Intelligent system for computing power center equipment power supply of knowledge graph

Abstract

The invention belongs to the technical field of crossing of artificial intelligence and computer system architecture, particularly relates to a computing power center equipment power supply intelligent system of a knowledge graph, and aims to solve the problem of low energy efficiency caused by extensive power supply management, unbalanced energy consumption and lagged load response. The system comprises a task analysis module, a dynamic load prediction module, a multi-source power supply scheduling module, an equipment-level power supply control module and a feedback optimization module, through load prediction and multi-source power supply cooperative scheduling driven by task semantics, voltage and current dynamic adjustment and sleep wake-up control are achieved, and a strategy is optimized online based on operation feedback. The scheme improves energy efficiency, equipment life and system stability.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and computer system architecture, specifically relating to an intelligent power supply system for knowledge graph computing center equipment. Background Technology

[0002] In the field of artificial intelligence and big data infrastructure, computing centers, as the core carriers supporting massive data processing and complex model training, directly affect the reliability and economy of the overall system through their operational stability and energy efficiency. With the exponential growth in computing power demand, the scale of computing center equipment continues to expand. As a critical infrastructure ensuring the continuous operation of equipment, the power system must not only meet the requirements of high availability and high redundancy, but also possess the ability to respond quickly to dynamic load changes. Among these requirements, power management has gradually evolved from the traditional passive power supply mode to an intelligent system integrating sensing, prediction, and control.

[0003] The introduction of knowledge graph technology provides a new technical approach for power management of computing center equipment. By structurally expressing multi-dimensional information such as equipment attributes, operating status, environmental parameters, and historical faults, knowledge graphs can construct semantic cognitive models of equipment power systems, thereby supporting more accurate energy consumption prediction, fault warning, and resource scheduling decisions. This technological direction aims to realize a paradigm shift in power systems from "response-based power supply" to "cognitive-driven power supply," improving overall energy utilization efficiency and the level of intelligent operation and maintenance.

[0004] Existing technologies for power management in computing centers still have significant shortcomings: power scheduling strategies generally lack a deep understanding of the semantic relationships and operational context of devices, making it difficult to achieve knowledge-driven dynamic power allocation; information silos exist between device status awareness data and power control commands, leading to delayed or misjudged adjustments; simultaneously, traditional power systems cannot effectively integrate historical operation and maintenance experience with real-time operational data, making it difficult to make collaborative optimization decisions in the event of sudden load surges or localized failures. These problems are particularly prominent in computing center environments with high-density deployments and high-concurrency tasks, easily leading to localized overloads, energy efficiency degradation, and even service interruptions. Therefore, there is an urgent need for an intelligent power system for computing center devices that integrates knowledge graphs and intelligent decision-making capabilities. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent power supply system for computing center equipment in knowledge graph computing scenarios, in order to solve the technical problems of low energy efficiency, shortened equipment life and insufficient system stability caused by crude power management, unbalanced energy consumption structure and delayed load response in current large-scale knowledge graph computing scenarios.

[0006] The technical solution of this invention is an intelligent power supply system for knowledge graph computing center equipment, comprising a knowledge graph task parsing module, a dynamic load prediction module, a multi-source power supply scheduling module, an equipment-level power control module, and a feedback optimization module. The knowledge graph task parsing module receives and parses the knowledge graph computing tasks to be executed, extracting the task's graph size, node type distribution, relationship complexity, real-time requirements, and historical execution characteristics. The dynamic load prediction module, based on the task characteristics output by the knowledge graph task parsing module, combines current computing center equipment operating status data, environmental temperature and humidity data, and historical load curves to construct a time-series prediction model and output the predicted equipment-level power load value within a future preset time window. The multi-source power supply scheduling module is used to adjust the dynamic load prediction based on the dynamic load characteristics. The load prediction module outputs the predicted power load value and schedules at least two power sources from the mains power, energy storage battery bank, photovoltaic power, and backup diesel generator to generate a power supply combination strategy that meets the load requirements and is cost-optimized. The device-level power control module receives the power supply combination strategy output by the multi-source power supply scheduling module and dynamically configures the power supply voltage, current limit, sleep threshold, and wake-up delay parameters of each computing node, storage unit, and network switching device according to the hardware topology and power interface specifications of the device. The feedback optimization module collects the actual power consumption, temperature changes, task completion time, and power switching events during device operation in real time and sends the data back to the dynamic load prediction module and the multi-source power supply scheduling module for online correction of prediction model parameters and power supply scheduling strategy.

[0007] Furthermore, the knowledge graph task parsing module includes a task semantic analysis submodule and a task feature vector submodule. The task semantic analysis submodule is used to parse the input knowledge graph task using natural language or structured query language, and to identify the entity types, relation path lengths, inference depths, and concurrent query counts involved in the task. The task feature vector submodule is used to quantize the identification results into multidimensional feature vectors, which include the number of graph nodes, edge density, average degree, query complexity index, and real-time level encoding.

[0008] Furthermore, the dynamic load prediction module adopts a time-series prediction architecture based on a long short-term memory network. Its inputs include the multi-dimensional feature vector, the current CPU utilization of the device, the GPU memory utilization, the memory bandwidth utilization, the disk I / O throughput, and the ambient temperature of the data center. The dynamic load prediction module outputs power load prediction values ​​for three time scales: 5 minutes, 15 minutes, and 60 minutes, and assigns a confidence interval weight to each time scale.

[0009] Furthermore, the multi-source power supply scheduling module incorporates a cost optimization engine. This engine constructs a mixed-integer linear programming model based on real-time electricity price signals, remaining battery power, predicted photovoltaic power generation, and diesel generator start-up costs. The mixed-integer linear programming model uses minimizing total power supply cost as the objective function and constrains equipment load demand, maximum power output, battery charge / discharge rate limits, and grid connection capacity to solve for the optimal power output allocation scheme for each power source within the scheduling cycle.

[0010] Furthermore, the device-level power control module includes a voltage regulation unit, a current limiting unit, a sleep management unit, and a wake-up scheduling unit. The voltage regulation unit dynamically adjusts the core power supply voltage of the CPU and GPU according to the task computation intensity to achieve dynamic voltage frequency scaling. The current limiting unit sets the maximum output current of each power rail to prevent hardware protection shutdown caused by instantaneous overcurrent. The sleep management unit switches non-critical computing nodes to a low-power sleep state during task idle periods and records their context state. The wake-up scheduling unit wakes up sleep nodes in a preset order according to task priority and node dependency after receiving a new task scheduling instruction, ensuring the continuity of task execution and data consistency.

[0011] Furthermore, the feedback optimization module includes an anomaly detection submodule and a strategy correction submodule. The anomaly detection submodule is used to compare the deviation between the actual power consumption and the predicted load. When the deviation exceeds a preset threshold, an anomaly event is triggered. The strategy correction submodule is used to use the anomaly event marker, task completion delay data, and power switching count as feedback signals, and to update the network weights in the dynamic load prediction module using an online gradient descent algorithm, and adjust the constraint relaxation coefficient of the cost optimization engine in the multi-source power supply scheduling module.

[0012] Furthermore, the system operates in a hierarchical control architecture, which includes a global scheduling layer, a regional coordination layer, and a device execution layer. The global scheduling layer is deployed on the main control server of the computing center and is responsible for task allocation and power policy generation across racks. The regional coordination layer is deployed on the edge controllers of each rack and is responsible for the aggregation of local power device status and instruction forwarding. The device execution layer is integrated into the baseboard management controller of each computing node and is responsible for executing specific voltage, current, and sleep control instructions.

[0013] Furthermore, the energy storage battery pack uses lithium iron phosphate batteries and is equipped with a battery health status monitoring unit; the battery health status monitoring unit is used to collect the battery's internal resistance, charge-discharge cycle count and temperature gradient in real time, calculate the remaining service life, and use the remaining service life as a correction factor for the battery's available capacity in the multi-source power supply scheduling module.

[0014] Furthermore, the photovoltaic power source is connected to the system through a maximum power point tracking controller. The maximum power point tracking controller dynamically adjusts the operating point according to the light intensity and component temperature to ensure that the photovoltaic power generation efficiency is maintained above 95%. The photovoltaic power generation prediction value is generated by the weather sensors deployed on the roof and historical irradiance data, and is updated to the multi-source power supply scheduling module at a cycle of 10 minutes.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: This invention achieves end-to-end intelligent scheduling from task semantics to power strategy by constructing a mapping relationship between knowledge graph task features and power load. This overcomes the shortcomings of traditional computing center power management, which relies solely on hardware monitoring indicators and ignores upper-layer application semantics. The dynamic load prediction module integrates task features and equipment status, significantly improving the accuracy and timeliness of load prediction and providing a reliable basis for power scheduling. The multi-source power supply scheduling module introduces a mixed-integer linear programming model, achieving global optimization of power supply costs while ensuring task execution, effectively reducing the operating energy consumption of the computing center. The device-level power control module, through fine-tuning voltage, current, and sleep parameters, not only extends the service life of hardware devices but also avoids computational interruptions caused by power fluctuations. The closed-loop control mechanism formed by the feedback optimization module enables the system to have continuous learning and adaptive capabilities, meeting the challenges of the ever-evolving complexity of knowledge graph tasks. The overall system achieves synergy between global optimization and local execution under a layered architecture, balancing scheduling efficiency and control accuracy, and providing a high-efficiency, high-reliability, and highly intelligent power support system for large-scale knowledge graph computing. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention; Figure 2 This is a schematic diagram of the core principle framework of dynamic load prediction and multi-source power supply scheduling collaborative optimization in this invention. Detailed Implementation

[0017] Example 1 Please refer to Figure 1 and Figure 2This invention provides an intelligent power supply system for knowledge graph computing center equipment. Its core purpose is to solve the technical problems of low energy efficiency, shortened equipment life and insufficient system stability caused by crude power management, unbalanced energy consumption structure and load response lag in large-scale knowledge graph computing scenarios. By tightly coupling the abstract knowledge graph computing tasks with the underlying power management strategy, end-to-end intelligent optimization from task semantics to power configuration is achieved. The system runs in a hierarchical control architecture, which includes a global scheduling layer, a regional coordination layer and a device execution layer, ensuring the coordinated unity of scheduling efficiency and control accuracy.

[0018] This system mainly includes a knowledge graph task parsing module, a dynamic load prediction module, a multi-source power supply scheduling module, a device-level power control module, and a feedback optimization module. The system's workflow begins with the knowledge graph task parsing module receiving and deeply analyzing the knowledge graph computation task to be executed, extracting its inherent graph characteristics and computational requirements. The parsing results are then transmitted to the dynamic load prediction module, which combines the current equipment operating status and environmental data to construct a high-precision time-series prediction model, generating device-level power load prediction values ​​for future time windows. Based on these prediction values, the multi-source power supply scheduling module intelligently schedules at least two power sources from mains power, energy storage battery banks, photovoltaic power, and backup diesel generators, considering cost, energy consumption, and reliability, to form an optimal power supply combination strategy. The device-level power control module receives this power supply strategy and performs refined dynamic configuration of the power supply parameters for each computing node, storage unit, and network switching equipment within the computing center. Simultaneously, the feedback optimization module collects real-time equipment operating data and task completion status, sending these as feedback signals back to the dynamic load prediction module and the multi-source power supply scheduling module, forming a closed-loop optimization mechanism that enables the system to continuously learn and adapt.

[0019] The knowledge graph task parsing module aims to efficiently and accurately understand the deep semantic and structural features of the knowledge graph computing task to be executed, providing a precise task profile for subsequent intelligent power scheduling. This module is deployed on the main control server of the global scheduling layer in the computing center, serving as the first entry point into the task processing flow. Its primary responsibility is to receive task requests from users or upper-layer applications. Task requests are typically presented in the form of natural language queries, structured query languages ​​such as SPARQL and Cypher, or domain-specific graph query languages. The receiving interface supports multiple standard data exchange formats, such as JSON, XML, or Protobuf. Upon receiving a task request, the data is encapsulated into an internal task message packet, containing a unique task identifier, request timestamp, user credentials, and the original query text. The task message packet is then forwarded to the task semantic analysis submodule for further processing. The task semantic analysis submodule's main function is to perform comprehensive semantic analysis on the input knowledge graph task. When the task is in natural language form, the submodule employs a deep learning-based natural language processing model, such as the Transformer architecture model, to perform word segmentation, part-of-speech tagging, named entity recognition, and relation extraction on the query statement, transforming unstructured text into a structured semantic representation. This model has been pre-trained on large-scale knowledge graph query logs and manually annotated data, enabling it to identify the core entity types involved in the task, such as people, places, events, and concepts; the length of the relation path, such as the number of relation hops required to obtain the result; the depth of reasoning, such as whether multiple steps of logical reasoning or ontology reasoning are required to reach a conclusion; and the number of concurrent queries. For example, how many identical or similar task requests are made within the same time period; for tasks using structured query language or graph query language, the submodule uses a syntax parser and semantic analyzer to convert them into an abstract syntax tree or logical query plan, from which key information related to the graph structure is extracted; during the parsing process, the submodule maintains a task type and feature mapping table to map different types of query statements to predefined feature dimensions; for example, an entity recognition query task may only require shallow entity matching, while a multi-hop inference query requires traversing a broader graph structure; the parsing results are output in the form of structured metadata, including a task attribute dictionary, where key-value pairs contain the list of identified entities, the list of relationships, the inference depth value, and the estimated concurrency.

[0020] The task feature vector submodule quantizes the structured metadata output by the task semantic analysis submodule into a unified format of multi-dimensional feature vectors. This submodule maintains a feature mapping dictionary, which defines all possible quantized feature dimensions and their encoding rules. For example, identified entity types are mapped to one-hot encodings or embedding vectors; relation path lengths are quantized as integer values; inference depth can be encoded as discrete levels, such as 0 for simple queries, 1 for single-step inference, and 2 for multi-step inference; the number of concurrent queries is directly converted to a numerical value. In addition, the submodule also calculates the number of knowledge graph nodes (the number of nodes in the knowledge graph subgraphs that the task may access), edge density (the ratio of the number of edges to the square of the number of nodes in the subgraph), and average degree. The number of nodes in the subgraph is the average number of connections. The query complexity index is a comprehensive indicator, calculated by weighting relation path length, inference depth, and concurrent query count. The real-time performance level is typically an enumerated value, such as 0 for offline batch processing, 1 for near real-time queries, and 2 for high real-time online queries, with each level corresponding to different latency sensitivities. These quantized features are combined into a fixed-length floating-point multidimensional feature vector. Each dimension of the feature vector is standardized to eliminate dimensional differences, such as min-max normalization or Z-score normalization. This feature vector is one of the key inputs to the dynamic load prediction module, ensuring that the computational characteristics of the task can be effectively understood by the machine learning model.

[0021] The dynamic load prediction module's core function is to construct a high-precision time-series prediction model based on the task features output by the knowledge graph task parsing module, combined with real-time operational data and environmental data from the computing center equipment. This model outputs predicted device-level power load values ​​within a preset future time window. The module is deployed on independent high-performance computing nodes in the global scheduling layer to ensure the timeliness and accuracy of the predictions. The module employs a time-series prediction architecture based on Long Short-Term Memory (LSTM) networks. LSTM networks are recurrent neural networks adept at processing and predicting events in time series and can capture long-term dependencies. Specifically, the architecture includes an input layer, multiple stacked LSTM network hidden layers, and an output layer. Each LSTM network unit contains an input gate, a forget gate, and an output gate. These gating mechanisms, through Sigmoid and Tanh activation functions, precisely control the flow of information at each time step, enabling the model to effectively remember key information and forget irrelevant data, thereby improving its ability to learn complex time-series patterns. The input vector for this module is a combination of information from multiple dimensions: First, there is the multi-dimensional feature vector output by the knowledge graph task parsing module, which represents the computational requirements of the current or upcoming task; second, there is the current CPU utilization of the device, collected from each computing node through operating system-level monitoring tools such as cAdvisor and Prometheus, reflecting the real-time load of the processor; third, there is the GPU memory utilization, obtained through driver interfaces provided by graphics processor manufacturers such as NVIDIA SMI, revealing the usage of graphics processor resources; fourth, there is the memory bandwidth utilization, monitored through system performance counters such as perf-stat, assessing the throughput pressure of the memory subsystem; fifth, there is the disk I / O throughput, obtained through the file system or storage controller interface, measuring the read and write performance of the storage device; and sixth, there is the data center ambient temperature, collected in real time by IoT temperature and humidity sensors deployed inside the racks and in various areas of the data center; all these input data are sampled, synchronized, timestamped, and normalized to a unified numerical range to form the input sequence of the Long Short-Term Memory network. The dynamic load forecasting module outputs predicted power load values ​​for three different time scales: 5 minutes, 15 minutes, and 60 minutes before the next power consumption at the device level. To reflect the uncertainty of the forecast, each time scale is assigned confidence interval weights. These weights are derived through statistical analysis of the prediction error during model training. For example, multiple quantiles of the predicted distribution can be directly output by training a quantile regression long short-term memory network or using the Monte Carlo Dropout method, thereby calculating the confidence interval. The confidence interval weight represents the probability that the predicted value falls within a certain interval; the higher the weight value, the higher the confidence level of the predicted value. The unit of the predicted value is usually watts or kilowatts, accurate to two decimal places. These predicted values, along with their confidence interval weights, will serve as the core basis for energy planning by the multi-source power supply scheduling module. The prediction model will be periodically retrained online using the latest collected actual power consumption data and corresponding input features to adapt to the dynamic changes in the task load mode of the computing center and the degradation of equipment performance, ensuring the continuous accuracy of the prediction. The retraining process adopts an incremental learning strategy, combining historical data with new data for small-batch gradient updates to avoid catastrophic forgetting problems.

[0022] The multi-source power supply scheduling module aims to intelligently schedule heterogeneous power sources within the computing center, including mains power, energy storage battery packs, photovoltaic power sources, and backup diesel generators, based on the power load forecast values ​​output by the dynamic load forecasting module. This generates a power supply combination strategy that meets future load demands and optimizes total cost. The module is deployed on the main control server of the global scheduling layer and has the ability to macro-control all energy assets. The module has a built-in cost optimization engine, which uses a mixed-integer linear programming model as its core algorithm. The cost optimization engine constructs a mixed-integer linear programming model with the objective function of minimizing the total power supply cost. Under the premise of satisfying all power supply constraints, it solves for the optimal output power allocation scheme of each power source within the scheduling cycle. The objective function comprehensively considers multiple cost factors: grid power purchase cost, which changes dynamically based on real-time electricity price signals, such as peak and off-peak electricity prices; charging and discharging loss cost and cycle life loss cost of energy storage batteries; operation and maintenance cost of photovoltaic power sources and cost of underutilized power generation; and startup cost, operating fuel cost and environmental emission cost of backup diesel generators. All these costs are uniformly quantified into monetary units. The constraints of the mixed-integer linear programming model include: Equipment load requirement constraint: At each time step, the sum of the output power of all power supplies must be greater than or equal to the power load prediction value provided by the dynamic load prediction module to ensure that the power supply to the computing center is uninterrupted.

[0023] Maximum power output constraints: There is an upper limit to the capacity of mains power access; there are limits on the maximum discharge power and maximum charging power of energy storage battery packs; there are limits on the instantaneous maximum power generation of photovoltaic power sources; and there are limits on the rated output power of standby diesel generators.

[0024] Battery charge / discharge rate limits: The charging and discharging power of the energy storage battery pack must not exceed its safe operating range to prevent overcharging and over-discharging.

[0025] Battery capacity constraint: The remaining capacity of the energy storage battery pack must always be between the minimum and maximum threshold of its capacity.

[0026] Grid access capacity constraints: There is an upper limit to the power interaction between the computing center and the external power grid, whether it is purchasing or selling electricity.

[0027] Standby diesel generator start-up constraints: Diesel generator start-up takes time and each start-up is accompanied by additional costs. The model needs to introduce binary integer variables to represent the generator start-up and shutdown states, and consider the minimum running time and minimum downtime.

[0028] Power ramp constraint: The rate of power change of each power source between consecutive time steps cannot exceed its physical limit. For example, a diesel generator cannot instantly change from shutdown to full load.

[0029] Battery health status correction: The remaining capacity of the energy storage battery is affected by the remaining lifespan correction factor provided by the battery health status monitoring unit. The actual usable capacity is the product of the nominal capacity and the correction factor.

[0030] The model is solved using dedicated optimization solvers such as CPLEX and Gurobi. The solution cycle is synchronized with the load forecast cycle, for example, the power supply combination strategy is updated every 5 minutes. The solution result is a detailed power allocation plan, specifying the amount of electricity purchased or sold from the grid, the charging and discharging power of the energy storage battery, the utilization power of the photovoltaic power source, and the start-stop status and output power of the diesel generator, accurate to the kilowatt level, and accompanied by an execution timestamp. This strategy is then passed to the device-level power control module for specific execution.

[0031] Lithium iron phosphate (LFP) batteries are the mainstream technology for energy storage battery packs due to their high safety, long cycle life, and excellent wide-temperature performance, making them ideal as backup power for computing centers and peak shaving / valley filling tools. The energy storage battery pack is equipped with a battery health monitoring unit. This unit is an embedded system that uses high-precision sensors to collect real-time data on the voltage, current, temperature, and charge / discharge cycle count of each battery cell in the array, based on coulomb counts. The collected raw data undergoes digital signal processing and filtering before being uploaded to a battery management server deployed at the regional coordination layer. The battery management server uses a complex battery health status algorithm, such as one based on Kalman filtering. The state estimator, combined with a machine learning model, analyzes the trend of battery internal resistance change and predicts its degradation characteristics under different load and temperature conditions. Based on this data, the remaining lifespan of the battery is calculated. The remaining lifespan is a percentage value, representing the ratio of the battery's current capacity to its initial capacity. For example, when the calculated remaining lifespan is 80%, it means that the actual usable capacity of the battery is 80% of its nominal capacity. This remaining lifespan value will be used as a correction factor for the battery's usable capacity in the multi-source power supply scheduling module to ensure that the scheduling strategy does not exceed the battery's actual carrying capacity, thereby extending the overall lifespan of the battery and ensuring the long-term reliability of the system.

[0032] The photovoltaic (PV) power supply is connected to the system via a maximum power point tracking (MPPT) controller. The MPPT controller is a highly efficient DC-DC converter whose core function is to dynamically adjust the operating voltage and current of the PV array based on changes in solar irradiance and PV module temperature. This ensures the PV panels always operate at their maximum power point, maximizing PV power generation efficiency, typically maintaining it above 95%. The controller integrates a microcontroller and runs MPPT algorithms such as the perturbation-observation method or the incremental conductance method. The predicted PV power output is generated by combining weather sensors deployed on the roof of the computing center with historical irradiance data. The weather sensors include a high-precision solar intensity meter and an ambient temperature sensor. The system uses wind speed sensors to collect real-time environmental parameters. This real-time data, combined with a historical irradiance database, is input into a photovoltaic power prediction model based on recurrent neural networks or support vector regression. The model considers the impact of various meteorological factors such as cloud cover and aerosols on power generation, and outputs a photovoltaic power generation prediction curve for the next 10 minutes to 2 hours. This prediction is updated to the multi-source power supply scheduling module every 10 minutes for it to optimize power supply combination strategies. This module also includes a fault detection mechanism that compares actual power generation with predicted power generation and monitors abnormal component temperatures to promptly detect and report potential faults in the photovoltaic array, such as component damage or inverter failure.

[0033] The device-level power control module's main function is to receive the power combination strategy output by the multi-source power supply scheduling module, and to perform fine-grained dynamic configuration of the power supply voltage, current limit, sleep threshold, and wake-up delay parameters for each computing node, storage unit, and network switching device based on the hardware topology and power interface specifications of the computing center equipment. This module is deployed in the edge controller of the regional coordination layer and the baseboard management controller of the device execution layer, realizing fine-grained distributed power management. The voltage regulation unit dynamically adjusts the core power supply voltage of the CPU and GPU based on the computational intensity of the task. This unit communicates with the baseboard management controller of each computing node, utilizing its integrated power management IC (PMIC) or voltage regulation module (VRM) to achieve dynamic voltage and frequency scaling. When the knowledge graph task parsing module reports a high computational intensity, such as complex graph traversal or a large number of floating-point operations, the voltage regulation unit increases the core voltage to allow the processor to run at a higher frequency, thereby accelerating task execution. Conversely, when the task intensity is low or the device is idle, the voltage regulation unit reduces the core voltage and frequency, significantly reducing power consumption. The voltage and frequency adjustment steps and ranges are preset according to the processor manufacturer's specifications and adjusted in real time using a precise power consumption model to achieve the best balance between performance and energy efficiency. The adjustment latency is controlled at the microsecond level, ensuring rapid response to load changes.

[0034] The current limiting unit sets the maximum output current for each power rail to prevent hardware-protective shutdowns caused by instantaneous overloads or faults. This unit also controls the power management integrated circuit or hot-swap controller through the baseboard management controller or intelligent power distribution unit. Each computing node, storage device, or even a single power rail, such as 12V, 5V, or 3.3V, can be independently configured with a current limit. When the actual current instantaneously exceeds the preset threshold, the unit immediately triggers a current limiting mechanism, such as reducing the output voltage or limiting the power output, and reports the overcurrent event to the feedback optimization module. This mechanism effectively protects sensitive hardware components from damage and avoids unexpected system interruptions or instability caused by power transients.

[0035] The hibernation management unit is used to switch non-critical computing nodes to a low-power hibernation state during the idle period of knowledge graph tasks. The definition of "non-critical computing node" is dynamically determined by the global scheduling layer based on the current task queue, node health status, and resource utilization. Hibernation states cover a variety of low-power modes, such as operating system-level ACPI S1-S5 states or processor-level C-states such as C6 and C7. Before a node enters hibernation, the hibernation management unit records its current context state, including data in memory, register contents, and network connection status. This context information can be stored on local non-volatile memory such as a solid-state drive or synchronized to a distributed cache system. Hibernation thresholds, such as CPU utilization below 5% or network traffic below a specific threshold, are dynamically adjusted by the feedback optimization module to ensure an optimal balance between energy saving and fast response. This unit effectively reduces the energy consumption of idle resources.

[0036] The wake-up scheduling unit, upon receiving a new task scheduling instruction, wakes up nodes in a pre-defined order based on task priority and node dependencies. Wake-up latency parameters, such as 500 milliseconds from S3 state and several seconds from S5 state, are precisely recorded and managed. When a new knowledge graph task needs to be assigned to a dormant node in the global scheduling layer, the wake-up scheduling unit selects the most suitable node based on the task's real-time requirements, computational resource needs, the dormant node's context state, and the wake-up latency. The wake-up process follows a strict protocol to ensure nodes can quickly return to their working state and reload previously saved context data, thus ensuring task execution continuity and data consistency. For example, a task dependent on a specific dataset might prioritize waking up the dormant storage node that stores the dataset's context to avoid data transfer overhead. Wake-up instructions are issued through the baseboard management controller and transmitted via network wake-up or a dedicated control bus.

[0037] The feedback optimization module is responsible for collecting real-time data on actual power consumption, temperature changes, task completion time, and power switching events during equipment operation. This key data is then transmitted back to the dynamic load prediction module and the multi-source power supply scheduling module to correct prediction model parameters and power supply scheduling strategies online, thereby achieving closed-loop adaptive optimization of the system. This module is deployed in a distributed manner, with the data acquisition agent located at the device execution layer and the data aggregation and analysis unit located at the regional coordination layer and the global scheduling layer. The anomaly detection submodule compares the real-time collected actual power consumption data with the predicted load value output by the dynamic load prediction module. The actual power consumption data is collected at a rate of seconds via the PMBus interface of the intelligent power distribution unit or server power supply. When the deviation between the actual power consumption and the predicted load exceeds a preset threshold, such as a deviation exceeding 10% within three consecutive sampling periods, an anomaly event is triggered. The preset threshold is not a fixed value, but is dynamically adjusted based on historical data and system operating experience using statistical methods such as the 3-standard deviation rule or an anomaly detection algorithm based on isolated forests. The anomaly event marker includes the anomaly type, such as excessive power consumption, excessively low power consumption, abnormal fluctuation, occurrence timestamp, anomaly severity quantification value, and the device identifier involved. These markers are immediately written to the system log and used as key feedback signals.

[0038] The strategy correction submodule uses abnormal event markers, task completion delay data, and power switching counts as core feedback signals to update the network weights in the dynamic load prediction module using an online gradient descent algorithm, and adjusts the constraint relaxation coefficient of the cost optimization engine in the multi-source power supply scheduling module. Online gradient descent is an iterative optimization algorithm that updates model parameters each time a new data sample or a small batch is received. For the dynamic load prediction module, the submodule uses the error between the actual power consumption and the predicted power consumption to calculate the gradient of the loss function, such as the mean squared error, and updates the weights and biases of the long short-term memory network model in the opposite direction of the gradient. The learning rate is adaptively adjusted according to the training progress, for example, by using the Adam optimizer to prevent overfitting or underfitting. The update frequency matches the data acquisition frequency, for example, a small batch update is performed every 5 minutes. For the multi-source power supply scheduling module, the strategy correction submodule adjusts the constraint relaxation coefficient of the cost optimization engine based on feedback signals. For example, if power outages or task delays occur frequently due to insufficient power capacity, the constraint on the maximum output power of the power supply will be appropriately relaxed, or the startup priority of the backup power supply will be increased. If power switching is too frequent, leading to increased equipment wear, the penalty term for power switching costs will be increased. These adjustments are achieved by modifying the parameters or weights in the mixed-integer linear programming model, with the goal of enabling the model to generate a more stable power supply strategy that better meets actual operating requirements in the next solution. The feedback mechanism ensures that the system can continuously learn and optimize its power management strategy when dealing with new knowledge graph task types, equipment performance degradation, or external environmental changes such as electricity price fluctuations.

[0039] The system operates entirely within a hierarchical control architecture. The global scheduling layer is deployed on the main control server of the computing center. This layer is responsible for task allocation and power supply strategy generation at the macro level of the entire computing center. Its core functions include a task scheduler, which allocates knowledge graph computing tasks to appropriate computing nodes or clusters based on the task characteristics output by the knowledge graph task parsing module; a resource manager, which monitors the global status of all computing, storage, and network resources; and a dynamic load prediction module and a multi-source power supply scheduling module defined in this invention, which are responsible for generating globally optimal power supply strategies. The global scheduling layer communicates with the regional coordination layer through a high-performance network such as 10 Gigabit Ethernet to send task scheduling instructions and power supply strategies.

[0040] The regional coordination layer consists of edge controllers deployed in each rack. These edge controllers are embedded high-performance computing devices deployed close to the computing nodes. This layer is responsible for aggregating the status of local power devices and forwarding commands. Its functions include a local data aggregator, which collects real-time data such as power consumption, temperature, and fan speed from the intelligent power distribution unit, baseboard management controller, and sensors in the region, performs preliminary processing and aggregation, and then reports it to the global scheduling layer; a command distributor, which receives task scheduling commands and power supply strategies from the global scheduling layer and converts them into specific control commands recognizable by the device execution layer; and a battery management server as defined in this invention, such as a server managing the battery health status monitoring unit of the energy storage battery pack. The regional coordination layer communicates with the device execution layer via industrial Ethernet or a dedicated control bus.

[0041] The device execution layer is integrated into the baseboard management controller of each computing node. The baseboard management controller is a microcontroller on the server motherboard, responsible for the hardware monitoring and management of the server. This layer is directly responsible for executing specific voltage, current, and sleep control commands. Its functions include a sensor interface to directly read parameters such as temperature, voltage, and current of the CPU and GPU; a power management unit responsible for dynamically adjusting the core voltage of the CPU and GPU through power management integrated circuits or voltage regulation modules according to the instructions issued by the regional coordination layer, limiting the power rail current, and controlling the sleep and wake-up of computing nodes; and an alarm processing unit that immediately reports to the regional coordination layer when a hardware failure such as over-temperature or over-voltage occurs. The device execution layer interacts directly with the physical hardware and is the layer at which the power strategy is ultimately implemented.

[0042] The above-described modules of this invention can also be applied in the following fields: The primary application: within the server power supply industry; The second largest application: photovoltaic energy storage; The third major application: switching power supplies and adapter power supplies used in computers, printers, and industrial control computers.

[0043] This invention constructs an end-to-end, closed-loop, adaptive intelligent power supply system for knowledge graph computing centers through the collaborative work of the aforementioned modules and the precise coordination of the layered architecture. The system no longer relies solely on traditional rules of thumb or passive responses, but achieves intelligent, refined, and forward-looking power management through a deep understanding of knowledge graph tasks, combined with high-precision dynamic load prediction and cost-optimized multi-source power supply scheduling. Through device-level power control, the system can dynamically adjust voltage and current and intelligently manage sleep modes, significantly improving energy efficiency, reducing energy consumption, and extending hardware lifespan. The feedback optimization mechanism endows the system with the ability to continuously learn and adapt to environmental changes. Overall, this invention provides a highly energy-efficient, highly reliable, and highly intelligent power support system for large-scale knowledge graph computing, effectively solving the technical problems of low energy efficiency, shortened equipment lifespan, and insufficient system stability in existing technologies, thereby achieving dynamic resource scheduling: optimizing equipment power allocation based on computing power demand and power load, reducing energy consumption, and improving computing power utilization.

Claims

1. A power supply intelligent system for the computing center equipment of a knowledge graph, characterized in that, include: The knowledge graph task parsing module is used to receive and parse the knowledge graph computing tasks to be executed, and extract the graph size, node type distribution, relationship complexity, real-time requirements and historical execution characteristics of the tasks. The dynamic load prediction module is used to construct a time-series prediction model based on the task features output by the knowledge graph task parsing module, combined with the current operating status data of the computing center equipment, environmental temperature and humidity data and historical load curves, and output the predicted value of the equipment-level power load within a future preset time window. The multi-source power supply scheduling module is used to schedule at least two of the following power sources, namely, mains power, energy storage battery pack, photovoltaic power and backup diesel generator, based on the power load prediction value output by the dynamic load prediction module, to generate a power supply combination strategy that meets the load requirements and is cost-optimal. The device-level power control module is used to receive the power combination strategy output by the multi-source power supply scheduling module, and dynamically configure the power supply voltage, current limit, sleep threshold and wake-up delay parameters of each computing node, storage unit and network switching device according to the hardware topology and power interface specifications of the device. The feedback optimization module is used to collect the actual power consumption, temperature changes, task completion time and power switching events during the operation of the equipment in real time, and transmit the data back to the dynamic load prediction module and the multi-source power supply scheduling module for online correction of prediction model parameters and power supply scheduling strategies.

2. The intelligent power supply system for the computing center equipment of the knowledge graph according to claim 1, characterized in that, The knowledge graph task parsing module includes a task semantic analysis submodule and a task feature vector submodule. The task semantic analysis submodule is used to parse the input knowledge graph task using natural language or structured query language, and to identify the entity types, relation path lengths, inference depths, and concurrent query counts involved in the task. The task feature vector submodule is used to quantize the identification results into multidimensional feature vectors, which include the number of graph nodes, edge density, average degree, query complexity index, and real-time level encoding.

3. The intelligent power supply system for the computing center equipment of the knowledge graph according to claim 2, characterized in that, The dynamic load prediction module adopts a time-series prediction architecture based on long short-term memory networks. Its inputs include the multi-dimensional feature vector, the current CPU utilization of the device, the GPU memory utilization, the memory bandwidth utilization, the disk I / O throughput, and the ambient temperature of the data center. The dynamic load prediction module outputs power load prediction values ​​for three time scales: 5 minutes, 15 minutes, and 60 minutes, and assigns a confidence interval weight to each time scale.

4. The intelligent power supply system for the computing center equipment of the knowledge graph according to claim 3, characterized in that, The multi-source power supply scheduling module has a built-in cost optimization engine. The cost optimization engine constructs a mixed integer linear programming model based on real-time electricity price signals, remaining energy storage battery capacity, photovoltaic power generation prediction, and diesel generator start-up cost. The mixed integer linear programming model takes minimizing the total power supply cost as the objective function and uses equipment load demand, maximum power output of the power source, battery charge and discharge rate limit, and grid access capacity as constraints to solve for the optimal output power allocation scheme of each power source within the scheduling cycle.

5. The intelligent power supply system for the computing center equipment of the knowledge graph according to claim 4, characterized in that, The device-level power control module includes a voltage regulation unit, a current limiting unit, a sleep management unit, and a wake-up scheduling unit. The voltage regulation unit dynamically adjusts the core power supply voltage of the CPU and GPU according to the task's computational intensity to achieve dynamic voltage frequency scaling. The current limiting unit sets the maximum output current of each power rail to prevent hardware-protected shutdown caused by instantaneous overcurrent. The sleep management unit switches non-critical computing nodes to a low-power sleep state during task idle periods and records their context state. The wake-up scheduling unit wakes up sleep nodes in a preset order according to task priority and node dependencies after receiving a new task scheduling instruction, ensuring the continuity of task execution and data consistency.

6. The intelligent power supply system for the computing center equipment of the knowledge graph according to claim 5, characterized in that, The feedback optimization module includes an anomaly detection submodule and a strategy correction submodule; the anomaly detection submodule is used to compare the deviation between the actual power consumption and the predicted load, and when the deviation exceeds a preset threshold, an anomaly event is triggered. The strategy correction submodule is used to use abnormal event markers, task completion delay data, and power switching counts as feedback signals to update the network weights in the dynamic load prediction module using an online gradient descent algorithm, and to adjust the constraint relaxation coefficient of the cost optimization engine in the multi-source power supply scheduling module.

7. The intelligent power supply system for the computing center equipment of the knowledge graph according to claim 1, characterized in that, The system operates in a hierarchical control architecture, which includes a global scheduling layer, a regional coordination layer, and a device execution layer. The global scheduling layer is deployed on the main control server of the computing center and is responsible for task allocation and power policy generation across racks. The regional coordination layer is deployed on the edge controllers of each rack and is responsible for the aggregation of local power device status and instruction forwarding. The device execution layer is integrated into the baseboard management controller of each computing node and is responsible for executing specific voltage, current, and sleep control instructions.

8. The intelligent power supply system for the computing center equipment of the knowledge graph according to claim 7, characterized in that, The energy storage battery pack uses lithium iron phosphate batteries and is equipped with a battery health status monitoring unit. The battery health status monitoring unit is used to collect the battery's internal resistance, charge-discharge cycle count and temperature gradient in real time, calculate the remaining service life, and use the remaining service life as a correction factor for the battery's available capacity in the multi-source power supply scheduling module.

9. The intelligent power supply system for the computing center equipment of the knowledge graph according to claim 8, characterized in that, The photovoltaic power source is connected to the system through a maximum power point tracking controller. The maximum power point tracking controller dynamically adjusts the operating point according to the light intensity and component temperature to ensure that the photovoltaic power generation efficiency is maintained above 95%. The photovoltaic power generation prediction value is generated by the weather sensors deployed on the roof and historical irradiance data, and is updated to the multi-source power supply scheduling module at a cycle of 10 minutes.

10. The intelligent power supply system for the computing center equipment of the knowledge graph according to claim 4, characterized in that, The constraints of the mixed-integer linear programming model also include battery capacity constraints, standby diesel generator start-up constraints, and power supply ramp constraints. The battery capacity constraints are used to ensure that the remaining power of the energy storage battery pack is between its minimum and maximum capacity thresholds at any time. The standby diesel generator start-up constraints introduce binary integer variables to represent the generator's start-up and shutdown states, and consider the minimum running time and minimum downtime. The power supply ramp constraints are used to limit the rate of power change of each power source between consecutive time steps to not exceed its physical limit.

Images