Power supply energy-saving system and method based on load prediction intelligent dormancy strategy

Abstract

The invention relates to the technical field of network load prediction, and particularly discloses a power supply energy saving system and method based on a load prediction intelligent dormancy strategy, and the method comprises the steps: collecting base station service load and power supply power data, and employing a neural network to accurately fit the nonlinear relation between the two for adjusting the power; constructing a multi-mode service load prediction model, generating a power supply power adjustment instruction in advance, and adjusting the power in time when the load prediction deviation exceeds a threshold value; frequent power fluctuation is avoided through a dynamic threshold mechanism based on hysteresis comparison; a sleep and wake-up strategy and a buffer mechanism are established to ensure that the state of the power supply is switched only when the load changes continuously, so that the switching accuracy and stability are improved; self-adaptive sleep duration and progressive wake-up adjustment are introduced, so that smooth transition and intelligent energy conservation of power supply operation are realized; the method can improve the energy efficiency of the power supply and reduce the energy consumption.

Description

Technical Field

[0001] The present application relates to the technical field of network load prediction, and specifically to a power energy-saving system and method based on a load prediction intelligent sleep strategy. Background Art

[0002] Existing 5G base station power sleep technology usually adopts a simple threshold judgment method, that is, when the business load is detected to be lower than the preset threshold, the power is triggered to sleep, and when the load exceeds the wake-up threshold, the power is reactivated.

[0003] However, due to the significant short-term fluctuations in 5G network communication traffic load, especially within critical load ranges, traffic volume can fluctuate frequently around the sleep threshold, causing the power supply system to repeatedly switch between sleep and wakeup within a short period of time. This frequent state switching has negative consequences. On the one hand, the frequent starting and stopping of power supply equipment accelerates hardware aging, shortens equipment lifespan, and increases maintenance costs. On the other hand, because the power supply requires a certain amount of time to resume normal operation from sleep mode, frequent switching may result in an inability to respond promptly to sudden service calls, affecting network service quality.

[0004] The root cause of this problem is that existing technologies lack the ability to deeply predict the dynamic changes in 5G network traffic load. Decisions rely solely on instantaneous real-time load values, failing to distinguish between short-term load fluctuations and long-term trends. Existing solutions also lack effective buffering mechanisms to filter out the impact of short-term load fluctuations, resulting in overly sensitive power hibernation decisions. Therefore, a method is urgently needed that can accurately predict traffic load trends, effectively filter out short-term fluctuations, and achieve stable and reliable power hibernation control to achieve efficient energy conservation in 5G base stations.

[0005] In view of this, the present application proposes a power energy saving system and method based on a load prediction intelligent sleep strategy. Summary of the Invention

[0006] To achieve the above objectives, the present application provides a power energy saving system and method based on a load prediction intelligent sleep strategy. The specific technical solutions are as follows:

[0007] The power saving method based on load prediction intelligent sleep strategy includes:

[0008] Collect business load and power data, and use neural networks to fit the nonlinear relationship between business load and power;

[0009] A business load prediction model is built based on a neural network model to predict future business loads. A power supply adjustment timing instruction is generated based on the predicted business load. When the difference between the actual business load and the predicted business load exceeds a load difference threshold, the power supply adjustment mechanism is triggered.

[0010] Build a power adjustment mechanism that dynamically sets upper and lower comparison thresholds for adjusting power based on the hysteresis comparison principle, forming a hysteresis control interval. Only when the predicted load continuously crosses the hysteresis control interval and the change trend is consistent, will the power adjustment instruction be executed.

[0011] Create power sleep and wake-up policies, which will enable the system to sleep when the service load is continuously below the sleep threshold and wake up when the service load is above the wake-up threshold.

[0012] Create a sleep buffer mechanism, set an adaptive sleep buffer time, and determine whether to enter the sleep state based on the buffer time;

[0013] Build an adaptive sleep duration mechanism to dynamically adjust the power sleep duration;

[0014] Build a progressive power adjustment mechanism for power wake-up, and adjust the power to the target power of the target business load when waking up the power.

[0015] Preferably, the collected service load and power data are statistically eliminated for outliers and interpolated for missing values, and normalized according to the base station design capacity to eliminate dimensional differences between different data types;

[0016] A three-layer feedforward neural network is used to fit the nonlinear mapping relationship between business load and power supply power. The input corresponds to the parameters of each dimension of business load, and the output generates normalized power supply power. The hidden layer uses a nonlinear activation function.

[0017] Based on the trained three-layer feedforward neural network, the optimal power supply power corresponding to the business load is calculated, and a mapping table between the business load and the optimal power supply power is generated for adjusting the optimal power supply power.

[0018] Preferably, three independent feature extraction channels are constructed to process multimodal features of time series features, periodic features and emergency event features respectively, and a multimodal business load prediction model is constructed based on the multimodal features to predict future business loads;

[0019] The multimodal traffic load prediction model is constructed based on a long short-term memory network (LSTM) model. The LSTM model is used to fuse multimodal features extracted from three feature channels: time series, periodicity, and emergencies, and learn the dynamic change pattern of traffic load in the time dimension.

[0020] The LSTM model is used to predict future business loads. Based on the predicted business load sequence and the established power mapping table, a power supply power adjustment timing instruction set is generated to adjust the power supply power.

[0021] Preferably, the difference between the actual business load and the predicted business load is compared in real time;

[0022] A load difference threshold is set. When the load deviation of multiple consecutive sampling points is greater than the load difference threshold, the prediction is judged to have a significant deviation. At the same time, the deviation trend indicator is calculated. If the deviation trend indicator diverges, the power adjustment mechanism is triggered.

[0023] Preferably, based on the current power supply power and business load fluctuations, upper and lower thresholds for power supply power regulation are dynamically set to construct a hysteresis control interval;

[0024] The power supply adjustment mechanism is triggered only when the predicted power continuously crosses the hysteresis control interval and the load change direction is consistent with the crossing direction.

[0025] Preferably, an adaptive power sleep threshold is established based on the statistical characteristics of historical business load data, and the power sleep threshold is dynamically adjusted in combination with time period characteristics and date type;

[0026] Build a power wake-up threshold and set an emergency wake-up mechanism. When the instantaneous load exceeds the threshold, the power wake-up is triggered immediately.

[0027] Preferably, the sleep buffer time is determined by dynamic calculation based on historical sleep success data, including the success rate and the business load rebound time when failure occurs;

[0028] Implement a sleep buffer state monitoring mechanism. After the power supply receives a sleep command, it enters the sleep buffer state and maintains the current power output unchanged.

[0029] Monitor the business load in real time. When the business load meets the preset load rebound judgment conditions, the power supply exits the sleep buffer state and cancels the sleep operation.

[0030] Preferably, a sleep confirmation and state transition mechanism is established. When the sleep buffer time ends and no load rebound occurs, the sleep execution is confirmed. Based on the comparison result of the average business load during the buffer period and the dynamic sleep threshold, the power supply is confirmed to enter the sleep state, and the number of sleep successes is updated to optimize subsequent decisions.

[0031] Build an adaptive sleep duration calculation model based on load forecasting, dynamically determine the optimal sleep duration by analyzing and predicting the changing trend of business load and historical sleep patterns; define the basic sleep duration and adjust it according to the business load reduction rate.

[0032] Preferably, a wake-up advance calculation model is established, wherein the wake-up advance calculation model takes into account the power supply startup time and combines the prediction error to dynamically determine the power supply wake-up advance;

[0033] Build a progressive power supply wake-up power regulation mechanism, and use a multi-stage power regulation strategy to increase the power supply output power when the power supply wakes up from sleep state.

[0034] A power energy-saving system based on a load prediction intelligent sleep strategy, which is used to implement the power energy-saving method based on a load prediction intelligent sleep strategy, includes: a business-power data fitting module, a business load prediction module, a power supply power adjustment mechanism module, a power state switching strategy module, a sleep buffer module, an adaptive sleep duration adjustment module, and a progressive wake-up module;

[0035] The service-power data fitting module collects base station service load and power data, and uses a neural network to fit the nonlinear relationship between service load and power;

[0036] The service load prediction module constructs a service load prediction model, predicts future service loads, generates power supply power adjustment timing instructions based on the predicted service loads, and triggers the power supply power adjustment mechanism when the difference between the actual service load and the predicted service load exceeds a load difference threshold;

[0037] The power supply adjustment mechanism module constructs a power supply adjustment mechanism, dynamically sets upper and lower comparison thresholds for adjusting the power supply based on the hysteresis comparison principle, forms a hysteresis control interval, and executes the power supply adjustment instruction only when the predicted load continuously crosses the hysteresis control interval and the change trend is consistent;

[0038] The power state switching strategy module creates a power sleep and wake-up strategy, which sleeps when the service load is less than the sleep threshold and wakes up when the service load is greater than the wake-up threshold;

[0039] The sleep buffer module creates a sleep buffer mechanism, sets an adaptive sleep buffer time, and determines whether to enter the sleep state according to the buffer time;

[0040] The adaptive sleep duration adjustment module constructs an adaptive sleep duration mechanism to dynamically adjust the power sleep duration;

[0041] The progressive awakening module constructs a progressive power supply awakening power adjustment mechanism, and adjusts the power supply power to the target power of the target business load when the power supply is awakened.

[0042] The beneficial effects of this application: This application establishes a nonlinear mapping relationship between business load and power supply power, which can improve the scientific nature of power allocation, reduce excessive power consumption configuration, and improve energy efficiency; by perceiving the business load change trend in advance, forward-looking power supply power adjustment can be achieved, effectively avoiding service quality degradation or energy waste due to lagging adjustment.

[0043] This application can effectively filter out misjudgments caused by short-term fluctuations, reduce unnecessary frequent adjustments, and ensure the stability and reliability of adjustment decisions; it realizes on-demand power supply operation, automatically enters sleep mode during low-load periods to reduce energy consumption, and wakes up in time when the load increases to ensure service continuity.

[0044] This application can effectively avoid frequent sleep and wake-up caused by short-term load fluctuations, and improve the accuracy and reliability of sleep decisions; it can intelligently adjust the sleep duration according to future load trends, make full use of low-load periods, further improve energy saving effects, and take into account business service needs; it realizes a smooth transition of power supply from sleep to normal operation, avoids the impact of large power jumps on power supply equipment, and can extend the life of power supply equipment.

[0045] This application realizes refined and intelligent energy-saving management of 5G base station power supply by integrating multiple intelligent technologies such as neural network modeling, multimodal load prediction, hysteresis control, intelligent sleep and buffering, dynamic duration and gradual wake-up.

[0046] This application solution can flexibly adjust power output based on actual business load changes and is suitable for managing the power supply of 5G base stations. It achieves the technical effect of significantly improving energy efficiency, reducing energy consumption, reducing equipment loss, and extending power supply life, while ensuring high reliability and continuity of business services. Compared with traditional static or linear energy-saving strategies, this method is suitable for the management level of 5G base station power supply and has significant advantages in energy saving effects, adjustment accuracy, and operational stability. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Flowchart of the power saving method based on load prediction intelligent sleep strategy provided by this application;

[0048] Figure 2 Flowchart of business load and power supply correlation modeling provided for this application;

[0049] Figure 3 A flow chart of multi-modal load prediction and power supply instruction generation provided in this application;

[0050] Figure 4 Flowchart of the dynamic power adjustment mechanism in the hysteresis control interval provided by this application;

[0051] Figure 5 A flowchart for building an intelligent sleep and wake-up strategy provided for this application;

[0052] Figure 6 Flowchart of the adaptive sleep buffer mechanism provided by this application;

[0053] Figure 7 Flowchart of the adaptive sleep duration optimization mechanism provided by this application;

[0054] Figure 8 This is a structural diagram of the power energy-saving system based on load prediction intelligent sleep strategy provided by this application. DETAILED DESCRIPTION

[0055] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the specific implementation methods of the present application are described in detail below in conjunction with the drawings in the specification.

[0056] In the following description, many specific details are set forth to facilitate a full understanding of the present application. However, the present application can also be implemented in other ways different from those described herein. Those skilled in the art can make similar generalizations without violating the connotation of the present application. Therefore, the present application is not limited to the specific embodiments disclosed below.

[0057] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present application. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0058] Example 1

[0059] Reference Figure 1 , which is the first embodiment of the present application, provides a power saving method based on load prediction intelligent sleep strategy.

[0060] The technical solution of this application demonstrates a high degree of adaptability to 5G base station power management through load prediction, intelligent sleep strategy and dynamic power adjustment mechanism. 5G base stations have extremely stringent requirements on power management due to the strong volatility of business loads, complex power requirements and high real-time requirements. This application can provide comprehensive power management support based on the complex and changeable business load characteristics of 5G base stations through the comprehensive application of load prediction, intelligent sleep wake-up, dynamic power adjustment and other technical means. Whether in high-load or low-load scenarios, the power supply power can be adaptively adjusted to achieve the technical effect of both ensuring service quality and efficient use of electricity.

[0061] In addition, this solution has significant advantages in terms of the stability and continuity of power regulation. Through a global intelligent power management strategy, frequent energy consumption fluctuations and equipment losses are avoided, making it particularly suitable for environments such as 5G base stations that require long-term and efficient operation. Overall, the technical solution of this application shows extremely high adaptability to the operating characteristics of 5G base stations, from energy-saving efficiency, real-time response capabilities to equipment reliability, providing strong technical support for the sustainable development of 5G networks.

[0062] Step 1: Collect base station service load and power data, and use neural network to fit the nonlinear relationship between service load and power. Figure 2 , which is the flow chart of business load and power supply correlation modeling in this step.

[0063] The collected business load data and power data include the number of user connections, data throughput and resource block utilization; for example, the collection frequency is set to once every 5 minutes to form a business load vector L t =[n t ,d t ,r t ], where n t Indicates the number of user connections at time t, d t represents the data throughput at time t, r t The power data is obtained through the power monitoring interface of the power supply, including the real-time output power P t and power efficiency η t (Value range 0-1).

[0064] The collected raw data is cleaned and standardized. First, abnormal data points are identified and eliminated using the 3σ criterion, and missing data are supplemented using linear interpolation. Then, the business load data is normalized, and the number of user connections is normalized to where n max is the maximum number of users supported by the base station; the data throughput is normalized to where d max is the maximum throughput of the base station; resource block utilization rate r t It is already in the range of 0-1 and does not need to be normalized; the power data is normalized to Among them, P rated is the rated power of the power supply; the normalized data processing eliminates the influence of data of different dimensions, which can improve the convergence speed and stability of neural network training; These are the normalized values ​​of the number of user connections, data throughput, and power consumption data, respectively.

[0065] Construct a neural network model and use a three-layer feedforward neural network to fit the nonlinear mapping relationship between business load and power supply power. Specifically, the input layer of the three-layer feedforward neural network contains 3 neurons, corresponding to the normalized business load vector The hidden layer contains 20 neurons and uses the ReLU activation function; the output layer contains 1 neuron and outputs the predicted normalized power The network structure of the three-layer feedforward neural network can effectively capture the complex nonlinear relationship between various dimensions of business load and power supply power.

[0066] Train a three-layer feedforward neural network, using the Adam optimization algorithm for parameter updates. For example, the learning rate is set to 0.001, the batch size is 64, and the number of training epochs is 200. The dataset is split into training, validation, and test sets in a ratio of 7:2:1. During training, early stopping is triggered when the validation set loss does not decrease for 10 consecutive epochs to prevent overfitting.

[0067] Based on the trained neural network model, an optimal power mapping table is generated. The service load space is discretized. For example, the number of user connections is divided into 10 levels with a step size of 10% of the base station capacity, the data throughput is divided into 10 levels with a step size of 10% of the maximum throughput, and the resource block utilization is divided into 10 levels with a step size of 0.1, forming 1000 service load combination points. For each combination point, the neural network model is input to calculate the corresponding optimal power supply power, taking into account a 10% power margin: in is the predicted normalized power supply, P optimal The power supply power is calculated with a 10% power margin. The calculation results are stored in a lookup table to achieve a fast mapping of service load to optimal power supply power.

[0068] This step, by constructing a neural network-based nonlinear mapping model between service load and power, accurately quantifies the optimal power supply under varying network service loads in 5G base station scenarios. Compared to traditional linear power regulation methods, this technical solution more accurately matches the actual power requirements of 5G base stations, avoiding over-configuration of 5G base station power while ensuring the quality of 5G communication services.

[0069] Step 2: Build a business load prediction model to predict future business loads. Generate power adjustment timing instructions based on the predicted business loads. When the difference between the actual business load and the predicted business load exceeds the load difference threshold, the power adjustment mechanism is triggered. Figure 3 , which is the multi-modal load prediction and power supply power instruction generation flow chart provided in this step.

[0070] Three independent feature extraction channels are constructed to process the multimodal features of time series features, periodic features and emergency features respectively. A multimodal business load prediction model is constructed based on the multimodal features to predict future business load.

[0071] Construct a time series feature extraction channel and use a one-dimensional convolution layer to extract the short-term variation pattern of the business load. For example, the convolution kernel size is set to 5 and the number of convolution kernels is 32. The business load sequence X in the past 24 hours is s ={x s-287 ,x s-286,...,x s} for feature extraction, where x s Represents the traffic load vector of the s-th 5-minute time slice.

[0072] A periodic feature extraction channel is constructed, and Fourier transform is used to extract daily and weekly periodic features. The spectrum amplitude corresponding to the 24-hour period and the 7-day period is extracted as the periodic feature.

[0073] Construct an emergency feature extraction channel. For example, the emergency feature vector E is constructed by counting the load peak change rate in the same period in the past 7 days. s =[μ s ,σ s ,γ s ], where μ s is the mean, σ s is the standard deviation, γ s is the kurtosis coefficient.

[0074] A multimodal business load prediction model is constructed based on the long short-term memory network (LSTM) model. The LSTM model is used to fuse multimodal features extracted from three feature channels: time series, periodicity, and emergencies. The LSTM model uses deep learning to model the dynamic changes in business load in the time dimension.

[0075] The LSTM model structure consists of two stacked LSTM layers, each containing multiple hidden units. It can effectively capture long-term dependencies and short-term fluctuation characteristics, and adapt to the changing patterns of business loads in different periods of time. The state update process of the LSTM model includes structural modules such as the forget gate, input gate, candidate memory unit, memory unit, and output gate. The forget gate is used to control how much historical memory information is retained. The input gate determines the degree of update of the memory state by the current input. The candidate memory generates new memory content through a nonlinear function and jointly decides with the input gate whether to write it into the memory unit. The output gate controls the output value of the current hidden state.

[0076] Each gating structure of the LSTM model is implemented through the learned weight matrix and bias vector, where the sigmoid function is used to generate the gating coefficient, and the tanh function is used to activate the candidate state. Finally, the historical information is fused with the current input through element-by-element multiplication to achieve modeling of time series data.

[0077] The attention mechanism is introduced into the LSTM model to assign different weights to input features at different time steps. By calculating the correlation between historical time steps and the predicted target, the attention mechanism generates an attention weight vector and dynamically adjusts the influence of each time step's features on the prediction process. This mechanism enables the model to focus on the historical features that most contribute to the current prediction, effectively improving the accuracy and stability of predictions.

[0078] Output the future business load through the LSTM model. For example, taking a prediction every 5 minutes as an example, a business load prediction sequence for the next 1 hour (i.e. 12 time steps) is generated. in, Indicates the business load forecast result for the 12th time step in the future. The forecast process adopts a rolling forecast strategy, using the forecast output of the previous step as the input of the next step to form a recursive forecast chain.

[0079] The prediction results are converted into actual business load values ​​through denormalization and subjected to rationality checks to ensure that the predicted values ​​are within the physical capacity of the base station; the multi-step prediction mechanism provides sufficient time window for early adjustment of power supply.

[0080] Based on the predicted service load sequence and the established power mapping table, a power supply power adjustment timing instruction set Ω={(τ1,P1),(τ2,P2),...,(τ k ,P k ),...,(τ K ,P K )}, where τ k represents the kth adjustment moment, P k represents the corresponding target power value, and K is the total number of instructions.

[0081] The instruction generation follows the power smoothing principle. The interval between two adjacent power adjustments is not less than ζ minutes, and the power adjustment range does not exceed ξ times the current power. The instruction sequence is optimized by the dynamic programming algorithm. The objective function Among them, P required,k To predict the ideal power corresponding to the load, λ1 and λ2 are the power matching weight and smoothing adjustment weight respectively, λ1+λ2=1; power smoothing can ensure sufficient power supply while reducing the loss caused by frequent power adjustment.

[0082] Compare actual business load in real time actual and predicting business load The load deviation metric Δ is defined as: where ||·||2 represents the L2 norm.

[0083] Set the load difference threshold θ load , when continuous The load deviation of each sampling point is greater than the load difference threshold θ load When Δ τ+q-1 >θ load , determine that the forecast has a significant deviation; and calculate the deviation trend index at the same time when determining that the forecast has a significant deviation When Θτ >0 indicates that the deviation continues to expand, triggering the power adjustment mechanism, where q is the index value, q∈[1,Q], Δ τ+q-1 represents the qth moment starting from τ.

[0084] This step achieves proactive power regulation by building a multimodal traffic load prediction model and an intelligent power adjustment command generation mechanism. This approach is suitable for the highly volatile signaling techniques and relay equipment of 5G base stations. Compared to passive power-following strategies, this solution can predict traffic load trends in advance, allowing the power supply ample time to adjust power, avoiding power shortages or oversupply. This avoids energy waste while ensuring the quality of 5G base station signal service, achieving energy conservation.

[0085] Step 3: Build a power adjustment mechanism. Based on the hysteresis comparison principle, dynamically set the upper and lower comparison thresholds for adjusting the power supply, forming a hysteresis control interval. Only when the predicted load continuously crosses the hysteresis control interval and the change trend is consistent, execute the power adjustment instruction; see Figure 4 This is the flow chart of the dynamic power adjustment mechanism in the hysteresis control interval of this step.

[0086] In the hysteresis control interval construction stage, based on the current power supply power P current Based on the fluctuation characteristics of business load, the upper and lower thresholds of power regulation are dynamically set; the hysteresis control interval is defined as [Γ lower ,Γ upper ], where the lower threshold Γ lower =P current ×(1-δ down ), upper threshold Γ upper =P current ×(1+δ up ), δ down and δ up are the downward and upward hysteresis coefficients respectively.

[0087] According to the historical load fluctuation amplitude σ load Dynamic adjustment of hysteresis coefficient: δ down =κ1×σ load +κ2,δ up =κ3×σ load +κ4, where κ1, κ2, κ3, and κ4 are empirical coefficients. κ1 and κ2 are the sensitivity of the control to the fluctuation amplitude. The larger the κ1 and κ2 are, the more sensitive the hysteresis coefficient is to the fluctuation. κ3 and κ4 are baseline biases, which are used to ensure that the hysteresis control range will not shrink to zero even if the load is very stable.

[0088] Monitor and predict business load sequences The relationship with the hysteresis interval, where T pred To predict the duration, is the prediction result of the last time step in the prediction interval; define the span judgment function Ω(t):

[0089]

[0090] in To predict the required power corresponding to the load; calculate the continuous span duration Where i is a parameter, is the indicator function. cross ≥T th When a valid crossing occurs, T th is the minimum duration threshold.

[0091] The constructed continuity judgment mechanism can effectively filter out short-lived business load spikes or troughs, ensuring that power regulation decisions are based on stable load change trends.

[0092] In the consistency verification phase of the change trend, the first-order difference sequence of the predicted load is calculated To evaluate the trend of change. Define the trend consistency index Υ trend :

[0093]

[0094] Where sign(·) is the sign function. trend >υ th When , it indicates that the change direction of the predicted load is consistent with the span direction, where υ th is the consistency threshold; trend strength indicator is calculated at the same time when When it is determined that there is an obvious trend of change, is the minimum trend strength threshold.

[0095] The dual verification mechanism ensures that power regulation is only performed when the load has a clear and consistent change trend, avoiding invalid power adjustments during business load fluctuations.

[0096] In the power adjustment decision generation stage, the power adjustment decision is generated by integrating the hysteresis crossing judgment and trend consistency verification results; the power adjustment decision function is defined when When , a power increase command is generated; when When , a power reduction command is generated; when When the current power is kept unchanged.

[0097] The calculation process of the power adjustment amplitude ΔP is as follows: Among them, P targetis the target power, δ max This is the maximum adjustment range for a single power adjustment. This limited adjustment strategy ensures smooth power changes and prevents large adjustments from impacting power supply safety.

[0098] Step 3 achieves stable and efficient dynamic power regulation by constructing a power supply adjustment mechanism based on the hysteresis comparison principle. The constructed power supply adjustment mechanism adapts to different 5G signal load fluctuation characteristics through dynamic hysteresis control interval, and effectively avoids frequent and incorrect adjustments of 5G base station power supply through dual guarantees of continuous crossing judgment and trend consistency verification.

[0099] Step 4: Create a power sleep and wake-up policy. The system will sleep when the service load is continuously below the sleep threshold and wake up when the service load is above the wake-up threshold. Figure 5 This is the flowchart for building the intelligent sleep and wake-up strategy for this step.

[0100] Establish an adaptive sleep threshold based on the statistical characteristics of historical business load data; define the basic sleep threshold Λ sleep =η base ×L min +(1-η base )×L avg , where L min is the historical minimum load value, L avg is the historical average load value, η base is the basic weight coefficient.

[0101] Dynamically adjust the basic dormancy threshold based on the characteristics of the current time period: in Is the dynamically changing sleep threshold, time adjustment factor t hour is the current hour, ω amp To adjust the amplitude.

[0102] Different adjustment strategies are adopted for weekdays and holidays by letting the date type coefficient ι day Take different values ​​to distinguish between working days and holidays. For example, when ι day =0.8 represents working days, when ι day =1.0 indicates holidays; sleep threshold for: The multi-factor dynamic adjustment mechanism enables the sleep threshold to adapt to the load characteristics of different time periods and date types, improving the accuracy of sleep decisions.

[0103] In the wake-up threshold collaborative setting stage, the wake-up threshold associated with the sleep threshold is constructed and the wake-up threshold is defined. where πgap is the hysteresis interval coefficient; the interval coefficient is dynamically adjusted according to the fluctuation of business load: where π min and π max are the minimum and maximum interval coefficients, σ recent is the recent load standard deviation, σ ref is the reference standard deviation.

[0104] Establish time window determination rules to ensure that sleep and wake-up decisions are based on stable business load trends; for sleep determination, define persistence conditions:

[0105]

[0106] Among them, S sleep is the dormant state determination condition, T sleep is the sleep determination window length, is the indicator function, Y current-a is the business load of the ath sampling point before the current moment, a is a parameter; when S sleep =1, that is, continuous T sleep The sleep continuity condition is met only when the load of each sampling point is lower than the sleep threshold.

[0107] Similarly, the wakeup persistence condition is defined as:

[0108]

[0109] Among them, S wake is the awakening state judgment condition, T wake The window length is determined for wake-up; the window length is adaptively adjusted according to the rate of change of business load: in is the load change rate, L current-b is the load value at the bth time point forward from the current moment, T base The basic time length of the determination window. The persistence determination mechanism effectively filters out the interference of instantaneous load fluctuations and avoids abnormal wake-up caused by transient business load.

[0110] In the state trigger decision generation stage, the final sleep or wake-up trigger decision is generated by integrating multi-dimensional information and defining the sleep trigger function

[0111]

[0112] in is the predicted future average load, t active is the duration of the current active state, t min is the minimum active time requirement; when Enters sleep mode when Does not enter sleep mode.

[0113] Wake-up trigger function Defined as:

[0114]

[0115] where Q buffer is the current buffer queue length, Q th is the queue threshold, when Enter the awake state when Do not enter the awakening state;

[0116] Set up emergency wake-up mechanism, when instantaneous load L instant >Λ emergency When Λ emergency =ε1×Λ wake ,ε1 is the multiplier of the emergency wake-up mechanism to the wake-up threshold, which is generally greater than 2, triggering wake-up immediately without waiting for continuity judgment.

[0117] Step 4: Implement refined energy consumption management of 5G base station power supplies by building an intelligent power sleep and wake-up strategy. The power sleep and wake-up strategy adapts to the load characteristics of different scenarios through dynamic threshold setting, avoids frequent switching through strict continuity judgment, ensures decision accuracy through multi-dimensional triggering, and ensures service continuity through gradual state transition.

[0118] Step 5: Create a sleep buffer mechanism, set an adaptive sleep buffer time, and determine whether to enter the sleep state based on the buffer time; see Figure 6 Construct a flow chart of the adaptive sleep buffer mechanism for this step.

[0119] Establish a historical sleep success rate statistical model, obtain the execution status of the past N sleep cycles, and define the sleep success rate where N s Indicates the number of times the system successfully enters the sleep state, N t Indicates the total number of sleep attempts; at the same time, obtains the business load rebound time t when each sleep failure occurs r , construct the rebound time set T r ={t r1 ,t r2 ,...,t rc}, where c is the total number of sleep failures, t rc The rebound time of the business load after the cth sleep failure.

[0120] Calculate the adaptive sleep buffer time based on historical data, the sleep buffer time T bUsing dynamic calculation method: in Is the basic buffer time; R s is the dormancy success rate; α is the standard deviation weight coefficient; β is the failure rate weight coefficient; T max The maximum buffer time limit.

[0121] The buffer time is dynamically adjusted according to the actual operation situation. When the historical hibernation success rate is high, the buffer time is relatively short to improve the hibernation response speed. When the success rate is low or the rebound time fluctuates greatly, the buffer time is appropriately extended to improve the hibernation success rate.

[0122] Implement the sleep buffer state monitoring mechanism. After receiving the sleep command, the power supply immediately enters the sleep buffer state, during which the current power output remains unchanged and the real-time monitoring program is started. The monitoring program collects the current business load L every Δt seconds. c (t), and the sleep threshold Compare and define the load rebound judgment condition as follows: Where γ is the rebound determination coefficient. If the dormant buffer time T b If the rebound condition is met at any time within the time limit, the system will immediately exit the sleep buffer state, cancel the current sleep operation, and record the rebound time t r =tt start , where t start is the time of entering the buffer state, and t is the current time.

[0123] Establish a sleep confirmation and state transition mechanism. When the sleep buffer time T b When the process ends and no load rebound occurs, confirm that the process is in sleep mode; calculate the average business load during the buffer period. Where n0 is the number of sampling times during the buffer period, n≤n0, L c (t n ) is the traffic load during the sampling period of the buffer period. If it is the dynamic sleep threshold, the power is confirmed to enter the sleep state and the number of successful sleep times N is updated at the same time. s =N s +1 for recalculating the sleep success rate.

[0124] The sleep buffer mechanism established in this step effectively avoids frequent sleep and wake-up switching caused by short-term fluctuations in service load, reducing energy loss associated with power state switching. The adaptive buffer time calculation method enables continuous optimization of the buffer strategy based on actual operating conditions, maximizing energy savings while ensuring service quality. The introduction of a real-time monitoring and confirmation mechanism improves the reliability of sleep decisions, reduces the risk of service interruptions caused by erroneous sleep, and overall enhances intelligent power management and energy efficiency.

[0125] Step 6: Build an adaptive sleep duration mechanism to dynamically adjust the power sleep duration; see Figure 7 Construct a flow chart of the adaptive sleep duration optimization mechanism for this step.

[0126] Build an adaptive sleep duration calculation model based on load forecasting, dynamically determine the optimal sleep duration by analyzing and forecasting the changing trend of business load and historical sleep patterns; define the basic value of sleep duration in is the average business load in the predicted time window, φ1 is the load correlation coefficient, and φ2 is the duration bias constant; the basic duration is adjusted based on the load reduction rate to calculate the load reduction rate Among them L current is the current business load, To predict the traffic load at the end of the window, T pred Forecast duration.

[0127] Adaptive sleep duration D sleep Calculated as: Where ξ1 is the rate adjustment coefficient, v ref is the reference descent rate, v down The load decrease rate is calculated based on the load trend. The duration calculation mechanism can extend the sleep time when the load decreases rapidly, making full use of the low-load period to achieve higher energy saving effects.

[0128] Establish a wake-up advance calculation model: T advance =T startup +χ·σ pred , where T startup is the power startup time, χ is the safety factor, σ pred is the standard deviation of the prediction error; T advance The early wake-up mechanism ensures that the power supply can complete power preparation before the load increases, avoiding service interruption.

[0129] The adaptive sleep duration mechanism constructed in this step achieves intelligent optimization of sleep duration by comprehensively considering load prediction and sleep depth control, maximizing energy saving while ensuring business service quality. Compared with the fixed sleep duration solution, the dynamic sleep duration has better energy saving effect and better user experience.

[0130] Step 7: Build a progressive power adjustment mechanism for power wake-up to adjust the power to the target power of the target business load when waking up the power.

[0131] Build a progressive power ramp control model. When the power supply is awakened from sleep mode, a multi-stage power regulation strategy is used to gradually increase the output power. Define the power ramp curve. Among them, P sleep is the sleep power, P target is the target power, is the climbing rate control parameter, T advance It is the early wake-up time and also the total power ramp-up time.

[0132] The progressive power wake-up adjustment mechanism achieves a smooth transition from sleep to normal operation of the 5G base station power supply through precise power ramp control, dynamic target positioning, and segmented adjustment strategies. Compared with the traditional direct wake-up method, the progressive power wake-up adjustment mechanism causes almost no loss of power supply health during the wake-up process, which is conducive to the long-term and stable operation of the power supply.

[0133] Example 2

[0134] Reference Figure 8 , which is the second embodiment of the present application, provides a power energy-saving system based on load prediction intelligent sleep strategy.

[0135] The system includes: a business-power data fitting module, a business load prediction module, a power adjustment mechanism module, a power state switching strategy module, a sleep buffer module, an adaptive sleep duration adjustment module and a progressive wake-up module.

[0136] The service-power data fitting module collects base station service load and power data, and uses a neural network to fit the nonlinear relationship between the service load and the power.

[0137] The business load prediction module builds a business load prediction model, predicts future business load, generates power supply power adjustment timing instructions based on the predicted business load, and triggers the power supply power adjustment mechanism when the difference between the actual business load and the predicted business load exceeds the load difference threshold.

[0138] The power supply power adjustment mechanism module constructs a power supply power adjustment mechanism, dynamically sets the upper and lower comparison thresholds for adjusting the power supply power based on the hysteresis comparison principle, forms a hysteresis control interval, and executes the power supply power adjustment instruction only when the predicted load continuously crosses the hysteresis control interval and the change trend is consistent.

[0139] The power state switching strategy module creates a power sleep and wake-up strategy, which sleeps when the service load is less than the sleep threshold and wakes up when the service load is greater than the wake-up threshold.

[0140] The sleep buffer module creates a sleep buffer mechanism, sets an adaptive sleep buffer time, and determines whether to enter the sleep state according to the buffer time.

[0141] The adaptive sleep duration adjustment module constructs an adaptive sleep duration mechanism to dynamically adjust the power sleep duration.

[0142] The progressive awakening module constructs a progressive power supply awakening power adjustment mechanism, and adjusts the power supply power to the target power of the target business load when the power supply is awakened.

[0143] In the embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some communication interface, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0144] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the purpose of this application and the scope of protection of the claims, which are all within the protection of this application.

Claims

1. A power energy saving method based on load prediction intelligent sleep strategy, characterized in that: include: Collect business load and power data, and use neural networks to fit the nonlinear relationship between business load and power; A business load prediction model is built based on a neural network model to predict future business loads. A power supply adjustment timing instruction is generated based on the predicted business load. When the difference between the actual business load and the predicted business load exceeds a load difference threshold, the power supply adjustment mechanism is triggered. Build a power adjustment mechanism that dynamically sets upper and lower comparison thresholds for adjusting power based on the hysteresis comparison principle, forming a hysteresis control interval. Only when the predicted load continuously crosses the hysteresis control interval and the change trend is consistent, will the power adjustment instruction be executed. Create power sleep and wake-up policies, which will enable the system to sleep when the service load is continuously below the sleep threshold and wake up when the service load is above the wake-up threshold. Create a sleep buffer mechanism, set an adaptive sleep buffer time, and determine whether to enter the sleep state based on the buffer time; Build an adaptive sleep duration mechanism to dynamically adjust the power sleep duration; Build a progressive power adjustment mechanism for power wake-up, and adjust the power to the target power of the target business load when waking up the power.

2. The power saving method based on load prediction intelligent sleep strategy according to claim 1 is characterized in that: The collected service load and power data are statistically removed for outliers and interpolated for missing values. They are then normalized according to the base station design capacity to eliminate dimensional differences between different data types. A three-layer feedforward neural network is used to fit the nonlinear mapping relationship between business load and power supply power. The input corresponds to the parameters of each dimension of business load, and the output generates normalized power supply power. The hidden layer uses a nonlinear activation function. Based on the trained three-layer feedforward neural network, the optimal power supply power corresponding to the business load is calculated, and a mapping table between the business load and the optimal power supply power is generated for adjusting the optimal power supply power.

3. The power saving method based on load prediction intelligent sleep strategy according to claim 2 is characterized in that: Three independent feature extraction channels are constructed to process multimodal features of time series features, periodic features, and emergency features respectively. Based on these multimodal features, a multimodal traffic load prediction model is constructed to predict future traffic load. The multimodal traffic load prediction model is constructed based on a long short-term memory network (LSTM) model. The LSTM model is used to fuse multimodal features extracted from three feature channels: time series, periodicity, and emergencies, and learn the dynamic change pattern of traffic load in the time dimension. The LSTM model is used to predict future business loads. Based on the predicted business load sequence and the established power mapping table, a power supply power adjustment timing instruction set is generated to adjust the power supply power.

4. The power saving method based on load prediction intelligent sleep strategy according to claim 3 is characterized in that: Compare the difference between actual business load and predicted business load in real time; A load difference threshold is set. When the load deviation of multiple consecutive sampling points is greater than the load difference threshold, the prediction is judged to have a significant deviation. At the same time, the deviation trend indicator is calculated. If the deviation trend indicator diverges, the power adjustment mechanism is triggered.

5. The power saving method based on load prediction intelligent sleep strategy according to claim 4 is characterized in that: Based on the current power supply and business load fluctuations, the upper and lower thresholds of power supply adjustment are dynamically set to build a hysteresis control range; The power supply adjustment mechanism is triggered only when the predicted power continuously crosses the hysteresis control interval and the load change direction is consistent with the crossing direction.

6. The power energy saving method based on load prediction intelligent sleep strategy according to claim 5, characterized in that: Establish an adaptive power sleep threshold based on the statistical characteristics of historical business load data, and dynamically adjust the power sleep threshold based on time period characteristics and date type; Build a power wake-up threshold and set an emergency wake-up mechanism. When the instantaneous load exceeds the threshold, the power wake-up is triggered immediately.

7. The power energy saving method based on load prediction intelligent sleep strategy according to claim 6, characterized in that: The sleep buffer time is dynamically calculated and determined based on historical sleep success data, including the success rate and the business load rebound time when failure occurs; Implement a sleep buffer state monitoring mechanism. After the power supply receives a sleep command, it enters the sleep buffer state and maintains the current power output unchanged. Monitor the business load in real time. When the business load meets the preset load rebound judgment conditions, the power supply exits the sleep buffer state and cancels the sleep operation.

8. The power energy saving method based on load prediction intelligent sleep strategy according to claim 7, characterized in that: Establish a sleep confirmation and state transition mechanism. When the sleep buffer time expires and no load rebound occurs, sleep is confirmed. Based on the comparison of the average service load during the buffer period with the dynamic sleep threshold, the power supply is confirmed to enter the sleep state. The number of successful sleep attempts is also updated to optimize subsequent decisions. Build an adaptive sleep duration calculation model based on load forecasting, dynamically determine the optimal sleep duration by analyzing and predicting the changing trend of business load and historical sleep patterns; define the basic sleep duration and adjust it according to the business load reduction rate.

9. The power saving method based on load prediction intelligent sleep strategy according to claim 8, characterized in that: Establishing a wake-up advance calculation model, wherein the wake-up advance calculation model considers the power supply startup time and combines the prediction error to dynamically determine the power supply wake-up advance; Build a progressive power supply wake-up power regulation mechanism, and use a multi-stage power regulation strategy to increase the power supply output power when the power supply wakes up from sleep state.

10. A power energy saving system based on a load prediction intelligent sleep strategy, which is used to implement a power energy saving method based on a load prediction intelligent sleep strategy according to any one of claims 1 to 9, characterized in that: include: Business-power data fitting module, business load prediction module, power supply adjustment mechanism module, power state switching strategy module, sleep buffer module, adaptive sleep duration adjustment module and progressive wake-up module; The service-power data fitting module collects base station service load and power data, and uses a neural network to fit the nonlinear relationship between service load and power; The service load prediction module constructs a service load prediction model, predicts future service loads, generates power supply power adjustment timing instructions based on the predicted service loads, and triggers the power supply power adjustment mechanism when the difference between the actual service load and the predicted service load exceeds a load difference threshold; The power supply adjustment mechanism module constructs a power supply adjustment mechanism, dynamically sets upper and lower comparison thresholds for adjusting the power supply based on the hysteresis comparison principle, forms a hysteresis control interval, and executes the power supply adjustment instruction only when the predicted load continuously crosses the hysteresis control interval and the change trend is consistent; The power state switching strategy module creates a power sleep and wake-up strategy, which sleeps when the service load is less than the sleep threshold and wakes up when the service load is greater than the wake-up threshold; The sleep buffer module creates a sleep buffer mechanism, sets an adaptive sleep buffer time, and determines whether to enter the sleep state according to the buffer time; The adaptive sleep duration adjustment module constructs an adaptive sleep duration mechanism to dynamically adjust the power sleep duration; The progressive awakening module constructs a progressive power supply awakening power adjustment mechanism, and adjusts the power supply power to the target power of the target business load when the power supply is awakened.

Images