Base station energy saving method and device based on deep learning, equipment and medium

Abstract

The invention provides a base station energy saving method and device based on deep learning, equipment and a medium, and effectively solves the defects of poor prediction effect, insufficient global optimization capability, insufficient fine management and the like in the base station energy saving technology at the present stage. The method comprises the following steps: acquiring target historical business data; controlling a service prediction model to extract time features and spatial features of the target historical service data, and fusing the time features and the spatial features to predict service data of the plurality of base stations in a future time period; an optimal power configuration model is obtained through a pre-improved particle swarm optimization algorithm in combination with a power optimization constraint condition, and service data of a plurality of base stations in a plurality of prediction time periods are processed through the optimal power configuration model to obtain optimal power configuration schemes of the corresponding base stations in the plurality of prediction time periods. And adjusting the transmitting power of the corresponding base station based on the optimal power configuration scheme, and monitoring the adjusted operation state of the base station.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of base station optimization, in particular to a base station energy saving method and device based on deep learning, equipment and medium. BACKGROUND

[0002] With the rapid development and large-scale deployment of 5G technology, the number of base stations has increased explosively, and the energy consumption problem of communication networks has become increasingly prominent. The energy consumption of 5G base stations is 2-3 times that of 4G base stations. In the context of global energy shortage and the "double carbon" target, reducing the energy consumption of base stations has become a key problem that needs to be solved in the communication industry. Traditional base station power control strategies are mostly based on static rules or simple threshold judgments, which cannot adapt to the dynamic changes of traffic, resulting in high power operation during low load periods and causing energy waste. At present, base station energy saving technologies are divided into traditional base station energy saving technologies and intelligent base station energy saving technologies.

[0003] Among them, the traditional base station energy saving technology has the following shortcomings: 1. Impact on network service quality: Traditional energy management strategies lack sufficient feedback mechanisms and are difficult to match actual traffic demand, which may lead to a decline in service quality. For example, some base station dynamic switching strategies are mainly implemented when user traffic decreases or increases significantly. This reactive strategy cannot respond to instantaneous changes in traffic and may affect user communication experience.

[0004] 2. Lack of fine-grained management: Current energy management strategies are usually performed at a high system level and cannot be fine-tuned for specific network conditions, limiting the maximization of energy saving effects. For example, traditional base station energy saving methods rely on simple shutdown strategies, such as shutting down services in some cells at night or during specific time periods. This extensive energy saving method cannot adapt to diverse and dynamic scenario requirements.

[0005] 3. Dependence on manual operation and fixed rules: Some existing base station energy saving technologies do not respond to traffic fluctuations in a timely manner, and the activation mechanism is lagging, and historical and real-time data are not fully utilized for optimized scheduling. For example, some energy saving technologies rely on fixed rules and cannot be adjusted in real time according to actual conditions, requiring a large amount of manpower to analyze data and develop strategies. As the number of sites and data volume increases, the cost and efficiency of manual analysis cannot meet the demand.

[0006] 4. Problems in monitoring and management: Base stations are distributed and lack effective means for power management, such as not being able to learn about power supply failures in a timely manner, distorted use of electricity information analysis and statistics, and lack of systematic statistical data on base station energy consumption. This lack of effective basis for selecting energy-saving base station power schemes, monitoring energy consumption anomalies, and testing energy saving effects.

[0007] The intelligent-based base station energy-saving technology has the following shortcomings: 1. Poor global optimization capability: The base station power optimization problem is essentially a nonlinear optimization problem subject to multiple constraints. Traditional energy management strategies and some simple intelligent algorithms often rely on fixed rules or lag in responding to traffic changes, cannot adjust energy-saving strategies in a timely manner based on real-time traffic data, cannot adapt to rapidly changing network environments, are inefficient in handling large-scale, complex constraint problems, and may result in decreased service quality or poor energy-saving effectiveness.

[0008] 2. Poor traffic prediction effectiveness: Traffic prediction is the premise and basis for base station energy saving. By accurately predicting traffic distribution in a future period of time, the power of the base station can be adjusted in advance to achieve on-demand allocation, ensuring service quality while reducing energy consumption. However, traditional traffic prediction methods such as ARIMA, SVR, LGBM, etc. mainly focus on the prediction of individual base stations, ignoring the spatial correlation between base stations and failing to accurately capture complex spatio-temporal dependencies, resulting in large traffic prediction errors and affecting the effective implementation of base station energy-saving strategies. SUMMARY

[0009] Therefore, the purpose of the present application is to provide a deep learning-based base station energy-saving method, device, equipment and medium, which effectively solves the defects of poor prediction effectiveness, insufficient global optimization capability, and insufficient fine management of current base station energy-saving technology.

[0010] In a first aspect, the embodiments of the present application provide a deep learning-based base station energy-saving method, which comprises: Collecting historical business data of multiple base stations, inputting the historical business data into a pre-trained business prediction model to preprocess the historical business data and obtain target historical business data; the business prediction model is constructed based on a deep learning network; the business prediction model is constructed based on a deep learning network; Controlling the business prediction model to extract the time characteristics and spatial characteristics of the target historical business data, and fusing the time characteristics and spatial characteristics to predict the business data of the multiple base stations in multiple prediction time periods; Obtaining an optimal power allocation model by combining a pre-improved particle swarm optimization algorithm with power optimization constraint conditions, processing the business data of the multiple base stations in the multiple prediction time periods by the optimal power allocation model, and obtaining an optimal power allocation scheme for the corresponding base stations in the multiple prediction time periods; Adjusting the transmission power of the corresponding base station based on the optimal power allocation scheme, and monitoring the adjusted operating state of the base station.

[0011] With reference to the first aspect, in a first possible implementation of the first aspect, the processing, by the optimal power configuration model, the service data of the plurality of base stations in the plurality of predicted time periods to obtain the optimal power configuration scheme of the corresponding base station in the plurality of predicted time periods comprises: mapping the power configuration scheme in the plurality of predicted time periods to the corresponding particle population and determining each particle in the particle population initialized according to a hierarchical strategy; controlling each initialized particle to perform iteration based on a pre-set target iteration mechanism to iteratively obtain the optimal power configuration scheme.

[0012] With reference to the first aspect, in a second possible implementation of the first aspect, the controlling each initialized particle to perform iteration based on the pre-set target iteration mechanism comprises: calculating an adaptive score of each particle in the current iteration, and updating an individual optimal particle and a global optimal particle according to the adaptive score; updating a speed and a position of each particle based on the updated individual optimal particle and the global optimal particle and an inertia weight until an iteration termination condition is met.

[0013] With reference to the first aspect, in a third possible implementation of the first aspect, the processing, by the optimal power configuration model, the service data of the plurality of base stations in the plurality of predicted time periods to obtain the optimal power configuration scheme of the corresponding base station in the plurality of predicted time periods comprises: setting a power optimization target and a plurality of power optimization constraint conditions for the base station based on a plurality of characteristics of the base station; processing the service data by the improved power configuration model to fuse the power optimization target and the plurality of power optimization constraint conditions, and generating the optimal power configuration scheme corresponding to the plurality of predicted time periods respectively.

[0014] With reference to the first aspect, in a fourth possible implementation of the first aspect, the controlling the service prediction model to extract the time feature and the space feature of the target historical service data comprises: filtering historical time sequence information of the target historical service data, and capturing key information of each time step of the historical time sequence information; the key information corresponds to fluctuation of the target historical service data; focusing on long-time dependence between the key information of different time steps to output the time feature corresponding to the different time steps.

[0015] In conjunction with the first aspect, this application provides a fifth possible implementation of the first aspect, wherein the step of controlling the business prediction model to extract the temporal and spatial features of the target historical business data further includes: The dynamic association features of multiple base stations are captured by a dynamic relationship matrix, and the fixed spatial features of the target historical service data are extracted by an adjacency matrix. By integrating the dynamic correlation features and fixed spatial features, spatial features corresponding to multiple base stations are obtained.

[0016] In conjunction with the first aspect, this application provides a sixth possible implementation of the first aspect, wherein predicting service data of multiple base stations in a future time period based on the time and spatial features includes: Collect external environmental factors corresponding to multiple base stations, and extract external environmental features from the external environmental factors; The business features are obtained by integrating the time features, spatial features, and external environment features, so that the business prediction model can process the business features to obtain the business data.

[0017] Secondly, embodiments of this application provide a base station energy-saving device based on deep learning, the device comprising: The input module is used to collect historical service data from multiple base stations and input the historical service data into a pre-trained service prediction model to preprocess the historical service data to obtain target historical service data; the service prediction model is built based on a deep learning network. The control module is used to control the service prediction model to extract the temporal and spatial features of the target historical service data, and to fuse the temporal and spatial features to predict the service data of multiple base stations in multiple prediction time periods. The processing module is used to obtain the optimal power configuration model by combining the pre-improved particle swarm optimization algorithm with power optimization constraints, and to process the service data of multiple base stations in multiple prediction time periods through the optimal power configuration model to obtain the optimal power configuration scheme of the corresponding base station in multiple prediction time periods. The monitoring module is used to adjust the transmit power of the corresponding base station based on the optimal power configuration scheme and monitor the adjusted operating status of the base station.

[0018] Thirdly, embodiments of this application provide an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of any one of the deep learning-based base station energy-saving methods described above.

[0019] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of any one of the deep learning-based base station energy-saving methods described above.

[0020] This application provides a deep learning-based base station energy-saving method. The method first collects historical service data from multiple base stations and inputs this data into a pre-trained service prediction model to preprocess the historical service data and obtain target historical service data. The service prediction model is constructed based on a deep learning network. Next, the service prediction model extracts the temporal and spatial features of the target historical service data and fuses these features to predict the service data of multiple base stations in multiple prediction time periods. Then, an optimal power configuration model is obtained by combining a pre-improved particle swarm optimization algorithm with power optimization constraints. This optimal power configuration model is used to process the service data of multiple base stations in multiple prediction time periods to obtain the optimal power configuration scheme for the corresponding base station in the multiple prediction time periods. Finally, the transmit power of the corresponding base station is adjusted based on the optimal power configuration scheme, and the adjusted operating status of the base station is monitored. Based on the above methods, an optimal power configuration scheme suitable for base stations is provided, which solves the defects of current base station energy-saving technologies such as poor prediction effect, insufficient global optimization capability, and insufficient fine management. It ensures the network service quality for base stations, and also provides fine management for base stations based on multiple power constraints, realizing the effectiveness and accuracy of the obtained power configuration scheme. Furthermore, it ensures effective monitoring of base stations by monitoring the adjusted operating status of the base stations. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart illustrating a deep learning-based base station energy-saving method provided in an embodiment of this application is shown. Figure 2 This illustrates the overall structural framework of the business forecasting model provided in the embodiments of this application; Figure 3 A schematic diagram of the process for obtaining the optimal power configuration scheme provided in an embodiment of this application is shown; Figure 4A structural block diagram of a deep learning-based base station energy-saving device provided in an embodiment of this application is shown. Figure 5 A structural block diagram of an electronic device provided in an embodiment of this application is shown. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in this application are for illustrative and descriptive purposes only and are not intended to limit the scope of protection of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0024] Furthermore, the described embodiments are merely some, not all, of the embodiments of this application. The components of the embodiments of this application described and illustrated herein can typically be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0025] It should be noted that the term "comprising" will be used in the embodiments of this application to indicate the presence of the features declared thereafter, but does not exclude the addition of other features.

[0026] At present, base station energy-saving technologies are divided into two types: traditional base station energy-saving technologies and intelligent base station energy-saving technologies. Among them, traditional base station energy-saving technologies have the following shortcomings: they affect network service quality, lack refined management, rely on manual operation and maintenance and fixed rules, and have problems with monitoring and management. Intelligent base station energy-saving technologies have the following shortcomings: insufficient global optimization capabilities and poor traffic prediction effects.

[0027] Based on this, this application provides a base station energy-saving method, apparatus, device, and medium based on deep learning, which will be described below through embodiments.

[0028] Example 1 To facilitate understanding of this embodiment, a deep learning-based base station energy-saving method disclosed in this application will first be described in detail. For example...Figure 1 The flowchart shown represents a base station energy-saving method based on deep learning. This application provides a base station energy-saving method based on deep learning, the method comprising: S101. Collect historical service data from multiple base stations, and input the historical service data into a pre-trained service prediction model to preprocess the historical service data to obtain target historical service data; the service prediction model is constructed based on a deep learning network; the service prediction model is constructed based on a deep learning network. S102. Control the service prediction model to extract the time and spatial features of the target historical service data, and fuse the time and spatial features to predict the service data of multiple base stations in multiple prediction time periods. S103. The optimal power configuration model is obtained by combining the pre-improved particle swarm optimization algorithm with power optimization constraints. The optimal power configuration model is used to process the service data of multiple base stations in multiple prediction time periods to obtain the optimal power configuration scheme of the corresponding base station in multiple prediction time periods. S104. Adjust the transmit power of the corresponding base station based on the optimal power configuration scheme, and monitor the adjusted operating status of the base station.

[0029] In step S101, this application needs to collect historical service data from multiple base stations in advance. The historical service data includes various types, such as: service data, such as the average number of connected users for RRC connections, user uplink traffic, user downlink traffic, etc.; time data, such as days of the week, whether it is a holiday, whether it is a working day, etc.; environmental data, such as regional activities, such as concerts, exhibitions, etc.; network performance index data, such as the success rate of handover within the system, the success rate of NR to LTE handover, the success rate of handover at the same frequency, and the success rate of RRC connection establishment; home scenario data, such as scenario type and scenario level; actual transmit power data of base station carriers, configured output power data, and latitude and longitude coordinate data. The collected historical service data is then input into a pre-trained service prediction model. This model, built on a deep learning network, is used to predict service data for future time periods based on the historical service data of base stations. The deep learning network includes a spatiotemporal graph convolutional network (ST-GCN) and a long short-term memory network (LSTM). ST-GCN is a convolutional neural network that studies graph data. The essential purpose of GCN is to use graph convolution to extract the spatial features of non-Euclidean graph data. LSTM is a variant of recurrent neural networks. LSTM is a neural network with memory capabilities, which can effectively utilize historical information of time series to process sequences. Predicting service data from multiple base stations is the prerequisite and foundation for energy saving at base stations. By accurately predicting service data for a future period, base station power can be adjusted in advance to achieve on-demand allocation, reducing energy consumption while ensuring service quality.

[0030] Because the collected historical service data may contain non-standard, invalid, or erroneous data, affecting the analysis of the temporal and spatial characteristics of historical service data from multiple base stations, preprocessing of the historical service data from multiple base stations is essential. Therefore, preprocessing of the historical service data, including data cleaning and normalization, yields the target historical service data. Data cleaning is the process of detecting and correcting (or deleting) damaged or inaccurately recorded historical service data from multiple base stations. It refers to identifying incomplete, incorrect, inaccurate, or irrelevant parts of the historical service data and then replacing, modifying, or deleting them. This includes missing value cleaning and outlier cleaning. Missing value cleaning addresses the presence of null values ​​in the historical service data by directly deleting missing data or using imputation methods. Outlier historical service data differs from the format or value range of the corresponding column of outlier historical service data; in this case, direct deletion of outlier data or substitution with historical average values ​​can be performed. Normalization refers to the process where historical service data... After centering by the minimum value and then scaling by the range, the final data is shifted by the minimum value units and will converge to the range [0,1]. The normalization formula is shown in formula (1): (1); in, This represents the maximum value of the historical business data sample. The normalized historical business data, which is the minimum sample value of historical business data, can obtain the optimal solution of gradient descent more quickly and greatly improve the accuracy of the business prediction model. It is an essential step in the process of building a business prediction model.

[0031] In step S102, the overall structural framework of the business prediction model is as follows: Figure 2 As shown, in the service prediction model, the LSTM network and ST-GCN network receive the target historical service data to extract the temporal and spatial features of the target historical service data based on the service prediction model. The LSTM network consists of several LSTM modules, each containing a cell state, input gate, forget gate, and output gate, used to process the temporal features of the target historical service data of multiple base stations to obtain high-dimensional features at different time steps. The ST-GCN network describes the relationship between base station locations through an adjacency matrix, extracts the spatial features of the target historical service data, analyzes the non-Euclidean structure of the data, and establishes a relationship matrix based on traffic data with different time distributions. It uses a graph convolutional network to obtain the spatial distribution features of different prediction time periods under the current state and fuses the temporal and spatial features to predict the service data of multiple base stations in multiple prediction time periods. The prediction time period is the future time period of the base station that needs to be predicted.

[0032] The training of the spatiotemporal ST-GCN+LSTM service prediction model requires sufficient historical service data from multiple base stations, which are then divided into training, validation, and test sets. Supervised learning is used for model training, updating model parameters by minimizing the loss function between predicted and actual values. Commonly used loss functions include mean squared error (MSE), mean absolute error (MAE), and mean absolute percentage error (MAPE), expressed as formulas (2), (3), and (4), respectively. (2); (3); (4); in, It is the actual value. Here, N is the predicted value, and N is the number of samples. Business prediction models are typically trained using stochastic gradient descent (SGD) and its variants, such as Adagrad, Adadelta, and Adam. The choice of learning rate is crucial for model training; strategies such as fixed learning rate, learning rate decay, or adaptive learning rate can be employed. To prevent overfitting in business prediction models, the following techniques can be used: 1) Regularization: Add L1 or L2 regularization terms to the loss function to constrain the size of the parameters of the business prediction model; 2) Dropout: Randomly discards some neurons during the training of the business prediction model to reduce co-adaptation between neurons; 3) Early stopping: Stop training when the loss on the validation set no longer improves to prevent overfitting; 4) Data Augmentation: Training data is augmented through methods such as time shifting and noise injection to improve the generalization ability of the business prediction model. During the training process, changes in training and validation losses need to be monitored, and the parameters and hyperparameters of the business prediction model need to be adjusted to obtain optimal predictive performance. After the business prediction model is trained, it needs to be evaluated on a test set, and various performance metrics, such as MSE, MAE, and MAPE, need to be calculated to verify the effectiveness of the business prediction model.

[0033] In a specific implementation of step S102, one embodiment is as follows: the control of the business prediction model to extract the temporal and spatial features of the target historical business data includes: A1. Filter out the historical time-series information of the target historical business data, and capture the key information of the historical time-series information at each time step; the key information corresponds to the fluctuation of the target historical business data. A2. Focus on the long-term dependencies between key information at different time steps to output the temporal features corresponding to different time steps.

[0034] In steps A1-A2, firstly, the forget gate of the LSTM network determines which information to discard or retain in the target historical service data, filtering out irrelevant abnormal time period data to continue propagation, thereby selecting the historical time series information of the target historical service data. Let the current time be... , This represents the cell's state at the previous moment. Enter the value for the current moment. , It is a weight matrix; For bias terms, If it is a Sigmoid function, then the information continues to propagate downwards. For formula (5): (5); Secondly, the input gate determines what is input to. New information from neurons at any given time is also generated by... The decision to discard or retain information is made by using a Sigmoid function to determine the importance of data at the current time step, such as the number of users or uplink / downlink traffic in a given hour, outputting a weight between 0 and 1. A weight close to 1 indicates critical information, while a weight close to 0 indicates negligible information. Then, a tanh function is used to generate candidate features for the current time period, such as the magnitude and duration of traffic spikes during peak hours. After being selected according to weights, these features are injected into the LSTM cell state. The weight matrix is ​​as follows: , , , The bias term is , This process can be described by formulas (6) and (7): (6); (7); Then, by combining the previous steps to update the cell state, the past and present states are integrated to achieve accurate updates of the current effective information and avoid interference from irrelevant noise. This can be expressed as formula (8): (8); Finally, the output gate focuses on the long-term dependencies of base station service data, such as the impact of previous low-peak periods on the current off-peak period and the correlation of consecutive peak periods. The output gate determines the information state flowing to the next neuron, which can be expressed as formula (9) and formula (10): (9); (10); First, the sigmoid function is used to filter the information in the cell state, retaining only the parts related to time series, such as the correlation strength between the current time period's flow and the previous 1-3 time periods. Then, the tanh function is used to map the filtered cell state to the [-1,1] interval, and the high-dimensional key features of the current time step are output as the temporal features. This lays the foundation for subsequent fully connected network integration and fusion with spatial features. , This is the weight matrix. This is a bias term.

[0035] In a specific implementation of step S102, another embodiment is as follows: the step of controlling the business prediction model to extract the temporal and spatial features of the target historical business data further includes: B1. Capture the dynamic association features of multiple base stations using a dynamic relationship matrix and extract the fixed spatial features of the target historical service data using an adjacency matrix. B2. By fusing the dynamic correlation features and fixed spatial features, spatial features corresponding to multiple base stations are obtained.

[0036] In steps B1-B2, the essential purpose of the ST-GCN network is to utilize graph convolution to extract spatial features from non-Euclidean graph data. Input signal and output signal The processing method adopted by the ST-GCN network Defined as formula (11): (11); in, This indicates the number of nodes in the graph. ,and Describe the set of edges. It is the adjacency matrix of the graph. Adjacency matrix elements in Representation diagram Middle node and The connection relationships between them, i.e., the adjacency matrix A The fixed spatial features extracted from the target historical business data are given by the forward propagation formula for graph convolution as shown in formula (12): Formula (12); In the formula, That is, the spatial characteristics corresponding to multiple base stations respectively. , It is the size of The identity matrix; It is a diagonal matrix. That is, the dynamic relationship matrix captures the dynamic correlation characteristics of multiple base stations; Indicates the first The output value of the layer, where , This represents the activation function, enhancing the ST-GCN network's ability to fit complex spatial relationships, such as the nonlinear collaborative relationships of peak traffic in base station clusters. Indicates the first Layer parameter values.

[0037] In the specific implementation of step S102, there is another embodiment where: predicting the service data of multiple base stations in the future time period based on the time and spatial features includes: C1. Collect external environmental factors corresponding to multiple base stations, and extract the external environmental features from the external environmental factors; C2. The time features, spatial features, and external environment features are integrated to obtain business features, so that the business prediction model processes the business features to obtain the business data.

[0038] In steps C1-C2, after obtaining the temporal and spatial features, these features are input into the fully connected layer of the service prediction model. The fully connected layer also collects external environmental factors corresponding to multiple base stations from the target historical service data and extracts external environmental features from these factors. These external environmental features can also be normalized and fused with the temporal, spatial, and external environmental features through a combination of concatenation and attention weighting to obtain service features. This introduces a feature attention mechanism, assigning dynamic weights to the three types of features. For example, during regional events, the weight of external environmental features is increased to 0.3; during peak periods, the weight of temporal features is increased to 0.45. This allows the service prediction model to process the service features and obtain the service data, addressing the shortcomings of traditional models that neglect cross-base station correlation and environmental interference. The result is service data that not only conforms to the temporal patterns of individual base stations but also matches the collaborative characteristics of regional base stations, while adapting to the sudden impact of external factors such as concerts and rainfall. The prediction accuracy is significantly better than that of single-feature models.

[0039] In step S103, this application pre-obtains an optimal power configuration model for base station power configuration by combining a pre-improved particle swarm optimization algorithm with power optimization constraints. The power optimization constraints are varied to ensure the accuracy of the optimal power configuration model in obtaining the optimal power configuration scheme for each base station in multiple prediction time periods based on service data from multiple base stations. The optimal power configuration model has the objective of minimizing base station energy consumption and various power optimization constraints, thereby transforming the base station power optimization problem into a constrained nonlinear optimization problem suitable for solving using intelligent optimization algorithms such as particle swarm optimization. Therefore, a power configuration model based on particle swarm optimization (PSO) improvement is established based on the power optimization model, and the service data from the multiple prediction time periods is input into the improved power configuration model. The optimal power configuration model makes predictions based on the service data to obtain the optimal power configuration schemes for multiple base stations in multiple prediction time periods; multiple base stations have corresponding optimal power configuration schemes; that is, each base station has a corresponding optimal power configuration scheme.

[0040] After obtaining the optimal power configuration scheme for the base station, considering that constant power base stations may operate at a fixed value for a long time, and that direct dynamic adjustment throughout the entire time period carries a high risk, a phased output strategy is adopted for the optimal power configuration scheme: 1) Pilot phase (1-2 weeks): The optimized dynamic power (reduced power) is only implemented during low service demand periods (such as 1:00-6:00), while it remains unchanged during high service demand periods to verify the energy-saving effect and service quality; 2) Promotion phase (3-4 weeks): If the data rate compliance rate is ≥98% and the energy consumption reduction rate is ≥10% during the pilot phase, it will be extended to off-peak periods; 3) Full phase (after 1 month): Dynamic power is implemented throughout the entire time period, and the final "time period-power" (one power value for each time period) is output.

[0041] In a specific implementation of step S103, one embodiment is as follows: processing service data of multiple base stations in multiple prediction time periods through the optimal power configuration model to obtain the optimal power configuration scheme of the corresponding base station in multiple prediction time periods includes: D1. Based on the various characteristics of the base station, set power optimization targets and various power optimization constraints for the base station; D2. By integrating the power optimization objective and multiple power optimization constraints through the improved power configuration model, the business data is processed to generate the optimal power configuration schemes corresponding to multiple prediction time periods.

[0042] In steps D1-D2, this application pre-constructs a base station power optimization model based on the goal of minimizing base station energy consumption while simultaneously meeting user service quality requirements. Therefore, the goal of minimizing base station energy consumption in the power configuration model is expressed by formula (13): (13); Where T is the number of time periods (a single time period can be set to 1 hour). It is the first The power configuration output power is configured for a given time period, and the power configuration model needs to satisfy the following power optimization constraints when performing power optimization: 1) Power dynamic range constraint: The original constant power value is denoted as Set the dynamic adjustment range: ,in To adjust the coefficient, an initial value of 0.2 is used, allowing for fluctuations of ±20%, which will be gradually increased to 0.5 based on service quality. Furthermore, the base station's transmission power cannot be lower than the minimum value or exceed the maximum value. .in, It is the first Minimum transmit power of each base station It is the first The maximum transmit power of each base station.

[0043] 2) Power smoothing constraint: To avoid damage to equipment or interference with signals caused by sudden power changes, the power smoothing constraint is expressed as formula (14): (14); in, for Configure output power at any time. for Configure output power at any time. This is the threshold for the maximum allowable power change.

[0044] 3) Service Requirements - Power Matching Constraint: Based on the service data values ​​predicted by ST-GCN+LSTM, the influence relationship between power and service data is set: when the predicted service data value > 120% of the historical average service data value at the corresponding time, (This indicates a need to increase power); When the predicted business data value is less than 50% of the historical average business data value at the corresponding time point, (This indicates a need to reduce power), ensuring that there is a clear basis for dynamically adjusting the power.

[0045] In the specific implementation of step S103, another embodiment exists as follows: Figure 3 As shown, the step of processing service data from multiple base stations over multiple prediction time periods using the optimal power configuration model to obtain the optimal power configuration scheme for the corresponding base station over multiple prediction time periods includes: E1. Map the corresponding particle population based on the power configuration scheme of multiple prediction time periods, and determine each particle in the particle population to be initialized according to the hierarchical strategy. E2. After initialization, each particle is controlled to iterate based on a pre-set target iteration mechanism to find the optimal power configuration scheme.

[0046] In steps E1-E2, this application uses the power configuration schemes for the multiple prediction time periods as a scheme set, and maps the scheme set to obtain the corresponding particle population. That is, each power configuration scheme in the scheme set is a particle. For constant power base stations, the particles need to search for the optimal solution within the dynamic range that allows adjustment, to avoid exceeding the safety range of the hardware or initial debugging. The encoding method is adjusted as follows: .in, To optimize the number of time periods (e.g., dividing a day into 24 periods of 1 hour each), =24); each (No. The first particle The power of the cycle must satisfy: and This ensures that particles search within a safe range. Particle initialization needs to guarantee diversity; for constant power scenarios, particles need to be deliberately guided to deviate from their designated paths. To avoid the initial values ​​still clustering around a constant value: 1) Layered initialization strategy: ① High-demand time periods (e.g., 8:00-11:00, 18:00-21:00): 30% of particles are initialized in [ ×1.1, ×1.2] (Increase configuration power); ② Low business demand time period (e.g., 1:00-6:00): 30% particle initialization in [ ×0.8, [×0.9] (Reduce configuration power); ③ Off-peak business demand period group (other time periods): 40% of particles are in [ ×0.9, [×1.1] Random initialization.

[0047] 2) Feasibility Screening: After initialization, identify particles that "do not meet business requirement matching constraints" (e.g., particle power is still ≥ 1 during periods of low business demand). The population is regenerated to ensure that more than 80% of the particles in the initial population deviate from a constant value and meet the constraints; and each particle after control initialization iterates based on a pre-set target iteration mechanism, which includes inertia weights. w By balancing global and local searches, combining non-dominated sorting and congestion assessment, Pareto optimal solutions are selected to iterate and find the optimal power configuration scheme.

[0048] In the specific implementation of step E2, one embodiment is as follows: each particle after control initialization iterates based on a pre-set target iteration mechanism, including: E21. Calculate the fitness score of each particle in the current iteration, and update the individual best particle and the global best particle based on the fitness score; E22. Based on the updated individual optimal particle, global optimal particle, and inertia weight, update the velocity and position of each particle until the iteration termination condition is met.

[0049] In steps E21-E22, in each iteration, the fitness score of each particle in the current iteration is calculated first. The fitness score is calculated based on the fitness function. The fitness function is the core criterion for evaluating the quality of particles in the particle swarm optimization algorithm. It directly determines the search direction of the algorithm and the effectiveness of the optimization results. In view of the requirement of "balancing energy consumption, service quality, interference and adapting to dynamic network environment" for base station power optimization, this application designs the fitness function to strengthen "dynamic adjustment compliance" and adds "dynamic power adjustment compliance penalty item" to ensure that the particles not only meet the energy consumption target, but also break the constant value. The fitness function is shown in formula (15): Formula (15) in: : Represents the basic energy consumption item (total power consumption); Indicates the power smoothing penalty coefficient. : Indicates a violation value of the power smoothing constraint; , This represents the penalty coefficient for violations within a dynamic range. : Indicates the lower limit of the violation value in the dynamic range. : Indicates the upper limit of the dynamic range violation value; This represents the penalty coefficient for matching business data constraints. : Indicates a violation value in business data matching constraints.

[0050] Business data matching constraints are the core of ensuring that dynamic power adjustment matches business needs, and the calculation of violation values ​​needs to be defined according to different scenarios: 1) High business demand scenarios (predicting business data) , (This is the average of historical business data) (Power configuration must be ≥1.1×) Otherwise, the violation value will be calculated based on the difference.

[0051] 2) Low business demand scenarios (predicting business data) , (This is the average of historical business data) (Power requirement: ≤0.8×) Otherwise, the violation value will be calculated based on the excess amount.

[0052] final This ensures that the power at each time step matches the business requirements.

[0053] The priority of the penalty coefficient is as follows: Highest priority: Ensure that power adjustments do not affect the user experience; , Secondly: Ensure that the power adjustment is within the safe debugging range. The penalty forces particles to converge within the interval, avoiding the risk of hardware overload due to excessive power during the initial debugging phase. Together, they form a "dual-boundary protection" for the dynamic range; Lowest priority: Balance equipment lifespan with dynamic adjustment needs.

[0054] Particles update themselves by tracking two "extremes": the first is the best solution found by the particle itself, called the individual extreme point, or the individual optimum. The other extreme point is the best solution found so far for the entire population, called the global extreme point, or global optimum. The position is represented by a subset of the population, while local PSO uses a subset of the population as neighbors instead of the entire population. The optimal solution among all neighbors is the local extremum. The position is indicated by the equation (16) and (17) below. After finding these two optimal solutions, the particle updates its velocity and position according to the following equations (16) and (17).

[0055] particle Information can be represented by a D-dimensional vector, and position is represented as... The speed is Then the velocity and position update equations are: (16); (17); From the particle velocity update formula, the first part represents the influence of the particle's current velocity on its flight, providing the propulsion for the particle's flight in the search space. The second part is the so-called "individual cognition," representing the particle's personal experience, prompting it to move towards the best position it has ever experienced. The third part is the so-called "collective cognition," representing the influence of collective experience on the particle's flight trajectory, prompting it to move towards the best position discovered by the collective. Among these: It is a particle In the In the nth iteration The current position of the dimension; It is a particle In the In the nth iteration The current position of the dimension; It is a particle In the The location of the individual extreme point of a dimension; Is the entire population in the first The location of the global extremum point in dimension; and It is a random number between [0, 1]; and This is the acceleration factor (or learning factor), which adjusts the particle's self-experience and social experience. Its role in the particle's motion indicates how it propels each particle towards... and The weight of the statistical acceleration term for position: low values ​​allow particles to linger outside the target area before being pulled back, while high values ​​cause particles to suddenly rush towards or cross the target area. If =0, then the particle has no cognitive ability. Through particle interactions, not only can new search spaces be reached, but it is also easy to get trapped in local extrema; if =0, there is no social information sharing among particles, and the algorithm becomes a multi-starting-point random search: if = =0, the particle will continue to fly at its current speed until it reaches the boundary. Normally , Between [0, 4], it is generally taken as 2.

[0056] To maintain the particle's inertia and give it a tendency to expand its search space and explore new regions, inertial weights can be added to the velocity update equation. As shown in formula (18).

[0057] (18); if =0, because velocity itself has no memory, it only depends on the particle's current position and its historical best position. and Therefore, the particle swarm will shrink to the current global best position, making it more like a local algorithm; if ≠0, particles tend to expand the search space, exhibiting global search capability. Inertia weights are used to control the influence of previous velocities on the current velocity; larger ones... This can enhance PSO's global search capabilities, while smaller... It can enhance local search capabilities; basic PSO can be seen as... =1, therefore, the local search capability is lacking in the later stages of the iteration. That is, the target iteration mechanism of this application is the scoring and update loop mechanism, that is, the above steps are executed in a loop.

[0058] During the operation of the power configuration model based on the improved PSO, the following indicators are continuously monitored to determine whether the optimization termination condition is met: 1) Iteration count judgment: The maximum number of iterations is preset. When the number of iterations of the model reaches the preset maximum value, the optimization process is considered to be terminated.

[0059] 2) Pareto front stability judgment: If, during 20 consecutive iterations, the average crowding rate of the solutions in the external archive used to store the Pareto optimal solution is less than or equal to 0.05, it indicates that the Pareto front has become stable, there is no obvious optimization space, and the termination condition is met.

[0060] 3) Stability judgment of the number of Pareto optimal solutions: When the number of Pareto optimal solutions in the external archive remains unchanged for 10 consecutive iterations, it means that no new Pareto optimal solutions are generated, and the optimization can be stopped.

[0061] 4) If, in 15 consecutive iterations, the "power standard deviation" of the optimal particle is ≥ If the power consumption is ×0.1 and the energy consumption reduction rate is ≥8%, then the iteration is terminated early, indicating that the optimization is effective. At this point, the power has been dynamically adjusted and energy is saved.

[0062] If any of the above conditions are met, the iteration ends and the optimal power configuration scheme is obtained; if not, the optimization parameters are adjusted and the optimization calculation based on the improved PSO power configuration model is performed again.

[0063] In step S104, after receiving the phased output optimal power configuration scheme, the base station adjusts the transmission power of the corresponding base station based on the optimal power configuration scheme. The optimal power configuration of the corresponding base station is determined according to the determined optimal power configuration scheme, and the transmission power of each base station at different times is adjusted to ensure it operates according to the configured value. Simultaneously, a monitoring system is deployed to monitor the adjusted operating status of the base stations and collect key data during base station operation in real time, including but not limited to base station energy saving rate (energy saving rate = (energy consumption at constant power - energy consumption at dynamic power) / (energy consumption at constant power) × 100%), signal-to-interference-plus-noise ratio (SINR) on the user side, and throughput. Through continuous monitoring of this real-time data, it is determined whether the network operation meets the expected performance indicators after the base station power configuration adjustment, providing data support for subsequent feedback and update processes.

[0064] The feedback update process compares and analyzes real-time data collected during base station execution and monitoring with expected targets. If monitoring reveals an increase in user complaints after power reduction during periods of low service demand, the scope will be narrowed. (If the value is reduced from 0.2 to 0.1), retrain the ST-GCN+LSTM model to adjust the power demand prediction; if the energy saving rate is not met, then expand the... Simultaneously, the fitness function penalty coefficient of the PSO model is optimized and improved, such as reducing... , This allows for a wider range of adjustments, enabling the entire base station power optimization system to better adapt to dynamically changing network environments and providing more accurate models and parameters for the next round of power optimization.

[0065] Example 2 This application also provides a base station energy-saving device based on deep learning, such as... Figure 4 The diagram shows a block diagram of a deep learning-based base station energy-saving device. The functions implemented by this device correspond to the steps described above in executing a deep learning-based base station energy-saving method on a terminal device. This device can be understood as a server component including a processor. The deep learning-based base station energy-saving device described in this application includes: The input module 401 is used to collect historical service data from multiple base stations and input the historical service data into a pre-trained service prediction model to preprocess the historical service data to obtain target historical service data; the service prediction model is constructed based on a deep learning network. The control module 402 is used to control the service prediction model to extract the time and spatial features of the target historical service data, and to fuse the time and spatial features to predict the service data of multiple base stations in multiple prediction time periods. The processing module 403 is used to obtain the optimal power configuration model by combining the pre-improved particle swarm optimization algorithm with power optimization constraints, and to process the service data of multiple base stations in multiple prediction time periods through the optimal power configuration model to obtain the optimal power configuration scheme of the corresponding base station in multiple prediction time periods. The monitoring module 404 is used to adjust the transmit power of the corresponding base station based on the optimal power configuration scheme and monitor the adjusted operating status of the base station.

[0066] In one feasible implementation, the processing module includes: The mapping module is used to map the corresponding particle population based on the power configuration scheme of multiple prediction time periods, and determine each particle in the particle population to be initialized according to the hierarchical strategy. The iteration module controls each particle after initialization to iterate according to a pre-set target iteration mechanism in order to find the optimal power configuration scheme.

[0067] In one feasible implementation, the processing module further includes: The calculation module is used to calculate the fitness score of each particle in the current iteration, and update the individual best particle and the global best particle based on the fitness score; The update module is used to update the velocity and position of each particle based on the updated individual optimal particle, the global optimal particle, and the inertia weight, until the iteration termination condition is met.

[0068] In one feasible implementation, the processing module also includes: The setting module is used to set power optimization targets and various power optimization constraints for the base station based on various characteristics of the base station. The generation module is used to integrate the power optimization objective and multiple power optimization constraints through the improved power configuration model, process the business data, and generate the optimal power configuration schemes corresponding to multiple prediction time periods.

[0069] In one feasible implementation, the control module includes: The filtering module is used to filter out the historical time-series information of the target historical business data and capture the key information of the historical time-series information at each time step; the key information corresponds to the fluctuation of the target historical business data. The output module is used to focus on the long-term dependencies between key information at different time steps, so as to output the temporal features corresponding to different time steps.

[0070] In one feasible implementation, the control module further includes: The capture module is used to capture the dynamic association features of multiple base stations through a dynamic relationship matrix and extract the fixed spatial features of the target historical service data through an adjacency matrix. The fusion module is used to fuse the dynamic correlation features and fixed spatial features to obtain spatial features corresponding to multiple base stations.

[0071] In one feasible implementation, the control module also includes: The collection module is used to collect external environmental factors corresponding to multiple base stations and extract external environmental features from the external environmental factors; The module is used to fuse the time features, spatial features, and external environment features to obtain business features, so that the business prediction model can process the business features to obtain the business data.

[0072] Example 3 This application also provides an electronic device, such as Figure 5 As shown, it includes: a processor 501, a memory 502, and a bus 503. The memory 502 stores machine-readable instructions that can be executed by the processor 501. When the electronic device is running, the processor 501 and the memory 502 communicate through the bus 503. When the machine-readable instructions are executed by the processor 501, the steps of any one of the deep learning-based base station energy-saving methods described above are executed.

[0073] Example 4 This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of any one of the deep learning-based base station energy-saving methods described above.

[0074] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and devices described above can be referred to the corresponding processes in the method embodiments, and will not be repeated here. In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple modules or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection can be through some communication interfaces; the indirect coupling or communication connection of devices or modules can be electrical, mechanical, or other forms.

[0075] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0076] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0077] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, a platform server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, ROM, RAM, magnetic disks, or optical disks.

[0078] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A base station energy-saving method based on deep learning, characterized in that, The method includes: Historical service data from multiple base stations is collected and input into a pre-trained service prediction model to preprocess the historical service data to obtain target historical service data; the service prediction model is constructed based on a deep learning network. The service prediction model is controlled to extract the temporal and spatial features of the target historical service data, and the temporal and spatial features are fused to predict the service data of multiple base stations in multiple prediction time periods. The optimal power configuration model is obtained by combining the pre-improved particle swarm optimization algorithm with power optimization constraints. The optimal power configuration model is then used to process the service data of multiple base stations in multiple prediction time periods to obtain the optimal power configuration scheme of the corresponding base stations in multiple prediction time periods. The transmit power of the corresponding base station is adjusted based on the optimal power configuration scheme, and the operating status of the base station after adjustment is monitored.

2. The method according to claim 1, characterized in that, The step of processing service data from multiple base stations over multiple prediction time periods using the optimal power configuration model to obtain the optimal power configuration scheme for the corresponding base stations over multiple prediction time periods includes: Based on the power configuration schemes for multiple prediction time periods, the corresponding particle population is mapped, and each particle in the particle population is determined to be initialized according to the hierarchical strategy. After initialization, each particle is controlled to iterate based on a pre-set target iteration mechanism to find the optimal power configuration scheme.

3. The method according to claim 2, characterized in that, Each particle after control initialization iterates based on a pre-set target iteration mechanism, including: Calculate the fitness score for each particle in the current iteration, and update the individual best particle and the global best particle based on the fitness score; Based on the updated individual optimal particle, global optimal particle, and inertia weights, update the velocity and position of each particle until the iteration termination condition is met.

4. The method according to claim 1, characterized in that, The step of processing service data from multiple base stations over multiple prediction time periods using the optimal power configuration model to obtain the optimal power configuration scheme for the corresponding base stations over multiple prediction time periods includes: Based on the various characteristics of the base station, power optimization targets and various power optimization constraints are set for the base station. By integrating the power optimization objective and multiple power optimization constraints through the improved power configuration model, the business data is processed to generate optimal power configuration schemes for multiple prediction time periods.

5. The method according to claim 1, characterized in that, The control of the business prediction model to extract the temporal and spatial features of the target historical business data includes: The historical time-series information of the target historical business data is filtered out, and key information of the historical time-series information is captured at each time step; the key information corresponds to the fluctuation of the target historical business data. Focusing on the long-term dependencies between key information at different time steps, we can output the temporal features corresponding to different time steps.

6. The method according to claim 1, characterized in that, The process of controlling the business prediction model to extract the temporal and spatial features of the target historical business data also includes: The dynamic association features of multiple base stations are captured by a dynamic relationship matrix, and the fixed spatial features of the target historical service data are extracted by an adjacency matrix. By integrating the dynamic correlation features and fixed spatial features, spatial features corresponding to multiple base stations are obtained.

7. The method according to claim 1, characterized in that, The prediction of service data for multiple base stations in the future time period based on the aforementioned time and spatial features includes: Collect external environmental factors corresponding to multiple base stations, and extract external environmental features from the external environmental factors; The business features are obtained by integrating the time features, spatial features, and external environment features, so that the business prediction model can process the business features to obtain the business data.

8. A deep learning-based base station energy-saving device, characterized in that, The device includes: The input module is used to collect historical service data from multiple base stations and input the historical service data into a pre-trained service prediction model to preprocess the historical service data to obtain target historical service data; the service prediction model is built based on a deep learning network. The control module is used to control the service prediction model to extract the temporal and spatial features of the target historical service data, and to fuse the temporal and spatial features to predict the service data of multiple base stations in multiple prediction time periods. The processing module is used to obtain the optimal power configuration model by combining the pre-improved particle swarm optimization algorithm with power optimization constraints, and to process the service data of multiple base stations in multiple prediction time periods through the optimal power configuration model to obtain the optimal power configuration scheme of the corresponding base station in multiple prediction time periods. The monitoring module is used to adjust the transmit power of the corresponding base station based on the optimal power configuration scheme and monitor the adjusted operating status of the base station.

9. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of a deep learning-based base station power-saving method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of a deep learning-based base station energy-saving method as described in any one of claims 1 to 7.

Images