Modular power supply system configuration method and system for micro-module computer room

Abstract

The present invention provides a modular power supply system configuration method and system for a micro-module computer room, which relates to the field of electric power technology, including collecting power consumption status data and load distribution data of functional equipment in the micro-module computer room, establishing a power supply network topology model, dividing the power supply system into primary load, secondary load and tertiary load, and configuring independent power supply circuits and data acquisition units in the primary load, secondary load and tertiary load respectively; updating the power supply network topology model according to the power consumption data collected by the data acquisition unit, using a reinforcement learning algorithm to generate an adaptive power allocation strategy, setting a modular energy storage unit, and providing backup power for the primary load and the secondary load; constructing a machine learning fault prediction model based on the power consumption data, and dynamically adjusting the power supply path and power allocation plan of the power supply circuit.

Description

Technical Field

[0001] The present invention relates to electric power technology, and in particular to a modular power supply system configuration method and system for a micro-module computer room. Background Art

[0002] As the scale of data centers continues to expand, traditional computer room power supply systems have the following problems: First, the power supply system configuration lacks flexibility and is difficult to dynamically adjust according to load changes; second, the redundancy configuration of the power supply system is unreasonable, resulting in excessively high system investment costs or insufficient power supply reliability; third, the backup power supply configuration solution is single and has limited response capabilities when dealing with different types of loads; finally, there is a lack of effective fault prediction and early warning mechanisms, which affects the reliability of system operation.

[0003] The current micro-module room power supply system usually adopts a unified power supply solution, and fails to adopt differentiated power supply strategies for loads of different importance. The topology of the power supply system is fixed and it is difficult to adapt to dynamic changes in loads. The energy management of the energy storage device lacks an intelligent scheduling mechanism, resulting in inefficient use of the energy storage unit. At the same time, the existing system generally lacks fault prediction capabilities based on data analysis, making it difficult to detect and deal with potential fault hazards in a timely manner.

[0004] In addition, the existing micro-module room power supply system often over-pursues power supply reliability during design and adopts a fully redundant configuration solution, resulting in a large number of idle devices and energy waste. At the same time, the power supply system lacks intelligent management methods and cannot achieve on-demand power supply and precise protection of load equipment. Traditional fault handling methods mainly rely on manual experience judgment and regular maintenance, which cannot effectively prevent system failures, nor can they quickly locate and handle faults. Summary of the invention

[0005] The embodiments of the present invention provide a modular power supply system configuration method and system for a micro-module computer room, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A modular power supply system configuration method for a micro-module computer room is provided, including:

[0008] Collect power consumption status data and load distribution data of functional equipment in the micromodule room, and establish a power supply network topology model based on the power consumption status data and load distribution data; analyze the power consumption characteristics of functional equipment based on the power supply network topology model, divide the power supply system into primary load, secondary load and tertiary load, and configure independent power supply circuits and data acquisition units in the primary load, secondary load and tertiary load respectively;

[0009] The power supply network topology model is updated according to the power consumption data collected by the data acquisition unit. Based on the updated power supply network topology model, a reinforcement learning algorithm is used to generate an adaptive power allocation strategy to determine the power supply path and power allocation scheme of the power supply circuit; a modular energy storage unit is set based on the power supply path and power allocation scheme of the power supply circuit, wherein the modular energy storage unit includes a supercapacitor module and a lithium battery module, and the charging and discharging process of the supercapacitor module and the lithium battery module is controlled by an intelligent energy management algorithm to provide backup power for the primary load and the secondary load;

[0010] A machine learning fault prediction model is built based on electricity consumption data, and the fault risk probability is used as the trigger condition for emergency response. A graded warning threshold is set. When the warning level changes, the power supply path and power distribution plan of the power supply circuit are dynamically adjusted.

[0011] In an optional embodiment,

[0012] The steps of collecting power consumption status data and load distribution data of functional equipment in the micromodule computer room, and establishing a power supply network topology model according to the power consumption status data and the load distribution data; analyzing the power consumption characteristics of the functional equipment based on the power supply network topology model, dividing the power supply system into primary load, secondary load and tertiary load, and configuring independent power supply circuits and data collection units in the primary load, secondary load and tertiary load respectively include:

[0013] The power consumption status data includes voltage effective value, current effective value, active power, reactive power and power factor, and the load distribution data includes equipment spatial position coordinates, rated power and power change data; median filtering is performed on the power consumption status data to obtain filtered data, and a power supply network topology model is constructed based on the filtered data and the load distribution data. The power supply network topology model includes power supply nodes and power supply lines, and the weight parameters of the power supply lines include line impedance parameters and the maximum allowable current of the lines;

[0014] Analyze the power consumption characteristics of functional equipment based on the power supply network topology model, extract characteristic data including rated power, power fluctuation coefficient, allowable interruption time and backup power demand coefficient from the power consumption characteristics, and build a fuzzy clustering model based on the characteristic data; calculate the membership matrix of functional equipment through the fuzzy clustering model, update the cluster center according to the membership matrix, and divide the power supply system into primary load, secondary load and tertiary load based on the cluster center;

[0015] Based on the power supply network topology model, a dual-bus dual-way power supply architecture and an online uninterruptible power supply are configured for the first-level load, a single-bus dual-way power supply architecture and a backup uninterruptible power supply are configured for the second-level load, and a single-way power supply architecture is configured for the third-level load, and a power supply architecture configuration scheme is generated; layout parameters of the power supply line are calculated according to the power supply architecture configuration scheme and the power supply network topology model, the layout parameters include cable specifications, wiring paths and protection device parameters, and independent power supply circuits and data acquisition units are respectively configured in the first-level load, the second-level load and the third-level load according to the layout parameters.

[0016] In an optional embodiment,

[0017] The steps of updating the power supply network topology model according to the power consumption data collected by the data acquisition unit, generating an adaptive power allocation strategy based on the updated power supply network topology model by using a reinforcement learning algorithm, and determining the power supply path and power allocation scheme of the power supply circuit include:

[0018] The power consumption data collected by the data acquisition unit is standardized to obtain standardized data; a three-dimensional state tensor including a voltage matrix, a current matrix and a power matrix is ​​constructed based on the standardized data, and the three-dimensional state tensor is updated by using a double sliding window mechanism to obtain an updated power supply network topology model;

[0019] The voltage deviation rate, line load rate, power allocation ratio and equipment operation status index in the updated power supply network topology model are constructed as a state vector, and the power adjustment amount of the main power supply, the power adjustment amount of the backup power supply and the power allocation adjustment amount between loops are constructed as an action vector; a multi-objective combined reward function including power supply reliability reward items, energy efficiency reward items and stability reward items is constructed, and an adaptive power allocation strategy is trained based on the state vector, action vector and multi-objective combined reward function. The adaptive power allocation strategy is implemented using a deep deterministic policy gradient algorithm, and the deep deterministic policy gradient algorithm includes an action evaluation network and an action execution network, wherein the action evaluation network adopts a dual-Q network structure, and the action execution network adopts a three-layer neural network structure, and the network parameters of the action evaluation network and the action execution network are updated based on priority sampling of the temporal difference error;

[0020] A long short-term memory network is used to predict the load demand for a preset time in the future. The power supply path of the power supply circuit is determined based on the load demand and the adaptive power allocation strategy. The equipment in each circuit is divided into primary load, secondary load and tertiary load according to their importance. The power ratio of the main power supply and the backup power supply is determined according to the power supply path and load level, and power is allocated to each device.

[0021] In an optional embodiment,

[0022] The steps of constructing a multi-objective combined reward function including power supply reliability reward items, energy efficiency reward items and stability reward items include:

[0023] The power supply reliability reward item is calculated based on the voltage deviation rate and power loss of each level of load, the energy efficiency reward item is calculated based on the deviation between the line load rate and the optimal load rate, and the stability reward item is calculated based on the power adjustment amount at adjacent moments;

[0024] A judgment matrix is ​​constructed by using a hierarchical analysis method and a characteristic vector is calculated to obtain a basic weight value of each reward item in the multi-objective combination reward function, and a weight adjustment range of each reward item is set based on the basic weight value;

[0025] The operating status of the modular power supply system is obtained, including the degree of voltage deviation exceeding the limit, the degree of load rate deviation from the optimal value, and the power fluctuation amplitude. Based on the operating status of the modular power supply system, the Sigmoid mapping function is used to dynamically adjust the weight of each reward item. When the degree of voltage deviation exceeding the limit is greater than the voltage threshold, the weight of the power supply reliability reward item is increased; when the degree of load rate deviation from the optimal value is greater than the load rate threshold, the weight of the energy efficiency reward item is increased; when the power fluctuation amplitude is greater than the power threshold, the weight of the stability reward item is increased.

[0026] In an optional embodiment,

[0027] The step of training the adaptive power allocation strategy based on the state vector, the action vector and the multi-objective combined reward function comprises:

[0028] Obtain the current state of the power supply system, generate power adjustment actions based on the current strategy, execute the power adjustment actions and observe the response of the power supply system, calculate the multi-objective combined reward function including power supply reliability reward items, energy efficiency reward items and stability reward items, and record the current state of the power supply system, power adjustment actions, calculation results of the multi-objective combined reward function and the state of the power supply system after response as state transition data and store them in the experience pool;

[0029] The training samples are obtained from the experience pool according to the priority of the temporal difference error, the temporal difference error is calculated based on the training samples, the temporal difference error is input into the action evaluation network and the action execution network respectively, the network weights of the action evaluation network and the network parameters of the action execution network are updated, and the target network is updated by a soft update method;

[0030] During the training process, exploration noise is set to conduct strategy exploration, the network learning rate is dynamically adjusted and the training batch size is optimized; the average reward value change trend, network parameter update amplitude and strategy performance indicators during the training process are monitored and recorded. When the training round reaches the preset value and the reward value converges and stabilizes, the verification data set is used to verify the training results. When the verification result meets the performance requirements, the training process is completed.

[0031] In an optional embodiment,

[0032] A modular energy storage unit is set based on the power supply path and power distribution scheme of the power supply circuit. The modular energy storage unit includes a super capacitor module and a lithium battery module. The charging and discharging process of the super capacitor module and the lithium battery module are controlled by an intelligent energy management algorithm. The steps of providing backup power for the primary load and the secondary load include:

[0033] According to the peak power demand, energy demand and backup time requirement of the primary load and the secondary load, an energy storage capacity optimization model is established, and the capacity configuration of the supercapacitor module and the lithium battery module is determined based on the energy storage capacity optimization model, wherein the constraint conditions of the energy storage capacity optimization model include: the total energy of the supercapacitor module and the lithium battery module is greater than or equal to the total energy required by the system, and the total power of the supercapacitor module and the lithium battery module is greater than or equal to the peak power demand;

[0034] The supercapacitor module is composed of single supercapacitors connected in series and in parallel to form a supercapacitor group, and the lithium battery module is composed of lithium iron phosphate batteries connected in series to form a lithium battery group; the supercapacitor module and the lithium battery module are both connected to the busbar through a bidirectional DC / DC converter, wherein the supercapacitor module adopts an interleaved parallel Buck-Boost topology, the lithium battery module adopts a full-bridge LLC resonant conversion topology, and the busbar adopts a three-level converter topology;

[0035] The main power supply is monitored in real time using a fault detection method based on voltage sag. When a main power supply fault is detected, the supercapacitor module is triggered to discharge, and the lithium battery module is started to enter the standby state, and power switching is achieved based on synchronous phase locking technology.

[0036] An intelligent energy management algorithm is designed, which includes power distribution control and mode switching control; the power distribution control adopts a fuzzy adaptive PID control strategy, establishes a fuzzy rule base based on the load power change rate, the supercapacitor module voltage deviation and the lithium battery module charge state deviation, realizes adaptive adjustment of the PID controller through online parameter setting, and dynamically distributes the output power of the supercapacitor module and the lithium battery module; the mode switching control includes a normal operating mode, a fast response mode and an energy recovery mode, wherein in the normal operating mode, the supercapacitor module and the lithium battery module distribute power according to a preset ratio, in the fast response mode, the supercapacitor module preferentially responds to load power mutations, and in the energy recovery mode, the supercapacitor module preferentially absorbs regenerative braking energy.

[0037] In an optional embodiment,

[0038] A machine learning fault prediction model is built based on power consumption data. The fault risk probability is used as the trigger condition for emergency response. A graded warning threshold is set. When the warning level changes, the power supply path and power distribution plan of the power supply circuit are dynamically adjusted. The steps include:

[0039] Collect electricity consumption data including voltage fluctuation rate, current harmonic content and power factor, perform feature extraction and data preprocessing on the electricity consumption data, input the preprocessed electricity consumption data into the LSTM-GRU hybrid network to establish a machine learning fault prediction model, wherein the LSTM unit extracts the long-term dependency features of the electricity consumption data, and the GRU unit extracts the short-term fluctuation features. An attention mechanism is introduced into the network to assign weight coefficients to features of different time scales, and a new feature representation is obtained by weighted calculation of the attention weight and the original feature vector, and the new feature representation is nonlinearly mapped through the Softmax function to obtain a fault probability value;

[0040] The fault probability values ​​are divided into different warning levels, and a risk assessment indicator system is established for each warning level. According to the risk assessment results, corresponding load transfer strategies, power switching strategies and power limitation strategies are set to form a hierarchical warning response mechanism.

[0041] Monitor changes in warning levels in real time. When the warning level changes, dynamically adjust the power supply path and power allocation plan of the power supply circuit according to the preset priority strategy and constraints based on the real-time operating status of the power supply system, the importance of the equipment and the load fluctuation characteristics.

[0042] According to a second aspect of the embodiments of the present invention,

[0043] Provide modular power supply system configuration system for micro-module room, including:

[0044] The first unit is used to collect power consumption status data and load distribution data of functional equipment in the micromodule room, and establish a power supply network topology model based on the power consumption status data and load distribution data; analyze the power consumption characteristics of functional equipment based on the power supply network topology model, divide the power supply system into primary load, secondary load and tertiary load, and configure independent power supply circuits and data acquisition units in the primary load, secondary load and tertiary load respectively;

[0045] The second unit is used to update the power supply network topology model according to the power consumption data collected by the data collection unit, and based on the updated power supply network topology model, use the reinforcement learning algorithm to generate an adaptive power allocation strategy to determine the power supply path and power allocation scheme of the power supply circuit; based on the power supply path and power allocation scheme of the power supply circuit, set a modular energy storage unit, the modular energy storage unit includes a super capacitor module and a lithium battery module, and the charging and discharging process of the super capacitor module and the lithium battery module is controlled by an intelligent energy management algorithm to provide backup power for the primary load and the secondary load;

[0046] The third unit is used to build a machine learning fault prediction model based on electricity consumption data, use the fault risk probability as the trigger condition for emergency response, set the graded warning threshold, and dynamically adjust the power supply path and power distribution plan of the power supply circuit when the warning level changes.

[0047] According to a third aspect of the embodiments of the present invention,

[0048] An electronic device is provided, comprising:

[0049] processor;

[0050] a memory for storing processor-executable instructions;

[0051] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0052] According to a fourth aspect of the embodiments of the present invention,

[0053] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0054] The present invention realizes load classification management and differentiated power supply strategies by establishing a dynamic power supply network topology model, effectively improving the rationality and economy of the power supply system configuration; adopts a reinforcement learning algorithm to generate an adaptive power allocation strategy, so that the power supply system can be dynamically adjusted in real time according to load changes, significantly improving the system operation efficiency and energy utilization rate; the fault prediction model based on machine learning realizes intelligent monitoring of the system operation status and fault warning, improving the reliability and safety of the power supply system; and enhances the scalability and maintenance convenience of the power supply system through modular system architecture design, reducing the system construction and operation and maintenance costs.

[0055] The modular energy storage unit provided in the present invention adopts a hybrid energy storage solution of supercapacitor modules and lithium battery modules, and realizes coordinated control and optimized scheduling of energy storage equipment through intelligent energy management algorithms, thereby improving the response speed and service life of the backup power supply; the differentiated power supply strategy and dynamically optimized power allocation solution significantly improve the energy utilization efficiency of the power supply system; the data-driven fault prediction and early warning mechanism greatly improves the stability of the system operation; the intelligent energy storage management solution improves the utilization efficiency of the energy storage system; the modular system structure enhances the adaptability and scalability of the power supply system under load changes.

[0056] The present invention realizes the intelligent operation and maintenance of the power supply system, reduces the frequency of manual intervention, and improves the operation and maintenance efficiency and management level; the prediction model based on machine learning improves the system's fault prevention capability and emergency response speed; the optimized power supply path and power distribution scheme improve the overall power supply quality of the system; through real-time monitoring and dynamic adjustment strategies, the system's ability to cope with emergencies is enhanced; the intelligent management scheme reduces the system's operating costs and improves power supply reliability; the modular design concept improves the system's standardization level and engineering implementation efficiency; the decision support based on data analysis improves the scientific nature of the system's operation and the foresight; and the multi-level linkage protection measures improve the system's safety performance and fault recovery capability. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A schematic diagram of a flow chart of a method for configuring a modular power supply system of a micro-module computer room according to an embodiment of the present invention;

[0058] Figure 2 It is a structural schematic diagram of a modular power supply system configuration system of a micro-module computer room according to an embodiment of the present invention. DETAILED DESCRIPTION

[0059] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. 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 creative work are within the scope of protection of the present invention.

[0060] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0061] Figure 1 FIG. 1 is a flow chart of a method for configuring a modular power supply system of a micro-module computer room according to an embodiment of the present invention. Figure 1As shown, the method includes:

[0062] S1. Collect power consumption status data and load distribution data of functional equipment in the micromodule room, and establish a power supply network topology model based on the power consumption status data and load distribution data; analyze the power consumption characteristics of functional equipment based on the power supply network topology model, divide the power supply system into primary load, secondary load and tertiary load, and configure independent power supply circuits and data acquisition units in the primary load, secondary load and tertiary load respectively;

[0063] S2. Update the power supply network topology model according to the power consumption data collected by the data acquisition unit, and based on the updated power supply network topology model, use the reinforcement learning algorithm to generate an adaptive power allocation strategy to determine the power supply path and power allocation scheme of the power supply circuit; set a modular energy storage unit based on the power supply path and power allocation scheme of the power supply circuit, wherein the modular energy storage unit includes a supercapacitor module and a lithium battery module, and the charging and discharging process of the supercapacitor module and the lithium battery module is controlled by an intelligent energy management algorithm to provide backup power for the primary load and the secondary load;

[0064] S3. Build a machine learning fault prediction model based on electricity consumption data, use the fault risk probability as the trigger condition for emergency response, set a graded warning threshold, and dynamically adjust the power supply path and power distribution plan of the power supply circuit when the warning level changes.

[0065] Exemplarily, the modular power supply system configuration method of the micro-module computer room first collects the power consumption status data and load distribution data of the functional equipment in the computer room. The power consumption status data includes the real-time power, voltage, current and other parameters of each device, which are collected through smart meters or power monitors. The load distribution data includes the physical location of each device, power supply line and other information, which are obtained through the computer room management system. For example, the real-time power of server A in a cabinet is 2kW, the voltage is 220V, the current is 9.1A, it is located at position 1 of the cabinet, and is connected to power supply line 1.

[0066] Based on the collected data, a power supply network topology model is constructed. The model is represented by a graph structure, where nodes represent devices, edges represent power supply lines, and both nodes and edges carry relevant attribute information. For example, node A in the model represents server A, which has attributes such as power and location, and is connected to node B through edge e, which represents power supply line No. 1 and has attributes such as rated current.

[0067] Then, the power consumption characteristics of functional equipment are analyzed, and the power supply system is divided into three levels. The first-level load is critical equipment, such as core switches, important servers, etc., which require uninterrupted power supply; the second-level load is important equipment, such as ordinary servers, storage devices, etc., which allow short-term power outages; the third-level load is non-critical equipment, such as lighting, air conditioning, etc., which can be powered off for a long time. Independent power supply circuits and data acquisition units are configured in each level of load. For example, the first-level load is configured with dual power supplies, each with an uninterruptible power supply; the second-level load is configured with a single power supply and a backup power supply; the third-level load is configured with ordinary AC power supply.

[0068] The data acquisition unit continuously collects power consumption data and updates the power supply network topology model. Based on the updated model, a reinforcement learning algorithm is used to generate an adaptive power allocation strategy. The algorithm aims to minimize energy consumption and continuously optimizes the power supply path and power allocation plan through trial and error learning. For example, when a power supply line is found to be overloaded, some equipment will be automatically switched to the backup line.

[0069] According to the generated power supply scheme, modular energy storage units are set up, including supercapacitor modules and lithium battery modules. The supercapacitor module is used to cope with instantaneous high power demand, and the lithium battery module provides long-term backup power. The intelligent energy management algorithm is used to control the charging and discharging process of the two modules, and the charging and discharging strategy is dynamically adjusted according to load changes and grid status. For example, when a fluctuation in the mains power is detected, the supercapacitor module is immediately started to supply power; after the grid returns to normal, the supercapacitor module is charged first.

[0070] Based on historical electricity consumption data, a machine learning fault prediction model is constructed. The model uses a random forest algorithm, and the input features include equipment operation time, load rate, ambient temperature, etc., and the output is the probability of fault occurrence. Set multi-level warning thresholds, such as a red warning when the probability of fault is greater than 80%, an orange warning when it is 60%-80%, and a yellow warning when it is 40%-60%. When the warning level changes, the power supply plan is adjusted dynamically. For example, when a device enters an orange warning, the system automatically increases its backup power capacity and adjusts the power supply path to reduce the scope of the fault.

[0071] The present invention significantly improves the system's anti-interference ability and power supply continuity through multi-level load division, modular energy storage unit configuration and dynamic power distribution. It adopts reinforcement learning algorithm to generate adaptive power distribution strategy, realizes intelligent management of power supply system, effectively reduces energy waste, and realizes early identification and warning of fault risks based on machine learning fault prediction model, providing decision support for system maintenance and fault handling.

[0072] In an optional embodiment,

[0073] The steps of collecting power consumption status data and load distribution data of functional equipment in the micromodule computer room, and establishing a power supply network topology model according to the power consumption status data and the load distribution data; analyzing the power consumption characteristics of the functional equipment based on the power supply network topology model, dividing the power supply system into primary load, secondary load and tertiary load, and configuring independent power supply circuits and data collection units in the primary load, secondary load and tertiary load respectively include:

[0074] The power consumption status data includes voltage effective value, current effective value, active power, reactive power and power factor, and the load distribution data includes equipment spatial position coordinates, rated power and power change data; median filtering is performed on the power consumption status data to obtain filtered data, and a power supply network topology model is constructed based on the filtered data and the load distribution data. The power supply network topology model includes power supply nodes and power supply lines, and the weight parameters of the power supply lines include line impedance parameters and the maximum allowable current of the lines;

[0075] Analyze the power consumption characteristics of functional equipment based on the power supply network topology model, extract characteristic data including rated power, power fluctuation coefficient, allowable interruption time and backup power demand coefficient from the power consumption characteristics, and build a fuzzy clustering model based on the characteristic data; calculate the membership matrix of functional equipment through the fuzzy clustering model, update the cluster center according to the membership matrix, and divide the power supply system into primary load, secondary load and tertiary load based on the cluster center;

[0076] Based on the power supply network topology model, a dual-bus dual-way power supply architecture and an online uninterruptible power supply are configured for the first-level load, a single-bus dual-way power supply architecture and a backup uninterruptible power supply are configured for the second-level load, and a single-way power supply architecture is configured for the third-level load, and a power supply architecture configuration scheme is generated; layout parameters of the power supply line are calculated according to the power supply architecture configuration scheme and the power supply network topology model, the layout parameters include cable specifications, wiring paths and protection device parameters, and independent power supply circuits and data acquisition units are respectively configured in the first-level load, the second-level load and the third-level load according to the layout parameters.

[0077] Exemplarily, first, a sensor network is deployed in the micromodule room to collect power status data and load distribution data of functional equipment. Power status data includes voltage RMS, current RMS, active power, reactive power and power factor. Load distribution data includes equipment spatial position coordinates, rated power and power change data. Intelligent power meters are used to monitor each device in real time, with a sampling frequency of 1 time / second, and power status data is collected continuously for 24 hours. A laser rangefinder is used to obtain the three-dimensional coordinates of the equipment, and the load distribution data is obtained by combining the equipment nameplate information and historical power consumption data.

[0078] Next, the collected power consumption data is processed by median filtering to remove outliers and noise interference. Specifically, a 5-point median filtering algorithm is used, that is, the middle value of every 5 data points is taken as the effective data at that moment. For example, if the current data of a device within 5 seconds is [10.2, 10.1, 15.3, 10.3, 10.2] A, then the effective current value at that moment is 10.2 A.

[0079] Then, the power supply network topology model is constructed based on the filtered power consumption status data and load distribution data. The model consists of two parts: power supply nodes and power supply lines. Power supply nodes represent power distribution cabinets, PDUs and other power supply equipment, and power supply lines represent the connection relationship between nodes. The weight parameters of the line include line impedance and maximum allowable current. All nodes and lines are traversed through the depth-first search algorithm to generate a complete topology structure.

[0080] Based on the constructed power supply network topology model, the power consumption characteristics of functional equipment are analyzed. Feature data are extracted from the power consumption characteristics, including rated power, power fluctuation coefficient, allowable interruption time, and backup power demand coefficient. The rated power is obtained directly from the equipment nameplate; the power fluctuation coefficient is calculated by the ratio of the standard deviation to the mean of the 24-hour power data; the allowable interruption time is divided according to the equipment type, such as 0 minutes for servers and 15 minutes for storage devices; the backup power demand coefficient is determined by whether redundant power is configured, with redundancy being 1 and no redundancy being 2.

[0081] The fuzzy clustering model is constructed using the extracted feature data. First, the feature data is normalized, and the fuzzy C-means clustering algorithm is used to calculate the membership matrix of each device to the three categories. The cluster centers are continuously updated during the iterative optimization process until convergence. Finally, three cluster centers are obtained, representing the first-level load, the second-level load, and the third-level load. Each device is divided into the category with the largest membership.

[0082] Based on the clustering results and the power supply network topology model, the corresponding power supply architecture is configured for different levels of loads. The first-level load is configured with a dual-bus dual-circuit power supply architecture and an online UPS to achieve 2N redundancy; the second-level load is configured with a single-bus dual-circuit power supply architecture and a backup UPS to achieve N+1 redundancy; the third-level load is configured with a single-circuit power supply architecture without a UPS. According to the load power and redundancy requirements, select the UPS and power distribution equipment with appropriate capacity.

[0083] Finally, the layout parameters of the power supply line are calculated according to the power supply architecture configuration scheme and the topology model. The cable specifications are selected according to the maximum allowable current; the wiring path is optimized by the Dijkstra shortest path algorithm; the protection device parameters include the circuit breaker setting value, etc. For example, a 100A power supply line uses a 4×25mm² copper core cable, and the circuit breaker setting current is 125A. Independent power supply circuits are configured in the primary, secondary and tertiary loads, and each circuit is equipped with an intelligent power meter as a data acquisition unit.

[0084] The present invention realizes accurate classification of loads and improves the reliability and economy of the power supply system by constructing a power supply network topology model and analyzing power consumption characteristics. It adopts a fuzzy clustering method to classify loads, overcomes the subjectivity of traditional manual classification methods, improves the scientificity and accuracy of classification, configures a differentiated power supply architecture based on the load classification results, avoids resource waste, and at the same time ensures the power supply reliability of important loads and improves the overall power supply efficiency.

[0085] In an optional embodiment,

[0086] The steps of updating the power supply network topology model according to the power consumption data collected by the data acquisition unit, generating an adaptive power allocation strategy based on the updated power supply network topology model by using a reinforcement learning algorithm, and determining the power supply path and power allocation scheme of the power supply circuit include:

[0087] The power consumption data collected by the data acquisition unit is standardized to obtain standardized data; a three-dimensional state tensor including a voltage matrix, a current matrix and a power matrix is ​​constructed based on the standardized data, and the three-dimensional state tensor is updated by using a double sliding window mechanism to obtain an updated power supply network topology model;

[0088] The voltage deviation rate, line load rate, power allocation ratio and equipment operation status index in the updated power supply network topology model are constructed as a state vector, and the power adjustment amount of the main power supply, the power adjustment amount of the backup power supply and the power allocation adjustment amount between loops are constructed as an action vector; a multi-objective combined reward function including power supply reliability reward items, energy efficiency reward items and stability reward items is constructed, and an adaptive power allocation strategy is trained based on the state vector, action vector and multi-objective combined reward function. The adaptive power allocation strategy is implemented using a deep deterministic policy gradient algorithm, and the deep deterministic policy gradient algorithm includes an action evaluation network and an action execution network, wherein the action evaluation network adopts a dual-Q network structure, and the action execution network adopts a three-layer neural network structure, and the network parameters of the action evaluation network and the action execution network are updated based on priority sampling of the temporal difference error;

[0089] A long short-term memory network is used to predict the load demand for a preset time in the future. The power supply path of the power supply circuit is determined based on the load demand and the adaptive power allocation strategy. The equipment in each circuit is divided into primary load, secondary load and tertiary load according to their importance. The power ratio of the main power supply and the backup power supply is determined according to the power supply path and load level, and power is allocated to each device.

[0090] Exemplarily, the collected power consumption data is first standardized, a three-dimensional state tensor is constructed, and a double sliding window mechanism is used to update the power supply network topology model. Then, based on the updated topology model, the state vector, action vector, and multi-objective combined reward function are constructed, and the deep deterministic policy gradient algorithm is used to train the adaptive power allocation strategy. Finally, the power supply path and power allocation scheme are determined in combination with the load demand prediction results. The power consumption data collected by the data acquisition unit is standardized. Specifically, the collected raw data such as voltage, current, and power are standardized according to a preset normalization formula so that the data is distributed in the interval [0,1].

[0091] Next, a three-dimensional state tensor is constructed based on the standardized data. The tensor contains three dimensions: voltage matrix, current matrix, and power matrix. The rows of each matrix represent time series and the columns represent different nodes. For example, for a system with 100 nodes, a sampling interval of 5 minutes, and a data window of 24 hours, a 288x100x3 three-dimensional tensor can be constructed.

[0092] Then, a double sliding window mechanism is used to update the three-dimensional state tensor. Two time windows, long and short, are set. The short window (such as 1 hour) is used to capture instantaneous changes, and the long window (such as 24 hours) is used to capture long-term trends. Every time new data arrives, the short window moves right as a whole and updates the latest data, while the long window retains historical data and adds new data at the end. In this way, historical information can be retained and system changes can be reflected in a timely manner, and an updated power supply network topology model can be obtained.

[0093] Based on the updated topology model, the state vector, action vector and multi-objective combined reward function are constructed. The state vector contains the voltage deviation rate, line load rate, power allocation ratio and equipment operation status indicators. For example, the voltage deviation rate can be expressed as the percentage of the deviation between the actual voltage and the rated voltage. The action vector contains the power adjustment of the main power supply, the power adjustment of the backup power supply and the power allocation adjustment between the circuits. The multi-objective combined reward function contains power supply reliability reward items, energy efficiency reward items and stability reward items, and these reward items can be combined by weighted summation.

[0094] The deep deterministic policy gradient algorithm is used to train the adaptive power allocation strategy. The algorithm includes an action evaluation network and an action execution network. The action evaluation network adopts a dual Q network structure, which contains two Q networks with the same structure but independent parameters, and is used to evaluate the value of the action. The action execution network adopts a three-layer neural network structure, with the input layer as the state vector, the hidden layer using the ReLU activation function, and the output layer as the action vector. During the training process, the network parameters are updated based on the priority sampling of the temporal difference error. Specifically, an experience replay buffer is maintained, a small batch of data is sampled from it for each training, the temporal difference error is calculated and the sample priority is updated, and then the network parameters are updated based on the error gradient.

[0095] The long short-term memory network is used to predict the load demand of a preset time (such as 24 hours) in the future. The network contains an LSTM layer and a fully connected layer. The input is historical load data and the output is the future load forecast value. The sliding window method is used during training, such as using the data of the past 7 days to predict the load of the next day.

[0096] Based on the predicted load demand and the trained adaptive power allocation strategy, the power supply path of the power supply circuit is determined. First, the power supply demand of each circuit is judged according to the load prediction results. Then, the adaptive power allocation strategy is used to determine the optimal power supply path by considering factors such as voltage deviation and line load.

[0097] The equipment in each circuit is divided into primary load, secondary load and tertiary load according to the importance. Primary load is the most important equipment, such as safety system; secondary load is the less important equipment, such as production equipment; tertiary load is the general equipment, such as lighting system. According to the power supply path and load level, determine the power ratio of the main power supply and the backup power supply. For example, for a circuit with high importance, the ratio of the backup power supply can be increased to improve reliability.

[0098] Finally, power is allocated to each device. According to the load level of the device and the predicted load demand, combined with the adaptive power allocation strategy, the specific power supply of each device is determined. For example, the power supply of the first-level load can be guaranteed first, and then allocated to the second-level and third-level loads within the remaining power range.

[0099] The present invention can respond to load changes in a timely manner, optimize power supply paths and power distribution, effectively reduce voltage fluctuations and line overloads, and improve the overall reliability and stability of the system by updating the power supply network topology model in real time and adopting an adaptive power distribution strategy; based on the deep reinforcement learning algorithm, the system can autonomously learn the optimal power distribution strategy, adapt to different load conditions and network topology changes, and realize intelligent scheduling and control without human intervention; through accurate prediction of load demand and reasonable power distribution, energy waste is avoided and the overall energy efficiency of the power supply network is improved. At the same time, the importance of the equipment is taken into consideration, the power supply quality of key loads is guaranteed, and the optimal configuration of energy utilization is realized.

[0100] In an optional embodiment,

[0101] The steps of constructing a multi-objective combined reward function including power supply reliability reward items, energy efficiency reward items and stability reward items include:

[0102] The power supply reliability reward item is calculated based on the voltage deviation rate and power loss of each level of load, the energy efficiency reward item is calculated based on the deviation between the line load rate and the optimal load rate, and the stability reward item is calculated based on the power adjustment amount at adjacent moments;

[0103] A judgment matrix is ​​constructed by using a hierarchical analysis method and a characteristic vector is calculated to obtain a basic weight value of each reward item in the multi-objective combination reward function, and a weight adjustment range of each reward item is set based on the basic weight value;

[0104] The operating status of the modular power supply system is obtained, including the degree of voltage deviation exceeding the limit, the degree of load rate deviation from the optimal value, and the power fluctuation amplitude. Based on the operating status of the modular power supply system, the Sigmoid mapping function is used to dynamically adjust the weight of each reward item. When the degree of voltage deviation exceeding the limit is greater than the voltage threshold, the weight of the power supply reliability reward item is increased; when the degree of load rate deviation from the optimal value is greater than the load rate threshold, the weight of the energy efficiency reward item is increased; when the power fluctuation amplitude is greater than the power threshold, the weight of the stability reward item is increased.

[0105] For example, first, calculate the power supply reliability bonus. Calculate the voltage deviation rate and power loss of each level of load in the modular power supply system. Take the 10kV distribution line as an example, select 5 key nodes, measure their voltage values ​​and compare them with the rated voltage of 10kV to calculate the voltage deviation rate. Measure the line power loss at the same time. Perform a weighted sum of the voltage deviation rate and the power loss to obtain the power supply reliability bonus value.

[0106] Secondly, calculate the energy efficiency bonus. Measure the load rate of each line in the system and compare it with the preset optimal load rate (such as 70%) to calculate the load rate deviation. Taking a 10kV line as an example, the actual load rate is 60%, and the deviation from the optimal load rate of 70% is 10%. Take the weighted average of the load rate deviations of each line to obtain the energy efficiency bonus value.

[0107] Next, calculate the stability bonus. Compare the power output of each power generation unit in the system at two adjacent moments and calculate the power adjustment. Take a photovoltaic power generation unit as an example. The output power at time t is 500kW, and the output power at time t+1 is 550kW, so the power adjustment is 50kW. Take the weighted sum of the power adjustment of all power generation units to get the value of the stability bonus.

[0108] Next, the judgment matrix is ​​constructed and the eigenvector is calculated using the hierarchical analysis method to obtain the basic weight values ​​of the three reward items. Assume that the basic weight values ​​of the three reward items of power supply reliability, energy efficiency, and stability are calculated to be 0.5, 0.3, and 0.2, respectively. Based on these basic weight values, the weight adjustment range of each reward item is set, such as the weight adjustment range of the power supply reliability reward item is 0.4-0.6, the energy efficiency reward item is 0.2-0.4, and the stability reward item is 0.1-0.3.

[0109] Then, the real-time operating status of the modular power supply system is obtained, including the voltage deviation exceeding the limit, the load rate deviation from the optimal value and the power fluctuation range. Taking the system status at a certain moment as an example, the maximum voltage deviation rate is 6% (exceeding the 5% threshold), the maximum load rate deviation is 15% (exceeding the 10% threshold), and the maximum power fluctuation range is 8% (not exceeding the 10% threshold).

[0110] Finally, based on the above operating status, the Sigmoid mapping function is used to dynamically adjust the weights of each reward item. Since the voltage deviation and load rate deviation both exceed the threshold, the weights of the power supply reliability reward item and the energy efficiency reward item are increased. After the Sigmoid function mapping, the final weights of the three reward items are adjusted to 0.55, 0.35, and 0.1, respectively. Multiply these weights with the previously calculated values ​​of each reward item and sum them up to get the final multi-objective combined reward function value.

[0111] The present invention constructs a multi-objective combined reward function that comprehensively reflects the operating quality of a modular power supply system by comprehensively considering the three aspects of power supply reliability, energy efficiency and stability, thereby providing a scientific basis for system optimization control; a hierarchical analysis method is used to determine the basic weights, and a weight adjustment range is set, thereby ensuring the rationality of the weights of each reward item and providing flexibility for dynamic adjustment, so that the reward function can adapt to different system operating states; a Sigmoid mapping function is introduced to realize dynamic adjustment of the weights, and the weights of each reward item are finely adjusted according to the real-time operating state of the system, thereby improving the sensitivity and adaptability of the reward function to changes in the system state, thereby better guiding system optimization control.

[0112] In an optional embodiment,

[0113] The step of training the adaptive power allocation strategy based on the state vector, the action vector and the multi-objective combined reward function comprises:

[0114] Obtain the current state of the power supply system, generate power adjustment actions based on the current strategy, execute the power adjustment actions and observe the response of the power supply system, calculate the multi-objective combined reward function including power supply reliability reward items, energy efficiency reward items and stability reward items, and record the current state of the power supply system, power adjustment actions, calculation results of the multi-objective combined reward function and the state of the power supply system after response as state transition data and store them in the experience pool;

[0115] The training samples are obtained from the experience pool according to the priority of the temporal difference error, the temporal difference error is calculated based on the training samples, the temporal difference error is input into the action evaluation network and the action execution network respectively, the network weights of the action evaluation network and the network parameters of the action execution network are updated, and the target network is updated by a soft update method;

[0116] During the training process, exploration noise is set to conduct strategy exploration, the network learning rate is dynamically adjusted and the training batch size is optimized; the average reward value change trend, network parameter update amplitude and strategy performance indicators during the training process are monitored and recorded. When the training round reaches the preset value and the reward value converges and stabilizes, the verification data set is used to verify the training results. When the verification result meets the performance requirements, the training process is completed.

[0117] For example, the current state information of the power supply system is first obtained, including parameters such as node voltage, line power flow, generator output, etc. Taking a 110kV distribution network as an example, there are 10 nodes, 5 lines, and 2 distributed generators. The current state vector contains 10 node voltage values, 5 line power flow values, and 2 generator output values, a total of 17 parameters.

[0118] Then, based on the currently trained power adjustment strategy, the corresponding power adjustment action is generated. The power adjustment action includes the adjustment amount of the output of the two distributed generators. For example, the generated action vector is [+0.5MW, -0.3MW], which means that the output of generator No. 1 is increased by 0.5MW and the output of generator No. 2 is reduced by 0.3MW.

[0119] Execute the generated power adjustment action and observe the response of the power supply system. Record the changes of parameters such as node voltage and line power flow after adjustment.

[0120] Next, calculate the multi-objective combined reward function. This function contains three items: Power supply reliability reward item: calculated based on node voltage deviation, line overload, etc. The higher the voltage qualification rate and the less overload, the higher the reward. Energy efficiency reward item: calculated based on network loss, power generation cost, etc. The lower the network loss and the lower the power generation cost, the higher the reward. Stability reward item: calculated based on voltage fluctuation, power fluctuation, etc. The smaller the fluctuation, the higher the reward.

[0121] The weighted sum of the three rewards is used to get the final reward value. For example, assuming the weights are 0.4, 0.4, and 0.2, and the three reward values ​​are 0.8, 0.6, and 0.7, the final reward value is 0.4*0.8+0.4*0.6+0.2*0.7=0.7.

[0122] The current state of the power supply system, the power adjustment action, the calculated reward value, and the adjusted system state are stored in the experience replay pool as a set of state transition data. The capacity of the experience pool is set to 10,000 data.

[0123] Sample training samples from the experience pool. When sampling, set the priority according to the size of the time series difference error. The larger the error, the higher the probability of being sampled. For example, if the sampling batch size is set to 32, the 32 sets of data with the largest time series difference error are selected from the experience pool as training samples.

[0124] The temporal difference error is calculated based on the sampled training samples. The temporal difference error reflects the gap between the prediction of future rewards and the actual rewards. The calculated temporal difference error is input into the action evaluation network and the action execution network respectively.

[0125] The action evaluation network is used to evaluate the value of the action. Its input is the state vector and its output is the Q value of the action. The action execution network is used to generate specific actions. Its input is the state vector and its output is the action vector. Both networks use a multi-layer perceptron structure, which contains 3 hidden layers and 64 neurons in each layer.

[0126] The weights of the action evaluation network and the parameters of the action execution network are updated based on the temporal difference error. The stochastic gradient descent method is used to update the parameters, and the learning rate is initially set to 0.001. After every 100 updates, the learning rate is multiplied by 0.9 for decay.

[0127] The target network is updated by soft update. The target network is a backup of the action evaluation network and the action execution network, and is used to calculate the target Q value. Each time an update is performed, the parameters of the evaluation network and the execution network are copied to the target network at a ratio of 0.001.

[0128] During the training process, exploration noise is set to explore the strategy. Gaussian noise is used, and the initial standard deviation is set to 0.3, which gradually decays as the training progresses. The addition of noise can increase the exploratory nature of the strategy and avoid falling into the local optimum.

[0129] Dynamically adjust the network learning rate and optimize the training batch size. When the reward value does not increase significantly for 5 consecutive cycles, reduce the learning rate to 0.8 times the original. When the reward value fluctuates significantly, increase the training batch size, for example, from 32 to 64.

[0130] Monitor and record key indicators during the training process, including the trend of average reward value changes, the amplitude of network parameter updates, and the performance indicators of the strategy in the test environment. Draw relevant curves to facilitate the analysis of the training process.

[0131] When the number of training rounds reaches the preset value (such as 10,000 rounds) and the average reward value fluctuation in the last 100 rounds is less than 5%, the reward value is considered to be converged and stable. At this time, an independent validation data set is used to verify the training results. The validation data set contains 1,000 sets of different initial states.

[0132] The trained strategy is run on the validation data set to calculate performance indicators such as average reward value, power supply reliability index, energy efficiency index, etc. When these indicators meet the pre-set requirements (for example, the average reward value is greater than 0.8, and the power supply reliability index is greater than 0.95), the training process is completed and the final adaptive power allocation strategy is obtained.

[0133] The present invention realizes the comprehensive optimization of multiple objectives such as power supply reliability, energy efficiency and stability through the design of a multi-objective combined reward function, thereby improving the overall performance of the power allocation strategy; adopts an experience replay mechanism based on priority sampling to improve the utilization efficiency of training samples and accelerate the convergence speed of the strategy; introduces exploration noise and dynamic adjustment mechanisms to enhance the adaptability and robustness of the strategy, so that it can better cope with complex and changeable power supply environments.

[0134] In an optional embodiment,

[0135] A modular energy storage unit is set based on the power supply path and power distribution scheme of the power supply circuit. The modular energy storage unit includes a super capacitor module and a lithium battery module. The charging and discharging process of the super capacitor module and the lithium battery module are controlled by an intelligent energy management algorithm. The steps of providing backup power for the primary load and the secondary load include:

[0136] According to the peak power demand, energy demand and backup time requirement of the primary load and the secondary load, an energy storage capacity optimization model is established, and the capacity configuration of the supercapacitor module and the lithium battery module is determined based on the energy storage capacity optimization model, wherein the constraint conditions of the energy storage capacity optimization model include: the total energy of the supercapacitor module and the lithium battery module is greater than or equal to the total energy required by the system, and the total power of the supercapacitor module and the lithium battery module is greater than or equal to the peak power demand;

[0137] The supercapacitor module is composed of single supercapacitors connected in series and in parallel to form a supercapacitor group, and the lithium battery module is composed of lithium iron phosphate batteries connected in series to form a lithium battery group; the supercapacitor module and the lithium battery module are both connected to the busbar through a bidirectional DC / DC converter, wherein the supercapacitor module adopts an interleaved parallel Buck-Boost topology, the lithium battery module adopts a full-bridge LLC resonant conversion topology, and the busbar adopts a three-level converter topology;

[0138] The main power supply is monitored in real time using a fault detection method based on voltage sag. When a main power supply fault is detected, the supercapacitor module is triggered to discharge, and the lithium battery module is started to enter the standby state, and power switching is achieved based on synchronous phase locking technology.

[0139] An intelligent energy management algorithm is designed, which includes power distribution control and mode switching control; the power distribution control adopts a fuzzy adaptive PID control strategy, establishes a fuzzy rule base based on the load power change rate, the supercapacitor module voltage deviation and the lithium battery module charge state deviation, realizes adaptive adjustment of the PID controller through online parameter setting, and dynamically distributes the output power of the supercapacitor module and the lithium battery module; the mode switching control includes a normal operating mode, a fast response mode and an energy recovery mode, wherein in the normal operating mode, the supercapacitor module and the lithium battery module distribute power according to a preset ratio, in the fast response mode, the supercapacitor module preferentially responds to load power mutations, and in the energy recovery mode, the supercapacitor module preferentially absorbs regenerative braking energy.

[0140] Exemplarily, first, according to the peak power demand, energy demand and backup time requirements of the primary load and the secondary load, an energy storage capacity optimization model is established. The constraints of the model include: the total energy of the supercapacitor module and the lithium battery module is greater than or equal to the total energy required by the system, and the total power of the supercapacitor module and the lithium battery module is greater than or equal to the peak power demand. Taking a certain actual system as an example, the peak power of the primary load is 100kW, the energy demand is 50kWh, and the backup time is 30 minutes; the peak power of the secondary load is 50kW, the energy demand is 25kWh, and the backup time is 15 minutes. Based on these parameters, it can be concluded that the total energy required by the system is 75kWh and the peak power demand is 150kW.

[0141] Based on the above energy storage capacity optimization model, the capacity configuration of the supercapacitor module and the lithium battery module is determined through iterative calculation. The calculation results show that the capacity of the supercapacitor module is 10kWh and the capacity of the lithium battery module is 65kWh, which can meet the system requirements.

[0142] Next, the supercapacitor module and lithium battery module are specifically designed. The supercapacitor module is composed of single supercapacitors connected in series and in parallel to form a supercapacitor group. Taking a 3000F, 2.7V single supercapacitor as an example, 160 are connected in series to form a submodule, and then 10 submodules are connected in parallel to form a supercapacitor group with a total capacity of about 10kWh. The lithium battery module is composed of lithium iron phosphate batteries connected in series to form a lithium battery group. Select 100Ah, 3.2V single batteries, and connect 200 in series to form a lithium battery group with a total capacity of 65kWh.

[0143] Both the supercapacitor module and the lithium battery module are connected to the busbar through a bidirectional DC / DC converter. The supercapacitor module adopts an interleaved parallel Buck-Boost topology to achieve bidirectional energy flow and wide-range voltage conversion. The lithium battery module adopts a full-bridge LLC resonant conversion topology with soft switching characteristics to reduce switching losses. The busbar adopts a three-level converter topology to improve power density and efficiency.

[0144] During system operation, a voltage sag-based fault detection method is used to monitor the main power supply in real time. In specific implementation, the voltage sag threshold can be set to 90% of the rated voltage, and the sampling period is 1ms. When the voltage of the main power supply is detected to be lower than the threshold and lasts for more than 5ms, it is determined that a fault has occurred. At this time, the supercapacitor module is triggered to discharge, and the lithium battery module is started to enter the standby state. In order to achieve smooth switching, synchronous phase locking technology is used. This technology tracks the phase and frequency information of the main power supply through a phase-locked loop circuit, maintains the continuity of the output voltage waveform at the moment of switching, and thus avoids the impact of voltage jumps on the load.

[0145] In order to achieve efficient coordinated control of modular energy storage units, an intelligent energy management algorithm is designed. The algorithm includes two parts: power distribution control and mode switching control.

[0146] The power distribution control adopts the fuzzy adaptive PID control strategy. First, a fuzzy rule base is established. The input variables include the load power change rate, the voltage deviation of the supercapacitor module and the charge state deviation of the lithium battery module. The output variables are the proportional, integral and differential parameters of the PID controller. Taking the load power change rate as an example, it can be divided into five fuzzy subsets: "large negative", "small negative", "zero", "small positive" and "large positive". The adjustment amount of the PID parameters is obtained through fuzzy reasoning to realize the online parameter setting of the controller.

[0147] Based on the adjusted PID controller, the output power of the supercapacitor module and the lithium battery module are dynamically allocated. Taking the load power of 120kW at a certain moment as an example, after fuzzy adaptive PID control, the output power of the supercapacitor module is 80kW and the output power of the lithium battery module is 40kW.

[0148] Mode switching control includes normal operation mode, fast response mode and energy recovery mode. In normal operation mode, the supercapacitor module and the lithium battery module distribute power according to a preset ratio, such as 6:4. When a sudden change in load power is detected, switch to the fast response mode. In this mode, the supercapacitor module responds to load power changes first, such as assuming 80% of the power change. When the system is in regenerative braking state, switch to energy recovery mode. At this time, the supercapacitor module absorbs regenerative braking energy first, such as absorbing 90% of the recovered energy.

[0149] This paper can effectively cope with load power fluctuations and main power failures and ensure the continuous power supply capacity of the system by reasonably configuring the capacity of supercapacitor modules and lithium battery modules and combining them with intelligent energy management algorithms; it adopts fuzzy adaptive PID control strategy to realize power distribution, and can adjust control parameters in real time according to the system status to improve the dynamic response speed and control accuracy of the system; through multi-mode switching control, the power of supercapacitor modules and lithium battery modules can be reasonably allocated under different working conditions, giving full play to their respective advantages and realizing efficient utilization and recovery of energy.

[0150] In an optional embodiment,

[0151] A machine learning fault prediction model is built based on power consumption data. The fault risk probability is used as the trigger condition for emergency response. A graded warning threshold is set. When the warning level changes, the power supply path and power distribution plan of the power supply circuit are dynamically adjusted. The steps include:

[0152] Collect electricity consumption data including voltage fluctuation rate, current harmonic content and power factor, perform feature extraction and data preprocessing on the electricity consumption data, input the preprocessed electricity consumption data into the LSTM-GRU hybrid network to establish a machine learning fault prediction model, wherein the LSTM unit extracts the long-term dependency features of the electricity consumption data, and the GRU unit extracts the short-term fluctuation features. An attention mechanism is introduced into the network to assign weight coefficients to features of different time scales, and a new feature representation is obtained by weighted calculation of the attention weight and the original feature vector, and the new feature representation is nonlinearly mapped through the Softmax function to obtain a fault probability value;

[0153] The fault probability values ​​are divided into different warning levels, and a risk assessment indicator system is established for each warning level. According to the risk assessment results, corresponding load transfer strategies, power switching strategies and power limitation strategies are set to form a hierarchical warning response mechanism.

[0154] Monitor changes in warning levels in real time. When the warning level changes, dynamically adjust the power supply path and power allocation plan of the power supply circuit according to the preset priority strategy and constraints based on the real-time operating status of the power supply system, the importance of the equipment and the load fluctuation characteristics.

[0155] For example, first collect power consumption data such as voltage fluctuation rate, current harmonic content and power factor. Perform feature extraction and preprocessing on the collected data, including removing outliers and normalization. Input the preprocessed data into the LSTM-GRU hybrid network for training to build a fault prediction model.

[0156] In the LSTM-GRU hybrid network, the LSTM unit is mainly responsible for extracting the long-term dependent features of electricity consumption data, such as seasonal change trends. The GRU unit focuses on extracting short-term fluctuation features, such as sudden load changes. The attention mechanism is introduced into the network to assign different weight coefficients to features of different time scales. Specifically, the additive attention mechanism can be used to calculate the similarity between the query vector and the key vector to obtain the attention score, and then perform weighted summation on the value vector. The new feature representation is obtained by weighted calculation of the attention weight and the original feature vector. Finally, the feature representation is nonlinearly mapped through the Softmax function to output the fault probability value.

[0157] For example, for the historical power consumption data of a substation, the voltage, current, power factor and other data were collected every 15 minutes in the past three years. After preprocessing, they were input into the LSTM-GRU hybrid network, which contains 2 layers of LSTM, 2 layers of GRU and 1 layer of fully connected layer. The network was trained by the back propagation algorithm, and finally achieved a fault prediction accuracy of 95% on the test set.

[0158] Next, the predicted failure probability values ​​are divided into different warning levels, such as low risk (0-30%), medium risk (30%-70%) and high risk (70%-100%). For each warning level, a corresponding risk assessment indicator system is established. The low risk level mainly considers long-term factors such as equipment aging and maintenance cycle; the medium risk level adds medium-term factors such as load fluctuation and power quality; the high risk level further incorporates short-term indicators such as real-time operating status and environmental factors.

[0159] Based on the risk assessment results, corresponding load transfer strategies, power switching strategies and power limitation strategies are formulated to form a hierarchical early warning response mechanism. For example, for low-risk warnings, daily maintenance plans can be adjusted; for medium-risk warnings, backup power supplies can be started to transfer some non-critical loads; for high-risk warnings, high-risk lines need to be urgently cut off and the use of high-power loads must be limited.

[0160] In actual operation, the system monitors the changes in the warning level in real time. When the warning level changes, the power supply path and power distribution plan of the power supply circuit are dynamically adjusted according to the preset priority strategy and constraints based on the real-time operation status of the power supply system, the importance of the equipment and the load fluctuation characteristics.

[0161] Specifically, first obtain the current system topology, the load of each circuit, the status of the equipment and other information. Then, according to the change of the warning level, determine the target that needs to be adjusted. If the warning level rises from low to medium, it may be necessary to increase the proportion of backup power supply. Then, based on graph theory algorithms such as Dijkstra algorithm, calculate the feasible power supply path plan. At the same time, use optimization algorithms such as genetic algorithms to solve the optimal power allocation plan that meets various constraints. Finally, send the calculated new plan to the relevant control equipment to complete the dynamic adjustment of the power supply path and power allocation.

[0162] For example, the power supply system of an industrial park includes 2 main transformers, 3 busbars and 10 outgoing line circuits. When the system detects that the failure probability of main transformer No. 1 has increased from 25% to 65%, the warning level has increased from low risk to medium risk. The system immediately starts the adjustment process, transferring part of the load of main transformer No. 1 to main transformer No. 2, while increasing the output of the backup generator to ensure power supply reliability. The specific adjustment plan includes: transferring the power supply of outgoing line circuit No. 2 from main transformer No. 1 to main transformer No. 2; reducing the power supply of circuits No. 3 and No. 4 by 20% each, supplemented by backup generators; other circuits remain unchanged. The entire adjustment process was completed within 3 minutes, effectively reducing the risk of system failure.

[0163] The present invention effectively extracts the long-term and short-term characteristics of electricity consumption data by adopting the LSTM-GRU hybrid network and attention mechanism, and improves the model's ability to identify various fault types. At the same time, real-time monitoring and dynamic early warning mechanisms ensure that the system can detect potential risks in a timely manner; by establishing a graded early warning mechanism and risk assessment system, differentiated response strategies are adopted for different risk levels, which not only avoids safety hazards caused by insufficient response, but also prevents waste of resources caused by excessive response, and improves the operating efficiency of the system; the dynamic adjustment mechanism based on real-time status enables the system to quickly adjust the power supply plan according to the changes in the early warning level, and effectively respond to various emergency situations. This flexible scheduling strategy significantly improves the system's anti-interference ability and reliability.

[0164] Figure 2 FIG. 1 is a schematic diagram of a modular power supply system configuration system of a micro-module computer room according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0165] The first unit is used to collect power consumption status data and load distribution data of functional equipment in the micromodule room, and establish a power supply network topology model based on the power consumption status data and load distribution data; analyze the power consumption characteristics of functional equipment based on the power supply network topology model, divide the power supply system into primary load, secondary load and tertiary load, and configure independent power supply circuits and data acquisition units in the primary load, secondary load and tertiary load respectively;

[0166] The second unit is used to update the power supply network topology model according to the power consumption data collected by the data collection unit, and based on the updated power supply network topology model, use the reinforcement learning algorithm to generate an adaptive power allocation strategy to determine the power supply path and power allocation scheme of the power supply circuit; based on the power supply path and power allocation scheme of the power supply circuit, set a modular energy storage unit, the modular energy storage unit includes a super capacitor module and a lithium battery module, and the charging and discharging process of the super capacitor module and the lithium battery module is controlled by an intelligent energy management algorithm to provide backup power for the primary load and the secondary load;

[0167] The third unit is used to build a machine learning fault prediction model based on electricity consumption data, use the fault risk probability as the trigger condition for emergency response, set the graded warning threshold, and dynamically adjust the power supply path and power distribution plan of the power supply circuit when the warning level changes.

[0168] According to a third aspect of the embodiments of the present invention,

[0169] An electronic device is provided, comprising:

[0170] processor;

[0171] a memory for storing processor-executable instructions;

[0172] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0173] According to a fourth aspect of the embodiments of the present invention,

[0174] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0175] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0176] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A modular power supply system configuration method for a micro-module computer room, characterized in that: include: Collect power consumption status data and load distribution data of functional equipment in the micromodule room, and establish a power supply network topology model based on the power consumption status data and load distribution data; analyze the power consumption characteristics of functional equipment based on the power supply network topology model, divide the power supply system into primary load, secondary load and tertiary load, and configure independent power supply circuits and data acquisition units in the primary load, secondary load and tertiary load respectively; The power supply network topology model is updated according to the power consumption data collected by the data acquisition unit. Based on the updated power supply network topology model, a reinforcement learning algorithm is used to generate an adaptive power allocation strategy to determine the power supply path and power allocation scheme of the power supply circuit; a modular energy storage unit is set based on the power supply path and power allocation scheme of the power supply circuit, wherein the modular energy storage unit includes a supercapacitor module and a lithium battery module, and the charging and discharging process of the supercapacitor module and the lithium battery module is controlled by an intelligent energy management algorithm to provide backup power for the primary load and the secondary load; Build a machine learning fault prediction model based on power consumption data, use the fault risk probability as the trigger condition for emergency response, set graded warning thresholds, and dynamically adjust the power supply path and power distribution plan of the power supply circuit when the warning level changes; The steps of updating the power supply network topology model according to the power consumption data collected by the data acquisition unit, generating an adaptive power allocation strategy based on the updated power supply network topology model by using a reinforcement learning algorithm, and determining the power supply path and power allocation scheme of the power supply circuit include: The power consumption data collected by the data acquisition unit is standardized to obtain standardized data; a three-dimensional state tensor including a voltage matrix, a current matrix and a power matrix is ​​constructed based on the standardized data, and the three-dimensional state tensor is updated by using a double sliding window mechanism to obtain an updated power supply network topology model; The voltage deviation rate, line load rate, power allocation ratio and equipment operation status index in the updated power supply network topology model are constructed as a state vector, and the power adjustment amount of the main power supply, the power adjustment amount of the backup power supply and the power allocation adjustment amount between loops are constructed as an action vector; a multi-objective combined reward function including power supply reliability reward items, energy efficiency reward items and stability reward items is constructed, and an adaptive power allocation strategy is trained based on the state vector, action vector and multi-objective combined reward function. The adaptive power allocation strategy is implemented using a deep deterministic policy gradient algorithm, and the deep deterministic policy gradient algorithm includes an action evaluation network and an action execution network, wherein the action evaluation network adopts a dual-Q network structure, and the action execution network adopts a three-layer neural network structure, and the network parameters of the action evaluation network and the action execution network are updated based on priority sampling of the temporal difference error; A long short-term memory network is used to predict the load demand for a preset time in the future. The power supply path of the power supply circuit is determined based on the load demand and the adaptive power allocation strategy. The equipment in each circuit is divided into primary load, secondary load and tertiary load according to their importance. The power ratio of the main power supply and the backup power supply is determined according to the power supply path and load level, and power is allocated to each device.

2. The method according to claim 1, characterized in that The steps of collecting power consumption status data and load distribution data of functional equipment in the micromodule computer room, and establishing a power supply network topology model according to the power consumption status data and the load distribution data; analyzing the power consumption characteristics of the functional equipment based on the power supply network topology model, dividing the power supply system into primary load, secondary load and tertiary load, and configuring independent power supply circuits and data collection units in the primary load, secondary load and tertiary load respectively include: The power consumption status data includes voltage effective value, current effective value, active power, reactive power and power factor, and the load distribution data includes equipment spatial position coordinates, rated power and power change data; median filtering is performed on the power consumption status data to obtain filtered data, and a power supply network topology model is constructed based on the filtered data and the load distribution data. The power supply network topology model includes power supply nodes and power supply lines, and the weight parameters of the power supply lines include line impedance parameters and the maximum allowable current of the lines; Analyze the power consumption characteristics of functional equipment based on the power supply network topology model, extract characteristic data including rated power, power fluctuation coefficient, allowable interruption time and backup power demand coefficient from the power consumption characteristics, and build a fuzzy clustering model based on the characteristic data; calculate the membership matrix of functional equipment through the fuzzy clustering model, update the cluster center according to the membership matrix, and divide the power supply system into primary load, secondary load and tertiary load based on the cluster center; Based on the power supply network topology model, a dual-bus dual-way power supply architecture and an online uninterruptible power supply are configured for the first-level load, a single-bus dual-way power supply architecture and a backup uninterruptible power supply are configured for the second-level load, and a single-way power supply architecture is configured for the third-level load, and a power supply architecture configuration scheme is generated; layout parameters of the power supply line are calculated according to the power supply architecture configuration scheme and the power supply network topology model, the layout parameters include cable specifications, wiring paths and protection device parameters, and independent power supply circuits and data acquisition units are respectively configured in the first-level load, the second-level load and the third-level load according to the layout parameters.

3. The method according to claim 1, characterized in that The steps of constructing a multi-objective combined reward function including power supply reliability reward items, energy efficiency reward items and stability reward items include: The power supply reliability reward item is calculated based on the voltage deviation rate and power loss of each level of load, the energy efficiency reward item is calculated based on the deviation between the line load rate and the optimal load rate, and the stability reward item is calculated based on the power adjustment amount at adjacent moments; A judgment matrix is ​​constructed by using a hierarchical analysis method and a characteristic vector is calculated to obtain a basic weight value of each reward item in the multi-objective combination reward function, and a weight adjustment range of each reward item is set based on the basic weight value; The operating status of the modular power supply system is obtained, including the degree of voltage deviation exceeding the limit, the degree of load rate deviation from the optimal value, and the power fluctuation amplitude. Based on the operating status of the modular power supply system, the Sigmoid mapping function is used to dynamically adjust the weight of each reward item. When the degree of voltage deviation exceeding the limit is greater than the voltage threshold, the weight of the power supply reliability reward item is increased; when the degree of load rate deviation from the optimal value is greater than the load rate threshold, the weight of the energy efficiency reward item is increased; when the power fluctuation amplitude is greater than the power threshold, the weight of the stability reward item is increased.

4. The method according to claim 1, characterized in that: The step of training the adaptive power allocation strategy based on the state vector, the action vector and the multi-objective combined reward function comprises: Obtain the current state of the power supply system, generate power adjustment actions based on the current strategy, execute the power adjustment actions and observe the response of the power supply system, calculate the multi-objective combined reward function including power supply reliability reward items, energy efficiency reward items and stability reward items, and record the current state of the power supply system, power adjustment actions, calculation results of the multi-objective combined reward function and the state of the power supply system after response as state transition data and store them in the experience pool; The training samples are obtained from the experience pool according to the priority of the temporal difference error, the temporal difference error is calculated based on the training samples, the temporal difference error is input into the action evaluation network and the action execution network respectively, the network weights of the action evaluation network and the network parameters of the action execution network are updated, and the target network is updated by a soft update method; During the training process, exploration noise is set to conduct strategy exploration, the network learning rate is dynamically adjusted and the training batch size is optimized; the average reward value change trend, network parameter update amplitude and strategy performance indicators during the training process are monitored and recorded. When the training round reaches the preset value and the reward value converges and stabilizes, the verification data set is used to verify the training results. When the verification result meets the performance requirements, the training process is completed.

5. The method according to claim 1, characterized in that A modular energy storage unit is set based on the power supply path and power distribution scheme of the power supply circuit. The modular energy storage unit includes a super capacitor module and a lithium battery module. The charging and discharging process of the super capacitor module and the lithium battery module are controlled by an intelligent energy management algorithm. The steps of providing backup power for the primary load and the secondary load include: According to the peak power demand, energy demand and backup time requirement of the primary load and the secondary load, an energy storage capacity optimization model is established, and the capacity configuration of the supercapacitor module and the lithium battery module is determined based on the energy storage capacity optimization model, wherein the constraint conditions of the energy storage capacity optimization model include: the total energy of the supercapacitor module and the lithium battery module is greater than or equal to the total energy required by the system, and the total power of the supercapacitor module and the lithium battery module is greater than or equal to the peak power demand; The supercapacitor module is composed of single supercapacitors connected in series and in parallel to form a supercapacitor group, and the lithium battery module is composed of lithium iron phosphate batteries connected in series to form a lithium battery group; the supercapacitor module and the lithium battery module are both connected to the busbar through a bidirectional DC / DC converter, wherein the supercapacitor module adopts an interleaved parallel Buck-Boost topology, the lithium battery module adopts a full-bridge LLC resonant conversion topology, and the busbar adopts a three-level converter topology; The main power supply is monitored in real time using a fault detection method based on voltage sag. When a main power supply fault is detected, the supercapacitor module is triggered to discharge, and the lithium battery module is started to enter the standby state, and power switching is achieved based on synchronous phase locking technology. An intelligent energy management algorithm is designed, which includes power distribution control and mode switching control; the power distribution control adopts a fuzzy adaptive PID control strategy, establishes a fuzzy rule base based on the load power change rate, the supercapacitor module voltage deviation and the lithium battery module charge state deviation, realizes adaptive adjustment of the PID controller through online parameter setting, and dynamically distributes the output power of the supercapacitor module and the lithium battery module; the mode switching control includes a normal operating mode, a fast response mode and an energy recovery mode, wherein in the normal operating mode, the supercapacitor module and the lithium battery module distribute power according to a preset ratio, in the fast response mode, the supercapacitor module preferentially responds to load power mutations, and in the energy recovery mode, the supercapacitor module preferentially absorbs regenerative braking energy.

6. The method according to claim 1, characterized in that A machine learning fault prediction model is built based on power consumption data. The fault risk probability is used as the trigger condition for emergency response. A graded warning threshold is set. When the warning level changes, the power supply path and power distribution plan of the power supply circuit are dynamically adjusted. The steps include: Collect electricity consumption data including voltage fluctuation rate, current harmonic content and power factor, perform feature extraction and data preprocessing on the electricity consumption data, input the preprocessed electricity consumption data into the LSTM-GRU hybrid network to establish a machine learning fault prediction model, wherein the LSTM unit extracts the long-term dependency features of the electricity consumption data, and the GRU unit extracts the short-term fluctuation features. An attention mechanism is introduced into the network to assign weight coefficients to features of different time scales, and a new feature representation is obtained by weighted calculation of the attention weight and the original feature vector, and the new feature representation is nonlinearly mapped through the Softmax function to obtain a fault probability value; The fault probability values ​​are divided into different warning levels, and a risk assessment indicator system is established for each warning level. According to the risk assessment results, corresponding load transfer strategies, power switching strategies and power limitation strategies are set to form a hierarchical warning response mechanism. Monitor changes in warning levels in real time. When the warning level changes, dynamically adjust the power supply path and power allocation plan of the power supply circuit according to the preset priority strategy and constraints based on the real-time operating status of the power supply system, the importance of the equipment and the load fluctuation characteristics.

7. A modular power supply system configuration system for a micro-module computer room, used to implement the method described in any one of claims 1 to 6, characterized in that: include: The first unit is used to collect power consumption status data and load distribution data of functional equipment in the micromodule room, and establish a power supply network topology model based on the power consumption status data and load distribution data; analyze the power consumption characteristics of functional equipment based on the power supply network topology model, divide the power supply system into primary load, secondary load and tertiary load, and configure independent power supply circuits and data acquisition units in the primary load, secondary load and tertiary load respectively; The second unit is used to update the power supply network topology model according to the power consumption data collected by the data collection unit, and based on the updated power supply network topology model, use the reinforcement learning algorithm to generate an adaptive power allocation strategy to determine the power supply path and power allocation scheme of the power supply circuit; based on the power supply path and power allocation scheme of the power supply circuit, set a modular energy storage unit, the modular energy storage unit includes a super capacitor module and a lithium battery module, and the charging and discharging process of the super capacitor module and the lithium battery module is controlled by an intelligent energy management algorithm to provide backup power for the primary load and the secondary load; The third unit is used to build a machine learning fault prediction model based on electricity consumption data, use the fault risk probability as the trigger condition for emergency response, set the graded warning threshold, and dynamically adjust the power supply path and power distribution plan of the power supply circuit when the warning level changes.

8. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 6 is implemented.

Images