5g base station planning optimization energy saving whole cycle closed loop intelligent management and control method and device

Abstract

The application discloses a 5G base station planning optimization energy-saving whole-cycle closed-loop intelligent management and control method and device, and belongs to the technical field of mobile communication networks. The application builds a three-layer closed-loop architecture of a planning layer pre-energy-saving, an optimization layer dynamic tuning and an energy-saving layer precise scheduling, constructs a time-space business energy consumption prediction model, establishes a four-dimensional joint target function of coverage, capacity, energy consumption and redundant resources, and outputs an optimal planning scheme through a multi-objective algorithm; generates individualized energy-saving thresholds based on real-time network state parameters, and executes hierarchical energy-saving scheduling and two-stage fallback protection; adopts a layered iteration strategy of bottom layer freezing and top layer incremental updating, and forms whole-life-cycle closed-loop optimization through running data backflow. The application breaks through the whole-link data of planning, optimization and energy-saving, significantly improves the comprehensive energy-saving rate of 5G base stations under the premise of guaranteeing network coverage quality and operation stability, and is suitable for energy consumption fine management and control of multi-scene 5G access networks.

Description

Technical Field

[0001] This invention belongs to the field of mobile communication network technology, specifically relating to a closed-loop intelligent management and control method and device for energy saving throughout the entire life cycle of 5G base station planning optimization, which is applicable to the refined management and control of energy consumption throughout the entire life cycle of 5G access networks in multiple scenarios. Definition of core terms

[0002] To clearly define the scope of the technical parameters of this application, the core terms are uniformly explained as follows: Overlap coverage: refers to the proportion of sampling points in a cell measurement sampling point where the difference between the received reference signal power (RSRP) of 3 or more neighboring cells and the RSRP of the serving cell is within 6dB, and is used to characterize the degree of signal overlap among multiple base stations in the area.

[0003] Redundant resource ratio: refers to the proportion of physical resource blocks (PRBs) not occupied by services to the total number of available resource blocks in a cell during a low-service period when the cell load rate is below 30% for more than 1 consecutive hour. It is used to characterize the idleness of base station resources.

[0004] User distribution density: refers to the normalized ratio of the hourly average number of RRC connected users in a single cell to the average number of users in all cells within the planning area, used to characterize the differences in service carrying pressure among different cells.

[0005] Interference level: refers to the average interference signal strength of the uplink operating frequency band of the cell, measured in dBm, and is used to characterize the noise floor interference level of the network.

[0006] Layered incremental update: This refers to a model update strategy that freezes all parameters of the bottom feature extraction network and only fine-tunes the weights of the top output layer during model iteration. Background Technology

[0007] With the large-scale deployment of 5G networks, the high power consumption of base stations has become a key factor restricting telecommunications operators from reducing costs and increasing efficiency. 5G base stations adopt Massive MIMO and high-frequency transmission architecture, and the power consumption of a single base station is 3 to 4 times that of a traditional 4G base station. The energy consumption of the entire network's base stations accounts for more than 30% of the operator's annual operating costs.

[0008] Existing 5G base station energy-saving technologies have the following main shortcomings: 1. Disconnect between planning and energy conservation: Traditional network planning focuses on coverage and capacity, failing to incorporate energy consumption and redundancy resource constraints during the planning phase. This results in a large amount of redundant resources remaining unused after base stations are built. According to industry statistics, the average redundancy rate of 5G base stations in China reaches 25% to 30%.

[0009] 2. Conservative Energy-Saving Threshold Settings: Existing commercial energy-saving solutions use PRB load as the sole trigger condition, failing to consider network status factors such as overlapping coverage, interference, and user distribution. In engineering implementation, a safety buffer of 20% to 30% is typically reserved, resulting in limited overall energy-saving rates. Actual network measurement data shows that the overall energy-saving rate of traditional solutions is mostly concentrated in the 15% to 20% range.

[0010] 3. Lack of a full lifecycle iteration mechanism: Traditional base station planning is a one-time solution. After the network is put into operation, it cannot be continuously optimized according to changes in regional services. Model updates often adopt the method of full retraining, which has problems such as model instability and large fluctuations in KPIs after updates.

[0011] 4. Insufficient scenario adaptability: Existing solutions mostly use unified parameter configuration, which cannot effectively adapt to the different business tidal characteristics of different scenarios such as residential areas, industrial parks, and universities. Sudden parameter changes during scenario switching can easily lead to network fluctuations. Summary of the Invention

[0012] I. Technical problems to be solved

[0013] The technical problem to be solved by this invention is that in the prior art, data is not shared between network planning, optimization and energy saving, the energy saving threshold is set conservatively resulting in insufficient energy saving depth, there is a lack of a stable optimization mechanism throughout the entire life cycle, and the ability to adapt to different scenarios is poor and the switching is prone to fluctuations. Technical solution

[0014] To address the technical problems in existing technologies, such as data incompatibility between network planning, optimization, and energy saving stages; insufficient energy saving depth due to conservative energy saving threshold settings; network KPI fluctuations easily caused by full model updates; and poor scenario adaptability and susceptibility to fluctuations during switching, this invention adopts the following technical solution: This invention designs a closed-loop operation mechanism of "planning layer pre-energy saving - optimization layer dynamic tuning - energy saving layer precise scheduling - data feedback iteration", and builds a three-layer closed-loop linkage architecture. Through the deep integration of algorithm improvement and communication network management, it reduces redundant deployment from the planning source and achieves a balance between precise energy saving and network quality during the operation phase. Specifically, it includes four processes: S1 Planning Layer Front-End Energy Saving Multi-dimensional basic data of the target area is collected, and an LSTM spatiotemporal service energy consumption prediction model is constructed based on the spatiotemporal tidal characteristics of 5G network services. The model accurately outputs the service load and power consumption prediction values ​​of each cell for future time periods. To address the shortcomings of traditional planning that only considers coverage and capacity, a four-dimensional joint objective function is established, which includes coverage quality, network capacity, total base station energy consumption, and redundant resource ratio. The NSGA-II multi-objective genetic algorithm is used to solve the Pareto optimal planning scheme, thereby reducing redundant base station deployment from the source and lowering the energy consumption base of the network throughout its entire life cycle.

[0015] In this step, the LSTM model designs a special feature vector for the time periodicity and spatial distribution characteristics of 5G services, achieving high-precision prediction of load and power consumption, providing reliable data support for multi-objective planning, and solving the problems of traditional planning relying on experience and lacking energy consumption constraints.

[0016] S2 optimization layer dynamic tuning It collects real-time network operation data and air interface latency data, and performs three types of network optimization in parallel: overlapping coverage optimization, load balancing optimization, and interference coordination optimization. It outputs four core network status parameters: cell load rate, overlapping coverage, user distribution density, and interference level, and dynamically refreshes energy-saving thresholds.

[0017] This step provides accurate network status input for energy-saving scheduling through real-time optimization of network parameters, solving the problems of traditional energy-saving solutions that rely solely on a single load as a trigger condition, fail to consider the actual operating status of the network, resulting in conservative thresholds and the inability to unleash energy-saving potential.

[0018] S3 Energy Saving Layer Precise Scheduling Based on four core network state parameters, a personalized graded energy-saving threshold is generated for each cell through a nonlinear mapping formula. The formula is applicable to the normal interference range of the 5G network. A graded energy-saving strategy of carrier shutdown or channel shutdown is implemented according to the load range. A two-level backoff protection mechanism with early warning and emergency is configured, and multiple indicators are set to trigger priority to maximize energy-saving benefits while ensuring network KPI stability.

[0019] In this step, the multi-factor dynamic threshold algorithm deeply binds overlapping coverage, user distribution, interference level and energy-saving trigger conditions, replacing the traditional uniform conservative threshold, and improving the depth of energy saving while ensuring coverage and service quality; the two-level rollback mechanism reduces unnecessary full rollback through hierarchical control, balancing energy-saving benefits and user experience.

[0020] S4 Full Lifecycle Closed-Loop Iteration Regularly collect network-wide operational data and adopt a hierarchical iterative strategy of "freezing bottom-level features + incremental updating top-level data" to update the prediction model: During daily periodic updates, freeze the parameters of the LSTM bottom-level hidden layers and update the weights of the top-level fully connected layers only through the FTRL online learning algorithm. This avoids catastrophic forgetting caused by full parameter updates and solves the technical problems of network KPI fluctuations, increased disconnection rates, and latency jitter caused by traditional full retraining. Quarterly, global model retraining is performed based on the full sample set to ensure the long-term adaptability of the model and form a closed-loop mechanism for continuous optimization.

[0021] Furthermore, this invention adopts a scene-adaptive weight switching mechanism. The system automatically identifies three types of scenes: residential areas, industrial parks, and universities, and loads the corresponding weight coefficient combinations. When switching scenes, a linear and smooth transition is adopted to avoid network fluctuations caused by parameter mutations, thereby further improving the energy-saving effect and network stability under different scenes.

[0022] III. Beneficial Effects Compared with the prior art, the present invention has the following beneficial effects, and each technical improvement corresponds to a specific effect on the communication network technology: 1. End-to-end closed-loop architecture reduces energy consumption at the source. By using a three-layer closed-loop architecture to connect data across the entire chain of planning, optimization, and energy saving, and by introducing energy consumption and redundant resource constraints during the planning stage, the traditional planning method that only focuses on coverage capacity can be replaced. This can reduce the deployment of redundant base stations by about 28%, thereby reducing energy consumption and operation and maintenance costs throughout the network's entire life cycle from the source of construction.

[0023] 2. Multi-factor dynamic threshold algorithm enhances energy saving depth and network security. Based on four network status parameters—overlapping coverage, user distribution density, and interference level—a personalized energy-saving threshold for each cell is dynamically generated, replacing the traditional conservative threshold scheme based on a single load. While ensuring network coverage quality, service rate, and handover performance, the overall energy saving rate of the entire network is increased from 15%~20% in the traditional scheme to 32%~38%. At the same time, a two-level hierarchical fallback protection mechanism is configured, with a rapid fallback of 100ms in emergency situations to ensure network operation security.

[0024] 3. A hierarchical incremental update strategy addresses the KPI fluctuation problem during model updates. To address the technical pain points of network parameter mutations, increased disconnection rates, and increased latency jitter caused by full model retraining, a layered iterative strategy of "bottom-level hidden layer freezing + daily FTRL incremental updates + quarterly full retraining" is adopted: bottom-level freezing preserves the stability of general spatiotemporal characteristics, top-level incremental updates adapt to short-term changes in business, reducing network KPI fluctuations caused by model updates by more than 85%, while shortening the time for edge-side model updates by 85%, adapting to deployment scenarios with limited edge cloud computing power.

[0025] 4. Smooth scene adaptation and switching, improving multi-scene adaptability. To address the different business tidal characteristics in scenarios such as residential areas, industrial parks, and universities, target weight combinations are set for different scenarios. A linear smooth transition mechanism is adopted when switching scenarios to avoid network fluctuations caused by parameter mutations. KPI fluctuations are reduced by more than 80% during scenario switching, achieving the optimal balance between energy saving and network quality in different business scenarios.

[0026] This solution has been verified through large-scale live network pilot testing, and all network indicators meet the commercial requirements of operators, demonstrating its value for large-scale promotion. Attached Figure Description

[0027] Figure 1 This is a diagram illustrating the overall architecture of the closed-loop management and control of the entire lifecycle of a 5G base station according to the present invention. Figure 2 This is a flowchart of the pre-planning energy-saving implementation process of the present invention; Figure 3 This is a flowchart illustrating the implementation process of dynamic optimization of the optimization layer in this invention. Figure 4 This is a flowchart illustrating the precise scheduling implementation of the energy-saving layer in this invention. Figure 5 This is a schematic diagram of the closed-loop iteration throughout the entire lifecycle of the present invention; Figure 6 This is a schematic diagram of the adaptive weight switching mechanism of the present invention.

[0028] Figure 7 This is the network structure and data flow diagram for LSTM feature splicing and FTRL hierarchical update in this invention.

[0029] Figure labeling: 1-Data acquisition server; 2-Planning and optimization server; 3-Energy-saving scheduling server; 4-Base station side execution unit; 5-Scene recognition module; 6-Weight parameter library; 7-Smooth transition module. Detailed Implementation

[0030] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0031] The device of this invention is compatible with 5G BBU / RRU hardware equipment from mainstream manufacturers such as Huawei, ZTE, and Ericsson. The invention will be described in detail below through embodiments.

[0032] Example 1: System Overall Architecture As shown in Figure 1, the device of this invention adopts a cloud + edge cloud distributed architecture, including four hardware units: The data acquisition server (1) is deployed on the operator's core network side. The standard configuration is an 8-core 16G CPU and 1TB SSD storage. It connects to the entire network base station through the OMC northbound interface to collect data such as MR measurement reports, signaling, cell load, total power consumption, handover, and OTT speed measurement. The collection cycle can be configured to 4~16 minutes, and the instruction scheduling response time is controlled within 1 second.

[0033] The planning and optimization server (2) is deployed in the cloud. The standard configuration is an 8-core 32G cloud host. It communicates with other servers through the HTTP / HTTPS encrypted protocol. It has built-in data preprocessing module, LSTM prediction module, four-dimensional planning module, NSGA-Ⅱ multi-objective optimization module, and network optimization module. The modules are interconnected through the high-speed bus inside the server. The data synchronization delay does not exceed 200ms. It is responsible for site planning, network optimization parameter calculation and quarterly global model retraining at the global level.

[0034] The energy-saving scheduling server (3) is deployed on the edge cloud node, close to the base station cluster. The standard configuration is a 16-core 32G CPU and a local Redis 6.0 memory database. It is directly connected to the base station through the operator's intranet dedicated line. The delay of the base station control command issuance is ≤1 second. It has built-in threshold calculation module, cluster energy-saving scheduling module, hierarchical rollback protection module, incremental learning module, and cross-vendor command adaptation layer. The modules achieve high-speed data communication through shared memory. It is responsible for the real-time energy-saving scheduling, rollback protection and daily model incremental update of the local base station cluster.

[0035] The base station side execution unit (4) consists of a 5G base station BBU baseband processing unit and an RRU radio frequency remote unit. It receives control commands issued by the energy-saving scheduling server (3), performs energy-saving operations such as radio frequency channel shutdown, carrier shutdown, and cell merging, and transmits the execution results and real-time status back to the energy-saving scheduling server (3).

[0036] Cross-vendor adaptation mechanism: This invention sets up a vendor instruction adaptation layer in the energy-saving scheduling server (3) to uniformly encapsulate the shutdown instruction format of base stations from different vendors, provides a standardized shutdown / start instruction interface upwards, and adapts to the private instruction formats of different vendors such as Huawei, ZTE, and Ericsson downwards, thereby realizing unified management and control of cross-vendor equipment.

[0037] Example 2: Pre-planning energy-saving process (S1) like Figure 2 As shown, the pre-planning energy conservation specifically includes the following steps: Step S101 Multi-source data acquisition and preprocessing Collect the following data from the target area: GIS terrain data includes digital elevation models, land feature types, and building vector data; Population distribution data: Grid-level population density data based on operator signaling data statistics; Existing base station coverage data: Engineering parameters such as the location, antenna height, azimuth, downtilt angle, and transmit power of existing base stations; Business tidal data: PRB utilization rate, RRC connection count, and traffic data for each cell in the region for more than 7 consecutive days and 24 hours; Power consumption parameters: Actual measured power consumption of the base station, carrier power consumption, and channel power consumption for each model.

[0038] Data preprocessing workflow: Data cleaning: Missing values ​​are filled with data from adjacent periods or historical data from the same period, and outliers are removed using the 3σ criterion; Data normalization: The max-min normalization method is used to map all feature values ​​to the interval [0,1]. Data alignment: Unify alignment to hourly granularity to construct time series samples.

[0039] Regional population activity calculation method: The total number of user terminal attachments, handovers, and signaling interactions within a cell within a unit of time (1 hour) is counted, and the average value of all cells in the region is used as a benchmark to normalize and obtain the population activity index of each cell.

[0040] Data Compliance and Privacy Protection Statement: In this embodiment, the signaling data, OTT speed test data, and population distribution data collected are all legally generated network management statistics during the operation of the operator's network. Irreversible anonymization and desensitization processes have been strictly implemented at the source. Specifically, the system only extracts statistical or grid-level (e.g., 500m×500m grid) aggregation features, completely removing personal privacy information that can identify specific natural persons, such as IMEI, IMSI, and mobile phone numbers. All data collection, transmission, and storage processes comply with the requirements of the National Data Security Law, the Personal Information Protection Law, and the data security management standards of the communications industry, and do not involve any data processing behaviors that violate social morality or public interest.

[0041] Step S102: Construction of LSTM Spatiotemporal Service Energy Consumption Prediction Model A prediction model is constructed using a dual-hidden-layer LSTM network, and the network structure parameters are shown in Table 1.

[0042]

[0043] Model training parameter settings: Optimizer: Adam, initial learning rate 0.001; Training rounds: Initial training for 100 rounds, using an early stopping mechanism (stop if the validation set loss does not decrease for 10 consecutive rounds); Batch size: 64; Loss function: Mean Squared Error (MSE); Dataset partitioning: Partitioned chronologically, with 70% training set, 15% validation set, and 15% test set; Weight initialization: Xavier uniform distribution initialization.

[0044] Weather feature coding method: Temperature, precipitation and wind force are normalized separately and concatenated into a 3D weather feature vector to correct the impact of outdoor travel on business volume; Time period label adopts 24-hour one-hot encoding to represent the business tidal characteristics of different time periods.

[0045] Feature Engineering and Physical Meaning Mapping Logic: To address the lack of physical interpretability of input features in traditional AI models, this embodiment features a targeted design for 10-dimensional spatiotemporal feature vectors. For discrete features like 'time period labels' and 'weather environment', instead of simple one-hot encoding, the model uses an embedding layer to transform them into low-dimensional dense vectors, capturing the nonlinear impact of temporal periodicity and external environment on spatial service distribution. For continuous features like 'traffic' and 'PRB load', Min-Max normalization is employed. Finally, these dense vectors and continuous features are concatenated along the channel dimension, enabling the LSTM model to simultaneously learn the 'temporal tidal effect' and 'spatial hotspot drift' patterns of 5G base station services, thereby significantly improving the prediction accuracy of spatiotemporal service energy consumption.

[0046] Step S103 Construction of the four-dimensional joint objective function Establish a four-dimensional joint objective function: F = αC + βV - γE - δR in: C represents the coverage quality index: the percentage of sampling points with RSRP ≥ -105dBm, with a value range of [0,1]. V is a network capacity indicator: the average downlink rate of cell users, normalized to the [0,1] interval; E represents the total energy consumption of the base station: the sum of the power consumption of all base stations within the planned area, normalized to the [0,1] interval; R represents the percentage of redundant resources: the percentage of unused PRB resources during off-peak hours (a continuous hour when the load rate is below 30%), with a value range of [0,1]. α, β, γ, and δ are weighting coefficients that satisfy α+β+γ+δ=1. These are fixed values ​​preset offline through orthogonal experiments for different scenarios and do not participate in the algorithm optimization.

[0047] Criteria for determining cover voids: Using a 10-meter grid, any 25 or more consecutive grids with RSRP < -105dBm are considered as contiguous cover voids, which serve as a planning constraint.

[0048] Step S104 NSGA-II Multi-Objective Optimization The non-dominated sorting genetic algorithm (NSGA-II) with elitist strategy is used to solve the Pareto optimal programming scheme. Different parameter configurations are adopted for different scenarios, and the specific parameters are shown in Table 2.

[0049]

[0050] Fitness function: A four-dimensional objective function value plus a constraint penalty term is used. The penalty function has the following form: P =λ1 max(0,C req -C real )+λ2 max(0,O req -O real )+λ3 N hole M in: P is the penalty value for the coverage scheme; C req To meet coverage requirements; C real This represents the actual coverage rate of the plan. O req This represents the upper limit of overlap coverage; O real This represents the actual overlap coverage of the scheme. N hole To cover the number of holes; M is the large number penalty factor, with a value of 1000; λ1, λ 2、 λ3 is the weight coefficient of the penalty term; The algorithm outputs a Pareto optimal solution set, and the TOPSIS method is used to select the compromise optimal solution as the final base station planning scheme.

[0051] Example 3: Dynamic tuning process of the optimization layer (S2) like Figure 3 As shown, the dynamic tuning of the optimization layer specifically includes the following steps: Step S201 Multi-source data acquisition The following data was collected via the OMC northbound interface: MR measurement report data: sampling period 4~16 minutes, including RSRP, RSRQ, and neighbor cell measurement data; Signaling data: RRC connection establishment, handover, and release signaling; Load data: cell PRB utilization, CPU utilization, traffic data; Data switching: Number of switching attempts, number of successful switching attempts, and reason for switching; OTT speed test data: User downlink rate sampling data.

[0052] An aggregated statistical report is generated every 15 minutes.

[0053] Step S202: Parallel execution of three types of network optimization 1. Overlapping Coverage Optimization: Statistically analyze the cell overlap coverage (the percentage of sampling points where the RSRP difference between three or more strong neighboring cells is within 6dB). For areas with overlap coverage exceeding 30%, adjust the antenna downtilt angle (in 2° increments) and power offset (in +3dB increments) to control the overlap coverage within the target range of 15% to 30%.

[0054] 2. Load balancing optimization: Statistically analyze the physical layer load rate of cells. For high-load cells with a load rate exceeding 70%, switch edge users to neighboring cells with lower load. For low-load cells with a load rate below 30%, appropriately absorb edge users from neighboring cells to balance the load rate of all cells in the network to the range of 30% to 70%.

[0055] 3. Interference Coordination Optimization: Statistically analyze the uplink interference level of cells. For cells with interference levels exceeding -110dBm, suppress the interference level to below -110dBm through PCI optimization, power control, ICIC interference coordination, and other means.

[0056] Step S203 Core Parameter Synchronization The optimized four core parameters—cell load rate, overlap coverage, user distribution density, and interference level—are synchronized in real time to the energy-saving scheduling server (3) via the internal bus, and the energy-saving trigger thresholds of each cell are dynamically refreshed.

[0057] User distribution density calculation method: The average number of users connected to RRC at the hourly level is counted for each cell. The average number of users in all cells within the planning area is used as the benchmark, and the user distribution density D value of each cell is obtained by normalization.

[0058] Example 4: Precise scheduling process for the energy-saving layer (S3) like Figure 4 As shown, the precise scheduling of the energy-saving layer specifically includes the following steps: Step S301 Personalized grading threshold calculation Based on four core network state parameters, personalized carrier shutdown thresholds and channel shutdown thresholds are calculated for each cell. The calculation formulas are as follows: T = T0 × (1 + k1×O + k2×D – k3×I) The formula is applicable to the normal operating range of 5G networks with cell interference levels ranging from -130dBm to -100dBm. The parameters are defined as follows: T0 is the baseline threshold: carrier shutdown baseline T0=30%, channel shutdown baseline T0=70%, which is the unified initial calculation baseline for the entire network; O is the normalized value for overlap coverage: O = actual overlap coverage / 30%; D is the normalized value of user distribution density: D = average number of users per cell / average number of users per area; I is the normalized value of the interference level: I = (-110dBm) / actual interference level; K1, k2, and k3 are correction coefficients with values ​​of 0.3, 0.2, and 0.25, respectively, calibrated through orthogonal experiments.

[0059] Threshold constraint rule: To ensure network security, boundary values ​​are used when calculation results exceed a reasonable range. Carrier shutdown threshold constraint range: 20%~40% load rate; Channel shutdown threshold constraint range: 50%~80% load rate.

[0060] Correction coefficient orthogonal test calibration method: Three factors, K1, K2, and K3, were selected, and each factor was set with 5 levels (0.1, 0.2, 0.3, 0.4, and 0.5). The L25 (5-factor) method was used. 3 An orthogonal array design experiment was conducted, using the weighted sum of energy saving rate and KPI retention rate as the evaluation index. The experiment was carried out on a dataset of 23,000 existing stations. The optimal parameter combination was determined to be k1=0.3, k2=0.2, and k3=0.25 through range analysis and variance analysis.

[0061] Step S302 Base station cluster collaborative energy saving Using the personalized carrier shutdown threshold and personalized channel shutdown threshold calculated above as the sole dividing points, the cell load is divided into three categories and corresponding operations are performed: Low load range (load rate < calculated carrier shutdown threshold): Carrier shutdown operation is performed when the RSRP of the strongest neighboring cell is ≥ -105dBm. Medium load range (carrier shutdown threshold ≤ load rate < channel shutdown threshold): Perform single-channel RF shutdown operation; High load range (load rate ≥ channel shutdown threshold): No shutdown operation will be performed.

[0062] Based on the tidal patterns of business operations, redundant micro and macro base station auxiliary carriers with sufficient neighboring cell coverage but no effective service carrying capacity are shut down in batches during low-load periods.

[0063] Differentiated strategies are set for specific business scenarios: When the regional service load is consistently below 20% load rate for a long period and the air interface latency jitter rate is below 10%, a deep sleep strategy is triggered to shut down the radio frequency channels of microcells without service carrying; when the regional service load exhibits short-cycle high-frequency pulse characteristics and the uplink throughput accounts for more than 40%, an industrial-grade low latency guarantee strategy is triggered to dynamically increase the base station transmit power and disable the symbol-level shutdown function to ensure the stability of uplink service latency.

[0064] Step S303 Two-level graded backoff protection Real-time monitoring of network KPIs and configuration of a two-level rollback mechanism are shown in Table 3. The rollback threshold configuration for general scenarios is shown in Table 3.

[0065]

[0066] For industrial park scenarios, an additional air interface latency indicator is added: below 8ms is considered normal, above 8ms triggers an alert, and above 10ms triggers an emergency rollback. The triggering priority is consistent with the dropout rate. For university scenarios, during winter and summer vacations, the user's average downlink speed alert threshold is adjusted to 40Mbps, and the emergency threshold is adjusted to 30Mbps.

[0067] Trigger priority rules: Emergency thresholds have higher priority than warning thresholds; triggering any emergency threshold will immediately trigger a full rollback. When multiple early warning indicators are triggered simultaneously, the control measures are executed in the following priority order: disconnection rate > handover success rate > coverage rate > average speed. Within 100ms after the emergency threshold is triggered, all energy-saving strategies are revoked, and the base station resumes full-power transmission.

[0068] Rollback and recovery mechanism: After all monitoring indicators recover to above the warning threshold and remain stable for 5 minutes, the energy-saving operation is gradually restarted in the order of "first restoring the carrier shutdown strategy, then restoring the channel shutdown strategy" to avoid network fluctuations caused by frequent strategy switching.

[0069] Example 5: Full Lifecycle Closed-Loop Iterative Process (S4) like Figure 5 As shown, the full lifecycle closed-loop iteration specifically includes the following steps: Step S401 Daily Data Collection The system will automatically execute the following from 2:00 AM to 4:00 AM daily: Summarize the energy consumption data, KPI data, and energy-saving operation log data of all base stations across the network from the previous day; Complete data validation: remove outliers and fill in missing values; After standardizing the format, it is stored in the historical database.

[0070] Step S402 FTRL Daily Incremental Update The LSTM model is updated using the FTRL online incremental learning algorithm (daily batch update mode), and the algorithm parameters are configured as shown in Table 4.

[0071]

[0072] Parameter freezing strategy: During the daily incremental update phase, all weights and bias parameters of the two hidden layers of the LSTM are set to non-trainable (trainable=False), freezing the spatiotemporal feature extraction capability of the underlying layer and preventing daily small-sample updates from damaging the general feature representation; during the quarterly global retraining phase, all network parameters are unfrozen, and full training is completed using the full dataset. Incremental update: Only the top fully connected output layer is set to trainable, and the output layer weights are updated using the new data from the previous day. After the update, KPI fluctuations are reduced by more than 85% compared to a full update. Sample management: Automatically remove historical samples older than 3 months to maintain a stable training sample library size.

[0073] Step S403: Quarterly global retraining Global model retraining is automatically triggered by the planning and optimization server at 2:00 AM on the last day of each quarter. The entire LSTM network (including all layers) was retrained using all valid samples from the past three months. The accuracy and stability of the new model were validated on an independent test set. Accuracy was measured by the mean squared error (MSE) between the predicted and actual values, and stability was measured by the variance of network KPI fluctuations. If the overall performance of the new model (accuracy × 0.7 + stability × 0.3) is better than the current online model, then a gray-scale replacement will be implemented: first, select 10% of base stations for trial operation for 24 hours, and after confirming that there are no abnormalities in the KPIs, the model will be fully deployed; otherwise, the current model will continue to be used, and retraining will be carried out in the next quarter.

[0074] Step S404: Model Iteration Closure The updated model parameters are fed back to the planning layer and enter the next round of planning optimization process, forming a complete closed loop.

[0075] Explanation of the technical mechanism for freezing the bottom hidden layers: This solution innovatively proposes a layered iterative strategy of "freezing the two bottom hidden layers of the LSTM during the daily update phase and updating only the top fully connected layer; and conducting full-parameter training during the quarterly retraining phase." The technical mechanism lies in the fact that the bottom hidden layers of the LSTM are mainly responsible for extracting the general periodicity and temporal dependence features of the spatiotemporal sequence (such as day-night peak-valley patterns). These features exhibit high generalization and stability under different seasons or sudden scenarios. The top fully connected layer is responsible for mapping these general features to specific load and power consumption values. When the 5G network faces data distribution drift caused by holidays or sudden data spikes, updating only the top weights can quickly adapt to recent local data changes, effectively avoiding the "catastrophic forgetting" phenomenon caused by full-parameter updates. This ensures a smooth transition of network KPIs (such as drop rate and latency) during energy-saving strategy adjustments and eliminates the risk of drastic KPI fluctuations.

[0076] Example 6: Scene Adaptive Weight Switching Mechanism like Figure 6 As shown, the scene adaptive weight switching includes the following modules: Scene recognition module (5) Regional attribute identification: Based on the latitude and longitude of the base station, a geographic information database is matched to identify the type of region where the base station is located (residential area, industrial park, university). Calendar tag recognition: Identifies calendar tags such as weekdays, weekends, winter and summer vacations, and holidays based on system dates; Determine the scenario type based on comprehensive factors.

[0077] The weight parameter library (6) stores the optimal combination of pre-calibrated weight coefficients for three types of scenarios: Residential area scenario: α=0.3, β=0.3, γ=0.25, δ=0.15 (balanced coverage, capacity and energy consumption); Industrial park scenario: α=0.28, β=0.27, γ=0.23, δ=0.20 (focusing on energy consumption and redundancy optimization to ensure low latency); University setting: α=0.26, β=0.26, γ=0.26, δ=0.22 (balancing the four indicators to adapt to significant business fluctuations during winter and summer vacations).

[0078] Orthogonal test calibration method for weighting coefficients: For each scenario, four weight coefficients (α, β, γ, δ) are assigned (satisfying α + β + γ + δ = 1), with four levels for each factor, using L16(4)4(1)2 ... 4 An orthogonal array design experiment was conducted, using the weighted sum of coverage compliance rate, speed compliance rate, energy saving rate, and redundancy reduction rate as a comprehensive evaluation index. 100 base stations were selected for each scenario for a 7-day experiment, and the optimal weight combination for each scenario was determined through range analysis.

[0079] When the smooth transition module (7) switches scenes, it uses a 24-hour linear interpolation method to smooth the transition of weight parameters, avoiding network index fluctuations caused by parameter mutations. The KPI fluctuation during the switching process is reduced by more than 80%. The weight calculation formula at time t is:

[0080] In the formula: w(t) represents the service scheduling weight at time t; w old The steady-state scheduling weight before the switchover is determined by the load percentage of the previous scheduling cycle; w new The target steady-state weight after the switch is determined by the load forecast results of the current period; t0 is the start time of the weight switching, in hours; t represents the current calculation time, in hours; T S The total period for the smooth transition of weights is 24 h in this embodiment.

[0081] Example 7: Typical Scenario Application Effects Three typical scenarios—residential areas, industrial parks, and universities—were selected for pilot verification. The pilot results for each scenario are shown in Table 5.

[0082]

[0083] Combination Figure 7 As shown, this embodiment breaks away from the "black box" structure of traditional AI models. During the feature engineering stage, discrete features are mapped into dense vectors through the Embedding layer, and then concatenated with normalized continuous features at the Concat node at the channel level. This allows the model to explicitly learn the spatiotemporal coupling patterns of 5G services. More importantly, during the online model update stage (such as...), Figure 7 (As shown in the dashed box), the system uses the FTRL algorithm to update the gradient of only the weight matrix of the top fully connected layer, while completely freezing the parameter matrix of the bottom hidden layer. This 'layered decoupling' network structure design ensures the stability of the model's underlying general features from a physical architecture perspective, which is the core technical support for this solution to maintain stable KPIs during data distribution drift.

[0084] Application in residential areas: A pilot project was conducted in a residential area of ​​0.8 square kilometers, with a planned population of 25,000. Using residential area weights α=0.3, β=0.3, γ=0.25, and δ=0.15, the NSGA-II algorithm output a planning scheme of 3 macro-stations + 4 micro-stations, reducing the number of micro-stations by 4 compared to the traditional 8-micro-station scheme. During operation, total energy consumption decreased by 32%, while all KPI indicators remained stable.

[0085] Industrial park application scenarios: A pilot project was conducted in an industrial park. Using industrial park weights α=0.28, β=0.27, γ=0.23, and δ=0.20, the planned scheme consisted of 3 macro base stations + 5 micro base stations, reducing the number of micro base stations by 4 compared to the traditional 9 micro base station scheme. Air interface latency monitoring was added, resulting in a 35% reduction in total energy consumption during operation. The air interface latency remained stable within 8ms, meeting the low-latency requirements of industrial control.

[0086] Application in university settings: A pilot program was conducted on a university campus. Using university weights of α=0.26, β=0.26, γ=0.26, and δ=0.22, the planned scheme consisted of 3 macro base stations + 3 micro base stations, reducing the number of micro base stations by 4 compared to the traditional 7 micro base station scheme. During winter and summer vacations, 80% of the inactive micro base stations went into hibernation, resulting in a 38% reduction in total energy consumption.

[0087] Cross-vendor verification: ZTE and Ericsson equipment were deployed separately under the same conditions for verification. The energy saving rate of ZTE equipment pilot was 32.0%, and the energy saving rate of Ericsson equipment pilot was 32.0%. All KPI indicators remained stable, verifying the cross-vendor universality of the solution.

Claims

1. A 5G base station planning optimization energy-saving full-cycle closed-loop intelligent management and control method for 5G base station full-life cycle energy consumption management and control, characterized in that, A three-layer closed-loop linkage architecture is established, consisting of "planning layer pre-energy saving, optimization layer dynamic tuning, and energy saving layer precise scheduling". The planning layer outputs standardized base station planning schemes to guide the optimization layer parameter tuning. The optimization layer outputs four core network status parameters, namely cell load rate, overlapping coverage, user distribution density, and interference level, to support the precise scheduling of the energy saving layer. The energy consumption and KPI data generated by the implementation of the energy saving layer are fed back to the planning layer to complete the model iteration and update. The overall process comprises four stages: S1 planning layer for pre-emptive energy saving, S2 optimization layer for dynamic tuning, S3 energy-saving layer for precise scheduling, and S4 full lifecycle closed-loop iteration. S1 Planning Layer Pre-Energy Saving: Collects multi-source basic data on the target area, including geographical topography, population distribution, existing base station coverage, service tidal volume, and base station power consumption. After preprocessing, it constructs a spatiotemporal service energy consumption prediction model and outputs the service load and base station power consumption prediction values ​​for each cell in the target area. It establishes a joint objective function that includes four dimensions: coverage quality, network capacity, total base station energy consumption, and redundant resource ratio. It uses a multi-objective intelligent optimization algorithm to globally optimize the objective function and outputs the Pareto optimal base station planning scheme. S2 Optimization Layer Dynamic Tuning: Periodically collects measurement reports, signaling, load, handover, and network speed test data from all cells in the network and generates aggregated statistical reports; performs overlapping coverage optimization, load balancing optimization, and interference coordination optimization in parallel, and synchronizes the four optimized core network status parameters to the energy-saving scheduling unit in real time; S3 Energy Saving Layer Precise Scheduling: Based on four core network status parameters, personalized carrier shutdown thresholds and channel shutdown thresholds are calculated for each cell; the cell load is divided into three levels: low, medium, and high according to the thresholds, and a graded energy saving strategy of carrier shutdown, single channel radio frequency shutdown, and normal operation is executed accordingly. Configure a two-tiered fallback protection mechanism with early warning and emergency levels to ensure network operation quality; S4 Full Lifecycle Closed-Loop Iteration: Regularly collect energy consumption, KPI, and energy-saving operation data of all base stations in the network; adopt a hierarchical incremental update strategy to periodically update the prediction model, freeze the bottom feature extraction parameters of the model during the update process, and only update the top output layer parameters; complete global model retraining based on all valid samples at a longer period, and the updated model is fed back to the planning layer to enter the next round of planning optimization process, forming a full lifecycle closed-loop iteration.

2. The control method according to claim 1, characterized in that, The S1 uses a dual-hidden-layer LSTM network to construct a spatiotemporal service energy consumption prediction model. The model input is a spatiotemporal feature vector containing real-time user count, uplink and downlink traffic, peak load during busy hours, resource block occupancy rate during idle hours, regional population activity, weather environment, time period label, average interference of surrounding base stations, and historical daily average power consumption. The model output is the predicted service load and base station power consumption of each cell in the target area for future time periods. The joint objective function is: F=αC+βV-γE-δR Where C is the coverage quality index, V is the network capacity index, E is the total energy consumption of the base station, R is the proportion of redundant resources, and α, β, γ, δ are the corresponding weight coefficients and satisfy the condition that their sum is 1; the NSGA-II non-dominated sorting genetic algorithm with elite strategy is used for global optimization, and the algorithm adopts a real number encoding method of base station coordinates, transmit power, and antenna downtilt angle.

3. The control method according to claim 2, characterized in that, In S1: a 10-meter grid is used as the standard for dividing the grid, and no less than 25 continuous grids are used as the criterion for identifying continuous coverage holes; the NSGA-II algorithm has 100-150 iterations, a population size of 50-80, a crossover probability of 0.8, and a mutation probability of 0.

1. The weighting coefficients were determined through offline screening using orthogonal experiments. For residential areas, the coefficients were α=0.3, β=0.3, γ=0.25, and δ=0.15; for industrial parks, the coefficients were α=0.28, β=0.27, γ=0.23, and δ=0.20; and for universities, the coefficients were α=0.26, β=0.26, γ=0.26, and δ=0.

22.

4. The control method according to claim 1, characterized in that, In S2: data is collected at a cycle of 4 to 16 minutes through the OMC northbound interface, and an aggregate report and air interface latency data are generated every 15 minutes; the target range for overlap coverage adjustment is 15% to 30%, the target range for cell load rate balancing is 30% to 70%, and the target for interference level suppression is below -110dBm.

5. The control method according to claim 1, characterized in that, The threshold calculation formula in S3 is as follows: T = T0 × (1 + k1 × O + k2 × D - k3 × I) The formula is applicable to the normal operating range of 5G networks with cell interference levels ranging from -130dBm to -100dBm. In the formula, T0 is the reference threshold, the carrier shutdown reference is 30%, and the channel shutdown reference is 70%; O is the normalized value of overlap coverage, which is calculated as the ratio of actual overlap coverage to 30%; D is the normalized value of user distribution density, which is calculated as the ratio of the average number of users in the cell to the average number of users in the area. I is the normalized value of the interference level, which is calculated as the ratio of -110dBm to the actual interference level. Correction factors k1=0.3, k2=0.2, k3=0.25; The calculated thresholds are subject to mandatory constraints: the carrier shutdown threshold is limited to the range of 20% to 40%, and the channel shutdown threshold is limited to the range of 50% to 80%. When the thresholds are exceeded, the boundary values ​​are taken.

6. The control method according to claim 1, characterized in that, In S3: a load rate below the personalized carrier shutdown threshold is a low load range, a load rate between the personalized carrier shutdown threshold and the personalized channel shutdown threshold is a medium load range, and a load rate above the personalized channel shutdown threshold is a high load range. Carrier shutdown is performed on low-load cells under the condition that the RSRP of the strongest neighboring cell is ≥-105dBm; the hierarchical backoff protection mechanism uses coverage, average downlink rate of users, drop rate, and handover success rate as monitoring indicators, and sets early warning thresholds and emergency thresholds respectively; When multiple early warning indicators are triggered simultaneously, control is performed in descending order of priority: call drop rate, handover success rate, coverage rate, and average rate. When the emergency threshold is triggered, the base station resumes full power operation within 100ms. After all indicators have recovered to above the early warning threshold and stabilized for 5 minutes, energy-saving operation is resumed in the order of carrier shutdown followed by channel shutdown.

7. The control method according to claim 1, characterized in that, The S3 also includes a scenario-based energy-saving strategy: when it is detected that the regional service load is continuously lower than the preset low threshold for a long period of time and the air interface latency jitter rate is lower than the preset threshold, a deep sleep strategy is triggered to shut down the micro base station radio frequency channel. When the regional service load is identified to exhibit short-cycle high-frequency pulse characteristics and the uplink throughput ratio exceeds the preset ratio, a low-latency protection strategy is triggered to dynamically increase the base station's transmit power and disable the symbol-level shutdown function. The S4 system uses the FTRL online incremental learning algorithm for daily batch updates. The system retains valid sample data for the past three months on a rolling basis and completes global model retraining on the last day of each quarter. The system automatically identifies scene types and matches corresponding weight coefficients. When switching scenes, linear interpolation is used to smoothly transition weight parameters.

8. A 5G base station planning optimization energy-saving closed-loop intelligent management and control device, characterized in that, It adopts a cloud + edge cloud distributed architecture, including a data acquisition server, a planning and optimization server, an energy-saving scheduling server, and a base station-side execution unit that interact with data sequentially. Data acquisition server: Deployed on the operator's core network side, it connects to the base station through the OMC northbound interface to collect network-wide operational data; Planning and optimization server: Deployed in the cloud, with built-in data preprocessing module, prediction module, multi-objective planning module, and network optimization module, used to execute the S1 planning layer pre-energy saving, S2 optimization layer dynamic tuning and global model retraining process in claim 1; Energy-saving dispatch server: Deployed at edge cloud nodes and directly connected to base stations via dedicated intranet lines of the operator; The system includes a built-in threshold calculation module, a cluster energy-saving scheduling module, a hierarchical rollback protection module, and an incremental update module, which are used to execute the precise scheduling and periodic incremental update process of the S3 energy-saving layer in claim 1. Base station side execution unit: including 5G base station BBU baseband processing unit and RRU radio frequency remote unit, receives instructions from energy-saving scheduling server, performs radio frequency channel shutdown and carrier shutdown operations, and sends the execution results back to energy-saving scheduling server.

9. The control device according to claim 8, characterized in that, In the planning and optimization server: the data preprocessing module completes the cleaning, normalization, and outlier removal of multi-source collected data; the prediction module adopts a LSTM network structure with 64 neurons in two hidden layers; the multi-objective planning module runs a four-dimensional joint objective function and the NSGA-II multi-objective optimization algorithm; and the network optimization module performs three types of optimization in parallel: overlapping coverage, load balancing, and interference coordination. The energy-saving scheduling server is equipped with a cross-vendor instruction adaptation layer, which uniformly encapsulates the shutdown instruction formats of base stations from different vendors, and enables high-speed data communication between modules through shared memory.

10. A non-volatile computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is run by the processor, it implements the 5G base station planning optimization energy-saving full-cycle closed-loop intelligent management and control method as described in any one of claims 1 to 7.

Images