A multi-level energy management system based on multi-dimensional data

Abstract

The present invention discloses a multi-level energy management system based on multi-dimensional data, which relates to the field of energy intelligent management technology. The system includes a multi-source data acquisition module for real-time collection of electricity consumption, environment and equipment status data; a data fusion processing module for processing outliers and fusing multi-scale data features through algorithms; an energy status assessment module for implementing equipment status assessment and early warning using fusion algorithms and prediction models; a multi-level energy scheduling module for dynamically adjusting strategies using optimization algorithms to balance energy costs, production efficiency and carbon emissions; an energy performance analysis module for developing analysis tools and evaluation models; and a decision support module for configuring expert knowledge bases and developing fault diagnosis systems and knowledge graphs. Through multi-dimensional data acquisition and multi-level management, the present invention significantly reduces energy data errors, significantly reduces comprehensive energy costs and carbon emissions, reduces operating costs for participating enterprises, and improves the accuracy, safety and sustainability of energy management.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent energy management, and in particular to a multi-level energy management system based on multi-dimensional data. Background Art

[0002] Driven by the global energy transition, traditional energy management models are no longer able to meet the needs of modern industrial and social development. Existing energy management systems rely on single-point data collection, often using fixed-cycle manual meter reading or low-precision sensors. This results in errors in electricity and energy consumption data as high as ±5%-±8%, failing to accurately reflect real-time energy consumption. For example, due to data lags and errors, a large industrial park was unable to promptly detect abnormal energy consumption fluctuations in its production lines, resulting in a 12% overspending on monthly electricity bills. Furthermore, the data source is limited to power parameters, lacking the coordinated collection of multi-dimensional information such as ambient temperature and humidity and equipment operating status. This makes it difficult to build a complete energy consumption analysis model and provide a comprehensive basis for energy conservation optimization.

[0003] Energy scheduling and decision-making processes lack intelligence. Most companies still rely on experience-driven energy scheduling strategies, which prevent them from quickly adjusting energy allocation in the face of electricity price fluctuations, equipment failures, or intermittent renewable energy supply. Traditional energy performance evaluations rely on manual statistics and simple comparisons, failing to deeply analyze the relationship between energy consumption and production processes and equipment efficiency, making it difficult to identify high-energy-consuming areas. A manufacturing company was unable to quantify the energy contribution of each workshop, resulting in insufficiently targeted energy-saving measures and a low return on investment.

[0004] There are significant conflicts between energy security and collaborative management. Existing systems lack full lifecycle health monitoring for energy equipment. Equipment failure warnings rely on manual inspections, resulting in less than 30% early detection of faults. One thermal power plant suffered an unplanned 48-hour shutdown due to a turbine hazard not being promptly detected, resulting in direct economic losses exceeding 5 million yuan. Furthermore, energy data is isolated between enterprises, preventing upstream and downstream sectors of the industrial chain from sharing energy resources and collaboratively optimizing them, leading to overall low efficiency in the regional energy system. Furthermore, carbon footprint accounting often relies on crude estimation methods, which cannot accurately quantify carbon emissions throughout a product's lifecycle, making it difficult to meet regulatory requirements for carbon trading and green development. Summary of the Invention

[0005] The present invention proposes a multi-level energy management system based on multi-dimensional data to solve the problems mentioned in the above-mentioned prior art.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a multi-level energy management system based on multi-dimensional data, comprising the following modules:

[0007] Multi-source data acquisition module: Configure electricity meters to collect electricity consumption data, deploy IoT sensors to monitor environmental parameters, integrate industrial controllers to obtain equipment operating status data, access meteorological data APIs, and build a multi-source heterogeneous data acquisition system;

[0008] Data fusion processing module: This module uses a time synchronization algorithm to align the spatiotemporal data of multiple sources, processes outliers through a combination of wavelet transform and Kalman filtering, builds a time series feature extraction model based on graph neural networks, and designs a three-level feature fusion architecture.

[0009] Energy Status Assessment Module: Establishes an energy equipment health evaluation index system, develops a multi-source information fusion algorithm based on DS evidence theory, constructs a Markov chain prediction model to predict the probability of energy equipment failure, quantifies equipment uncertainty by calculating the state information entropy value, and implements fault tracing by combining Bayesian networks;

[0010] Multi-level energy scheduling module: Construct a three-level scheduling architecture, use a multi-objective optimization algorithm to balance energy costs, production efficiency and carbon emissions, and propose a multi-level energy collaborative scheduling algorithm based on the improved NSGA-III: the objective function is , f(x) is a multi-objective function vector, consisting of m objective functions Composition, x is the decision variable, and the inequality constraint is is an inequality constraint, i ranges from 1 to p, indicating that there are p inequality constraints, and the equality constraints are ;

[0011] Energy Performance Analysis Module: Develop an energy flow analysis tool, build an energy efficiency evaluation model based on data envelopment analysis, and establish an energy cost decomposition model. By tracking the energy conversion path in the production process, the energy efficiency loss in each link is quantified, and the Shapley value method is used to achieve fair cost allocation.

[0012] Decision support module: configure expert knowledge base, build energy management knowledge graph, design symbolic reasoning engine to process deterministic knowledge, connectionist reasoning engine to process uncertain knowledge, and realize cross-domain knowledge transfer through knowledge graph embedding technology.

[0013] Furthermore, the multi-source data acquisition module further includes: using an adaptive sampling algorithm based on variational mode decomposition, and the objective function is: ;

[0014] The constraints are: , K is the number of modes; is the modal function, is the center frequency; is the Dirac delta function; represents the partial derivative with respect to time t; j is an imaginary unit; f(t) is the original signal; this module configures gas sensors to monitor greenhouse gas emissions, deploys a drone inspection system to obtain factory energy facility status data, integrates blockchain technology to ensure that data cannot be tampered with, and develops edge computing nodes to implement local data preprocessing and feature extraction.

[0015] Furthermore, the data fusion processing module also includes: developing a multimodal data fusion algorithm based on the attention mechanism, building a time series anomaly detection model, and deploying a federated learning framework to protect data privacy; this module realizes the transfer of energy management knowledge between enterprises by constructing a teacher-student model architecture without leaking the original data.

[0016] Furthermore, the energy status assessment module also includes: establishing a digital twin model of energy equipment, developing an abnormal sound recognition system for equipment based on convolutional neural networks, deploying a fiber optic distributed temperature sensing system to monitor the temperature of key equipment in real time, building an energy system vulnerability assessment model, and performing parameter estimation through the fusion of physical models and data-driven models.

[0017] Furthermore, the multi-level energy scheduling module also includes: developing a short-term energy scheduling algorithm based on model predictive control, building a demand response management system, deploying a distributed energy collaborative control platform, realizing the interaction between renewable energy consumption and power grid, and building a virtual energy storage system to participate in power grid auxiliary services by aggregating adjustable load resources.

[0018] Furthermore, the energy performance analysis module also includes: developing an energy consumption prediction model based on machine learning, building an energy efficiency benchmarking analysis tool, deploying an energy management dashboard, realizing the visualization and dynamic monitoring of energy performance, displaying three-dimensional indicators through three-dimensional space, and combining interactive drill-down functions to support in-depth data analysis.

[0019] Furthermore, it also includes: energy carbon footprint accounting module, using life cycle assessment method to build carbon footprint accounting model, developing carbon emission reduction path optimization algorithm, deploying blockchain carbon trading platform, realizing carbon emission accounting and trading, and quantitatively analyzing the flow path and conversion efficiency of carbon in the energy system by constructing energy-carbon flow coupling model.

[0020] Furthermore, it also includes: energy security management module, development of energy system protection system based on network security situation awareness, construction of energy equipment failure emergency response model, deployment of energy system disaster recovery system, design of "active immune" security architecture, training of defense models through simulated attack scenarios, and realization of early warning and defense against unknown threats.

[0021] Furthermore, it also includes: an energy knowledge graph module, which builds a knowledge graph that includes energy equipment, process flows, and management specification entities, develops a knowledge reasoning engine based on graph neural networks, deploys a question-answering system, and automatically updates the knowledge graph structure and parameters by continuously learning new data and expert experience.

[0022] Furthermore, it also includes: energy ecological collaboration module, development of energy data sharing platform, construction of cross-enterprise energy collaborative optimization model, deployment of blockchain contract system, realization of collaborative management and value co-creation of upstream and downstream of the energy industry chain; automatic execution of energy transactions and value distribution through blockchain contracts, and construction of a win-win energy ecosystem for all parties.

[0023] Compared with the existing technology, the beneficial effects of the present invention are:

[0024] The multi-source data acquisition module builds a comprehensive data collection system, integrating multiple types of data, including electricity consumption, environment, and equipment. Through adaptive sampling algorithms and edge computing, it ensures data accuracy and real-time performance, providing a solid data foundation for subsequent management. The data fusion processing module uses multiple algorithms to handle outliers, fuse features, and incorporates federated learning to protect data privacy, effectively integrating heterogeneous data from multiple sources.

[0025] The energy status assessment module leverages a variety of technologies to accurately assess equipment health and provide fault warnings. Technologies like digital twins further enhance assessment reliability and reduce the risk of equipment failure. The multi-level energy scheduling module builds a hierarchical scheduling architecture and employs a multi-objective optimization algorithm to balance energy costs and carbon emissions while ensuring production efficiency, thereby enhancing renewable energy absorption capacity.

[0026] The Energy Performance Analysis module utilizes a variety of analytical tools and models to provide in-depth analysis and visualization of energy consumption, helping companies pinpoint energy consumption issues. New modules, such as energy carbon footprint accounting and security management, further enhance system functionality, ensuring the stable operation of energy systems from carbon emissions management to network security protection.

[0027] Through multi-module collaboration and technological innovation, this application can not only significantly reduce corporate energy costs, improve energy utilization efficiency and equipment management level, but also promote regional energy collaborative optimization and help enterprises achieve green and low-carbon transformation. It has significant economic and social value in the field of energy management and is of great significance to promoting the intelligent development of the energy industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a schematic block diagram of a multi-level energy management system based on multi-dimensional data proposed by the present invention;

[0029] Figure 2The following is a bar chart comparing the power consumption prediction accuracy of the traditional method and the system method over time;

[0030] Figure 3 A radar chart comparing the traditional management scores and the scores of this system under different evaluation indicators (maximum 100 points). DETAILED DESCRIPTION

[0031] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0032] The entire system involves multiple modules, including multi-source data acquisition, time synchronization, data processing, and feature fusion, all of which operate in close coordination and collaboration. After the multi-source data acquisition module acquires various raw data, the NTP time synchronization module ensures data time accuracy. The wavelet transform and adaptive Kalman filter modules pre-process the data to remove noise and correct anomalies. The feature fusion architecture then further explores the data's value. These modules collaborate to provide reliable support for energy system status assessment and optimized scheduling. The present invention will be further described in detail below, with reference to the accompanying drawings.

[0033] Reference Figures 1 to 3 :A multi-level energy management system based on multi-dimensional data, including the following modules:

[0034] Multi-source data acquisition module: Huawei smart meters with 0.2S-level accuracy are used for electricity data collection. These meters are installed on each of the company's power branches and collect voltage, current, and power data at a frequency of 1 minute. These data are transmitted to edge computing nodes via the Modbus-TCP protocol. DHT22 temperature and humidity sensors are deployed for environmental parameter monitoring, evenly distributed across the factory in a 50m x 50m grid. Temperature accuracy reaches ±0.1°C and humidity accuracy reaches ±3%. Data is aggregated via the LoRa wireless communication network. Industrial equipment operating status data collection integrates Siemens S7-1200 series PLC controllers, which use the Profinet protocol to obtain real-time information such as equipment start / stop, speed, and load, with a response time of less than 10ms. Meteorological data, such as wind speed and light intensity, is updated every 15 minutes via the China Meteorological Administration's API and transmitted to edge computing nodes. These different data collection components work together under the unified scheduling of the edge computing nodes. The edge computing nodes allocate resources based on the operating status and data transmission requirements of each device, ensuring timely and accurate data collection. For example, when industrial equipment is in a critical stage of operation, the edge computing node will prioritize processing data from the PLC to ensure real-time monitoring of the equipment's operating status; when environmental parameters suddenly change, the temperature and humidity sensor data will be processed in a timely manner so that the company can respond quickly.

[0035] To achieve adaptive sampling, the variational mode decomposition (VMD) algorithm is used. The optimization problem of this algorithm includes the objective function and constraints. The objective function is Variables need to be (modal function) and (center frequency) is adjusted so that the objective function Take the minimum value. Accumulate the items from k=1 to k=K (K is the number of modes). A measure of signal energy. Take the partial derivative with respect to time t and act on The Dirac delta function δ(t) is infinite at t=0 and 0 at other times, and ; The imaginary unit j satisfies ; ∗ is the convolution operation; is the kth mode function; is a complex exponential function, is the center frequency of the kth mode, which determines the frequency domain distribution of the signal.

[0036] The constraints are That is, to optimize the objective function under this condition, the K modal functions obtained by decomposition are required to be The sum of the two equals the original signal f(t), ensuring that the decomposed signal accurately restores the original signal. The VMD algorithm dynamically adjusts the sampling frequency based on the acquired signal. When the signal fluctuates significantly and is highly complex, the sampling frequency is automatically increased to accurately capture signal characteristics; when the signal is stable, the sampling frequency is correspondingly reduced. This adaptive sampling method compresses the original data volume by 8 times while ensuring data integrity, reducing data transmission pressure and storage costs. The VMD algorithm works closely with the multi-source data acquisition module. After receiving multi-source data, the edge computing node uses the VMD algorithm to process the data. For electricity usage data, when enterprises enter peak production periods and current and voltage signals fluctuate complexly, the VMD algorithm increases the sampling frequency to accurately capture data changes, providing more accurate data for electricity safety monitoring and energy consumption analysis. For industrial equipment operating status data, the sampling frequency is adjusted promptly during key operating conditions such as equipment startup and shutdown to accurately obtain information such as equipment speed and load, facilitating equipment fault warning and maintenance. Furthermore, the data processed by the edge computing node is stored using blockchain evidence storage technology. Blockchain's distributed ledger ensures the security and immutability of multi-source data from collection, processing, and storage. The real-time data processing capabilities of edge computing combined with the security of blockchain evidence storage provide reliable protection for industrial enterprise data assets. This multi-technology integration model represents an innovative approach to data management within the Industrial Internet of Things (IIoT), distinguishing it from traditional data management methods.

[0037] Data Fusion Processing Module: The data fusion processing module first performs spatiotemporal alignment of multi-source data, using the NTP network time protocol to synchronize each sensor with the server, with timestamp errors within 50μs. The NTP time synchronization module provides a precise time reference for the entire system. During the multi-source data acquisition process, each sensor's data is accurately timestamped. This enables subsequent wavelet transform and adaptive Kalman filter modules to process data according to precise time sequence, ensuring both sequential and accurate data processing. For example, for electricity consumption data collected at different times, the wavelet transform module can effectively decompose and denoise the data within the appropriate time window based on accurate timestamps. The adaptive Kalman filter module can also accurately identify and correct abnormal data based on the accurate time sequence. The specific operation process is as follows: First, when the data fusion processing module starts, it initializes the NTP network time protocol configuration. The server enables a high-precision clock service, and each sensor connects to the network as a client. At regular intervals, the client sends a time synchronization request packet to the server, which carries the timestamp information of its current record. After receiving the request, the server generates a response packet containing a precise timestamp based on the accurate time recorded by its own high-precision clock and returns it to the client. After receiving the response packet, the client calculates the round-trip delay between the request packet's send time and the response packet's receive time, and corrects the local time by half of this round-trip delay. Through repeated iterations of this process, the client's time is continuously adjusted to ensure that the timestamp error remains stable within 50μs. To address outliers, the data is first decomposed into eight layers using a wavelet transform to remove high-frequency noise. Then, an adaptive Kalman filter dynamically adjusts the Q / R matrix to accurately identify and correct outliers. The collected multi-source data is first input into the wavelet transform module. This module processes the multi-source data using its time-frequency analysis characteristics. Specifically, it selects an appropriate wavelet basis and performs an eight-layer decomposition on the data. At each decomposition layer, the data is separated into a low-frequency approximation component and a high-frequency detail component. High-frequency noise is typically concentrated in the high-frequency detail component. Through layer-by-layer decomposition, the high-frequency noise component is gradually removed. The wavelet-transformed data is then input into the adaptive Kalman filter module. The adaptive Kalman filter module initially sets the initial values ​​for the process noise covariance matrix Q and the measurement noise covariance matrix R, and can dynamically adjust these parameters based on the actual data. For example, taking industrial equipment operating status data as an example, after initial wavelet transform processing of data such as the equipment's speed and load, it is input into the module. During the filtering process, the module calculates the error between the current measured value and the predicted value from the previous moment. Based on this error, it adaptively adjusts the Q and R matrices. For example, if the equipment's operating status suddenly changes and the measured value fluctuates significantly, the adaptive Kalman filter module increases the weight of the process noise covariance matrix Q, relying more on the current measured value to correct the predicted value, thereby accurately identifying anomalous data points.This iterative process enables accurate identification and correction of abnormal data, ensuring that the final output data accurately reflects the actual operating status of industrial equipment. Initially, the process noise covariance matrix Q and the measurement noise covariance matrix R are set. During the filtering process, the residual between the current measured value and the predicted value is calculated in real time. When the residual exceeds a preset threshold, the current data is considered abnormal. At this point, the Q and R matrices are dynamically adjusted based on the characteristics of the residual and the current system operating status. If data fluctuations increase abnormally, the weights of the corresponding state variables in the process noise covariance matrix Q are appropriately increased, allowing the filter to pay more attention to changes in the system state. Simultaneously, the measurement noise covariance matrix R is adjusted based on factors such as changes in the accuracy of the measuring equipment. Through continuous iterative calculations, accurate identification and correction of abnormal data are achieved. The wavelet transform module works closely with the adaptive Kalman filter module. The wavelet transform module performs preliminary data processing, removing high-frequency noise and providing relatively clean data input to the adaptive Kalman filter module. Based on this, the adaptive Kalman filter module further accurately identifies and corrects abnormal data. The data processed by these two modules provides a high-quality data foundation for the subsequent three-level feature fusion architecture. For example, when processing the operating status data of industrial equipment, the wavelet transform module first filters out interference noise, and then the adaptive Kalman filter module corrects abnormal fluctuations, so that the data input into the feature fusion architecture can more realistically reflect the equipment operation status and improve the accuracy of feature extraction.

[0038] The graph neural network-based feature extraction model was built using the PyTorchGeometric framework, with a hidden layer dimension of 256. The specific process involved collecting data from various devices within the park's energy system, including power equipment, thermal networks, and refrigeration units. This data was preprocessed using one-hot encoding for discrete data, such as the start and stop status of equipment. For continuous data, such as voltage and temperature, the Z-score method was used to normalize the data scale by adjusting the mean to 0 and the variance to 1.

[0039] In the PyTorchGeometric framework, energy devices are defined as nodes, energy flow relationships are defined as edges to construct an association graph, and edge weights are set according to the physical relationships of energy flow.

[0040] Next, a three-layer feature fusion architecture was constructed. The bottom layer's sliding window was set to 24 hours with a step size of 1 hour. Data within each window was extracted sequentially in chronological order. Within each window, LSTM units were used to process the data, capturing trends and patterns of change over time, and outputting a 128-dimensional time series feature vector. The middle layer used a graph convolutional network (GCN) to process the energy device association graph, convolving the graph's adjacency matrix with node features to explore coupling relationships between devices and generate a 128-dimensional coupling feature vector. The top layer's attention mechanism dynamically assigns weights to the feature vectors output by the bottom and middle layers based on the importance of the device and the diversity of real-time data features. For example, feature vectors of critical power equipment are given higher weights to highlight their importance in the energy system. Finally, the weighted feature vectors from the bottom and middle layers, after being weighted by the attention mechanism, are fused to output a 256-dimensional global feature vector, providing comprehensive and accurate feature information for subsequent energy system status assessment and optimized scheduling. This three-layer feature fusion architecture operates in tandem with other modules. The bottom-level LSTM units, middle-level GCN, and top-level attention mechanism each implement their respective functions based on processed data, supporting energy system assessment and scheduling, capturing data trends based on accurate time series. The energy device association graph processed by the middle-level GCN is constructed based on the device operational relationships reflected in pre-processed multi-source data, mining the coupling relationships between devices. The top-level attention mechanism dynamically assigns weights to the feature vectors output by the bottom and middle layers based on device importance and data feature differentiation. The determination of device importance also relies in part on the analysis of device operational data by the pre-processing module. The resulting fused feature vector provides key information for subsequent energy system status assessment and optimized scheduling. It is closely integrated with the overall system data processing flow, achieving a complete closed loop from raw data acquisition to final decision support.

[0041] The edge weight calculation for constructing the energy equipment association graph is based on the physical laws of energy flow: the edge weight from electricity to heat is assigned by the product of the electricity-to-heat conversion efficiency and the real-time power; the edge weight from heat to refrigeration is dynamically updated based on the heat exchange volume and the heat loss coefficient of the pipeline network.

[0042] Energy Status Assessment Module: This module constructs an equipment health evaluation index system covering 12 core indicators, including equipment efficiency, energy intensity, and maintenance cycle. In practical applications, data for these 12 core indicators must first be obtained from various data sources. For example, for equipment efficiency, the system regularly reads actual output power data from high-precision smart meters. It also retrieves equipment nameplate parameters and rated power data from the manufacturer's technical documentation from the equipment management database. The equipment efficiency index is calculated by real-time calculation of the ratio of actual output power to rated power. For energy intensity, the system connects with the production work order system to obtain information such as product type, batch, and corresponding production quantity. Then, based on the total energy consumption recorded by energy meters (such as gas and electricity meters), energy consumption is allocated according to specific allocation rules (such as the proportion of production hours), resulting in the energy consumption per unit product, or energy intensity index. The maintenance cycle index is determined by integrating historical equipment maintenance records from the database and querying the manufacturer's equipment maintenance manuals. This approach accurately captures data for each indicator, providing a foundation for subsequent equipment health assessments.

[0043] In the data processing phase, data from multiple sources, including vibration and temperature sensors, are fused using the DS evidence theory. Wavelet transforms are first used to reduce noise on the data collected by each sensor. For the multi-kilohertz time-domain signals output by the vibration sensors, an 8-layer decomposition using the db4 wavelet basis is performed, and soft threshold parameters are determined through cross-validation. For the temperature sensor data, a 5-layer decomposition using the sym8 wavelet basis is used, combined with hard thresholding to remove low-frequency noise. After noise reduction, features such as the root mean square (RMS) and spectral entropy of the vibration signal are calculated using a 100-point sliding window. The temperature data is trend-removed and the mean and variance are extracted. These parameters are then mapped to the fault mode space to construct a basic probability distribution function. When the degree of conflict between different pieces of evidence exceeds 0.7, a weighted average method is used to adjust the probability distribution, thereby accurately identifying potential equipment faults. This innovative application of the DS evidence theory to multi-source sensor data fusion is demonstrated in industrial equipment monitoring scenarios. Unlike traditional single-sensor fault detection methods, this fusion approach comprehensively considers data from multiple sensors, providing a more comprehensive picture of equipment operating status. The module introduces the concept of "device health entropy," continuously monitoring device operating data through a sliding time window to quantify the uncertainty of device state changes. Specifically, the system sets a sliding time window. Within each time window, it collects various device operating data, such as efficiency, temperature, and vibration. Based on this data, the device's state changes within that time window are calculated. For example, by calculating parameters such as the rate of change of device efficiency and the range of temperature fluctuation, and combining them with a specific entropy calculation model, these state change parameters are converted into a device health entropy value. The system monitors the rate of change of the device health entropy value in real time by calculating the ratio of the difference between the device health entropy values ​​in two adjacent time windows to the time interval. When this rate of change exceeds a pre-set threshold, the system immediately triggers an alert mechanism and sends an alert notification to the device operation and maintenance personnel. The notification includes information such as the device number, current health entropy value, and rate of change, allowing operators to promptly monitor the device status.

[0044] The Markov chain prediction model breaks down equipment status into 10 dimensions, including temperature, vibration, and pressure, with each dimension further divided into five status levels. Based on historical equipment operating data, the model first divides the time window into weekly windows, normalizes the data, and then applies Laplace smoothing. The model then trains the state transition probability matrix over 100 iterations, converging when the difference between adjacent iterations is less than 0.01. This model then predicts future trends in the equipment's status. In industrial equipment status prediction, the model innovatively breaks down equipment status into multiple dimensions, assigning multiple status levels to each dimension. The model then trains the state transition probability matrix over numerous iterations and finalizes the model with strict convergence conditions, ensuring more reliable prediction results. This customized modeling approach, tailored to the complex operating conditions of industrial equipment, differentiates from conventional general-purpose prediction models and better adapts to the dynamic state changes of equipment in industrial scenarios. The model incorporates a Bayesian network for fault tracing, constructing a directed acyclic graph consisting of equipment components, failure modes, and symptoms. Fault tree analysis identifies causal relationships between nodes, such as the association between motor bearing components and poor lubrication and abnormal vibration symptoms. A conditional probability table is trained using 200 historical fault samples. By propagating and updating evidence across the network, a joint tree algorithm is employed to optimize inference efficiency. When a vibration sensor detects an anomaly, node sorting is used to rapidly calculate the posterior probability of each faulty node, allowing the root cause of the equipment failure to be quickly identified. In the area of ​​fault tracing, a directed acyclic graph (DAG) consisting of equipment components, failure modes, and symptoms is constructed, combined with fault tree analysis to determine node causal relationships. This innovation represents a significant improvement over traditional fault analysis methods. This structured, graphical approach clearly illustrates the inherent connections between various factors within complex equipment systems. Furthermore, by utilizing a large number of historical fault samples to train conditional probability tables and employing a joint tree algorithm to optimize inference efficiency, the accuracy and speed of fault tracing are enhanced. This approach is particularly effective in scenarios where multiple industrial equipment failures occur simultaneously and the causes are complex. Compared to traditional troubleshooting methods, this method can more quickly and accurately identify the root cause, providing strong support for efficient operation and maintenance of industrial equipment.

[0045] Multi-level energy scheduling module: This module builds a three-level scheduling architecture at the device, workshop, and factory levels, utilizing a multi-objective optimization algorithm to balance energy costs, production efficiency, and carbon emissions. During device-level scheduling, data such as the device's energy consumption rate, operating time, and production task progress are collected in real time. Using this data, a dynamic programming algorithm is used to develop energy allocation plans for each device based on its priority and energy demand characteristics. For example, if a high-priority device is detected to be nearing energy depletion and impacting production continuity, energy is allocated to that device first, while the operating parameters of lower-priority devices are adjusted to balance energy usage.

[0046] Shop-level scheduling takes into account the combined energy demands of all equipment within the shop, the shop's production plan, and the upper limit of energy supply. Using integer programming algorithms, we determine the optimal energy allocation for each piece of equipment within energy supply constraints, with the goal of maximizing shop production efficiency. For example, if a shop has multiple production lines, energy allocation is optimized based on the urgency of each production line's production tasks and energy efficiency, maximizing overall shop production efficiency.

[0047] Factory-level scheduling is based on the multi-level energy collaborative scheduling algorithm of the improved NSGA-III, which takes energy cost, production efficiency and carbon emissions as multi-objective functions. Combined with the factory's long-term production plan, market energy price fluctuations and carbon emission index restrictions, the Pareto optimal solution set is found through iterative calculation (the number of iterations is set to 50 times). The most appropriate scheduling scheme is selected from it to achieve the optimal distribution of energy among different workshops and different equipment in the factory. For example, when energy prices are high and carbon emission indicators are tight, clean energy equipment is scheduled first to reduce the use of high-cost, high-carbon emission energy; when production tasks are urgent, energy cost restrictions are appropriately relaxed to ensure production efficiency; and reinforcement learning strategy networks (action space dimension 20) are deployed to dynamically adjust scheduling strategies to achieve the optimal configuration of energy resources. A multi-level energy collaborative scheduling algorithm based on the improved NSGA-Ⅲ is proposed, and the relevant formulas are as follows: The objective function is Among them, f(x) is a multi-objective function vector, consisting of m objective functions Composition, x is the decision variable.

[0048] The inequality constraints are here is an inequality constraint, i ranges from 1 to p, indicating that there are p inequality constraints.

[0049] The equality constraints are in is an equality constraint, and j ranges from 1 to q, indicating the existence of q equality constraints. In terms of algorithm design, the non-dominated sorting mechanism of the NSGA-III algorithm was optimized to address the complex and dynamic nature of multiple objectives in industrial energy scheduling scenarios. When processing millions of decision variables, the traditional NSGA-III algorithm has a large amount of non-dominated sorting computation and low efficiency. The improved algorithm introduces an adaptive weight allocation strategy, which dynamically adjusts the sorting weights based on the importance of different objectives such as energy costs, production efficiency, and carbon emissions under different working conditions. This significantly reduces the amount of computation and effectively improves the algorithm's operational efficiency under large-scale decision variables.

[0050] This algorithm works closely with other components of the multi-level energy scheduling module. The "dynamic constraint relaxation" mechanism works in conjunction with the improved NSGA-III algorithm. Under extreme operating conditions, the algorithm can optimize energy allocation in real time based on the constraint boundaries adjusted by the "dynamic constraint relaxation" mechanism. The distributed computing architecture (16 nodes) provides powerful computing power for the improved NSGA-III algorithm, enabling it to rapidly process millions of decision variables and achieve real-time optimal allocation of energy resources. Furthermore, the optimized scheduling solution output by the algorithm also provides a reference for the deployed reinforcement learning policy network, assisting it in dynamically adjusting its scheduling strategy. Together, they promote the optimal allocation of energy resources, forming an integrated whole with the entire multi-level energy scheduling module and improving the overall operational efficiency of the energy system.

[0051] Energy Performance Analysis Module: This module develops an energy flow diagram analysis tool that supports visualization of eight energy media, including electricity, steam, and natural gas. This tool uses a layered coloring scheme, with blue (#1E88E5) for electricity, orange (#FF9800) for steam, and green (#4CAF50) for natural gas. Saturation is adjusted using the HSL color model to distinguish different pressure levels. Arrow widths are divided into five levels based on flow rate: 1px for flows below 100kW and 5px for flows above 1000kW. Dynamic scaling is achieved using the D3.js path interpolation algorithm. At the data processing level, cubic spline interpolation is used to upsample steam data to account for the differences in sampling frequency (1 minute for electricity, 5 minutes for steam, and 15 minutes for natural gas). Linear interpolation is used to fill gaps in natural gas data. Timestamps are aligned, and missing values ​​are processed using a weighted average of the preceding and following points. Ultimately, the diagram is unified to a minute-level granularity, ensuring the real-time and accuracy of the energy flow diagram.

[0052] The "energy footprint" analysis method constructs an energy consumption tree structure. It first extracts total energy consumption data from the ERP system. Energy consumption is then allocated to each workshop based on the proportion of work hours according to the workshop schedule. Production line operating hours are then obtained from the MES system and broken down into specific equipment using the ratio of equipment power × operating time. For electrophoresis equipment, for example, the heating system's energy consumption is linked to real-time current and voltage data (sampling frequency 10Hz). Edge computing nodes calculate instantaneous power and accumulate it to form daily power consumption. Data envelopment analysis (DEA) is used to evaluate the energy efficiency of 50 decision-making units (DMUs). Energy inputs for the seven selected indicators are converted to standard coal (0.1229 kgce / kWh for electricity and 0.1314 kgce / kg for steam). Human resource input is calculated as the average number of workers per day × work hours in the workshop, and output is calculated as product yield × output. When calculating efficiency using the CCR model, a slack variable tolerance of 0.001 is set. When the efficiency value is less than 1, projection analysis is used to determine input redundancy, thus constructing a multi-input, multi-output efficiency evaluation model. When combining the Shapley value method for cost allocation, the marginal contribution of each device to total energy consumption is considered to ensure fair cost distribution. Ultimately, 12 high-energy consumption bottlenecks, such as the electrophoresis equipment heating system, were accurately identified. The machine learning-based energy consumption prediction model uses the LightGBM algorithm and performs rolling window training on historical data. The model inputs include 15 features, including production plans, weather data, and equipment status. Key variables are automatically selected by feature importance ranking.

[0053] The energy management dashboard utilizes the "Energy Performance Cube" visualization framework, achieving 3D rendering based on WebGL technology. Energy efficiency, cost, and carbon emissions are mapped to the X, Y, and Z axes, respectively. Data points are displayed as cube icons (5-15px side length, scaled according to the indicator value). Colors are mapped using the HSV model (green = H = 120° on target, yellow = H = 60° warning, and red = H = 0° exceeding the target). Drag-and-drop operations are implemented using the mouse wheel and drag event listeners. When switching between workshops, the backend queries the InfluxDB database in real time, aggregates minute-level data for the selected time period (default is the last 30 days) into hourly averages, and pushes these averages to the frontend via WebSocket to update the view. The system has a built-in anomaly detection algorithm that calculates the mean μ and standard deviation σ based on the past 14 days of data, with a threshold of μ ± 3σ. When real-time data exceeds the threshold, an alert message is sent via WebSocket to the manager's mobile device, containing the abnormal indicator, time point, and deviation magnitude. The abnormal point is also highlighted in the corresponding location on the dashboard. Through multi-dimensional analysis and visualization, this module provides enterprises with a comprehensive energy performance assessment tool, helping them achieve energy conservation and consumption reduction goals.

[0054] Decision Support Module: The expert knowledge base built by the Decision Support Module systematically organizes the knowledge system in the energy field and contains 5,000 structured rules covering multiple key areas such as equipment fault diagnosis, energy scheduling strategies, and energy efficiency optimization solutions. These rules are expressed using a production rule method and stored in a standard "IF (condition) - THEN (conclusion)" format, facilitating rapid retrieval and updating, providing a solid knowledge foundation for system decision-making.

[0055] The Bayesian network-based fault diagnosis system uses a hierarchical modeling approach to construct the network structure. Taking a motor system as an example, the top layer is the "motor fault" root node, the middle layer contains "electrical fault" and "mechanical fault" subnodes, and the bottom layer contains specific fault mode nodes such as "winding short circuit" and "bearing wear." Symptom nodes include "abnormal current" and "excessive vibration." During training, 2,000 motor fault records were collected from an industrial park over a three-year period, including 1,500 normal operation data and 500 fault data. The training and test sets were divided into an 8:2 ratio. Using the maximum likelihood estimation method, each conditional probability table was optimized 100 times, stopping when the change in probability between successive iterations was less than 0.001. Ultimately, the accuracy of the conditional probability table for bearing wear faults was controlled to 0.008. During real-time diagnosis, if the vibration sensor data exceeds a threshold (10 mm / s) and the current exceeds 110% of the rated value, the system calculates the posterior probability of bearing wear to 0.87 using Bayes' theorem. If the probability exceeds the threshold of 0.7, an early warning is triggered, enabling rapid and accurate diagnosis of equipment faults.

[0056] The energy management knowledge graph uses the Neo4j graph database as its storage medium, constructing a complex network consisting of 10,000 entities (such as energy equipment, process flows, and management specifications) and 30 types of relationships. During graph construction, natural language processing techniques are used to extract entities and identify relationships from textual materials in the energy field. A graph embedding algorithm is then used to convert the knowledge graph into a vector representation, facilitating computation and reasoning. The system employs a "dual-engine" reasoning mechanism, integrating symbolic and connectionist reasoning. The symbolic reasoning engine, based on rules from an expert knowledge base, employs a reverse reasoning strategy for handling transformer oil overtemperature issues: first assuming a "transformer winding overheating" fault, it then verifies that conditions such as "oil temperature exceeding 95°C" and "load factor >80%" are met. The system matches rules at a speed of up to 100 entries per second. The connectionist reasoning engine uses an LSTM-Attention model to handle power outage scenarios. Training uses 5,000 historical emergency repair cases. Input features include 12-dimensional data such as fault time, weather, and equipment parameters. An 8-layer LSTM network extracts temporal features, and the attention mechanism focuses on key features. The model achieves a solution generation accuracy of 0.89% on the test set. When a sudden power outage occurs in a workshop, the two engines work together: the symbolic engine first uses rules to determine whether the circuit breaker has tripped, while the connectionist engine simultaneously retrieves similar power outage scenarios from the case library and generates a comprehensive solution, including inspecting the transformer windings and switching to a backup power source. Response time is kept within 30 seconds, significantly reducing response time for complex issues and providing efficient and reliable support for enterprise energy management decisions.

[0057] In the present invention, the multi-source data acquisition module also includes: configuring high-precision gas sensors to monitor greenhouse gas emissions, with a detection accuracy of ±0.5ppm, which can capture subtle changes in gas concentrations in real time. Deploy a drone inspection system to obtain status data of plant energy facilities, with an inspection cycle of 7 days / time, and use the onboard infrared thermal imager and high-definition camera to scan the equipment's operating status in all directions. Integrate blockchain technology to build a distributed data storage architecture, use hash algorithms and consensus mechanisms to ensure that data cannot be tampered with, and ensure data integrity from the source. Develop edge computing nodes to implement local preprocessing and feature extraction of data, and use embedded processors to perform filtering, noise reduction, feature dimensionality reduction and other operations on the collected data. This module innovatively designs a "data twin" mechanism, builds a lightweight data model at the edge, and realizes data interaction and collaborative processing between the cloud and the edge through message queues and real-time synchronization protocols, significantly reducing data transmission delays and cloud computing pressure.

[0058] In the present invention, the data fusion processing module also includes: developing a multimodal data fusion algorithm based on the attention mechanism, setting up 8 heads to achieve parallel extraction of multi-dimensional data features, strengthening the weight of key information through self-attention calculation, and effectively integrating heterogeneous data such as electricity consumption, equipment operation, and environmental monitoring. The constructed time series anomaly detection model uses a long short-term memory network (LSTM) combined with a Transformer architecture. After training and optimization with a large amount of labeled data, the F1 score exceeds 0.95 and can accurately identify abnormal fluctuations in energy data. The deployed federated learning framework uses layered encryption and secure aggregation protocols to support 100 participants to collaborate on modeling without leaving the local data, and further enhances data privacy protection through differential privacy technology.

[0059] This module innovatively designs a "knowledge distillation under privacy protection" mechanism, which adopts a "cloud-based teacher model + local student model" architecture: the teacher model is trained using desensitized data from 100 companies and contains a 12-layer Transformer structure; the student model is a lightweight LSTM network deployed on the company's local edge node. During knowledge transfer, the cloud transmits the intermediate layer output of the teacher model using homomorphic encryption. After receiving it, the local student model decrypts it and calculates the distillation loss. The loss function is the KL divergence between the predicted probabilities of the teacher model and the student model. At the same time, gradient compression technology is used to sparsely process the uploaded gradient parameters, retaining only 30% of the critical gradients, with a compression ratio of 3:1. Taking the motor fault diagnosis task as an example, after 10 rounds of knowledge distillation, the student model's fault recognition accuracy is improved, and the company's original data does not leave the local server throughout the entire process.

[0060] In this invention, the energy status assessment module also includes: establishing a digital twin model of energy equipment. Through high-precision three-dimensional modeling and real-time data-driven operation, a 1:1 replica of the equipment's operating status is created in a virtual space, enabling visual management of the equipment's entire lifecycle. A convolutional neural network-based abnormal sound recognition system for equipment is developed. This system uses mel-spectrograms to convert sound signals into image features. Key features are extracted through multi-layer convolution and pooling operations, enabling rapid identification of abnormal operating conditions such as equipment wear and looseness. A fiber-optic distributed temperature sensing system is deployed, leveraging Raman scattering to achieve distributed, long-distance, real-time temperature monitoring of key equipment. Temperature measurement accuracy reaches ±0.5°C, effectively preventing the risk of equipment overheating. An energy system vulnerability assessment model is constructed, encompassing 20 evaluation indicators, including equipment reliability, network topology, and operating conditions. The analytic hierarchy process is used to determine indicator weights, combined with a fuzzy comprehensive evaluation method to quantify the system's vulnerability level. The 20 evaluation indicators in the energy system vulnerability assessment model cover dimensions such as equipment reliability (e.g., mean time between failures, failure rate), network topology (e.g., node degree centrality, line load factor), and operating conditions. When using the hierarchical analysis method to determine the weights, a 1-9 level judgment matrix between indicators is first constructed, and the weight vector is calculated by the eigenvalue method. The final weight is determined after the consistency test (CI<0.1). In the fuzzy comprehensive evaluation method, the vulnerability level is divided into four levels: "safe", "mildly vulnerable", "moderately vulnerable", and "severely vulnerable". The vulnerability membership of each indicator is quantified by triangular fuzzy quantification, and finally the comprehensive vulnerability index of the system is calculated by the weighted average operator. For parameters that are difficult to measure directly, a basic model is first constructed based on physical principles such as thermodynamics and kinetics, and then the historical operation data is trained and optimized through machine learning algorithms. The model parameters are dynamically corrected to achieve accurate estimation of key parameters such as internal stress of the equipment and fluid flow rate, significantly improving the comprehensiveness and accuracy of equipment status assessment.

[0061] In the present invention, the multi-level energy scheduling module also includes: developing a short-term energy scheduling algorithm based on model predictive control, setting a 24-hour prediction time domain, combining weather forecasts, historical load data and equipment operating status, and using a rolling optimization strategy to dynamically generate an energy scheduling plan. The constructed demand response management system interacts with the user terminal in real time through smart meters to achieve precise regulation and response of demand-side loads. The deployed distributed energy collaborative control platform adopts a microservice architecture, supporting access to distributed photovoltaic, wind power generation, energy storage equipment and other energy resources with 500 access points. Through a unified data interface and communication protocol, it realizes the coordinated operation and power optimization distribution of various energy equipment, and promotes the efficient consumption of renewable energy and the stable interaction of the power grid.

[0062] This module aggregates and manages distributed flexible loads by tapping into adjustable load resources such as industrial equipment, commercial buildings, and residential users. Using cluster analysis and load characteristic modeling techniques, it categorizes and integrates different load types to construct a virtual energy storage system. This system can rapidly respond to ancillary service instructions such as peak shaving and frequency regulation based on the grid's real-time needs. By flexibly adjusting loads, it simulates the charging and discharging characteristics of energy storage, enhancing the flexibility and reliability of the energy system and achieving efficient coordinated operation of the power grid, load, and storage.

[0063] The present invention also includes: an energy carbon footprint accounting module, which strictly follows the ISO14040 standard, takes the life cycle assessment (LCA) as the core framework, and adopts an improved input-output method to construct an accounting model. In specific operations, the production activities of the enterprise are first disassembled and divided into 128 basic units. Each unit is set with a unique identification code, and the ERP system, MES system and supply chain management platform are connected through the interface to collect data such as energy consumption, material input and output, carbon emission factors, etc. of each unit to construct an input-output matrix containing 5000+ data nodes. Each node in the matrix corresponds to the material or energy flow relationship between units, and the transmission relationship between energy consumption and carbon emissions in each link is quantified through matrix operations. The dynamic update mechanism is triggered by a scheduled task (3 am every day), automatically capturing real-time supply chain data (such as carbon emission data of raw materials of suppliers), and storing it on the chain after verification by the blockchain node to ensure that the data cannot be tampered with, solving the data lag problem of traditional methods.

[0064] During the accounting process, a multi-level analysis system was constructed. At the macro level, the industrial park was considered as a holistic carbon system, analyzing the overall balance of carbon input, conversion, and output. At the meso level, this was broken down to workshops and production lines, where carbon flow topologies were mapped to clearly define the carbon flow paths within each production unit. At the micro level, key equipment was focused on, with sensors collecting real-time energy consumption data. Combined with material balance models, this accurately calculated the carbon emission intensity per unit product. A library of over 200 evaluation indicators was established for the entire life cycle, encompassing aspects such as raw material procurement, production and processing, and product transportation, covering dimensions such as fossil energy consumption, process carbon emissions, and transportation losses. The developed carbon emission reduction path optimization algorithm utilizes a multi-objective genetic algorithm (MOGA) architecture. At startup, the algorithm first encodes 10 technical routes (such as waste heat recovery and photovoltaic substitution) into genetic sequences. Fifty candidate solutions were initialized and evaluated using a fitness function. The energy cost target calculated the equipment procurement and operation and maintenance costs of the solution implementation; the emission reduction effectiveness target calculated the emission reduction amount based on baseline carbon emissions; and the technical feasibility target evaluated constraints such as site and funding through expert scoring (1-10). To assess technical feasibility, an evaluation team comprised of experienced industry engineers, energy management experts, and financial specialists was assembled. Ten technical approaches were comprehensively evaluated based on factors such as site suitability, reasonable capital investment, and technological maturity, drawing on actual project experience and industry standards. For example, for waste heat recovery technology, the team conducted an on-site inspection of the park's existing site layout and, based on the equipment installation dimensions, determined whether the site met the requirements for equipment installation and subsequent operation and maintenance. From a financial perspective, the team assessed the rationality of the capital investment-output ratio based on equipment procurement costs, expected operating and maintenance expenses, and potential returns. Regarding technological maturity, the team considered indicators such as the number of application cases and operational stability under similar operating conditions in the same industry, providing a comprehensive assessment rather than relying solely on expert scores. After 300 generations of evolution, combined with Monte Carlo simulation (1,000 random scenario tests), the three most stable solutions were selected. Monte Carlo simulation was employed during multiple rounds of evolutionary selection. Multiple random scenario simulation tests were conducted to account for uncertainties in the park's energy system operation, such as energy price fluctuations and production load variations. During the simulation process, different variable combinations such as energy price ranges and production task volume fluctuation ranges are set to simulate various complex situations that may arise in actual operation.

[0065] The module also features a visual decision-making platform, showcasing the park's carbon emission hotspots via a three-dimensional carbon footprint map, allowing users to drill down to view detailed data for each link. The built-in carbon accounting accuracy verification module utilizes cross-validation and error compensation algorithms. The cross-validation method involves dividing the park's 2023 carbon emission data into four quarterly groups. Three groups of data are used to train an LSTM error compensation model (with energy consumption data as input and accounting error as output). The fourth group of data is then used to make prediction corrections. Model parameters include two hidden layers (128 neurons each), a learning rate of 0.01, and 200 training epochs, meeting ISO14040 standards. This minimizes carbon footprint accounting errors, providing a scientific basis for enterprises to set carbon reduction targets and participate in carbon trading markets, thereby contributing to the realization of a green and low-carbon transition.

[0066] The present invention also includes an energy security management module and a deep learning-based intrusion detection system (IDS) deployed at the network layer. This system uses a convolutional neural network (CNN) to extract network traffic features and combines it with a long short-term memory (LSTM) network to analyze traffic temporal changes and identify abnormal network behavior in real time. The intrusion prevention system (IPS) utilizes a dynamic rule base containing over 5,000 security policies, supporting automatic updates and policy conflict detection, enabling real-time interception of malicious traffic.

[0067] Defense model training combines simulated attacks with real-world drills. A virtual simulation environment is constructed to simulate over 20 typical attack scenarios, including DDoS attacks, APT attacks, and data theft, collecting over 100,000 attack sample data points. Generative adversarial networks (GANs) are used to generate a diverse set of unknown attack samples for enhanced defense model training. The model employs a transfer learning strategy, transferring weight parameters pre-trained on public datasets to specific energy industry scenarios for fine-tuning, effectively improving its detection capabilities against unknown attacks. The energy equipment fault emergency response model, based on an event-driven architecture, establishes an emergency knowledge base encompassing over 300 fault scenarios. When a device fails, the system integrates sensor data, equipment operation logs, environmental monitoring data, and other information through multi-source data fusion technology. Using a Bayesian network, the system quickly diagnoses the fault type and impact scope. In conjunction with pre-defined emergency response plans, the emergency response process is initiated within 10 minutes, automatically adjusting energy scheduling strategies.

[0068] The energy system's disaster recovery system utilizes a remote active-active architecture, establishing a high-speed data transmission channel between the primary data center and the disaster recovery center. Real-time data replication ensures data consistency. The disaster recovery center deploys a complete replica of the energy management system, supporting both automatic failover and manual takeover. The system also includes a built-in disaster recovery drill module that regularly simulates major disaster scenarios such as fires and earthquakes for disaster recovery and recovery drills. In the event of a major failure, automated orchestration tools enable system recovery and service failover within 45 minutes, minimizing the impact of the failure on energy supply and ensuring the safe and stable operation of the energy system.

[0069] The present invention also includes an energy knowledge graph module. A web crawler system is deployed on the data collection end. The web crawler is implemented using the Python Scrapy framework. Crawling rules are set for the National Energy Administration's official website (e.g., only PDF files under the "Policy Documents" section are captured). IEEE Xplore papers are batch-accessed using an API (a scheduled task is triggered every Monday at 3:00 AM). Crawled unstructured text (e.g., PDF policy documents) is parsed into text format using the PyPDF2 library and stored in a distributed file system along with Excel-formatted equipment parameter tables from the internal ERP system and JSON-formatted maintenance records from the MES system. During the cleaning phase, the spaCyNLP library loads the pretrained "en_core_web_sm" model to remove special characters such as '\n' and '\t', as well as garbled text. The TextBlob library performs word segmentation on the cleaned text (e.g., splitting "high-voltage motor maintenance" into "high voltage," "motor," and "maintenance") and annotates parts of speech (noun, verb, etc.) to facilitate subsequent entity recognition. A named entity recognition model based on BERT-CRF was used to extract entities. The model was deployed on a GPU server (NVIDIA A100). During training, 100,000 annotated data items (with entity types such as "equipment," "policy," and "fault type") were split into a training set and a validation set with an 8:2 ratio. The Adam optimizer (learning rate 5e-5) was used for 30 epochs, and training was terminated when the F1 score stabilized at 0.92. During actual extraction, a piece of equipment manual text (e.g., "Centrifugal pumps should have their lubricant changed every 300 hours") was input. The model automatically identified entities such as "centrifugal pump" (equipment entity) and "300 hours" (time entity). During the relationship extraction phase, the GraphSAGE model took entity co-occurrence windows (e.g., the co-occurrence of "motor" and "bearing" within five consecutive sentences) as input. Through neighbor node sampling and feature aggregation, it output relationship types such as "motor-contains-bearing."

[0070] The knowledge update process utilizes an incremental learning strategy. After new data is verified by the quality assessment module, atomic-level updates to the knowledge graph are achieved through the graph database's transaction management mechanism. Version control is introduced to record the time, content, and operator of each update, supporting backtracking of historical versions of the knowledge graph. An expert review process is also implemented, allowing enterprise energy experts to manually verify automatically updated knowledge through a visual interface to ensure accuracy.

[0071] This knowledge-based reasoning engine, based on a graph neural network and employing a graph attention network (GAT) architecture, processes equipment selection. For example, a user enters "Production requirements: Daily steel production 1,000 tons; Energy consumption standard: ≤500 kWh / ton; Budget: 5 million yuan" on the system interface. The engine converts the input conditions into vectors and matches them to entities such as "steelmaking equipment" in the knowledge graph. Using GAT's multi-layer attention mechanism, the engine calculates the degree of match between the entities and the conditions (e.g., "arc furnace" with a weight of 0.8 in the energy consumption dimension, and "converter" with a weight of 0.7 in the cost dimension). The engine then outputs the three highest-scoring solutions and their reasons for recommendation. For energy-saving retrofit scenarios, a reinforcement learning algorithm uses "annual energy savings after retrofit" as a reward value and iteratively optimizes action spaces such as "photovoltaic panel installation area." It simulates energy savings under different light intensities 1,000 times and generates a payback period forecast (e.g., "Initial investment 2 million yuan, annual electricity savings 500,000 yuan, payback period 4 years"). This multi-layer attention mechanism captures the complex relationships between entities in the knowledge graph. When solving equipment selection problems, the engine uses production requirements, energy consumption standards, and budget constraints as input criteria, searches the knowledge graph for related entities and relationship paths, and through node feature aggregation and message passing algorithms, calculates the matching score for each candidate device and outputs the optimal selection solution. For energy-saving retrofit scenarios, the inference engine combines historical project case data with reinforcement learning algorithms to simulate and deduces different retrofit options, evaluating energy savings and return on investment to inform decision-making. The intelligent question-answering system is based on a pre-trained BERT model and fine-tuned for domain adaptation using the company's internal energy knowledge corpus. A dedicated dataset of over 200,000 question-answer pairs was constructed, covering common questions such as equipment operation, troubleshooting, and policy interpretation. The system utilizes a hybrid architecture combining retrieval and generation. It first uses an inverted index to quickly retrieve relevant knowledge documents, then uses a fine-tuned BERT model to generate accurate answers. It supports multi-turn conversations and uses a memory mechanism to record contextual information and understand complex user questions. The system also includes an answer quality assessment module that scores generated answers based on accuracy, completeness, and comprehensibility to ensure reliability. The entire module operation process is as follows: collect internal and external data regularly every week → clean and extract entities / relationships → incrementally update the knowledge graph (including expert review) → the inference engine and question-answering system call the graph data to respond to business needs (such as equipment selection and fault consultation).

[0072] This invention also includes an energy ecosystem collaboration module that utilizes the PBFT (Practical Byzantine Fault Tolerance) consensus algorithm, ensuring data consistency while reducing block confirmation times to seconds. The platform encrypts data throughout its lifecycle using the nationally recognized SM2 / SM3 / SM4 algorithm system. Data transmission utilizes the TLS 1.3 protocol, and storage utilizes homomorphic encryption technology. This ensures the security and privacy of 50 participating companies sharing electricity usage data, equipment parameters, production plans, and other information through APIs. The blockchain smart contract system, developed in Solidity, includes three core contracts: energy trading, value distribution, and settlement and liquidation. Through sharding and off-chain computing optimization, the system supports high-concurrency transaction processing at 10,000 transactions per second. The energy trading contract incorporates a dynamic matching algorithm. When a company generates a surplus of photovoltaic power, the system automatically scans the demand pool and quickly locates the demander based on a dual-priority matching strategy based on price and distance. The value distribution contract automatically calculates the profit share of each party based on their contribution to energy production, transmission, and consumption, using the Shapley value method combined with blockchain-stored evidence data. The settlement and clearing contract is connected to the commercial banking system, triggering cross-border blockchain payments through smart contracts to achieve fund settlement in seconds.

[0073] The platform also features a visual energy collaborative management interface, allowing businesses to view energy supply and demand dynamics, transaction records, and revenue details in real time. The built-in collaborative optimization model, based on multi-agent game theory, comprehensively considers factors such as energy price fluctuations, policy subsidies, and carbon emission costs, generating optimal collaborative solutions through Monte Carlo simulation. Through data sharing and automated execution of smart contracts, this effectively reduces trust costs and transaction friction among businesses, optimizes resource allocation across the energy supply chain, and promotes the coordinated development and mutually beneficial outcomes of regional energy ecosystems.

[0074] The above are only preferred specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with this technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solutions and inventive concepts of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A multi-level energy management system based on multi-dimensional data, characterized in that: Includes the following modules: Multi-source data acquisition module: Configure electricity meters to collect electricity consumption data, deploy IoT sensors to monitor environmental parameters, integrate industrial controllers to obtain equipment operating status data, access meteorological data APIs, and build a multi-source heterogeneous data acquisition system; Data fusion processing module: This module uses a time synchronization algorithm to align the spatiotemporal data of multiple sources, processes outliers through a combination of wavelet transform and Kalman filtering, builds a time series feature extraction model based on graph neural networks, and designs a three-level feature fusion architecture. Energy Status Assessment Module: Establishes an energy equipment health evaluation index system, develops a multi-source information fusion algorithm based on DS evidence theory, constructs a Markov chain prediction model to predict the probability of energy equipment failure, quantifies equipment uncertainty by calculating the state information entropy value, and implements fault tracing by combining Bayesian networks; Multi-level energy scheduling module: Construct a three-level scheduling architecture, use a multi-objective optimization algorithm to balance energy costs, production efficiency and carbon emissions, and propose a multi-level energy collaborative scheduling algorithm based on the improved NSGA-III: the objective function is , f(x) is a multi-objective function vector, consisting of m objective functions Composition, x is the decision variable, and the inequality constraint is is an inequality constraint, i ranges from 1 to p, indicating that there are p inequality constraints, and the equality constraints are ; Energy Performance Analysis Module: Develop an energy flow analysis tool, build an energy efficiency evaluation model based on data envelopment analysis, and establish an energy cost decomposition model. By tracking the energy conversion path in the production process, the energy efficiency loss in each link is quantified, and the Shapley value method is used to achieve fair cost allocation. Decision support module: Configure expert knowledge base, build energy management knowledge graph, design symbolic reasoning engine to process deterministic knowledge, design connectionist reasoning engine to process uncertain knowledge, and realize cross-domain knowledge transfer through knowledge graph embedding technology; The multi-source data acquisition module further includes: using an adaptive sampling algorithm based on variational mode decomposition, and the objective function is: ; The constraints are: , K is the number of modes; is the modal function, is the center frequency; is the Dirac delta function; represents the partial derivative with respect to time t; j is an imaginary unit; f(t) is the original signal. This module configures gas sensors to monitor greenhouse gas emissions, deploys a drone inspection system to obtain plant energy facility status data, integrates blockchain technology to ensure data is tamper-proof, and develops edge computing nodes to implement local data preprocessing and feature extraction. The energy status assessment module also includes: establishing a digital twin model of energy equipment, developing an abnormal sound recognition system for equipment based on convolutional neural networks, deploying a fiber optic distributed temperature sensing system to monitor the temperature of key equipment in real time, building an energy system vulnerability assessment model, and performing parameter estimation through the fusion of physical models and data-driven models.

2. A multi-level energy management system based on multi-dimensional data according to claim 1, characterized in that: The data fusion processing module also includes: developing a multimodal data fusion algorithm based on the attention mechanism, building a time series anomaly detection model, and deploying a federated learning framework to protect data privacy; this module realizes the transfer of energy management knowledge between enterprises by constructing a teacher-student model architecture without leaking the original data.

3. The multi-level energy management system based on multi-dimensional data according to claim 1, characterized in that: The multi-level energy scheduling module also includes: developing a short-term energy scheduling algorithm based on model predictive control, building a demand response management system, deploying a distributed energy collaborative control platform, realizing the interaction between renewable energy consumption and power grid, and building a virtual energy storage system to participate in power grid auxiliary services by aggregating adjustable load resources.

4. The multi-level energy management system based on multi-dimensional data according to claim 1, characterized in that: The energy performance analysis module also includes: developing an energy consumption prediction model based on machine learning, building an energy efficiency benchmarking analysis tool, deploying an energy management dashboard, realizing the visualization and dynamic monitoring of energy performance, displaying three-dimensional indicators in three-dimensional space, and combining interactive drill-down functions to support in-depth data analysis.

5. The multi-level energy management system based on multi-dimensional data according to claim 1, characterized in that: Also includes: The energy carbon footprint accounting module uses the life cycle assessment method to construct a carbon footprint accounting model, develops a carbon emission reduction path optimization algorithm, deploys a blockchain carbon trading platform, realizes the accounting and trading of carbon emissions, and quantitatively analyzes the flow path and conversion efficiency of carbon in the energy system by constructing an energy-carbon flow coupling model.

6. The multi-level energy management system based on multi-dimensional data according to claim 1, characterized in that: Also includes: The energy security management module develops an energy system protection system based on network security situational awareness, builds an energy equipment failure emergency response model, deploys an energy system disaster recovery system, designs an "active immune" security architecture, and trains defense models through simulated attack scenarios to achieve early warning and defense against unknown threats.

7. The multi-level energy management system based on multi-dimensional data according to claim 1, characterized in that: Also includes: The energy knowledge graph module builds a knowledge graph that includes energy equipment, process flows, and management specification entities, develops a knowledge reasoning engine based on graph neural networks, deploys a question-answering system, and automatically updates the knowledge graph structure and parameters by continuously learning new data and expert experience.

8. The multi-level energy management system based on multi-dimensional data according to claim 1, characterized in that: Also includes: The energy ecosystem collaboration module develops an energy data sharing platform, builds a cross-enterprise energy collaborative optimization model, and deploys a blockchain contract system to achieve collaborative management and value co-creation across the upstream and downstream of the energy industry chain; automatically executes energy transactions and value distribution through blockchain contracts, and builds a win-win energy ecosystem for all parties.

Images