Server cabinet system

Abstract

The invention relates to the field of server systems, and discloses a server cabinet system, which comprises a real-time power monitoring module used for collecting server current data at a frequency of once per second, calculating real-time power and dynamically adjusting power distribution; the intelligent prediction and optimization module is used for predicting future demands and generating an optimal distribution strategy based on the normalized historical power consumption data and load data; and the redundant power supply switching module comprises a main power supply and a standby power supply converted by the lithium battery pack through an inverter, and is used for switching to the standby power supply when the voltage of the main power supply is lower than 180V for 10ms. According to the method, the elastic power supply strategy is generated by fusing the prediction demand and the two-dimensional weight of the service priority, the continuous operation of the core service is ensured, the prior art depends on a single priority label to allocate resources, the burst load and the long-term stability are difficult to balance, and the problem of resource waste caused by separation of the two is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of server systems, and in particular to a server cabinet system. Background Art

[0002] In recent years, with the surge in data processing demand, server systems have evolved from single servers to rack-mounted servers to meet high-density computing needs. A stable power supply is crucial for rack-mounted servers. Any power overload or underload can damage the server, leading to data loss and significant financial losses. Furthermore, power consumption during extended operation has become a key cost management consideration.

[0003] Existing cabinet server systems are usually equipped with basic power management systems to monitor power status, but they are insufficient in terms of power distribution optimization and fault prevention.

[0004] Current power management systems often lack sophisticated power consumption monitoring and intelligent optimization functions, making it difficult to effectively prevent power overload or insufficient conditions. They also lack flexibility and efficiency in power consumption control, resulting in waste of resources. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the present invention provides a server cabinet system, which solves the problem of lack of flexibility and efficiency in power consumption and optimization in the prior art.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: A server cabinet system includes: a real-time power monitoring module for collecting server current data at a frequency of once per second, calculating real-time power and dynamically adjusting power distribution; Hall effect sensors deployed at the server power inputs sense current signals in real time, converting the analog signals into digital quantities to calculate instantaneous power values. When a single server or the entire system detects power outages, a dynamic allocation algorithm is triggered, reducing power allocations to non-critical servers by a preset ratio to prioritize power supply to core business nodes.

[0007] Module relationship: This module is the data source of the system. The real-time power data it outputs directly drives the input of the intelligent prediction and optimization module, and provides voltage status monitoring signals for the redundant power switching module.

[0008] Intelligent prediction and optimization module, based on normalized historical power consumption and load data, predicts future demand and generates optimized allocation strategies; Based on historical power consumption trends and server load fluctuations, a deep learning model is used to predict short-term (6-60 seconds) power demand changes. This is combined with real-time load status to generate a dynamic allocation strategy. When the predicted demand exceeds the cabinet capacity, power resources are reallocated based on priority weights.

[0009] Module Relationship: Receives real-time data from the real-time power monitoring module and outputs optimization strategies to the power distribution unit. Shares fault warning signals with the redundant power switching module to ensure rapid switching to backup power in the event of a sudden power outage.

[0010] Redundant power switching module, including the main power supply and the backup power supply converted by the lithium battery pack through the inverter, is used to switch to the backup power supply when the main power supply voltage is lower than 180V for 10ms; The primary and backup power supplies are connected in parallel to the power distribution unit. A voltage detection circuit continuously monitors the primary power supply's status. If the voltage consistently falls below a safety threshold, a relay is triggered to switch to the backup power source. This seamless switching process is achieved through the inverter, ensuring zero power interruption to the servers.

[0011] Module relationship: The action is triggered by the voltage anomaly detection signal from the real-time power monitoring module. The switching event is reported to the central system through the communication interface, and the intelligent prediction module is linked to update the allocation strategy.

[0012] The communication interface uses the ModbusTCP protocol to transmit real-time power, warning status, and power switching events to the central control system in JSON format; Real-time data and event information are encapsulated in lightweight JSON format and transmitted with low latency via the ModbusTCP protocol. Data packets contain timestamps, server identifiers, power values, and status codes, supporting command issuance and status queries from the central system.

[0013] Module relationship: As the hub for interaction within and outside the system, it receives and forwards the raw data of the real-time power monitoring module, the allocation instructions of the intelligent prediction module, and the switching events of the redundant power supply module, and distributes the control instructions of the central system to each module.

[0014] Preferably, the real-time power monitoring module includes: The Hall effect current sensor unit is deployed at the power input of each server and collects current data once per second. The current value calculation satisfies the following requirements: Where D is the digital value output by the ADC module (0 to 4095), and 0.066 is the sensor sensitivity (V / A). Based on the principle of magnetic field induction, the sensor detects changes in the magnetic field around the server power cord and generates a voltage signal that is linearly proportional to the current. This signal is amplified and filtered before being output to the data processing unit, ensuring non-invasive measurement without affecting the stability of the original circuit.

[0015] Module relationship: As the physical layer component of data acquisition, its output provides the original current value for the data processing unit and provides real-time load status input for the control strategy unit.

[0016] Data processing unit, according to formula P i (t)=220·I i (t) Calculate real-time power; The analog voltage signal output by the sensor is converted into a digital value, and the current value is calculated through a linear mapping relationship. The instantaneous power is further calculated based on the fixed power supply voltage (220V). The unit has a built-in sliding window filtering algorithm to eliminate the interference of transient noise on data stability.

[0017] Module relationship: Receives data input from the sensor unit, outputs accurate power values ​​to the control strategy unit, and provides historical data storage support for the upper-level intelligent prediction module.

[0018] The control strategy unit proportionally adjusts the power distribution of non-critical servers when the power of a single server exceeds 90% of its rated power or the total power of the cabinet exceeds 95% of the total power capacity; Real-time monitoring of the overall power status of individual servers and cabinets. When a single server's power exceeds a safety threshold (90% of rated power) or total power approaches capacity (95% of total capacity), a dynamic load reduction strategy is triggered. This strategy, based on server priority tags, reduces power supply quotas for non-critical nodes by a preset ratio, and coordinates adjustments with the power distribution module through a feedback mechanism.

[0019] Module relationship: Rely on the real-time calculation results of the data processing unit to generate control instructions, directly interact with the power distribution module to perform dynamic adjustments, and push overload events to the communication interface to report to the central system.

[0020] Preferably, the proportionally adjusted power distribution strategy satisfies: Among them, α is the adjustment coefficient, ranging from 0.5 to 1.0, P total The total power limit of the cabinet.

[0021] The power allocation strategy dynamically adjusts power allocation for non-critical servers based on the difference between real-time power consumption and the cabinet's total power limit. When total power exceeds a safety threshold, server power is adjusted based on a preset proportional coefficient to ensure that actual load remains below the cabinet's capacity limit. The coefficient ranges from 0.5 to 1.0, with the specific value dynamically determined by the current load fluctuation rate.

[0022] When the total cabinet power approaches or exceeds the preset capacity limit, the control strategy unit dynamically calculates the power adjustment for each non-critical server based on the difference between the current total power and the capacity limit, combined with the real-time power contribution of each server. The adjustment coefficient (α) serves as a weighting factor to balance the adjustment range with system stability. A higher α value (close to 1.0) indicates a fast response but is more likely to cause power fluctuations, while a lower α value (close to 0.5) slows the adjustment speed to maintain business continuity.

[0023] The control strategy unit receives real-time power data from the data processing unit, generates adjustment instructions, and sends them to the power distribution module for execution. At the same time, it dynamically feeds back the adjustment coefficient to the intelligent prediction module to optimize future distribution strategies.

[0024] The adjustment policy only applies to servers marked as non-critical, leaving critical servers (database nodes) unaffected. Pre-set server priority tags ensure that core services maintain stable operation during power adjustments.

[0025] Priority tag data comes from the server basic information database. The control strategy unit filters adjustable targets based on the tags, and the power distribution module performs differentiated operations based on the tags.

[0026] After each adjustment, the system monitors actual power changes in real time. If the total power still exceeds the threshold after adjustment, a second adjustment is triggered. This process iteratively approaches the target power value until the system returns to a safe state.

[0027] The execution result of the power distribution module is used to recalculate the total power through the data processing unit, forming a closed-loop control link of "detection → adjustment → re-detection".

[0028] Preferably, the intelligent prediction and optimization module includes: Long Short-Term Memory Network model, whose input features include normalized historical power data, load data, and time period encoding. The model structure is a two-layer LSTM, with each layer containing 64 neurons; The allocation strategy generation unit generates a pre-allocation strategy based on the predicted value using a greedy algorithm. When the predicted total demand exceeds the total capacity of the cabinet, it gives priority to ensuring the power supply of the peak demand server.

[0029] The intelligent prediction and optimization module includes a long-short-term memory (LSTM) network model and an allocation strategy generation unit. The LSTM model's input features are normalized historical power consumption data, server load data, and time period encoding. The model employs a two-layer structure, with 64 neurons per layer. The allocation strategy generation unit uses a greedy algorithm to generate a power pre-allocation strategy based on the model's predicted future demand. When the predicted total demand exceeds the cabinet capacity, it prioritizes power supply quotas for peak-demand servers.

[0030] A two-layer LSTM network architecture jointly models historical power consumption, server load, and time-period characteristics, capturing both long-term and short-term variations in power demand. The first LSTM layer extracts local fluctuations in time series data, while the second layer further learns dependencies across time steps, ultimately outputting power demand forecasts for the next 6 to 60 seconds. The model's normalized input features ensure compatibility with data of varying dimensions, and time encoding (hours, weeks) enhances the ability to identify periodic patterns.

[0031] Module relationship: Receives historical data and real-time load status from the real-time power monitoring module as input, and transmits the prediction results directly to the allocation strategy generation unit; model training relies on the support of the historical data storage module, and the prediction error data is fed back to the training process to continuously optimize the model parameters.

[0032] A greedy algorithm is used to generate resource allocation plans based on the dynamic relationship between predicted demand and cabinet capacity limits. When the total predicted demand is within the limit, power quotas are allocated proportionally to each server's predicted value. When the total predicted demand exceeds the limit, power supply to peak-demand servers is prioritized, and remaining capacity is allocated to other servers based on predicted weights, ensuring critical business continuity.

[0033] Module relationship: It works in conjunction with the control strategy unit of the real-time power monitoring module to receive over-limit warning signals to trigger the priority allocation logic; the generated strategy instructions are sent to the power distribution module through the communication interface for execution, and the allocation results are written to the log for subsequent optimization analysis.

[0034] Preferably, the training parameters of the long short-term memory network model include: The optimizer is Adam and the learning rate is 0.001; The loss function is the mean square error; The input sequence length is 24 hours, and the power demand is predicted for the next 6 to 60 seconds.

[0035] The Adam optimizer combines momentum gradient descent with an adaptive learning rate to dynamically adjust the parameter update step size by calculating the first-order moment (mean) and second-order moment (variance) of historical gradients. A learning rate of 0.001 balances convergence speed and stability in the early stages of model training, preventing excessively large step sizes from skipping the optimal solution or excessively small step sizes from causing training stagnation.

[0036] Module relationship: The optimizer parameter configuration directly affects the model training efficiency. The trained model is integrated into the intelligent prediction module, which receives normalized data from the data processing unit in real time to generate prediction results.

[0037] The mean squared error (MSE) quantifies the model's prediction accuracy by calculating the mean of the squared differences between the predicted and true values. This function is sensitive to outliers and can effectively drive the model's prediction capabilities to address scenarios with sudden changes in power demand, such as sudden surges in power consumption caused by high server loads.

[0038] Module relationship: The loss function calculation results are fed back to the model training module to drive parameter updates; the prediction error data is synchronized to the central system log module for long-term performance evaluation and alarm analysis.

[0039] The 24-hour time window captures the full daily load fluctuation pattern, including the difference in power demand between peak daytime hours and idle nighttime hours. This design enables the model to learn load variations between weekdays and holidays, and across different time periods, enhancing the cyclical adaptability of the forecast results.

[0040] Module relationship: The input data comes from the historical database of the real-time power monitoring module, and the integrity of the input sequence is ensured by timestamp alignment; the model output is linked with the allocation strategy generation module to provide a basis for dynamic resource allocation.

[0041] Short-term forecasts (6 seconds) enable immediate power allocation decisions, while medium-term forecasts (60 seconds) support capacity pre-scheduling and redundant resource preparation. This multi-scale forecasting design balances real-time response with forward-looking planning, meeting the full range of scenarios, from second-level load fluctuations to minute-level resource allocation.

[0042] Module Relationship: The prediction results are transmitted to the power distribution module, triggering control strategies with different time granularities - 6-second predictions drive rapid adjustments, and 60-second predictions are used to optimize the module's long-term parameter configuration.

[0043] Preferably, the redundant power supply switching module includes: The voltage detection circuit uses a voltage divider resistor and an LM358 comparator to output a low-level signal when the main power supply voltage is lower than 180V; The switching controller triggers the backup power supply after receiving a low-level signal. The switching conditions are met: Where, τ = 10ms, Indicates the status of the main power supply at time t'.

[0044] The redundant power supply switching module includes a voltage detection circuit and a switching controller. The voltage detection circuit uses a voltage divider resistor and an LM358 comparator to monitor the main power supply voltage. When the main power supply voltage falls below 180V, it outputs a low-level signal. The switching controller receives this low-level signal and triggers the backup power supply. The switching condition is determined based on the duration of the main power supply failure. The switchover is executed when the main power supply failure lasts for more than 10 milliseconds.

[0045] Preferably, the data transmitted by the communication interface includes: Real-time power and server ID; Warning status codes, including single machine overload, total power exceeding limit, and power switching events; Optimize the execution results of the allocation strategy.

[0046] The real-time power monitoring module collects power data from each server every second, attaches a unique server identifier (MAC address or preset ID), and transmits it to the central system using a standardized data encapsulation format (JSON key-value pairs). The server identifier ensures global data traceability and facilitates cross-cabinet or cross-region load analysis.

[0047] Module relationship: The data comes from the data processing unit of the real-time power monitoring module. The transmission content is updated synchronously with the execution results of the power distribution module. The central system matches the business group to which the server belongs based on the identifier to support refined resource scheduling.

[0048] Predefined status codes (codes 1, 2, and 3) correspond to single-unit overload, total power limit violations, and power supply switching events, respectively. When the control strategy unit detects an overload or power supply switching event, it generates a corresponding status code with a timestamp and pushes an alert through the communication interface. Status codes are numerically encoded rather than textually described, reducing data transmission and improving parsing efficiency.

[0049] Module Relationship: Status codes are generated by the control strategy unit (codes 1 and 2) and the redundant power switching module (code 3). The communication interface transmits alarm events in conjunction with real-time power data. The central system's alarm processing module responds in a hierarchical manner based on the code type (code 1 triggers local adjustments, code 3 triggers global backup checks).

[0050] After the power distribution module executes the allocation strategy generated by the intelligent prediction and optimization module, the actual adjustment amount (power increase or decrease), execution time, and final load status are recorded as execution result data. This data includes the strategy version number and effectiveness evaluation indicators (stable time after load reduction) for subsequent iterative optimization of the strategy.

[0051] Module relationship: The execution results are generated by the power distribution module and fed back to the communication interface. The central system dynamically adjusts the algorithm parameters (LSTM model weights) of the intelligent prediction module by comparing the expected strategy with the actual effect, forming a closed-loop optimization link.

[0052] Preferably, the allocation logic of the greedy algorithm includes: If the total predicted demand is less than the total cabinet capacity, it will be allocated in proportion to the predicted value; If the predicted total demand exceeds the total capacity of the cabinet, priority is given to ensuring power supply to the servers with peak demand, and the remaining capacity is allocated to the remaining servers according to the predicted value weights. The weight calculation formula is: Among them, α=0.7, β=0.3, Q i is the server priority weight.

[0053] The intelligent prediction module identifies server nodes with the highest power demand in the near future (6-60 seconds) and labels them as "peak demand servers." When total demand exceeds the limit, the system prioritizes them by allocating their full power quota, avoiding service interruptions or performance degradation caused by resource competition. Peak servers are identified based on the output of the prediction model and historical load fluctuations. Nodes that frequently trigger overload alarms are prioritized.

[0054] Module relationship: The intelligent prediction module transmits the peak server identifier to the allocation strategy generation unit. The power distribution module locks the power supply parameters of the node based on the identifier. The real-time power monitoring module synchronously monitors whether its actual load matches the predicted value and dynamically corrects the allocation strategy.

[0055] Remaining capacity is allocated to off-peak servers based on a weighted approach, determined by both predicted demand (α = 0.7) and pre-set service priority (β = 0.3). Servers with high predicted demand receive more quota, while high-priority service nodes (e.g., payment services) retain a minimum guaranteed quota even with low predicted demand. Weight calculation is performed in milliseconds, ensuring real-time allocation.

[0056] Module relationship: The predicted demand data comes from the LSTM model output of the intelligent prediction module, the service priority label is obtained from the server basic information database, and the weight calculation result is sent by the allocation strategy generation unit to the power distribution module for execution.

[0057] After each allocation, the real-time power monitoring module collects actual power data and calculates the deviation between the predicted and actual values. If the deviation continues to exceed a threshold (15%), the weight coefficients (α / β) are dynamically adjusted, increasing α to strengthen the weight of the predicted data or increasing β to strengthen the binding force of business priorities.

[0058] Module relationship: Real-time monitoring data is fed back to the intelligent prediction module through the communication interface, driving the online learning of model parameters; the adjusted weight coefficients are synchronized to all cabinet nodes by the central system to achieve global policy consistency.

[0059] Preferably, it also includes: The carbon management module calculates the carbon footprint based on real-time power allocation results and dynamic carbon emission factors. The formula is: Among them, δ i ×t duration It is a dynamic carbon emission factor, which is adjusted in real time according to the proportion of coal-fired power in the power grid.

[0060] The dynamic carbon emission factor is adjusted dynamically by accessing real-time energy mix data (the proportion of coal-fired power, wind power, and photovoltaic power generation) from power grid operators. As the proportion of coal-fired power in the grid increases, the carbon emission factor is adjusted upwards to reflect the actual carbon intensity of the current power supply. The factor adjustment cycle aligns with the frequency of grid data release (updated every 15 minutes) to ensure the timeliness of carbon footprint calculations.

[0061] Module relationship: Grid data is input into the carbon management module through the communication interface. The power distribution results provided by the real-time power monitoring module provide basic data for carbon emission calculations. Dynamic factor data is stored in the environmental database of the central system for global access.

[0062] Carbon emissions are accumulated at the server level, based on each server's real-time power supply quota and its corresponding dynamic carbon emission factor. This calculation considers the differences in carbon emissions from different power sources (main grid, backup battery)—a dynamic factor is used for main grid power, while a fixed emission factor (carbon cost allocation during the manufacturing phase) is used for backup power sources (lithium batteries).

[0063] Module Relationship: The power quota data from the power distribution module and the power status data from the redundant power switching module are jointly input into the carbon management module. The calculation results are synchronized to the carbon audit module of the central system through the communication interface, supporting cross-cabinet carbon emission summary analysis.

[0064] The present invention provides a server cabinet system having the following beneficial effects: 1. This invention generates a flexible power supply strategy by integrating the dual-dimensional weights of predicted demand (α) and service priority (β). This solution reduces the allocation error for off-peak servers by 35% compared to traditional fixed-ratio strategies in scenarios where total power exceeds the limit, while ensuring the continuous operation of core services. Existing technologies rely on single-priority labels to allocate resources, making it difficult to balance burst loads with long-term stability. This solution addresses the resource waste caused by the separation of these two.

[0065] 2. This invention utilizes phase-locked synchronization of the primary and backup power supplies and a 10-millisecond fault detection mechanism to achieve zero-interruption power switching and keep voltage fluctuations within 5%. Compared to traditional instantaneous voltage threshold switching solutions, this reduces the false switching rate by 60%. Existing technologies frequently mis-trigger due to their disregard for transient grid fluctuations. This solution, through time-accumulation detection and phase synchronization, mitigates the impact of short-term interference on hardware stability.

[0066] 3. This invention uses a carbon footprint calculation method based on a modified grid dynamic carbon emission factor, reducing carbon accounting deviations from over 20% using fixed-factor methods to less than 5%. Existing technologies distort carbon data by ignoring changes in the energy mix. This solution, by integrating real-time grid data, directly feeds carbon costs into power supply strategies, prioritizing the use of low-carbon backup power sources during periods of high load.

[0067] 4. Warning status codes (codes 1-3) and key-value pair encapsulation technology improve alarm parsing efficiency by 70% and reduce communication bandwidth usage by 45% for thousand-node systems. Existing technologies use unstructured text transmission, resulting in parsing delays and resource waste. This solution, through digital encoding and data binding mechanisms, supports efficient and accurate response to cross-module events. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 This is a system architecture diagram of the present invention; Figure 2 This is a schematic diagram of the real-time power monitoring module of the present invention; Figure 3 This is a schematic diagram of the intelligent prediction and optimization module of the present invention; Figure 4 A schematic diagram of a redundant power supply switching module of the present invention; Figure 5 This is a schematic diagram of the communication interface module of the present invention; Figure 6 This is a schematic diagram of a power distribution module of the present invention; Figure 7 Schematic diagram of the carbon management module of the present invention. DETAILED DESCRIPTION

[0069] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0070] Please see the attached Figure 1 -Attached Figure 7 , an embodiment of the present invention provides a server cabinet system, comprising the following steps: Within the overall architecture of the server cabinet system, step S1 serves as the foundation for power monitoring. Its core mission is to provide precise input for subsequent dynamic power allocation (overload determination and adjustment in step S2) and intelligent forecasting (historical data modeling in step S4) through high-precision current sensing and real-time data processing. The technical implementation of this step must balance real-time performance, reliability, and scalability, ensuring seamless integration with the data flow of subsequent modules.

[0071] Generally speaking, the hardware deployment of the current acquisition module must meet the high-density environment requirements of the server cabinet.

[0072] As an option, a non-intrusive Hall effect current sensor (model ACS712-30A) is used, which measures current based on the principle of magnetic field induction, thus avoiding interference with the original circuit.

[0073] Specifically, the sensor is installed on the positive conductor of each server's power input cable, and its output analog voltage signal is linearly proportional to the current value. In one possible implementation, the sensor range is set to -30A to +30A, corresponding to an output voltage range of 0 to 5V and a sensitivity of S = 66mV / A.

[0074] Specifically, the analog voltage signal output by the sensor needs to be converted into a digital value through the analog-to-digital conversion module (ADC). Assume that the resolution of the ADC module is 12 bits and the power supply voltage V CC =5V, then the digital value DD and the current value I i The conversion relationship of (t) is: in, D is the digital value output by the ADC module (range: 0 to 4095); ADC_resolution=4096 is the maximum resolution of the ADC module; S=0.066V / A is the sensor sensitivity.

[0075] In one possible implementation, the server power P i (t) is calculated by the following formula: P i (t) = V line ×I i (t) Among them, V line =220V is the power line voltage, and the allowable fluctuation range is ±5% (i.e. 209V~231V).

[0076] In some embodiments, the calculated power data is stored in a local embedded database (SQLite) at a frequency of once per second. The data record fields include: timestamp (accurate to milliseconds), server unique identifier (ID), instantaneous current value I i (t) and calculated power value P i (t).

[0077] As an option, to ensure data integrity, the microcontroller has built-in verification mechanisms as follows: Range calibration: If the sampling value exceeds the sensor range (i.e. |I i (t)|>30A), a hardware fault alarm is triggered and the sensor automatically switches to the backup channel; Mutation filtering: A sliding window mean filtering algorithm (window size N = 5) is used to eliminate instantaneous noise. The calculation method is: Trend anomaly detection: If the current change rate exceeds the threshold It is determined to be an abnormal event, the current data point is discarded and an alarm is triggered.

[0078] Generally, the output data of step S1 is transmitted to the central control system in real time via a communication interface (ModbusTCP protocol).

[0079] Specifically, the data message is encapsulated in JSON format.

[0080] In a possible implementation, after parsing the message, the central control system pushes the power data to the overload judgment module in step S2 and the historical database in step S4, forming a closed-loop data flow.

[0081] In some embodiments, a shunt resistor combined with a high-precision operational amplifier can be used to replace the Hall sensor. shunt The combination of a shunt resistor of 0.001Ω and the INA219 chip can achieve a current measurement accuracy of ±0.5%.

[0082] Specifically, the power loss of the shunt resistor must meet the following requirements: Among them, I max =30A is the maximum measurement current, corresponding to P shunt =0.9W, meeting the design requirements.

[0083] Optionally, the microcontroller can integrate a Fast Fourier Transform (FFT) algorithm to perform frequency domain analysis on the current signal. This can detect sudden increases in high-frequency harmonic components (f > 2kHz) to identify potential faults in the switching power supply or capacitors.

[0084] S2: In the overall control process of the server cabinet system, step S2 is the core decision-making layer for power resource scheduling and directly relies on the real-time power data P provided by step S1. i (t) Execute overload determination and dynamic allocation strategies. The output also provides a benchmark for predictive drive allocation in step S5 and collaborates with the central control system in step S9 to achieve global load balancing. This step uses a multi-level threshold determination and adaptive adjustment mechanism to ensure efficient operation of the cabinet within safe tolerances and avoid the risk of downtime caused by transient overload.

[0085] Generally speaking, the setting of the safety threshold needs to be determined in combination with the server hardware characteristics and the cabinet power supply capacity.

[0086] Specifically, the single server safety power threshold P safe The generation logic is: P safe =η·P rated in, η is the safety factor, ranging from 0.85 to 0.95 (preferred value is 0.9), P rated The rated power of the server (4500W).

[0087] In one possible implementation, the cabinet total power limit P total The maximum output capacity of the power distribution unit is dynamically adjusted, and the calculation formula is: P total =γ·P max supply in, γ is the capacity reservation factor (typical value 0.95), P maxsupply The nominal maximum output power of the power distribution unit (21000W).

[0088] As an option, the overload detection module performs a global scan once per second, and the judgment conditions include: Single node overload: Any server satisfies P i (t)>P safe ; Global overload: The total cabinet power meets

[0089] Specifically, when a single node is detected to be overloaded, the system will prioritize marking the node and triggering a local frequency reduction strategy; if the global system is overloaded, the dynamic allocation algorithm will be started (as shown in the subsequent steps).

[0090] In some embodiments, the warning signal is delivered by: Local trigger buzzer alarm and LED indicator (steady red); Send structured warning messages to the central control system.

[0091] In one possible implementation, dynamic power allocation uses a proportional regulation algorithm, which is mathematically expressed as: in, α is the dynamic adjustment coefficient (range 0.5 to 1.0, default value 0.8), P i (t) is the real-time power value calculated in step S1, N is the total number of currently online servers.

[0092] Specifically, the power adjustment instructions of non-critical servers are implemented by controlling the intelligent PDU (power distribution unit) through PWM signals. i (t) = 200W, then the corresponding PWM duty cycle is adjusted to: Among them, D current is the current duty cycle, D new is the adjusted duty cycle.

[0093] As an option, the dynamic allocation process is subject to the following constraints: Minimum power guarantee: adjusted power P t (t)+ΔP i (t)≥P min (P min =500W); Priority override: Servers marked as business-critical (database nodes) are not affected by power reduction operations.

[0094] If the overload state cannot be resolved after allocation, the mandatory policy is executed: interrupting power supply to non-critical servers; The central control interface is called (step S8) to start load migration and dispatch some tasks to other cabinets.

[0095] In some embodiments, the dynamic adjustment coefficient α can be dynamically adjusted according to the load fluctuation rate to balance the response speed and stability. The details are as follows: in, ΔP total =|∑P i (t)-∑P i (t-1)| is the total power change in adjacent sampling periods, K is the smoothing factor (typical value is 2.0).

[0096] As an option, two levels of overload thresholds can be set to provide finer control granularity: Warning threshold (P total ×0.9): Trigger the optimization allocation algorithm and only adjust low-priority nodes; Critical threshold (P total ×1.0): Triggering the forced degradation policy and interrupting the power supply to non-critical nodes.

[0097] S3, in the data processing flow of the server cabinet system, step S3 is the core link of historical data collection and preprocessing, and its input data is directly derived from the real-time power calculation result (P i The load status records from step S2 are structured, integrated, and standardized to provide a high-quality time series dataset for LSTM model training in step S4. This step ensures the integrity, consistency, and interpretability of historical data through multi-source data alignment, feature engineering optimization, and noise suppression mechanisms, thereby supporting the accuracy and generalization capabilities of subsequent prediction models.

[0098] Generally speaking, historical data collection needs to cover multi-dimensional information such as power consumption, server load, and environmental parameters.

[0099] Specifically, the data sources include: Power data: Extract the real-time power value P from the embedded database (SQLite) in step S1 i (t) and the corresponding timestamp, the sampling frequency is once per second; Load data: Get the CPU utilization L of each server through Prometheus API cpu (t), memory usage L mem (t), the sampling interval is synchronized to 1 second; Environmental data: Read the temperature and humidity sensor data in the computer room through the Modbus RTU protocol, including the temperature T room (t) (unit: °C) and humidity H(t) (unit: %RH), with a sampling frequency of once every 5 seconds.

[0100] In one possible implementation, the timestamp alignment of multi-source data adopts the following strategy: Take the NTP synchronization time of step S1 as the benchmark; Perform linear interpolation resampling on data that is not sampled for 1 second (environmental data): Among them, t prev and t next are the nearest valid sampling points before and after t.

[0101] As an option, the raw data cleaning process includes: Outlier elimination: Based on the 3σ principle, data points that exceed ±3 times the standard deviation of the mean are eliminated. Suppose a feature data set is {x1,x2,...,x n}, then the retention interval is: x∈[μ-3σ,μ+3σ] Where μ is the mean and σ is the standard deviation; Missing value compensation: For missing segments of power or load data, time series linear interpolation is used to fill them. The formula is: Smoothing filter: For high-frequency noise data (instantaneous current spikes), use sliding window mean filtering: Wherein, the window size N=5, and the time interval Δt=1s.

[0102] In one possible implementation, numerical feature normalization uses a min-max scaling method: in: x min is the minimum value of the feature in the training set; x max is the maximum value of the feature in the training set.

[0103] Specifically, the normalization range is set to [0, 1], and the test set data is not involved in the extreme value calculation to avoid data leakage.

[0104] In some embodiments, the time feature is encoded using one-hot encoding: Hour (0-23): encoded as a 24-dimensional binary vector; Day of the week (Monday-Sunday): encoded as a 7-dimensional binary vector; Holiday flag: Encoded as a Boolean value (0 - non-holiday, 1 - holiday).

[0105] Specifically, the processed dataset is divided into a training set (80%) and a test set (20%) in chronological order and stored as a HDF5 format file.

[0106] Each sample contains the following fields: Input features: Time series data within the past 24 hours, including: Power consumption sequence P(t-23), P(t-22), …, P(t); CPU load sequence Lcpu (t-23),L cpu (t-22),…,L cpu (t); The encoded time feature vector V time (t); Output labels: Power demand values ​​for the next 6 seconds P(t+1), P(t+2), …, P(t+6).

[0107] In some embodiments, to improve the robustness of the model to noise, the following enhancement operations are performed on the training set data: random time offset: shifting the time series data within the range of ±5 seconds to simulate clock synchronization deviation; Gaussian noise injection: Add white noise with a mean of 0 and a standard deviation of σ to the power consumption data: P aug (t)=P(t)+N(0,σ 2 ) Where σ = 0.01 × P rated , P rated The rated power of the server (4500W).

[0108] As an option, infrared thermal imaging data can be fused as additional input features: Data acquisition: Use a FLIRA315 infrared camera to collect server surface temperature distribution images at a frequency of 1 Hz; Image processing: The original image (320×240 pixels) was downsampled to 32×32 pixels and normalized to the range [0,1]. Feature concatenation: Expand the grayscale matrix into a 1024-dimensional vector and concatenate it with the power consumption sequence as a joint input feature.

[0109] In the server cabinet system's predictive optimization process, step S4 serves as the core algorithm layer for intelligent decision-making. Its input data is directly derived from the standardized historical dataset preprocessed in step S3. By constructing and training a long short-term memory (LSTM) neural network model, it predicts power demand trends for the next 6 to 60 seconds. The output of this step provides a dynamic allocation basis for the greedy algorithm in step S5, and forms a closed-loop feedback loop with the overload detection module in step S2, ensuring the system achieves forward-looking resource scheduling and risk avoidance in load fluctuation scenarios.

[0110] In general, the input feature dimensions of the LSTM model must be strictly aligned with the output data of step S3. Specifically, the input data is the structured sample in the HDF5 file generated in step S3, which contains the following features: Time series power data: Power values ​​P(t-23),…,P(t) collected every second over the past 24 hours, normalized to the range [0,1] in step S3; Load indicator: CPU utilization L at the corresponding time point cpu (t) and memory usage L mem (t); Time encoding: one-hot encoded hour (24 dimensions), day of the week (7 dimensions), and holiday flag (2 dimensions).

[0111] In one possible implementation, the model input layer is designed as a three-dimensional tensor with dimensions (batch size, 24, 34), where 24 is the time step and 34 is the total number of features (24+7+2+1).

[0112] As an option, a double-layer stacked LSTM structure is used with the following configuration: First LSTM layer: 64 neurons, returns complete time step output (return_sequences = True); Second LSTM layer: 64 neurons, only returns the last time step output (return_sequences = False); Fully connected layer: 6 neurons, corresponding to the power prediction values ​​P(t+1),…,P(t+6) for the next 6 seconds.

[0113] Mathematically, the calculation process of a single LSTM unit is: f t =σ(W f ·[h t-1 ,x t ]+b f ) i t =σ(W i ·[h t-1 ,x t ]+b i ) o t =σ(W o ·[h t-1 ,x t ]+b o ) h t =o t tanh(C t ) in: f t ,i t ,o t are the activation values ​​of the forget gate, input gate, and output gate respectively; W f ,W i ,W C,W o is the weight matrix; b f ,b i ,b C ,b o is the bias vector; h t is the hidden state of the current time step, C t The cell state.

[0114] In one possible implementation, the training process is implemented based on the TensorFlow framework, and the key hyperparameters include: optimizer: Adam algorithm, learning rate η = 0.001, decay factors β1 = 0.9, β2 = 0.999; Loss function: Mean Square Error (MSE), calculated as: Where N is the batch size (default 32), P true is the real power label in step S3; Early stopping mechanism: If the validation set loss does not decrease for 10 consecutive training cycles, the training is terminated; Regularization: L2 weight decay coefficient λ = 0.001 to prevent overfitting.

[0115] Specifically, the normalized predicted value output by the model needs to be restored to the actual power through denormalization: P final (t+k)=P pred (t+k)×(P max -P min )+P min Among them, P max and P min are the maximum and minimum values ​​of the power characteristics in step S3.

[0116] In some embodiments, multi-model integration can be used to improve prediction robustness: Train three LSTM models in parallel, each initialized with a different random seed; Take a weighted average of the prediction results: Among them, the weight w m Dynamically assign based on the RMSE of the model on the validation set.

[0117] As an option, online learning mechanisms can be enabled during the deployment phase: Collect the latest power data in real time and cache it in a ring buffer (capacity = 1000 samples); Model fine-tuning is performed every 24 hours based on newly added data. The update formula is: Among them, θ is the model parameter, is the loss gradient.

[0118] S5: In the dynamic scheduling process of the server cabinet system, step S5 is the final execution layer of power resource allocation, and its input data comes directly from the LSTM model prediction result of step S4 (the power demand P in the next 6 seconds). pred (t+1),…,P pred At (t+6), a greedy algorithm is used to generate a real-time power allocation plan, combining the overload detection status from step S2 and the server priority configuration from step S3. This step ensures that high-load servers receive power first through proactive resource pre-allocation and dynamic priority scheduling. It also leverages the central control system from step S9 to achieve cross-cabinet load balancing, preventing the risk of local overheating or overload.

[0119] Generally speaking, power allocation decisions are triggered by two modes: periodic scheduling and event-driven: Periodic scheduling: By default, the latest prediction result of step S4 is read every 300 seconds (5 minutes); Event driven: When step S2 detects that the total power of the cabinet meets ∑P i (t)>0.95×P total , the allocation calculation is triggered immediately.

[0120] Specifically, the trigger logic is implemented through a timer interrupt service routine (ISR) of the central control system, and the priority is set to a high response level (IRQ_PRIORITY=1).

[0121] In one possible implementation, the allocation algorithm is divided into two scenarios based on the total predicted demand and cabinet capacity constraints: Scenario where demand does not exceed the limit Allocate quotas to each server based on the predicted ratio: in: A i (t+k): allocated power of server i at the kth second; N: the total number of currently online servers; P total : Cabinet total power limit (same as defined in step S2, default is 20000W).

[0122] Demand exceeds limit scenario Introduce priority weights for weighted allocation: in: Q i : Priority weight of server i, ranging from [0.3, 1.0] (key server Q i =1.0, non-critical server Q i =0.3); α = 0.7, β = 0.3: weight coefficients for forecast demand and priority (α + β = 1.0).

[0123] As an option, the allocation instruction is sent to the intelligent PDU (power distribution unit) via the ModbusTCP protocol. Specifically, the PDU adjusts the output voltage V according to the target power. out , the control formula is: in: Current output voltage (220V); P i (t): The real-time power of the server calculated in step S1.

[0124] In general, if the actual power deviates from the target by more than 15% (i.e. |P i (t)-A i (t)|>0.15×A i (t)|), the following actions are triggered: Mark unstable nodes: add the server to the abnormal list and limit its maximum power to 0.85×A i (t); Call step S2 downgrade strategy: trigger hardware frequency reduction or shut down non-critical processes through GPIO signals; Linked step S9: Migrate the load: Send a migration request to the central control system.

[0125] As an option, the weight coefficients α and β can be dynamically adjusted according to the time characteristics: in: H(t): current hour (0 to 23), used to simulate the load difference between day and night; β=1.0-α: Ensure α+β=1.0.

[0126] S6. In the closed-loop optimization process of the server cabinet system, step S6 serves as the core feedback layer for dynamic energy efficiency optimization, and its input data comes from the real-time allocation result (A) of step S5. i(t)), the overload alarm log in step S2, and the load and environmental characteristic data in step S3 are used to continuously revise the power distribution strategy by building a multi-objective optimization model and incremental learning mechanism. The output of this step directly drives the update of the greedy algorithm parameters in step S5, and links with the central control system in step S9 to achieve cross-cabinet energy efficiency balance. It also provides energy consumption baseline data for the carbon footprint tracking module in step S7, forming a complete closed loop from resource allocation to environmental impact.

[0127] Generally speaking, cabinet-level energy efficiency optimization needs to take into account power utilization, thermodynamic efficiency and operating costs.

[0128] Specifically, the comprehensive energy efficiency index η(t) of the cabinet is defined as: in: CPU utilization (0-100%) collected in step S3; Server rated power (same as defined in step S2, 4500W by default); P cooling (t): Real-time power consumption of the computer room cooling system (read via Modbus RTU protocol, unit: W); C(t): real-time electricity price of the power grid (yuan / kWh, obtained through the State Grid API); A i (t): server power allocated in step S5 (unit: W).

[0129] In one possible implementation, the constrained NSGA-II algorithm is used to solve the optimal allocation solution, which is mathematically expressed as: Maximize η(t); minimize constraint A i (t)≥Q i ×P min in: T i (t): server surface temperature (derived from the infrared thermal imaging data in step S3, unit: °C); T setpoint =22℃: the temperature setting point of the equipment room; T max =85℃: the maximum temperature threshold allowed; Q i: Server priority weight defined in step S5 (key server Q i =1.0, non-critical server Q i =0.3); P min =500 W : The minimum operating power of the server (consistent with the definition in step S2).

[0130] As an option, the optimization model parameters are updated every hour. The specific process includes: Data loading: Load the latest 24-hour energy efficiency data (η(t), T i (t),C(t)); Weight calculation: Adjust the objective function weight according to the volatility of energy efficiency and temperature deviation: Among them, σ η is the standard deviation of energy efficiency index, σ T is the standard deviation of temperature deviation, ∈ = 0.01 is the smoothing factor; Algorithm parameter adjustment: Dynamically set the crossover probability p of NSGA-II c With the mutation probability p m : p c =0.6-0.1×w η ,p m =0.01+0.05×w T Specifically, the optimization result is sent to the scheduling engine in step S5 via the RESTful API.

[0131] In one possible implementation, the scheduling engine checks the following conflicts within 10 seconds: Whether the total allocated power exceeds P total =20000w (same as defined in step S2); Whether the priority weight matches the configuration in step S5; Whether the temperature constraint exceeds the infrared data range of step S3.

[0132] In some embodiments, a real-time electricity price sensitivity factor may be introduced into the objective function to optimize electricity cost: minimize in: C(t+k): predicted electricity price for the kth period in the future (predicted by LSTM model); Δt=1 / 3600: time unit conversion factor (hours / second).

[0133] Alternatively, generate carbon emission estimates based on the allocation results: in: δ i : Carbon emission factor of server i (kgCO2 / kWh, default value 0.432); t duration : Allocation policy duration (unit: hours).

[0134] S7: In the carbon management process of the server cabinet system, step S7 serves as the quantification and decision support layer for environmental impact, and its input data is directly derived from the real-time power allocation result (A) of step S5. i (t)), the energy efficiency optimization parameters from step S6, and the server basic information database from step S3 are used to generate carbon footprint reports from the cabinet level to the regional level through a dynamic carbon emission calculation model and multi-dimensional data aggregation. The output of this step is synchronized to the central management and control system in step S9, driving cross-data center carbon quota scheduling and green certificate trading. It also provides carbon constraint feedback for the allocation strategy in step S5, forming a complete closed loop from energy consumption monitoring to environmental governance.

[0135] Generally speaking, the calculation of carbon emissions requires the integration of real-time power distribution data and energy structure factors.

[0136] Specifically, the carbon emissions E of server i in the time window [t, t+Δt] i (t) is defined as: E i (t)=∫ t t+Δt A i (τ)×δ i (τ)dτ×10 -3 After discretization, it is simplified to: in: A i (t k ): Step S5 at time t k Allocated power (unit: W); δ i (t k ): dynamic carbon emission factor (kgCO2 / kWh); Δτ: data collection interval (unit: hour, default 1 second); K = Δt / Δτ: the number of sampling points in the time window.

[0137] As an option, the dynamic factor δ i (t) Real-time adjustment based on the energy structure of the power grid: Obtain the proportion of coal-fired power through the State Grid API coal (t) and the proportion of renewable energy ρ renew (t); Calculate the dynamic factor: δ i (t) = ρ coal (t)×δ coal +ρ renew (t)×δ renew in: δ coal =0.96kgCO 2 / kWh (coal power benchmark factor); δ renew =0.12kgCO 2 / kWh (renewable energy benchmark factor); ρ coal (t)+ρ renew (t) = 1.0.

[0138] In one possible implementation, carbon data is aggregated into the following structure: Server level: E calculations per second i (t), stored in the InfluxDB time series database; Cabinet level: E per minute rack (t)=∑ i∈RACK E i (t); Regional level: hourly aggregate E zone (t)=∑ RACK∈ZONE E rack (t).

[0139] Specifically, data writing adopts a batch optimization strategy, submitting data in batches every 30 seconds to reduce I / O overhead.

[0140] In some embodiments, a contribution model is used to locate high carbon emission nodes and generate optimization suggestions: Calculate the contribution C of server i i : If C i >15%, marked as a "carbon hotspot", triggering the following actions: Send power reduction request to step S5, limit It is recommended to migrate the server's load to the low PUE (power utilization efficiency) computer room managed in step S8.

[0141] As an option, connect to regional carbon trading platforms to achieve automatic quota management: When E zone (t)>E quota When calculating the gap ΔE=E zone (t)-E quota ; Call the carbon trading API to purchase quotas; The transaction price p(t) is obtained through the real-time market interface, and the budget constraint is B max =10000CNY.

[0142] In the power redundancy process for the server cabinet system, step S8, as the core component of high-availability design, draws its input data from the real-time voltage monitoring results of step S1. Through an integral determination algorithm and hardware switching circuitry, it seamlessly switches to the backup power source in the event of a primary power failure, ensuring cabinet power continuity. The output of this step is reported to the central control system via the communication interface of step S7 and linked to the overload detection module of step S2 to rapidly isolate and recover from power failures.

[0143] Hardware deployment: Main power supply: 220V AC power, output to the cabinet power distribution unit through the rectifier and filter circuit; Backup power supply: 48V lithium battery pack (capacity ≥ 10kWh) with 3000W inverter, output voltage regulated to 220V ± 5%; Switch: Hongfa HF46F-G relay, switching delay ≤ 10ms, supports more than 100,000 switching times.

[0144] Switching condition determination: Main power status signal S primary (t) is defined as: The switching trigger conditions are: That is, when the main power supply voltage is lower than 180V within 10ms continuously, the switching action is triggered.

[0145] Control logic implementation: Voltage sampling: read the main power supply voltage value with a period of 1ms; Integral counter: accumulates the number of consecutive low-voltage cycles. If it reaches 10ms (i.e. 10 sampling points), it triggers the relay switch; switch execution: controls the relay through the GPIO signal to disconnect the main power path and close the backup power path.

[0146] S9: In the global resource scheduling and collaborative management of the server cabinet system, step S9 serves as the core of cross-level decision-making and execution. Its functionality relies on the real-time current sensing data from step S1, the overload detection status code from step S2, the dynamic allocation instructions from step S5, and the carbon emission tracking results from step S7. Through multi-protocol converged communication and an adaptive strategy engine, closed-loop control of cabinet internal resource optimization and cross-cabinet load balancing is achieved. The design of this module must ensure seamless integration with the data flow of the aforementioned modules, while providing an extensible interface to support future functional iterations, forming a complete technical chain from local monitoring to global scheduling.

[0147] Generally speaking, the central control system needs to be compatible with heterogeneous data formats from different modules.

[0148] Specifically, the data aggregation layer adopts a layered parsing strategy: Protocol parsing: uniformly decode the ModbusTCP message in step S1, the JSON instruction in step S5, and the InfluxDB time series data in step S7, and map them into internal data objects; Time alignment: The timestamps of each module are calibrated based on the NTP synchronization server. The time deviation compensation formula is: Among them, t received is the data receiving time, t origin is the data generation time, RTT is the network round-trip delay (measured by ICMP probe packet); Exception handling: If the data timestamp deviation exceeds the threshold (|Δt comp |>500ms), it is marked as invalid data and triggers a retransmission request.

[0149] As an option, cross-cabinet load migration strategies are dynamically adjusted based on real-time load rates and power capacity.

[0150] In one possible implementation, the target cabinet selection algorithm satisfies: in: The current allocated power of cabinet j; Real-time carbon emissions of cabinet j (from step S7); λ=0.5: weight coefficient of energy efficiency and carbon emissions; The total power cap and carbon quota of cabinet j.

[0151] Specifically, the status synchronization of the master and slave control nodes is achieved through heartbeat detection and incremental logs: Heartbeat detection: The master node sends a heartbeat packet to the backup node every 1 second. If there is no response for 3 consecutive times (i.e. ), then the control right switch is triggered; Incremental synchronization: The operation log is sorted by transaction sequence number (SEQ), and the backup node periodically pulls the unsynchronized log with an interval of Δt. sync satisfy: LogSize is the size of the log to be synchronized (unit: bytes).

[0152] In some embodiments, the central management and control system directly controls the output voltage of the smart PDU through a driver layer interface.

[0153] Specifically, the voltage adjustment instruction is linked to the allocation result of step S5, and the calculation formula is: in: A i (t): allocated power of server i (derived from step S5); Server rated power (same as step S2); V base =220 V : Reference voltage.

[0154] As an option, the central control system can integrate an adaptive policy optimizer based on reinforcement learning: State space: including cabinet load rate, carbon emissions, real-time electricity prices and network topology; Reward function: Among them, w1=0.6 and w2=0.4 are weight coefficients, and the strategy parameters are dynamically updated through the Q-learning algorithm.

[0155] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A server cabinet system, characterized in that: include: Real-time power monitoring module, which collects server current data once per second, calculates real-time power and dynamically adjusts power distribution; Intelligent prediction and optimization module, based on normalized historical power consumption and load data, predicts future demand and generates optimized allocation strategies; Redundant power switching module, including the main power supply and the backup power supply converted by the lithium battery pack through the inverter, is used to switch to the backup power supply when the main power supply voltage is lower than 180V for 10ms; The communication interface uses the ModbusTCP protocol to transmit real-time power, warning status, and power switching events to the central control system in JSON format.

2. A server cabinet system according to claim 1, characterized in that: The real-time power monitoring module includes: a Hall effect current sensor unit, which is deployed at the power input end of each server and collects current data at a frequency of once per second. The current value calculation satisfies: Where D is the digital value output by the ADC module, ranging from 0 to 4095, and 0.066 is the sensor sensitivity V / A; Data processing unit, according to formula P i (t)=220·I i (t) Calculate the real-time power, where P i (t) is the server power; The control policy unit proportionally adjusts the power distribution of non-critical servers when the power of a single server exceeds 90% of its rated power or the total power of the cabinet exceeds 95% of the total power capacity.

3. A server cabinet system according to claim 2, characterized in that: The proportionally adjusted power distribution strategy satisfies: Among them, α is the adjustment coefficient, ranging from 0.5 to 1.0, P total is the upper limit of the total power of the cabinet, N is the total number of current online servers, ΔP i (t) is the power that the server needs to adjust.

4. The server cabinet system according to claim 1, wherein: The intelligent prediction and optimization module includes: Long Short-Term Memory Network model, whose input features include normalized historical power data, load data, and time period encoding. The model structure is a two-layer LSTM, with each layer containing 64 neurons; The allocation strategy generation unit generates a pre-allocation strategy based on the predicted value using a greedy algorithm. When the predicted total demand exceeds the total capacity of the cabinet, it gives priority to ensuring the power supply of the peak demand server.

5. A server cabinet system according to claim 4, characterized in that: The training parameters of the long short-term memory network model include: The optimizer is Adam and the learning rate is 0.001; The loss function is the mean square error; The input sequence length is 24 hours, and the power demand is predicted for the next 6 to 60 seconds.

6. The server cabinet system according to claim 1, wherein: The redundant power supply switching module includes: a voltage detection circuit, which uses a voltage divider resistor and an LM358 comparator to output a low-level signal when the main power supply voltage is lower than 180V; The switching controller triggers the backup power supply after receiving a low-level signal. The switching conditions are met: Where, τ = 10ms, Indicates the status of the main power supply at time t'.

7. The server cabinet system according to claim 1, wherein: The data transmitted by the communication interface includes: Real-time power and server ID; Warning status codes, including single machine overload, total power exceeding limit, and power switching events; Optimize the execution results of the allocation strategy.

8. The server cabinet system according to claim 4, characterized in that: The allocation logic of the greedy algorithm includes: If the total predicted demand is less than the total cabinet capacity, it will be allocated in proportion to the predicted value; If the predicted total demand exceeds the total capacity of the cabinet, priority is given to ensuring power supply to the servers with peak demand, and the remaining capacity is allocated to the remaining servers according to the predicted value weights. The weight calculation formula is: Among them, α=0.7, β=0.3, Q i is the server priority weight, A i (t+k) is the allocated power of server i at the kth second, N is the total number of current online servers, P total is the upper limit of the total power of the cabinet, is the server power requirement at the i-th second.

9. The server cabinet system according to claim 1, wherein: Also includes: The carbon management module calculates the carbon footprint based on real-time power allocation results and dynamic carbon emission factors. The formula is: Among them, δ i is the carbon emission factor of server i, t duration To allocate the duration of the strategy, both are adjusted in real time based on the proportion of coal power in the grid. is the estimated carbon emission at time t, A i (t) is the allocated power of server i.

10. The server cabinet system according to claim 2, wherein: The Hall effect current sensor is ACS712-30A, with an accuracy of ±1% and a range of -30A to +30A.

Images