Energy-saving optimization control method and system for communication base station

Abstract

The invention relates to the technical field of communication base station energy-saving control, and discloses an energy-saving optimization control method and system for a communication base station. The method comprises the following steps: collecting and preprocessing traffic energy consumption data of a communication base station to obtain standardized characteristics; inputting the features into a double-flow space-time convolutional network, and carrying out parallel calculation through time and space branches to predict a load; performing hierarchical quantitative pruning on the network, and distributing bit width according to the influence degree to obtain a compression structure; and processing the real-time data based on the compression network, and generating an energy-saving control strategy for radio frequency resource configuration and power regulation. Based on accurate load prediction and differentiated resource scheduling strategies, the optimal balance between energy saving and service quality is realized, the problems of passive response and coarse-grained control in a traditional method are effectively solved, and the energy saving effect is maximized on the premise of ensuring communication quality.

Description

Technical Field

[0001] The present application relates to the technical field of energy-saving control of communication base stations, and in particular to an energy-saving optimization control method and system for communication base stations. Background Art

[0002] With the rapid development of mobile communication networks and the widespread adoption of 5G technology, the number of communication base stations has increased dramatically, and energy consumption has become an increasingly prominent issue. According to statistics, base station energy consumption accounts for approximately 60%-80% of the total energy consumption of mobile communication networks, becoming a major component of operators' operating costs. Traditional base station energy-saving technologies mainly fall into two categories: hardware-level energy conservation and software-level energy conservation. Hardware-level energy conservation is primarily achieved through equipment improvements such as high-efficiency power amplifiers and natural cooling; software-level energy conservation is achieved through resource management methods such as carrier dormancy and power adjustment. Load-aware dynamic resource management has become a mainstream energy-saving method. It typically uses a preset threshold trigger method. When the base station load falls below a certain threshold, the energy-saving mechanism is automatically activated, reducing energy consumption by shutting down some radio frequency units and reducing transmit power.

[0003] However, existing base station energy-saving technologies have significant shortcomings. First, the traditional preset threshold trigger method can only respond passively to load changes and lacks foresight, resulting in resource adjustments lagging behind load changes, which can easily cause service quality fluctuations when the load changes rapidly. Second, fixed threshold strategies are difficult to adapt to the differences in load characteristics of different base stations and often require manual experience adjustment, lacking flexibility and adaptability. Third, existing energy-saving methods generally lack in-depth mining of load data and fail to fully utilize the temporal and spatial correlation characteristics of the load for accurate prediction and resource scheduling. Fourth, there is a coupling problem between hardware resources and control strategies, the accuracy of energy-saving control is limited, and it is difficult to achieve fine-grained energy consumption control. Finally, there is a contradiction between energy-saving effects and service quality. Excessive energy saving may lead to a decline in user experience, while conservative energy saving has limited effect.

[0004] To address the difficulty in balancing energy conservation and service quality due to insufficient load forecasting accuracy, an innovative neural network architecture simultaneously captures the temporal variation of load and the spatial correlation characteristics between parameters, significantly improving load forecasting accuracy. Furthermore, layered quantization and pruning techniques are used to lightweight the network, adapting to the resource constraints of base station processors and solving the challenge of deploying complex models at the edge. At the energy-saving control level, prediction-driven differentiated strategy generation and adaptive execution mechanisms are employed to achieve precise energy conservation while ensuring service quality. Summary of the Invention

[0005] This application provides an energy-saving optimization control method and system for communication base stations. By using a dual-stream spatiotemporal convolutional neural network to simultaneously capture the temporal variation of load and the spatial correlation characteristics between parameters, combined with layered quantization and pruning technology, this method significantly reduces model complexity while ensuring prediction accuracy, solving the challenge of deploying complex prediction models in base stations. Based on precise load prediction and differentiated resource scheduling strategies, it achieves an optimal balance between energy conservation and service quality, effectively overcoming the passive response and coarse-grained control issues of traditional methods, and maximizing energy savings while ensuring communication quality.

[0006] In the first aspect, the present application provides an energy-saving optimization control method for a communication base station, which includes: collecting and preprocessing the traffic data and energy consumption data of the communication base station to obtain standardized multidimensional feature data; inputting the standardized multidimensional feature data into a dual-stream spatiotemporal convolutional neural network, performing parallel calculations through the time feature processing branch and the spatial feature processing branch to obtain fused load prediction data; performing layered quantization and pruning on the dual-stream spatiotemporal convolutional neural network, allocating different bit widths according to the degree of influence of each network layer on the accuracy of the fused load prediction data, and obtaining a compressed network structure suitable for base station deployment; processing real-time load data based on the compressed network structure to generate a base station dynamic energy-saving control strategy including radio frequency resource configuration instructions and power adjustment parameters.

[0007] In a second aspect, the present application provides an energy-saving optimization control system for a communication base station, the energy-saving optimization control system for the communication base station comprising:

[0008] The acquisition module is used to collect and pre-process the traffic data and energy consumption data of the communication base station to obtain standardized multi-dimensional feature data;

[0009] An input module, configured to input the standardized multi-dimensional feature data into a dual-stream spatiotemporal convolutional neural network, and perform parallel calculations through a temporal feature processing branch and a spatial feature processing branch to obtain fused load prediction data;

[0010] A pruning module is used to perform layered quantization pruning on the dual-stream spatiotemporal convolutional neural network, allocating different bit widths according to the degree of influence of each network layer on the accuracy of the fused load prediction data, and obtaining a compressed network structure suitable for base station deployment;

[0011] A generation module is used to process real-time load data based on the compressed network structure and generate a base station dynamic energy-saving control strategy including radio frequency resource configuration instructions and power adjustment parameters.

[0012] In the third aspect, an energy-saving optimization control device for a communication base station is provided, comprising: a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the energy-saving optimization control device for the communication base station executes the above-mentioned energy-saving optimization control method for the communication base station.

[0013] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, enables the computer to execute the above-mentioned energy-saving optimization control method for a communication base station.

[0014] In the technical solution provided by this application, standardized multi-dimensional feature data is obtained by collecting and preprocessing the traffic data and energy consumption data of the communication base station, which solves the problems of strong heterogeneity and large noise interference of the original data, and lays a data foundation for subsequent accurate prediction. The standardized multi-dimensional feature data is input into the dual-stream spatiotemporal convolutional neural network, and parallel calculation is performed through the time feature processing branch and the spatial feature processing branch. It not only overcomes the limitation of the traditional single model that it is difficult to simultaneously process time series changes and feature interactions, but also effectively improves the computing efficiency through the dual-stream parallel architecture, realizes the comprehensive capture of load change rules and the relationship between features, greatly improves the accuracy of the fused load prediction data, and makes energy-saving control decisions more forward-looking. The dual-stream spatiotemporal convolutional neural network is layered and quantized and pruned, and different bit widths are allocated according to the degree of influence of each network layer on the accuracy of the fused load prediction data, to obtain a compressed network structure suitable for base station deployment, which effectively solves the problem that complex neural network models are difficult to deploy in base station environments with limited computing resources. While ensuring prediction accuracy, it significantly reduces storage space and computational complexity, enabling intelligent algorithms to run efficiently in edge computing environments. Based on the compressed network structure, real-time load data is processed to generate a base station dynamic energy-saving control strategy containing radio frequency resource configuration instructions and power adjustment parameters. Through differentiated resource scheduling and refined power control, the optimal balance between energy saving and service quality is achieved, avoiding the defects of service quality fluctuations or limited energy saving effects caused by the one-size-fits-all strategy in traditional methods. The actual load characteristics of the communication base station guide the design of the dual-stream spatiotemporal convolutional network, enabling the algorithm to extract the spatiotemporal pattern of the load in a targeted manner; the layered quantization pruning technology performs customized compression on the network according to the hardware constraints of the base station, so that the algorithm adapts to the deployment environment; the dynamic energy-saving control strategy formulates differentiated execution plans based on communication quality requirements and resource scheduling characteristics, and converts the algorithm output into actual energy-saving effects. The mutual support of this technical feature and the algorithm feature gives the present invention a significant advantage in solving the problem of energy saving in communication base stations, and can achieve a significant reduction in base station energy consumption while ensuring the quality of communication services. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0016] Figure 1 Schematic diagram of an embodiment of an energy-saving optimization control method for a communication base station in an embodiment of the present application;

[0017] Figure 2 This is a schematic diagram of an embodiment of an energy-saving optimization control system for a communication base station in an embodiment of the present application;

[0018] Figure 3 It is a schematic block diagram of the structure of an energy-saving optimization control device for a communication base station in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The present application embodiment provides an energy-saving optimization control method and system for a communication base station. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or precedence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0020] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 , an embodiment of the energy-saving optimization control method for a communication base station in the embodiment of the present application includes:

[0021] Step S101: collecting and preprocessing traffic data and energy consumption data of a communication base station to obtain standardized multi-dimensional feature data;

[0022] Step S102: Input the standardized multi-dimensional feature data into a dual-stream spatiotemporal convolutional neural network, perform parallel calculations through the temporal feature processing branch and the spatial feature processing branch, and obtain fused load prediction data;

[0023] Step S103: performing layered quantization and pruning on the dual-stream spatiotemporal convolutional neural network, allocating different bit widths according to the degree of influence of each network layer on the accuracy of the fused load prediction data, and obtaining a compressed network structure suitable for base station deployment;

[0024] Step S104: Process the real-time load data based on the compressed network structure to generate a base station dynamic energy-saving control strategy including radio frequency resource configuration instructions and power adjustment parameters.

[0025] It is understandable that the execution subject of this application can be an energy-saving optimization control system for a communication base station, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking a server as the execution subject as an example.

[0026] Specifically, the base station's RF power, channel usage, number of connected users, data transmission volume, power consumption, and temperature are recorded every minute. In practice, the data acquisition module retrieves this raw data from the base station management system interface and automatically stores it in a raw data table. Raw data often contains noise and missing values, necessitating data cleaning. A statistical threshold detection method is used to identify outliers outside the normal range, and then the values are replaced with the average values of the preceding and following time points. For example, if the RF power value at a specific moment is abnormally high, exceeding three times the average value of the preceding and following 10 minutes, it is replaced with the average value of the preceding and following data to ensure data continuity and rationality. The complete data series is then normalized. The mean value of each feature is subtracted from the corresponding feature value and then divided by its standard deviation to convert it to a zero-mean unit variance distribution, addressing the incomparability of data of different dimensions. For this normalized data, linear relationships between features are further calculated, and a feature-to-feature relationship strength matrix is constructed to form a parameter association table. Based on this association table, the importance score of each feature is analyzed, and redundant and low-value features are eliminated, ultimately generating standardized multidimensional feature data. The core innovation of this method lies in feeding standardized multidimensional feature data into a two-stream spatiotemporal convolutional neural network. This network consists of two parallel processing branches: a temporal feature processing branch and a spatial feature processing branch. The data is first split into a time series and a feature relationship component, which are then fed into these two processing branches. The temporal feature processing branch employs a one-dimensional convolutional chain structure with a variable-length receptive field, using a dilated convolution mechanism to capture load fluctuation patterns over different time spans. Dilated convolution is a special convolution operation that increases the receptive field by inserting holes into the convolution kernel while maintaining the number of parameters and computational complexity. In this method, multiple levels of dilated convolutional layers are designed with dilation rates of 1, 4, and 16, enabling the network to simultaneously capture short-term, medium-term, and long-term load variation characteristics. Simultaneously, the spatial feature processing branch processes feature relationship data using an autocorrelation graph structure to extract implicit dependencies between different monitoring parameters. This autocorrelation graph constructs a map of the strength of connections between features through cross-channel feature mapping, forming a parameter relationship matrix. After being processed by their respective branches, temporal and spatial features are fed into the temporal gating unit and feature reconstruction network for further processing, generating adaptive temporal features and nonlinear spatial features. Finally, these two features are combined through a bidirectional attention fusion mechanism to form fused load prediction data.

[0027] Layer-wise quantization and pruning of the two-stream spatiotemporal convolutional neural network is designed to address the large size and high computational complexity of the neural network model, enabling efficient operation on resource-constrained base station controllers. This step first tests the network's sensitivity using real-world base station traffic data to calculate the impact of each layer's parameters on load forecast errors during peak and off-peak periods. Parameter sensitivity is determined through perturbation testing, where small perturbations are added to each layer's parameters to observe the impact on the forecast results. Based on the test results, the network is divided into three layers: a high-sensitivity layer responsible for capturing burst traffic, a medium-sensitivity layer for identifying daily fluctuations, and a low-sensitivity layer for extracting underlying load trends. Different quantization strategies are employed for layers of varying sensitivity. The high-sensitivity layer retains full burst recognition functionality and uses eight-bit fixed-point quantization; the medium-sensitivity layer optimizes parameters for identifying daily traffic patterns and uses six-bit fixed-point quantization; and the low-sensitivity layer simplifies parameters for extracting long-term trends and uses four-bit fixed-point quantization. This differentiated processing approach significantly reduces the model's storage and computational requirements while ensuring forecast accuracy, resulting in a compressed network structure suitable for base station deployment. This compressed network structure processes real-time load data to generate dynamic energy-saving control strategies for base stations. First, real-time load data is input into the compression network structure for forward calculation, generating load forecast data for future time periods. Based on the forecast results, the time period is divided into high-load, medium-load, and low-load intervals to form a load status table. Then, by determining whether the duration of the low-load interval exceeds a threshold (such as 30 minutes), a decision is made to trigger deep energy-saving mode or basic energy-saving mode. For the determined energy-saving mode, a corresponding resource scheduling plan is generated, including carrier switching status, antenna configuration, and channel allocation instructions. At the same time, based on the changing trends of the load forecast data, the optimal transmit power curve for each time period is calculated, and the power adjustment parameters are generated. Finally, the resource scheduling plan and power control plan are organized in chronological order, and execution conditions and a safety fallback mechanism are added to form a complete base station dynamic energy-saving control strategy.

[0028] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0029] Record the radio frequency power value, channel usage number, number of access users, data transmission volume, power consumption value and temperature value of the communication base station every minute to obtain an original data table;

[0030] Replace the values in the original data table that exceed the upper and lower thresholds, and fill the missing values with the average values of the previous and next time points to obtain a complete data series;

[0031] For each value of the complete data series, the corresponding characteristic mean is subtracted and then divided by the standard deviation to convert it into a zero-mean unit variance distribution to obtain uniform scale data;

[0032] Calculate the linear relationship measurement value for the unified scale data pairwise, construct the relationship strength matrix between features, and obtain the parameter association table;

[0033] Analyze the importance scores based on the parameter association table, remove feature columns with importance scores lower than the set threshold, and obtain the key feature set;

[0034] The key feature sets are organized in chronological order, and time period identifiers and load status labels are added to obtain standardized multi-dimensional feature data.

[0035] Specifically, key operating parameters of the communication base station are recorded every minute through the data interface of the base station management system, including radio frequency power, number of channels in use, number of connected users, data transmission volume, power consumption, and temperature. Radio frequency power refers to the real-time transmit power of the base station's radio frequency unit (RFU), typically measured in watts; channel use indicates the number of currently active wireless channels; connected users record the number of terminal devices connected to the base station; data transmission volume measures the total amount of uplink and downlink data per unit time, typically measured in bits; power consumption records the base station's overall energy consumption, including the combined energy consumption of the RFU, baseband processing unit, transmission unit, and auxiliary equipment; and temperature records the base station's ambient temperature, a factor influencing energy consumption. This data is organized and stored in timestamp order to form a raw data table. Raw data tables often contain outliers and missing values, requiring data cleaning. For values exceeding upper and lower thresholds, a sliding window is constructed, typically with a window size of 30 data points (i.e., 30 minutes). The mean μ and standard deviation σ of the data within the window are calculated, and data points that deviate from the mean by more than three standard deviations (μ ± 3σ) are marked as outliers. For example, if the number of connected users at a certain moment suddenly jumps from 100 to 900, while the average number of connected users 30 minutes before and after is 120 with a standard deviation of 30, this value clearly exceeds the normal fluctuation range and requires replacement. This replacement method uses the average of the preceding and following time points—that is, the arithmetic mean of the five valid data points before and after the outlier. Missing values are similarly filled with the average of the preceding and following time points to ensure data continuity. This complete data series eliminates noise interference and more accurately reflects the changing trends of base station load. The complete data series contains features of different dimensions. Direct comparison and calculation will cause features with large values to dominate the model training process. Therefore, normalization is necessary. The mean of each feature is subtracted from its value and then divided by its standard deviation to convert it to a zero-mean, unit-variance distribution. This involves calculating the mean and standard deviation of each feature column and then applying the Z-score normalization formula to each data point. For example, if the average data transmission volume of a base station over a 24-hour period is 500MB and the standard deviation is 150MB, then a data point with an original value of 800MB will have a normalized value of (800-500) / 150=2, indicating that the data transmission volume at that point is 2 standard deviations higher than the average. This transformation ensures that all features have similar value ranges and are equally weighted, resulting in uniformly scaled data.

[0036] Inter-feature correlation analysis is performed on the uniformly scaled data, and linear relationship metrics are calculated pairwise to construct a feature-to-feature relationship strength matrix. The linear relationship metric uses the Pearson correlation coefficient to calculate the degree of correlation between any two features. For feature X and feature Y, their respective means are first calculated, followed by their covariances and standard deviations, and finally the correlation coefficient. The correlation coefficient ranges from -1 to 1, with values close to 1 indicating a strong positive correlation, -1 indicating a strong negative correlation, and 0 indicating no correlation. For example, if the correlation coefficient between the number of connected users and data transmission volume is 0.92, it indicates a high positive correlation between the two features. A correlation coefficient of -0.15 between temperature and RF power indicates a weak correlation. By calculating the correlation coefficients between all feature pairs, an n×n relationship strength matrix, or parameter association table, is constructed.

[0037] Feature importance scores are analyzed based on the parameter association table to identify and remove redundant or low-value features. Importance analysis uses the information gain method to calculate the information gain of each feature with respect to the target variable (usually energy consumption data). Features with high information gain have a significant impact on the prediction results and should be retained; features with low information gain contribute little and can be considered for removal. Furthermore, for feature pairs with a correlation greater than 0.85, the feature with the higher information gain is retained and the other is removed to reduce redundancy. For example, if the correlation coefficient for the number of connected users and data transmission volume is 0.92, and their information gains for energy consumption are 0.65 and 0.58, respectively, the number of connected users is retained and the data transmission volume feature is removed. An importance threshold is set and feature columns with scores below the threshold are removed to obtain a key feature set. The key feature set is reorganized in chronological order, with time period identifiers and load status labels added. Time period identifiers are categorized by time of day into morning peak (7:00-10:00), daily period (10:00-17:00), evening peak (17:00-21:00), and nighttime period (21:00-7:00). Load status labels mark each time point as high, medium, or low load based on the values of key load indicators (such as the number of connected users or data transmission volume). For example, when the normalized value of data transmission volume at a certain moment is greater than 1.5, it is marked as high load; between -0.5 and 1.5, it is marked as medium load; and less than -0.5, it is marked as low load. These time period identifiers and load status labels provide important reference information for subsequent load prediction and resource scheduling. The resulting standardized multidimensional feature data contains a cleaned, standardized, and filtered feature set, as well as additional time period and load status labels.

[0038] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0039] Split the standardized multi-dimensional feature data into two parts: time series and feature relationship, and input them into the asymmetric dual-stream processing structure of the parallel architecture to obtain the initial dual-stream input;

[0040] The time series portion of the initial dual-stream input is processed through a one-dimensional convolution chain with a variable-length receptive field. Using the dilated convolution mechanism, the load fluctuation patterns of different time spans are captured to obtain a multi-granularity time series representation.

[0041] The autocorrelation graph structure is applied to the feature relationship part of the initial dual-stream input, and the implicit dependency between parameters is extracted through the feature mapping of the cross channel to obtain the parameter relationship matrix;

[0042] The multi-granularity time series representation is input into the time series gating unit, and the weight ratio of different time windows is adjusted according to the historical load change trend to obtain adaptive time features;

[0043] The parameter relationship matrix is analyzed through the feature recombination network to dynamically extract the highly influential feature combinations and interaction patterns to obtain nonlinear spatial features.

[0044] Bidirectional attention fusion is performed on the adaptive temporal features and nonlinear spatial features, and spatiotemporal collaborative expression is constructed through the complementary information enhancement mechanism to obtain fused load prediction data.

[0045] Specifically, a two-stream spatiotemporal convolutional neural network is the core module of the energy-saving optimization control method for communication base stations. It first requires splitting the standardized multidimensional feature data into two components: a time series component and a feature relationship component. The time series component contains the feature values arranged in chronological order, capturing the temporal patterns of load variation; the feature relationship component focuses on the interactions and dependencies between different features. During this splitting process, the continuous values of all features along the time dimension are extracted from the standardized multidimensional feature data to form the time series component, while the correlation strength information between features is organized into the feature relationship component. Specifically, for each time point t, all feature values at that time point are extracted to form a vector Xt. The vectors for all time points are sequentially concatenated into the sequence {X1, X2, ..., Xt, ...} as the time series component. Simultaneously, a feature relationship matrix R is constructed based on the parameter association table generated in the previous step. Each element Rij in the matrix represents the correlation strength between feature i and feature j. These two components of data are input separately into the asymmetric two-stream processing structure with a parallel architecture to form the initial two-stream input.

[0046] To process the time series portion of the initial two-stream input, a variable-length receptive field (VRF) one-dimensional convolution chain is employed. This unique neural network architecture is capable of simultaneously capturing patterns across different time spans. This variable-length receptive field is achieved through dilated convolution, a convolution operation that introduces gaps (or holes) in addition to standard convolution. In standard convolution, kernel elements are applied to adjacent elements of the input; in dilated convolution, fixed gaps are inserted between kernel elements, thereby expanding the receptive field. The dilation rate determines the size of the gaps. A dilation rate of 1 is equivalent to a standard convolution, while increasing the dilation rate increases the receptive field exponentially. In this method, multiple levels of VRF layers are constructed with dilation rates of 1, 4, and 16, respectively, to capture short-term (minutes), medium-term (hours), and long-term (days) load variation patterns. Specifically, a one-dimensional convolution with a dilation rate of 1 is first applied to the time series data to extract relationships between adjacent time points. The resulting convolution is then fed into a convolution layer with a dilation rate of 4 to capture patterns across a wider time range. Finally, a convolution layer with a dilation rate of 16 captures long-term trends. Through this stacked structure, feature representations of different time scales are formed, namely multi-granularity temporal representation.

[0047] The autocorrelation graph structure is applied to process the feature relationship part in the initial dual-stream input. The autocorrelation graph structure is a computational structure that specifically processes the relationship between features. It constructs a feature relationship graph by treating the features as nodes in the graph and the correlation strength as the weight of the edge. The processing process includes: first, constructing the initial graph structure based on the feature relationship matrix; then, updating the relationship between each pair of features through the cross-channel feature mapping operation. Cross-channel feature mapping means that the influence of feature i on feature j depends not only on the direct correlation Rij between them, but also on the indirect influence of other features. Through multiple rounds of iterative updates, the high-order relationships between features are calculated to form a more accurate parameter relationship matrix. During the update process, each round of iteration applies a nonlinear transformation to the relationship matrix R: R'=σ(RWR T ), where W is a learnable weight matrix and σ is a nonlinear activation function such as ReLU or tanh. This update method can mine the implicit dependencies between features, thereby obtaining a more comprehensive parameter relationship matrix.

[0048] The multi-granularity time series representation is further processed by a time gating unit, a mechanism that dynamically adjusts the weights of different time windows based on historical load trends. This unit receives the multi-granularity time series representation as input, analyzes the importance of features at different time scales, and assigns appropriate weights to them. Specifically, the time gating unit first extracts the characteristics of the historical load change trend, then calculates the correlation between each time scale feature and the historical trend, with time scales with high correlation receiving greater weights. For example, when the historical load exhibits obvious periodic changes, time scale features that capture the corresponding period length will receive higher weights; when the load exhibits sudden changes, the weights of short-term time scale features will increase. This dynamic weight assignment mechanism enables the model to adaptively focus on the most relevant time scale information, thereby forming adaptive time features.

[0049] The parameter relationship matrix is further analyzed and processed by the feature recombination network. The feature recombination network is a computational structure used to analyze the importance of feature combinations. It extracts highly influential feature interaction patterns through multi-layer nonlinear transformations. The network first applies a threshold to each element in the parameter relationship matrix to retain strong correlations. It then performs a nonlinear transformation on the filtered matrix using a multi-layer perceptron to extract high-order feature combinations. Finally, significance analysis is used to identify key feature interaction patterns. This significance analysis uses an attention mechanism to calculate the contribution of each feature combination to load prediction, with combinations with high contributions receiving higher attention weights. In this way, the feature recombination network dynamically extracts the most valuable feature combinations and interaction patterns for load prediction, forming nonlinear spatial features. A bidirectional attention fusion mechanism combines adaptive temporal features with nonlinear spatial features to generate fused load forecast data. Bidirectional attention fusion considers mutual reinforcement between features, unlike simple feature concatenation or weighted averaging. This mechanism involves attention computation in two directions: attention of temporal features on spatial features, and attention of spatial features on temporal features. Specifically, the method first calculates the attention weight of each element in the adaptive temporal feature on the nonlinear spatial feature to form a temporally guided spatial feature. Simultaneously, the method calculates the attention weight of the nonlinear spatial feature on the adaptive temporal feature to form a spatially enhanced temporal feature. These two enhanced features are then adaptively fused through a gating mechanism to generate a fused feature that contains both temporal and spatial information. Finally, a fully connected layer maps the fused feature to the target dimension to generate fused load prediction data, which serves as the basis for subsequent energy-saving strategy generation.

[0050] In a 5G communication base station, a two-stream spatiotemporal convolutional neural network processed seven days of historical data containing five features: RF power, channel usage, number of connected users, power consumption, and temperature. After data segmentation, the time series component formed a sequence of 10,080 time points (one point per minute), and the feature relationship component formed a 5×5 correlation matrix. Processing the time series with three levels of dilated convolutions generated a multi-granularity time series representation containing short-term (last four hours), medium-term (last 24 hours), and long-term (7 days) features. For example, the dilated convolutions discovered that the base station's load exhibited a distinct bimodal pattern on weekdays (9:00 AM and 8:00 PM) and a unimodal pattern on weekends (around 3:00 PM). After analyzing these patterns, the temporal gating unit (TGM) prioritized short-term and daily periodic patterns when predicting weekday load, while prioritizing periodic patterns when predicting weekend load. Furthermore, feature correlation analysis revealed a strong correlation between the number of connected users and RF power and channel usage. The feature reconstruction network then extracted these highly correlated feature combinations as key predictors. Ultimately, through bidirectional attention fusion, the model accurately predicted base station load trends over the next 24 hours, providing a reliable basis for formulating energy-saving control strategies. This prediction method, which integrates both temporal and spatial features, overcomes the limitations of traditional single models, which struggle to simultaneously handle temporal variations and feature interactions, significantly improving prediction accuracy and robustness.

[0051] In a specific embodiment, the process of performing the step of passing the variable-length receptive field one-dimensional convolution chain on the time series portion of the initial dual-stream input may specifically include the following steps:

[0052] A multi-level dilated convolution structure is constructed for the time series part, and the convolution kernel interval is set according to the increasing dilation rate to cover the load pattern from short time to long time, and the original dilated feature group is obtained;

[0053] Apply the residual connection mechanism to the original dilated feature group, add and fuse the input features with the convolution output features, retain the original temporal information, and obtain enhanced dilated features;

[0054] The enhanced dilated features are processed by group convolution to extract different frequency features in groups to obtain a frequency separation feature set;

[0055] Multi-head temporal encoding is performed on the frequency separation feature set to extract periodic patterns, trend patterns, and burst patterns, respectively, and obtain three types of temporal semantic features;

[0056] Based on the three types of temporal semantic features, a feature importance graph is constructed, and the weight distribution of different time windows and different feature types is calculated to obtain a weighted feature map;

[0057] The weighted feature maps are fused across windows, and the self-attention mechanism is used to associate feature expressions at different time scales to obtain multi-granularity temporal representation.

[0058] Specifically, the time series is divided into three time granularities: minute, hour, and day segments, forming a multi-scale input sequence. For these input sequences, a three-level cascade of dilated convolutional layers is designed, with dilation rates set to one, four, and sixteen, respectively. Dilated convolution is a special convolution operation that increases the receptive field by inserting holes between convolution kernels while keeping the number of parameters unchanged. The dilation rate represents the spacing between convolution kernel elements. A dilation rate of one is equivalent to a standard convolution. As the dilation rate increases, the receptive field increases exponentially. The first-level dilated convolutional layer uses a convolution kernel with a dilation rate of one to process minute-scale sequences. The convolution kernel slides directly between adjacent time points, capturing short-term load variations, such as sudden increases and decreases in traffic. The second-level dilated convolutional layer uses a convolution kernel with a dilation rate of four to process hour-scale sequences. The convolution kernel skips four time points each time during calculation, capturing medium-term load variations, such as peak and trough cycles. The third-level dilated convolutional layer uses a convolution kernel with a dilation rate of sixteen to process day-scale sequences. The convolution kernel spans a larger time range and captures long-term load trends, such as the difference in traffic between weekdays and weekends. This multi-level structure with increasing dilation rates effectively covers a wide range of load patterns, from short-term to long-term. The output features of the three dilated convolutional layers are combined to form the original dilated feature set. Although the original dilated feature set captures features at different time scales, it is prone to losing the original temporal information during multiple convolutions. To address this issue, the original dilated feature set is enhanced using a residual connection mechanism. Residual connections are a technique that directly adds input features to the convolution output features, implemented through shortcut connections. The specific operation is to add the input features and output features of each dilated convolution layer element-wise after the output. For example, for the first-level dilated convolution layer, if the input feature is X1 and the convolution output is F1(X1), the result after residual connection is X1+F1(X1). Similarly, the second and third-level dilated convolution layers also apply the same residual connection mechanism. The result of the second level is X2+F2(X2), and the result of the third level is X3+F3(X3). This residual connection mechanism effectively retains the original time series information, prevents the gradient vanishing problem in deep network training, and enables the network to learn the residual mapping between input and output to form enhanced dilated features.

[0059] The enhanced dilated features contain a mixture of information from multiple time and frequency domains. To more precisely analyze the features at different frequencies, the enhanced dilated features are processed using grouped convolution. Grouped convolution is a technique that divides the input feature channels into multiple groups, each undergoing independent convolution operations. In this method, the enhanced dilated features are divided into three groups based on their frequency characteristics: high-frequency, medium-frequency, and low-frequency. The high-frequency group corresponds to features with short-term, rapid changes, the medium-frequency group to features with medium-term fluctuations, and the low-frequency group to features with long-term trends. Each group is processed using a convolution kernel with different parameters: the high-frequency group uses a kernel with a smaller receptive field to capture detailed changes, the medium-frequency group uses a kernel with a medium receptive field to extract periodic patterns, and the low-frequency group uses a kernel with a larger receptive field to extract long-term trends. The output features of the grouped convolution are combined by group to form a frequency-separated feature set, which clearly distinguishes the load characteristics at different frequencies. Multi-head temporal coding is applied to the frequency-separated feature set to extract temporal patterns from different perspectives. Multi-head temporal coding is a technique for processing multiple patterns in parallel and consists of three heads: a periodic pattern head, a trend pattern head, and a burst pattern head. The periodic pattern head focuses on extracting cyclical variations in load, using Fourier transform and autocorrelation analysis to identify periodic components in the signal. Examples include differences in load between weekdays and weekends, and patterns of high and low peaks at different times of the day. The trend pattern head focuses on long-term load trends, using methods such as moving average and linear regression to filter out short-term fluctuations and extract overall trends. The burst pattern head specifically identifies unconventional burst load events, employing anomaly detection algorithms to calculate the rapid rate of change and deviation of the signal and identify points where the load suddenly increases or decreases. These three heads process the frequency separation feature set in parallel, outputting three types of temporal semantic features: periodic features, trend features, and burst features.

[0060] A feature importance map is constructed based on three types of temporal semantic features to assess the importance of different time windows and feature types. The feature importance map is a two-dimensional matrix, with the horizontal axis representing different time windows (such as the last hour, the last four hours, or the last day) and the vertical axis representing different feature types (periodic, trending, or bursty). Each element in the matrix represents the importance weight of the corresponding time window and feature type. Importance weights are calculated based on the contribution of the feature to load prediction using an attention score. For each time window w and feature type t, the conditional mutual information (I(y; f_t|w)) with respect to the load prediction y is calculated. Higher mutual information values indicate greater importance of the feature for prediction in that time window. The softmax function is then used to normalize the mutual information values into weights, forming a complete feature importance map. Based on this, the feature vectors at each time point are weighted and summed according to their importance weights to generate a weighted feature map.

[0061] Cross-window fusion of weighted feature maps is the final step in forming a multi-granularity time series representation. Cross-window fusion uses a self-attention mechanism to allow features at different time scales to correlate and enhance each other. The self-attention mechanism is a technique for calculating the correlation between elements in a sequence, implemented through a query-key-value computation framework. In this method, features from different time windows are treated as sequence elements, and attention scores are calculated between them. First, a query vector Q, a key vector K, and a value vector V are generated for each time window feature. Then, the dot product of the query vector and all key vectors is calculated to obtain an attention score. The attention scores are then softmax-normalized. Finally, the normalized scores are weighted and summed over the value vectors to obtain the fused feature representation. This mechanism allows windows with strong correlations to enhance each other, thereby integrating information from different time scales and forming a comprehensive multi-granularity time series representation.

[0062] In a specific embodiment, the process of executing the step of constructing a multi-stage dilated convolution structure for the time series portion may specifically include the following steps:

[0063] The time series is divided into three types of sequence segments according to the time granularity: minute-level, hour-level, and day-level, to obtain a multi-scale input sequence;

[0064] A first-level dilated convolution layer is constructed for the multi-scale input sequence. The dilation rate is set to 1, so that the convolution kernel slides between adjacent time points, capturing the load relationship between adjacent moments and obtaining short-term dilation features.

[0065] The short-term expansion features are input into the second-level expansion convolution layer, and the expansion rate is set to four, so that the convolution kernel slides between four time points, capturing the load changes across time periods and obtaining the medium-term expansion features.

[0066] The mid-term expansion features are input into the third-level expansion convolution layer. The expansion rate is set to 16, so that the convolution kernel slides at intervals of 16 time points, capturing the long-term load trend and obtaining the long-term expansion features.

[0067] The short-term expansion features, medium-term expansion features and long-term expansion features are merged through channel splicing to retain the information of each time scale and obtain multi-view temporal features;

[0068] The feature importance scores are calculated for the multi-view time series features, and weight coefficients are assigned according to the correlation between each time scale feature and the historical load change to obtain the original expanded feature group.

[0069] Specifically, the time series is segmented into three types of segments based on time granularity: minute, hour, and day. The time series segment is a chronological sequence of feature values from the standardized multidimensional feature data. It contains records of time-varying parameters such as base station radio frequency power, channel usage, and number of connected users. This segmentation process involves resampling and aggregating the original sequence based on time granularity. The minute-level segment maintains the original sampling frequency, typically one data point per minute, fully preserving short-term fluctuations. The hour-level segment is obtained by aggregating the original sequence hourly, taking the average, maximum, and minimum values of 60 minutes of data per hour to form features reflecting hourly-scale variations. The day-level segment aggregates the original sequence daily, taking statistics from 24 hours of data per day to reflect daily load variations. This multi-scale segmentation yields input sequences of varying time granularity, providing the data foundation for subsequent feature extraction at different scales. A first-level dilated convolutional layer is constructed on the multi-scale input sequence to capture load relationships between adjacent moments. Dilated convolution is a convolution operation that introduces a dilation hole based on the standard convolution. The dilation rate determines the size of the dilation hole. The dilation rate of the first-level dilated convolution layer is set to one, that is, there are no holes, which is equivalent to standard convolution, and the convolution kernel slides directly between adjacent time points. In specific implementation, a one-dimensional convolution kernel of length 5 is used to perform convolution operations on minute-level sequence fragments, and the convolution kernel weights are learned through the backpropagation algorithm. The convolution operation is to multiply the convolution kernel with the corresponding position of the input sequence and sum them. For example, for time point t, the output feature is calculated by multiplying the feature values of t and the two time points before and after it, a total of 5 points, by the convolution kernel weights and then summing them. This convolution operation that slides directly between adjacent time points can effectively capture the load change relationship within the local time range, such as burst traffic, short-term fluctuations and other features, forming short-term dilation features that express the law of short-term load changes.

[0070] The short-term dilated features are then fed into the second-level dilated convolutional layer to further extract medium-term load variation patterns. The dilation rate of the second-level dilated convolutional layer is set to four, which means that the convolution kernel elements are spaced three positions apart, thus expanding the actual time range covered by the convolution kernel. If a convolution kernel of length 5 is still used, the actual time span covered is 1 + (5 - 1) × 4 = 17 time points, which is much larger than the coverage of the first-level convolutional layer. In practice, the convolution kernel is no longer convolved continuously with the input sequence, but instead takes a value every four time points. For example, at time point t, the output feature calculation considers the feature values at time points t, t-4, t-8, t+4, and t+8, multiplied by the convolution kernel weights, and then summed. This sliding convolution operation between four time points can capture medium-term load variation patterns over a wider time range, such as morning and evening peaks and traffic differences between working and non-working hours, forming medium-term dilated features that express medium-term load variation patterns. The medium-term dilated features are then fed into the third-level dilated convolutional layer to capture long-term load trends. The expansion rate of the third-level dilated convolution layer is set to sixteen, which means that the convolution kernel elements are separated by fifteen positions, further expanding the time span covered by the convolution kernel. Using the same convolution kernel of length 5, the actual time span covered is 1+(5-1)×16=65 time points, which is sufficient to capture load variation trends spanning multiple hours or even days. Specifically, for time point t, the output feature calculation will consider the feature values of five time points that are far apart: t, t-16, t-32, t+16, and t+32. This convolution operation that slides between sixteen time points can capture long-term trends such as traffic differences between weekdays and weekends and special holiday patterns, forming long-term expansion features that express long-term load variation patterns.

[0071] After three levels of dilated convolution, short-term, medium-term, and long-term dilated features are obtained, representing load characteristics at different time scales. To comprehensively utilize information from these three time scales, channel splicing is used to combine them to form multi-view temporal features. Channel splicing involves concatenating multiple feature vectors into a single, longer vector along the channel dimension, preserving all information from the original features. Specifically, the short-term, medium-term, and long-term dilated features are superimposed along the channel dimension. Assuming the dimensions of the three features are [batch size, time duration, number of short-term feature channels], [batch size, time duration, number of medium-term feature channels], and [batch size, time duration, number of long-term feature channels], the dimensions of the spliced multi-view temporal features are [batch size, time duration, number of short-term feature channels + number of medium-term feature channels + number of long-term feature channels]. This channel splicing method ensures that information from different time scales is fully preserved.

[0072] Feature importance scores are calculated for multi-view time series features to assign weights based on the importance of different timescale features to prediction. Feature importance analysis is based on the correlation between each timescale feature and historical load changes. The specific calculation process involves first extracting the historical load change series, typically taking the load change rate within the most recent period (e.g., 48 hours). Then, the correlation coefficients between the short-term, medium-term, and long-term expansion features and this load change series are calculated. A higher correlation coefficient indicates a more important timescale feature for prediction. Finally, the correlation coefficients are normalized into weight coefficients, ensuring that the sum of all weights is 1. The weight coefficients are calculated using the softmax function, which first exponentially calculates each correlation coefficient and then divides it by the sum of all exponentials to obtain a normalized weight between 0 and 1. Through this correlation-based weighting method, the different timescale features in the multi-view time series features are weighted and combined according to their importance, forming the original expansion feature group that represents the multi-scale temporal characteristics. In the practical application of energy-saving optimization control for communication base stations, the multi-level dilated convolution analysis process for historical load data from a specific 5G base station is as follows: First, the load data collected minute by minute over a week (a total of 10,080 data points) is divided into sequence segments at three time granularities. The minute level maintains the original sampling frequency; the hourly level is aggregated into a 168-point sequence by taking the average, maximum, and minimum values of 60 points per hour; and the daily level is aggregated into a 7-point sequence. A first-level dilated convolution layer (dilation rate = 1) is then constructed to directly convolve the minute-level sequence with a kernel length of 5, extracting load variation patterns within a local time window, such as the sudden increase in traffic at the start of the office at 9:00 AM. A second-level dilated convolution layer (dilation rate = 4) is then constructed to process the features output by the first level, expanding the receptive field to 17 time points, capable of capturing patterns such as the decrease and then increase in traffic during lunch time from 12:00 to 1:00 PM. Then comes the third-level dilated convolution layer (dilation rate = 16), which expands the receptive field to 65 time points, capturing patterns over longer time spans, such as the evening peak load pattern from 18:00 to 22:00 every day. These three levels of dilated features are combined through channel stitching to form comprehensive multi-perspective time series features. Finally, by analyzing the correlation between the three time scale features and historical load changes, they are assigned weights. For example, when predicting daily load on weekdays, medium-term features (hourly patterns) have the highest correlation and receive a weight of 0.6; when predicting load during special events on weekends, short-term features (burst patterns) have increased correlation and receive a weight of 0.5; when predicting load during long holidays, long-term features (day-level patterns) have the highest correlation and receive a weight of 0.7. This dynamic weight allocation based on correlation ensures that the model can flexibly adjust its focus according to different prediction scenarios, thereby improving prediction accuracy. Through the design and implementation of the above-mentioned multi-level dilated convolution structure, the technical problem that traditional load forecasting methods cannot simultaneously capture load characteristics at different time scales is solved.This method effectively expands the receptive field of view while maintaining the same number of parameters through the design of convolutional layers with increasing dilation rates. This overcomes the high computational complexity and large model size of traditional methods when processing long time series. The multi-perspective feature extraction and feature importance analysis mechanism enable the model to adaptively focus on the most relevant timescale features in different prediction scenarios, improving the accuracy and robustness of predictions. This flexible and efficient prediction mechanism provides a reliable decision-making basis for energy-saving optimization and control of communication base stations, addressing the problems of low energy efficiency and unstable service quality caused by inaccurate load forecasting.

[0073] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0074] A sensitivity test of the two-stream spatiotemporal convolutional neural network was conducted on real communication base station traffic data. The influence of each layer parameter on the load prediction error during peak and off-peak periods was calculated, and the sensitivity distribution related to base station scenarios was obtained.

[0075] Based on the sensitivity distribution related to base station scenarios, the convolutional network is divided into a high-sensitivity layer responsible for capturing burst traffic, a medium-sensitivity layer for identifying daily fluctuation patterns, and a low-sensitivity layer for extracting basic load trends, thereby obtaining a hierarchical labeling of communication load characteristics.

[0076] The high-sensitivity layer in the communication load characteristic level marking retains all RF burst traffic recognition functions, uses eight-bit fixed-point quantization to ensure the accuracy of burst traffic warning, and obtains the burst traffic perception structure;

[0077] The parameters for identifying daily business patterns are optimized for the medium-sensitive layer in the communication load characteristic level marking, and six-bit fixed-point quantization is used to ensure resource allocation efficiency during regular periods, thereby obtaining a regular traffic processing structure.

[0078] The long-term trend extraction parameters of the low-sensitivity layer in the communication load characteristic level mark are simplified, and four-bit fixed-point quantization is used to meet the basic load prediction needs to obtain the basic load evaluation structure;

[0079] The burst traffic perception structure, regular traffic processing structure and basic load assessment structure are integrated into a lightweight prediction engine for base station controllers. The calculation process is optimized according to the computing power constraints of the communication base station processor to obtain a compressed network structure suitable for base station deployment.

[0080] Specifically, a sensitivity test was conducted on the two-stream spatiotemporal convolutional neural network. Sensitivity testing evaluates the importance of parameters at each layer of the neural network, quantifying the impact of parameter changes on prediction results. The process involves selecting historical traffic data from real base station networks, including samples from typical peak hours (such as weekday peaks between 9:00 AM and 5:00 PM and 7:00 PM) and trough hours (such as the trough between 2:00 AM and 5:00 AM). These samples are then fed into the trained two-stream spatiotemporal convolutional neural network for prediction, and the original prediction results are recorded. The network parameters are then perturbed slightly by adding a small Gaussian noise, with the standard deviation typically set to 1% of the standard deviation of the layer's parameters. After the perturbation, predictions are performed again using the same samples, comparing the prediction results before and after the perturbation, and calculating the change in prediction error. This perturbation-prediction-comparison cycle is repeated multiple times for each layer, and the average change in error is taken as the sensitivity index for that layer. A higher sensitivity index indicates a greater impact and importance of the layer's parameters on the prediction results. This method tests all network layers, ultimately yielding a complete base station scenario-specific sensitivity distribution—the quantified sensitivity values of each layer under different load scenarios. Based on this base station scenario-specific sensitivity distribution, the layers of the two-stream spatiotemporal convolutional neural network are then classified by function and sensitivity. This classification combines the layer's functional role and sensitivity metric, dividing the network into three hierarchical levels. Highly sensitive layers, defined as those in the top 30% by sensitivity, are typically located in shallow layers of the network and in specific feature extraction modules. They are responsible for capturing bursty traffic patterns, such as a sudden increase in user connection requests or traffic surges caused by large-scale events. These layers are crucial for identifying short-term, sudden load fluctuations. Their high sensitivity stems from the fact that even small parameter changes can significantly impact the accuracy of sudden event predictions. Moderately sensitive layers, defined as those in the middle 40% by sensitivity, are typically located in the middle of the network and are responsible for identifying daily fluctuation patterns, such as the difference in traffic between weekdays and weekends, and periodic peaks and lows within a day. These layers are important for capturing regular variations in base station load, but require less robust parameter stability than for sudden events. Low-sensitivity layers, representing the bottom 30% of the sensitivity ranking, are primarily located deep within the network and are responsible for extracting fundamental load trends, such as stable features like long-term user growth and seasonal variations. These layers contribute significantly to understanding long-term trends, but because long-term trends themselves change slowly, the parameter accuracy requirements for these layers are relatively low. This sensitivity- and function-based classification establishes a hierarchical labeling of communication load characteristics, providing a basis for subsequent differentiated compression.

[0081] When processing the highly sensitive layers in the communication load characteristic hierarchical marking, a conservative compression strategy is employed, preserving full RF burst recognition capabilities. Specifically, the network structure of these layers remains unchanged, while only the parameters are compressed using eight-bit fixed-point quantization. Fixed-point quantization converts floating-point parameters into fixed-point representations. Eight-bit fixed-point quantization uses 8 binary bits (1 sign bit, 7 value bits) to represent each parameter. The quantization process first determines the parameter's dynamic range [min, max], then calculates the quantization step size. Finally, the floating-point parameter value is converted to an integer quantization value qval = round((val - min) / step). Only the quantization value qval and the step size step need to be recorded for storage. This eight-bit fixed-point representation maintains high numerical precision, ensuring the ability to sensitively capture burst changes. Through this refined quantization process, the highly sensitive layers can maintain accurate recognition of burst events, forming a burst-aware structure that provides reliable predictions for rapid base station response under burst load conditions.

[0082] A moderate compression strategy is employed for the medium-sensitivity layer, improving efficiency by optimizing parameters for identifying daily traffic patterns. The medium-sensitivity layer is primarily responsible for identifying periodic changes and regular fluctuations in base station load. Six-bit fixed-point quantization provides sufficient accuracy for these characteristics. Six-bit fixed-point quantization uses a 6-bit binary number (1 sign bit, 5 value bits) to represent each parameter. The quantization process is similar to that of the high-sensitivity layer, but with a larger step size and slightly lower accuracy. In addition to quantization, the medium-sensitivity layer undergoes sparsification by setting a small threshold (typically 5% of the mean absolute value of the parameter) and setting parameters with absolute values below the threshold to zero. This sparsification operation reduces the number of non-zero parameters, further reducing storage and computational requirements, while having a limited impact on the identification of regular daily load patterns. Through this optimization, the medium-sensitivity layer forms an efficient regular traffic processing structure that accurately identifies the regular fluctuations in base station load, providing a basis for regular resource scheduling.

[0083] A more aggressive compression strategy is implemented for the low-sensitivity layer, reducing computational complexity by simplifying parameters for long-term trend extraction. The low-sensitivity layer is primarily responsible for extracting long-term trends in base station load. These trends typically change slowly and smoothly, requiring lower parameter accuracy. Therefore, four-bit fixed-point quantization is used for compression, representing each parameter using a four-bit binary number (1 sign bit, 3 value bits). While this four-bit representation offers lower precision, it is sufficient to capture the slow changes in long-term trends. In addition to quantization, the low-sensitivity layer undergoes structural simplification by examining the importance of each neuron, using the L1 norm (the sum of the absolute values of the parameters) as a metric. Unimportant neurons (the 20%-30% with the smallest L1 norm) are removed, and connectivity is adjusted accordingly. This structural simplification significantly reduces the number of model parameters and computational complexity, while minimally impacting the accuracy of long-term trends. Through these simplifications, the low-sensitivity layer forms a lightweight basic load assessment structure that effectively extracts long-term trends in base station load, providing a reference for long-term resource planning.

[0084] The final step is to integrate the burst traffic sensing structure, regular traffic processing structure, and basic load assessment structure into a lightweight prediction engine for base station controllers. This integration process first establishes the data flow relationship between the three structures to ensure the correct forward propagation order of the network. The computational process is then optimized, primarily through batch processing optimization, which reduces repeated operations by feeding multiple samples into the network at once; memory access optimization, which adjusts the data storage layout to reduce memory access conflicts; and parallel computing optimization, which adjusts the computational task allocation based on the parallel computing capabilities of the base station processor. Furthermore, algorithm-level optimizations are performed, combining multiple consecutive linear operations into a single one to reduce the storage and loading of intermediate results. The integrated network structure undergoes final verification to ensure that prediction accuracy is maintained after compression, particularly under burst load conditions. This series of integration and optimization operations ultimately results in a compressed network structure suitable for base station deployment. This structure significantly reduces storage requirements and computational complexity while maintaining prediction accuracy, enabling efficient execution on the base station's embedded processor.

[0085] In a practical application scenario, a 5G communication base station used this method to compress a two-stream spatiotemporal convolutional neural network. The process was as follows: First, the base station's load data from the past 30 days was selected, specifically samples from weekday peak hours (8-10 AM and 6-8 PM, when traffic reaches over 150% of its peak value) and nighttime off-peak hours (1-5 AM, when traffic drops below 30% of the daily average) were extracted for sensitivity testing. Gaussian noise with a standard deviation of 1% of the parameter value was added to each of the network's 18 layers. The results showed that the first two convolutional and attention layers in the front-end temporal feature processing branch had the highest sensitivity, with prediction error changes exceeding 15%. The intermediate feature fusion module and some fully connected layers had moderate sensitivity, with prediction error changes between 5% and 15%. The deeper output network and some auxiliary layers had lower sensitivity, with prediction error changes of less than 5%. Based on this sensitivity distribution, the network is divided into three layers: the high-sensitivity layer, comprising the first two convolutional layers and the attention layer of the temporal feature processing branch, is responsible for identifying burst traffic; the medium-sensitivity layer, comprising the feature fusion module and the key fully connected layer, is responsible for identifying daily load patterns; and the low-sensitivity layer, comprising the output network and auxiliary function layers, is responsible for extracting long-term trends. 8-bit fixed-point quantization is applied to the high-sensitivity layer, converting floating-point parameters to integers in the range [-128, 127]. 6-bit fixed-point quantization and sparsification are applied to the medium-sensitivity layer, zeroing out parameters less than 5% of their mean, ultimately reducing the proportion of non-zero parameters to approximately 60%. 4-bit fixed-point quantization and structural simplification are applied to the low-sensitivity layer, removing 25% of unimportant neurons. After integrating these three structures, computational optimizations were performed to target the characteristics of the base station's ARM processor, merging consecutive linear operations and optimizing memory access patterns. The resulting compressed network reduces model size by 85% and inference time by 78%, while maintaining a maximum 2% decrease in burst load prediction accuracy, meeting the resource constraints of the base station controller.

[0086] The above-mentioned layered quantization pruning and differentiated compression strategies have solved the technical problems faced by traditional neural network models in communication base station deployment. The differentiated quantization strategy ensures high-precision retention of important layers and effective compression of secondary layers, resolving the contradiction between model accuracy and resource usage. The function- and sensitivity-oriented compression method ensures the accuracy of the model in key load prediction scenarios, especially the early warning capability for burst traffic, overcoming the problem of insufficient prediction accuracy of the compressed model in existing technologies. The optimization of the calculation process based on the characteristics of the base station processor solves the problem of low operating efficiency of complex models in resource-constrained environments. This optimization method, which maintains accuracy and improves efficiency, provides reliable decision-making support for base station energy-saving control, effectively solving problems such as low energy-saving efficiency and unstable network service quality.

[0087] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0088] Input the real-time load data into the compression network structure, generate the load change curve of the future period through forward calculation, and obtain the load forecast data;

[0089] The load forecast data is divided into states, and the time period is divided into high load interval, medium load interval and low load interval according to the load rate threshold to obtain a load state table;

[0090] Determine whether the duration of the low load interval in the load status table exceeds thirty minutes, if so, trigger the deep energy saving mode, if not, trigger the basic energy saving mode, and obtain the energy saving mode identifier;

[0091] According to the energy-saving mode identifier and load status table, the radio frequency resource demand of the time period is calculated, the carrier switch state, antenna configuration and channel allocation instructions are generated, and the resource scheduling plan is obtained;

[0092] Based on the resource scheduling plan and the change trend of load forecast data, the transmit power curve of the time period is calculated, the power level, adjustment time point and step size parameters are generated, and the power control plan is obtained;

[0093] The resource scheduling scheme and power control scheme are organized in chronological order, and execution conditions and fallback mechanisms are added to form a dynamic energy-saving control strategy for base stations.

[0094] Specifically, real-time load data is input into the compression network structure for processing. The real-time load data includes base station operating parameters such as RF power values, number of channels in use, number of connected users, and data transmission volume. These data are standardized to match the training data and then input into the network. The compression network structure is a two-stream spatiotemporal convolutional neural network that has undergone layered quantization and pruning. It contains three sensitivity layers: high, medium, and low, responsible for capturing burst traffic, identifying daily patterns, and extracting long-term trends, respectively. The network generates load forecast data through a forward computation process. Forward computation is the process of transferring data from the input layer to the output layer: time series and feature relationship data pass through the temporal feature processing branch and the spatial feature processing branch respectively, and undergo a series of operations such as convolutional layers, pooling layers, and fully connected layers to ultimately output the load forecast value for the future time period. The forecast range is usually for the next 6 hours, with 5-minute intervals, forming a load change curve.

[0095] Classifying predicted load data into different states is the foundation for developing differentiated energy-saving strategies. This classification process uses a load rate threshold method. The load rate is the ratio of a base station's actual load to its designed capacity. It is typically calculated based on key metrics such as the ratio of connected users to the maximum number of users or the ratio of data throughput to maximum capacity. The classification criteria are set at three levels: periods with a load rate greater than 70% are classified as high-load intervals, where high user demand requires full resource activation; periods with a load rate between 30% and 70% are classified as medium-load intervals, suitable for partial energy conservation; and periods with a load rate below 30% are classified as low-load intervals, where resources are severely idle and suitable for deep energy conservation. The classification operation involves iterating through each predicted time point, comparing its load rate with a threshold, determining its state, and then merging adjacent time points with the same state into continuous intervals. This ultimately generates a load state table containing multiple time intervals and their corresponding load states.

[0096] Energy-saving mode judgment is based on the load status table, with the focus on evaluating the persistence of low-load intervals. The judgment process first extracts all low-load intervals from the load status table, and then calculates the duration of each interval. When the duration of any low-load interval exceeds the set threshold of thirty minutes, the deep energy-saving mode (marked as "deep") is triggered; when the duration of all low-load intervals does not exceed thirty minutes, the basic energy-saving mode (marked as "basic") is triggered. The setting of the duration threshold takes into account the resource switching overhead: although deep energy saving has a good energy-saving effect, it takes time to switch the carrier and antenna configuration. Too frequent switching will increase energy consumption. Therefore, it is only enabled when the low load is stable and lasts for a long time. This duration-based mode judgment ensures the stability and applicability of the energy-saving strategy.

[0097] Based on the energy-saving mode identifier and the load status table, the required radio resource requirements for each time period are calculated and a resource scheduling plan is generated. The calculation process first estimates the number of radio resource units (RRUs) required for each time period based on the predicted load and quality of service requirements. The calculation method converts the predicted number of users and data volume into the required number of resource blocks (RBs). Then, based on the RB capacity available for each carrier and antenna configuration, the minimum resource allocation that meets the requirements is determined. Differentiated resource allocation strategies are adopted for different load ranges: In high-load ranges, all resources remain active; in medium-load ranges, the number of carriers remains unchanged but power is adjusted in basic energy-saving mode, while some carriers may be shut down in deep energy-saving mode. In low-load ranges, power and channel allocation are primarily adjusted in basic energy-saving mode, while some carriers may be shut down and the MIMO order reduced in deep energy-saving mode. Resource scheduling instructions include the on / off status of each carrier, antenna configuration status, and channel allocation plan, organized in chronological order to form a complete resource scheduling plan. Based on the resource scheduling plan and predicted load trends, a transmit power control curve is calculated and a power control plan is generated. A base power level is first determined to ensure coverage requirements are met even at the lowest power level. The calculation method is based on the link budget, taking into account coverage radius, path loss, and receive sensitivity, to determine the transmit power that meets the minimum signal requirements. The power adjustment timing and step size are then determined based on the predicted load change trend. The timing is chosen based on the lead time of the load transition point, typically starting with power increases 5-10 minutes before the load increases. The step size is calculated based on the load change rate, with larger step sizes used for rapid changes and smaller step sizes for slower changes. The power control plan includes the power level, adjustment timing, and step size parameters at each time point, and works in conjunction with the resource scheduling plan to form a comprehensive energy consumption control strategy.

[0098] Finally, the resource scheduling and power control schemes are integrated into a complete base station dynamic energy-saving control strategy. The integration process begins by organizing the instructions of the two schemes in chronological order to ensure consistent resource allocation and power adjustment. Execution conditions and fallback mechanisms are then added to ensure the security of policy execution. Execution conditions include prediction credibility checks, such as setting a threshold for deviation between the predicted and actual load, pausing execution if exceeded. Fallback mechanisms include service quality monitoring triggers, such as setting a threshold for monitoring key performance indicators to immediately terminate energy-saving mode if the indicator deteriorates. The resulting base station dynamic energy-saving control strategy is a complete control scheme that can adaptively adjust base station resource allocation based on load forecasts while ensuring service quality.

[0099] The above describes the energy-saving optimization control method for a communication base station in an embodiment of the present application. The following describes the energy-saving optimization control system for a communication base station in an embodiment of the present application. Figure 2 In one embodiment of the present application, an energy-saving optimization control system for a communication base station includes:

[0100] The acquisition module 201 is used to collect and pre-process the traffic data and energy consumption data of the communication base station to obtain standardized multi-dimensional feature data;

[0101] Input module 202, used to input the standardized multi-dimensional feature data into the dual-stream spatiotemporal convolutional neural network, and perform parallel calculations through the temporal feature processing branch and the spatial feature processing branch to obtain fused load prediction data;

[0102] A pruning module 203 is configured to perform layered quantization pruning on the dual-stream spatiotemporal convolutional neural network, assigning different bit widths to each network layer according to the degree of influence of each network layer on the accuracy of the fused load prediction data, and obtaining a compressed network structure suitable for base station deployment;

[0103] The generating module 204 is configured to process the real-time load data based on the compressed network structure and generate a base station dynamic energy-saving control strategy including radio frequency resource configuration instructions and power adjustment parameters.

[0104] above Figure 2 The energy-saving optimization control system for a communication base station in an embodiment of the present invention is described in detail from the perspective of modular functional entities. The energy-saving optimization control device for a communication base station in an embodiment of the present invention is described in detail from the perspective of hardware processing.

[0105] Figure 3 : This is a structural diagram of an energy-saving optimization control device for a communication base station provided by an embodiment of the present invention. The energy-saving optimization control device 300 for a communication base station may have relatively large differences due to different configurations or performances, and may include one or more processors (central processing units, CPU) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 (for example, one or more mass storage devices) storing application programs 333 or data 332. Among them, the memory 320 and the storage medium 330 can be temporary storage or permanent storage. The program stored in the storage medium 330 may include one or more modules (not shown in the figure), each module may include a series of instruction operations in the energy-saving optimization control device 300 for the communication base station. Furthermore, the processor 310 can be configured to communicate with the storage medium 330, and execute a series of instruction operations in the storage medium 330 on the energy-saving optimization control device 300 for the communication base station to implement the steps of the above-mentioned energy-saving optimization control method for the communication base station.

[0106] The energy-saving optimization control device 300 for a communication base station may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The energy-saving optimization control device structure for a communication base station shown does not constitute a limitation on the energy-saving optimization control device for a communication base station provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0107] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium. When the instructions are executed on a computer, the computer executes the steps of the energy-saving optimization control method for a communication base station.

[0108] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0109] If the integrated unit is implemented in the form of 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 the present invention is essentially 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. The computer software product is stored in a storage medium and includes several instructions for enabling an energy-saving optimization control device for a communication base station (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk.

[0110] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for energy-saving optimization control of a communication base station, characterized in that: The energy-saving optimization control method for a communication base station includes: Collect and preprocess the traffic data and energy consumption data of communication base stations to obtain standardized multi-dimensional feature data; Inputting the standardized multi-dimensional feature data into a dual-stream spatiotemporal convolutional neural network, performing parallel calculations through a temporal feature processing branch and a spatial feature processing branch to obtain fused load prediction data; Performing layered quantization and pruning on the dual-stream spatiotemporal convolutional neural network, allocating different bit widths according to the degree of influence of each network layer on the accuracy of the fused load prediction data, and obtaining a compressed network structure suitable for base station deployment; The real-time load data is processed based on the compressed network structure to generate a base station dynamic energy-saving control strategy including radio frequency resource configuration instructions and power adjustment parameters.

2. The energy-saving optimization control method for a communication base station according to claim 1, characterized in that: The flow data and energy consumption data of the communication base station are collected and preprocessed to obtain standardized multi-dimensional feature data, including: Record the radio frequency power value, channel usage number, number of access users, data transmission volume, power consumption value and temperature value of the communication base station every minute to obtain an original data table; Replace the values in the original data table that exceed the upper and lower thresholds, and fill the missing values with the average values of the previous and next time points to obtain a complete data series; Subtract the corresponding characteristic mean from each value of the complete data sequence and divide it by the standard deviation to convert it into a zero-mean unit variance distribution to obtain unified scale data; Calculating linear relationship metrics for the unified scale data pairwise, constructing a feature-to-feature relationship strength matrix, and obtaining a parameter association table; Analyzing the importance scores based on the parameter association table, removing feature columns with importance scores lower than a set threshold, and obtaining a key feature set; The key feature set is organized in chronological order, and a time period identifier and a load status label are added to obtain the standardized multi-dimensional feature data.

3. The energy-saving optimization control method for a communication base station according to claim 1, characterized in that: The standardized multi-dimensional feature data is input into a dual-stream spatiotemporal convolutional neural network, and parallel calculation is performed through a temporal feature processing branch and a spatial feature processing branch to obtain fused load prediction data, including: Splitting the standardized multi-dimensional feature data into two parts: a time series and a feature relationship, and inputting the two parts into an asymmetric dual-stream processing structure of a parallel architecture to obtain an initial dual-stream input; The time series portion of the initial dual-stream input is subjected to a variable-length receptive field one-dimensional convolution chain, using a dilated convolution mechanism to capture load fluctuation patterns of different time spans, thereby obtaining a multi-granularity time series representation. Applying an autocorrelation graph structure to the feature relationship part of the initial dual-stream input, extracting implicit dependencies between parameters through cross-channel feature mapping, and obtaining a parameter relationship matrix; Inputting the multi-granularity time series representation into a time series gating unit, adjusting the weight ratios of different time windows according to historical load change trends, and obtaining an adaptive time feature; The parameter relationship matrix is subjected to significance analysis through a feature recombination network, and high-impact feature combinations and interaction patterns are dynamically extracted to obtain nonlinear spatial features; The adaptive time feature and the nonlinear spatial feature are bidirectionally fused with attention, and a spatiotemporal collaborative expression is constructed through a complementary information enhancement mechanism to obtain the fused load prediction data.

4. The energy-saving optimization control method for a communication base station according to claim 3, characterized in that: The time series part of the initial dual-stream input is subjected to a variable-length receptive field one-dimensional convolution chain, and a dilated convolution mechanism is used to capture the load fluctuation patterns of different time spans to obtain a multi-granularity time series representation, including: Constructing a multi-level dilated convolution structure for the time series portion, setting the convolution kernel interval according to the increasing dilation rate, covering the load pattern from short time to long time, and obtaining the original dilated feature group; Applying a residual connection mechanism to the original expanded feature group, adding and fusing the input features with the convolution output features, retaining the original temporal information, and obtaining enhanced expanded features; The enhanced expansion features are processed by group convolution to extract different frequency features in groups to obtain a frequency separation feature set; Performing multi-head temporal coding on the frequency separation feature set to extract periodic patterns, trend patterns, and burst patterns respectively, and obtaining three types of temporal semantic features; Constructing a feature importance graph based on the three types of temporal semantic features, calculating the weight distribution of different time windows and different feature types, and obtaining a weighted feature map; The weighted feature maps are fused across windows, and a self-attention mechanism is used to associate feature expressions at different time scales to obtain the multi-granularity temporal representation.

5. The energy-saving optimization control method for a communication base station according to claim 4, characterized in that: The multi-level dilated convolution structure is constructed for the time series portion, and the convolution kernel interval is set according to the increasing dilation rate to cover the load mode from short time to long time, thereby obtaining the original dilated feature group, including: Divide the time series into three sequence segments at the minute, hour, and day levels according to the time granularity to obtain a multi-scale input sequence; Constructing a first-level dilated convolution layer for the multi-scale input sequence, with the dilation rate set to 1, so that the convolution kernel slides between adjacent time points, capturing the load relationship between adjacent moments, and obtaining short-term dilation features; The short-term expansion features are input into the second-level expansion convolution layer, and the expansion rate is set to four, so that the convolution kernel slides between four time points to capture the load changes across time periods and obtain the medium-term expansion features; The mid-term expansion feature is input into the third-level expansion convolution layer, and the expansion rate is set to 16, so that the convolution kernel slides at intervals of 16 time points to capture the long-term load trend and obtain the long-term expansion feature; The short-term expansion feature, the medium-term expansion feature, and the long-term expansion feature are combined by channel splicing, retaining information of each time scale, and obtaining a multi-view temporal feature; The feature importance scores are calculated for the multi-view temporal features, and weight coefficients are assigned according to the correlation between each time scale feature and the historical load change, so as to obtain the original expanded feature group.

6. The energy-saving optimization control method for a communication base station according to claim 1, characterized in that: The dual-stream spatiotemporal convolutional neural network is subjected to layered quantization and pruning, and different bit widths are allocated according to the degree of influence of each network layer on the accuracy of the fused load prediction data to obtain a compressed network structure suitable for base station deployment, including: The sensitivity test of the dual-stream spatiotemporal convolutional neural network is performed on real communication base station traffic data. The influence of each layer parameter on the load prediction error during peak and valley periods is calculated to obtain the sensitivity distribution related to the base station scenario. Based on the sensitivity distribution related to the base station scenario, the convolutional network is divided into a high-sensitivity layer responsible for capturing burst traffic, a medium-sensitivity layer for identifying daily fluctuation patterns, and a low-sensitivity layer for extracting basic load trends, thereby obtaining a hierarchical labeling of communication load characteristics; The high-sensitivity layer in the communication load characteristic level mark retains all radio frequency burst traffic recognition functions, adopts eight-bit fixed-point quantization to ensure the accuracy of burst traffic warning, and obtains a burst traffic perception structure; Optimizing daily business pattern recognition parameters for the medium-sensitive layer in the communication load characteristic level label, using six-bit fixed-point quantization to ensure resource allocation efficiency during regular periods, and obtaining a regular traffic processing structure; Simplifying the long-term trend extraction parameters of the low-sensitivity layer in the communication load characteristic level mark, using four-bit fixed-point quantization to meet the basic load prediction needs, and obtaining a basic load evaluation structure; The burst traffic perception structure, the regular traffic processing structure and the basic load evaluation structure are integrated into a lightweight prediction engine for the base station controller, and the calculation process is optimized according to the computing power constraints of the communication base station processor to obtain the compressed network structure suitable for base station deployment.

7. The energy-saving optimization control method for a communication base station according to claim 1, characterized in that: The processing of real-time load data based on the compressed network structure to generate a base station dynamic energy-saving control strategy including radio frequency resource configuration instructions and power adjustment parameters includes: Input the real-time load data into the compression network structure, generate the load change curve of the future period through forward calculation, and obtain the load forecast data; The load forecast data is divided into states, and the time period is divided into high load interval, medium load interval and low load interval according to the load rate threshold to obtain a load state table; Determine whether the duration of the low load interval in the load status table exceeds thirty minutes, if so, trigger the deep energy saving mode, if not, trigger the basic energy saving mode, and obtain the energy saving mode identifier; According to the energy-saving mode identifier and load status table, the radio frequency resource demand of the time period is calculated, the carrier switch state, antenna configuration and channel allocation instructions are generated, and the resource scheduling plan is obtained; Based on the resource scheduling plan and the change trend of load forecast data, the transmit power curve of the time period is calculated, the power level, adjustment time point and step size parameters are generated, and the power control plan is obtained; The resource scheduling scheme and power control scheme are organized in chronological order, and execution conditions and fallback mechanisms are added to form a dynamic energy-saving control strategy for base stations.

8. An energy-saving optimization control system for a communication base station, used to implement the energy-saving optimization control method for a communication base station according to any one of claims 1 to 7, characterized in that: The energy-saving optimization control system for a communication base station includes: The acquisition module is used to collect and pre-process the traffic data and energy consumption data of the communication base station to obtain standardized multi-dimensional feature data; An input module, configured to input the standardized multi-dimensional feature data into a dual-stream spatiotemporal convolutional neural network, and perform parallel calculations through a temporal feature processing branch and a spatial feature processing branch to obtain fused load prediction data; A pruning module is used to perform layered quantization pruning on the dual-stream spatiotemporal convolutional neural network, allocating different bit widths according to the degree of influence of each network layer on the accuracy of the fused load prediction data, and obtaining a compressed network structure suitable for base station deployment; A generation module is used to process real-time load data based on the compressed network structure and generate a base station dynamic energy-saving control strategy including radio frequency resource configuration instructions and power adjustment parameters.

9. A computer device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the energy-saving optimization control method for a communication base station according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium, characterized in that A computer program is stored thereon, and when the computer program is executed by a processor, the processor is caused to execute the energy-saving optimization control method for a communication base station according to any one of claims 1 to 7.

Images