Configuration method and system for overall electrical scheme of machine room

Abstract

The invention provides a configuration method and system for an overall electrical scheme of a machine room, and relates to the technical field of electrical planning, and the method comprises the steps: collecting multi-source heterogeneous data, carrying out feature extraction and semantic mapping, constructing a high-dimensional semantic association network, carrying out semantic enhancement, semantic modeling and logic association, extracting an abnormal mode, and obtaining a machine room operation state portrait; determining a machine room energy consumption rule and high-energy-consumption equipment, carrying out composite characteristic analysis to obtain a machine room electrical load curve, carrying out hierarchical progressive prediction in combination with exogenous variables, constructing a target function, solving to obtain optimal configuration parameters, and generating a power distribution system design scheme; a self-learning controller is added in a power distribution unit, the electrical load of a machine room is optimized, a power distribution equipment degradation mechanism model is constructed, service life prediction is performed, a power distribution equipment service life management strategy is output, equipment running abnormally is identified, abnormal reasons are judged by referring to historical abnormal data, and an abnormal report is generated and sent to equipment maintenance personnel.

Description

Technical Field

[0001] The present invention relates to the technical field of electrical planning, and in particular to a configuration method and system for an overall electrical solution of a computer room. Background Art

[0002] With the rapid development of information technology, data centers, as important infrastructure in the information age, are growing in size and energy consumption. How to reduce the energy consumption cost of data centers and improve the efficiency and reliability of the electrical system in computer rooms has become an urgent problem to be solved in the industry.

[0003] Traditional data center electrical design mainly uses empirical formulas and simulation analysis methods, which have problems such as inaccurate load forecasting, over-capacity configuration, and low energy efficiency. At the same time, traditional computer room operation and maintenance mainly rely on manual inspections and regular maintenance, which have defects such as large investment, low efficiency, and high risk.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for configuring an overall electrical solution for a computer room, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a method for configuring an overall electrical solution of a computer room, comprising:

[0007] Based on the distributed data architecture, an intelligent sensor gateway is deployed in the computer room to collect multi-source heterogeneous data corresponding to the current computer room in real time. Feature extraction and semantic mapping are performed on the multi-source heterogeneous data through a pre-set deep learning model to build a high-dimensional semantic association network. Semantic enhancement is performed based on the high-dimensional semantic association network and cross-modal representation learning technology. The semantically enhanced multi-source heterogeneous data is semantically modeled and logically associated through knowledge graph technology to obtain a data tracing knowledge graph and a data flow trajectory. Based on the data flow trajectory and the stream data intelligent processing framework, abnormal patterns in the current data are extracted to obtain a portrait of the computer room operation status.

[0008] Based on the data tracing knowledge graph and the operating status portrait of the computer room, a digital portrait model of the energy consumption of the computer room is constructed, the energy consumption pattern and high-energy consumption equipment of the computer room are determined, and the composite characteristic analysis and abnormal diagnosis of the high-energy consumption equipment are performed to obtain the power load curve of the computer room. The operating status portrait of the computer room and the power load curve of the computer room are used as inputs, and hierarchical progressive prediction is performed in combination with the exogenous variables corresponding to the current computer room to obtain an optimized load prediction result. The optimized load prediction result and the data tracing knowledge graph are used as inputs, and an objective function is constructed with the goal of minimizing cost and maximizing energy utilization, and the optimal configuration parameters are obtained and a distribution system design scheme is generated;

[0009] Based on the power distribution system design, a self-learning controller is added to the power distribution unit, and the power load of the computer room is optimized in combination with the reinforcement learning algorithm. Based on the multi-source heterogeneous data, a degradation mechanism model of the distribution equipment is constructed through a machine learning algorithm, and the life of the distribution equipment is predicted. The life management strategy of the distribution equipment is output, and according to the collected computer room operation data and the energy consumption law of the computer room, the data analysis algorithm is used to identify abnormal operating equipment, extract abnormal features and refer to historical abnormal data to determine the cause of the abnormality, generate an abnormal report and send it to the equipment maintenance personnel.

[0010] In an optional embodiment,

[0011] Based on the distributed data architecture, an intelligent sensor gateway is deployed in the computer room to collect multi-source heterogeneous data corresponding to the current computer room in real time. Feature extraction and semantic mapping of the multi-source heterogeneous data are performed through a pre-set deep learning model to build a high-dimensional semantic association network. Semantic enhancement based on the high-dimensional semantic association network and cross-modal representation learning technology includes:

[0012] According to the characteristics of the computer room environment and task requirements, an intelligent sensor gateway is deployed in the electrical computer room in combination with a distributed data architecture. The multi-source heterogeneous data corresponding to the current computer room is collected in real time according to the intelligent sensor gateway, wherein the multi-source heterogeneous data includes power consumption data of the computer room, equipment status data and environmental monitoring data. The high-frequency noise and abnormal points in the multi-source heterogeneous data are removed based on the Kalman filter algorithm, and the missing points in the multi-source heterogeneous data are completed through matrix completion to obtain pre-processed multi-source heterogeneous data.

[0013] Based on a preset deep learning model, feature extraction is performed on the preprocessed multi-source heterogeneous data. For the multi-source heterogeneous data, the corresponding time-frequency domain features are extracted by a moving average algorithm. For the environmental image corresponding to the environmental monitoring data, the corresponding multi-scale texture features are extracted by a convolutional neural network. Combined with the attention mechanism, the time-frequency domain features and the multi-scale texture features are mapped to a common semantic space. In the common semantic space, for each different feature, the semantic association strength of each feature pair is determined based on cosine similarity, and a high-dimensional semantic association network corresponding to the multi-source heterogeneous data is constructed in combination with an improved graph attention network.

[0014] Based on the high-dimensional semantic association network, a cross-modal semantic embedding framework is set up, and cross-modal positive and negative sample pairs are constructed in combination with contrastive learning. A consistent semantic representation space is obtained by maximizing the similarity of positive samples and the difference of negative samples. An ontology knowledge base is constructed according to the computer room field, and the multi-source heterogeneous data is semantically annotated in combination with the concepts and relationships in the ontology knowledge base, and semantic enhancement is performed in combination with ontology knowledge constraints.

[0015] In an optional embodiment,

[0016] The semantically enhanced multi-source heterogeneous data is semantically modeled and logically associated through knowledge graph technology to obtain data traceability knowledge graph and data flow trajectory. Based on the data flow trajectory and stream data intelligent processing framework, abnormal patterns in the current data are extracted to obtain the operating status portrait of the computer room, including:

[0017] Based on the enhanced multi-source semantic data, knowledge representation is performed based on the pre-built ontology knowledge base, data entities, attributes and relationships are defined, feature triples are generated and a composite knowledge association framework based on rules and graph embedding is designed. Logical reasoning is performed on the knowledge-represented data according to the pre-acquired domain rule base to generate hidden associations between different ontologies. Based on the hidden associations, a data traceability knowledge graph is constructed by combining the temporal constraints represented by temporal logic. Through the knowledge graph reasoning technology, the paths between different nodes are determined to obtain the data flow trajectory.

[0018] Based on the data flow trajectory and the data traceability knowledge graph, the dependencies between different electrical equipment and the fault propagation mechanism between electrical equipment are encoded as constraint rules, and a directed acyclic graph is constructed through a stream data intelligent processing framework, wherein nodes represent multi-source heterogeneous data, and edges represent dependencies between data. The constraint rules are used as prior knowledge, and anomaly point detection is performed through the isolation forest algorithm. If an anomaly point exists, all root cause paths are determined from the current anomaly point through the shortest path algorithm of the graph. For each root cause path, automatic scoring is performed in combination with domain knowledge, and the root cause path with the highest score is selected as the anomaly mode corresponding to the current anomaly point to generate a portrait of the computer room operation status.

[0019] In an optional embodiment,

[0020] Based on the data tracing knowledge graph and the operating status portrait of the computer room, a digital portrait model of the energy consumption of the computer room is constructed, the energy consumption law and high-energy consumption equipment of the computer room are determined, and the composite characteristic analysis and abnormal diagnosis of the high-energy consumption equipment are performed to obtain the power load curve of the computer room. The operating status portrait of the computer room and the power load curve of the computer room are used as inputs, and hierarchical progressive prediction is performed in combination with the exogenous variables corresponding to the current computer room to obtain an optimized load prediction result. The optimized load prediction result and the data tracing knowledge graph are used as inputs, and an objective function is constructed with the goal of minimizing cost and maximizing energy utilization. The optimal configuration parameters are obtained and a distribution system design scheme is generated, including:

[0021] Based on the data traceability knowledge graph and the operating status portrait of the computer room, the time series data corresponding to the energy consumption information is extracted, the support of each data point in the time series data is determined according to the association rule mining algorithm, the frequent item set is determined based on the support and the preset minimum support threshold, the frequent contribution mode and sequence mode of the energy consumption of electrical equipment are obtained based on the frequent item set, the energy consumption of the computer room is obtained, the energy consumption data corresponding to each device is grouped and aggregated according to the device type and device location, the outliers in the energy consumption data are detected in combination with the box plot, the high-energy consumption equipment is determined, the correlation coefficient matrix between the energy consumption of the equipment and its own attributes and operating conditions is calculated, the correlation strength is determined and a regression model of influencing factors is constructed, the degree of influence of the interaction effect between each influencing factor on the energy consumption is determined in combination with the interaction term, and the power load curve of the computer room is obtained;

[0022] Collect load power data from the operating status portrait of the computer room, construct a load time series data set, perform detrending and standardization operations on each load time series data, fit the load time series with a long short-term memory network, obtain time series dynamic laws, collect weather and holiday information at the location of the computer room as exogenous variables, standardize the exogenous variables and align them with the load time series data, align the aligned exogenous variables with the load time series data to obtain a comprehensive feature set, perform long-term trend modeling and short-term fluctuation modeling based on the comprehensive feature set, generate long-term trend prediction results and short-term fluctuation prediction results and fuse them, and combine grid search strategy and attention mechanism to obtain the optimized load prediction result;

[0023] Based on the optimized load forecast results and the data traceability knowledge graph, a cost objective function and an efficiency objective function are constructed, and a comprehensive objective function corresponding to the distribution system is obtained through multi-objective weighted construction. For the comprehensive objective function, decision variables are determined, and constraints are set according to the power grid topology and equipment parameter variables. A mixed integer programming model is constructed, and the comprehensive objective function is solved by the branch and bound method to obtain the optimal configuration parameters. The distribution system design scheme is generated according to the optimal configuration parameters.

[0024] In an optional embodiment,

[0025] The comprehensive objective function corresponding to the distribution system is obtained through multi-objective weighted construction as shown in the following formula:

[0026] ;

[0027] Among them, MF represents the objective function value, represents the fixed cost associated with the ith device, C i0 represents the baseline cost of the ith device, represents the variable cost associated with the ith device, ρ represents the weight coefficient of the network loss function, and L j (x) represents the jth network loss function, which is related to power flow and power loss, μ represents the weight coefficient of renewable energy utilization, and H(x) represents the utilization rate of renewable energy.

[0028] In an optional embodiment,

[0029] Based on the power distribution system design, a self-learning controller is added to the power distribution unit, and the power load of the computer room is optimized in combination with the reinforcement learning algorithm. Based on the multi-source heterogeneous data, a degradation mechanism model of the power distribution equipment is constructed through a machine learning algorithm, and the life of the power distribution equipment is predicted. The life management strategy of the power distribution equipment is output. According to the collected computer room operation data and the energy consumption law of the computer room, the data analysis algorithm is used to identify the abnormal operation equipment, extract the abnormal features and refer to the historical abnormal data to determine the abnormal cause, generate an abnormal report and send it to the equipment maintenance personnel, including:

[0030] Based on the power distribution system design scheme, a self-learning controller is added to each power distribution unit, wherein the self-learning controller includes a perception layer, a decision layer and an execution layer, wherein the perception layer is used to obtain real-time operation data of the power distribution system, including voltage, current, power and environmental information. Based on the real-time operation data, the decision layer maps the state of the electrical equipment into continuous control actions through a deep deterministic policy gradient algorithm, sets the corresponding state space in combination with a reinforcement learning algorithm, constructs a discrete action space based on the adjustment action of each controllable electrical equipment, and constructs a reward objective function with the goal of minimizing system energy consumption, maximizing voltage qualification rate and minimizing equipment loss, solves the reward objective function and optimizes the power load of the computer room according to the function value;

[0031] Based on the optimized power load of the computer room and the multi-source heterogeneous data, the equipment status data in the multi-source heterogeneous data is encoded into a time series through a machine learning algorithm, the equipment degradation time series is generated and degradation features are extracted, based on the degradation features, the Shapley value corresponding to each degradation feature is calculated in combination with a long short-term memory network, and the degradation features are ranked in importance according to the Shapley value, the degradation features in the top 10% are taken as key degradation features, a degradation mechanism model is generated, and the parameters of the degradation mechanism model are initialized according to the key degradation features;

[0032] Based on the degradation mechanism model, combined with the pre-acquired historical fault data and expert experience, the equipment health is defined and the warning threshold and failure threshold of the equipment health are set. The equipment remaining life is modeled according to the random process theory, and the equipment degradation process is modeled as a Markov chain of continuous-time discrete states. The state transition probability is determined by the degradation mechanism model, and the absorption time distribution of the Markov chain is solved to obtain the equipment remaining life. Based on the equipment remaining life and the key degradation characteristics, the equipment maintenance decision is determined in combination with the historical fault data and the distribution equipment life management strategy is generated;

[0033] Based on the collected computer room operation data and the energy consumption pattern of the computer room, abnormal data is identified through a data analysis algorithm, and the operation abnormality features in the abnormal data are extracted. The current abnormal data is attribution analyzed in combination with a few-sample learning algorithm, the cause of the abnormality is determined, and an abnormal report is generated and sent to the equipment maintenance personnel.

[0034] In an optional embodiment,

[0035] Solving the absorption time distribution of the Markov chain, the remaining life of the equipment is obtained as shown in the following formula:

[0036] ;

[0037] Among them, R(t) represents the expected value of the remaining life of the current equipment, π s represents the initial state distribution, P represents the state transition probability matrix, e is a vector of all 1s, represents the cumulative probability calculation, ζ represents the scale parameter of the Weibull distribution, and β represents the shape parameter of the Weibull distribution.

[0038] A second aspect of an embodiment of the present invention provides a system for configuring an overall electrical solution for a computer room, including:

[0039] The first unit is used to deploy an intelligent sensor gateway in the computer room based on a distributed data architecture, collect multi-source heterogeneous data corresponding to the current computer room in real time, perform feature extraction and semantic mapping on the multi-source heterogeneous data through a preset deep learning model, build a high-dimensional semantic association network, perform semantic enhancement based on the high-dimensional semantic association network and cross-modal representation learning technology, perform semantic modeling and logical association on the semantically enhanced multi-source heterogeneous data through knowledge graph technology, obtain data tracing knowledge graph and data flow trajectory, extract abnormal patterns in current data based on the data flow trajectory and stream data intelligent processing framework, and obtain a portrait of the computer room operation status;

[0040] The second unit is used to construct a digital portrait model of energy consumption in the computer room based on the data tracing knowledge graph and the operating status portrait of the computer room, determine the energy consumption law and high-energy consumption equipment in the computer room, perform composite characteristic analysis and abnormal diagnosis on the high-energy consumption equipment, obtain the power load curve of the computer room, take the operating status portrait of the computer room and the power load curve of the computer room as input, perform hierarchical progressive prediction in combination with the exogenous variables corresponding to the current computer room, obtain the optimized load prediction result, take the optimized load prediction result and the data tracing knowledge graph as input, construct an objective function with the goal of minimizing cost and maximizing energy utilization, solve to obtain the optimal configuration parameters and generate a distribution system design plan;

[0041] The third unit is used to add a self-learning controller in the distribution unit based on the distribution system design scheme, optimize the power load of the computer room in combination with the reinforcement learning algorithm, build a distribution equipment degradation mechanism model based on the multi-source heterogeneous data through a machine learning algorithm and predict the life of the distribution equipment, output the distribution equipment life management strategy, identify abnormal operating equipment through a data analysis algorithm according to the collected computer room operation data and the energy consumption law of the computer room, extract abnormal features and refer to historical abnormal data to determine the cause of the abnormality, generate an abnormal report and send it to the equipment maintenance personnel.

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

[0043] An electronic device is provided, comprising:

[0044] processor;

[0045] a memory for storing processor-executable instructions;

[0046] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0047] A fourth aspect of the embodiments of the present invention is:

[0048] 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.

[0049] In the present invention, load characteristics, abnormal diagnosis, exogenous factors, etc. are comprehensively considered to realize hierarchical progressive prediction and multi-objective optimization of the power load in the computer room, thereby improving the design accuracy and energy-saving and emission-reduction effect of the power distribution system. The intelligent control method can dynamically adjust the control strategy according to load fluctuations, equipment status, environmental changes, etc., thereby improving the flexibility, robustness and efficiency of the power distribution system. Predictive maintenance can arrange maintenance in time before the performance of the equipment decreases significantly, thereby avoiding the cost waste caused by excessive maintenance and preventing the expansion of losses caused by sudden failures, effectively extending the service life of the equipment and improving the reliability of the power distribution system. The data analysis algorithm is used to detect abnormalities in the computer room operation data, automatically identify abnormal fluctuations and over-limit conditions of key performance indicators, and extract abnormal features. The expert knowledge base determines the possible causes of abnormalities, generates abnormality reports and disposal suggestions, and realizes real-time monitoring of the operating status of the distribution system and fault warning. Intelligent abnormality diagnosis can effectively solve problems such as monitoring blind spots, discovery lags, and subjective judgments, reducing the expansion of accidents and impact time, and ensuring power supply safety and power quality. In summary, the present invention realizes the mapping integration of physics, information, and knowledge through digital twins, achieves balanced improvement of system efficiency, economy, and reliability through intelligent optimization, and realizes active perception and rapid response of operating risks through abnormal diagnosis, forming a complete, full-life cycle intelligent distribution health management system, which provides key technical support for smart grids and smart energy, and is of great value in promoting the digital and intelligent transformation of the energy and power sectors. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 A schematic diagram of a flow chart of a method for configuring an overall electrical solution for a computer room according to an embodiment of the present invention;

[0051] Figure 2 It is a structural schematic diagram of the configuration system of the overall electrical solution of the computer room according to an embodiment of the present invention. DETAILED DESCRIPTION

[0052] 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.

[0053] 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.

[0054] Figure 1FIG. 1 is a flow chart of a method for configuring an overall electrical solution for a computer room according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0055] S1. Deploy intelligent sensor gateways in the computer room based on the distributed data architecture, collect multi-source heterogeneous data corresponding to the current computer room in real time, perform feature extraction and semantic mapping on the multi-source heterogeneous data through a pre-set deep learning model, build a high-dimensional semantic association network, perform semantic enhancement based on the high-dimensional semantic association network and cross-modal representation learning technology, perform semantic modeling and logical association on the semantically enhanced multi-source heterogeneous data through knowledge graph technology, obtain data tracing knowledge graph and data flow trajectory, extract abnormal patterns in the current data based on the data flow trajectory and stream data intelligent processing framework, and obtain a portrait of the computer room operation status;

[0056] The intelligent sensor gateway refers to a device that integrates multiple sensors, which is used to collect different types of data and transmit them to the background processing system through the network, supporting real-time monitoring and management of multiple sensor data. The high-dimensional semantic association network is a complex network structure used to represent and process semantic associations between data. It strengthens the semantic understanding between different data sources by capturing the connections between multi-dimensional features. The logical association refers to the causal relationship between data or events derived through logical rules or relationships, which is used to reveal potential system behaviors or event development patterns. The data flow trajectory refers to the path of data transmission and processing between different nodes, which helps track the entire life cycle of data and ensure the transparency and controllability of data flow. The data traceability knowledge graph is used to build a knowledge system of data sources, processing processes and flows, which is convenient for tracing and verification in the data processing chain. The computer room operation status portrait is an analysis and visualization of the global or local status of the operation of the data center computer room, and usually generates a picture reflecting the health status of the computer room based on real-time data.

[0057] In an optional embodiment,

[0058] Based on the distributed data architecture, an intelligent sensor gateway is deployed in the computer room to collect multi-source heterogeneous data corresponding to the current computer room in real time. Feature extraction and semantic mapping of the multi-source heterogeneous data are performed through a pre-set deep learning model to build a high-dimensional semantic association network. Semantic enhancement based on the high-dimensional semantic association network and cross-modal representation learning technology includes:

[0059] According to the characteristics of the computer room environment and task requirements, an intelligent sensor gateway is deployed in the electrical computer room in combination with a distributed data architecture. The multi-source heterogeneous data corresponding to the current computer room is collected in real time according to the intelligent sensor gateway, wherein the multi-source heterogeneous data includes power consumption data of the computer room, equipment status data and environmental monitoring data. The high-frequency noise and abnormal points in the multi-source heterogeneous data are removed based on the Kalman filter algorithm, and the missing points in the multi-source heterogeneous data are completed through matrix completion to obtain pre-processed multi-source heterogeneous data.

[0060] Based on a preset deep learning model, feature extraction is performed on the preprocessed multi-source heterogeneous data. For the multi-source heterogeneous data, the corresponding time-frequency domain features are extracted by a moving average algorithm. For the environmental image corresponding to the environmental monitoring data, the corresponding multi-scale texture features are extracted by a convolutional neural network. Combined with the attention mechanism, the time-frequency domain features and the multi-scale texture features are mapped to a common semantic space. In the common semantic space, for each different feature, the semantic association strength of each feature pair is determined based on cosine similarity, and a high-dimensional semantic association network corresponding to the multi-source heterogeneous data is constructed in combination with an improved graph attention network.

[0061] Based on the high-dimensional semantic association network, a cross-modal semantic embedding framework is set up, and cross-modal positive and negative sample pairs are constructed in combination with contrastive learning. A consistent semantic representation space is obtained by maximizing the similarity of positive samples and the difference of negative samples. An ontology knowledge base is constructed according to the computer room field, and the multi-source heterogeneous data is semantically annotated in combination with the concepts and relationships in the ontology knowledge base, and semantic enhancement is performed in combination with ontology knowledge constraints.

[0062] The high-frequency noise refers to high-frequency interference that occurs in signal processing, which usually affects the quality of the signal and requires noise reduction through filtering technology. The moving average algorithm is an algorithm for smoothing time series data, which eliminates short-term fluctuations by calculating the average value of adjacent data points. The graph attention network is a neural network based on a graph structure, which uses the attention mechanism to highlight the features of important nodes or edges in the graph, thereby improving the model's ability to understand complex graph structures. The cross-modal semantic embedding framework is used to map data from different modalities (such as text, images, etc.) to a unified semantic space for the fusion and processing of multimodal information. The positive sample similarity refers to the measure of the similarity between samples of the same type during the training process of the model, and the negative sample difference refers to the difference between samples of different types.

[0063] According to the characteristics of the computer room environment and task requirements, an intelligent sensor gateway is deployed in the electrical computer room. The intelligent sensor gateway includes temperature and humidity sensors, power parameter sensors, equipment status sensors, etc., which can collect multi-source heterogeneous data of the computer room in real time. Specifically, the temperature and humidity sensors collect computer room temperature and humidity data at intervals of 5 minutes, the power parameter sensors collect power consumption data such as the total power of the computer room and the power of each device in real time, and the equipment status sensors collect the operating status data of each device in real time. In addition, cameras are installed in the computer room to take an image of the computer room environment every 10 minutes.

[0064] The collected multi-source heterogeneous data are preprocessed, and the Kalman filter algorithm is used to remove high-frequency noise and abnormal points in the data. Taking temperature data as an example, assuming that the temperature measured at a certain moment is 35°C, and the temperature before and after are around 25°C, 35°C can be determined as an abnormal point and removed, and the missing points in the data are completed. For example, if the temperature data at a certain moment is missing, the temperature value at that moment can be estimated based on the temperature data before and after through the matrix completion algorithm to obtain the preprocessed multi-source heterogeneous data.

[0065] Feature extraction is performed on the preprocessed multi-source heterogeneous data. For time series data such as temperature, humidity, and electricity consumption, a moving average algorithm with a 10-minute window is used to extract time-frequency domain features. For example, the statistical features such as the average, variance, maximum, and minimum values ​​of the temperature within 10 minutes, as well as the frequency domain features after Fourier transform are calculated. For environmental images, a pre-trained ResNet50 convolutional neural network is used to extract multi-scale texture features to obtain a 2048-dimensional feature vector.

[0066] The extracted features of different modalities are mapped to the common semantic space. The fully connected layer is used to reduce the dimension of each feature and map it to a 256-dimensional common semantic space. The attention mechanism is introduced to calculate the importance weights of different features in the semantic space. Specifically, for each feature vector, the softmax function is used to calculate its similarity with all other feature vectors to obtain the attention weight. The original feature vector is multiplied by the attention weight to obtain the enhanced feature representation.

[0067] A high-dimensional semantic association network is constructed in the public semantic space. The cosine similarity between all feature vectors is calculated to obtain a similarity matrix. The threshold is set to 0.7. When the similarity between two feature vectors is greater than 0.7, a connection is established between the two features. The initial network constructed is optimized using an improved graph attention network. Specifically, for each node, the information of its neighboring nodes is aggregated and the feature representation of the node is updated. By stacking multiple layers of graph attention networks, an optimized high-dimensional semantic association network is finally obtained.

[0068] Based on the high-dimensional semantic association network, cross-modal semantic enhancement is performed, and a cross-modal semantic embedding framework is set up to map the features of different modalities to a unified semantic space, and construct cross-modal positive and negative sample pairs. Taking temperature and humidity features as an example, if the temperature feature and the humidity feature are directly connected in the semantic association network, they constitute a positive sample pair, otherwise they constitute a negative sample pair. Through contrastive learning, the similarity of positive sample pairs is maximized and the similarity of negative sample pairs is minimized, so as to obtain a consistent semantic representation space.

[0069] Combine domain knowledge for semantic enhancement and build a computer room domain ontology knowledge base, which includes concepts such as equipment, parameters, status and the relationships between them. Based on the knowledge base, semantic annotation is performed on multi-source heterogeneous data. For example, temperature data is annotated as "environmental parameter-temperature" and equipment power data is annotated as "equipment parameter-power". The conceptual hierarchical relationship and attribute relationship in the ontology knowledge are used to further semantically enhance the data features. For example, since "temperature" is a sub-concept of "environmental parameter", the temperature feature can be associated with other environmental parameter features.

[0070] In this embodiment, by deploying an intelligent sensor gateway in the computer room, multi-source heterogeneous data such as power consumption, equipment status, and environmental monitoring of the computer room can be collected in real time, providing a data basis for subsequent applications such as energy efficiency optimization of the computer room and equipment health management. By preprocessing and extracting features from multi-source heterogeneous data, the fusion representation of cross-modal data is realized. Kalman filtering is used to remove data noise, matrix completion is used to repair data missing, moving average is used to extract time-frequency features, and convolutional neural networks are used to extract multi-scale texture features. These technologies effectively improve the data quality and feature representation capabilities, provide high-quality input for subsequent machine learning algorithms, and map different modal features to a common semantic space through the attention mechanism. The intrinsic semantic associations between different data modalities are mined through the graph attention network, which lays the foundation for a deep understanding of the internal mechanism of the computer room system. By constructing a computer room domain ontology knowledge base and using it to semantically annotate and constrain multi-source heterogeneous data, external domain knowledge is introduced to enhance the data semantics and improve the accuracy and consistency of the data semantic representation. In summary, this embodiment realizes the quality improvement, feature enhancement, semantic association and cross-modal fusion of multi-source heterogeneous data in the computer room environment, provides high-quality and semantically rich data support for subsequent intelligent applications in the computer room, and improves the performance and generalization ability of data-driven intelligent applications.

[0071] In an optional embodiment,

[0072] The semantically enhanced multi-source heterogeneous data is semantically modeled and logically associated through knowledge graph technology to obtain data traceability knowledge graph and data flow trajectory. Based on the data flow trajectory and stream data intelligent processing framework, abnormal patterns in the current data are extracted to obtain the operating status portrait of the computer room, including:

[0073] Based on the enhanced multi-source semantic data, knowledge representation is performed based on the pre-built ontology knowledge base, data entities, attributes and relationships are defined, feature triples are generated and a composite knowledge association framework based on rules and graph embedding is designed. Logical reasoning is performed on the knowledge-represented data according to the pre-acquired domain rule base to generate hidden associations between different ontologies. Based on the hidden associations, a data traceability knowledge graph is constructed by combining the temporal constraints represented by temporal logic. Through the knowledge graph reasoning technology, the paths between different nodes are determined to obtain the data flow trajectory.

[0074] Based on the data flow trajectory and the data traceability knowledge graph, the dependencies between different electrical equipment and the fault propagation mechanism between electrical equipment are encoded as constraint rules, and a directed acyclic graph is constructed through a stream data intelligent processing framework, wherein nodes represent multi-source heterogeneous data, and edges represent dependencies between data. The constraint rules are used as prior knowledge, and anomaly point detection is performed through the isolation forest algorithm. If an anomaly point exists, all root cause paths are determined from the current anomaly point through the shortest path algorithm of the graph. For each root cause path, automatic scoring is performed in combination with domain knowledge, and the root cause path with the highest score is selected as the anomaly mode corresponding to the current anomaly point to generate a portrait of the computer room operation status.

[0075] The feature triple is a triple form used to express the relationship between data entities, usually composed of entities, attributes and attribute values, and is used to enrich the semantic expression of data. The composite knowledge association framework based on rules and graph embedding is to achieve efficient association between different knowledge points and enhance knowledge reasoning and understanding capabilities by combining rules and graph embedding technology in the field. The domain rule base refers to a knowledge base containing rules in a specific field, which helps to solve problems or reason in the field. The timing constraints expressed in combination with temporal logic are used to describe and restrict the behavior of the system or process in the time dimension to ensure the temporal consistency of the system. The fault propagation mechanism describes the process of how a fault spreads from one component to other components in a complex system, which helps to understand and predict possible faults in the system. The stream data intelligent processing framework is a system architecture for real-time processing of large-scale, continuously arriving data streams, which usually integrates functions such as data collection, filtering, analysis and decision-making. The shortest path algorithm of the graph is used to find the shortest path between two points in the graph structure, and is often used to optimize resource allocation or path selection in the network. The root cause path refers to the cause and influencing link of an event or fault in the system. By analyzing the path, the root cause of the fault can be traced back.

[0076] Based on the enhanced multi-source semantic data, the pre-built ontology knowledge base is used for knowledge representation. The ontology knowledge base contains domain concepts and their relationships in the fields of computer room equipment, operating parameters, fault types, etc. For example, "air conditioning" is defined as an equipment entity, "temperature" is defined as a parameter attribute, and "refrigeration" is defined as a function relationship. For the input semantic data, the entity, attribute and relationship information are extracted to generate feature triples. For example, the temperature of air conditioning 1 is 25℃.

[0077] A composite knowledge association framework based on rules and graph embedding is designed. The rule part uses predefined domain rules to reason about triples, such as "if the air conditioner temperature is abnormal and the humidity is normal, there may be a refrigeration failure." The graph embedding part maps entities and relationships to a low-dimensional vector space and discovers potential associations through vector operations. Combining the two methods can obtain more comprehensive knowledge association results.

[0078] Logical reasoning is performed on the data represented by knowledge according to the pre-acquired domain rule base to generate hidden associations between different ontologies. For example, through reasoning, the hidden relationship "UPS1 supplies power to air conditioner 1" can be obtained. Based on these associations and combined with the timing constraints represented by temporal logic, a data traceability knowledge graph is constructed. The timing constraints can describe the order in which events occur, such as "UPS failure precedes air conditioner failure".

[0079] By utilizing the constructed knowledge graph and using knowledge graph reasoning technology to determine the paths between different nodes, we can obtain the data flow trajectory. For example, we can track the data flow process from UPS to air conditioner and then to temperature and humidity sensor. These trajectories reflect the dependencies between devices and the data propagation paths.

[0080] Based on the obtained data flow trajectory and data traceability knowledge graph, the dependencies between different power-consuming devices and the fault propagation mechanism between power-consuming devices are encoded as constraint rules. For example, "if the UPS fails, it may cause the air conditioner to fail" can be encoded as a constraint rule, and a directed acyclic graph is constructed through the stream data intelligent processing framework, where nodes represent multi-source heterogeneous data, such as UPS voltage, air conditioner temperature, etc., and edges represent the dependencies between data.

[0081] The encoded constraint rules are used as prior knowledge, and the isolation forest algorithm is used to detect anomalies in the input real-time data stream. The isolation forest algorithm isolates the anomalies by randomly constructing decision trees. If an anomaly is detected, such as abnormal air-conditioning temperature, all possible root cause paths are determined starting from the anomaly through the shortest path algorithm of the graph.

[0082] For each root cause path, domain knowledge is combined to automatically score. The scoring criteria may include factors such as path length, equipment importance, and failure frequency. For example, the path of UPS→air conditioner→temperature sensor may get a higher score. The root cause path with the highest score is selected as the abnormal mode corresponding to the current abnormal point.

[0083] Based on the selected abnormal mode and combined with the real-time status data of related equipment, a profile of the computer room operation status is generated. The profile contains information such as abnormal equipment, root cause of the fault, and scope of impact. For example, "UPS1 failure causes abnormal temperature of air conditioner 1, affecting the operation of the server in cabinet 1"

[0084] In this embodiment, by designing a composite knowledge association framework based on rules and graph embedding, the association fusion between different data and knowledge is realized, the semantic consistency and richness of the data are improved, the hidden associations are helpful to discover the causal mechanism and evolution law behind the data, and provide a new perspective for in-depth understanding of system behavior. The abnormal points are automatically detected through the isolation forest algorithm, and the automatic identification and root cause tracing of abnormal patterns are realized through the shortest path algorithm of the graph and domain knowledge scoring, and a portrait of the operating status of the computer room is generated. The knowledge graph provides a natural carrier for the visualization of the operation and maintenance knowledge of the computer room, and enhances human-computer interaction and cognitive understanding. In summary, this embodiment enhances the semantics and value of the data, improves the timeliness and accuracy of fault discovery and location, provides an intuitive decision-making reference for operation and maintenance personnel, assists in the intelligent operation and maintenance of the computer room, and provides new ideas and methods for knowledge modeling and abnormal analysis of complex industrial systems, which has certain theoretical contributions and application value.

[0085] S2. Construct a digital energy consumption portrait model of the computer room based on the data tracing knowledge graph and the operating status portrait of the computer room, determine the energy consumption pattern and high-energy consumption equipment in the computer room, perform composite characteristic analysis and abnormal diagnosis on the high-energy consumption equipment, and obtain the power load curve of the computer room. Take the operating status portrait of the computer room and the power load curve of the computer room as input, and perform hierarchical progressive prediction in combination with the exogenous variables corresponding to the current computer room to obtain the optimized load prediction result. Take the optimized load prediction result and the data tracing knowledge graph as input, construct an objective function with the goal of minimizing cost and maximizing energy utilization, solve to obtain the optimal configuration parameters and generate a distribution system design plan;

[0086] The digital energy consumption portrait model of the computer room is a tool for digitally modeling the energy consumption behavior of the computer room of the data center. Based on historical energy consumption data and real-time monitoring data, it generates an intuitive energy consumption portrait through multi-dimensional feature extraction and analysis to help optimize energy management. The composite characteristic analysis refers to a comprehensive analysis of the characteristics of data or systems from multiple angles and dimensions to reveal the interactive relationships and characteristic performances within complex systems. The power load curve of the computer room is a curve chart that reflects the power usage of the computer room in different time periods, which helps analyze the power demand and load change patterns of the computer room, and then optimizes power distribution. The exogenous variable refers to an external factor that affects the behavior or results of the system but is not directly controlled by the internal mechanism of the system, such as ambient temperature or market price. The hierarchical progressive prediction is a method of decomposing data or problems hierarchically and making predictions step by step to improve the accuracy and stability of the prediction.

[0087] In an optional embodiment,

[0088] Based on the data tracing knowledge graph and the operating status portrait of the computer room, a digital portrait model of the energy consumption of the computer room is constructed, the energy consumption law and high-energy consumption equipment of the computer room are determined, and the composite characteristic analysis and abnormal diagnosis of the high-energy consumption equipment are performed to obtain the power load curve of the computer room. The operating status portrait of the computer room and the power load curve of the computer room are used as inputs, and hierarchical progressive prediction is performed in combination with the exogenous variables corresponding to the current computer room to obtain an optimized load prediction result. The optimized load prediction result and the data tracing knowledge graph are used as inputs, and an objective function is constructed with the goal of minimizing cost and maximizing energy utilization. The optimal configuration parameters are obtained and a distribution system design scheme is generated, including:

[0089] Based on the data traceability knowledge graph and the operating status portrait of the computer room, the time series data corresponding to the energy consumption information is extracted, the support of each data point in the time series data is determined according to the association rule mining algorithm, the frequent item set is determined based on the support and the preset minimum support threshold, the frequent contribution mode and sequence mode of the energy consumption of electrical equipment are obtained based on the frequent item set, the energy consumption of the computer room is obtained, the energy consumption data corresponding to each device is grouped and aggregated according to the device type and device location, the outliers in the energy consumption data are detected in combination with the box plot, the high-energy consumption equipment is determined, the correlation coefficient matrix between the energy consumption of the equipment and its own attributes and operating conditions is calculated, the correlation strength is determined and a regression model of influencing factors is constructed, the degree of influence of the interaction effect between each influencing factor on the energy consumption is determined in combination with the interaction term, and the power load curve of the computer room is obtained;

[0090] Collect load power data from the operating status portrait of the computer room, construct a load time series data set, perform detrending and standardization operations on each load time series data, fit the load time series with a long short-term memory network, obtain time series dynamic laws, collect weather and holiday information at the location of the computer room as exogenous variables, standardize the exogenous variables and align them with the load time series data, align the aligned exogenous variables with the load time series data to obtain a comprehensive feature set, perform long-term trend modeling and short-term fluctuation modeling based on the comprehensive feature set, generate long-term trend prediction results and short-term fluctuation prediction results and fuse them, and combine grid search strategy and attention mechanism to obtain the optimized load prediction result;

[0091] Based on the optimized load forecast results and the data traceability knowledge graph, a cost objective function and an efficiency objective function are constructed, and a comprehensive objective function corresponding to the distribution system is obtained through multi-objective weighted construction. For the comprehensive objective function, decision variables are determined, and constraints are set according to the power grid topology and equipment parameter variables. A mixed integer programming model is constructed, and the comprehensive objective function is solved by the branch and bound method to obtain the optimal configuration parameters. The distribution system design scheme is generated according to the optimal configuration parameters.

[0092] The association rule mining algorithm is an algorithm used to discover frequent item sets from a data set and derive the association relationship between data items. It is often used in scenarios such as shopping basket analysis, and helps decision optimization by discovering implicit association rules. The box plot is a statistical chart used to describe the distribution of a data set. By displaying the minimum value, lower quartile, median, upper quartile and maximum value of the data, it helps identify the central tendency and outliers of the data. The correlation coefficient matrix is ​​a matrix used to show the degree of correlation between different variables. Each element in the matrix represents the correlation between two variables, which is usually quantified by the Pearson correlation coefficient. The interaction effect refers to the mutual influence between two or more factors in multifactor analysis. The combined effect may be different from the effect of each factor acting alone. The mixed integer programming model is a type of mathematical model used to solve optimization problems involving continuous variables and integer variables. It is widely used in fields such as resource allocation, scheduling and planning. The branch and bound method is an algorithm used to solve integer programming problems. By recursively dividing the solution space of the problem, the search range is gradually narrowed to finally find the optimal solution.

[0093] Based on the data traceability knowledge graph and the computer room operation status portrait, a digital portrait model of the computer room energy consumption is constructed, and the time series data corresponding to the energy consumption information is extracted. Taking a data center computer room as an example, the power consumption, operation status and other time series data of each device are extracted from the knowledge graph, such as the hourly power consumption data of Air Conditioner 1 from January 1 to January 31, 2023. The Apriori algorithm is used for association rule mining, and the minimum support threshold is set to 0.3 to obtain frequent item sets, such as {Air Conditioner 1, UPS 1}, {Air Conditioner 2, lighting system}, etc. Based on the frequent item set analysis, the frequent contribution pattern of energy consumption of electrical equipment is obtained, such as "the energy consumption is higher when Air Conditioner 1 and UPS 1 are running at the same time". The sequential pattern is obtained through the sequence pattern mining algorithm, such as "turn on the air conditioner first, and turn on the server 15 minutes later".

[0094] Energy consumption data is grouped and aggregated according to equipment type and location. Equipment is grouped according to types such as air conditioners, UPS, lighting, and IT equipment, as well as different areas in the computer room. The average energy consumption, maximum energy consumption, and other statistical values ​​of each group are calculated. The box plot method is used to detect abnormal values. For example, if the daily energy consumption of an air conditioner exceeds the average value by more than 3 times, it is judged to be abnormal. High-energy-consuming equipment is identified through comparative analysis, such as equipment ranked in the top 20% in energy consumption.

[0095] Calculate the correlation coefficient matrix between the energy consumption of the equipment and its properties and operating conditions. Take the air conditioner as an example to analyze the correlation between its energy consumption and factors such as cooling capacity, outdoor temperature, and set temperature. Use multiple linear regression to build an influencing factor model, such as air conditioning energy consumption = a*cooling capacity + b*outdoor temperature + c*set temperature + d. By introducing interaction terms, analyze the interaction effects between factors, such as the interaction term between cooling capacity and outdoor temperature. Finally, get the power load curve of the computer room, showing the load changes within 24 hours.

[0096] Based on the operating status portrait of the computer room, hourly load power data within a year is collected to construct a time series of 8760 data points. The data is detrended and the moving average is subtracted to obtain the residual sequence. The data is standardized and scaled to the interval [-1, 1]. The processed sequence is fit using a long short-term memory network (LSTM). The network contains an LSTM layer (64 neurons) and a fully connected layer. The Adam optimizer is used to train for 50 epochs to obtain the time series dynamic law of load changes.

[0097] The weather data (such as temperature and humidity) and holiday information of the computer room are collected as exogenous variables. The exogenous variables are also standardized and then aligned with the load time series in the time dimension to obtain a comprehensive feature set including load data and exogenous variables. Based on the comprehensive feature set, a long-term trend model and a short-term fluctuation model are constructed respectively. The long-term trend model uses the Prophet algorithm to consider the annual and seasonal change trends. The short-term fluctuation model uses the ARIMA model to capture the short-term autoregressive and moving average characteristics.

[0098] The prediction results of the two models are weighted and fused, and the weights are used to determine the optimal value through grid search. The attention mechanism is introduced to assign different weights to predictions of different time scales. For example, for the prediction of the next 24 hours, the weight of recent data is higher, and the optimized load forecast results are obtained, including the hourly load forecast value for the next week.

[0099] Based on the optimization load forecast results and data traceability knowledge graph, the cost objective function and efficiency objective function of the distribution system are constructed. The cost objective function takes into account the equipment investment cost, operation and maintenance cost, and energy cost. The efficiency objective function measures the energy utilization rate of the system. The two objective functions are merged into a comprehensive objective function through a weighted method.

[0100] Determine the decision variables, including transformer capacity, line specifications, switch settings, etc., set constraints according to the grid topology, such as power balance constraints, voltage limit constraints, etc., and build a mixed integer programming model, in which integer variables represent equipment selection and continuous variables represent operating parameters.

[0101] The optimization problem is solved using the branch and bound method. The integer variables are relaxed into continuous variables to solve the linear programming problem, and the initial lower bound is obtained. An integer variable is selected for branching, and the constraints of rounding up and rounding down are added respectively to form two sub-problems. The sub-problems are solved recursively, and the upper and lower bounds are updated until the convergence conditions are reached. Finally, the optimal configuration parameters are obtained, including the model, capacity and operating parameter settings of each device.

[0102] Generate a power distribution system design plan based on the optimal configuration parameters, including system topology, equipment list, installation location diagram, etc. For example, determine that two 800kVA transformers, six 400mm2 cables, 15 800A air switches are needed. Give the specific installation location and connection relationship of each device, as well as the main operating parameter settings.

[0103] In this embodiment, the energy consumption data is detected for anomalies in combination with the box plot, and the inherent influence mechanism of the equipment energy consumption and its attributes and working conditions is revealed through correlation analysis and regression model, which provides a decision-making basis for energy-saving transformation. The time series dynamic law is modeled through the long short-term memory network, and the long-term trend and short-term fluctuation are predicted separately, which improves the spatiotemporal expression ability of load prediction. The model is optimized through grid search and attention mechanism, which improves the prediction accuracy and generalization performance. The objective function is solved by the branch and bound algorithm, and the optimal configuration parameters that take into account both cost and efficiency are obtained, and the optimal design scheme of the distribution system is generated, which improves the power utilization efficiency and power supply reliability. The power consumption law is grasped through load prediction, and the energy efficiency is improved through distribution optimization. A closed-loop energy efficiency management process from data collection, anomaly diagnosis, trend prediction to optimization control is formed, which improves the systematicness and sustainability of energy saving and emission reduction in the computer room. In summary, this embodiment realizes the refinement, prediction and optimization of the energy efficiency management of the computer room, improves the energy efficiency and resource utilization of the computer room infrastructure, and enriches the energy efficiency optimization theory and method under the background of energy Internet.

[0104] In an optional embodiment,

[0105] The comprehensive objective function corresponding to the distribution system is obtained through multi-objective weighted construction as shown in the following formula:

[0106] ;

[0107] Among them, MF represents the objective function value, represents the fixed cost associated with the ith device, C i0 represents the baseline cost of the ith device, represents the variable cost associated with the ith device, ρ represents the weight coefficient of the network loss function, and L j(x) represents the jth network loss function, which is related to power flow and power loss, μ represents the weight coefficient of renewable energy utilization, and H(x) represents the utilization rate of renewable energy.

[0108] In this embodiment, through a multi-objective weighted approach, the main optimization objectives such as fixed costs, variable costs, network losses, and renewable energy utilization rates are quantified and linearly combined to form a comprehensive evaluation index, thereby realizing multi-objective balanced optimization of the distribution system. The adjustable weight mechanism enhances the flexibility of the optimization model and can adapt to changing design requirements and preferences. After integrating multiple optimization objectives into a single comprehensive objective function, the original multi-objective optimization problem is transformed into a single-objective optimization problem, avoiding contradictions and conflicts among multiple objectives, reducing the dimensions of decision variables and constraints, simplifying the structure and solution process of the optimization model, and improving computational efficiency and convergence speed. In summary, this embodiment realizes multi-objective optimization of distribution system design, improves the flexibility and solution efficiency of the optimization model, clarifies the influence mechanism of various factors on system performance, and forms complementary advantages with the previous data analysis, reflecting the synergistic enhancement of domain knowledge and data intelligence, which is of great significance to further improving the planning and design level of distribution systems.

[0109] S3. Based on the power distribution system design, a self-learning controller is added to the power distribution unit, and the power load of the computer room is optimized in combination with the reinforcement learning algorithm. Based on the multi-source heterogeneous data, a degradation mechanism model of the distribution equipment is constructed through a machine learning algorithm, and the life of the distribution equipment is predicted. The life management strategy of the distribution equipment is output. According to the collected computer room operation data and the energy consumption law of the computer room, the data analysis algorithm is used to identify abnormal operating equipment, extract abnormal features and refer to historical abnormal data to determine the cause of the abnormality, generate an abnormal report and send it to the equipment maintenance personnel.

[0110] The self-learning controller is an intelligent system that can continuously learn and optimize its own control strategy through interaction with the environment. It does not require pre-set fixed rules and can adaptively adjust control parameters based on real-time feedback. The distribution equipment degradation mechanism model is used to describe the process of gradual degradation of distribution equipment over time or changes in the use environment, helping to predict the health status of the equipment and assess possible future failures. The life management strategy refers to a management method that extends the life of the equipment and reduces the failure rate through monitoring, maintenance and optimization operations throughout the life cycle of the equipment. The data analysis algorithm is an algorithm used to extract useful information from large amounts of data and reveal trends or patterns, providing a basis for decision-making.

[0111] In an optional embodiment,

[0112] Based on the power distribution system design, a self-learning controller is added to the power distribution unit, and the power load of the computer room is optimized in combination with the reinforcement learning algorithm. Based on the multi-source heterogeneous data, a degradation mechanism model of the power distribution equipment is constructed through a machine learning algorithm, and the life of the power distribution equipment is predicted. The life management strategy of the power distribution equipment is output. According to the collected computer room operation data and the energy consumption law of the computer room, the data analysis algorithm is used to identify the abnormal operation equipment, extract the abnormal features and refer to the historical abnormal data to determine the abnormal cause, generate an abnormal report and send it to the equipment maintenance personnel, including:

[0113] Based on the power distribution system design scheme, a self-learning controller is added to each power distribution unit, wherein the self-learning controller includes a perception layer, a decision layer and an execution layer, wherein the perception layer is used to obtain real-time operation data of the power distribution system, including voltage, current, power and environmental information. Based on the real-time operation data, the decision layer maps the state of the electrical equipment into continuous control actions through a deep deterministic policy gradient algorithm, sets the corresponding state space in combination with a reinforcement learning algorithm, constructs a discrete action space based on the adjustment action of each controllable electrical equipment, and constructs a reward objective function with the goal of minimizing system energy consumption, maximizing voltage qualification rate and minimizing equipment loss, solves the reward objective function and optimizes the power load of the computer room according to the function value;

[0114] Based on the optimized power load of the computer room and the multi-source heterogeneous data, the equipment status data in the multi-source heterogeneous data is encoded into a time series through a machine learning algorithm, the equipment degradation time series is generated and degradation features are extracted, based on the degradation features, the Shapley value corresponding to each degradation feature is calculated in combination with a long short-term memory network, and the degradation features are ranked in importance according to the Shapley value, the degradation features in the top 10% are taken as key degradation features, a degradation mechanism model is generated, and the parameters of the degradation mechanism model are initialized according to the key degradation features;

[0115] Based on the degradation mechanism model, combined with the pre-acquired historical fault data and expert experience, the equipment health is defined and the warning threshold and failure threshold of the equipment health are set. The equipment remaining life is modeled according to the random process theory, and the equipment degradation process is modeled as a Markov chain of continuous-time discrete states. The state transition probability is determined by the degradation mechanism model, and the absorption time distribution of the Markov chain is solved to obtain the equipment remaining life. Based on the equipment remaining life and the key degradation characteristics, the equipment maintenance decision is determined in combination with the historical fault data and the distribution equipment life management strategy is generated;

[0116] Based on the collected computer room operation data and the energy consumption pattern of the computer room, abnormal data is identified through a data analysis algorithm, and the operation abnormality features in the abnormal data are extracted. The current abnormal data is attribution analyzed in combination with a few-sample learning algorithm, the cause of the abnormality is determined, and an abnormal report is generated and sent to the equipment maintenance personnel.

[0117] The deep deterministic policy gradient algorithm is a reinforcement learning algorithm for solving continuous action space. By directly optimizing the policy gradient, efficient learning and decision-making of complex control tasks are achieved. The continuous control action refers to the continuous value of the action variable in the control system, which can be precisely adjusted, such as the control of temperature or voltage. The discrete action space refers to the action variable can only take a limited number of discrete values, which is usually used to simplify decision problems. The equipment degradation time series is a data series that records the degradation process of the equipment over time, which is usually used to analyze the changing trend of the equipment health status. The degradation feature refers to the key indicator or parameter that can reflect the degradation status of the equipment, which can evaluate the health of the equipment. The Shapley value is an allocation method derived from game theory, which is often used to explain the importance of features in model prediction. The marginal contribution of each feature in different combinations is calculated to evaluate its impact on the overall result. The equipment health is a quantitative evaluation of the current state of the equipment, which is usually used to measure whether the equipment is in normal operation. The warning threshold refers to the warning line set in equipment monitoring. When the equipment operation status reaches or exceeds this threshold, a warning will be triggered to indicate potential risks. The failure threshold is the maximum limit that the equipment can withstand during operation. When the threshold is exceeded, the equipment may malfunction or fail. The Markov chain is a mathematical model used to describe random processes, in which the future state of the system depends only on the current state and is independent of the past state. It is often used for modeling and predicting equipment degradation or health status. The attribution analysis is used to explain the main reasons behind a certain result or phenomenon, usually by analyzing the impact of various factors on the final result to help identify the root cause of the problem.

[0118] A self-learning controller is added to each distribution unit, which includes three main modules: perception layer, decision layer and execution layer. The perception layer is responsible for acquiring the operating data of the distribution system in real time, including voltage, current, power, ambient temperature and humidity, etc. After being collected by various sensors, the data is preprocessed and feature extracted to form a standardized data format for transmission to the decision layer.

[0119] Define the state space, including the operating parameters and load requirements of various electrical equipment, and then construct a discrete action space, including the adjustment actions of various controllable electrical equipment, such as transformer tap adjustment, reactive compensation switching, etc., with the optimization goals of minimizing system energy consumption, maximizing voltage qualification rate and minimizing equipment loss, construct a reward function, and continuously learn through interaction with the environment, update the policy network parameters, and finally output the optimal control strategy. For example, a distribution unit contains 1 transformer, 2 sets of reactive compensation devices and 10 important loads. Define the state space as a 15-dimensional vector, including the transformer secondary voltage, reactive compensation switching state, and the power of each load; define the action space as a 4-dimensional vector, corresponding to the transformer tap adjustment (-2, -1, 0, +1, +2) and the switching of 2 sets of reactive compensation devices (0, 1). After 1000 rounds of interaction with the simulation environment, the controller learns the optimal adjustment strategy under load changes, which increases the system voltage qualification rate from 90% to 99.5% and reduces energy consumption by 5%.

[0120] The execution layer is responsible for converting the control instructions output by the decision layer into actual equipment operation signals, such as transformer tap switching instructions, switch control signals, etc., and feeding back the execution results to the perception layer to form a closed-loop control. Based on the optimized power load of the computer room and multi-source heterogeneous data, a machine learning algorithm is used to construct a degradation mechanism model of distribution equipment, encode the equipment status data into a time series, and generate an equipment degradation time series. For example, the aging degree of the insulating paper of a transformer can be characterized by the furfural content. Its time-varying data is encoded into a degradation time series, and multi-dimensional degradation features such as statistical features, time domain features, and frequency domain features are extracted from the degradation time series.

[0121] The extracted degradation features are used as network input and the remaining life of the equipment as output. The long short-term memory network model is trained. The trained model is used to sort the degradation features by importance by calculating the contribution of each feature to the output (i.e., the Shapley value). The top 10% of the features are selected as key degradation features for subsequent modeling. For example, the key degradation features of a transformer may include the furfural content of insulating paper, the dissolved gas content in oil, the local discharge intensity, etc.

[0122] Based on the key degradation features, a degradation mechanism model is constructed. A deep learning model such as a recurrent neural network can be used to take the key degradation features as input to predict the future degradation trend of the equipment. The model parameters can be initialized using a pre-training method. The model can be pre-trained using historical data to improve the model convergence speed and accuracy. The equipment health index is defined by combining historical failure data and expert experience. The health can be expressed as a value between 0 and 1, with 1 representing a brand new state and 0 representing complete failure. A warning threshold (such as 0.3) and a failure threshold (such as 0.1) are set. When the health is lower than the warning threshold, an early warning is issued. When it is lower than the failure threshold, the equipment is judged to be failed.

[0123] The equipment degradation process is modeled as a Markov chain of continuous-time discrete states. The state transition probability is given by the degradation mechanism model. By solving the absorption time distribution of the Markov chain, the probability distribution of the remaining life of the equipment is obtained. For example, the remaining life prediction result of a transformer may be: an average remaining life of 5 years, and a 90% confidence interval of 3-7 years.

[0124] Based on the remaining life prediction results and key degradation characteristics, combined with historical fault data, equipment maintenance decisions are made. A risk-based maintenance strategy can be adopted to seek a balance between equipment failure risk and maintenance cost. For example, when the predicted remaining life of the transformer is less than 2 years and the partial discharge intensity exceeds the warning value, it is recommended to carry out a major overhaul. When the remaining life is between 2-5 years, it is recommended to strengthen monitoring and preventive maintenance, and finally generate a distribution equipment life management strategy, including maintenance cycles, maintenance items, and replacement time.

[0125] According to the collected data and energy consumption patterns of the computer room, abnormal data is identified through data analysis algorithms. First, a normal operation mode is established based on historical data, including the power curve of each device, temperature change pattern, etc. Then, the real-time data is compared with the normal mode to detect abnormal points that deviate from the normal range. For example, the power consumption of an air-conditioning device suddenly increases by 50%, or the temperature of a server rises abnormally, which may be identified as abnormal data. Operation abnormality features are extracted from the abnormal data, such as the size of the abnormal value, duration, and change rate. Combined with the few-sample learning algorithm, the current abnormality is attributed and analyzed. Specifically, a knowledge base is constructed using historical abnormal cases, and each case contains abnormal features and corresponding causes. For newly emerging abnormalities, the most similar cases are found by calculating the similarity with historical cases, and the possible abnormal causes are inferred. For example, the temperature of a server rises abnormally, and the extracted abnormal features include: temperature rise rate of 0.5℃ / minute, duration of 30 minutes, and the highest temperature exceeds the threshold of 10℃. By comparing with historical cases, it is found that it is highly similar to the "cooling fan failure" case, so it is judged that the current abnormal cause may be a failure of the server cooling fan.

[0126] Generate an exception report and send it to the equipment maintenance personnel. The exception report includes abnormal equipment information, abnormal feature description, inferred cause and recommended treatment measures. Maintenance personnel can quickly locate the problem and take corresponding maintenance measures based on the report, thereby improving equipment maintenance efficiency and reducing failure downtime.

[0127] In this embodiment, by adding a self-learning controller in each distribution unit, and using a deep deterministic policy gradient algorithm and a reinforcement learning algorithm, the state of the electrical equipment is mapped to the continuous control action, and the energy consumption is minimized, the voltage qualification rate is maximized, and the equipment loss is minimized as the optimization goal, and a reward function is constructed for load optimization, so that the adaptive regulation and real-time optimization of the distribution system are realized, and the economy and reliability of the system operation are improved. The degradation mechanism model is initialized based on the key characteristics, and the concept of equipment health is introduced. The equipment degradation process is modeled as a Markov chain. By solving the state transition probability and the absorption time distribution, the remaining life of the equipment is accurately predicted. Predictive maintenance can be inspected and repaired before the equipment performance deteriorates, avoiding sudden failures and unplanned downtime, extending the service life of the equipment, reducing maintenance costs, and improving the availability and operation efficiency of the distribution system. In summary, this embodiment realizes the coordinated management and control of the entire life cycle of the distribution equipment, and plays an important role in improving power supply reliability, reducing operation and maintenance costs, and ensuring power quality. It has a demonstrative significance for promoting the intelligent transformation of the energy and power industry, and has important value for building a safe, economical, green and efficient modern distribution system.

[0128] In an optional embodiment,

[0129] Solving the absorption time distribution of the Markov chain, the remaining life of the equipment is obtained as shown in the following formula:

[0130] ;

[0131] Among them, R(t) represents the expected value of the remaining life of the current equipment, π s represents the initial state distribution, P represents the state transition probability matrix, e is a vector of all 1s, represents the cumulative probability calculation, ζ represents the scale parameter of the Weibull distribution, and β represents the shape parameter of the Weibull distribution.

[0132] In this embodiment, the probabilistic prediction method is more in line with the random characteristics of equipment degradation, can quantify the uncertainty of remaining life, and provide probabilistic risk information for maintenance decisions. The Markov chain model makes full use of the complete information of the degradation process through multi-state modeling, improves the long-term consistency and robustness of the remaining life prediction, and is particularly applicable to equipment with obvious stages in the degradation process. By combining data-driven and model-driven methods, it not only takes into account the authenticity and timeliness of the data, but also incorporates the guidance and explanatory nature of domain knowledge, thereby improving the accuracy and reliability of the remaining life prediction. In summary, this embodiment realizes the probabilistic prediction of the remaining life, improves the accuracy and robustness of the long-term prediction, integrates the data-driven and model-driven modeling paradigms, simplifies the calculation process of the remaining life distribution, provides an important basis for formulating a quantitative predictive maintenance strategy, and provides a core model support for the full life cycle management of equipment. It is of great significance to improve equipment reliability, reduce maintenance costs, and ensure system safety.

[0133] Figure 2 FIG. 1 is a schematic diagram of the structure of the configuration system of the overall electrical solution of the computer room according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0134] The first unit is used to deploy an intelligent sensor gateway in the computer room based on a distributed data architecture, collect multi-source heterogeneous data corresponding to the current computer room in real time, perform feature extraction and semantic mapping on the multi-source heterogeneous data through a preset deep learning model, build a high-dimensional semantic association network, perform semantic enhancement based on the high-dimensional semantic association network and cross-modal representation learning technology, perform semantic modeling and logical association on the semantically enhanced multi-source heterogeneous data through knowledge graph technology, obtain data tracing knowledge graph and data flow trajectory, extract abnormal patterns in current data based on the data flow trajectory and stream data intelligent processing framework, and obtain a portrait of the computer room operation status;

[0135] The second unit is used to construct a digital portrait model of energy consumption in the computer room based on the data tracing knowledge graph and the operating status portrait of the computer room, determine the energy consumption law and high-energy consumption equipment in the computer room, perform composite characteristic analysis and abnormal diagnosis on the high-energy consumption equipment, obtain the power load curve of the computer room, take the operating status portrait of the computer room and the power load curve of the computer room as input, perform hierarchical progressive prediction in combination with the exogenous variables corresponding to the current computer room, obtain the optimized load prediction result, take the optimized load prediction result and the data tracing knowledge graph as input, construct an objective function with the goal of minimizing cost and maximizing energy utilization, solve to obtain the optimal configuration parameters and generate a distribution system design plan;

[0136] The third unit is used to add a self-learning controller in the distribution unit based on the distribution system design scheme, optimize the power load of the computer room in combination with the reinforcement learning algorithm, build a distribution equipment degradation mechanism model based on the multi-source heterogeneous data through a machine learning algorithm and predict the life of the distribution equipment, output the distribution equipment life management strategy, identify abnormal operating equipment through a data analysis algorithm according to the collected computer room operation data and the energy consumption law of the computer room, extract abnormal features and refer to historical abnormal data to determine the cause of the abnormality, generate an abnormal report and send it to the equipment maintenance personnel.

[0137] According to a third aspect of the embodiments of the present invention,

[0138] An electronic device is provided, comprising:

[0139] processor;

[0140] a memory for storing processor-executable instructions;

[0141] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0142] A fourth aspect of the embodiments of the present invention is:

[0143] 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.

[0144] 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.

[0145] 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. The configuration method of the overall electrical scheme of the computer room is characterized by: include: Based on the distributed data architecture, an intelligent sensor gateway is deployed in the computer room to collect multi-source heterogeneous data corresponding to the current computer room in real time. Feature extraction and semantic mapping are performed on the multi-source heterogeneous data through a pre-set deep learning model to construct a high-dimensional semantic association network. Semantic enhancement is performed based on the high-dimensional semantic association network and cross-modal representation learning technology. The semantically enhanced multi-source heterogeneous data is semantically modeled and logically associated through knowledge graph technology to obtain a data tracing knowledge graph and a data flow trajectory. Based on the data flow trajectory and the stream data intelligent processing framework, abnormal patterns in the current data are extracted to obtain a portrait of the computer room operation status. Based on the data tracing knowledge graph and the operating status portrait of the computer room, a digital portrait model of the energy consumption of the computer room is constructed, the energy consumption pattern and high-energy consumption equipment of the computer room are determined, and the composite characteristic analysis and abnormal diagnosis of the high-energy consumption equipment are performed to obtain the power load curve of the computer room. The operating status portrait of the computer room and the power load curve of the computer room are used as inputs, and hierarchical progressive prediction is performed in combination with the exogenous variables corresponding to the current computer room to obtain an optimized load prediction result. The optimized load prediction result and the data tracing knowledge graph are used as inputs, and an objective function is constructed with the goal of minimizing cost and maximizing energy utilization, and the optimal configuration parameters are obtained and a distribution system design scheme is generated; Based on the power distribution system design, a self-learning controller is added to the power distribution unit, and the power load of the computer room is optimized in combination with the reinforcement learning algorithm. Based on the multi-source heterogeneous data, a degradation mechanism model of the distribution equipment is constructed through a machine learning algorithm, and the life of the distribution equipment is predicted. The life management strategy of the distribution equipment is output, and according to the collected computer room operation data and the energy consumption law of the computer room, the data analysis algorithm is used to identify abnormal operating equipment, extract abnormal features and refer to historical abnormal data to determine the cause of the abnormality, generate an abnormal report and send it to the equipment maintenance personnel.

2. The method according to claim 1, characterized in that Based on the distributed data architecture, an intelligent sensor gateway is deployed in the computer room to collect multi-source heterogeneous data corresponding to the current computer room in real time. Feature extraction and semantic mapping of the multi-source heterogeneous data are performed through a pre-set deep learning model to build a high-dimensional semantic association network. Semantic enhancement based on the high-dimensional semantic association network and cross-modal representation learning technology includes: According to the characteristics of the computer room environment and task requirements, an intelligent sensor gateway is deployed in the electrical computer room in combination with a distributed data architecture. The multi-source heterogeneous data corresponding to the current computer room is collected in real time according to the intelligent sensor gateway, wherein the multi-source heterogeneous data includes power consumption data of the computer room, equipment status data and environmental monitoring data. The high-frequency noise and abnormal points in the multi-source heterogeneous data are removed based on the Kalman filter algorithm, and the missing points in the multi-source heterogeneous data are completed through matrix completion to obtain pre-processed multi-source heterogeneous data. Based on a preset deep learning model, feature extraction is performed on the preprocessed multi-source heterogeneous data. For the multi-source heterogeneous data, the corresponding time-frequency domain features are extracted by a moving average algorithm. For the environmental image corresponding to the environmental monitoring data, the corresponding multi-scale texture features are extracted by a convolutional neural network. Combined with the attention mechanism, the time-frequency domain features and the multi-scale texture features are mapped to a common semantic space. In the common semantic space, for each different feature, the semantic association strength of each feature pair is determined based on cosine similarity, and a high-dimensional semantic association network corresponding to the multi-source heterogeneous data is constructed in combination with an improved graph attention network. Based on the high-dimensional semantic association network, a cross-modal semantic embedding framework is set up, and cross-modal positive and negative sample pairs are constructed in combination with contrastive learning. A consistent semantic representation space is obtained by maximizing the similarity of positive samples and the difference of negative samples. An ontology knowledge base is constructed according to the computer room field, and the multi-source heterogeneous data is semantically annotated in combination with the concepts and relationships in the ontology knowledge base, and semantic enhancement is performed in combination with ontology knowledge constraints.

3. The method according to claim 1, characterized in that The semantically enhanced multi-source heterogeneous data is semantically modeled and logically associated through knowledge graph technology to obtain data traceability knowledge graph and data flow trajectory. Based on the data flow trajectory and stream data intelligent processing framework, abnormal patterns in the current data are extracted to obtain the operating status portrait of the computer room, including: Based on the enhanced multi-source semantic data, knowledge representation is performed based on the pre-built ontology knowledge base, data entities, attributes and relationships are defined, feature triples are generated and a composite knowledge association framework based on rules and graph embedding is designed. Logical reasoning is performed on the knowledge-represented data according to the pre-acquired domain rule base to generate hidden associations between different ontologies. Based on the hidden associations, a data traceability knowledge graph is constructed by combining the temporal constraints represented by temporal logic. Through the knowledge graph reasoning technology, the paths between different nodes are determined to obtain the data flow trajectory. Based on the data flow trajectory and the data traceability knowledge graph, the dependencies between different electrical equipment and the fault propagation mechanism between electrical equipment are encoded as constraint rules, and a directed acyclic graph is constructed through a stream data intelligent processing framework, wherein nodes represent multi-source heterogeneous data, and edges represent dependencies between data. The constraint rules are used as prior knowledge, and anomaly point detection is performed through the isolation forest algorithm. If an anomaly point exists, all root cause paths are determined from the current anomaly point through the shortest path algorithm of the graph. For each root cause path, automatic scoring is performed in combination with domain knowledge, and the root cause path with the highest score is selected as the anomaly mode corresponding to the current anomaly point to generate a portrait of the computer room operation status.

4. The method according to claim 1, characterized in that: Based on the data tracing knowledge graph and the operating status portrait of the computer room, a digital portrait model of the energy consumption of the computer room is constructed, the energy consumption law and high-energy consumption equipment of the computer room are determined, and the composite characteristic analysis and abnormal diagnosis of the high-energy consumption equipment are performed to obtain the power load curve of the computer room. The operating status portrait of the computer room and the power load curve of the computer room are used as inputs, and hierarchical progressive prediction is performed in combination with the exogenous variables corresponding to the current computer room to obtain an optimized load prediction result. The optimized load prediction result and the data tracing knowledge graph are used as inputs, and an objective function is constructed with the goal of minimizing cost and maximizing energy utilization. The optimal configuration parameters are obtained and a distribution system design scheme is generated, including: Based on the data traceability knowledge graph and the operating status portrait of the computer room, the time series data corresponding to the energy consumption information is extracted, the support of each data point in the time series data is determined according to the association rule mining algorithm, the frequent item set is determined based on the support and the preset minimum support threshold, the frequent contribution mode and sequence mode of the energy consumption of electrical equipment are obtained based on the frequent item set, the energy consumption of the computer room is obtained, the energy consumption data corresponding to each device is grouped and aggregated according to the device type and device location, the outliers in the energy consumption data are detected in combination with the box plot, the high-energy consumption equipment is determined, the correlation coefficient matrix between the energy consumption of the equipment and its own attributes and operating conditions is calculated, the correlation strength is determined and a regression model of influencing factors is constructed, the degree of influence of the interaction effect between each influencing factor on the energy consumption is determined in combination with the interaction term, and the power load curve of the computer room is obtained; Collect load power data from the operating status portrait of the computer room, construct a load time series data set, perform detrending and standardization operations on each load time series data, fit the load time series with a long short-term memory network, obtain time series dynamic laws, collect weather and holiday information at the location of the computer room as exogenous variables, standardize the exogenous variables and align them with the load time series data, align the aligned exogenous variables with the load time series data to obtain a comprehensive feature set, perform long-term trend modeling and short-term fluctuation modeling based on the comprehensive feature set, generate long-term trend prediction results and short-term fluctuation prediction results and fuse them, and combine grid search strategy and attention mechanism to obtain the optimized load prediction result; Based on the optimized load forecast results and the data traceability knowledge graph, a cost objective function and an efficiency objective function are constructed, and a comprehensive objective function corresponding to the distribution system is obtained through multi-objective weighted construction. For the comprehensive objective function, decision variables are determined, and constraints are set according to the power grid topology and equipment parameter variables. A mixed integer programming model is constructed, and the comprehensive objective function is solved by the branch and bound method to obtain the optimal configuration parameters. The distribution system design scheme is generated according to the optimal configuration parameters.

5. The method according to claim 4, characterized in that The comprehensive objective function corresponding to the distribution system is obtained through multi-objective weighted construction as shown in the following formula: ； Among them, MF represents the objective function value, represents the fixed cost associated with the i-th equipment, C i0 represents the baseline cost of the ith device, represents the variable cost associated with the i-th device, ρ represents the weight coefficient of the network loss function, and L j (x) represents the jth network loss function, which is related to power flow and power loss, μ represents the weight coefficient of renewable energy utilization, and H(x) represents the utilization rate of renewable energy.

6. The method according to claim 1, characterized in that Based on the power distribution system design, a self-learning controller is added to the power distribution unit, and the power load of the computer room is optimized in combination with the reinforcement learning algorithm. Based on the multi-source heterogeneous data, a degradation mechanism model of the power distribution equipment is constructed through a machine learning algorithm, and the life of the power distribution equipment is predicted. The life management strategy of the power distribution equipment is output. According to the collected computer room operation data and the energy consumption law of the computer room, the data analysis algorithm is used to identify the abnormal operation equipment, extract the abnormal features and refer to the historical abnormal data to determine the abnormal cause, generate an abnormal report and send it to the equipment maintenance personnel, including: Based on the power distribution system design scheme, a self-learning controller is added to each power distribution unit, wherein the self-learning controller includes a perception layer, a decision layer and an execution layer, wherein the perception layer is used to obtain real-time operation data of the power distribution system, including voltage, current, power and environmental information. Based on the real-time operation data, the decision layer maps the state of the electrical equipment into continuous control actions through a deep deterministic policy gradient algorithm, sets the corresponding state space in combination with a reinforcement learning algorithm, constructs a discrete action space based on the adjustment action of each controllable electrical equipment, and constructs a reward objective function with the goal of minimizing system energy consumption, maximizing voltage qualification rate and minimizing equipment loss, solves the reward objective function and optimizes the power load of the computer room according to the function value; Based on the optimized power load of the computer room and the multi-source heterogeneous data, the equipment status data in the multi-source heterogeneous data is encoded into a time series through a machine learning algorithm, the equipment degradation time series is generated and degradation features are extracted, based on the degradation features, the Shapley value corresponding to each degradation feature is calculated in combination with a long short-term memory network, and the degradation features are ranked in importance according to the Shapley value, the degradation features in the top 10% are taken as key degradation features, a degradation mechanism model is generated, and the parameters of the degradation mechanism model are initialized according to the key degradation features; Based on the degradation mechanism model, combined with the pre-acquired historical fault data and expert experience, the equipment health is defined and the warning threshold and failure threshold of the equipment health are set. The equipment remaining life is modeled according to the random process theory, and the equipment degradation process is modeled as a Markov chain of continuous-time discrete states. The state transition probability is determined by the degradation mechanism model, and the absorption time distribution of the Markov chain is solved to obtain the equipment remaining life. Based on the equipment remaining life and the key degradation characteristics, the equipment maintenance decision is determined in combination with the historical fault data and the distribution equipment life management strategy is generated; Based on the collected computer room operation data and the energy consumption pattern of the computer room, abnormal data is identified through a data analysis algorithm, and the operation abnormality features in the abnormal data are extracted. The current abnormal data is attribution analyzed in combination with a few-sample learning algorithm, the cause of the abnormality is determined, and an abnormal report is generated and sent to the equipment maintenance personnel.

7. The method according to claim 6, characterized in that Solving the absorption time distribution of the Markov chain, the remaining life of the equipment is obtained as shown in the following formula: ； Among them, R(t) represents the expected value of the remaining life of the current equipment, π s represents the initial state distribution, P represents the state transition probability matrix, e is a vector of all 1s, represents the cumulative probability calculation, ζ represents the scale parameter of the Weibull distribution, and β represents the shape parameter of the Weibull distribution.

8. A system for configuring the overall electrical solution of a computer room, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to deploy an intelligent sensor gateway in the computer room based on a distributed data architecture, collect multi-source heterogeneous data corresponding to the current computer room in real time, perform feature extraction and semantic mapping on the multi-source heterogeneous data through a preset deep learning model, build a high-dimensional semantic association network, perform semantic enhancement based on the high-dimensional semantic association network and cross-modal representation learning technology, perform semantic modeling and logical association on the semantically enhanced multi-source heterogeneous data through knowledge graph technology, obtain data tracing knowledge graph and data flow trajectory, extract abnormal patterns in current data based on the data flow trajectory and stream data intelligent processing framework, and obtain a portrait of the computer room operation status; The second unit is used to construct a digital portrait model of energy consumption in the computer room based on the data tracing knowledge graph and the operating status portrait of the computer room, determine the energy consumption law and high-energy consumption equipment in the computer room, perform composite characteristic analysis and abnormal diagnosis on the high-energy consumption equipment, obtain the power load curve of the computer room, take the operating status portrait of the computer room and the power load curve of the computer room as input, perform hierarchical progressive prediction in combination with the exogenous variables corresponding to the current computer room, obtain the optimized load prediction result, take the optimized load prediction result and the data tracing knowledge graph as input, construct an objective function with the goal of minimizing cost and maximizing energy utilization, solve to obtain the optimal configuration parameters and generate a distribution system design plan; The third unit is used to add a self-learning controller in the distribution unit based on the distribution system design scheme, optimize the power load of the computer room in combination with the reinforcement learning algorithm, build a distribution equipment degradation mechanism model based on the multi-source heterogeneous data through a machine learning algorithm and predict the life of the distribution equipment, output the distribution equipment life management strategy, identify abnormal operating equipment through a data analysis algorithm according to the collected computer room operation data and the energy consumption law of the computer room, extract abnormal features and refer to historical abnormal data to determine the cause of the abnormality, generate an abnormal report and send it to the equipment maintenance personnel.

9. 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 described in any one of claims 1 to 7.

10. 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 7 is implemented.

Images