer) to generate the regional communication base station load prediction result. This prediction layer employs a two-layer fully connected structure. The first fully connected layer has 32 neurons, uses ReLU as the activation function, initializes the weights using Xavier (a standard method for initializing neural network weights), initializes the bias term to 0, and sets the dropout rate to 0.2, compressing and mapping the 64-dimensional fused feature vector into a 32-dimensional feature vector. The second fully connected layer has one neuron, no additional activation function, and initializes the weights and bias term using the same rules, mapping the 32-dimensional feature vector into a one-dimensional normalized predicted value (i.e., the load prediction value).

[0071] Optionally, before using the target load prediction model for prediction, it needs to be trained. End-to-end joint training can be adopted to achieve collaborative optimization of parameters of each branch, feature fusion layer, and prediction layer. Specifically, historical load data of communication base stations in the target area are selected as training samples. The sample data is input into the overall model in time sequence to complete multi-scale feature extraction and interaction enhancement of the TCN branch, long-term time-series dependency feature analysis of the xLSTM branch, weight allocation of the CA branch, and feature fusion of the feature fusion layer. Throughout the process, the parameters of each branch and feature fusion layer are passed in series. The Adam optimizer is selected as the training optimizer, with an initial learning rate set to 0.001. The learning rate is dynamically adjusted using a cosine annealing strategy, the batch size is set to 32, and the mean squared error loss function is selected. The deviation between the load prediction value output by the prediction layer and the actual load value is used as the basis for loss calculation. During the training process, the backpropagation chain rule is adopted to synchronously update the network parameters of all layers and branches. The model training effect is monitored in real time using the validation set. Training is terminated when the validation set loss value does not decrease for 20 consecutive epochs, the optimal parameters of the model are saved, and the joint training of the overall model is completed.

[0072] To accurately determine multiple target time periods, the method for determining the load of a regional communication base station provided in Embodiment 1 of this application traverses a historical load data sequence and performs time-series gradient analysis on each historical load data in the historical load data sequence to obtain load data gradient values ​​at multiple time points; compares each load data gradient value with a preset time-series gradient threshold to obtain all time points where the load data gradient value is greater than or equal to the preset time-series gradient threshold; constructs a set of filtered time periods based on all time points; assigns values ​​to time points within a preset historical time window based on the set of filtered time periods to construct a two-dimensional assignment coordinate system, wherein the horizontal axis of the two-dimensional assignment coordinate system is the time axis, the vertical axis is the number of assignments, and the two-dimensional assignment coordinate system includes at least multiple assignment points; performs time period diffusion analysis on the multiple assignment points to determine multiple target time periods.

[0073] In this embodiment of the invention, the power load monitoring data of the communication base station within a preset historical time window is first collected by the power monitoring module and conventional data acquisition equipment such as smart meters, with an hourly time granularity. The collected load data is then arranged in the order of timestamps to form a continuous and time-series complete historical load data sequence. Each data point corresponds to a unique timestamp, providing basic data for the time-series analysis of subsequent core periods.

[0074] The first-order difference method of numerical gradient calculation is used to perform time-series gradient calculation on the historical load data sequence step by step (i.e. time point), and obtain the load data gradient value corresponding to each time step. The gradient value represents the rate of change of the load data at the corresponding time step. The larger the absolute value of the gradient value, the more significant the load data fluctuation at that time step.

[0075] To obtain the preset time-series gradient threshold, the arithmetic mean and standard deviation of the gradient values ​​for all time steps can be calculated first. Then, the mean plus 1.5 times the standard deviation is used as the preset time-series gradient threshold. Subsequently, the load data gradient value for each time step is compared with the preset time-series gradient threshold one by one. All time steps with gradient values ​​greater than or equal to the threshold are selected. Adjacent time steps that meet the selection criteria are merged to form continuous time segments. All such continuous time segments together form a set of selected time periods, thereby completing the initial screening of periods with significant fluctuations in base station load.

[0076] Then, based on the set of filtered time periods, a two-dimensional time point assignment method is used for each time point within the preset historical time window. First, the initial assignment value of all time points is set to 0. Then, all time periods in the set of filtered time periods are traversed, and the time points contained in each time period are counted one by one. Each time point is incremented by 1 for each time point that appears in a filtered time period. Through this counting method, each time point in the window will get a unique assignment value. The larger the assignment value of a time point, the more times that time point appears in periods of significant load fluctuation, the higher the frequency of load data fluctuation, and the greater the possibility that the time period to which that time point belongs is a core time period. After assigning values ​​to all time points, a two-dimensional assignment coordinate system is constructed. This system uses the time axis of a preset historical time window as the horizontal axis, with each scale point corresponding to a time point within the window. The vertical axis uses the value assigned at each time point as the vertical axis, with the scale equally divided according to the natural number sequence of the assigned values. The timestamps and corresponding assigned values ​​of all time points are used as coordinate parameters to mark all assignment points in the two-dimensional coordinate system. The coordinate position of each assignment point uniquely corresponds to the time characteristics and load fluctuation characteristics of that time point. Subsequently, time period diffusion analysis is performed on multiple assignment points to determine multiple target time periods.

[0077] Optionally, while performing each step of the core time period time series analysis, the CA (attention weight allocation module) branch is fully constructed and trained based on the operation process of each step. Regarding the construction of the CA branch: First, an attention weight calculation layer is constructed. This layer adopts a two-layer fully connected network structure. The input is the gradient value obtained from time series gradient analysis and the assigned value obtained from the two-dimensional time point assignment. The hidden layer has 32 neurons, the activation function is the Sigmoid function, and the output is the attention weight value in the range of 0 to 1, realizing a non-linear mapping from load fluctuation characteristics to attention weights. Second, a causal attention mask module is designed. This module constructs a lower triangular mask matrix based on the number of time points within a preset historical time window. The matrix dimension is consistent with the number of time points. The lower triangular elements of the matrix are 1, and the upper triangular elements are 0. This mask matrix is ​​applied to the output of the attention weight calculation layer, retaining only the attention weight information of the current time point and historical time points, and masking information of future time points, ensuring the temporal causality of attention weight allocation. Finally, an attention focusing module was designed, which consists of a weight normalization layer and a feature weighting layer. The weight normalization layer uses the Softmax function to normalize the weight values ​​output by the attention weight calculation layer, so that the sum of the attention weights at all time points is 1. The feature weighting layer multiplies the normalized attention weights element by element with the corresponding base station load data features, thereby strengthening the weights of features in the core time period and naturally weakening the features in the non-core time period.

[0078] In terms of CA branch training, the model parameters are first initialized. The weight parameters of the attention weight calculation layer are initialized using the Xavier normal initialization method, and the bias parameters are uniformly initialized to 0. The matrix parameters of the causal attention mask module are fixed values ​​and do not need to participate in training. The weight parameters of the attention focusing module are initialized using a random uniform initialization method in the range of 0 to 1. Next, attention weight mapping rules are constructed. In the temporal gradient analysis stage, the gradient values ​​obtained by the first-order difference method are mapped to the range of 0 to 1 using the minimum-maximum normalization method. The normalized gradient values ​​are used as the initial attention weights at that time point. In the two-dimensional time point assignment stage, the assigned values ​​at each time point are also mapped to the range of 0 to 1 using the minimum-maximum normalization method. The normalized assigned values ​​are then multiplied element-wise with the initial attention weights at that time point to obtain the corrected attention weight values, thus realizing a two-dimensional weight mapping based on load fluctuation rate and fluctuation frequency.

[0079] Then, an iterative optimization strategy was constructed, using stochastic gradient descent as the optimizer. The learner rate was set to 0.001, the batch size to 32, and the maximum number of training epochs to 100. An early stopping mechanism was also set up: if the loss function value does not decrease significantly after 10 consecutive training epochs, the training is terminated early to avoid model overfitting. Finally, a loss function was designed and training iterations were completed. The mean squared error loss function was used as the training loss function for the CA branch. The set of selected time periods obtained from the core time period time series analysis was used as the training labels. The label values ​​of time points within the selected time periods were set to 1, and the label values ​​of time points outside the selected time periods were set to 0. The attention weights output by the CA branch were used as the predicted values. The mean squared error between the label values ​​and the predicted values ​​was calculated as the loss function value. Model training was carried out synchronously in each step of the core time period time series analysis. After the time series gradient analysis, the initial attention weights were calculated and mapped. After the two-dimensional time point assignment, the attention weights were corrected. After the two-dimensional assignment coordinate system was constructed, the causal attention mask module was adapted. One round of training iterations was completed before the time period diffusion analysis. The iterations were repeatedly optimized until the loss function value converged to the preset minimum value of 0.001. At this time, the core training of the CA branch was completed. The trained CA branch can achieve accurate attention focusing on the core time period of base station load.

[0080] By combining numerical statistics and time series analysis methods, the preliminary analysis of the core period of regional communication base station load and the construction of a two-dimensional assigned coordinate system were completed. At the same time, based on the operation steps of the core period analysis, the basic construction and core training of the CA branch were completed, which enabled the accurate preliminary screening of the period of significant base station load fluctuation. It also laid a complete parameter and rule foundation for the CA branch to focus on the core impact period in load prediction.

[0081] To improve the accuracy of determining multiple target time periods, in the method for determining the load of a regional communication base station provided in Embodiment 1 of this application, multiple assignment points are screened based on the maximum value of the assignment number to determine multiple initial diffusion assignment points; each initial diffusion assignment point is used as a starting point, and based on each starting point, the assignment points are diffused over time periods to obtain initial diffusion time periods; based on a preset time period length threshold, all initial diffusion time periods are screened to determine multiple target time periods.

[0082] Optionally, the maximum values ​​of the horizontal coordinates of multiple assignment points in the two-dimensional assignment coordinate system are screened to determine multiple initial diffusion assignment points. The local neighborhood maximum traversal method can be used for the screening operation. Combined with the hourly time granularity of the regional communication base station load data, the local neighborhood of each assignment point is set as the range of assignment points corresponding to the left and right of the horizontal time axis. All assignment points are traversed in ascending order of the horizontal coordinate. The vertical coordinate assignment value of each assignment point is compared with the assignment values ​​of other assignment points in its local neighborhood. If the assignment value of the assignment point is the maximum value in its local neighborhood, the assignment point is marked as an initial diffusion assignment point. If it is not the maximum value in the local neighborhood, it is not marked. After completing the traversal and comparison of all assignment points, all marked assignment points are integrated to form an initial diffusion assignment point set. The assignment points in this set are all the peak points of load fluctuation within the local time range in the two-dimensional assignment coordinate system, which are the key starting points for load core period analysis. Then, each initial diffusion assignment point in the initial diffusion assignment point set is taken as the starting point, and based on each starting point, the assignment point is diffused over a time period to obtain the initial diffusion time period.

[0083] Then, based on the actual operating patterns of load fluctuations at the target communication base stations, the minimum duration of effective fluctuation periods in historical load data is statistically analyzed. This minimum duration is set as a preset period length threshold, which is compatible with the hourly time granularity of the load data, ensuring that the selected periods are effective periods with actual load fluctuation significance. Subsequently, each initial diffusion period in the initial diffusion period set is traversed, and the duration of each initial diffusion period on the horizontal time axis is extracted. Initial diffusion periods with a duration greater than or equal to the preset period length threshold are retained, while initial diffusion periods with a duration less than the preset period length threshold are directly removed to avoid interference from short-duration invalid periods caused by data noise on the core period analysis results. After completing the screening of all initial diffusion periods, all retained initial diffusion periods are integrated to finally determine multiple core periods of regional communication base station load.

[0084] Optionally, during the entire core time period parsing process, which involves filtering for the maximum horizontal coordinate of the assignment points, and filtering for time period diffusion and duration thresholds, the CA branch and the core time period parsing steps are simultaneously adapted and trained. After the initial diffusion assignment points are selected, the attention focusing module of the CA branch strengthens the attention weights corresponding to these assignment points, increasing the feature weight ratio of the core starting point. During multiple rounds of neighborhood diffusion, the CA branch dynamically adjusts the attention weights of each time point in the neighborhood based on the distribution density of assignment points within the diffusion neighborhood. The higher the density of assignment points, the greater the weight adjustment, achieving feature focusing on areas with dense load fluctuations. After completing the core time period selection, the CA branch normalizes the attention weights of all time points within each core time period, forming the overall attention weight distribution for each core time period. This ensures that the attention weights output by the CA branch perfectly match the final determined core time periods, providing accurate weight support for subsequent load feature extraction. Through the above steps, combined with the CA branch, the dynamic adjustment, strengthening, and normalization of attention weights for core assignment points and core time periods are achieved, accurately identifying the core fluctuation periods of regional communication base station load and completing the attention weight focusing of features corresponding to the core time periods.

[0085] To accurately determine the initial diffusion period, in the method for determining the regional communication base station load provided in Embodiment 1 of this application, for each starting point, diffusion is performed on the starting point based on a preset diffusion bandwidth to obtain an initial diffusion neighborhood; based on the preset diffusion bandwidth, neighborhood edge diffusion is performed on the initial diffusion neighborhood to obtain a stage diffusion neighborhood; based on a preset time period diffusion stopping strategy, the initial diffusion neighborhood and the stage diffusion neighborhood are analyzed to obtain analysis results; if the analysis result indicates that diffusion has stopped, the time period corresponding to the initial diffusion neighborhood is determined as the initial diffusion period; if the analysis result indicates that diffusion continues, diffusion is continued on the stage diffusion neighborhood based on the preset diffusion bandwidth until the analysis result indicates that diffusion has stopped, and the time period corresponding to the stage diffusion neighborhood is determined as the initial diffusion period.

[0086] In this embodiment of the invention, a preset diffusion bandwidth is first set for each initial diffusion assignment point (i.e., the starting point). This bandwidth can be set to 2 hours, taking into account the hourly time granularity of the regional communication base station load data. This corresponds to the interval between two time points on the horizontal time axis of the two-dimensional assignment coordinate system. A coordinate traversal method is used to perform bidirectional diffusion operations on multiple initial diffusion assignment points. Specifically, based on the time axis number of each initial diffusion assignment point on the horizontal axis, two adjacent time point numbers are traversed to the left and right respectively. The assignment points corresponding to all the traversed time point numbers are integrated into a neighborhood. Each initial diffusion assignment point generates an independent neighborhood range. After completing the diffusion operation for all initial diffusion assignment points, an initial diffusion neighborhood corresponding one-to-one with each initial diffusion assignment point is obtained. Each initial diffusion neighborhood has a clearly defined horizontal time axis number range, which can be directly mapped to the actual time interval.

[0087] After obtaining multiple initial diffusion neighborhoods, the preset diffusion bandwidth remains unchanged, and the coordinate traversal method is used to perform neighborhood edge diffusion operations on each initial diffusion neighborhood to obtain multiple stage diffusion neighborhoods. Specifically, the left and right edge time point numbers of each initial diffusion neighborhood on the horizontal time axis are extracted first. The two edge point numbers are used as new diffusion starting points, and unidirectional traversal diffusion is performed outward from the initial diffusion neighborhood according to the preset diffusion bandwidth. The left edge point traverses two time point numbers to the left, and the right edge point traverses two time point numbers to the right. The time point numbers covered by this edge diffusion are merged with the time point numbers of the original initial diffusion neighborhood to form a new neighborhood range. Each initial diffusion neighborhood corresponds to a stage diffusion neighborhood, ensuring the continuity and integrity of the stage diffusion neighborhood on the time axis.

[0088] After constructing multiple initial diffusion neighborhoods and stage diffusion neighborhoods, a preset time-period diffusion stopping mechanism (i.e., a preset time-period diffusion stopping strategy) is invoked to perform mapping analysis on the two, thereby determining the execution method of subsequent diffusion operations. Specifically, all initial diffusion neighborhoods and their corresponding stage diffusion neighborhoods are first mapped to a two-dimensional assignment coordinate system. According to their respective time axis numbering ranges, the number of assignment points in each initial diffusion neighborhood and the number of assignment points in the corresponding stage diffusion neighborhood are precisely counted. After completing the one-to-one correspondence of assignment point counts, the statistical results are input into a preset time-period diffusion cessation mechanism for judgment and analysis. The preset time-period diffusion cessation mechanism can be as follows: when the number of assignment points in multiple initial diffusion neighborhoods is greater than or equal to the number of assignment points in multiple corresponding stage diffusion neighborhoods, the analysis result is diffusion cessation; when the number of assignment points in multiple initial diffusion neighborhoods is less than the number of assignment points in multiple corresponding stage diffusion neighborhoods, and the difference in the number of assignment points is less than or equal to a preset difference threshold, the analysis result is diffusion cessation; when the number of assignment points in multiple initial diffusion neighborhoods is less than the number of assignment points in multiple corresponding stage diffusion neighborhoods, and the difference in the number of assignment points is greater than a preset difference threshold, the analysis result is diffusion cessation.

[0089] Before adopting the preset time period diffusion cessation mechanism, it is necessary to first combine the actual data characteristics of the target communication base station load core time period parsing and use statistical analysis to determine the preset difference threshold. For example, statistical analysis can be used to determine the difference distribution of the number of assignment points between the initial diffusion neighborhood and the stage diffusion neighborhood in the historical load data during multiple time period diffusion processes, calculate the arithmetic mean of all differences, and set this average value as the preset difference threshold.

[0090] When the preset diffusion stop mechanism is invoked, the number of assignment points for each initial diffusion neighborhood and its corresponding stage diffusion neighborhood is first counted. According to the time axis number range in the two-dimensional assignment coordinate system, all assignment points in the neighborhood are traversed one by one and counted. The total number of assignment points in each initial diffusion neighborhood and the total number of assignment points in the corresponding stage diffusion neighborhood are recorded. After the number of all neighborhood pairs is counted, the number of assignment points in the initial diffusion neighborhood and the stage diffusion neighborhood are counted one by one, providing accurate basic data for subsequent layer determination.

[0091] Based on the above statistical results, the first-level judgment of the execution period diffusion cessation mechanism directly compares the number of assigned points in each initial diffusion neighborhood with the number of assigned points in the corresponding stage diffusion neighborhood. If the number of assigned points in an initial diffusion neighborhood is greater than or equal to the number of assigned points in its corresponding stage diffusion neighborhood, it is determined that the diffusion operation of the initial diffusion assigned point has reached the cessation condition. At this time, continuing diffusion will include low assigned points with no load fluctuation characteristics into the neighborhood, which will dilute the characteristics of the core fluctuation period. Therefore, the subsequent diffusion operation of the initial diffusion assigned point is directly terminated, and the corresponding initial diffusion period is determined based on the current initial diffusion neighborhood.

[0092] If, after the first-level determination, the number of assigned points in the initial diffusion neighborhood is less than the number of assigned points in its corresponding stage diffusion neighborhood, then the difference calculation method is used to calculate the difference between the two. The number of assigned points in the stage diffusion neighborhood is subtracted from the number of assigned points in the initial diffusion neighborhood to obtain the absolute difference in the number of assigned points between the two neighborhoods. Then, this absolute difference is compared with a preset difference threshold, and the second-level determination of the time-period diffusion stop mechanism is executed. If the calculated difference in quantity is less than or equal to the preset difference threshold, it indicates that the stage diffusion only includes a small number of newly added assignment points, and the core load fluctuation area has been basically covered. Continuing diffusion cannot effectively extract more core features, and diffusion is judged to stop. The corresponding initial diffusion period is determined based on the current stage diffusion neighborhood. If the calculated difference in quantity is greater than the preset difference threshold, it indicates that the current stage diffusion includes a large number of newly added assignment points. Continuing diffusion will include a large number of non-core fluctuation assignment points in the neighborhood, causing the core period features to be diluted. Diffusion is also judged to stop, and subsequent diffusion operations are terminated. The diffusion neighborhood of the previous stage is determined as the initial diffusion period. If the previous stage is the initial diffusion neighborhood, then the initial diffusion neighborhood is determined as the initial diffusion period.

[0093] After obtaining the diffusion cessation result through the preset diffusion cessation mechanism analysis, the horizontal axis time number range corresponding to each initial diffusion neighborhood is converted into an actual time interval, which is then used as the initial diffusion period corresponding to each initial diffusion assignment point. If the mechanism analysis results in continued diffusion, the current stage diffusion neighborhood is used as the new diffusion basis, and the edge diffusion operation is continued according to the preset diffusion bandwidth to generate a new stage diffusion neighborhood. The steps of neighborhood mapping, assignment point number statistics, and mechanism analysis are repeated until the time period diffusion cessation mechanism outputs the diffusion cessation analysis result. Then, the time interval corresponding to the previous stage diffusion neighborhood of the final diffusion neighborhood is determined as the initial diffusion period, completing the time period diffusion operation for all initial diffusion assignment points.

[0094] By constructing a hierarchical diffusion cessation mechanism, the cessation determination of the diffusion operation at the initial diffusion assignment point is completed in stages. At the same time, relying on CA branch, the attention weight is adjusted in real time and finally optimized during the determination process. The diffusion range of the initial diffusion assignment point is accurately controlled, effectively avoiding the problems of insufficient or excessive diffusion. This provides accurate diffusion neighborhood range and optimized attention weight support for the precise selection of subsequent core time periods.

[0095] Optionally, during the entire process of performing multiple rounds of neighborhood diffusion on the initial diffusion assignment points and generating the initial diffusion period, the CA branch and diffusion steps are simultaneously trained for adaptation and the weights are dynamically updated. After generating the initial diffusion neighborhood, the attention weight calculation layer of the CA branch performs a basic allocation of attention weights for all assignment points within the neighborhood, dividing the weights proportionally according to the assignment values ​​of the assignment points to ensure basic weight coverage of load fluctuation characteristics within the neighborhood. After generating the stage diffusion neighborhood, the CA branch compares the changes in the number of assignment points between the initial diffusion neighborhood and the stage diffusion neighborhood, initializes the attention weights of newly assigned points, and slightly strengthens the weights of high-value assignment points in the original neighborhood to highlight the characteristics of core fluctuation points. During the iterative diffusion process, the CA branch captures the neighborhood changes after each round of diffusion in real time, dynamically adjusts the attention weights of each assignment point within the neighborhood, and reduces the weight proportion of low-value noise points included in the diffusion process. After diffusion stops and the initial diffusion period is determined, the CA branch performs a normalization calibration on all attention weights within each initial diffusion period, so that the sum of the weights within the period is 1, ensuring that the attention weight distribution of the initial diffusion period is highly matched with the load fluctuation characteristics.

[0096] In this embodiment of the invention, the initial diffusion assignment point is multi-round neighborhood diffusion is realized by coordinate traversal method. The initial diffusion period is determined by combining neighborhood mapping statistics and iterative judgment. At the same time, the basic allocation, dynamic adjustment and normalization calibration of attention weights during diffusion are realized by relying on CA branch. The initial diffusion period corresponding to the load fluctuation characteristics of regional communication base stations is accurately obtained, and the attention weight distribution of load characteristics within the initial diffusion period is optimized.

[0097] In order to accurately obtain the first historical multi-scale interactive enhancement features and the second historical multi-scale interactive enhancement features, in the method for determining the regional communication base station load provided in Embodiment 1 of this application, features are extracted from the historical load data sequence through the first branch to obtain a historical multi-scale feature set; interactive enhancement is performed on the historical multi-scale feature set to obtain the first historical multi-scale interactive enhancement features; features are extracted from multiple historical load data sub-sequences through the second branch to obtain multiple target time period historical multi-scale feature sets; interactive enhancement is performed on each target time period historical multi-scale feature set to obtain the second historical multi-scale interactive enhancement features.

[0098] Optionally, the overall construction of the temporal convolutional network (TCN branch) needs to be completed first. This branch includes parallel feature extraction networks for the entire sequence and core time period subsequences, adopting a multi-scale convolutional structure as the core, and constructing a feature interaction enhancement module. The multi-scale convolutional structure sets three different kernel sizes: 3, 5, and 7, corresponding to capturing the short, medium, and long-term temporal scale changes in base station load. The dilation rates of each convolutional layer are set to 1, 2, and 4 respectively. By expanding the convolution, the receptive field is increased, allowing coverage of longer temporal data without increasing the number of network layers. Each convolutional layer is configured with residual connections, adding the input and output of the convolutional layer element-wise before inputting it into the activation function, mitigating the gradient vanishing problem in deep networks and ensuring the effectiveness of feature extraction. The parallel feature extraction network contains two structurally identical branches: a full-sequence feature extraction branch (i.e., the first branch), used to process the complete historical load data sequence; and a core time period subsequence feature extraction branch (i.e., the second branch), used to process historical load data subsequences for each core time period. The two branches share the parameters of the multi-scale convolutional structure, improving the consistency and efficiency of feature extraction. The integration of the TCN branch and the feature interaction enhancement module is achieved by using a pooling layer plus a fully connected layer. The features output by the multi-scale convolutional structure are compressed in dimension by a global average pooling layer, and then mapped by the fully connected layer to the feature dimension that matches the feature interaction enhancement module, thus achieving seamless feature transfer.

[0099] After constructing the TCN branch, systematic training can begin. First, convolutional layer parameters are initialized. All weights are initialized using the Xavier initialization method to ensure a uniform distribution and gradient stability during forward and backward propagation. Bias parameters are uniformly initialized to 0. Pooling and fully connected layers use the same initialization rule. The multi-scale feature extraction training strategy employs mini-batch stochastic gradient descent, with a batch size of 32 and an initial learning rate of 0.001. A learning rate decay strategy is used, multiplying the learning rate by 0.9 every 100 iterations until it reaches 0.0001 and remains constant. Training samples are selected from historical load data of regional communication base stations and their corresponding labeled features. The gradient update rule for the feature interaction enhancement stage uses a backpropagation chain rule, propagating the enhanced feature loss value backward to the fully connected, pooling, and multi-scale convolutional layers of the TCN branch, updating the parameters of each layer sequentially. A gradient clipping threshold of 1.0 is set to prevent gradient explosion during training.

[0100] The training loss function uses the mean squared error (MSE) loss function, calculating the MSE between the multi-scale features extracted by the TCN branch and the actual load features. This MSE is used as the training loss metric. Training is terminated when the loss function value no longer decreases and converges to 0.0001 after 20 consecutive iterations, completing the entire training process of the TCN branch. Simultaneously, the core parameters for feature processing in the TCN branch are determined. The ReLU activation function is selected for the convolutional analysis kernel to enhance the non-linear expressive power of the features. The feature combination sampling rule is set to continuous sampling with a stride of 1 to ensure the temporal continuity of the features. The scaling factor for normalization is set to 0.1 to distribute the feature values ​​within a reasonable range, improving training efficiency.

[0101] Optionally, the historical load data sequence and multiple historical load data subsequences can be preprocessed using standardization. This standardization preprocessing can employ the Z-score standardization method, which calculates the mean and standard deviation of all data points in the historical load data sequence. The standardized value of each data point is equal to the value of that data point minus the mean, then divided by the standard deviation. The historical load data subsequences can be standardized using the same mean and standard deviation as the historical load data sequence, eliminating the impact of dimensional differences on feature extraction. The standardized historical load data sequence is then input into the full-sequence feature extraction branch (i.e., the first branch). After convolution, activation, residual connection, and pooling processing using a multi-scale convolutional structure, the features extracted by different convolution kernels and dilation rates are concatenated according to dimension to obtain a historical multi-scale feature set. Next, each standardized historical load data subsequence is sequentially input into the core time period subsequence feature extraction branch (i.e., the second branch), using the same feature extraction and concatenation methods to generate a corresponding target time period historical multi-scale feature set for each core time period historical load data subsequence.

[0102] Then, interactive enhancement operations can be performed on the extracted historical multi-scale feature sets to obtain the first historical multi-scale interactive enhanced features. The same interactive enhancement operations can also be performed on the historical multi-scale feature sets of multiple target time periods to obtain multiple second historical multi-scale interactive enhanced features.

[0103] By constructing a TCN branch that includes a multi-scale convolutional structure with residual connections and a parallel feature extraction network, and combining Xavier initialization and mini-batch stochastic gradient descent to complete the system training of the TCN branch, feature extraction of historical load data sequences and core time period subsequences is achieved through Z-score normalization and multi-scale convolutional concatenation. Finally, two types of interactive enhancement features are obtained through unified interactive enhancement operations, which effectively captures the changing characteristics of regional communication base station load at different time scales, strengthens the correlation between features, and improves the expressive power of load features.

[0104] To accurately obtain the first historical multi-scale interactive enhancement features, in the method for determining the regional communication base station load provided in Embodiment 1 of this application, every two historical multi-scale features in the historical multi-scale feature set are combined to generate multiple historical multi-scale feature combinations to be interacted, wherein the historical multi-scale feature combinations to be interacted correspond to a generation order; according to the generation order, interactive enhancement is performed on each historical multi-scale feature combination to be interacted to obtain the first-stage historical multi-scale interactive enhancement features, and a first-stage historical multi-scale interactive enhancement feature set is constructed based on all the first-stage historical multi-scale interactive enhancement features; the mean of the first-stage historical multi-scale interactive enhancement feature set is calculated to obtain the first historical multi-scale interactive enhancement features.

[0105] In this embodiment of the invention, the historical multi-scale feature set extracted by the TCN branch is preprocessed. First, the number of features and the dimensionality of each feature are analyzed to unify the dimensionality specifications of all features and ensure the feasibility of feature combinations. Then, a pairwise enumeration method without repetition is used to enumerate and combine the historical multi-scale feature set. Following the order of features in the set, different features are selected sequentially for pairwise combinations, without repeating the same feature pair. After completing the combination matching of all features, each feature combination is treated as an independent feature unit to be interacted with. All feature units to be interacted with are integrated to obtain multiple historical multi-scale feature combinations to be interacted with, achieving comprehensive association matching between features with different scale loads.

[0106] Following the generation order of the historical multi-scale feature combinations to be interacted with, an interaction enhancement operation is performed on each historical multi-scale feature combination to be interacted with, resulting in the first-stage historical multi-scale interaction enhancement features. Then, an initially empty set of first-stage historical multi-scale interaction enhancement features is pre-constructed. The obtained first-stage historical multi-scale interaction enhancement features are added to this set according to the extraction order, completing the storage of the interaction enhancement results of a single set of features. The first-stage historical multi-scale interaction enhancement features obtained by interaction enhancement of other historical multi-scale feature combinations to be interacted with are all added to this set in the same way to ensure the integrity and orderliness of the features within the set.

[0107] Then, the arithmetic mean method is used to calculate the mean of all features in the first-stage historical multi-scale interactive enhancement feature set. First, the common dimension of all features in the set is determined. Then, the mean of all feature values ​​in each dimension is calculated according to the dimension order of the features. The mean results of each dimension are integrated in order to form a new feature vector with the same dimension as the original feature. This feature vector is the first historical multi-scale interactive enhancement feature obtained after interactive enhancement. The information fusion of multiple sets of interactive enhancement features is realized through mean calculation, thereby improving the comprehensive expressive ability of the features.

[0108] By employing a pairwise enumeration method without repetition, a sequential extraction method, and an arithmetic mean method, interactive enhancement processing of historical multi-scale feature sets was completed. This achieved comprehensive correlation and information fusion among base station load features at different scales, effectively strengthening the expressive power of historical multi-scale features and the correlation between features.

[0109] To accurately determine the first-stage historical multi-scale interactive enhancement features, in the method for determining the regional communication base station load provided in Embodiment 1 of this application, similarity calculation is performed on features belonging to the same feature type in the historical multi-scale feature combination to be interacted, resulting in an interactive enhancement similarity set. The feature type is the category to which each historical multi-scale feature in the historical multi-scale feature set belongs, and the category is determined by the value of the convolution kernel in the first branch. Each interactive enhancement similarity in the interactive enhancement similarity set is normalized, and each normalized interactive enhancement similarity is filled into a preset initial matrix to obtain an interactive enhancement matrix. Based on the interactive enhancement matrix, a convolution operation is performed on each historical multi-scale feature combination to be interacted, resulting in an interactive historical multi-scale feature combination. Based on the feature dimension, multiple means of the interactive historical multi-scale feature combination are calculated to obtain the mean feature, and the mean feature is determined as the first-stage historical multi-scale interactive enhancement feature.

[0110] In this embodiment of the invention, the feature types of the multi-scale feature combination of the interaction history to be performed first. This clarifies the feature types included in the combination, such as intraday fluctuations, intraweekly cyclical changes, and long-term trends (corresponding to the short, medium, and long-term timescale changes in base station load captured by three different sets of convolutional kernels (e.g., 3, 5, and 7)). Sub-features within the combination are then classified according to feature type, grouping sub-features of the same type together to ensure that the dimensions and feature attributes of sub-features within the same group are completely consistent. The cosine similarity calculation method is used to calculate the similarity of each group of sub-features of the same type. For each pair of sub-feature vectors within a group, the cosine value of the angle between them is calculated, and this cosine value is used as the similarity value between the two sub-features. The closer the cosine value is to 1, the higher the similarity of the sub-features; the closer it is to 0, the lower the similarity. After completing the similarity calculation for all sub-features of the same type within the multi-scale feature combination of the interaction history to be performed, all the obtained similarity values ​​are integrated according to the order of feature type to form an interaction enhancement similarity set, providing a quantitative basis for feature association in subsequent matrix construction.

[0111] Normalization is performed on each interaction enhancement similarity in the interaction enhancement similarity set. The min-max normalization method can be used to map all similarity values ​​in the set to a numerical range of 0 to 1. That is, the maximum and minimum values ​​of similarity values ​​in the set are calculated, the minimum value is subtracted from each similarity value, and then the result is divided by the difference between the maximum and minimum values ​​to obtain the normalized result for each similarity value. This method eliminates the numerical differences between different similarity values ​​and improves the stability of subsequent matrix operations.

[0112] An initial empty matrix (i.e., a preset initial matrix) can be pre-constructed to match the feature dimensions of the multi-scale feature combination of the history to be interacted with. The number of rows and columns of the matrix corresponds to the number of dimensions and feature types of the sub-features, respectively. According to the order of feature types, the normalized similarity values ​​are sequentially filled into the corresponding positions of the initial empty matrix. After all values ​​are filled, the interaction enhancement matrix is ​​obtained, which can accurately represent the degree of correlation between different sub-features.

[0113] Then, the interaction enhancement matrix is ​​used as the convolution kernel, and the one-dimensional convolution analysis method is used to perform convolution operation on the multi-scale feature combination of the interaction history. The convolution stride is set to 1, so that the convolution kernel slides continuously on the feature vector of the feature combination. At each sliding position, the value in the convolution kernel is multiplied element-wise with the value at the corresponding position of the feature vector, and then all the multiplication results are summed to obtain the convolution calculation value at that position. After completing the convolution sliding calculation of the entire feature vector, the convolution feature vector is obtained.

[0114] Finally, the above one-dimensional convolution analysis operation is performed sequentially on all sub-feature vectors within the multi-scale feature combination of the interaction history. All convolved feature vectors are then integrated to obtain the multi-scale feature combination of the interaction history. The arithmetic mean method is used to calculate the mean of this feature combination. The average value of each feature vector within the combination is calculated according to its corresponding dimension. The mean results of each dimension are then integrated in the order of the feature dimensions to form a new feature vector with the same dimensions as the original feature vector. This feature vector is the first-stage historical multi-scale interaction enhancement feature.

[0115] By completing the interactive enhancement processing of multi-scale feature combinations of interaction history, the correlation degree between sub-features of the same type within the feature combination is quantified and the feature is enhanced, which effectively improves the nonlinear expression capability and temporal correlation of multi-scale features of base station load, and provides more discernible feature support for subsequent load feature fusion and prediction.

[0116] Figure 3 This is a flowchart of a regional communication base station load prediction method that integrates multi-scale temporal features according to an embodiment of this application, such as... Figure 3As shown, firstly, historical load data sequences of regional communication base stations within a preset historical time window are collected. Time series parsing of core time periods is performed to determine multiple core time periods (i.e., target time periods). Based on these multiple core time periods, data extraction is performed on the historical load data sequences to obtain multiple core time period historical load data subsequences. Multi-scale feature extraction is then performed on the historical load data sequences and the multiple core time period historical load data subsequences through the TCN branch, and interactive enhancement is performed separately to obtain historical multi-scale interactive enhancement features (i.e., the first historical multi-scale interactive enhancement features) and multiple core time period historical interactive enhancement features (i.e., the second historical multi-scale interactive enhancement features). The historical load data sequences are then input into the xLSTM branch, and dual-mechanism collaborative feature analysis is performed using the sLSTM and mLSTM mechanisms of the xLSTM branch to obtain long-term time-series dependency features. Load prediction is then performed based on the historical multi-scale interactive enhancement features, the multiple core time period historical interactive enhancement features, and the long-term time-series dependency features to obtain the regional communication base station load prediction results.

[0117] Figure 4 This is a structural diagram of a regional communication base station load prediction system that integrates multi-scale temporal features according to an embodiment of this application, as shown below. Figure 4 As shown, the system includes: a core time period acquisition module, a historical load data acquisition module, an interaction enhancement feature acquisition module, a long-term time-series dependency feature acquisition module, and a load prediction result acquisition module. The core time period acquisition module collects historical load data sequences of regional communication base stations within a preset historical time window, performs core time period time-series parsing, and determines multiple core time periods. The historical load data acquisition module extracts data from the historical load data sequences based on multiple core time periods to obtain multiple core time period historical load data sub-sequences. The interaction enhancement feature acquisition module performs multi-scale feature extraction on the historical load data sequences and multiple core time period historical load data sub-sequences through a TCN branch, and performs interaction enhancement respectively to obtain historical multi-scale interaction enhancement features and multiple core time period historical interaction enhancement features. The long-term time-series dependency feature acquisition module inputs the historical load data sequences into an xLSTM branch, and performs dual-mechanism collaborative feature analysis using sLSTM and mLSTM in the xLSTM branch to obtain long-term time-series dependency features. The load prediction result acquisition module performs load prediction based on historical multi-scale interaction enhancement features, multiple core time period historical interaction enhancement features, and long-term time-series dependency features to obtain the regional communication base station load prediction result.

[0118] The method for determining the load of regional communication base stations provided in this application can accurately analyze and focus on the core period of load through the CA branch, solving the problem of ambiguous positioning of key influencing factors. It relies on the TCN branch to extract multi-scale load features and complete interactive enhancement, making up for the deficiency of insufficient multi-scale feature capture. Through the dual mechanism of xLSTM branch for collaborative analysis, it solves the problems of weak long-term dependency modeling ability and gradient vanishing in deep training. Finally, it integrates three types of features to carry out load prediction, which improves the accuracy and stability of regional communication base station load prediction, clearly represents the load change pattern at different time scales, and provides accurate technical support for base station energy management, resource allocation and scheduling decisions. It can adapt to the actual engineering needs of day-ahead / short-term load prediction.

[0119] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0120] Example 2

[0121] This application also provides an apparatus for determining the load of a regional communication base station. It should be noted that this apparatus can be used to execute the method for determining the load of a regional communication base station provided in this application. The apparatus for determining the load of a regional communication base station provided in this application will be described below.

[0122] According to an embodiment of this application, an apparatus for implementing the above-described method for determining the load of a regional communication base station is also provided. Figure 5 This is a schematic diagram of a device for determining the load of a regional communication base station according to an embodiment of this application, as shown below. Figure 5 As shown, the device for determining the load of communication base stations in the area may include: a parsing unit 50, an extraction unit 51, an input unit 52, a processing unit 53, and a determination unit 54.

[0123] The parsing unit 50 is used to collect the historical load data sequence of the target communication base station in the target area within a preset historical time window, and to parse the historical load data sequence to obtain multiple target time periods within the preset historical time window.

[0124] Extraction unit 51 is used to extract data from historical load data sequences based on multiple target time periods to obtain multiple historical load data subsequences;

[0125] Input unit 52 is used to input historical load data sequences and multiple historical load data subsequences into the target load prediction model to obtain load prediction values ​​within a preset time period. The target load prediction model is used to extract the first historical multi-scale interactive enhancement features of the historical load data sequence, extract the second historical multi-scale interactive enhancement features of each historical load data subsequence, and extract the time-series dependency features of the historical load data sequence. Based on the first historical multi-scale interactive enhancement features, all second historical multi-scale interactive enhancement features, and time-series dependency features, a fused feature vector is generated, and the fused feature vector is mapped to the load prediction value.

[0126] Processing unit 53 is used to standardize the historical load data sequence to obtain the mean and standard deviation;

[0127] The determination unit 54 is used to determine the target load value of the target communication base station based on the mean, standard deviation and load prediction value.

[0128] The regional communication base station load determination device provided in this application embodiment can collect historical load data sequences of target communication base stations in a target area within a preset historical time window through the parsing unit 50, and parse the historical load data sequences to obtain multiple target time periods within the preset historical time window. The extraction unit 51 can extract data from the historical load data sequences based on the multiple target time periods to obtain multiple historical load data subsequences. The input unit 52 can input the historical load data sequences and multiple historical load data subsequences into the target load prediction model to obtain the load prediction value within the preset time period. The processing unit 53 can perform standardization processing on the historical load data sequences to obtain the mean and standard deviation. The determination unit 54 can determine the target load value of the target communication base station based on the mean, standard deviation, and load prediction value.

[0129] Optionally, the network structure of the target load prediction model includes: a temporal convolutional network, an extended long short-term memory network, a feature fusion layer, and an output layer. The temporal convolutional network includes at least a first branch and a second branch. The first branch extracts the first historical multi-scale interactive enhancement features of the historical load data sequence, and the second branch extracts the second historical multi-scale interactive enhancement features of each historical load data subsequence. The extended long short-term memory network extracts the temporal dependency features of the historical load data sequence. The feature fusion layer includes at least an attention weight allocation module and a fully connected layer. The attention weight allocation module weights the first historical multi-scale interactive enhancement features, the second historical multi-scale interactive enhancement features, and the temporal dependency features. The fully connected layer fuses and compresses the weighted features to generate a fused feature vector. The output layer maps the fused feature vector to a one-dimensional load prediction value.

[0130] Optionally, the parsing unit 50 includes: a first analysis module, used to traverse the historical load data sequence and perform time-series gradient analysis on each historical load data in the historical load data sequence to obtain load data gradient values ​​at multiple time points; a first comparison module, used to compare each load data gradient value with a preset time-series gradient threshold to obtain all time points where the load data gradient value is greater than or equal to the preset time-series gradient threshold; a first construction module, used to construct a set of filtered time periods based on all time points; a second construction module, used to assign values ​​to time points in a preset historical time window based on the set of filtered time periods to construct a two-dimensional assignment coordinate system, wherein the horizontal axis of the two-dimensional assignment coordinate system is the time axis and the vertical axis is the number of assignments, and the two-dimensional assignment coordinate system includes at least multiple assignment points; and a first determination module, used to perform time period diffusion analysis on the multiple assignment points to determine multiple target time periods.

[0131] Optionally, the first determining module includes: a first filtering submodule, used to filter multiple assignment points based on the maximum value of the assignment number to determine multiple initial diffusion assignment points; a first diffusion submodule, used to take each initial diffusion assignment point as a starting point and perform time-period diffusion on the assignment points based on each starting point to obtain an initial diffusion time period; and a second filtering submodule, used to filter all initial diffusion time periods based on a preset time period length threshold to determine multiple target time periods.

[0132] Optionally, the first diffusion submodule includes: a first diffusion subunit, used to diffuse each starting point based on a preset diffusion bandwidth to obtain an initial diffusion neighborhood; a second diffusion subunit, used to perform neighborhood edge diffusion on the initial diffusion neighborhood based on the preset diffusion bandwidth to obtain a staged diffusion neighborhood; a first analysis subunit, used to analyze the initial diffusion neighborhood and the staged diffusion neighborhood based on a preset time-period diffusion stopping strategy to obtain analysis results; a first determination subunit, used to determine the time period corresponding to the initial diffusion neighborhood as the initial diffusion time period when the analysis result indicates that diffusion should stop; and a second determination subunit, used to continue diffusion on the staged diffusion neighborhood based on the preset diffusion bandwidth when the analysis result indicates that diffusion should continue, until the analysis result indicates that diffusion should stop, and to determine the time period corresponding to the staged diffusion neighborhood as the initial diffusion time period.

[0133] Optionally, the device for determining the load of regional communication base stations further includes: a first extraction module, used to extract features from historical load data sequences through a first branch to obtain a historical multi-scale feature set; a first enhancement module, used to perform interactive enhancement on the historical multi-scale feature set to obtain a first historical multi-scale interactive enhancement feature; a second extraction module, used to extract features from multiple historical load data sub-sequences through a second branch to obtain multiple historical multi-scale feature sets for target time periods; and a second enhancement module, used to perform interactive enhancement on the historical multi-scale feature set for each target time period to obtain a second historical multi-scale interactive enhancement feature.

[0134] Optionally, the first enhancement module includes: a first combination submodule, used to combine every two historical multi-scale features in the historical multi-scale feature set to generate multiple historical multi-scale feature combinations to be interacted with, wherein the historical multi-scale feature combinations to be interacted with correspond to a generation order; a first enhancement submodule, used to perform interactive enhancement on each historical multi-scale feature combination to be interacted with according to the generation order to obtain first-stage historical multi-scale interactive enhancement features, and to construct a first-stage historical multi-scale interactive enhancement feature set based on all first-stage historical multi-scale interactive enhancement features; and a first calculation submodule, used to calculate the mean of the first-stage historical multi-scale interactive enhancement feature set to obtain the first historical multi-scale interactive enhancement feature.

[0135] Optionally, the first enhancement submodule includes: a first calculation subunit, used to calculate the similarity of features belonging to the same feature type in the historical multi-scale feature combination to be interacted, to obtain an interaction enhancement similarity set, wherein the feature type is the category to which each historical multi-scale feature in the historical multi-scale feature set belongs, and the category is determined by the value of the convolution kernel in the first branch; a first processing subunit, used to normalize each interaction enhancement similarity in the interaction enhancement similarity set, and fill each normalized interaction enhancement similarity into a preset initial matrix to obtain an interaction enhancement matrix; a first operation subunit, used to perform convolution operation on each historical multi-scale feature combination to be interacted based on the interaction enhancement matrix to obtain an interaction historical multi-scale feature combination; and a second calculation subunit, used to calculate multiple means of the interaction historical multi-scale feature combination based on the feature dimension to obtain the mean feature, and determine the mean feature as the first-stage historical multi-scale interaction enhancement feature.

[0136] The aforementioned device for determining the load of regional communication base stations may further include a processor and a memory. The parsing unit 50, extraction unit 51, input unit 52, processing unit 53, determination unit 54, etc., are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0137] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and by adjusting kernel parameters, the target load value for the target communication base station can be determined based on the mean, standard deviation, and load prediction values.

[0138] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0139] It should be noted that the parsing unit 50, extraction unit 51, input unit 52, processing unit 53, and determining unit 54 mentioned above correspond to steps S201 to S205 in Embodiment 1. The instances and application scenarios implemented by the above units and corresponding steps are the same, but are not limited to the content disclosed in Embodiment 1. It should be noted that the above units can be hardware or software components stored in memory (e.g., memory 104) and processed by one or more processors (e.g., processors 102a, 102b, ..., 102n). The above units can also be part of a device and can run in the computer terminal 10 provided in Embodiment 1.

[0140] Example 3

[0141] Embodiments of this application may provide a computer terminal, which may be any computer terminal device in a group of computer terminals. Optionally, in this embodiment, the aforementioned computer terminal may also be replaced with a mobile terminal or an electronic device, etc.

[0142] Optionally, in this embodiment, the computer terminal may be located in at least one of a plurality of network devices in a computer network.

[0143] In this embodiment, the computer terminal described above can execute the program code for the following steps in the method for determining the load of a regional communication base station: collecting historical load data sequences of the target communication base station in the target area within a preset historical time window, and parsing the historical load data sequences to obtain multiple target time periods within the preset historical time window; extracting data from the historical load data sequences based on the multiple target time periods to obtain multiple historical load data subsequences; inputting the historical load data sequences and multiple historical load data subsequences into the target load prediction model to obtain load prediction values ​​within the preset time period, wherein the target load prediction model is used to extract the first historical multi-scale interactive enhancement features of the historical load data sequences, extract the second historical multi-scale interactive enhancement features of each historical load data subsequence, and extract the time-series dependency features of the historical load data sequences, and generate a fused feature vector based on the first historical multi-scale interactive enhancement features, all second historical multi-scale interactive enhancement features, and the time-series dependency features, and map the fused feature vector to load prediction values; standardizing the historical load data sequences to obtain the mean and standard deviation; and determining the target load value of the target communication base station based on the mean, standard deviation, and load prediction values.

[0144] Optionally, the aforementioned computer terminal can execute program code for the following steps in the method for determining the load of a regional communication base station: The network structure of the target load prediction model includes: a temporal convolutional network, an extended long short-term memory network, a feature fusion layer, and an output layer. The temporal convolutional network includes at least: a first branch and a second branch. The first branch extracts the first historical multi-scale interactive enhancement features of the historical load data sequence, and the second branch extracts the second historical multi-scale interactive enhancement features of each historical load data sub-sequence. The extended long short-term memory network extracts the temporal dependency features of the historical load data sequence. The feature fusion layer includes at least: an attention weight allocation module and a fully connected layer. The attention weight allocation module weights the first historical multi-scale interactive enhancement features, the second historical multi-scale interactive enhancement features, and the temporal dependency features. The fully connected layer fuses and compresses the weighted features to generate a fused feature vector. The output layer maps the fused feature vector to a one-dimensional load prediction value.

[0145] Optionally, the aforementioned computer terminal can execute the program code for the following steps in the method for determining the load of a regional communication base station: traversing the historical load data sequence and performing time-series gradient analysis on each historical load data in the historical load data sequence to obtain load data gradient values ​​at multiple time points; comparing each load data gradient value with a preset time-series gradient threshold to obtain all time points where the load data gradient value is greater than or equal to the preset time-series gradient threshold; constructing a set of filtered time periods based on all time points; assigning values ​​to time points within a preset historical time window based on the set of filtered time periods to construct a two-dimensional assignment coordinate system, wherein the horizontal axis of the two-dimensional assignment coordinate system is the time axis, the vertical axis is the number of assignments, and the two-dimensional assignment coordinate system includes at least: multiple assignment points; performing time period diffusion analysis on the multiple assignment points to determine multiple target time periods.

[0146] Optionally, the aforementioned computer terminal may execute the following steps in the method for determining the load of a regional communication base station: based on the maximum value of the assignment number, filter multiple assignment points to determine multiple initial diffusion assignment points; take each initial diffusion assignment point as a starting point, and based on each starting point, diffuse the assignment points over time periods to obtain initial diffusion time periods; based on a preset time period length threshold, filter all initial diffusion time periods to determine multiple target time periods.

[0147] Optionally, the aforementioned computer terminal may execute the following steps in the method for determining the load of a regional communication base station: For each starting point, based on a preset diffusion bandwidth, the starting point is diffused to obtain an initial diffusion neighborhood; based on the preset diffusion bandwidth, the initial diffusion neighborhood is diffused at its neighborhood edges to obtain a staged diffusion neighborhood; based on a preset time-period diffusion cessation strategy, the initial diffusion neighborhood and the staged diffusion neighborhood are analyzed to obtain an analysis result; if the analysis result indicates that diffusion has stopped, the time period corresponding to the initial diffusion neighborhood is determined as the initial diffusion time period; if the analysis result indicates that diffusion should continue, based on the preset diffusion bandwidth, the staged diffusion neighborhood is continued to diffuse until the analysis result indicates that diffusion has stopped, and the time period corresponding to the staged diffusion neighborhood is determined as the initial diffusion time period.

[0148] Optionally, the aforementioned computer terminal may execute the program code for the following steps in the method for determining the load of a regional communication base station: extracting features from the historical load data sequence through a first branch to obtain a historical multi-scale feature set; performing interactive enhancement on the historical multi-scale feature set to obtain a first historical multi-scale interactive enhancement feature; extracting features from multiple historical load data sub-sequences through a second branch to obtain multiple target time period historical multi-scale feature sets; and performing interactive enhancement on each target time period historical multi-scale feature set to obtain a second historical multi-scale interactive enhancement feature.

[0149] Optionally, the aforementioned computer terminal may execute the following steps in the method for determining the load of a regional communication base station: combining every two historical multi-scale features in the historical multi-scale feature set to generate multiple historical multi-scale feature combinations to be interacted, wherein the historical multi-scale feature combinations to be interacted correspond to a generation order; performing interactive enhancement on each historical multi-scale feature combination to be interacted according to the generation order to obtain the first-stage historical multi-scale interactive enhancement features, and constructing a first-stage historical multi-scale interactive enhancement feature set based on all the first-stage historical multi-scale interactive enhancement features; calculating the mean of the first-stage historical multi-scale interactive enhancement feature set to obtain the first historical multi-scale interactive enhancement feature.

[0150] Optionally, the aforementioned computer terminal may execute the following steps in the method for determining the load of a regional communication base station: calculate the similarity of features belonging to the same feature type in the historical multi-scale feature combination to be interacted, to obtain an interaction enhancement similarity set, wherein the feature type is the category to which each historical multi-scale feature in the historical multi-scale feature set belongs, and the category is determined by the value of the convolution kernel in the first branch; normalize each interaction enhancement similarity in the interaction enhancement similarity set, and fill each normalized interaction enhancement similarity into a preset initial matrix to obtain an interaction enhancement matrix; based on the interaction enhancement matrix, perform convolution operation on each historical multi-scale feature combination to be interacted, to obtain an interaction historical multi-scale feature combination; based on the feature dimension, calculate multiple means of the interaction historical multi-scale feature combination to obtain the mean feature, and determine the mean feature as the first-stage historical multi-scale interaction enhancement feature.

[0151] Optionally, Figure 6 This is a structural block diagram of an electronic device according to an embodiment of this application. Figure 6 As shown, the electronic device may include: one or more ( Figure 6 (Only one is shown) Processor 602, memory 604, memory controller, and peripheral interface, wherein the peripheral interface is connected to the radio frequency module, audio module and display.

[0152] The memory can be used to store software programs and modules, such as the program instructions / modules corresponding to the method and apparatus for determining the load of a regional communication base station in this embodiment. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory, thereby realizing the aforementioned method for determining the load of a regional communication base station. The memory may include high-speed random access memory, and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory may further include memory remotely located relative to the processor, and these remote memories can be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, corporate intranets, local area networks, mobile communication networks, and combinations thereof.

[0153] The processor can call the information and application program stored in the memory through the transmission device to execute the above steps in the method for determining the load of the regional communication base station.

[0154] The embodiments of this application provide a scheme for determining the load of regional communication base stations. By accurately capturing long-term time-series load dependencies, effectively extracting multi-scale load features, focusing on the core period of load prediction, the accuracy and stability of base station load prediction are improved, providing technical support for base station energy management, resource allocation and scheduling decisions, thereby solving the technical problem of low accuracy in predicting base station load in related technologies.

[0155] Those skilled in the art will understand that Figure 6 The structure shown is for illustrative purposes only. Electronic devices can also be terminal devices such as smartphones, tablets, PDAs, and mobile internet devices (MIDs). Figure 6 This does not limit the structure of the aforementioned electronic device. For example, electronic devices may also include components that are more... Figure 6 The more or fewer components shown (such as network interfaces, display devices, etc.), or having the same Figure 6 The different configurations shown.

[0156] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing the hardware related to the terminal device. The program can be stored in a computer-readable storage medium, which may include: flash drive, read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0157] Example 4

[0158] Embodiments of this application also provide a storage medium. Optionally, in this embodiment, the storage medium can be used to store the program code executed by the method for determining the load of a regional communication base station provided in Embodiment 1.

[0159] Optionally, in this embodiment, the storage medium may be located in any computer terminal in a group of computer terminals in a computer network, or in any mobile terminal in a group of mobile terminals.

[0160] This application also provides a computer program product, which, when executed on a data processing device, is suitable for performing the steps of a method for determining the load of a regional communication base station.

[0161] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0162] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0163] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0164] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0165] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0166] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard drive, magnetic disk, or optical disk.

[0167] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for determining the load of a regional communication base station, characterized in that, include: Collect historical load data sequences of target communication base stations in the target area within a preset historical time window, and parse the historical load data sequences to obtain multiple target time periods within the preset historical time window; Based on the multiple target time periods, data extraction is performed on the historical load data sequence to obtain multiple historical load data sub-sequences; The historical load data sequence and the multiple historical load data subsequences are input into the target load prediction model to obtain the load prediction value within a preset time period. The target load prediction model is used to extract the first historical multi-scale interaction enhancement feature of the historical load data sequence, extract the second historical multi-scale interaction enhancement feature of each historical load data subsequence, and extract the time-series dependency feature of the historical load data sequence. Based on the first historical multi-scale interaction enhancement feature, all the second historical multi-scale interaction enhancement features, and the time-series dependency feature, a fused feature vector is generated, and the fused feature vector is mapped to the load prediction value. The historical load data sequence is standardized to obtain the mean and standard deviation; The target load value of the target communication base station is determined based on the mean, the standard deviation, and the load prediction value.

2. The method for determining the load of a regional communication base station according to claim 1, characterized in that, The network structure of the target load prediction model includes: The system comprises a temporal convolutional network, an extended long short-term memory network, a feature fusion layer, and an output layer. The temporal convolutional network includes at least a first branch and a second branch. The first branch extracts the first historical multi-scale interaction enhancement feature of the historical load data sequence, and the second branch extracts the second historical multi-scale interaction enhancement feature of each historical load data sub-sequence. The extended long short-term memory network extracts the temporal dependency feature of the historical load data sequence. The feature fusion layer includes at least an attention weight allocation module and a fully connected layer. The attention weight allocation module weights the first historical multi-scale interaction enhancement feature, the second historical multi-scale interaction enhancement feature, and the temporal dependency feature. The fully connected layer fuses and compresses the weighted features to generate the fused feature vector. The output layer maps the fused feature vector to a one-dimensional load prediction value.

3. The method for determining the load of a regional communication base station according to claim 1, characterized in that, The step of parsing the historical load data sequence to obtain multiple target time periods within the preset historical time window includes: Traverse the historical load data sequence and perform time-series gradient analysis on each historical load data in the historical load data sequence to obtain load data gradient values ​​at multiple time points; Each load data gradient value is compared with a preset time-series gradient threshold to obtain all time points where the load data gradient value is greater than or equal to the preset time-series gradient threshold. Based on all the aforementioned time points, construct a set of filter time periods; Based on the set of filtered time periods, values ​​are assigned to time points in the preset historical time window to construct a two-dimensional assignment coordinate system. The horizontal axis of the two-dimensional assignment coordinate system is the time axis, and the vertical axis is the number of assignments. The two-dimensional assignment coordinate system includes at least multiple assignment points. Perform time-period diffusion analysis on the multiple assignment points to determine the multiple target time periods.

4. The method for determining the load of a regional communication base station according to claim 3, characterized in that, The steps of performing time-period diffusion analysis on the multiple assignment points to determine the multiple target time periods include: Based on the maximum value of the assigned number, the multiple assignment points are filtered to determine multiple initial diffusion assignment points; Each initial diffusion assignment point is used as a starting point, and based on each starting point, the assignment point is diffused over a time period to obtain the initial diffusion time period; Based on a preset time period length threshold, all the initial diffusion time periods are filtered to determine the multiple target time periods.

5. The method for determining the load of a regional communication base station according to claim 4, characterized in that, The step of performing time-period diffusion on the assigned points based on each starting point to obtain the initial diffusion time period includes: For each starting point, the starting point is diffused based on a preset diffusion bandwidth to obtain an initial diffusion neighborhood; Based on the preset diffusion bandwidth, the initial diffusion neighborhood is diffused at the neighborhood edge to obtain a staged diffusion neighborhood; Based on a preset diffusion cessation strategy, the initial diffusion neighborhood and the staged diffusion neighborhood are analyzed to obtain the analysis results; If the analysis result indicates that diffusion has stopped, the time period corresponding to the initial diffusion neighborhood is determined as the initial diffusion period. If the analysis result indicates continued diffusion, the diffusion of the stage diffusion neighborhood continues based on the preset diffusion bandwidth until the analysis result indicates cessation of diffusion, and the time period corresponding to the stage diffusion neighborhood is determined as the initial diffusion period.

6. The method for determining the load of a regional communication base station according to claim 2, characterized in that, The steps of extracting the first historical multi-scale interaction enhancement feature of the historical load data sequence through the first branch and extracting the second historical multi-scale interaction enhancement feature of each historical load data sub-sequence through the second branch include: The historical load data sequence is feature extracted by the first branch to obtain a historical multi-scale feature set. Perform interactive enhancement on the historical multi-scale feature set to obtain the first historical multi-scale interactive enhancement feature; By performing feature extraction on the multiple historical load data subsequences through the second branch, multiple target time period historical multi-scale feature sets are obtained; Interactive enhancement is performed on the historical multi-scale feature set for each target time period to obtain the second historical multi-scale interactive enhancement feature.

7. The method for determining the load of a regional communication base station according to claim 6, characterized in that, The step of performing interactive enhancement on the historical multi-scale feature set to obtain the first historical multi-scale interactive enhanced feature includes: Each pair of historical multi-scale features in the historical multi-scale feature set is combined to generate multiple historical multi-scale feature combinations to be interacted with, wherein the historical multi-scale feature combinations to be interacted with correspond to a generation order. According to the generation order, each of the proposed interactive historical multi-scale feature combinations is enhanced to obtain the first-stage historical multi-scale interactive enhancement features, and based on all the first-stage historical multi-scale interactive enhancement features, a first-stage historical multi-scale interactive enhancement feature set is constructed. Calculate the mean of the first stage historical multi-scale interaction enhancement feature set to obtain the first historical multi-scale interaction enhancement feature.

8. The method for determining the load of a regional communication base station according to claim 7, characterized in that, The steps of performing interaction enhancement on each of the proposed interactive historical multi-scale feature combinations to obtain the first-stage historical multi-scale interaction enhancement features according to the generation order include: Similarity calculation is performed on features of the same feature type in the historical multi-scale feature combination to be interacted with, to obtain an interaction enhancement similarity set, wherein the feature type is the category to which each historical multi-scale feature in the historical multi-scale feature set belongs, and the category is determined by the value of the convolution kernel in the first branch; Normalize each interaction enhancement similarity in the interaction enhancement similarity set, and fill each normalized interaction enhancement similarity into a preset initial matrix to obtain the interaction enhancement matrix; Based on the interaction enhancement matrix, a convolution operation is performed on each of the multi-scale feature combinations of the interaction history to be obtained; Based on the feature dimension, multiple mean values ​​of the interaction history multi-scale feature combination are calculated to obtain the mean feature, and the mean feature is determined as the first stage historical multi-scale interaction enhancement feature.

9. A device for determining the load of a regional communication base station, characterized in that, include: The parsing unit is used to collect historical load data sequences of target communication base stations in the target area within a preset historical time window, and to parse the historical load data sequences to obtain multiple target time periods of the preset historical time window; The extraction unit is used to extract data from the historical load data sequence based on the multiple target time periods to obtain multiple historical load data sub-sequences; An input unit is used to input the historical load data sequence and the multiple historical load data subsequences into a target load prediction model to obtain load prediction values ​​within a preset time period. The target load prediction model is used to extract a first historical multi-scale interaction enhancement feature of the historical load data sequence, extract a second historical multi-scale interaction enhancement feature of each historical load data subsequence, and extract a time-series dependency feature of the historical load data sequence. Based on the first historical multi-scale interaction enhancement feature, all second historical multi-scale interaction enhancement features, and the time-series dependency feature, a fused feature vector is generated, and the fused feature vector is mapped to the load prediction value. The processing unit is used to standardize the historical load data sequence to obtain the mean and standard deviation; The determining unit is used to determine the target load value of the target communication base station based on the mean, the standard deviation, and the load prediction value.

10. A computer program product, characterized in that, The method includes a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for determining the load of a regional communication base station as described in any one of claims 1 to 8.

11. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method for determining the load of a regional communication base station as described in any one of claims 1 to 8.

Images