Energy saving optimization control method and system for communication base station

Abstract

The application relates to the technical field of energy-saving control of communication base stations, 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 the communication base station to obtain standardized features; inputting the features into a double-flow space-time convolution network to predict the load through parallel calculation of time and space branches; performing hierarchical quantization pruning on the network, allocating bit widths according to the influence degree to obtain a compressed structure; and processing real-time data based on the compressed network to generate an energy-saving control strategy for radio frequency resource configuration and power adjustment. Based on accurate load prediction and a differentiated resource scheduling strategy, the application realizes the optimal balance between energy saving and service quality, effectively overcomes the problems of passive response and coarse-grained control in traditional methods, and maximizes the energy saving effect under the premise of guaranteeing the communication quality.

Description

Technical Field

[0001] This application relates to the field of energy-saving control technology for communication base stations, and in particular to an energy-saving optimization control method and system for communication base stations. Background Technology

[0002] With the rapid development of mobile communication networks and the widespread application of 5G technology, the number of communication base stations has increased dramatically, making energy consumption a growing concern. Statistics show that communication base stations account 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 saving and software-level energy saving. Hardware-level energy saving is primarily achieved through improvements to equipment such as high-efficiency power amplifiers and natural cooling; software-level energy saving is achieved through resource management methods such as carrier hibernation and power adjustment. Among these, load-aware dynamic resource management has become the mainstream energy-saving method. It typically employs a preset threshold triggering mechanism, automatically activating energy-saving mechanisms when the base station load falls below a certain threshold, reducing energy consumption by shutting down some radio frequency units and lowering transmission power.

[0003] However, existing base station energy-saving technologies have significant shortcomings. First, traditional preset threshold triggering methods can only passively respond to load changes, lacking foresight, resulting in resource adjustments lagging behind load changes and easily causing service quality fluctuations when the load changes rapidly. Second, fixed threshold strategies are difficult to adapt to the different load characteristics of different base stations, often requiring manual adjustment based on experience, lacking flexibility and adaptability. Third, existing energy-saving methods generally lack in-depth mining of load data, failing to fully utilize the spatiotemporal correlation characteristics of the load for accurate prediction and resource scheduling. Fourth, there is a coupling problem between hardware resources and control strategies, resulting in limited precision in energy-saving control and difficulty in achieving 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 effects.

[0004] To address the challenge of balancing energy saving and service quality caused by insufficient load prediction accuracy, an innovative neural network architecture is employed to simultaneously capture the temporal variation patterns of load and the spatial correlation characteristics between parameters, significantly improving the accuracy of load prediction. Building upon this, a hierarchical quantization pruning technique is used to lightweight the network, adapting to the resource constraints of base station processors and solving the problem of deploying complex models at the edge. At the energy-saving control level, a prediction-driven differentiated strategy generation and adaptive execution mechanism are adopted to achieve precise energy saving while ensuring service quality. Summary of the Invention

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

[0006] In a first aspect, this application provides an energy-saving optimization control method for communication base stations. The method includes: collecting and preprocessing 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 computation through temporal feature processing branches and spatial feature processing branches to obtain fused load prediction data; performing hierarchical 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, to obtain a compressed network structure suitable for base station deployment; and processing real-time load data based on the compressed network structure to generate a dynamic energy-saving control strategy for the base station that includes radio frequency resource configuration instructions and power adjustment parameters.

[0007] Secondly, this application provides an energy-saving optimization control system for a communication base station, the energy-saving optimization control system for a communication base station comprising:

[0008] The data acquisition module is used to collect and preprocess traffic data and energy consumption data of communication base stations to obtain standardized multidimensional feature data.

[0009] The input module is used to input the standardized multidimensional feature data into a dual-stream spatiotemporal convolutional neural network, and perform parallel computation through temporal feature processing branches and spatial feature processing branches to obtain fused load prediction data;

[0010] The trimming module is used to perform hierarchical quantization trimming on the dual-stream spatiotemporal convolutional neural network, and allocate different bit widths according to the degree of influence of each network layer on the accuracy of the fused load prediction data, so as to obtain a compressed network structure suitable for base station deployment.

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

[0012] Thirdly, an energy-saving optimization control device for a communication base station is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the energy-saving optimization control device for the communication base station to execute the above-described energy-saving optimization control method for the communication base station.

[0013] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored therein, which, when executed on a computer, cause the computer to perform the above-described energy-saving optimization control method for a communication base station.

[0014] The technical solution provided in this application collects and preprocesses traffic and energy consumption data from communication base stations to obtain standardized multidimensional feature data. This solves the problems of strong heterogeneity and high noise interference in the original data, laying a data foundation for subsequent accurate prediction. The standardized multidimensional feature data is input into a dual-stream spatiotemporal convolutional neural network, which performs parallel computation through temporal and spatial feature processing branches. This not only overcomes the limitations of traditional single models in simultaneously handling temporal changes and feature interactions, but also effectively improves computational efficiency through the dual-stream parallel architecture. It achieves comprehensive capture of load change patterns and inter-feature relationships, significantly improving the accuracy of fused load prediction data and making energy-saving control decisions more forward-looking. The dual-stream spatiotemporal convolutional neural network is then subjected to hierarchical quantization pruning. Different bit widths are allocated according to the influence of each network layer on the accuracy of fused load prediction data, resulting in a compressed network structure suitable for base station deployment. This effectively solves the problem of difficult deployment of complex neural network models 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. This invention processes real-time load data using a compressed network structure to generate a dynamic energy-saving control strategy for base stations, including RF resource configuration instructions and power adjustment parameters. Through differentiated resource scheduling and refined power control, it achieves an optimal balance between energy saving and service quality, avoiding the service quality fluctuations or limited energy-saving effects caused by the one-size-fits-all approach in traditional methods. The actual load characteristics of communication base stations guide the design of a dual-stream spatiotemporal convolutional network, enabling the algorithm to specifically extract the spatiotemporal patterns of the load. Hierarchical quantization pruning technology performs customized network compression based on base station hardware constraints, adapting the algorithm to the deployment environment. The dynamic energy-saving control strategy formulates differentiated execution schemes based on communication quality requirements and resource scheduling characteristics, transforming the algorithm output into actual energy-saving effects. This mutual support between technical and algorithmic features gives this invention a significant advantage in solving the energy-saving problem of communication base stations, enabling a substantial reduction in base station energy consumption while ensuring communication service quality. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a schematic diagram of one embodiment of the energy-saving optimization control method for communication base stations in this application.

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

[0018] Figure 3 This is a schematic block diagram of the energy-saving optimization control device for communication base stations in an embodiment of the present invention. Detailed Implementation

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

[0020] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the energy-saving optimization control method for communication base stations in this application includes:

[0021] Step S101: Collect and preprocess the traffic data and energy consumption data of the communication base station to obtain standardized multidimensional feature data;

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

[0023] Step S103: Perform hierarchical quantization pruning on the dual-stream spatiotemporal convolutional neural network, and allocate different bit widths 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.

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

[0025] It is understood that the executing entity of this application can be an energy-saving optimization control system for communication base stations, or it can be a terminal or a server; no specific limitation is made here. This application's embodiments use a server as an example for illustration.

[0026] Specifically, the system records the radio frequency power value, number of channels used, number of access users, data transmission volume, power consumption value, and temperature value of the communication base station every minute. In practical applications, the data acquisition module obtains this raw data from the base station management system interface and automatically stores it as a raw data table. Raw data often contains noise and missing values, thus requiring data cleaning. Outliers exceeding the normal range are identified using statistical threshold detection methods and then replaced with the average value of the preceding and following time points. For example, if the radio frequency power value at a certain 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 sequence is then standardized by subtracting the average value of the corresponding feature from the value of each feature and dividing by its standard deviation, converting it to a zero-mean unit variance distribution to address the incomparability of data with different dimensions. For the standardized data, the linear relationship between features is further calculated, constructing a feature relationship strength matrix to form a parameter correlation table. Based on this correlation table, the importance score of each feature is analyzed, redundant and low-value features are removed, and finally, standardized multidimensional feature data is obtained. The core innovation of this method lies in inputting standardized multidimensional feature data into a dual-stream spatiotemporal convolutional neural network. This network consists of two parallel processing branches: a temporal feature processing branch and a spatial feature processing branch. First, the data is split into two parts: time series and feature relationships, which are then input into the two processing branches respectively. The temporal feature processing branch employs a variable-length receptive field one-dimensional convolutional chain structure, capturing load fluctuation patterns across different time spans through a dilated convolution mechanism. Dilated convolution is a special convolution operation that increases the receptive field by inserting holes in the convolution kernel while maintaining the same 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 features. Simultaneously, the spatial feature processing branch processes feature relationship data through a self-association graph structure, extracting the implicit dependencies between different monitoring parameters. The self-association graph structure establishes a connection strength map between features through cross-channel feature mapping, forming a parameter relationship matrix. After being processed by their respective branches, the temporal and spatial features are input into the temporal gating unit and the feature reconstruction network for further processing, generating adaptive temporal features and nonlinear spatial features. Finally, these two types of features are combined through a bidirectional attention fusion mechanism to form fused load prediction data.

[0027] Layered quantization pruning of the dual-stream spatiotemporal convolutional neural network aims to address the issues of large model size and high computational complexity, enabling efficient operation on resource-constrained base station controllers. This step first involves sensitivity testing of the network on real communication base station traffic data, calculating the impact of each layer's parameters on load prediction errors during peak and off-peak periods. Parameter sensitivity is determined using a perturbation test method, adding small perturbations to each layer's parameters and observing their impact on prediction 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 identifying daily fluctuation patterns, and a low-sensitivity layer extracting basic load trends. Different quantization strategies are employed for layers of varying sensitivity: the high-sensitivity layer retains all burst traffic identification functions using 8-bit fixed-point quantization; the medium-sensitivity layer optimizes daily traffic pattern identification parameters using 6-bit fixed-point quantization; and the low-sensitivity layer simplifies long-term trend extraction parameters using 4-bit fixed-point quantization. This differentiated approach significantly reduces model storage and computational requirements while maintaining prediction accuracy, resulting in a compressed network structure suitable for base station deployment. Real-time load data is then processed based on this compressed network structure to generate dynamic energy-saving control strategies for base stations. First, real-time collected load data is input into the compressed network structure for forward computation to generate load prediction data for future time periods. Based on the prediction results, the time period is divided into high-load, medium-load, and low-load intervals, forming a load status table. Then, by determining whether the duration of the low-load interval exceeds a threshold (e.g., 30 minutes), it is decided whether to trigger a deep power-saving mode or a basic power-saving mode. For the determined power-saving mode, a corresponding resource scheduling scheme is generated, including carrier switching status, antenna configuration, and channel allocation instructions. Simultaneously, based on the changing trends of the load prediction data, the optimal transmit power curve for each time period is calculated, generating power adjustment parameters. Finally, the resource scheduling scheme and power control scheme are organized in chronological order, with execution conditions and a safety backoff mechanism added to form a complete dynamic power-saving control strategy for the base station.

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

[0029] Record the radio frequency power value, number of channels used, number of access users, data transmission volume, power consumption value and temperature value of the communication base station at a frequency of one minute to obtain the 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 preceding and following time points to obtain a complete data sequence;

[0031] Subtract the corresponding feature mean from each value in the complete data sequence and divide by the standard deviation to convert it into a zero-mean unit variance distribution, thus obtaining uniform-scale data.

[0032] Calculate the linear relationship measure for each pair of data at the same scale, construct the relationship strength matrix between features, and obtain the parameter association table;

[0033] Based on the parameter association table, the importance score is analyzed, and feature columns with importance scores below a set threshold are removed to obtain the key feature set;

[0034] Organize the key feature set in chronological order, add time period identifiers and load status labels to obtain standardized multidimensional 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. These parameters include radio frequency power, number of channels used, 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, usually measured in watts. The number of channels used indicates the number of currently active wireless channels. The number of connected users records the number of terminal devices connected to the base station. Data transmission volume is the total uplink and downlink data volume per unit time, usually measured in bits. Power consumption records the overall energy consumption of the base station, including the total energy consumption of the radio frequency unit, baseband processing unit, transmission unit, and auxiliary equipment. Temperature records the ambient temperature of the base station, as one of the factors affecting energy consumption. This data is organized and stored in timestamp order, forming a raw data table. The raw data table usually contains outliers and missing values, requiring data cleaning. For values ​​exceeding upper and lower thresholds, a sliding window is constructed, typically set to 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 deviating from the mean by more than three times the standard deviation (μ ± 3σ) are marked as outliers. For example, if the number of access users suddenly jumps from 100 to 900 at a certain moment, while the average number of access users over the preceding and following 30 minutes is 120 with a standard deviation of 30, this value clearly exceeds the normal fluctuation range and needs to be replaced. The 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. For missing values, the average of the preceding and following time points is also used to fill in the missing values ​​to ensure data continuity. The complete data sequence obtained after this processing eliminates noise interference and more accurately reflects the trend of base station load changes. The complete data sequence contains features with different dimensions, and direct comparison and calculation will lead to the features with larger values ​​dominating the model training process. Therefore, standardization is required. The value of each feature is subtracted from the mean of that feature and then divided by its standard deviation to convert it into a zero-mean unit variance distribution. Specifically, the mean and standard deviation of each column of features are calculated, and then the Z-score standardization formula is applied to each data point for transformation. For example, if a base station's average data transmission volume over 24 hours is 500MB with a standard deviation of 150MB, then a data point with an original value of 800MB will have a standardized 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 numerical ranges and are weighted equally, thus obtaining data with a uniform scale.

[0036] A correlation analysis between features is performed on data of a uniform scale. Linearity measures are calculated pairwise to construct a feature relationship strength matrix. The Pearson correlation coefficient is used to measure the correlation between any two features. For features X and Y, their means are calculated first, then their covariance and standard deviation are calculated, and finally, the correlation coefficient is calculated. The correlation coefficient ranges from -1 to 1, with values ​​close to 1 indicating a strong positive correlation, close to -1 indicating a strong negative correlation, and close to 0 indicating no correlation. For example, when calculating the correlation coefficient between the number of access users and the amount of data transmitted, a result of 0.92 indicates a high positive correlation between these two features; a correlation coefficient of -0.15 between temperature and radio frequency power indicates a weak correlation. By calculating the correlation coefficients between all feature pairs, an n×n relationship strength matrix, i.e., a parameter correlation table, is constructed.

[0037] Based on the parameter correlation table analysis, feature importance scores are used to identify and remove redundant or low-value features. The importance analysis employs the information gain method, calculating the information gain of each feature on 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 deletion. Simultaneously, for feature pairs with a correlation higher than 0.85, the one with the higher information gain is retained, and the other is deleted to reduce redundancy. For example, if the correlation coefficient between the number of connected users and the data transmission volume is 0.92, and their information gains on energy consumption are 0.65 and 0.58 respectively, then the number of connected users is retained, and the data transmission volume feature is deleted. An importance threshold is set, and feature columns with scores below the threshold are removed to obtain the key feature set. The key feature set is reorganized chronologically, adding time period identifiers and load status labels. Time period identifiers are divided according to the time of day into morning peak (7:00-10:00), daily hours (10:00-17:00), evening peak (17:00-21:00), and nighttime hours (21:00-7:00). Load status labels categorize each time point as high, medium, or low load based on key load metrics (such as the number of connected users or data transmission volume). For example, a normalized data transmission volume value greater than 1.5 at a given time is labeled as high load; between -0.5 and 1.5 as medium load; and less than -0.5 as low load. These time period identifiers and load status labels provide crucial reference information for subsequent load prediction and resource scheduling. The resulting standardized multidimensional feature data includes a cleaned, standardized, and filtered feature set, along with additional time period and load status labels.

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

[0039] The standardized multidimensional feature data is split into two parts: time series and feature relationship, and then input 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 captured by a one-dimensional convolution chain with variable receptive field and an expanded convolution mechanism to capture the load fluctuation patterns at different time spans, thus obtaining a multi-granularity time series representation.

[0041] A self-association graph structure is applied to the feature relationship part of the initial dual-stream input, and the implicit dependencies between parameters are extracted through the feature mapping of cross channels to obtain the parameter relationship matrix;

[0042] The multi-granularity time series characterization 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 for significance through a feature recombination network, and high-influence feature combinations and interaction patterns are dynamically extracted to obtain nonlinear spatial features.

[0044] A bidirectional attention fusion of adaptive temporal features and nonlinear spatial features is performed, and a spatiotemporal collaborative representation is constructed through a complementary information enhancement mechanism to obtain fused load prediction data.

[0045] Specifically, the dual-stream spatiotemporal convolutional neural network is the core module of the energy-saving optimization control method for communication base stations. First, it needs to split the standardized multidimensional feature data into two parts: a time series and feature relationships. The time series part contains the feature values ​​arranged in chronological order, used to capture the pattern of load changes over time; the feature relationship part focuses on the mutual influence and dependencies between different features. During the splitting process, continuous values ​​of all features in the time dimension are extracted from the standardized multidimensional feature data to form the time series part, while the correlation strength information between features is organized into the feature relationship part. Specifically, for each time point t, all feature values ​​at that time point are extracted to form a vector Xt, and the vectors from all time points are concatenated in order to form the sequence {X1, X2, ..., Xt, ...} as the time series part; simultaneously, a feature relationship matrix R is constructed based on the parameter correlation table generated in the previous step, where each element Rij in the matrix represents the correlation strength between feature i and feature j. These two parts of data are respectively input into the asymmetric dual-stream processing structure of the parallel architecture to form the initial dual-stream input.

[0046] When processing the time-series portion of the initial dual-stream input, a variable-length receptive field one-dimensional convolutional chain structure is employed. This is a special neural network structure capable of simultaneously capturing patterns across different time spans. The variable-length receptive field is achieved through dilated convolution, which introduces gaps (diplets) into the standard convolution operation. In standard convolution, elements in the kernel 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 gap; a dilation rate of 1 is equivalent to standard convolution, and the receptive field grows exponentially with increasing dilation rate. In this method, multiple levels of dilated convolutional layers are constructed with dilation rates of 1, 4, and 16 to capture short-term (minute-level), medium-term (hour-level), and long-term (day-level) load variation patterns, respectively. Specifically, a one-dimensional convolution with a dilation rate of 1 is first applied to the time-series data to extract the relationships between adjacent time points. The result is then input into a convolutional layer with a dilation rate of 4 to capture patterns over a wider time range, and finally, a convolutional layer with a dilation rate of 16 is used to obtain the long-term trend. This layered structure forms feature representations at different time scales, i.e., multi-granular temporal representation.

[0047] The feature relationship portion of the initial dual-stream input is processed using a self-associative graph structure. A self-associative graph structure is a computational structure specifically designed to handle relationships between features. It constructs a feature relationship graph by treating features as nodes in a graph and the association strength as the weight of the edges. The processing includes: first, constructing an initial graph structure based on the feature relationship matrix; then, updating the relationship between each pair of features through cross-channel feature mapping operations. Cross-channel feature mapping means that the influence of feature i on feature j depends not only on their direct association Rij but also on the indirect influence of other features. Through multiple rounds of iterative updates, higher-order relationships between features are calculated, forming a more accurate parametric relationship matrix. During the update process, a nonlinear transformation is applied to the relationship matrix R in each iteration: R' = σ(RWR). T ), where W is a learnable weight matrix and σ is a non-linear activation function such as ReLU or tanh. This update method can uncover the implicit dependencies between features, thus obtaining a more comprehensive parameter relationship matrix.

[0048] Multi-granularity time-series representations are further processed by a time-series gating unit, a mechanism that dynamically adjusts the weights of different time windows based on historical load change trends. This unit receives multi-granularity time-series representations as input, analyzes the importance of features at different time scales, and assigns them appropriate weights. Specifically, the time-series gating unit first extracts the historical load change trend features, then calculates the correlation between each time-scale feature and the historical trend, with time scales showing higher correlation receiving greater weight. For example, when historical loads exhibit significant periodic changes, time-scale features capturing the corresponding period length receive higher weights; when loads exhibit sudden changes, the weights of short-term time-scale features increase. This dynamic weight allocation mechanism allows the model to adaptively focus on the most relevant time-scale information, thus forming adaptive time features.

[0049] The parameter relationship matrix is ​​further analyzed and processed through a feature reconstruction network. The feature reconstruction network is a computational structure used to analyze the importance of feature combinations, extracting high-influence feature interaction patterns through multi-layer nonlinear transformations. The network first applies a threshold to each element in the parameter relationship matrix, retaining strong correlations; then, a multi-layer perceptron performs a nonlinear transformation on the filtered matrix to extract high-order feature combinations; finally, saliency analysis identifies key feature interaction patterns. The saliency analysis employs an attention mechanism, calculating the contribution of each feature combination to load prediction, with combinations of higher contributions receiving higher attention weights. In this way, the feature reconstruction network can dynamically extract the most valuable feature combinations and interaction patterns for load prediction, forming nonlinear spatial features. A bidirectional attention fusion mechanism combines adaptive temporal features and nonlinear spatial features to generate fused load prediction data. Bidirectional attention fusion is a fusion method that considers mutual reinforcement between features, unlike simple feature concatenation or weighted averaging. This mechanism includes attention calculations in two directions: attention of temporal features to spatial features and attention of spatial features to temporal features. Specifically, firstly, the attention weight of each element in the adaptive temporal feature to the nonlinear spatial feature is calculated to form a time-guided spatial feature; simultaneously, the attention weight of the nonlinear spatial feature to the adaptive temporal feature is calculated to form a spatially enhanced temporal feature; then, these two enhanced features are adaptively fused through a gating mechanism to generate a fused feature containing both temporal and spatial information. Finally, the fused feature is mapped to the target dimension through a fully connected layer 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 dual-stream spatiotemporal convolutional neural network processed seven days of historical data containing five features: RF power, channel usage, number of access users, power consumption, and temperature. After data splitting, the time series portion formed a sequence of 10,080 time points (one point per minute), and the feature relationship portion formed a 5×5 correlation matrix. After three levels of dilated convolution, the time series generated a multi-granularity time series representation containing short-term (last 4 hours), medium-term (last 24 hours), and long-term (7 days) features. For example, the dilated convolution revealed a clear bimodal load pattern on weekdays (9 AM and 8 PM), while weekends exhibited a unimodal pattern (around 3 PM). After analyzing these patterns, the time series gating unit assigned higher weights to short-term and daily periodic patterns when predicting weekday load, while focusing more on periodic patterns when predicting weekend load. Simultaneously, feature correlation analysis revealed a strong correlation between the number of access users and RF power and channel usage; the feature reorganization network extracted these highly correlated feature combinations as key predictive factors. Ultimately, through bidirectional attention fusion, the model accurately predicted the base station load change trend for 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 in simultaneously handling temporal changes and feature interactions, significantly improving the accuracy and robustness of predictions.

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

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

[0053] By applying a residual connection mechanism to the original expanded feature set, the input features are added and fused with the convolutional output features, preserving the original temporal information, and thus obtaining the enhanced expanded features.

[0054] The enhanced dilated features are processed by grouped convolution to extract features of different frequencies in groups, resulting in a frequency-separated feature set.

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

[0056] A feature importance graph is constructed based on three types of temporal semantic features. The weight distribution of different time windows and different feature types is calculated to obtain a weighted feature graph.

[0057] We perform cross-window fusion on the weighted feature maps and use a self-attention mechanism to associate feature representations at different time scales to obtain multi-granular temporal representations.

[0058] Specifically, the time series portion is divided into three sequence segments based on time granularity: minute-level, hour-level, and day-level, forming a multi-scale input sequence. For these input sequences, a three-level cascaded dilated convolutional layer structure is designed, with dilation rates set to 1, 4, and 16, respectively. Dilated convolution is a special type of convolution operation that increases the receptive field by inserting holes between the convolutional kernels while keeping the number of parameters constant. The dilation rate represents the spacing between kernel elements; a dilation rate of 1 is equivalent to standard convolution, and as the dilation rate increases, the receptive field grows exponentially. The first-level dilated convolutional layer uses a kernel with a dilation rate of 1 to process minute-level sequence segments. The kernel slides directly between adjacent time points, capturing short-term load changes, such as sudden increases or decreases in traffic. The second-level dilated convolutional layer uses a kernel with a dilation rate of 4 to process hour-level sequence segments. The kernel skips four time points each time for calculation, capturing medium-term load change patterns, such as cycles between peak and off-peak periods. The third-level dilated convolutional layer uses a kernel with a dilation rate of 16 to process day-level sequence segments. The kernel spans a larger time range, capturing long-term load trends, such as traffic differences between weekdays and weekends. This multi-level structure with increasing dilation rates effectively covers various load patterns from short to long time periods. The output features of the three-level dilated convolutions are merged to form the original dilated feature set. Although the original dilated feature set captures features at different time scales, it is prone to losing original temporal information during multiple convolutional processes. To address this issue, a residual connection mechanism is applied to enhance the original dilated feature set. Residual connections are a technique that directly adds input features to the convolutional output features, implemented through shortcut connections. Specifically, after each dilated convolutional layer outputs, the input and output features of that layer are added element-wise. For example, for the first dilated convolutional layer, if the input feature is X1 and the convolutional output is F1(X1), the result after the residual connection is X1+F1(X1). Similarly, the second and third dilated convolutional layers apply the same residual connection mechanism; the result for the second layer is X2+F2(X2), and the result for the third layer is X3+F3(X3). This residual connection mechanism effectively preserves the original temporal information, prevents the gradient vanishing problem in deep network training, and enables the network to learn the residual mapping between the input and output, forming enhanced dilated features.

[0059] Enhanced extended features contain a mixture of information from multiple time frequencies. To analyze different frequency features more precisely, these enhanced extended features are processed using grouped convolution. Grouped convolution is a technique that divides the input feature channels into multiple groups, each of which undergoes convolution operations independently. In this method, the enhanced extended features are divided into three groups according to their frequency characteristics: high-frequency group, mid-frequency group, and low-frequency group. The high-frequency group corresponds to short-term rapidly changing features, the mid-frequency group corresponds to mid-term fluctuating features, and the low-frequency group corresponds to long-term trend features. Each group is processed using convolution kernels with different parameters. The high-frequency group uses convolution kernels with smaller receptive fields to capture detailed changes, the mid-frequency group uses convolution kernels with medium receptive fields to extract periodic patterns, and the low-frequency group uses convolution kernels with large receptive fields to extract long-term trends. The output features of the grouped convolutions are combined to form a frequency-separated feature set, clearly distinguishing the load features of different frequencies. Multi-head temporal coding is performed on 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, consisting of three heads: a periodic pattern head, a trend pattern head, and a burst pattern head. The periodic pattern head focuses on extracting the cyclical variation patterns of load, using Fourier transform and autocorrelation analysis to identify periodic components in the signal. Examples include load differences between weekdays and weekends, and peak and off-peak patterns at different times of the day. The trend pattern head focuses on long-term load trends, using moving averages and linear regression to filter out short-term fluctuations and extract the overall trend. The burst pattern head specifically identifies unconventional, sudden load events, employing anomaly detection algorithms to calculate the rapid rate of change and deviation of the signal, identifying points of sudden increases or decreases in load. These three heads process the frequency-separated feature set in parallel, outputting three types of temporal semantic features: periodic features, trend features, and burst features.

[0060] The feature importance map, constructed based on three types of temporal semantic features, aims to evaluate 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 (e.g., the most recent hour, the most recent four hours, the most recent day, etc.) and the vertical axis representing different feature types (periodicity, trend, burstiness). Each element in the matrix represents the importance weight for the corresponding time window and feature type. The importance weight is calculated based on the feature's contribution to load prediction, using an attention score method. For each time window w and feature type t, the conditional mutual information I(y; f_t|w) for its contribution to load prediction y is calculated. The higher the mutual information value, the more important the feature is to the prediction within that time window. Then, the mutual information value is normalized into weights using the softmax function, 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 multi-granularity temporal representations. Cross-window fusion employs a self-attention mechanism, allowing features from 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 used 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 the attention score. Next, the attention score is normalized using softmax. Finally, the normalized score is used to perform a weighted summation of the value vectors to obtain the fused feature representation. This mechanism enables highly correlated windows to enhance each other, thereby integrating information from different time scales to form a comprehensive multi-granularity temporal representation.

[0062] In one specific embodiment, the process of constructing a multi-level dilated convolutional structure for the time series portion may specifically include the following steps:

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

[0064] A first-level dilated convolutional layer is constructed on the multi-scale input sequence, with the dilation rate set to one, so that the convolutional kernel slides between adjacent time points to capture the load relationship between adjacent time points and obtain short-term dilated features.

[0065] The short-term dilation feature is input into the second-level dilated convolutional layer with a dilation rate of four, so that the convolutional kernel slides between four time points to capture load changes across time periods and obtain the medium-term dilation feature.

[0066] The intermediate-term dilation feature is input into the third-level dilated convolutional layer with a dilation rate of sixteen. This allows the convolutional kernel to slide between sixteen time points to capture long-term load trends and obtain long-term dilation features.

[0067] The short-term expansion features, medium-term expansion features and long-term expansion features are merged by channel splicing to retain information at each time scale and obtain multi-perspective time series features.

[0068] The importance scores of the features from multiple perspectives are calculated, and weight coefficients are assigned according to the correlation between the features at each time scale and the changes in historical load, thus obtaining the original expanded feature set.

[0069] Specifically, the time series portion is divided into three sequence segments based on time granularity: minute-level, hour-level, and day-level. The time series portion is a sequence of feature values ​​arranged chronologically within standardized multidimensional feature data, including records of how parameters such as base station radio frequency power, channel usage, and number of access users change over time. The segmentation process involves resampling and aggregation of the original sequence according to time granularity. The minute-level sequence segment maintains the original sampling frequency, typically one data point per minute, fully preserving short-term fluctuation information. The hour-level sequence segment is obtained by aggregating the original sequence by hour, taking the average, maximum, and minimum values ​​of 60 minutes of data for each hour to form features reflecting hourly-scale changes. The day-level sequence segment aggregates the original sequence by day, taking statistics from 24 hours of data each day to reflect the daily load variation pattern. This multi-scale segmentation yields input sequences with different time granularities, providing a data foundation for subsequent feature extraction at different scales. A first-level dilated convolutional layer is constructed on the multi-scale input sequences to capture the load relationship between adjacent time points. Dilated convolution is a convolution operation that introduces holes into standard convolution; the dilation rate determines the size of the holes. The first-level dilated convolutional layer has a dilation rate of 1, meaning there are no holes, equivalent to standard convolution, and the convolutional kernel slides directly between adjacent time points. In practice, a one-dimensional convolutional kernel of length 5 is used to perform convolution operations on minute-level sequence segments. The kernel weights are learned through backpropagation. The convolution operation involves multiplying the kernel with corresponding positions in the input sequence and summing the results. 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) with the kernel weights and then summing the results. This convolutional operation, which slides directly between adjacent time points, effectively captures load variation relationships within a local time range, such as burst traffic and short-term fluctuations, forming short-term dilated features that express short-term load variation patterns.

[0070] The short-term dilated features are then input into the second-level dilated convolutional layer to further extract medium-term load change patterns. The dilation rate of the second-level dilated convolutional layer is set to four, indicating that there is a three-position interval between convolutional kernel elements, thus expanding the actual time range covered by the convolutional kernel. If a convolutional 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 convolutional kernel is no longer continuously convolved with the input sequence, but rather a value is taken every four time points for convolution. For example, for time point t, the output feature calculation will consider the feature values ​​of five time points: t, t-4, t-8, t+4, and t+8, multiplied by the convolutional kernel weights and then summed. This convolution operation, which slides between four time points, can capture load change patterns over a wider time range, such as the medium-term change patterns of morning and evening peak hours, and the difference in traffic between working and non-working hours, forming medium-term dilated features that express medium-term load change patterns. The medium-term dilated features are then input into the third-level dilated convolutional layer to capture long-term load trends. The dilation rate of the third-level dilated convolutional layer is set to sixteen, meaning there are fifteen positions between the convolutional kernel elements, further expanding the time span covered by the kernel. Using a kernel of length 5, the actual time span covered is 1 + (5-1) × 16 = 65 time points, sufficient to capture load change trends spanning multiple hours or even days. Specifically, for time point t, the output feature calculation considers the feature values ​​of five relatively distant time points: t, t-16, t-32, t+16, and t+32. This convolutional operation, sliding between sixteen time points, can capture long-term load change patterns, such as the difference in traffic between weekdays and weekends, and special patterns during holidays, forming long-term dilated features that express long-term load change patterns.

[0071] After three levels of dilated convolution, short-term, medium-term, and long-term dilated features representing load characteristics at different time scales are obtained. To comprehensively utilize the information from these three time scales, channel concatenation is used to merge them into multi-view temporal features. Channel concatenation refers to connecting multiple feature vectors into a longer vector along the channel dimension of the features, preserving all information in 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 length, number of short-term feature channels], [batch size, time length, number of medium-term feature channels], and [batch size, time length, number of long-term feature channels], then the dimensions of the concatenated multi-view temporal features are [batch size, time length, number of short-term feature channels + number of medium-term feature channels + number of long-term feature channels]. This channel concatenation method ensures that information from different time scales is completely preserved.

[0072] The purpose of calculating feature importance scores for multi-perspective time-series features is to assign weights based on the importance of features at different time scales to the prediction. Feature importance analysis is based on the correlation between features at each time scale and historical load changes. The specific calculation process includes: first, extracting historical load change sequences, typically the load change rate within the most recent period (e.g., 48 hours); then, calculating the correlation coefficients between short-term, medium-term, and long-term extended features and the load change sequence, with higher correlation coefficients indicating greater importance of the time-scale feature for prediction; finally, normalizing the correlation coefficients into weight coefficients to ensure that the sum of all weights is 1. The weight coefficients are calculated using the softmax function, i.e., first taking the exponent of each correlation coefficient, then dividing by the sum of all exponents to obtain a normalized weight between 0 and 1. Through this correlation-based weight allocation method, features at different time scales in the multi-perspective time-series features are weighted and combined according to their importance to form the original extended feature set expressing multi-scale time features. In the practical application of energy-saving optimization control for communication base stations, the process of performing multi-level dilated convolutional analysis on the historical load data of a 5G base station is as follows: First, the load data collected by minute within a week (a total of 10080 data points) is divided into sequence segments of three time granularities. The minute-level data maintains the original sampling frequency; the hour-level data is obtained by averaging, taking the maximum and minimum values ​​of 60 points per hour to obtain a sequence of 168 points; and the day-level data is aggregated into a sequence of 7 points. Then, a first-level dilated convolutional layer (dilation rate = 1) is constructed to directly convolve the minute-level sequence with a convolution kernel length of 5, extracting the load change pattern within a local time window, such as the sudden increase in traffic at the start of office work at 9:00 AM. Next, a second-level dilated convolutional layer (dilation rate = 4) is constructed to process the features output by the first level, expanding the actual receptive field to 17 time points, which can capture patterns such as the decrease and then increase in traffic during lunchtime from 12:00 to 13:00. Then comes the third-level dilated convolutional layer (dilation rate = 16), expanding the receptive field to 65 time points to capture patterns over longer time spans, such as the evening peak load pattern from 18:00 to 22:00 daily. These three levels of dilated features are concatenated and merged through channels to form comprehensive multi-perspective temporal features. Finally, weights are assigned to them by analyzing the correlation between the three time-scale features and historical load changes. For example, when predicting daily load on weekdays, the medium-term feature (hourly pattern) has the highest correlation and receives a weight of 0.6; while when predicting load during weekend special events, the short-term feature (burst pattern) has a stronger correlation and receives a weight of 0.5; and when predicting load during long holidays, the long-term feature (daily pattern) has the highest correlation and receives a weight of 0.7. This correlation-based dynamic weight allocation ensures that the model can flexibly adjust its focus according to different prediction scenarios, improving prediction accuracy. Through the design and implementation of the above multi-level dilated convolutional structure, the technical problem that traditional load prediction methods cannot simultaneously capture load features at different time scales is solved.This method effectively expands the receptive field by using convolutional layers with increasing dilation rates, while maintaining a constant number of parameters. This overcomes the drawbacks of traditional methods, such as high computational complexity and large model size, when processing long-term series. The multi-view feature extraction and feature importance analysis mechanism enables the model to adaptively focus on the most relevant time-scale features in different prediction scenarios, improving prediction accuracy and robustness. This flexible and efficient prediction mechanism provides a reliable decision-making basis for energy-saving optimization control of communication base stations, solving the problems of low energy efficiency and unstable service quality caused by inaccurate load prediction.

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

[0074] Sensitivity tests were conducted on the dual-stream spatiotemporal convolutional neural network using 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 the base station scenario 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, thus obtaining a hierarchical label for communication load characteristics.

[0076] The high-sensitivity layer in the communication load characteristic hierarchical labeling retains all radio frequency burst traffic identification functions, and eight-bit fixed-point quantization is used to ensure the accuracy of burst traffic early warning, thus obtaining the burst traffic sensing structure.

[0077] The daily business mode identification parameters of the medium-sensitive layer in the communication load characteristic hierarchy are optimized, and six-bit fixed-point quantization is used to ensure the efficiency of resource allocation during normal periods, resulting in a normal traffic processing structure.

[0078] The parameters for extracting long-term trends in the low-sensitivity layer of the communication load characteristic hierarchy are simplified, and four-bit fixed-point quantization is used to meet the basic load prediction needs, thus obtaining the basic load evaluation structure.

[0079] The burst traffic sensing 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 based on 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 is performed on the dual-stream spatiotemporal convolutional neural network. Sensitivity testing is the process of evaluating the importance of parameters at each layer of the neural network, quantifying the impact of parameter changes on the prediction results. The specific operation involves selecting historical traffic data from real communication base stations, including samples from typical peak periods (e.g., weekday peaks from 9-11 AM and 5-7 PM) and off-peak periods (e.g., 2-5 AM). These samples are input into the trained dual-stream spatiotemporal convolutional neural network for prediction, and the original prediction results are recorded. Then, a small perturbation is applied to the parameters of each layer of the network. The perturbation method involves adding a small amount of Gaussian noise to the parameters, with the noise standard deviation typically set to 1% of the standard deviation of the parameters at that layer. After adding the perturbation, predictions are performed again using the same samples, and the difference in prediction results before and after the perturbation is compared to calculate the change in prediction error. This perturbation-prediction-comparison operation is performed multiple times for each layer, and the average error change is taken as the sensitivity index for that layer. The higher the sensitivity index, the greater the impact of the layer's parameters on the prediction results, and the more important they are. This method tests all layers of the network, ultimately yielding a complete sensitivity distribution related to base station scenarios—that is, the quantified sensitivity values ​​of each layer under different load scenarios. Based on this sensitivity distribution, the layers of the dual-stream spatiotemporal convolutional neural network are then classified by function and sensitivity. The classification criteria combine the functional role of the layers and the sensitivity index, dividing the network into three levels. High-sensitivity layers are those ranking in the top 30% of sensitivity indicators. These layers are typically located in the shallow layers of the network and in specific feature extraction modules, responsible for capturing bursty traffic characteristics, such as a sudden increase in user connection requests or traffic surges caused by large events. These layers are crucial for identifying short-term, bursty load changes; their high sensitivity stems from the fact that even small parameter changes can significantly affect the accuracy of predictions for bursty events. Medium-sensitivity layers are those ranking in the middle 40% of sensitivity indicators. These layers are typically located in the middle part of the network, responsible for identifying daily fluctuation patterns, such as the difference in traffic between weekdays and weekends, and periodic changes like regular peaks and troughs within a day. These layers are important for capturing regular changes in base station load, but their parameter stability requirements are slightly lower compared to bursty events. The low-sensitivity layer comprises the bottom 30% of layers in terms of sensitivity metrics. Primarily located in the deeper parts of the network, these layers are responsible for extracting basic load trends, such as long-term user growth and stable characteristics like seasonal variations. While these layers contribute significantly to understanding long-term trends, their parameter precision requirements are relatively low due to the slow changes in these trends. This classification based on sensitivity and function creates hierarchical markers for communication load characteristics, providing a basis for subsequent differentiated compression.

[0081] When processing the highly sensitive layers in the communication load characteristic hierarchy, a conservative compression strategy is adopted to retain all radio frequency burst traffic identification functions. Specifically, the network structure of these layers remains unchanged, and only the parameters are compressed in precision using 8-bit fixed-point quantization. Fixed-point quantization is the process of converting floating-point parameters into fixed-point representations. Eight-bit fixed-point quantization uses 8 bits (1 sign bit, 7 value bits) to represent each parameter. The quantization process first determines the dynamic range [min, max] of the parameter, then calculates the quantization step size, and finally converts the floating-point parameter value into an integer quantized value qval = round((val-min) / step). During storage, only the quantized value qval and the step size step need to be recorded. This 8-bit fixed-point representation maintains high numerical precision, ensuring that the ability to sensitively capture burst traffic changes is not affected. Through this fine-grained quantization processing, the highly sensitive layers can maintain accurate identification of burst events, forming a burst traffic sensing structure and providing reliable predictions for the base station's rapid response under burst load conditions.

[0082] The processing of the medium-sensitivity layer employs a moderate compression strategy, improving efficiency by optimizing parameters identified through daily service patterns. The medium-sensitivity layer is primarily responsible for identifying periodic changes and regular fluctuation patterns in base station load. For these characteristics, six-bit fixed-point quantization provides sufficient accuracy. Six-bit fixed-point quantization uses 6 bits (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, sparsification is applied to the medium-sensitivity layer by setting a small threshold (typically 5% of the average absolute value of the parameters) and directly setting parameters with absolute values ​​less than the threshold to zero. This sparsity operation reduces the number of non-zero parameters, further reducing storage and computational requirements, while having limited impact on the identification of regular daily load patterns. Through this optimization, the medium-sensitivity layer forms an efficient regular traffic processing structure, accurately identifying the daily load variation patterns of base stations and providing a basis for regular resource scheduling.

[0083] A more aggressive compression strategy is applied to the low-sensitivity layer, reducing computational complexity by simplifying long-term trend extraction parameters. The low-sensitivity layer is primarily responsible for extracting long-term trends in base station load, which are typically slow and smooth, requiring lower parameter precision. Therefore, four-bit fixed-point quantization is used for compression, representing each parameter with 4 bits (1 sign bit, 3 value bits). While this four-bit representation has lower precision, it is sufficient for capturing 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, removing unimportant neurons (the 20%-30% with the smallest L1 norm), and adjusting connections accordingly. This structural simplification significantly reduces the number of model parameters and computational complexity without significantly impacting the capture of long-term trends. Through these simplifications, the low-sensitivity layer forms a lightweight basic load assessment structure, effectively extracting long-term trends in base station load and providing a reference for long-term resource planning.

[0084] The final step is to integrate the burst traffic sensing structure, the regular traffic processing structure, and the basic load assessment structure into a lightweight prediction engine for the base station controller. The integration process first establishes the data flow relationships between the three structures to ensure the correct forward propagation order of the network. Then, the computation process is optimized, primarily including: batch processing optimization, inputting multiple samples into the network for computation at once to reduce redundant operations; memory access optimization, adjusting the data storage layout to reduce memory access conflicts; and parallel computing optimization, adjusting the allocation of computational tasks based on the parallel computing capabilities of the base station processor. Furthermore, algorithm-level optimization is performed, merging multiple consecutive linear operations into one to reduce the storage and loading of intermediate results. The integrated network structure undergoes final verification to ensure that prediction accuracy is maintained even after compression, especially under burst load conditions. Through this series of integration and optimization operations, a compressed network structure suitable for base station deployment is finally obtained. This structure significantly reduces storage requirements and computational complexity while maintaining prediction accuracy, and can run efficiently on the embedded processor of the base station.

[0085] In a practical application scenario, the process of using this method to compress a dual-stream spatiotemporal convolutional neural network in a 5G communication base station is as follows: First, load data from the base station over the past 30 days is selected, with a particular focus on extracting samples from weekday morning and evening peak hours (traffic reaching over 150% of peak levels between 8-10 AM and 6-8 PM) and nighttime off-peak hours (traffic dropping below 30% of the daily average between 1-5 AM) for sensitivity testing. During the test, Gaussian noise with a standard deviation of 1% of the parameter values ​​was added to each of the network's 18 layers. It was found that the first two convolutional and attention layers in the initial temporal feature processing branch exhibited the highest sensitivity, with a prediction error change rate exceeding 15%; the intermediate feature fusion module and some fully connected layers showed moderate sensitivity, with prediction error change rates between 5% and 15%; while the deeper output network and some functional auxiliary layers showed lower sensitivity, with prediction error change rates not exceeding 5%. Based on this sensitivity distribution, the network is divided into three layers: a high-sensitivity layer, comprising the first two convolutional layers and an attention layer of the temporal feature processing branch, responsible for identifying burst traffic; a medium-sensitivity layer, comprising a feature fusion module and a key fully connected layer, responsible for identifying daily load patterns; and a low-sensitivity layer, comprising an output network and an auxiliary function layer, responsible for extracting long-term trends. 8-bit fixed-point quantization is applied to the high-sensitivity layer, quantizing floating-point parameters to integers within the range [-128, 127]. 6-bit fixed-point quantization and sparsification are applied to the medium-sensitivity layer, setting parameters less than 5% of the mean to zero, 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 optimization is performed for the characteristics of the ARM processor in the base station, merging continuous linear operations and optimizing memory access patterns. The final compressed network reduces the model size by 85% and the inference time by 78% without decreasing burst load prediction accuracy by more than 2%, meeting the resource constraints of the base station controller.

[0086] By employing the aforementioned hierarchical quantization pruning and differentiated compression strategies, the technical challenges faced by traditional neural network models in communication base station deployments are addressed. The differentiated quantization strategy ensures high precision preservation of important layers and effective compression of secondary layers, resolving the conflict between model accuracy and resource consumption. The function- and sensitivity-oriented compression method ensures the model's accuracy in critical load prediction scenarios, particularly its ability to warn of sudden traffic surges, overcoming the problem of insufficient prediction accuracy after compression in existing technologies. Optimization of the computational process tailored to the characteristics of base station processors solves the problem of low efficiency for complex models operating in resource-constrained environments. This optimization method, which maintains both accuracy and efficiency, provides reliable decision support for base station energy-saving control, effectively addressing issues such as low energy efficiency and unstable network service quality.

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

[0088] Real-time load data is input into a compressed network structure, and load change curves for future periods are generated through forward computation to obtain load prediction 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 the load status table.

[0090] Determine whether the duration of the low load interval in the load status table exceeds 30 minutes. If it does, trigger the deep energy-saving mode; if it does not, trigger the basic energy-saving mode and obtain the energy-saving mode identifier.

[0091] Based on the energy-saving mode identifier and load status table, calculate the radio frequency resource demand for the time period, generate carrier switching status, antenna configuration and channel allocation instructions, and obtain the resource scheduling scheme.

[0092] Based on the resource scheduling scheme and the load prediction data change trend, the transmit power curve for each period is calculated, and power level, adjustment time point and step size parameters are generated to obtain the power control scheme.

[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 a compressed network structure for processing. This real-time load data includes base station operating parameters such as RF power values, number of channels used, number of access users, and data transmission volume. This data is input into the network after standardization processing consistent with the training data. The compressed network structure is a dual-stream spatiotemporal convolutional neural network that has undergone hierarchical quantization pruning. It contains three types of sensitivity layers: high, medium, and low, responsible for capturing burst traffic, identifying daily patterns, and extracting long-term trends, respectively. The network generates load prediction data through a forward computation process, which is the process of data transfer from the input layer to the output layer: time series and feature relationship data are processed through temporal and spatial feature processing branches, respectively, and then through a series of operations including convolutional layers, pooling layers, and fully connected layers, ultimately outputting the load prediction value for the future period. The prediction range is typically 6 hours into the future, with 5-minute intervals, forming a load change curve.

[0095] Segmenting predicted load data into states is fundamental to developing differentiated energy-saving strategies. The segmentation process uses a load rate threshold method. The load rate is the ratio of the actual load of a base station to its designed capacity, typically calculated based on key indicators such as the proportion of access users to the maximum number of users or the proportion of data throughput to the maximum capacity. Three levels of segmentation criteria are set: periods with a load rate greater than 70% are classified as high-load intervals, indicating high user demand and requiring full resource activation; periods with a load rate between 30% and 70% are classified as medium-load intervals, suitable for partial energy saving; and periods with a load rate below 30% are classified as low-load intervals, indicating severely idle resources and suitable for deep energy saving. The segmentation operation involves traversing each predicted time point, comparing its load rate with the threshold to determine its state, and then merging adjacent time points with the same state into continuous intervals, ultimately generating a load state table containing multiple time intervals and their corresponding load states.

[0096] The power-saving mode determination based on the load status table focuses on assessing the persistence of low-load intervals. The determination 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 a set threshold of thirty minutes, a deep power-saving mode (marked "deep") is triggered; when the duration of all low-load intervals does not exceed thirty minutes, a basic power-saving mode (marked "basic") is triggered. The duration threshold is set to take into account resource switching overhead: while deep power-saving offers good energy savings, carrier and antenna configuration switching takes time, and excessively frequent switching can increase energy consumption. Therefore, it is only activated when the low load remains stable for a relatively long period. This duration-based mode determination ensures the stability and applicability of the power-saving strategy.

[0097] Based on the energy-saving mode identifier and load status table, the required radio frequency resources for each time period are calculated, and a resource scheduling scheme is generated. The calculation process first estimates the number of radio resource units required for each time period based on predicted load and service quality requirements. The calculation method converts the predicted number of users and data volume into the required number of resource blocks, and then determines the minimum resource configuration to meet the requirements based on the resource block capacity provided by each carrier and antenna configuration. Differentiated resource allocation strategies are adopted for different load intervals: in high-load intervals, all resources remain active; in medium-load intervals, the number of carriers remains unchanged but power is adjusted in basic energy-saving mode, and some carriers can be shut down in deep energy-saving mode; in low-load intervals, power and channel allocation are mainly adjusted in basic energy-saving mode, and some carriers can 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 scheme, organized chronologically to form a complete resource scheduling scheme. Based on the resource scheduling scheme and predicted load trends, a transmit power control curve is calculated, and a power control scheme is generated. First, a base power level is determined to ensure coverage requirements are met even at the lowest power level. The calculation method is based on link budget, considering coverage radius, path loss, and receiver sensitivity to derive the transmit power that meets the minimum signal requirements. Then, the timing and step size for power adjustments are determined based on the predicted load change trend. The timing is chosen considering the lead time for load state transitions, typically starting power increases 5-10 minutes before the load increases. The step size is calculated based on the load change rate; larger steps are used for rapid changes, and smaller steps for slow changes. The power control scheme includes the power level, adjustment timing, and step size parameters for each time point, which, in conjunction with the resource scheduling scheme, form a comprehensive energy consumption control strategy.

[0098] Finally, the resource scheduling scheme and power control scheme are integrated into a complete dynamic energy-saving control strategy for base stations. The integration process first organizes the instructions of the two schemes in chronological order to ensure consistency between resource allocation and power adjustment. Then, execution conditions and a fallback mechanism are added to ensure the security of strategy execution. Execution conditions include prediction reliability checks, such as setting a threshold for deviation between prediction and actual load, suspending execution if this threshold is exceeded; the fallback mechanism includes service quality monitoring triggers, such as setting monitoring thresholds for key performance indicators, immediately interrupting the energy-saving mode when indicators deteriorate. The final dynamic energy-saving control strategy for base stations is a complete control scheme that can adaptively adjust base station resource allocation based on load prediction while ensuring service quality.

[0099] The energy-saving optimization control method for communication base stations in the embodiments of this application has been described above. The energy-saving optimization control system for communication base stations in the embodiments of this application is described below. Please refer to [link / reference]. Figure 2 One embodiment of the energy-saving optimization control system for communication base stations in this application includes:

[0100] The acquisition module 201 is used to collect and preprocess traffic data and energy consumption data of communication base stations to obtain standardized multidimensional feature data;

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

[0102] The pruning module 203 is used to perform hierarchical quantization pruning on the dual-stream spatiotemporal convolutional neural network. Different bit widths are allocated according to the degree of influence of each network layer on the accuracy of the fused load prediction data, so as to obtain a compressed network structure suitable for base station deployment.

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

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

[0105] Figure 3 This is a schematic diagram of the structure of an energy-saving optimization control device 300 for a communication base station according to an embodiment of the present invention. The energy-saving optimization control device 300 for a communication base station can vary significantly due to different configurations or performance characteristics. It may include one or more central processing units (CPUs) 310 (e.g., one or more processors) and a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 can be temporary or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the diagram), each module including a series of instruction operations on the energy-saving optimization control device 300 for the communication base station. Furthermore, the processor 310 may be configured to communicate with the storage media 330 and execute the series of instruction operations in the storage media 330 on the energy-saving optimization control device 300 for the communication base station to implement the steps of the energy-saving optimization control method for the communication base station described above.

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

[0107] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform 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 sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0109] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an energy-saving optimization control device (which may be a personal computer, server, or network device, etc.) used for a communication base station to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0110] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. An energy-saving optimization control method for communication base stations, characterized in that, The energy-saving optimization control method for communication base stations includes: Traffic and energy consumption data from communication base stations are collected and preprocessed to obtain standardized multidimensional feature data; The standardized multidimensional feature data is input into a dual-stream spatiotemporal convolutional neural network, and parallel computation is performed through temporal feature processing branches and spatial feature processing branches to obtain fused load prediction data. The dual-stream spatiotemporal convolutional neural network is subjected to hierarchical quantization 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, resulting in a compressed network structure suitable for base station deployment. Based on the compressed network structure, real-time load data is processed to generate a dynamic energy-saving control strategy for base stations, including RF resource configuration instructions and power adjustment parameters. This includes: inputting real-time load data into the compressed network structure; generating a load change curve for future time periods through forward computation to obtain load prediction data; dividing the load prediction data into states, classifying time periods into high-load, medium-load, and low-load intervals according to load rate thresholds to obtain a load status table; determining whether the duration of the low-load interval in the load status table exceeds 30 minutes; triggering a deep energy-saving mode when it exceeds 30 minutes, and triggering a basic energy-saving mode when it does not exceed 30 minutes, thus obtaining an energy-saving mode identifier; calculating the RF resource demand for the time period based on the energy-saving mode identifier and the load status table, generating carrier switching status, antenna configuration, and channel allocation instructions to obtain a resource scheduling scheme; calculating the transmit power curve for the time period based on the resource scheduling scheme and the load prediction data change trend, generating power levels, adjustment time points, and step size parameters to obtain a power control scheme; and organizing the resource scheduling scheme and power control scheme in chronological order, adding execution conditions and a backoff mechanism to form the dynamic energy-saving control strategy for base stations.

2. The energy-saving optimization control method for communication base stations according to claim 1, characterized in that, The process of collecting and preprocessing traffic and energy consumption data from communication base stations yields standardized multidimensional feature data, including: Record the radio frequency power value, number of channels used, number of access users, data transmission volume, power consumption value and temperature value of the communication base station at a frequency of one minute to obtain the original data table; The values ​​in the original data table that exceed the upper and lower thresholds are replaced, and the missing values ​​are filled with the average value of the preceding and following time points to obtain a complete data sequence. Subtracting the corresponding feature mean from each value in the complete data sequence and then dividing by the standard deviation converts it into a zero-mean unit variance distribution, thus obtaining uniform-scale data. Calculate the linear relationship metric for each pair of the uniform scale data, construct the feature relationship strength matrix, and obtain the parameter association table; Based on the parameter association table, the importance score is analyzed, and feature columns with importance scores below a set threshold are removed to obtain the key feature set. The key feature set is organized in chronological order, and time period identifiers and load status labels are added to obtain the standardized multidimensional feature data.

3. The energy-saving optimization control method for communication base stations according to claim 1, characterized in that, The process involves inputting the standardized multidimensional feature data into a dual-stream spatiotemporal convolutional neural network, performing parallel computation through temporal feature processing branches and spatial feature processing branches to obtain fused load prediction data, including: The standardized multidimensional feature data is split into two parts: time series and feature relationship, and then input into the parallel architecture asymmetric dual-stream processing structure to obtain the initial dual-stream input. The time series portion of the initial dual-stream input is processed by a variable-length receptive field one-dimensional convolution chain, and the dilated convolution mechanism is used to capture the load fluctuation patterns at different time spans, thus obtaining a multi-granularity time series representation. An autocorrelation graph structure is applied to the feature relationship part of the initial dual-stream input, and the implicit dependencies between parameters are extracted through the feature mapping of cross channels to obtain the parameter relationship matrix; The multi-granularity time series characterization 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; The parameter relationship matrix is ​​subjected to saliency analysis through a feature recombination network to dynamically extract high-influence feature combinations and interaction patterns, thereby obtaining nonlinear spatial features. The adaptive temporal features and the nonlinear spatial features are fused by bidirectional attention, and a spatiotemporal collaborative representation is constructed through a complementary information enhancement mechanism to obtain the fused load prediction data.

4. The energy-saving optimization control method for communication base stations according to claim 3, characterized in that, The time-series portion of the initial dual-stream input is processed through a variable-length receptive field one-dimensional convolution chain, using a dilated convolution mechanism to capture the load fluctuation patterns across different time spans, resulting in a multi-granularity temporal representation, including: A multi-level dilated convolutional structure is constructed for the time series portion, and the kernel interval is set according to the increasing dilation rate to cover the load patterns from short time to long time, so as to obtain the original dilated feature group. Applying a residual connection mechanism to the original expanded feature set, the input features are added and fused with the convolutional output features, preserving the original temporal information, to obtain the enhanced expanded features; The enhanced dilated features are processed by grouped convolution to extract features of different frequencies in groups, resulting in a frequency-separated feature set. Multi-head temporal coding is performed on the frequency separation feature set to extract periodic patterns, trend patterns, and burst patterns, respectively, to obtain three types of temporal semantic features; Based on the three types of temporal semantic features, a feature importance map is constructed, and the weight distribution of different time windows and different feature types is calculated to obtain a weighted feature map. The weighted feature maps are fused across windows, and a self-attention mechanism is used to associate feature representations at different time scales to obtain the multi-granular temporal representation.

5. The energy-saving optimization control method for communication base stations according to claim 4, characterized in that, The construction of a multi-level dilated convolutional structure on the time series portion, setting the kernel interval according to an increasing dilation rate, covering load patterns from short to long time, yields the original dilated feature group, including: The time series portion is divided into three types of sequence segments according to time granularity: minute-level, hour-level, and day-level, to obtain a multi-scale input sequence; A first-level dilated convolutional layer is constructed on the multi-scale input sequence, with the dilation rate set to one, so that the convolutional kernel slides between adjacent time points to capture the load relationship between adjacent time points and obtain short-term dilated features. The short-term dilation feature is input into the second-level dilated convolutional layer with a dilation rate of four, so that the convolutional kernel slides between four time points to capture load changes across time periods and obtain the medium-term dilation feature. The intermediate-term dilation feature is input into the third-level dilated convolutional layer with a dilation rate of sixteen, so that the convolutional kernel slides between sixteen time points to capture long-term load trends and obtain long-term dilation features. The short-term expansion feature, the medium-term expansion feature, and the long-term expansion feature are merged by channel splicing to retain information at each time scale and obtain multi-view time series features; The feature importance score is calculated for the multi-view time series features, and weight coefficients are assigned according to the correlation between each time scale feature and historical load changes to obtain the original expanded feature group.

6. The energy-saving optimization control method for communication base stations according to claim 1, characterized in that, The hierarchical quantization pruning of the dual-stream spatiotemporal convolutional neural network, by allocating different bit widths according to the degree of influence of each network layer on the accuracy of the fused load prediction data, yields a compressed network structure suitable for base station deployment, including: The sensitivity of the dual-stream spatiotemporal convolutional neural network was tested 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 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, thus obtaining a communication load characteristic hierarchical label. The high-sensitivity layer in the communication load characteristic hierarchy retains all radio frequency burst traffic identification functions, and eight-bit fixed-point quantization is used to ensure the accuracy of burst traffic early warning, thus obtaining the burst traffic sensing structure. The daily service mode identification parameters of the medium-sensitive layer in the communication load characteristic hierarchy are optimized, and six-bit fixed-point quantization is used to ensure the efficiency of resource allocation during normal periods, resulting in a normal traffic processing structure. The parameters for extracting long-term trends of the low-sensitivity layer in the communication load characteristic hierarchy are simplified, and four-bit fixed-point quantization is used to meet the basic load prediction requirements to obtain the basic load evaluation structure. The burst traffic sensing structure, the regular traffic processing structure, and the basic load assessment structure are integrated into a lightweight prediction engine for the base station controller. 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. 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 as described in any one of claims 1-6, characterized in that, The energy-saving optimization control system for communication base stations includes: The data acquisition module is used to collect and preprocess traffic data and energy consumption data of communication base stations to obtain standardized multidimensional feature data. The input module is used to input the standardized multidimensional feature data into a dual-stream spatiotemporal convolutional neural network, and perform parallel computation through temporal feature processing branches and spatial feature processing branches to obtain fused load prediction data; The trimming module is used to perform hierarchical quantization trimming on the dual-stream spatiotemporal convolutional neural network, and allocate different bit widths according to the degree of influence of each network layer on the accuracy of the fused load prediction data, so as to obtain a compressed network structure suitable for base station deployment. The generation module is used to process real-time load data based on the compressed network structure and generate a dynamic energy-saving control strategy for base stations that includes radio frequency resource configuration instructions and power adjustment parameters.

8. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the energy-saving optimization control method for a communication base station as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when run by a processor, causes the processor to execute the energy-saving optimization control method for a communication base station as described in any one of claims 1 to 6.

Images