Multi-dimensional coupled base station energy consumption cooperative control method and system

Abstract

The application discloses a multi-dimensional coupling base station energy consumption cooperative control method and system, through digital twin pre-performance and reinforcement learning, no longer relying on single rules, but taking the lowest total energy consumption, the longest equipment life and user perception as multiple targets, realizing global optimal control, significantly reducing the base station energy consumption, combining real-time data and future load prediction, changing 'passive response' to 'active adjustment', avoiding energy waste or overheating risk caused by adjustment lag, realizing the linkage of energy supply, temperature control and business resources, and solving the low efficiency problem caused by independent control of each subsystem of the traditional base station.

Description

Technical Field

[0001] This invention relates to the field of base station control, and in particular to a multi-dimensional coupled base station energy consumption collaborative control method and system. Background Technology

[0002] Base stations are the core infrastructure of mobile communication networks. Simply put, they are the "bridge" or "relay station" connecting wireless terminals such as mobile phones with wired communication networks. Without base stations, mobile phones cannot make calls, send text messages, or access the Internet.

[0003] Traditional base station energy-saving measures are often isolated and focused on single-point optimizations. For example, optimizing only the air conditioning system or only the power amplifier equipment lacks systematic coordination and fails to maximize overall energy efficiency. It fails to deeply couple and coordinate the three key dimensions of energy supply, thermal management, and communication service load. Energy, heat, and information flow are usually managed independently, resulting in ineffective energy transfer over time (e.g., utilizing off-peak electricity or excess solar energy) and inefficient spatial allocation. Traditional control strategies are mostly reactive based on the current state, such as turning on the air conditioning only when the temperature is high, lacking AI-based forward-looking prediction of service load. This prevents the smooth adjustment of equipment power or coverage reconstruction in advance, easily leading to power consumption surges from frequent start-stop cycles, or maintaining high energy consumption even under low load. Traditional air cooling or simple temperature control methods fail to fully utilize natural cooling sources, such as comparing temperature differences while ignoring dew points, and do not recover waste heat or utilize phase change materials for thermal energy storage, resulting in high cooling energy consumption.

[0004] In summary, a multi-dimensional coupled base station energy consumption collaborative control method and system is needed to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a multi-dimensional coupled base station energy consumption collaborative control method and system, aiming to solve the aforementioned problems.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a multi-dimensional coupled base station energy consumption collaborative control method, comprising the following steps:

[0007] Step S1: Collect base station data by using multi-dimensional sensor fusion technology to collect real-time operating data of the base station;

[0008] Step S2: Simulate energy consumption changes. Based on the real-time operating data, construct a real-time digital twin model of the base station in the cloud or at the edge to simulate energy consumption and temperature field changes under different control strategies.

[0009] Step S3: Predict future base station load. Analyze historical call volume, holiday effects, and surrounding traffic and pedestrian flow data using deep learning algorithms to predict the base station load within a preset time period.

[0010] Step S4: Generate a collaborative control strategy. Based on a multi-objective reinforcement learning algorithm, the optimal collaborative control strategy is generated by using the lowest total energy consumption, the longest equipment lifespan, and the user perception rate not lower than a preset threshold as multiple reward functions, combined with the simulation results of the digital twin model and the base station load prediction.

[0011] Step S5: Execute the control strategy, and implement the optimal collaborative control strategy to coordinate and adjust the intelligent solar thermal integrated power supply unit, the fluid dynamics intelligent temperature control network, and business resources.

[0012] Optionally, step S1 includes the following sub-steps:

[0013] Step S11: Deployment and data acquisition of real-time radiation irradiance monitoring. A high-precision radiation irradiance sensor is integrated into the base station's solar dual-axis automatic tracking device to monitor the light intensity and spectral distribution of the current environment in real time, and the collected radiation data is transmitted to the control unit in real time.

[0014] Step S12: Precisely locate the junction temperature of the device using infrared thermal imaging. Deploy an infrared thermal imager at the key nodes of high-power core components to perform non-contact scanning on the device surface, generate a real-time temperature field distribution map, accurately locate local hot spots, and identify chips or modules with abnormal temperature rise.

[0015] Step S13: Data collection of micro-prediction probe for service traffic. Deploy the micro-prediction probe for service traffic to capture the current signaling interaction data, number of user connections and instantaneous throughput of the base station in real time. Combine the historical call volume database, holiday effect markers and surrounding traffic and pedestrian flow data to form a multi-dimensional input feature vector.

[0016] Step S14: Multi-source data fusion and digital twin mapping. The collected irradiance data, infrared thermal imaging temperature field data, and service traffic micro data are time-stamp aligned and spatial coordinate matched. Using a multi-dimensional sensor fusion algorithm, noise errors from a single sensor are eliminated, and a high-fidelity real-time operating status vector of the base station is constructed.

[0017] Optionally, step S2 includes the following sub-steps:

[0018] Step S21: Multidimensional data mapping and model initialization. Receive the fused real-time running data, and on the cloud server or edge computing node, call the pre-set base station high-fidelity geometric model and physical attribute library to map the real-time data to the corresponding node of the digital twin and initialize the digital twin state at the current moment.

[0019] Step S22: Construct a multiphysics coupled simulation environment, activate the multiphysics coupled solver in the digital twin model, simulate the solar spectral splitting efficiency, photovoltaic power generation output, heat absorption and release process of phase change materials and battery charging and discharging characteristics through the energy flow model, and simulate the airflow organization, heat transfer efficiency of liquid cooling circuit and heat exchange process of indirect evaporative cooling in the computer room based on the principle of computational fluid dynamics.

[0020] Step S23: Generate and inject a set of candidate control strategies. Based on the current state, the decision layer generates a set of candidate collaborative control strategies and injects them as control variables into the execution mechanism interface of the digital twin model.

[0021] Step S24: Parallel extrapolation and multi-dimensional index simulation. Multiple simulation instances are run in parallel in the digital space. Each instance corresponds to a candidate control strategy. The system evolution within a preset time window is extrapolated, and the power consumption curve and heat generation changes under different business volumes are simulated.

[0022] Step S25: Output the policy evaluation matrix, collect the output results of all parallel simulation instances, construct the policy effectiveness evaluation matrix, quantify the score of each policy, and transmit the evaluation matrix to the multi-objective reinforcement learning algorithm module as the basis for selecting the optimal policy.

[0023] Optionally, step S3 includes the following sub-steps:

[0024] Step S31: Multi-source heterogeneous data cleaning and feature engineering. Extract time series data of base station call volume from the past few months or even years, perform missing value imputation and outlier smoothing, construct calendar feature vectors, introduce time offset features before and after holidays, access traffic flow data and population heat map data in the base station coverage area, transform them into external covariate features with the same time granularity as call volume, strictly align them on the time axis, and construct a multi-dimensional input feature matrix.

[0025] Step S32: Construct a deep learning prediction model architecture. Adopt a deep learning network architecture and design a multi-branch input layer: receive historical call sequence branches, calendar feature branches and external traffic flow branches respectively. Introduce an attention mechanism to focus on the historical moments and key external events that have the greatest impact on the current load.

[0026] Step S33: Model training and spatiotemporal pattern mining. The deep learning model is trained offline using historical datasets. The network weights are optimized through backpropagation algorithm to learn and mine nonlinear relationships in the data.

[0027] Step S34: Load prediction execution. The latest data collected in real time is input into the trained model to perform rolling prediction and output a high-resolution load curve for the future preset time period.

[0028] Step S35: Confidence assessment and dynamic correction of prediction. Calculate the confidence interval of the prediction result, assess the uncertainty of the prediction, and if a significant deviation between the real-time data and the prediction trend is detected, trigger the online fine-tuning mechanism to quickly correct the prediction values ​​for subsequent time periods using the data from the latest few minutes.

[0029] Optionally, step S4 is implemented in the following manner:

[0030] Step S41: Construct a multi-objective reinforcement learning environment, define the state space and action space, and establish the interaction interface between the reinforcement learning agent and the digital twin simulation environment;

[0031] Step S42: Design a multi-reward function, introduce an energy consumption reward item, and set the negative reward to be proportional to the total energy consumption;

[0032] Step S43: Train multi-objective reinforcement learning by using a large-scale parallel simulation training in a digital twin environment driven by historical data. The agent tries millions of different control combinations in the virtual environment and learns the optimal strategy under different working conditions through trial and error.

[0033] Step S44: Online reasoning and policy generation. Deploy the trained policy model to the edge or cloud decision module, input the current real-time state and prediction information, and the policy network outputs the optimal action vector at the current moment.

[0034] Step S45: Policy security verification and smooth output. Before the action is issued, the generated policy is verified for security by the rule engine, and the control command is smoothed.

[0035] Optionally, step S5 includes the following sub-steps:

[0036] Step S51: The power supply unit is precisely controlled. It receives the angle command in the strategy, drives the high-precision stepper motor to adjust the dual-axis azimuth and elevation angle of the solar panel, automatically adjusts the spectral splitting device, efficiently guides the visible light band to the photovoltaic power generation module, and guides the infrared thermal band to the heat energy collection circuit. It controls the converter and bidirectional inverter, intelligently schedules the photovoltaic power to be stored in the lithium battery or directly supplied to the load, and controls the heat energy storage unit to balance the heat energy supply and demand.

[0037] Step S52: Intelligent temperature control network linkage response, according to the flow demand output by the strategy, adjusts the speed of the microchannel liquid cooling pump in real time, changes the coolant flow rate, adjusts the fan frequency and water valve opening of the indirect evaporative cooling unit, and automatically adjusts the opening of the air supply and return air outlets according to the dynamic temperature field fed back by thermal imaging.

[0038] Step S53: Dynamically adjust service resources. Based on load prediction results, during off-peak periods, gradually shut down redundant radio frequency carriers, symbol channels, or entire sectors according to policy instructions and enter deep sleep mode. Under the premise of ensuring that the user perceived rate is not lower than the preset threshold, fine-tune the output power of the power amplifier and link the baseband processing unit to dynamically adjust the working frequency of the CPU / GPU according to the real-time traffic volume.

[0039] Step S54: Millisecond-level synchronization and closed-loop feedback of multiple systems. Through the edge computing gateway, ensure that the actions of the three subsystems of solar thermal power supply, temperature control and regulation and business resource adjustment are executed synchronously on a millisecond time scale, collect feedback data from each actuator in real time, and transmit the new status data after execution.

[0040] Optionally, the adjustment of the intelligent solar thermal integrated energy supply unit in step S5 includes:

[0041] The solar dual-axis automatic tracking device is controlled to fine-tune the angle to maximize radiation reception. Spectral dispersion technology is used to collect visible light for photovoltaic power generation and collect infrared heat energy to drive an absorption chiller or to keep the computer room warm in winter. The dynamic wall of the phase change building material is controlled to absorb excess heat during the day to reduce the start-up of air conditioning and release heat from natural cold sources at night to solidify the phase change material.

[0042] Optionally, the adjustment of the fluid dynamics intelligent temperature control network in step S5 includes:

[0043] Depending on the base station load, single-phase immersion liquid cooling or cold plate liquid cooling is enabled in the high-power active antenna unit and baseband processing unit to directly transfer heat to the outside. The outdoor air dew point temperature is calculated. When the outdoor air dew point is lower than the set threshold and the temperature is suitable, indirect evaporative cooling is activated. When the outdoor air is dry but the temperature is slightly high, the waste heat of the liquid cooling system is used to drive the dehumidification wheel for dehumidification.

[0044] Optionally, the adjustment of business resources in step S5 includes:

[0045] Before the arrival of low traffic, the carrier power is reduced smoothly in advance to achieve pre-sleep. During low load periods, the antenna tilt and beamforming of neighboring base stations are adjusted in conjunction with the neighboring cells, and the coverage is taken over by the neighboring cells, and the local station is shut down at the symbol level or channel level.

[0046] A multi-dimensional coupled base station energy consumption collaborative control system, employing the multi-dimensional coupled base station energy consumption collaborative control method, includes a multi-dimensional data acquisition and fusion module, a digital twin simulation module, a base station load prediction module, a collaborative control strategy generation module, and an execution control and linkage adjustment module;

[0047] The multi-dimensional data acquisition and fusion module is used to monitor light intensity and spectral distribution through a high-precision irradiance sensor; to perform non-contact scanning of high-power components using an infrared thermal imager, generate temperature field distribution maps and locate hotspots, deploy microscopic prediction probes for service traffic, acquire real-time signaling interaction, user connection count and throughput data, and perform time-stamp alignment and spatial matching of multi-source data such as radiation, temperature and service traffic to eliminate noise and construct a high-fidelity real-time operating status vector for the base station.

[0048] The digital twin simulation module is used to build a real-time digital twin model of a base station in the cloud or at the edge, map geometric models to physical properties, simulate changes in energy flow and fluid dynamics, inject candidate control strategies in parallel, predict changes in power consumption and heat generation within future time windows, output a strategy performance evaluation matrix, and quantify the scores of different strategies.

[0049] The base station load prediction module is used to clean historical call volume data, introduce external covariates such as holidays and traffic flow, construct a multi-dimensional input feature matrix, use a deep learning network with attention mechanism to explore spatiotemporal patterns, predict the base station load in the future within a preset time period, evaluate the prediction confidence, and trigger an online fine-tuning mechanism to correct the prediction value when a significant deviation is detected.

[0050] The collaborative control strategy generation module is used to construct a multi-objective reinforcement learning environment. It uses the lowest total energy consumption, the longest device lifespan, and the user perception rate as multiple reward functions. It trains the intelligent agent through large-scale simulation, combines the simulation results of digital twins and load prediction information, and infers and outputs the optimal action vector at the current moment online. The generated strategy is then subjected to rule engine security verification and smoothing processing.

[0051] The execution control and linkage adjustment module is responsible for translating strategies into specific physical actions and collecting new state data after execution to form a closed-loop feedback.

[0052] The beneficial effects of this invention are:

[0053] 1. In this invention, through digital twin pre-playing and reinforcement learning, it no longer relies on a single rule, but takes the lowest total energy consumption, the longest equipment life and user perception as multiple objectives to achieve global optimal control, significantly reducing base station energy consumption. Combined with real-time data and future load prediction, it changes "passive response" to "active adjustment", avoiding energy waste or overheating risks caused by adjustment lag. It realizes the linkage of energy supply, temperature control and service resources, and solves the problem of low efficiency caused by independent control of each subsystem of traditional base stations.

[0054] 2. In this invention, external covariates such as holidays and traffic flow are introduced, and a deep learning model with attention mechanism is adopted, which has online fine-tuning capability. It takes into account the nonlinear impact of external environment (such as traffic flow) on traffic volume, which is more accurate than traditional time series prediction. It has online fine-tuning and confidence assessment mechanism, which can quickly correct the prediction when there are sudden changes in traffic, ensuring the real-time adaptability of the control strategy. The high-resolution load curve helps to plan the timing of resource hibernation and wake-up more finely, maximizing energy saving space.

[0055] 3. In this invention, multiple reward functions are used to balance energy saving with equipment lifespan and user speed, avoiding sacrificing network quality or accelerating equipment aging for energy saving. Rule engine verification and smoothing prevent aggressive control commands from causing system oscillations or service interruptions, ensuring the safe operation of the existing network. The intelligent agent can continuously adapt to new base station environments and service models through trial and error learning, and continuously optimize control strategies. Attached Figure Description

[0056] Figure 1 This is a schematic diagram of a method flow of the present invention.

[0057] Figure 2 This is a schematic diagram of step S1 of the present invention.

[0058] Figure 3 This is a schematic diagram of step S2 of the present invention.

[0059] Figure 4 This is a schematic diagram of step S3 of the present invention.

[0060] Figure 5 This is a schematic diagram of step S4 of the present invention.

[0061] Figure 6 This is a schematic diagram of step S5 of the present invention. Detailed Implementation

[0062] To more clearly illustrate the technical solutions in the embodiments of the invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0063] like Figures 1 to 6 As shown, a multi-dimensional coupled base station energy consumption collaborative control method includes the following:

[0064] Step S1: Collect base station data by using multi-dimensional sensor fusion technology to collect real-time operating data of the base station;

[0065] Specifically, it will be implemented in the following ways:

[0066] S1.1 Real-time irradiance monitoring deployment and data acquisition: A high-precision irradiance sensor is integrated into the base station's dual-axis automatic solar tracking device to monitor the current ambient light intensity and spectral distribution in real time. The collected radiation data is transmitted to the control unit in real time to fine-tune the angle of the solar panel to maximize radiation reception and to serve as input parameters for spectral splitting technology (visible light power generation, infrared thermal energy collection).

[0067] S1.2 Precise positioning of equipment junction temperature using infrared thermal imaging: Infrared thermal imagers are deployed near high-power core components (such as AAU active antenna units and BBU baseband processing units) or at key liquid-cooled / air-cooled nodes to perform non-contact scanning of the equipment surface and generate real-time temperature field distribution maps.

[0068] It accurately locates local hot spots and identifies chips or modules with abnormal temperature rise, using pixel-level junction temperature data as the direct control basis for the fluid dynamics intelligent temperature control network (liquid cooling pump frequency, air valve opening).

[0069] S1.3 Service Traffic Micro-Prediction Probe Data Collection: Deploy a service traffic micro-prediction probe to capture the current signaling interaction data, user connection count, and instantaneous throughput of the base station in real time. Combine this with historical call volume database, holiday effect markers, and surrounding traffic and pedestrian flow data (if available) to form a multi-dimensional input feature vector. Send the above real-time data stream to the AI ​​spatiotemporal load prediction module to provide basic data for subsequent deep learning algorithms, so as to achieve micro-prediction of the base station load in the next 15-30 minutes.

[0070] S1.4 Multi-source data fusion and digital twin mapping: The collected irradiance data, infrared thermal imaging temperature field data, and service traffic micro data are time-stamped and spatially matched. Using a multi-dimensional sensor fusion algorithm, noise errors from a single sensor are eliminated, and a high-fidelity real-time operating state vector of the base station is constructed. This state vector is then mapped to a "real-time digital twin model of the base station" in the cloud or at the edge, completing the real-time synchronization of the physical base station to the digital space. This provides accurate initial boundary conditions for subsequent simulation of energy consumption and temperature field changes under different control strategies.

[0071] Step S2: Simulate energy consumption changes. Based on the real-time operating data, construct a real-time digital twin model of the base station in the cloud or at the edge to simulate energy consumption and temperature field changes under different control strategies.

[0072] Specifically, it will be implemented in the following ways:

[0073] S2.1 Multidimensional data mapping and model initialization: Receive real-time operating data from the fusion in step S1, including irradiance, equipment junction temperature distribution, and service traffic characteristics. On the cloud server or edge computing node, call the pre-set high-fidelity geometric model and physical attribute library of the base station to map the real-time data to the corresponding node of the digital twin, initialize the digital twin state at the current moment, and ensure that the virtual model is completely consistent with the geometric structure, material properties (such as phase change wall state) and operating parameters of the physical base station.

[0074] S2.2 Construct a multiphysics coupling simulation environment and activate the "optical-thermal-electrical-information" multiphysics coupling solver in the digital twin model.

[0075] Establish an energy flow model to simulate the solar spectral splitting efficiency, photovoltaic power generation output, heat absorption and release processes of phase change materials, and battery charging and discharging characteristics.

[0076] Establish a thermal fluid model: Based on the principles of computational fluid dynamics (CFD), simulate the airflow organization, heat transfer efficiency of liquid cooling circuit, and heat exchange process of indirect evaporative cooling in the computer room.

[0077] Establish a service load model: simulate the power consumption curves and heat generation changes of AAU / BBU under different service volumes.

[0078] S2.3 Generate and inject a set of candidate control strategies. Based on the current state, the decision layer generates a set of candidate collaborative control strategy combinations. For example, strategy A is "maximize photovoltaic tracking + liquid cooling full speed + carrier sleep"; strategy B is "fixed photovoltaic angle + free cooling + full power operation", etc. These candidate strategies are used as control variables and injected into the actuator interface of the digital twin model respectively.

[0079] S2.4 Parallel extrapolation and multi-dimensional index simulation: Multiple simulation instances are run in parallel in the digital space, each instance corresponding to a candidate control strategy, to extrapolate the system evolution within a preset time window (e.g., the next 15-30 minutes).

[0080] Energy consumption simulation: Calculate the total energy consumption (grid power consumption - self-generated revenue) under each strategy.

[0081] Temperature field simulation: Predicting the junction temperature variation trend of key chips, the distribution of hot spots in the computer room, and the phase transition progress of phase change materials.

[0082] Performance simulation: Estimating user-perceived speed and network coverage quality under different resource scheduling strategies.

[0083] S2.5 outputs the strategy evaluation matrix and collects the output results of all parallel simulation instances.

[0084] Construct a "strategy-performance" evaluation matrix to quantify the scores of each strategy in three dimensions: "lowest total energy consumption", "longest device life (minimum temperature fluctuation)" and "user perceived rate meets the standard". Transmit the evaluation matrix to the multi-objective reinforcement learning algorithm module in step S4 as the basis for selecting the optimal strategy.

[0085] Step S3: Predict future base station load. Analyze historical call volume, holiday effects, and surrounding traffic and pedestrian flow data using deep learning algorithms to predict the base station load within a preset time period.

[0086] Implemented in the following ways:

[0087] S3.1 Multi-source heterogeneous data cleaning and feature engineering

[0088] Historical traffic volume processing: Extract time series data of base station traffic volume over the past few months or even years (including RRC connection count, PRB utilization, and throughput), and perform missing value imputation and outlier smoothing.

[0089] Holiday effect labeling: Construct calendar feature vectors to clearly mark the dates of weekdays, weekends, statutory holidays and special major events, and introduce time offset features of "before / after the holiday".

[0090] Surrounding traffic and pedestrian flow integration: Traffic flow data (such as vehicle speed and density) and pedestrian flow heat map data within the coverage area of ​​the access base station are transformed into external covariate features consistent with the time granularity of call volume.

[0091] Feature alignment: The three types of data mentioned above are strictly aligned on the time axis to construct a multi-dimensional input feature matrix that includes internal load history and external environmental factors.

[0092] S3.2 Construct a deep learning prediction model architecture, using a deep learning network architecture suitable for time series prediction (such as a variant of LSTM long short-term memory network, GRU gated recurrent unit, or Transformer architecture).

[0093] Design a multi-branch input layer: receive historical call sequence branches, calendar feature branches, and external pedestrian and traffic branches respectively.

[0094] Introduce an attention mechanism: This enables the model to automatically focus on historical moments that have the greatest impact on the current load (such as the same time last week) and key external events (such as sudden peaks in pedestrian traffic).

[0095] S3.3 Model Training and Spatiotemporal Pattern Mining: The deep learning model is trained offline using historical datasets. Backpropagation is used to optimize network weights, allowing the model to automatically learn and mine complex nonlinear relationships within the data.

[0096] Periodic patterns: Identify daily morning and evening peak hours, and weekly weekday and weekend fluctuation patterns.

[0097] Event-driven patterns: Learn about the impact of sudden traffic surges caused by holidays, large-scale events, or extreme weather.

[0098] Spatiotemporal correlation patterns: Capture the lagging or leading correlation between changes in surrounding traffic and pedestrian flow and base station traffic volume (e.g., traffic volume increases 15 minutes after the arrival of morning rush hour traffic).

[0099] S3.4 Rolling micro-load prediction execution inputs the latest real-time collected data into the trained model, performs rolling prediction, and outputs high-resolution load curves for a future preset time period (mentioned in the document as 15-30 minutes).

[0100] The forecast results specifically include: the expected number of users, service throughput requirements, and corresponding theoretical power consumption baseline for each time slice in the future.

[0101] S3.5 Prediction Confidence Assessment and Dynamic Correction: Calculate the confidence interval of the prediction results, assess the uncertainty of the prediction, and if a significant deviation is detected between the real-time data and the prediction trend (such as a sudden event causing a surge in traffic), trigger the online fine-tuning mechanism to quickly correct the prediction values ​​for subsequent time periods using the data from the latest few minutes, ensuring the accuracy of "forward-looking resource scheduling". Output the final determined load prediction result to step S4 as the key basis for the multi-objective reinforcement learning algorithm to formulate "pre-sleep" or "power smoothing" strategies.

[0102] Step S4: Generate a collaborative control strategy. Based on a multi-objective reinforcement learning algorithm, the optimal collaborative control strategy is generated by using the lowest total energy consumption, the longest equipment lifespan, and the user perception rate not lower than a preset threshold as multiple reward functions, combined with the simulation results of the digital twin model and the base station load prediction.

[0103] Implemented in the following ways:

[0104] S4.1 Constructing a multi-objective reinforcement learning environment:

[0105] State space definition: The current state of the digital twin model in step S2 (radiation, indoor and outdoor temperature difference, phase change material saturation, equipment junction temperature distribution) is spliced ​​with the load prediction result in step S3 (flow trend in the next 15-30 minutes) to form a high-dimensional state vector.

[0106] Action space definition: Defines a set of continuous or discrete control actions, including: solar panel tracking angle, liquid cooling pump speed / valve opening, evaporative cooling fan frequency, AAU carrier shutdown ratio, transmit power adjustment step size, etc.

[0107] Environment Interaction Interface: Establish an interaction interface between the reinforcement learning agent and the digital twin simulation environment constructed in step S2, so that the actions issued by the agent can drive the evolution of the simulation model in real time.

[0108] S4.2 Design a multi-reward function:

[0109] Energy consumption bonus: Set negative bonuses to be proportional to total energy consumption.

[0110] Energy bonus value:

[0111] ,

[0112] In the formula, Power consumed by mains electricity Photovoltaic power generation capacity, For grid electricity price, This is the green energy benefit coefficient.

[0113] Equipment lifespan bonus value:

[0114] ,

[0115] In the formula, α is the weighting coefficient for temperature fluctuation rate. β is the absolute value of the junction temperature change, and β is the over-temperature penalty weighting coefficient.

[0116] This refers to the temperature exceedance range.

[0117] The user perceived rate reward value is:

[0118] ,

[0119] In the formula, For the actual rate perceived by the user, γ is the preset rate threshold, and γ is the QoS penalty weight coefficient.

[0120] S4.3 Train a multi-objective reinforcement learning agent by selecting a multi-objective deep reinforcement learning algorithm suitable for continuous control, such as a variant of multi-objective PPO, SAC, or DDPG.

[0121] Offline pre-training: Large-scale parallel simulation training is performed using a digital twin environment driven by historical data. The agent tries millions of different control combinations in the virtual environment, learning the optimal strategy under different operating conditions (such as extreme heat or sudden traffic peaks) through trial and error.

[0122] Policy network optimization: Update the parameters of the policy network (Actor) and value network (Critic) to maximize the cumulative expected reward, enabling the agent to learn to find the optimal balance between "energy saving", "machine protection" and "experience preservation".

[0123] S4.4 Online Inference and Policy Generation: The trained policy model is deployed to the edge or cloud decision module. The current real-time state and prediction information are input, and the policy network outputs the optimal action vector at the current moment, i.e. the optimal cooperative control policy. Specifically, it includes: the optimal tracking angle and spectral splitting mode of the photothermal unit, the liquid cooling / air cooling ratio and fluid parameters of the temperature control network, and the carrier sleep timing and power adjustment range of the service resources.

[0124] S4.5 Policy Security Verification and Smooth Output: Before the action is issued, the generated policy is verified for security through the rule engine. For example, it prevents the liquid cooling pump from suddenly stopping and causing overheating, and prevents the carrier from being completely shut down at an instant and causing dropped calls.

[0125] The control commands are smoothed (e.g., a maximum rate of change limit is set) to avoid mechanical wear or power grid impact caused by frequent and violent movements of the actuators. Finally, the optimal strategy that has been verified and passed the smoothing process is sent to the execution layer in step S5 for physical execution.

[0126] Step S5: Execute the control strategy, and execute the optimal collaborative control strategy to coordinate and adjust the intelligent solar thermal integrated power supply unit, the fluid dynamics intelligent temperature control network, and business resources.

[0127] Specifically, it will be implemented in the following ways:

[0128] Precise control of the S5.1 intelligent solar thermal integrated energy supply unit

[0129] Dual-axis tracking angle execution: Receives angle commands from the strategy and drives a high-precision stepper motor to adjust the dual-axis azimuth and elevation angles of the solar panel, ensuring maximum light energy capture under real-time irradiance.

[0130] Dynamic switching of spectral splitting: Based on the spectral utilization mode set by the strategy, the spectral splitting device, such as a dichroic filter or a nano-coating switching device, is automatically adjusted to efficiently guide the visible light band to the photovoltaic power generation module, while guiding the infrared thermal band to the heat energy collection circuit.

[0131] Thermoelectric storage management: Based on the charge and discharge commands in the strategy, control the DC / DC converter and bidirectional inverter to intelligently schedule photovoltaic power to be stored in the lithium battery or directly supplied to the load; at the same time, control the heat absorption / heat release valves of the thermal energy storage unit to balance the supply and demand of thermal energy.

[0132] S5.2 Interlocking Response of Intelligent Temperature Control Network for Fluid Dynamics

[0133] Liquid cooling circuit frequency conversion regulation: Based on the flow requirements of the strategy output, the speed of the microchannel liquid cooling pump is adjusted in real time (PWM control) to change the coolant flow rate, accurately remove the heat generated by the AAU / BBU core chip, and eliminate local hot spots.

[0134] Air-cooled / evaporative cooling synergy: Adjust the fan frequency and water valve opening of the indirect evaporative cooling unit. When the outdoor dry-bulb temperature is low, the free cooling mode is activated first; during high-temperature periods, mechanical refrigeration and evaporative cooling are mixed in a strategic ratio to maintain the set temperature of the computer room with minimal energy consumption.

[0135] Airflow organization optimization: If the base station is equipped with adjustable air valves or intelligent air deflectors, the opening of the air supply and return air vents will be automatically adjusted according to the dynamic temperature field fed back by thermal imaging, optimizing the airflow path in the equipment room and avoiding short-circuiting of hot and cold air.

[0136] S5.3 Dynamic Hibernation and Power Smoothing of Service Resources

[0137] Smart Carrier / Channel Sleep: Based on load prediction results, during off-peak periods (such as late at night or when there is little foot traffic), redundant radio frequency carriers, symbol channels, or entire sectors are gradually shut down according to policy instructions, entering a deep sleep mode to reduce static power consumption.

[0138] Dynamic transmission power adjustment: While ensuring that the user's perceived rate is not lower than a preset threshold, the output power of the power amplifier (PA) is finely adjusted. In areas of overlapping coverage or low load, the transmission power is appropriately reduced to reduce ineffective radiation and energy consumption.

[0139] Elastic scaling of computing resources: The baseband processing unit (BBU) is linked to dynamically adjust the operating frequency of the CPU / GPU or the number of cores to wake up / sleep based on real-time traffic volume, so as to achieve on-demand matching of computing power and energy consumption.

[0140] S5.4 Millisecond-level Synchronization and Closed-Loop Feedback for Multiple Systems

[0141] Synchronous execution: Through the edge computing gateway, the actions of the three subsystems of solar thermal power supply, temperature control and regulation and business resource adjustment are ensured to be executed synchronously on a millisecond time scale, avoiding instantaneous overload or temperature fluctuation caused by asynchronous actions.

[0142] Real-time feedback of execution status: While performing actions, feedback data from each actuator (such as actual pump speed, actual number of carriers shut down, and actual power generation) is collected in real time.

[0143] Closed-loop correction mechanism: The new state data (new temperature field, new energy consumption value) after execution is immediately transmitted back to the perception layer of step S1 and the digital twin model of step S2, forming a complete closed loop of "perception-decision-execution-re-perception", providing the latest basis for the next round of strategy optimization.

[0144] A multi-dimensional coupled base station energy consumption collaborative control system, employing the multi-dimensional coupled base station energy consumption collaborative control method, includes a multi-dimensional data acquisition and fusion module, a digital twin simulation module, a base station load prediction module, a collaborative control strategy generation module, and an execution control and linkage adjustment module;

[0145] The multi-dimensional data acquisition and fusion module is used to monitor light intensity and spectral distribution through a high-precision irradiance sensor; to perform non-contact scanning of high-power components using an infrared thermal imager, generate temperature field distribution maps and locate hotspots, deploy microscopic prediction probes for service traffic, acquire real-time signaling interaction, user connection count and throughput data, and perform time-stamp alignment and spatial matching of multi-source data such as radiation, temperature and service traffic to eliminate noise and construct a high-fidelity real-time operating status vector for the base station.

[0146] The digital twin simulation module is used to build a real-time digital twin model of a base station in the cloud or at the edge, map geometric models to physical properties, simulate changes in energy flow and fluid dynamics, inject candidate control strategies in parallel, predict changes in power consumption and heat generation within future time windows, output a strategy performance evaluation matrix, and quantify the scores of different strategies.

[0147] The base station load prediction module is used to clean historical call volume data, introduce external covariates such as holidays and traffic flow, construct a multi-dimensional input feature matrix, use a deep learning network with attention mechanism to explore spatiotemporal patterns, predict the base station load in the future within a preset time period, evaluate the prediction confidence, and trigger an online fine-tuning mechanism to correct the prediction value when a significant deviation is detected.

[0148] The collaborative control strategy generation module is used to construct a multi-objective reinforcement learning environment. It uses the lowest total energy consumption, the longest device lifespan, and the user perception rate as multiple reward functions. It trains the intelligent agent through large-scale simulation, combines the simulation results of digital twins and load prediction information, and infers and outputs the optimal action vector at the current moment online. The generated strategy is then subjected to rule engine security verification and smoothing processing.

[0149] The execution control and linkage adjustment module is responsible for translating strategies into specific physical actions and collecting new state data after execution to form a closed-loop feedback.

[0150] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions or improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-dimensional coupled base station energy consumption collaborative control method, characterized in that, Includes the following steps: Step S1: Collect base station data by using multi-dimensional sensor fusion technology to collect real-time operating data of the base station; Step S2: Simulate energy consumption changes. Based on the real-time operating data, construct a real-time digital twin model of the base station in the cloud or at the edge to simulate energy consumption and temperature field changes under different control strategies. Step S3: Predict future base station load. Analyze historical base station load using deep learning algorithms to predict base station load within a preset time period. Step S4: Generate a cooperative control strategy. Based on a multi-objective reinforcement learning algorithm, set multiple reward functions, and combine the simulation results of the digital twin model and the base station load prediction to generate the optimal cooperative control strategy. Step S5: Execute the control strategy. Execute the optimal collaborative control strategy to coordinate and adjust the intelligent integrated solar and thermal power supply unit, the fluid dynamics intelligent temperature control network, and business resources within the digital twin model, sharing energy consumption adjustment parameters.

2. The multi-dimensional coupled base station energy consumption coordinated control method according to claim 1, characterized in that, Step S1 includes the following sub-steps: Step S11: Deployment and data acquisition of real-time radiation irradiance monitoring. A high-precision radiation irradiance sensor is integrated into the base station's solar dual-axis automatic tracking device to monitor the light intensity and spectral distribution of the current environment in real time, and the collected radiation data is transmitted to the control unit in real time. Step S12: Precisely locate the junction temperature of the device using infrared thermal imaging. Deploy an infrared thermal imager at the key nodes of high-power core components to perform non-contact scanning on the device surface, generate a real-time temperature field distribution map, accurately locate local hot spots, and identify chips or modules with abnormal temperature rise. Step S13: Data collection of micro-prediction probe for service traffic. Deploy the micro-prediction probe for service traffic to capture the current signaling interaction data, number of user connections and instantaneous throughput of the base station in real time. Combine the historical call volume database, holiday effect markers and surrounding traffic and pedestrian flow data to form a multi-dimensional input feature vector. Step S14: Multi-source data fusion and digital twin mapping. The collected irradiance data, infrared thermal imaging temperature field data, and service traffic micro data are time-stamp aligned and spatial coordinate matched. Using a multi-dimensional sensor fusion algorithm, noise errors from a single sensor are eliminated, and a high-fidelity real-time operating status vector of the base station is constructed.

3. The multi-dimensional coupled base station energy consumption coordinated control method according to claim 2, characterized in that, Step S2 includes the following sub-steps: Step S21: Multidimensional data mapping and model initialization. Receive the fused real-time running data, and on the cloud server or edge computing node, call the pre-set base station high-fidelity geometric model and physical attribute library to map the real-time data to the corresponding node of the digital twin and initialize the digital twin state at the current moment. Step S22: Construct a multiphysics coupled simulation environment, activate the multiphysics coupled solver in the digital twin model, simulate the solar spectral splitting efficiency, photovoltaic power generation output, heat absorption and release process of phase change materials and battery charging and discharging characteristics through the energy flow model, and simulate the airflow organization, heat transfer efficiency of liquid cooling circuit and heat exchange process of indirect evaporative cooling in the computer room based on the principle of computational fluid dynamics. Step S23: Generate and inject a set of candidate control strategies. Based on the current state, the decision layer generates a set of candidate collaborative control strategies and injects them as control variables into the execution mechanism interface of the digital twin model. Step S24: Parallel extrapolation and multi-dimensional index simulation. Multiple simulation instances are run in parallel in the digital space. Each instance corresponds to a candidate control strategy. The system evolution within a preset time window is extrapolated, and the power consumption curve and heat generation changes under different business volumes are simulated. Step S25: Output the policy evaluation matrix, collect the output results of all parallel simulation instances, construct the policy effectiveness evaluation matrix, quantify the score of each policy, and transmit the evaluation matrix to the multi-objective reinforcement learning algorithm module as the basis for selecting the optimal policy.

4. The multi-dimensional coupled base station energy consumption coordinated control method according to claim 1, characterized in that, Step S3 includes the following sub-steps: Step S31: Multi-source heterogeneous data cleaning and feature engineering. Extract time series data of base station call volume from the past few months or even years, perform missing value imputation and outlier smoothing, construct calendar feature vectors, introduce time offset features before and after holidays, access traffic flow data and population heat map data in the base station coverage area, transform them into external covariate features with the same time granularity as call volume, strictly align them on the time axis, and construct a multi-dimensional input feature matrix. Step S32: Construct a deep learning prediction model architecture. Adopt a deep learning network architecture and design a multi-branch input layer: receive historical call sequence branches, calendar feature branches and external traffic flow branches respectively. Introduce an attention mechanism to focus on the historical moments and key external events that have the greatest impact on the current load. Step S33: Model training and spatiotemporal pattern mining. The deep learning model is trained offline using historical datasets. The network weights are optimized through backpropagation algorithm to learn and mine nonlinear relationships in the data. Step S34: Load prediction execution. The latest data collected in real time is input into the trained model to perform rolling prediction and output a high-resolution load curve for the future preset time period. Step S35: Confidence assessment and dynamic correction of prediction. Calculate the confidence interval of the prediction result, assess the uncertainty of the prediction, and if a significant deviation between the real-time data and the prediction trend is detected, trigger the online fine-tuning mechanism to quickly correct the prediction values ​​for subsequent time periods using the data from the latest few minutes.

5. The multi-dimensional coupled base station energy consumption coordinated control method according to claim 1, characterized in that, Step S4 is implemented in the following manner: Step S41: Construct a multi-objective reinforcement learning environment, define the state space and action space, and establish the interaction interface between the reinforcement learning agent and the digital twin simulation environment; Step S42: Design a multi-reward function, introduce an energy consumption reward item, and set the negative reward to be proportional to the total energy consumption; Step S43: Train multi-objective reinforcement learning by using a large-scale parallel simulation training in a digital twin environment driven by historical data. The agent tries millions of different control combinations in the virtual environment and learns the optimal strategy under different working conditions through trial and error. Step S44: Online reasoning and policy generation. Deploy the trained policy model to the edge or cloud decision module, input the current real-time state and prediction information, and the policy network outputs the optimal action vector at the current moment. Step S45: Policy security verification and smooth output. Before the action is issued, the generated policy is verified for security by the rule engine, and the control command is smoothed.

6. The multi-dimensional coupled base station energy consumption coordinated control method according to claim 1, characterized in that, Step S5 includes the following sub-steps: Step S51: The power supply unit is precisely controlled. It receives the angle command in the strategy, drives the high-precision stepper motor to adjust the dual-axis azimuth and elevation angle of the solar panel, automatically adjusts the spectral splitting device, efficiently guides the visible light band to the photovoltaic power generation module, and guides the infrared thermal band to the heat energy collection circuit. It controls the converter and bidirectional inverter, intelligently schedules the photovoltaic power to be stored in the lithium battery or directly supplied to the load, and controls the heat energy storage unit to balance the heat energy supply and demand. Step S52: Intelligent temperature control network linkage response, according to the flow demand output by the strategy, adjusts the speed of the microchannel liquid cooling pump in real time, changes the coolant flow rate, adjusts the fan frequency and water valve opening of the indirect evaporative cooling unit, and automatically adjusts the opening of the air supply and return air outlets according to the dynamic temperature field fed back by thermal imaging. Step S53: Dynamically adjust service resources. Based on load prediction results, during off-peak periods, gradually shut down redundant radio frequency carriers, symbol channels, or entire sectors according to policy instructions and enter deep sleep mode. Under the premise of ensuring that the user perceived rate is not lower than the preset threshold, fine-tune the output power of the power amplifier and link the baseband processing unit to dynamically adjust the working frequency of the CPU / GPU according to the real-time traffic volume. Step S54: Millisecond-level synchronization and closed-loop feedback of multiple systems. Through the edge computing gateway, ensure that the actions of the three subsystems of solar thermal power supply, temperature control and regulation and service resource adjustment are executed synchronously on the millisecond time scale. Real-time collection of feedback data from each actuator, and transmission of new status data after execution to obtain energy consumption regulation parameters, which are then synchronized to the corresponding equipment in the base station.

7. The multi-dimensional coupled base station energy consumption coordinated control method according to claim 1, characterized in that, The adjustment of the intelligent solar thermal integrated power supply unit in step S5 includes: The solar dual-axis automatic tracking device is controlled to fine-tune the angle to maximize radiation reception. Spectral dispersion technology is used to collect visible light for photovoltaic power generation and collect infrared heat energy to drive an absorption chiller or to keep the computer room warm in winter. The dynamic wall of the phase change building material is controlled to absorb excess heat during the day to reduce the start-up of air conditioning and release heat from natural cold sources at night to solidify the phase change material.

8. The multi-dimensional coupled base station energy consumption coordinated control method according to claim 1, characterized in that, The adjustment of the fluid dynamics intelligent temperature control network in step S5 includes: Depending on the base station load, single-phase immersion liquid cooling or cold plate liquid cooling is enabled in the high-power active antenna unit and baseband processing unit to directly transfer heat to the outside. The outdoor air dew point temperature is calculated. When the outdoor air dew point is lower than the set threshold and the temperature is suitable, indirect evaporative cooling is activated. When the outdoor air is dry but the temperature is slightly high, the waste heat of the liquid cooling system is used to drive the dehumidification wheel for dehumidification.

9. The multi-dimensional coupled base station energy consumption coordinated control method according to claim 1, characterized in that, The adjustment of business resources in step S5 includes: Before the arrival of low traffic, the carrier power is reduced smoothly in advance to achieve pre-sleep. During low load periods, the antenna tilt and beamforming of neighboring base stations are adjusted in conjunction with the neighboring cells, and the coverage is taken over by the neighboring cells, and the local station is shut down at the symbol level or channel level.

10. A multi-dimensional coupled base station energy consumption collaborative control system, employing the multi-dimensional coupled base station energy consumption collaborative control method as described in any one of claims 1-9, characterized in that, It includes a multi-dimensional data acquisition and fusion module, a digital twin simulation module, a base station load prediction module, a collaborative control strategy generation module, and an execution control and linkage adjustment module; The multi-dimensional data acquisition and fusion module is used to monitor light intensity and spectral distribution through a high-precision irradiance sensor; to perform non-contact scanning of high-power components using an infrared thermal imager, generate temperature field distribution maps and locate hotspots, deploy microscopic prediction probes for service traffic, acquire real-time signaling interaction, user connection count and throughput data, and perform time-stamp alignment and spatial matching of multi-source data such as radiation, temperature and service traffic to eliminate noise and construct a high-fidelity real-time operating status vector for the base station. The digital twin simulation module is used to build a real-time digital twin model of a base station in the cloud or at the edge, map geometric models to physical properties, simulate changes in energy flow and fluid dynamics, inject candidate control strategies in parallel, predict changes in power consumption and heat generation within future time windows, output a strategy performance evaluation matrix, and quantify the scores of different strategies. The base station load prediction module is used to clean historical call volume data, introduce external covariates such as holidays and traffic flow, construct a multi-dimensional input feature matrix, use a deep learning network with attention mechanism to explore spatiotemporal patterns, predict the base station load in the future within a preset time period, evaluate the prediction confidence, and trigger an online fine-tuning mechanism to correct the prediction value when a significant deviation is detected. The collaborative control strategy generation module is used to construct a multi-objective reinforcement learning environment. It uses the lowest total energy consumption, the longest device lifespan, and the user perception rate as multiple reward functions. It trains the intelligent agent through large-scale simulation, combines the simulation results of digital twins and load prediction information, and infers and outputs the optimal action vector at the current moment online. The generated strategy is then subjected to rule engine security verification and smoothing processing. The execution control and linkage adjustment module is responsible for translating strategies into specific physical actions and collecting new state data after execution to form a closed-loop feedback.

Images