Intelligent micro-grid energy management system
By constructing a four-layer closed-loop architecture for a smart microgrid energy management system, and combining digital twins and deep reinforcement learning, the problems of fragmented and error-prone scheduling strategies in microgrids are solved. This enables efficient and accurate multi-timescale scheduling and deep load-side perception, thereby improving the operational economy and power supply reliability of microgrids.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- BEIJING GUOKE PRIME TECHNOLOGY CO LTD
- Filing Date
- 2026-05-07
- Publication Date
- 2026-06-05
AI Technical Summary
In existing microgrid energy management systems, the disconnect between digital twins and optimization decision-making, the coarse layering of multi-timescale scheduling, the accumulation of errors in the electric-hydrogen coupling model, and the lack of deep perception and decoupling on the load side lead to low accuracy and efficiency of scheduling strategies.
A smart microgrid energy management system is constructed, adopting a four-layer closed-loop architecture: equipment and sensing execution layer, edge computing and data preprocessing layer, cloud digital twin and intelligent decision-making layer, and application and interaction layer. Combining digital twin endogenous simulation and hierarchical deep reinforcement learning optimization, it realizes early warning-driven decision-making and real-time correction of deviation feedback. Through multi-timescale collaboration of edge-side photovoltaic power correction and cloud scheduling, it dynamically adapts to equipment aging and temperature changes. A non-intrusive load decoupling and electro-hydrogen coupling model is adopted to improve the predictability and accuracy of scheduling strategies.
It significantly improves the operational economy, power supply reliability, and carbon emission reduction benefits of microgrids. Through multi-timescale coordinated scheduling, it matches renewable energy fluctuations and equipment impacts, reduces operational losses, and achieves fine load classification and user privacy protection.
Smart Images
Figure CN122159387A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of microgrid technology, and more specifically to a smart microgrid energy management system. Background Technology
[0002] As the penetration rate of distributed renewable energy continues to increase, the role of industrial park microgrids in improving energy efficiency and power supply reliability is becoming increasingly prominent.
[0003] Currently, there are various energy management technology solutions. One type of solution focuses on day-ahead optimization scheduling. For example, the energy management method based on robust optimization transforms the uncertain variables of photovoltaic output and load into deterministic linear constraints to establish a day-ahead scheduling model. Another type of solution introduces digital twin technology. For example, the distributed energy management system for zero-carbon parks uses model predictive control to achieve source-grid-load-storage optimization scheduling and uses a digital twin platform to assist in fault diagnosis.
[0004] Some technical solutions combine multi-energy coordinated control with virtual synchronous generator technology to build a three-level virtual inertia support architecture with energy storage at the second level, flexible load at the minute level, and controllable power supply at the hour level. For example, a company uses fuzzy algorithms to dynamically identify priority targets in the photovoltaic-storage microgrid scenario, uses digital twins to update synchronously, and combines federated learning and cooperative game theory to generate control schemes.
[0005] In addition, technologies such as multi-agent reinforcement learning, game theory, and blockchain have also entered the field of microgrid scheduling. For example, rural microgrid clusters use spatiotemporal twin CNN-LSTM models combined with multi-agent reinforcement learning to achieve dynamic optimization, integrated energy microgrid clusters use multi-agent hierarchical game collaboration for scheduling, and hydrogen energy storage units are incorporated into the multi-objective optimization scheduling of microgrids.
[0006] However, in existing energy management solutions, digital twins mostly play an auxiliary role in state mirroring or fault diagnosis, lacking real-time two-way closed-loop interaction with optimization decisions; scheduling architectures typically only distinguish between day-ahead and real-time levels, making it difficult to accurately match different time scale characteristics such as minute-level fluctuations in photovoltaics, second-level impacts on high-power equipment, and safety constraints for electrolyzer start-up and shutdown; efficiency models in electric-hydrogen collaborative scheduling often use fixed empirical curves, which cannot adapt to efficiency drift caused by equipment aging and temperature changes; and in terms of non-intrusive load monitoring, lightweight solutions that balance identification accuracy and user privacy protection are still immature. Summary of the Invention
[0007] To address these issues, the present invention provides a smart microgrid energy management system to solve the problems of existing technologies, such as the disconnect between digital twins and optimization decision-making, coarse multi-time-scale scheduling, accumulation of errors in the electric-hydrogen coupling model, and insufficient deep perception and decoupling on the load side.
[0008] To achieve the above objectives, the present invention provides the following technical solution:
[0009] The smart microgrid energy management system includes:
[0010] The equipment and sensing execution layer collects and uploads sensing data from all energy units, and receives and executes control messages; the energy units include distributed generation units, battery energy storage systems, electro-hydrogen coupled energy storage systems, and flexible loads and charging pile groups;
[0011] The edge computing and data preprocessing layer preprocesses the received sensor data, runs the photovoltaic ultra-short-term power correction model and a non-intrusive anomaly screening model, generates power prediction values, feature basis matrices and event messages, and uploads them to the cloud digital twin and intelligent decision-making layer; at the same time, it parses the scheduling instructions from the cloud digital twin and intelligent decision-making layer into control messages, sends them to the device and sensor execution layer after security verification, and feeds back the execution results;
[0012] The cloud-based digital twin and intelligent decision-making layer includes a digital twin engine, a hierarchical deep reinforcement learning collaborative optimization engine, a non-intrusive load decoupling module, and an electric-hydrogen coupled energy storage collaborative control module.
[0013] The cloud-based digital twin and intelligent decision-making layer is used to comprehensively infer early warning, load classification results and efficiency constraints to generate scheduling strategies, and then distribute the scheduling strategies to the edge computing and data preprocessing layer.
[0014] The application and interaction layer provides interactive interfaces for features including 3D digital twin visualization, scheduling simulation and actual curve comparison, carbon asset management, and alarm work order management.
[0015] Furthermore: the digital twin engine performs future time-domain extrapolation based on the current state before each scheduling, compares the final extrapolation value with the actual state collected subsequently to obtain a deviation vector, and uses the deviation vector to correct the model parameters and the state equation in scheduling optimization online;
[0016] The upper-layer deep reinforcement learning agent of the hierarchical deep reinforcement learning collaborative optimization engine generates a coarse scheduling target in the first cycle, and the lower-layer model prediction controller decomposes the coarse scheduling target into fine power instructions for each energy unit in the second cycle.
[0017] The non-intrusive load decoupling module takes the feature basis matrix and event messages uploaded by the edge computing and data preprocessing layer as input, and outputs flexible load classification results and minute-level power curves.
[0018] The electro-hydrogen coupling energy storage collaborative control module dynamically updates the optimal efficiency operating parameters of the electro-hydrogen coupling device based on the digital twin model, and injects the constraints of the lower-level model predictive controller.
[0019] Furthermore, the edge computing and data preprocessing layer preprocesses the sensor data by: aligning, interpolating, and removing outliers from the received sensor data, standardizing it into structured data packets, asynchronously uploading it to the cloud message queue, and writing the n-minute aggregated data into the local time-series database.
[0020] Furthermore: The operation of the photovoltaic ultra-short-term power correction model in the edge computing and data preprocessing layer specifically involves: deploying a lightweight inference engine within the edge computing node, performing a local correction at a preset high frequency in conjunction with all-sky imager data, and outputting a power prediction value with an m-minute resolution.
[0021] Furthermore, the edge computing and data preprocessing layer operates a non-intrusive abnormal event screening model, which specifically involves separating the feature blind source of the real-time current waveform through rapid independent component analysis, using a shallow convolutional neural network for initial screening of device types, and only pushing suspected abnormal or start-up / shutdown events to the cloud in the form of messages.
[0022] Furthermore: the edge computing and data preprocessing layer includes a data cleaning and standardization module, a photovoltaic ultra-short-term power correction module, a non-intrusive event anomaly screening module, and an instruction parsing and security verification module;
[0023] The data cleaning and standardization module is used to perform the preprocessing; the instruction parsing and security verification module is used to parse the scheduling instruction into a control message and perform power limiting, and then send it out after performing security verification based on the voltage and current limits of the local device port.
[0024] Furthermore: The digital twin engine, based on the physical information neural network combined with the state space equation, dynamically models the entire system. Before each scheduling decision, it performs a time-domain extrapolation of the next 4 to 24 hours based on the current state. It compares the final extrapolation value with the actual state collected by subsequent sensors to obtain a deviation vector. After smoothing, the deviation vector is used to trigger online fine-tuning of the physical information neural network model parameters and as a correction term in the state equation during scheduling optimization.
[0025] Furthermore: the state space of the upper-layer deep reinforcement learning agent includes microgrid power deficit, average battery state of charge, hydrogen storage tank pressure, load and photovoltaic prediction curves, electricity price and carbon emission factor; the actions of the upper-layer deep reinforcement learning agent are total energy storage charging and discharging power, start-up and shutdown of electrolyzer and fuel cell and power, and power purchased and sold from the main grid; the reward function is a weighted sum of operating cost, carbon emissions and power supply reliability.
[0026] While receiving the coarse scheduling target, the lower-level model prediction controller uses the electrochemical model in the digital twin constructed by the digital twin engine to calculate the optimal electrolysis power point in real time and inject it as a constraint. When the cumulative deviation exceeds a preset threshold, it triggers the upper-level deep reinforcement learning agent to make an immediate re-decision.
[0027] Furthermore: the deep neural network in the non-intrusive load decoupling module is a deep neural network based on a bidirectional long short-term memory network and a self-attention mechanism;
[0028] The flexible load classification results include rigid loads, movable loads, and interruptible loads;
[0029] The deep neural network is trained using a federated learning framework, where each participant only uploads the model gradients and does not share the original load data.
[0030] Furthermore: the application and interaction layer includes:
[0031] A 3D digital twin visualization interface supports real-time rendering of operating status, historical playback, and device selection query.
[0032] The scheduling comparison view displays the actual execution curve and the digital twin projection curve on the same timeline;
[0033] The blockchain-based carbon asset management interface writes the net carbon emission reductions calculated at the edge layer into a distributed ledger; and includes alarm and work order management modules.
[0034] This invention has the following advantages: By constructing a four-layer closed-loop architecture of device-sensor-edge-cloud-application, it deeply integrates digital twin endogenous simulation with hierarchical deep reinforcement learning optimization. This significantly improves the predictability and accuracy of scheduling strategies by using deductive early warning to drive decision-making and deviation feedback to correct models in real time. It adopts multi-timescale collaboration of 15-second photovoltaic power correction at the edge and 15-minute coarse scheduling and 1-minute fine allocation in the cloud to effectively match the minute-level fluctuations of renewable energy and the second-level impact of high-power equipment. At the same time, it uses an electro-hydrogen coupled twin model to dynamically inject optimal efficiency constraints, adapting to equipment aging and temperature drift, reducing operating losses and carbon emissions. Non-intrusive load decoupling is based on a federated learning framework, achieving fine load classification while protecting user privacy, improving demand-side response capabilities, thereby comprehensively improving the economic efficiency, power supply reliability, and carbon emission reduction benefits of microgrid operation.
[0035] Other features and advantages of the present invention will be set forth in the following description. Attached Figure Description
[0036] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).
[0037] Fig. 1 This is an architecture diagram of the smart microgrid energy management system provided in the embodiments of this application.
[0038] Fig. 2 This is a flowchart illustrating the implementation process of the smart microgrid energy management system of the present invention. Detailed Implementation
[0039] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Those skilled in the art can make some non-essential improvements and adjustments to the present invention based on the above-described content.
[0040] Please see Figs. 1-2 The smart microgrid energy management system includes an equipment and sensing execution layer, an edge computing and data preprocessing layer, a cloud-based digital twin and intelligent decision-making layer, and an application and interaction layer.
[0041] The device and sensing execution layer collects sensing data (including electrical and environmental status data) from all energy units through non-intrusive load monitoring terminals and multi-protocol intelligent acquisition terminals deployed at key nodes, and uploads the collected sensing data to the edge computing and data preprocessing layer; at the same time, it receives and executes control messages from the edge computing and data preprocessing layer.
[0042] The energy unit includes distributed generation units, battery energy storage systems, electro-hydrogen coupled energy storage systems, and flexible loads and charging pile groups; among them, the distributed generation units include rooftop photovoltaic arrays and micro inverters, small wind turbine generators, and backup micro gas turbines, etc.
[0043] Edge computing nodes in the edge computing and data preprocessing layer perform timestamp alignment, interpolation, and outlier removal on the received sensor data, and standardize it into structured data packets for asynchronous upload to the cloud message queue. At the same time, they write 5-minute aggregated data into the local time-series database.
[0044] A lightweight inference engine is deployed within the edge computing node to run a photovoltaic ultra-short-term power correction model. Every 15 seconds, it performs a local correction based on all-sky imager data and outputs a 1-minute resolution power prediction value. A non-intrusive anomaly screening model is also run, which performs feature blind source separation on real-time current waveforms using FastICA and performs initial screening of device types using a shallow CNN. Only suspected anomalies or start-up / shutdown events are pushed to the cloud in the form of messages.
[0045] The edge computing and data preprocessing layer includes four functional modules: data cleaning and standardization module, photovoltaic ultra-short-term power correction module, NILM event anomaly screening module, and instruction parsing and security verification module.
[0046] The data cleaning and standardization module receives sensor data, and the processed sensor data is uploaded to the cloud-based digital twin and intelligent decision-making layer in the form of "standardized data / event messages". The feature base matrix / start-stop events generated by the NILM event anomaly screening module are also uploaded to the cloud.
[0047] Meanwhile, the edge computing node receives scheduling instructions from the cloud-based digital twin and intelligent decision-making layer, parses the scheduling instructions into control messages for each device, and sends them to the device and sensing execution layer after power limiting and voltage / frequency safety verification, and then feeds the execution results back to the cloud.
[0048] The cloud-based digital twin and intelligent decision-making layer includes a digital twin engine, a hierarchical deep reinforcement learning engine, a non-intrusive load decoupling module, and an electro-hydrogen coupled energy storage collaborative control module.
[0049] The digital twin engine, built on a physical information neural network, uses PINN combined with state-space equations to dynamically model the entire system. Before each scheduling decision, it performs a time-domain extrapolation of the next 4 to 24 hours based on the current state, forming the digital twin endogenous simulation-correction closed loop: comparing the final extrapolation value with the actual state collected by subsequent sensors to obtain a deviation vector. This deviation vector, after smoothing, is used to trigger online fine-tuning of PINN model parameters and as a correction term for the state equation in scheduling optimization.
[0050] The hierarchical deep reinforcement learning collaborative optimization engine has a two-layer architecture. The upper layer uses a deep reinforcement learning agent (RL) to generate coarse scheduling objectives with a period of 15 minutes, while the lower layer uses a model predictive controller (MPC) to solve for optimal power allocation with a period of 1 minute. The state space of the upper-layer deep reinforcement learning agent includes microgrid power deficit, average battery state of charge, hydrogen storage tank pressure, load and photovoltaic prediction curves, electricity price and carbon emission factor. The actions are total energy storage charging and discharging power, electrolyzer / fuel cell start-up and shutdown and power, and power purchased and sold from the main grid. The reward function is a weighted sum of operating costs, carbon emissions and power supply reliability, and the TD3 algorithm is used.
[0051] While receiving the coarse scheduling target, the lower-level MPC uses the electrochemical model in the digital twin constructed by the digital twin engine to calculate the optimal electrolysis power point in real time and inject it as a constraint, forming a hierarchical deep reinforcement learning collaborative optimization closed loop. When the accumulated deviation exceeds the limit, it triggers the upper-level deep reinforcement learning agent to make an immediate re-decision.
[0052] The non-intrusive load decoupling module deploys a deep neural network based on bidirectional LSTM and self-attention mechanism in the cloud. It takes the feature basis matrix and event messages uploaded by the edge computing and data preprocessing layer as input and outputs minute-level power curves for rigid loads, movable loads, and interruptible loads. The network is trained using a federated learning framework, where each participant only uploads the model gradient and does not share the original load data.
[0053] The electro-hydrogen coupled energy storage collaborative control module links the special operating constraints of the electrolyzer, hydrogen storage tank and fuel cell with the lower-level MPC. The optimal efficiency operating parameters are updated in real time by the digital twin electrochemical and thermodynamic models, and the MPC constraints are dynamically corrected.
[0054] The cloud-based digital twin and intelligent decision-making layer constructs a dynamic digital twin based on a physical information neural network. Before each scheduling, it predicts the system evolution over the next few hours and corrects the model online. The upper-layer deep reinforcement learning agent generates coarse scheduling targets, while the lower-layer model predicts and decomposes them into fine power setpoints for each device. It combines a federated learning framework to achieve non-intrusive load decoupling and flexible load classification. It also dynamically injects the optimal efficiency parameters of the electro-hydrogen coupling devices based on the twin model into optimization constraints. The final output of this cloud-based digital twin and intelligent decision-making layer is a multi-timescale scheduling strategy derived from the combined effects of prediction and early warning, load decoupling, and efficiency constraints. The scheduling strategy is then distributed to the edge computing and data preprocessing layer.
[0055] The application and interaction layer provides interactive interfaces for operators, such as 3D digital twin visualization, scheduling simulation and actual curve comparison, blockchain-based carbon asset trust management, and alarm work orders, making the scheduling decision-making process and carbon emission reduction benefits transparent and traceable.
[0056] Specifically, the application and interaction layer includes: a 3D digital twin visualization interface that supports real-time rendering of operating status, historical playback, and device selection query; a scheduling comparison view that displays the actual execution curve and the digital twin projection curve on the same timeline; a blockchain-based carbon asset management interface that writes the net carbon emission reductions calculated by the edge computing and data preprocessing layer into a distributed ledger, supporting compliance integration with the carbon trading market and export of verification reports; and alarm and work order management modules.
[0057] In this invention, physical layer sensing data is preprocessed by the edge computing and data preprocessing layer and screened at the edge before being uploaded to the cloud layer. The cloud layer generates scheduling instructions based on inference and optimization, which are then parsed and verified by the edge computing and data preprocessing layer before being sent to the physical layer for execution. The execution results of the physical layer are then fed back to the cloud layer via the edge computing and data preprocessing layer. In this process, three closed-loop circuits operate collaboratively across all layers: the digital twin endogenous simulation-correction closed loop runs within the cloud twin, continuously refining the model through inference and measured deviations; the hierarchical deep reinforcement learning collaborative optimization closed loop spans the cloud and edge computing and data preprocessing layers, performing coarse scheduling, fine allocation, and deviation-triggered re-decision coordination; and the macro-optimization scheduling closed loop runs through cloud decision-making, edge distribution, and physical execution, performing rolling optimization on a 15-minute cycle, and is protected locally by the edge computing and data preprocessing layer in emergency situations.
[0058] See also Fig. 2 The complete scheduling process of the system of the present invention shown below includes the following specific steps:
[0059] Step S1: The edge node cleans and standardizes the sensor data and uploads it to the cloud message queue.
[0060] Step S2: Based on the current system status, the digital twin engine performs a time-domain simulation of 4 to 24 hours to identify potential risk conditions such as voltage over-limit and energy storage overcharging / over-discharging in the future.
[0061] Step S3: The load decoupling module combines the NILM event messages uploaded by the edge computing and data preprocessing layer to classify flexible loads into rigid loads, movable loads, and interruptible loads, and outputs minute-level power estimates.
[0062] Step S4: Based on the electrochemical model in the digital twin, the electro-hydrogen co-operation module dynamically injects the current optimal efficiency operating parameters of the electrolyzer and fuel cell into the constraints of the lower-level MPC.
[0063] Step S5: The upper-layer RL agent integrates digital twin simulation and early warning information, load decoupling classification results, electricity-hydrogen efficiency constraints, time-of-use electricity prices, and carbon emission factors to generate coarse scheduling targets with a granularity of 15 minutes.
[0064] Step S6: The lower-level MPC decomposes the coarse scheduling target into fine power commands for each device in a 1-minute cycle, and sends them to the edge nodes after security verification.
[0065] Step S7: After parsing the scheduling instructions and performing local security checks, the edge node sends control instructions to each physical device and feeds back the execution results.
[0066] Step S8: Determine whether the cumulative deviation exceeds a preset threshold;
[0067] If the time limit is not exceeded, wait for the next MPC cycle to perform rolling optimization and return to step S6;
[0068] If the threshold is exceeded, an immediate re-decision is triggered by the upper-level RL agent, and the process returns to step S5.
[0069] The above process forms a closed loop, with a 1-minute rolling optimization nested within a 15-minute macro scheduling cycle, and a rapid response under abnormal operating conditions achieved through a deviation threshold triggering mechanism.
[0070] Furthermore, the collection of sensor data requires the consent of the data subject (including organizations, enterprises, or individuals) before collection to ensure that the data source complies with relevant laws and regulations. In addition, an independent privacy compliance audit log is generated to record the scope, time, data subjects involved, and purpose of data collection. This log is tamper-proof and is used by relevant enterprises or organizations to fulfill their compliance audit obligations as required by laws and regulations.
[0071] In this invention, the device and sensing execution layer output sensor data as the uplink. After processing by the edge computing and data preprocessing layer, the data is uploaded to the cloud. The scheduling instructions generated in the cloud are parsed by the edge computing and data preprocessing layer and converted into control messages for execution. The edge computing and data preprocessing layer feeds back execution deviations to the cloud, thus forming a macroscopic scheduling closed loop that spans four layers. Simultaneously, the digital twin engine and the hierarchical deep reinforcement learning engine form an endogenous closed loop that drives optimization through deduction and corrects deviations through feedback. The two closed loops work together to achieve a deep integration of simulation and optimization.
[0072] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A smart microgrid energy management system, characterized in that, include: The equipment and sensing execution layer collects and uploads sensing data from all energy units, and receives and executes control messages; the energy units include distributed generation units, battery energy storage systems, electro-hydrogen coupled energy storage systems, and flexible loads and charging pile groups; The edge computing and data preprocessing layer preprocesses the received sensor data, runs the photovoltaic ultra-short-term power correction model and a non-intrusive anomaly screening model, generates power prediction values, feature basis matrices and event messages, and uploads them to the cloud digital twin and intelligent decision-making layer; at the same time, it parses the scheduling instructions from the cloud digital twin and intelligent decision-making layer into control messages, sends them to the device and sensor execution layer after security verification, and feeds back the execution results; The cloud-based digital twin and intelligent decision-making layer includes a digital twin engine, a hierarchical deep reinforcement learning collaborative optimization engine, a non-intrusive load decoupling module, and an electric-hydrogen coupled energy storage collaborative control module. The cloud-based digital twin and intelligent decision-making layer is used to comprehensively infer early warning, load classification results and efficiency constraints to generate scheduling strategies, and then distribute the scheduling strategies to the edge computing and data preprocessing layer. The application and interaction layer provides interactive interfaces for features including 3D digital twin visualization, scheduling simulation and actual curve comparison, carbon asset management, and alarm work order management.
2. The smart microgrid energy management system according to claim 1, characterized in that, The digital twin engine performs future time-domain extrapolation based on the current state before each scheduling, compares the final extrapolation value with the actual state collected subsequently to obtain a deviation vector, and uses the deviation vector to correct the model parameters and the state equation in scheduling optimization online. The upper-layer deep reinforcement learning agent of the hierarchical deep reinforcement learning collaborative optimization engine generates a coarse scheduling target in the first cycle, and the lower-layer model prediction controller decomposes the coarse scheduling target into fine power instructions for each energy unit in the second cycle. The non-intrusive load decoupling module takes the feature basis matrix and event messages uploaded by the edge computing and data preprocessing layer as input, and outputs flexible load classification results and minute-level power curves. The electro-hydrogen coupling energy storage collaborative control module dynamically updates the optimal efficiency operating parameters of the electro-hydrogen coupling device based on the digital twin model, and injects the constraints of the lower-level model predictive controller.
3. The smart microgrid energy management system according to claim 1, characterized in that, The edge computing and data preprocessing layer preprocesses the sensor data by: aligning, interpolating, and removing outliers from the received sensor data, standardizing it into structured data packets, asynchronously uploading it to the cloud message queue, and writing the n-minute aggregated data into the local time-series database.
4. The smart microgrid energy management system according to claim 1, characterized in that, The operation of the photovoltaic ultra-short-term power correction model in the edge computing and data preprocessing layer is as follows: a lightweight inference engine is deployed in the edge computing node to perform a local correction at a preset high frequency in combination with all-sky imager data, and outputs a power prediction value with a resolution of m minutes.
5. The smart microgrid energy management system according to claim 1, characterized in that, The edge computing and data preprocessing layer operates a non-intrusive abnormal event screening model, which specifically involves separating the blind source of the real-time current waveform through rapid independent component analysis, using a shallow convolutional neural network for initial screening of equipment types, and only pushing suspected abnormal or start-up / shutdown events to the cloud in the form of messages.
6. The smart microgrid energy management system according to claim 1, characterized in that, The edge computing and data preprocessing layer includes a data cleaning and standardization module, a photovoltaic ultra-short-term power correction module, a non-intrusive event anomaly screening module, and an instruction parsing and security verification module. The data cleaning and standardization module is used to perform the preprocessing; the instruction parsing and security verification module is used to parse the scheduling instruction into a control message and perform power limiting, and then send it out after performing security verification based on the voltage and current limits of the local device port.
7. The smart microgrid energy management system according to claim 1, characterized in that, The digital twin engine dynamically models the entire system based on a physical information neural network combined with state-space equations. Before each scheduling decision, it performs a time-domain extrapolation of the next 4 to 24 hours based on the current state. It compares the final extrapolation value with the actual state collected by subsequent sensors to obtain a deviation vector. After smoothing, the deviation vector is used to trigger online fine-tuning of the physical information neural network model parameters and as a correction term in the state equation during scheduling optimization.
8. The smart microgrid energy management system according to claim 2, characterized in that, The state space of the upper-layer deep reinforcement learning agent includes microgrid power deficit, average battery state of charge, hydrogen storage tank pressure, load and photovoltaic prediction curves, electricity price and carbon emission factor; the actions of the upper-layer deep reinforcement learning agent are total energy storage charging and discharging power, start-up and shutdown and power of electrolyzer and fuel cell, and power purchased and sold from the main grid; the reward function is a weighted sum of operating cost, carbon emissions and power supply reliability. While receiving the coarse scheduling target, the lower-level model prediction controller uses the electrochemical model in the digital twin constructed by the digital twin engine to calculate the optimal electrolysis power point in real time and inject it as a constraint. When the cumulative deviation exceeds a preset threshold, it triggers the upper-level deep reinforcement learning agent to make an immediate re-decision.
9. The smart microgrid energy management system according to claim 1, characterized in that, The deep neural network in the non-invasive load decoupling module is a deep neural network based on a bidirectional long short-term memory network and a self-attention mechanism. The flexible load classification results include rigid loads, movable loads, and interruptible loads; The deep neural network is trained using a federated learning framework, where each participant only uploads the model gradients and does not share the original load data.
10. The smart microgrid energy management system according to claim 1, characterized in that, The application and interaction layer includes: A 3D digital twin visualization interface supports real-time rendering of operating status, historical playback, and device selection query. The scheduling comparison view displays the actual execution curve and the digital twin projection curve on the same timeline; The blockchain-based carbon asset management interface writes the net carbon emission reductions calculated at the edge layer into a distributed ledger; and includes alarm and work order management modules.