Multi-park energy consumption prediction scheduling method and control system based on digital twinning

Abstract

The invention provides a multi-park energy consumption prediction scheduling method based on digital twinning and a control system, and systematically solves the problem of the pain point of multi-park energy consumption management by constructing a technical chain from data perception to closed-loop optimization. The method comprises the following steps: firstly, by constructing a global unified digital twinborn model, standardized integration of dispersed and heterogeneous park assets and data is realized, and the problem of information islands is solved; secondly, prediction is carried out by adopting federated learning, cross-park knowledge sharing and joint modeling are realized on the premise of ensuring data privacy and security of each park, and the prediction precision of a single park under the condition of limited data is remarkably improved; and finally, through a'prediction-decision-execution-update 'closed-loop process, traditional passive and static energy consumption management is converted into active and dynamic prediction scheduling, so that the energy consumption peak can be stabilized prospectively, the energy distribution can be optimized, and the comprehensive energy consumption cost and carbon emission can be effectively reduced.

Description

Technical Field

[0001] This application relates to the field of multi-park management technology, and in particular to a multi-park energy consumption prediction and scheduling method and control system based on digital twins. Background Technology

[0002] Traditional energy management in industrial parks largely relies on deploying various sensors and smart meters to build isolated energy monitoring platforms, enabling data visualization and report statistics. However, such systems can only provide historical and real-time data, representing a reactive, "hindsight-based" approach. They cannot proactively predict future energy consumption trends, nor do they have the capability to actively intervene and regulate based on forecast results. Consequently, they struggle to achieve a closed-loop optimization of energy use throughout its entire lifecycle, encompassing "pre-planning, in-process control, and post-analysis."

[0003] To improve energy efficiency, several advanced solutions have emerged in existing technologies. Firstly, some systems attempt to introduce single prediction algorithms, such as time-series analysis based on historical data, to roughly predict the overall electricity consumption of the park. However, such methods generally suffer from insufficient accuracy, especially in dealing with the strong fluctuations brought about by the integration of renewable energy sources like photovoltaics. Furthermore, the coarse-grained predictions make it difficult to support refined management at the equipment level. Secondly, some single-park systems have begun to attempt simple automated control, such as starting and stopping lighting and adjusting air conditioning temperatures according to preset schedules. However, these control strategies are static and rigid, unable to be linked with real-time changes in energy supply and demand (such as photovoltaic power generation), dynamic energy demand, and cross-system collaboration, resulting in limited optimization effects. Moreover, most existing solutions are designed for single parks. When the management scope expands to multiple geographically dispersed parks, their inherent shortcomings are amplified dramatically: data from each park becomes isolated, and due to data privacy and security considerations, it is difficult to centrally process and jointly analyze the raw data, making it impossible to tap energy-saving potential at the cluster level and achieve cross-park energy collaborative scheduling and mutual assistance.

[0004] In recent years, digital twins and federated learning technologies have offered new solutions to the aforementioned problems. Digital twin technology promises to create high-fidelity virtual mappings of physical campuses, but its current application in campus management is largely limited to geometric model visualization and equipment status monitoring, failing to deeply integrate with business processes and lacking sufficient integration with energy consumption prediction and optimization control logic, resulting in models that are "good to look at but not practical." While federated learning technology can achieve joint modeling with "data not leaving the domain," existing research focuses on theory or single-function verification, and there are no mature cases of successfully integrating it deeply with digital twin technology and applying it to the complex scenario of multi-campus energy consumption prediction and scheduling. Furthermore, there is still a lack of effective linkage between various systems (such as energy management, maintenance work orders, and microgrid monitoring), forming "data silos" and "functional chimneys." For example, it is impossible to automatically translate abnormal energy consumption prediction results into maintenance actions, nor can it dynamically adjust energy consumption strategies based on the real-time status of the microgrid.

[0005] Therefore, there is an urgent need in this field for an innovative system that can break down the aforementioned barriers and achieve energy efficiency management across multiple parks. Summary of the Invention

[0006] To address the aforementioned technical issues, this application provides a multi-campus energy consumption prediction and scheduling method and control system based on digital twins.

[0007] A first aspect of this application provides a multi-campus energy consumption prediction and scheduling method based on digital twins, comprising the following steps: S100: Data acquisition steps, which collect real-time operation data, energy consumption data and environmental parameters of various energy-consuming systems and microgrids in multiple distributed parks in parallel; S200: Digital twin construction steps, based on the collected data, to construct digital twins for each park, including four-level topological relationships of project-building-space-equipment, and to summarize them into a unified model for the entire region; S300: Federated learning prediction step, based on the global unified model, drives the federated learning process, and generates energy consumption prediction data for future periods of each park by integrating the weight parameters of the local prediction models of each park. S400: Flexible decision-making steps compare the energy consumption forecast data of each park with the flexible energy consumption budget. If the difference exceeds the preset value, a set of energy-saving control instructions is generated, which includes at least one of the following devices: energy storage, HVAC, lighting and charging piles. S500: Instruction execution steps, parse the topology relationship of equipment in each park, issue and execute the energy-saving control instruction set, and collect feedback data; S600: Closed-loop update steps, using feedback data to update the digital twin status of each park and the weight parameters of the local prediction model respectively.

[0008] By constructing a technology chain from data perception to closed-loop optimization, the pain points of energy consumption management in multiple parks are systematically solved. First, by building a unified digital twin model across the entire domain, the standardized integration of dispersed and heterogeneous park assets and data is achieved, overcoming the problem of information silos. Second, federated learning is used for prediction, which, while ensuring the data privacy and security of each park, enables cross-park knowledge sharing and joint modeling, significantly improving the prediction accuracy of a single park with limited data. Finally, through a closed-loop process of "prediction-decision-execution-update", the traditional passive and static energy consumption management is transformed into proactive and dynamic predictive scheduling, which can proactively smooth out peak energy consumption, optimize energy allocation, and effectively reduce overall energy consumption costs and carbon emissions.

[0009] In one or more embodiments of this application, the specific process of constructing a digital twin in step S200 includes: automatically generating a project-building-space-device hierarchical path to which the device belongs based on device access information and spatial layout diagram to form a subordinate topology; simultaneously, automatically generating an electrical connection topology between devices based on the subordinate relationship data of smart meters, circuit breakers and distribution boxes; the digital twin is a fusion of the subordinate topology and the electrical connection topology.

[0010] In the above embodiments, by automatically generating and integrating the membership topology and electrical connection topology, the digital twin not only reflects the management affiliation of the equipment but also accurately depicts the actual energy flow path. This provides an irreplaceable data foundation for subsequent precise scheduling (such as preventing local overload) and avoids misconfiguration or failure of control commands due to inaccurate topology relationships.

[0011] In one or more embodiments of this application, when constructing the digital twins of each park in step S200, a digital energy consumption benchmark fingerprint is dynamically created and bound to each device. The energy consumption benchmark fingerprint is composed of static attributes, dynamic operating parameters, and historical statistical features of the device. The static attributes include at least the device type, rated power, manufacturing date, and energy efficiency rating. The dynamic operating parameters are power consumption data collected after the device has been running continuously for at least 24 hours under normal operating conditions. The historical statistical features are steady-state power consumption Gaussian distribution parameters calculated from the power consumption data, including the mean μ and variance σ. 2 .

[0012] In the above embodiments, the created "energy consumption benchmark fingerprint" establishes a dynamic energy consumption health profile for each device. It not only provides a high-precision interpolation basis for missing data, ensuring the quality of model input, but also serves as a quantitative benchmark for device energy efficiency degradation and early fault diagnosis, realizing a functional extension from simple energy consumption monitoring to device health management.

[0013] In one or more embodiments of this application, the federated learning process in step S300 adopts a hybrid architecture that combines long and short time windows; wherein, the long time window model uses long-term data from the past 7 to 30 days as the training set; the short time window model uses recent data from the past 1 to 24 hours as the training set; and an adaptive weight fusion mechanism is designed in the cloud server, the formula of which is expressed as: W = α W_long + (1-α) W_short; Where W represents the weight parameters of the fused prediction model; W_long represents the weights of the long-term window model, which includes long-term (7 to 30 days) seasonal and cyclical trend characteristics; W_short represents the weights of the short-window model, which includes short-term (1 to 24 hours) instantaneous fluctuations and abrupt changes. α is an adaptive weighting factor, a dynamically changing coefficient with a value range of [0, 1]. Its magnitude is determined by the variance of photovoltaic power fluctuation. If the variance is large, α is reduced to make the model more trusting of the short-window model; if the variance is small, α is increased to make the model more trusting of the long-window model.

[0014] In the above embodiments, the design of long and short time window models and adaptive weighting factors enables the prediction model to intelligently balance long-term patterns and short-term fluctuations. In particular, the dynamic correlation between the weighting factors and the variance of photovoltaic fluctuations greatly improves the response speed and prediction accuracy to sudden weather changes in scenarios with a high proportion of renewable energy access, and reduces scheduling errors caused by prediction lag.

[0015] In one or more embodiments of this application, the elastic energy consumption budget in step S400 is a mechanism with upper and lower boundaries and dynamically adjustable width. Its decision logic is as follows: the system monitors the operating status of each park's microgrid in real time, continuously calculates its current surplus photovoltaic power and the adjustable capacity of the energy storage system; if the park's energy consumption prediction data exceeds the budget limit, but the system determines that it currently has surplus photovoltaic power, then the strip widening strategy is automatically triggered, temporarily relaxing the elastic energy consumption budget to the park's budget limit constraint value, and giving priority to generating instructions to guide the consumption of local surplus photovoltaic power; if the prediction data exceeds the budget limit and the park's microgrid is in a state of insufficient photovoltaic output or dependent on grid power supply, then the tightening strategy is automatically triggered, temporarily tightening the elastic energy consumption budget to less than or equal to 70% of the park's budget limit constraint value, and generating instructions to reduce the operating intensity of energy-consuming equipment in the park.

[0016] In the above embodiments, the fixed energy consumption budget is transformed into a "flexible strip" that is linked to renewable energy in real time. This mechanism can dynamically adjust the intensity of energy-saving strategies, prioritizing the consumption of green electricity when there is a surplus of photovoltaic power and decisively suppressing the load when there is a shortage of photovoltaic power. In this way, while ensuring energy demand, the local consumption rate of renewable energy is maximized, achieving a balance between economic efficiency and environmental friendliness.

[0017] In one or more embodiments of this application, in step S500, before the energy-saving control instruction set is sent to a certain park for execution, a centralized instruction conflict arbitration and verification step is introduced: This step traces all instruction receiving devices based on the digital twin of the park. When it is detected that the same device is assigned two or more instructions with logical contradictions or mutually exclusive operations within the same time window, an arbitration rule is triggered. The arbitration rule is based on a preset energy consumption sensitivity ranking knowledge base. The knowledge base quantifies and ranks the sensitivity of all types of controlled devices and their control instructions to the global energy consumption. The arbitration engine prioritizes the execution of the instruction that contributes the most to the global energy saving effect and has the highest sensitivity, while the conflicting secondary instructions are temporarily stored in a cache queue and delayed until the next control cycle or until the preconditions change before being evaluated and sent out.

[0018] In the above embodiments, by introducing a centralized arbitration mechanism based on energy consumption sensitivity, the problem of instruction conflict that may occur when multiple automation systems are controlled in parallel is effectively solved. This ensures the coordination and consistency of control logic, avoids frequent actuator actions or strategy failures, and greatly improves the reliability and stability of complex automation systems.

[0019] In one or more embodiments of this application, in step S500, the energy-saving control instruction set supports cross-system scenario linkage based on digital twin topology. Specifically, when the system predicts that the total load of a certain park's charging piles is about to exceed the safety threshold of its upstream transformer capacity, the control logic does not simply cut off the charging power. Instead, it first queries the digital twin to accurately locate all smart lighting circuits that are in the same transformer power supply circuit as the group of charging piles. Then, it automatically generates instructions to uniformly lower the dimming value of these lighting circuits to a preset level without affecting safe lighting, thereby reducing the total load of the transformer. After the peak load of the charging piles has passed, it automatically restores the lighting dimming value to 100%.

[0020] In the above embodiments, by utilizing the precise electrical topology of a digital twin, an innovative approach is taken to use the intelligent lighting system as a flexible adjustment resource to cope with impact loads such as charging piles. This method achieves "soft expansion" of the local power grid at extremely low cost (compared to transformer expansion), ensuring power supply security and significantly saving on fixed asset investment.

[0021] In one or more embodiments of this application, step S600, which uses feedback data to update the digital twin state and local prediction model weight parameters of each park, includes: adopting a gradient truncation incremental learning strategy: when the system receives device execution feedback data, if it finds that the error between the data and the expected state of the model is greater than a preset value, the algorithm is updated; wherein, the update algorithm does not perform a global retraining of the entire federated learning model, but only performs local fine-tuning of the fully connected weights of the last layer or the last two layers of the model, using an extremely low learning rate for iteration; at the same time, the feature extraction weights of the intermediate layers of the model are kept completely frozen to prevent damage or overwriting of previously learned, well-generalized knowledge.

[0022] In the above embodiments, a gradient-truncation incremental learning strategy is used for model updates, enabling the system to adapt to new data changes while effectively protecting the learned core knowledge from being destroyed (avoiding forgetting). This robust update mechanism ensures the continuity and stability of the algorithm during long-term operation and reduces maintenance costs.

[0023] A second aspect of this application provides a digital twin-based multi-campus control system for implementing any of the methods described above, comprising: a cloud server, wherein the cloud server is configured with: The digital twin module is used to build and maintain a unified four-level topology digital twin model that includes hierarchical and electrical connection relationships based on project, building, space and equipment information in each park. The energy consumption prediction module is connected to the digital twin module and is used to access the energy consumption, environmental and microgrid operation data of each park, and generate energy consumption prediction data for future periods based on the federated learning algorithm. The predictive scheduling decision module is connected to the energy consumption prediction module and the digital twin module. It is used to compare the energy consumption prediction results with the preset energy consumption budget and generate a set of energy-saving control instructions for HVAC, lighting, energy storage and charging pile systems. The operation and maintenance linkage module is connected to the energy consumption prediction module and is used to automatically generate and dispatch inspection work orders to the operation and maintenance service module when abnormal energy consumption of equipment is predicted. Multiple edge acquisition and control nodes are deployed in each park to collect local data and upload it to the cloud server, as well as to receive and execute the energy-saving control instruction set issued by the cloud server. The microgrid monitoring module is integrated into the cloud server or the multiple edge acquisition and control nodes. It is used to monitor the operating status of photovoltaic, energy storage, charging piles and power distribution systems in real time, and provide the status data to the energy consumption prediction module and the prediction and scheduling decision module. The digital twin module, energy consumption prediction module, prediction scheduling decision module, operation and maintenance linkage module, microgrid monitoring module, and multiple edge acquisition and control nodes work together to form a cross-park data acquisition, prediction analysis, optimization decision and closed-loop execution control system.

[0024] In the above embodiments, the system implements the hardware architecture and software module entities of the above method. Through the distributed architecture design of cloud collaboration, it not only ensures the powerful computing and collaborative capabilities of the cloud, but also endows the edge side with real-time response and data privacy protection capabilities, providing efficient, reliable and scalable physical support for the implementation of the method.

[0025] In one or more embodiments of this application, the operation and maintenance linkage module is configured to: continuously monitor the device-level energy consumption anomaly prediction results and alarm events output by the prediction scheduling decision module; automatically call the operation and maintenance service work order creation function through the internal API interface, convert the abnormal events into executable predictive inspection work orders, and automatically assign inspection priorities for inspection.

[0026] In the above embodiments, by achieving automatic linkage between energy consumption prediction and maintenance work orders at the system architecture level, the data flow of "monitoring-prediction-maintenance" is streamlined, creating a new predictive maintenance model. This can transform energy efficiency anomalies into timely maintenance actions, eliminating sub-healthy equipment conditions at the source, while improving maintenance efficiency and energy usage efficiency. Attached Figure Description

[0027] Figure 1 A flowchart illustrating a multi-campus energy consumption prediction and scheduling method based on digital twins provided in an embodiment of this application; Figure 2 This is a structural diagram of a multi-campus control system based on digital twins provided in an embodiment of this application; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0028] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0029] To make the purpose, technical solution, and advantages of this application clearer, the following will be described in conjunction with the appendix. Figure 1-3 The following is an explanation using specific examples.

[0030] Please refer to Figure 1, Figure 1 A flowchart illustrating a multi-campus energy consumption prediction and scheduling method based on digital twins, provided as an embodiment of this application, includes the following steps: S100: Data acquisition steps, which collect real-time operation data, energy consumption data and environmental parameters of various energy-consuming systems and microgrids in multiple distributed parks in parallel; S200: Digital twin construction steps, based on the collected data, to construct digital twins for each park, including four-level topological relationships of project-building-space-equipment, and to summarize them into a unified model for the entire region; S300: Federated learning prediction step, based on the global unified model, drives the federated learning process, and generates energy consumption prediction data for future periods of each park by integrating the weight parameters of the local prediction models of each park. S400: Flexible decision-making steps compare the energy consumption forecast data of each park with the flexible energy consumption budget. If the difference exceeds the preset value, a set of energy-saving control instructions is generated, which includes at least one of the following devices: energy storage, HVAC, lighting and charging piles. S500: Instruction execution steps, parse the topology relationship of equipment in each park, issue and execute the energy-saving control instruction set, and collect feedback data; S600: Closed-loop update steps, using feedback data to update the digital twin status of each park and the weight parameters of the local prediction model respectively.

[0031] Understandingly, a digital twin refers to a high-fidelity virtual model created through digital means for a project, building, space, and equipment. This model not only reflects its geometric appearance but, more importantly, maps its real-time state (e.g., switching, energy consumption), behavior (e.g., operating modes), and rules (e.g., topological relationships), enabling simulation, analysis, and control. Multiple parks refer to geographically dispersed park entities that may be administratively independently managed. Federated learning is a distributed machine learning technique whose core idea is: data remains local (across parks), without raw data exchange; models are trained locally, and then the model parameters (e.g., weights, gradients) are encrypted and uploaded to a cloud server for aggregation, generating a more powerful global model. Flexible energy consumption budgeting is distinct from fixed, rigid budget indicators. It is a budget range or "band" that can dynamically adjust its upper and lower limits based on real-time conditions (especially renewable energy), essentially representing a dynamic resource allocation and decision boundary. Closed-loop update refers to the system's ability to re-input the results (feedback data) after the execution of control commands into the system for evaluating the effects and automatically adjusting model parameters or control strategies, forming a self-optimizing cycle of "perception-decision-execution-learning".

[0032] The principle of this embodiment is explained. (1) Data acquisition steps Perception layer hardware: Deploy various sensors (smart meters, temperature / humidity / light sensors, flow meters), IoT gateways, smart device controllers (such as DDC controllers), and microgrid monitoring equipment (photovoltaic inverter monitoring, energy storage BMS communication interface, charging pile controller) in each park.

[0033] Communication protocols: Industry standard protocols such as Modbus, BACnet, OPC UA, MQTT, etc. are adopted to aggregate data to local edge computing nodes in the park or directly upload it to the cloud server through IoT gateways.

[0034] The specific data collected includes: energy consumption data: instantaneous and cumulative values ​​of electricity, water, gas, etc. in each circuit; operation data: equipment start-up and shutdown status, operating frequency, power factor, and setpoints (such as temperature setpoint); environmental parameters: indoor and outdoor temperature and humidity, illuminance, and carbon dioxide concentration; microgrid status: real-time photovoltaic power generation, energy storage system SOC (state of charge), current load of charging piles, and power of grid feeders; key point: achieving "parallel" data collection across multiple parks means that a unified equipment access standard and communication framework are needed to ensure the timeliness and consistency of the data.

[0035] (2) Steps for constructing a digital twin Modeling platform: Uses a professional digital twin platform as the technical foundation.

[0036] Model definition: First, establish a standardized data model, defining entity types such as "project", "building", "space" and "equipment", as well as their respective attributes (such as name, ID, area), relationships and telemetry data points.

[0037] Instantiation: Based on the collected asset information, create digital twin instances of each entity on the platform, such as "Park A Energy Saving Project", "Building 1", "Conference Room 101", "Air Conditioner-1" and other instances, and establish connections according to the actual affiliation.

[0038] Data binding: The real-time data stream collected by S100 is bound to the corresponding digital twin instance attributes through a unique identifier (such as device ID), so that the virtual model can reflect the state of the physical entity in real time.

[0039] Cloud server aggregation: The digital twins of each park are registered and associated on the cloud server, forming a unified digital twin model network that covers all parks and can be globally queried and analyzed.

[0040] (3) Federated learning prediction steps Step 1: Cloud Initialization: The cloud server initializes a global prediction model and distributes it to all edge nodes participating in the park; Step 2: Local training: Each edge node in the park uses local historical energy consumption, environmental and other data to train the received global model for several rounds to obtain local model updates; Step 3: Parameter Upload: Each edge node encrypts the trained model weight parameters (not the training data) and uploads them to the cloud; Step 4: Cloud Aggregation: The cloud server collects the model weights of all parks and uses methods such as weighted averaging to aggregate and generate a new, improved global model; Step 5: Model Distribution: Distribute the new global model to each park. Iterative loop: Repeat steps 2-5 to continuously optimize the global model.

[0041] Prediction generation: Finally, each park uses the optimized global model or a model combined with local fine-tuning, inputs the latest real-time data, and generates energy consumption prediction curves for the next few hours, days, months, or even years at the park, building, and key equipment levels.

[0042] (4) Flexible decision-making steps Budget strip setting: The management console sets a basic energy consumption budget target value (such as monthly total electricity cost or total electricity consumption) and allows defining a dynamic floating range.

[0043] Decision Engine: Develop a rule engine or optimization algorithm module. The input to this module is the prediction curve generated in step S300 and the real-time microgrid status (PV output, energy storage SOC).

[0044] Decision logic: IF (predicted energy consumption > budget ceiling) AND (surplus photovoltaic power > threshold) THEN: trigger the "broadening" strategy to generate instructions such as "increase energy storage charging power" and "guide electric vehicle charging" to consume green electricity.

[0045] If (predicted energy consumption > budget limit) AND (insufficient photovoltaic output) THEN: trigger the "tightening" strategy and generate load reduction instructions such as "moderately increase the air conditioning temperature setpoint" and "reduce the lighting brightness in non-critical areas".

[0046] Instruction set generation: The output of the decision engine is a structured list of instructions, which clearly defines the instruction content, target device, execution time window, etc.

[0047] (5) Execution and closed-loop update Command sets are reliably sent to edge controllers in the designated area via message queues or industrial protocols. After execution, the controllers send confirmation receipts and new sensor readings back to the cloud.

[0048] Digital twin status update: Update the values ​​of corresponding attributes in the digital twin with real data returned from feedback (such as actual energy consumption).

[0049] Incremental model update: Using new data (predicted values ​​+ actual values) as training samples, the local parameters of the federated learning model are fine-tuned through online learning or mini-batch learning to correct prediction bias.

[0050] In the embodiments of this application, the above-described method realizes the transformation of multi-park energy consumption management from decentralized, isolated, passive response, and experience-driven to unified collaboration, proactive intervention, and data intelligence-driven. It effectively resolves the contradiction between data privacy and collaborative optimization, significantly improves prediction accuracy, achieves refined and forward-looking energy scheduling, and can continuously evolve through closed-loop learning, ultimately achieving significant energy conservation and consumption reduction, lower operating costs, and improved management efficiency.

[0051] In one or more embodiments of this application, the specific process of constructing a digital twin in step S200 includes: automatically generating a project-building-space-device hierarchical path to which the device belongs based on device access information and spatial layout diagram to form a subordinate topology; simultaneously, automatically generating an electrical connection topology between devices based on the subordinate relationship data of smart meters, circuit breakers and distribution boxes; the digital twin is a fusion of the subordinate topology and the electrical connection topology.

[0052] Membership topology refers to hierarchical relationships from an organizational management perspective, reflecting the logic of "who manages whom" and "who belongs to which place." For example, a piece of equipment belongs to a certain space, that space is located in a certain building, and that building belongs to a certain project. This is a tree-like structure. Electrical connection topology describes the connection relationships of the physical flow paths of energy, reflecting the physical facts of "who supplies power to whom" and "how current flows." For example, a distribution box supplies power to multiple lighting circuits, and these circuits form a star connection with the distribution box; transformers, busbars, and switchgear form a complex network structure.

[0053] Understandably, for a system designed for in-depth energy consumption analysis and optimization, a mere management topology is insufficient. A deep integration of the management dimension and the physical energy flow dimension is essential. Only by accurately understanding the actual path of current can precise scheduling based on real constraints (such as line capacity and transformer load), fault impact range analysis, and energy flow tracing be achieved. Linking these two topologies in a digital twin forms a comprehensive "energy information model."

[0054] The data for the topology primarily originates from asset management systems, spatial databases, or manual entry. It includes information such as equipment location codes and their respective departments. The data for the electrical connection topology comes from electrical design drawings (such as single-line diagrams), point configuration tables of intelligent power distribution systems, or is obtained through circuit identification technology.

[0055] The principle behind this embodiment is explained as follows: (1) Data import and processing: Import the two types of topological data mentioned above into the digital twin platform. Data cleaning and transformation may be required to ensure consistent naming conventions.

[0056] (2) Entity alignment: By using a unique device identifier (such as asset number), location information or circuit number, “device A” in the administrative topology and “load controlled by circuit breaker B” in the electrical topology are associated as the same entity.

[0057] (3) Relationship establishment: In the digital twin model, a device instance will have two relationship attributes at the same time. For example, "Air conditioner-01" has a relationship pointing to "Conference Room" (membership topology) and a relationship pointing to "Lighting Circuit" (electrical connection topology).

[0058] (4) Visualization and query: The platform should be able to display and query both types of topology at the same time. For example, clicking on a distribution box can display all the spaces and devices it manages (subordinate topology view) as well as all the electrical circuits and loads under it (electrical connection topology view).

[0059] In this implementation, the digital twin is upgraded from a "location map" to an "energy map" through dual topology fusion, enabling the system to perform advanced analysis based on real physical connections, such as accurate load forecasting, line loss analysis, and rapid fault location and isolation, thus avoiding control ineffectiveness or safety accidents caused by inaccurate models.

[0060] In one or more embodiments of this application, when constructing the digital twins of each park in step S200, a digital energy consumption benchmark fingerprint is dynamically created and bound to each device. The energy consumption benchmark fingerprint is composed of static attributes, dynamic operating parameters, and historical statistical features of the device. The static attributes include at least the device type, rated power, manufacturing date, and energy efficiency rating. The dynamic operating parameters are power consumption data collected after the device has been running continuously for at least 24 hours under normal operating conditions. The historical statistical features are steady-state power consumption Gaussian distribution parameters calculated from the power consumption data, including the mean μ and variance σ. 2 .

[0061] Understandably, an energy consumption benchmark fingerprint is a quantified, multi-dimensional energy consumption characteristic established for a specific device. It acts like the device's "energy consumption DNA," uniquely identifying its energy consumption pattern under normal operating conditions. Steady-state power consumption refers to the power consumption of a device when it is running stably near its rated load, excluding transient conditions such as startup and shutdown.

[0062] Gaussian distribution parameters: mean (μ) and variance (σ) 2 The mean represents the typical level of power consumption, while the variance represents the normal range of power consumption fluctuations. Using it to describe power consumption is based on the statistical assumption that normal random fluctuations exist during device operation.

[0063] In this implementation, this pattern is learned and quantified through historical data to establish a reliable baseline. Any significant and persistent deviation from this baseline can be considered abnormal.

[0064] The principle behind this embodiment is explained as follows: (1) Data preparation: Select high-frequency power consumption data (e.g., one point per minute) of the device under known health conditions and continuous operation for at least 24 hours (preferably covering different weekday / weekend modes).

[0065] (2) Feature calculation: This includes static attribute and dynamic parameter calculation. Static attributes are extracted directly from the equipment ledger and stored in database fields. Dynamic parameter calculation includes removing obvious outliers, identifying stable operating periods of the equipment through algorithms, and calculating the average (μ) and standard deviation (σ) and variance (σ²) of the power consumption data during the steady-state period. 2 This is the square of the standard deviation.

[0066] (3) Fingerprint binding and storage: The calculated μ and σ 2 The two attributes are stored and associated as a digital twin instance of the device.

[0067] (4) Imputation of missing data: When data is missing at a certain moment, N(μ, σ) can be used. 2 Generating a random number from the distribution or directly filling it with the μ value is more realistic than using a fixed value or the mean before and after interpolation.

[0068] (5) Anomaly detection: The deviation between the current power consumption and μ is calculated in real time. If the deviation continues to exceed K times σ (e.g., K=3, i.e., the 3σ principle), an anomaly alarm is triggered. This is more adaptable to the normal fluctuations of the equipment under different seasons and different loads than a fixed threshold.

[0069] In this implementation, energy consumption benchmark fingerprinting elevates equipment management from the level of "data availability" to "data quality and data value." The effects are reflected in: significantly improved data quality, providing cleaner input for predictive models; early and accurate warnings of equipment failures or energy efficiency degradation, supporting predictive maintenance and preventing energy waste and sudden equipment failures; and personalized, data-driven metrics for energy efficiency auditing and optimization.

[0070] In one or more embodiments of this application, the federated learning process in step S300 adopts a hybrid architecture that combines long and short time windows; wherein, the long time window model uses long-term data from the past 7 to 30 days as the training set; the short time window model uses recent data from the past 1 to 24 hours as the training set; and an adaptive weight fusion mechanism is designed in the cloud server, the formula of which is expressed as: W = α W_long + (1-α) W_short; Where W represents the weight parameters of the fused prediction model; W_long represents the weights of the long-term window model, which includes long-term (7 to 30 days) seasonal and cyclical trend characteristics; W_short represents the weights of the short-window model, which includes short-term (1 to 24 hours) instantaneous fluctuations and abrupt changes. α is an adaptive weighting factor, a dynamically changing coefficient with a value range of [0, 1]. Its magnitude is determined by the variance of photovoltaic power fluctuation. If the variance is large, α is reduced to make the model more trusting of the short-window model; if the variance is small, α is increased to make the model more trusting of the long-window model.

[0071] Understandably, long-window models are trained using historical data spanning a longer period (e.g., 7-30 days). Their advantage lies in learning macroscopic patterns of energy consumption, such as diurnal cycles, weekly cycles, and seasonal trends; the models are robust but slow to respond. Short-window models, on the other hand, are trained using recent data (e.g., 1-24 hours). Their advantage is the ability to quickly capture the latest changes, such as sudden weather events or load fluctuations caused by temporary activities; the models are sensitive but may lack a long-term perspective.

[0072] The adaptive weighting factor (α) is a dynamically changing fusion coefficient whose value is determined by external environmental conditions (in this case, photovoltaic volatility) and is used to intelligently allocate the "voice" of long and short time window models in the final prediction.

[0073] The implementation principle is explained as follows: (1) In each park, train two prediction models with the same structure but independent initial weights using long-term window data (such as the past 30 days) and short-term window data (such as the past 24 hours); (2) The cloud server continuously receives recent photovoltaic power sequences (such as data per second over the past hour) reported by each park and calculates the variance of the data within that time window. A large variance indicates drastic changes in sunlight and high uncertainty. (3) Pre-define a mapping function. For example, a function can be designed such that when the photovoltaic variance is 0 (extremely stable), α=1 (fully trust the long-term window model); when the photovoltaic variance exceeds a maximum threshold (extremely unstable), α=0 (fully trust the short-term window model); and the intermediate state is interpolated linearly or nonlinearly.

[0074] (4) After receiving the long and short time window model weights (W_long and W_short) from each park, the cloud server calculates the park-specific α value and applies the formula W = α. W_long + (1-α) The W_short values ​​are fused to obtain the weight parameters W of a customized prediction model for the park.

[0075] (5) The fused weight W is sent to the corresponding park for local prediction.

[0076] In this implementation, this adaptive fusion mechanism significantly enhances the predictive model's adaptability to the intermittency and volatility of renewable energy. The effects are: under stable weather conditions, the prediction results are smooth and consistent with long-term trends; under abrupt weather changes, the prediction results can quickly track instantaneous load changes, significantly reducing prediction errors. In particular, it avoids the prediction lag problem of traditional models when photovoltaic output changes abruptly, providing a more reliable time window and decision-making basis for subsequent precise scheduling.

[0077] In one or more embodiments of this application, the elastic energy consumption budget in step S400 is a mechanism with upper and lower boundaries and dynamically adjustable width. Its decision logic is as follows: the system monitors the operating status of each park's microgrid in real time, continuously calculates its current surplus photovoltaic power and the adjustable capacity of the energy storage system; if the park's energy consumption prediction data exceeds the budget limit, but the system determines that it currently has surplus photovoltaic power, then the strip widening strategy is automatically triggered, temporarily relaxing the elastic energy consumption budget to the park's budget limit constraint value, and giving priority to generating instructions to guide the consumption of local surplus photovoltaic power; if the prediction data exceeds the budget limit and the park's microgrid is in a state of insufficient photovoltaic output or dependent on grid power supply, then the tightening strategy is automatically triggered, temporarily tightening the elastic energy consumption budget to less than or equal to 70% of the park's budget limit constraint value, and generating instructions to reduce the operating intensity of energy-consuming equipment in the park.

[0078] It's understandable that surplus photovoltaic power refers to the remaining power after subtracting the current total load within the park from the current photovoltaic power generation. A positive surplus power indicates that green electricity is available for consumption; a negative surplus power indicates that electricity needs to be purchased from the grid or that energy storage needs to be utilized. The strip widening / tightening strategy is a control strategy that dynamically adjusts the intensity of budget constraints based on energy supply and demand. Widening means relaxing restrictions and encouraging electricity consumption; tightening means strengthening restrictions and requiring energy conservation.

[0079] The principle behind this embodiment is explained as follows: (1) Mechanism Settings: In the system management interface, the administrator sets a basic energy consumption budget target value (M) and a maximum allowable fluctuation percentage (Δ%). This forms an initial elastic band [M]. (1-Δ%), M (1+Δ%).

[0080] (2) Real-time monitoring and decision-making: Calculate surplus power: Photovoltaic surplus power = Photovoltaic power generation - Current total load.

[0081] (3) Decision-making logic: Scenario 1 (Expansion): IF (Predicted energy consumption > Budget cap) AND (PV surplus power > Threshold) THEN: Trigger the "expansion" strategy. The system temporarily adjusts the budget cap to the budget cap + budget expansion increment ΔE (ΔE is an increment proportional to surplus power) and generates instructions to guide the system to use this "extra" budget to consume surplus PV, such as: "Increase energy storage charging power to the maximum value" or "Increase the upper limit of electric vehicle charging power by 20%".

[0082] Scenario 2 (Tightening): IF (Predicted energy consumption > Budget limit) AND (PV surplus power < 0) THEN: Trigger the "tightening" strategy system to strictly enforce or even strengthen budget constraints, such as tightening the effective budget limit to 70% of the original limit and generating energy-saving instructions, such as: "Raise the air conditioning temperature in the office area by 1°C" and "Turn off the landscape lighting in non-main passageways".

[0083] Strategy execution: The generated instructions are sent to the S500 step for execution.

[0084] In this implementation, a flexible energy budget mechanism makes energy management strategies "intelligent" and "flexible." It avoids wasting green energy by rigidly adhering to budgets during periods of abundant sunshine, while also ensuring decisive energy-saving measures can be taken during energy shortages. Ultimately, throughout its lifecycle, the system significantly increases the penetration rate of renewable energy and reduces carbon emissions while adhering to overall energy control targets, achieving a win-win situation for both economic and environmental benefits.

[0085] In one or more embodiments of this application, in step S500, before the energy-saving control instruction set is sent to a certain park for execution, a centralized instruction conflict arbitration and verification step is introduced: This step traces all instruction receiving devices based on the digital twin of the park. When it is detected that the same device is assigned two or more instructions with logical contradictions or mutually exclusive operations within the same time window, an arbitration rule is triggered. The arbitration rule is based on a preset energy consumption sensitivity ranking knowledge base. The knowledge base quantifies and ranks the sensitivity of all types of controlled devices and their control instructions to the global energy consumption. The arbitration engine prioritizes the execution of the instruction that contributes the most to the global energy saving effect and has the highest sensitivity, while the conflicting secondary instructions are temporarily stored in a cache queue and delayed until the next control cycle or until the preconditions change before being evaluated and sent out.

[0086] Understandably, command conflict refers to a situation in automated systems where two or more independent control logics or strategies, based on different optimization objectives (such as comfort and energy saving, demand response), issue contradictory or mutually exclusive operational commands to the same actuator (such as an air conditioner) within the same time period (e.g., one command requires "cooling" while another requires "heating"). Arbitration rules are a predefined set of decision-making logic used to resolve disputes. They specify the criteria for selecting the optimal command from multiple candidate commands when a conflict occurs. Energy sensitivity ranking is a quantitative indicator used to assess the sensitivity and contribution of different devices or control commands to global energy consumption. For example, turning off a large centrifugal chiller has an energy-saving effect several orders of magnitude greater than turning off an LED light; the former's energy sensitivity is far higher than the latter.

[0087] The principle behind this embodiment is explained as follows: (1) Knowledge base construction: This needs to be completed jointly by domain experts and data analysts to assign a "sensitivity score" to all controllable device types in the system and their typical control commands (such as adjusting the air conditioner temperature by 1°C, dimming the lighting to 70%, and limiting the charging pile power to 50%). This score can be calculated based on historical data statistics (how many kilowatt-hours of electricity can be saved per hour on average after the command is executed) or physical models. The higher the score, the higher the energy-saving priority of the command.

[0088] (2) Conflict detection: At the instruction scheduling layer, an instruction-device mapping table is maintained. Before an instruction is issued, the system checks whether the same device has been assigned other instructions in the upcoming control cycle. If so, a conflict detection algorithm is started to determine whether these instructions are logically mutually exclusive (e.g., one instruction requires "on" while another requires "off").

[0089] (3) Arbitration Execution: Once a conflict is detected, the arbitration engine queries the knowledge base and compares the sensitivity scores of all conflicting instructions. The instruction with the highest score is selected and executed immediately. For instructions that are not selected, they are not simply discarded, but are placed in a delay queue and their triggering conditions are recorded. The system will re-evaluate whether the execution conditions of these cached instructions are still met in the next control cycle, or issue them again after the preconditions change (such as after the high-priority instructions have been executed).

[0090] (4) Logs and learning: All conflict events and arbitration results should be recorded for subsequent analysis, and may even be fed back to the knowledge base to optimize sensitivity scores.

[0091] This embodiment greatly improves the reliability and intelligence of complex automated systems through this mechanism, avoids frequent actuator actions due to receiving conflicting instructions, reduces equipment wear, extends equipment lifespan, ensures that the system can execute the most effective energy-saving actions under any circumstances, prevents mutual cancellation between optimization strategies, and guarantees the achievement of overall energy efficiency goals; it also provides crucial coordination guarantees for the collaborative work of multiple systems and strategies.

[0092] In one or more embodiments of this application, in step S500, the energy-saving control instruction set supports cross-system scenario linkage based on digital twin topology. Specifically, when the system predicts that the total load of a certain park's charging piles is about to exceed the safety threshold of its upstream transformer capacity, the control logic does not simply cut off the charging power. Instead, it first queries the digital twin to accurately locate all smart lighting circuits that are in the same transformer power supply circuit as the group of charging piles. Then, it automatically generates instructions to uniformly lower the dimming value of these lighting circuits to a preset level without affecting safe lighting, thereby reducing the total load of the transformer. After the peak load of the charging piles has passed, it automatically restores the lighting dimming value to 100%.

[0093] As we can understand, electrical topology refers to the electrical connections between various components in a power supply system (such as transformers, distribution cabinets, circuit breakers, and loads). It clearly describes the distribution path of electrical energy and the upstream and downstream relationships. Transformer capacity refers to the maximum apparent power (unit: kVA) that a transformer can safely carry over a long period of time, and it is a key bottleneck constraint of the power supply system. Flexible loads refer to loads whose operating power can be adjusted within a certain range without substantially affecting their main functions, or for which users are insensitive to power changes. Intelligent lighting is a typical example of a flexible load, where dimming within a certain range does not affect basic lighting functions.

[0094] The principle behind this embodiment is explained as follows: (1) Overload prediction: The system monitors the load rate of the transformer in real time and, in conjunction with the reservation information or historical data of the charging pile, predicts whether the total load will exceed the safety threshold (such as 95% of the transformer capacity) in a short period of time (such as 1 hour).

[0095] (2) Topology query: Once an overload risk is predicted, the system immediately queries the digital twin. Based on the electrical topology, it traces upwards to the risk transformer and then downwards to find all the power distribution circuits supplied by that transformer. It further selects the "intelligent lighting circuits" as an adjustable flexible resource pool.

[0096] (3) Strategy Formulation and Implementation: (4) Safety constraints: Ensure that the brightness of the lighting after adjustment is still higher than the minimum standard for safe lighting (e.g., 30%).

[0097] (5) Control command: The system automatically generates a command to uniformly set the dimming value of the selected lighting circuits to a preset, acceptable energy-saving level (e.g., from 100% to 70%). This reduction is calculated to be sufficient to offset the predicted overload.

[0098] (6) Command issuance: Dimming commands are issued through the control system.

[0099] (7) Recovery mechanism: The system continuously monitors the transformer load. When the peak load of the charging pile has passed and the total load has fallen below the safe level, the system automatically generates a recovery command to gradually or all at once restore the lighting dimming value to the normal level (100%).

[0100] The effects of this embodiment are remarkable: First, it solves the bottleneck problem of insufficient local power capacity in a highly cost-effective manner (nearly zero hardware cost), avoiding expensive transformer capacity expansion and saving significant expenditures. Second, it ensures the continuity of critical business operations (electric vehicle charging), avoiding the decline in user experience and revenue loss caused by direct power outages at charging stations. Finally, it fully explores and utilizes the adjustment potential of existing assets, serving as a model application for building a flexible, efficient, and resilient new power system.

[0101] In one or more embodiments of this application, step S600, which uses feedback data to update the digital twin state and local prediction model weight parameters of each park, includes: adopting a gradient truncation incremental learning strategy: when the system receives device execution feedback data, if it finds that the error between the data and the expected state of the model is greater than a preset value, the algorithm is updated; wherein, the update algorithm does not perform a global retraining of the entire federated learning model, but only performs local fine-tuning of the fully connected weights of the last layer or the last two layers of the model, using an extremely low learning rate for iteration; at the same time, the feature extraction weights of the intermediate layers of the model are kept completely frozen to prevent damage or overwriting of previously learned, well-generalized knowledge.

[0102] Incremental learning, as we understand it, refers to the ability of a machine learning model to continuously learn new knowledge from newly arriving data without forgetting existing knowledge. It is also known as continuous learning or online learning. Gradient truncation, however, does not refer to gradient clipping in optimization, but rather to cutting off or restricting the path of gradient backpropagation during model updates. That is, only allowing the error gradient to update a portion of the network's weights, while another portion of the weights is "frozen" and remains unchanged.

[0103] The principle behind this embodiment is explained as follows: (1) Model structure: Assume that the federated learning model used is a deep neural network, which includes an input layer, multiple hidden layers (intermediate layers) and an output layer (the last layer or several fully connected layers).

[0104] (2) Update trigger: The system receives feedback data after the device executes the update, compares the model's previous predicted value with the actual value, and calculates the error (such as mean square error). When the error exceeds a preset threshold (such as 10%), the incremental update process is triggered.

[0105] (3) Freeze operation: Before the backpropagation algorithm starts, set the weight parameters of all layers in the network except the last one or two fully connected layers to "untrainable" (i.e., frozen state). This means that in the subsequent optimization process, the gradients of these weights will be calculated but not updated.

[0106] (4) Local fine-tuning: Only the last few layers of weights that are not frozen are trained using new feedback data (usually a small batch) and a very low learning rate (e.g., 0.001). Since the number of parameters that need to be adjusted is greatly reduced and the learning rate is very low, this fine-tuning process is gentle and controlled.

[0107] (5) Model update: After fine-tuning, the original model weights are replaced with the updated weights. At this point, the core of the model's feature extraction remains basically unchanged, but the final decision boundary is slightly adjusted based on the new data.

[0108] This strategy in this embodiment ensures the stability and adaptability of the model during long-term deployment. It allows the model to evolve with the times, learning from new operational data and correcting potential biases to maintain predictive accuracy. More importantly, it effectively prevents "catastrophic forgetting," ensuring that the model doesn't suddenly forget valuable knowledge such as seasonal patterns learned over the past few months after learning new energy consumption patterns. This is crucial for industrial systems requiring stable 24 / 7 operation, avoiding predictive failures and scheduling chaos caused by improper model updates.

[0109] See Figure 2 In another embodiment of this application, a multi-park control system for implementing a multi-park energy consumption prediction and scheduling method based on digital twins is provided, comprising: a cloud server 20, wherein the cloud server 20 is configured with: a digital twin module 21, used to construct and maintain a unified four-level topology digital twin model containing hierarchical and electrical connection relationships based on project, building, space and equipment information of each park; an energy consumption prediction module 22, connected to the digital twin module 21, used to access energy consumption, environmental and microgrid operation data of each park, and generate energy consumption prediction data for future periods based on federated learning algorithms; a prediction and scheduling decision module 23, connected to the energy consumption prediction module 22 and the digital twin module 21, used to compare the energy consumption prediction results with a preset energy consumption budget, and generate an energy-saving control instruction set for HVAC, lighting, energy storage and charging pile systems; and an operation and maintenance linkage module 24, connected to the... The energy consumption prediction module 22 is used to automatically generate and dispatch inspection work orders to the operation and maintenance service module when abnormal equipment energy consumption is predicted. Multiple edge acquisition and control nodes 25 are deployed in each park to collect local data and upload it to the cloud server 20, as well as receive and execute the energy-saving control instruction set. The microgrid monitoring module 26 is integrated into the cloud server 20 or the multiple edge acquisition and control nodes 25 to monitor the operating status of photovoltaic, energy storage, charging piles and power distribution systems in real time, and provide the status data to the energy consumption prediction module 22 and the prediction and scheduling decision module 23. The digital twin module 21, energy consumption prediction module 22, prediction and scheduling decision module 23, operation and maintenance linkage module 24, multiple edge acquisition and control nodes 25 and microgrid monitoring module 26 work together to form a cross-park data acquisition, prediction analysis, optimization decision and closed-loop execution control system.

[0110] The multi-campus control system in this embodiment provides a high-performance, highly reliable, and scalable technological platform, enabling the implementation of all the aforementioned methods. It achieves optimized allocation of computing resources, ensures data security and business continuity, and, due to its modular design, is easy to maintain, upgrade, and expand, adapting to future business development and changes.

[0111] In one or more embodiments of this application, the operation and maintenance linkage module 24 is configured to: continuously monitor the device-level energy consumption anomaly prediction results and alarm events output by the prediction scheduling decision module 23; automatically call the work order creation function of the operation and maintenance service through the internal API interface, convert the abnormal events into executable predictive inspection work orders, and automatically allocate inspection priorities for inspection.

[0112] Understandably, coordinated operation and maintenance (O&M) is an advanced O&M strategy. It predicts future failure probabilities or performance degradation trends by monitoring and analyzing equipment status in real-time or near real-time, thus proactively scheduling maintenance activities before failures occur. An API (Application Programming Interface) is a set of predefined functions and protocols that allow different software applications to communicate and exchange data. Here, it's used to connect the energy consumption prediction system and the work order management system.

[0113] The principle behind this embodiment is explained as follows: Event Listening: This module acts as a background service, continuously subscribing to or listening to the message bus from the "Predictive Scheduling Decision Module". It is interested in the event type "Device-level Energy Consumption Anomaly Warning", which includes at least the device ID, prediction deviation value, and timestamp.

[0114] Work order generation logic: Upon receiving an alert event, the module's internal logic will make a judgment (such as whether the deviation has continuously exceeded a threshold for a certain period of time). If it is determined that a work order needs to be generated, the work order data will be prepared.

[0115] API Call: The module calls the RESTful API interface for creating work orders provided by the operation and maintenance service module (which may be an independent work order management system such as Maximo, ServiceNow, or a self-developed system), and transmits the required information in structured data (usually in JSON format). The transmitted data includes: work order title (e.g., "Air Conditioner AHU-1 Energy Consumption Anomaly Warning"), associated equipment, description information (including predicted abnormal data), priority (which can be automatically set to "high"), and suggested measures (e.g., "Check the filter, calibrate the sensor").

[0116] Work order dispatch: After receiving an API request, the operation and maintenance service module automatically creates a work order and dispatches it to the corresponding operation and maintenance team or engineer according to preset rules.

[0117] Closed-loop feedback (optional enhancement): The module can also provide a reverse interface to receive the processing result of the work order (such as "filter replaced"). This result can be used to update the device's "energy consumption baseline fingerprint" (see claim 3), forming a complete "monitoring-prediction-operation and maintenance-verification" closed loop.

[0118] This embodiment achieves proactive and intelligent operation and maintenance activities through coordinated operation and maintenance. It significantly advances the time of equipment failure detection, shifting from "reactive repair" to "proactive maintenance," thereby avoiding unplanned downtime and higher maintenance costs caused by minor issues escalating into major failures. Simultaneously, it reduces continuous energy waste caused by suboptimal or inefficient equipment operation, achieving synergistic optimization of energy conservation and asset management, and significantly improving overall operational efficiency and economic benefits.

[0119] See Figure 3 , Figure 3 This is a schematic block diagram of an electronic device provided according to an embodiment of this application. Figure 3 The electronic device 300 in this embodiment may include one or more processors 301, one or more input devices 302, one or more output devices 303, and one or more memories 304. The processors 301, input devices 302, output devices 303, and memories 304 communicate with each other via a communication bus 305. The memories 304 store computer programs, including program instructions. The processors 301 execute the program instructions stored in the memories 304. Specifically, the processors 301 are configured to invoke the program instructions to execute the aforementioned multi-campus energy consumption prediction and scheduling method based on digital twins.

[0120] It should be understood that, in the embodiments of this application, the processor 301 may be a central processing unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0121] Input device 302 may include a touchpad, a fingerprint sensor (for collecting the user's fingerprint information and fingerprint orientation information), a microphone, etc., and output device 303 may include a display (LCD, etc.), a speaker, etc.

[0122] The memory 304 may include read-only memory and random access memory, and provides instructions and data to the processor 301. A portion of the memory 304 may also include non-volatile random access memory. For example, the memory 304 may also store device type information.

[0123] In specific implementations, the processor 301, input device 302, and output device 303 described in the embodiments of this application can execute the implementation methods described in any embodiment of the multi-campus energy consumption prediction and scheduling method based on digital twins provided in the embodiments of this application, or they can execute the implementation methods of the electronic devices described in the embodiments of this application, which will not be repeated here.

[0124] In another embodiment of this application, a computer-readable storage medium is provided. This computer-readable storage medium stores a computer program, which includes program instructions. When executed by a processor, the program instructions implement all or part of the processes in the multi-campus energy consumption prediction and scheduling method based on digital twins described in the above embodiments. Alternatively, the computer program can instruct related hardware to implement these processes. The computer program can be stored in a computer-readable storage medium. When executed by a processor, the computer program can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include any entity or device capable of carrying computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0125] The computer-readable storage medium can be an internal storage unit of the electronic device in any of the foregoing embodiments, such as a hard disk or memory of the electronic device. The computer-readable storage medium can also be an external storage device of the electronic device, such as a plug-in hard disk, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device. Furthermore, the computer-readable storage medium can include both internal and external storage units of the electronic device. The computer-readable storage medium is used to store computer programs and other programs and data required by the electronic device. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0126] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0127] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the electronic devices and units described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0128] In the several embodiments provided in this application, it should be understood that the disclosed electronic devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces or units, or it may be an electrical, mechanical, or other form of connection.

[0129] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of this application, depending on actual needs.

[0130] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0131] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A multi-campus energy consumption prediction and scheduling method based on digital twins, characterized in that, Includes the following steps: S100: Data acquisition steps, which collect real-time operation data, energy consumption data and environmental parameters of various energy-consuming systems and microgrids in multiple distributed parks in parallel; S200: Digital twin construction steps, based on the collected data, to construct digital twins for each park, including four-level topological relationships of project-building-space-equipment, and to summarize them into a unified model for the entire region; S300: Federated learning prediction step, based on the global unified model, drives the federated learning process, and generates energy consumption prediction data for future periods of each park by integrating the weight parameters of the local prediction models of each park. S400: Flexible decision-making steps compare the energy consumption forecast data of each park with the flexible energy consumption budget. If the difference exceeds the preset value, a set of energy-saving control instructions is generated, which includes at least one of the following devices: energy storage, HVAC, lighting and charging piles. S500: Instruction execution steps, parse the topology relationship of equipment in each park, issue and execute the energy-saving control instruction set, and collect feedback data; S600: Closed-loop update steps, using feedback data to update the digital twin status of each park and the weight parameters of the local prediction model respectively.

2. The method according to claim 1, characterized in that, The specific process of constructing a digital twin as described in step S200 includes: Based on device access information and spatial layout diagram, the system automatically generates a project-building-space-device hierarchical path to which the device belongs to form a subordinate topology; at the same time, based on the subordinate relationship data of smart meters, circuit breakers and distribution boxes, the system automatically generates an electrical connection topology between devices; the digital twin is a fusion of the subordinate topology and the electrical connection topology.

3. The method according to claim 1 or 2, characterized in that, In step S200, when constructing digital twins for each park, a digital energy consumption benchmark fingerprint is dynamically created and bound to each device. The energy consumption benchmark fingerprint is composed of static attributes, dynamic operating parameters, and historical statistical features of the device. The static attributes include at least the equipment type, rated power, manufacturing date, and energy efficiency rating. The dynamic operating parameters are power consumption data collected after the equipment has been running continuously for at least 24 hours under normal operating conditions. The historical statistical characteristics are steady-state power consumption Gaussian distribution parameters calculated from the power consumption data, including the mean μ and variance σ. 2 .

4. The method according to claim 1 or 2, characterized in that, The federated learning process described in step S300 employs a hybrid architecture that combines long and short time windows. The long-time window model uses long-term data from the past 7 to 30 days as its training set; the short-time window model uses recent data from the past 1 to 24 hours as its training set. An adaptive weight fusion mechanism is designed in the cloud server, and its formula is expressed as follows: W = α W_long + (1-α) W_short; Where W represents the weight parameters of the fused prediction model; W_long represents the weights of the long-term window model, which includes long-term (7 to 30 days) seasonal and cyclical trend characteristics; W_short represents the weights of the short-window model, which includes short-term (1 to 24 hours) instantaneous fluctuations and abrupt changes. α is an adaptive weighting factor, a dynamically changing coefficient with a value range of [0, 1]. Its magnitude is determined by the variance of photovoltaic power fluctuation. If the variance is large, α is reduced to make the model more trusting of the short-window model; if the variance is small, α is increased to make the model more trusting of the long-window model.

5. The method according to claim 1 or 2, characterized in that, The flexible energy consumption budget mentioned in step S400 is a mechanism with upper and lower boundaries and a dynamically adjustable width. Its decision logic is as follows: The system monitors the operating status of each park's microgrid in real time, continuously calculates its current surplus photovoltaic power and the adjustable capacity of the energy storage system; if the park's energy consumption prediction data exceeds the budget limit, but the system determines that there is currently surplus photovoltaic power, it automatically triggers a strip widening strategy, temporarily relaxes the flexible energy consumption budget to the park's budget limit constraint value, and prioritizes generating instructions to guide the consumption of local surplus photovoltaic power; if the prediction data exceeds the budget limit and the park's microgrid is in a state of insufficient photovoltaic output or dependent on grid power supply, it automatically triggers a tightening strategy, temporarily tightens the flexible energy consumption budget to less than or equal to 70% of the park's budget limit constraint value, and generates instructions to reduce the operating intensity of energy-consuming equipment in the park.

6. The method according to claim 1 or 2, characterized in that, In step S500, before the energy-saving control instruction set is sent to a certain park for execution, a centralized instruction conflict arbitration and verification step is introduced: This step traces all instruction receiving devices based on the digital twin of the park. When it is detected that the same device is assigned two or more instructions with logical contradictions or mutually exclusive operations in the same time window, the arbitration rule is triggered. The arbitration rule is based on a set of preset energy consumption sensitivity ranking knowledge base. The knowledge base quantifies and ranks the sensitivity of all types of controlled devices and their control instructions to the impact on global energy consumption. The arbitration engine prioritizes the execution of the instruction that contributes the most to the global energy saving effect and has the highest sensitivity, while the conflicting secondary instructions are temporarily stored in the cache queue and delayed until the next control cycle or until the preconditions change before being evaluated and sent.

7. The method according to claim 1 or 2, characterized in that, In step S500, the energy-saving control instruction set supports cross-system scenario linkage based on digital twin topology. Specifically, when the system predicts that the total load of a certain park's charging piles is about to exceed the safety threshold of its upstream transformer capacity, the control logic does not simply cut off the charging power. Instead, it first queries the digital twin to accurately locate all smart lighting circuits that are in the same transformer power supply circuit as the group of charging piles. Then, it automatically generates instructions to uniformly lower the dimming value of these lighting circuits to a preset level without affecting safe lighting, thereby reducing the total load on the transformer. After the peak load of the charging piles has passed, it automatically restores the lighting dimming value to 100%.

8. The method according to claim 1 or 2, characterized in that, In step S600, the feedback data is used to update the state of the digital twins of each park and the weight parameters of the local prediction model, including: adopting a gradient truncation incremental learning strategy: when the system receives the device execution feedback data, if it finds that the error between the data and the expected state of the model is greater than a preset value, the algorithm is updated; wherein, the update algorithm does not perform a global retraining of the entire federated learning model, but only performs local fine-tuning on the fully connected weights of the last layer or the last two layers of the model, using an extremely low learning rate for iteration; at the same time, the feature extraction weights of the intermediate layers of the model are kept completely frozen to prevent damage or overwriting of previously learned, well-generalized knowledge.

9. A multi-campus control system based on digital twins for implementing the method of any one of claims 1-8, characterized in that, include: A cloud server, wherein the cloud server is configured with: The digital twin module is used to build and maintain a unified four-level topology digital twin model that includes hierarchical and electrical connection relationships based on project, building, space and equipment information in each park. The energy consumption prediction module is connected to the digital twin module and is used to access the energy consumption, environmental and microgrid operation data of each park, and generate energy consumption prediction data for future periods based on the federated learning algorithm. The predictive scheduling decision module is connected to the energy consumption prediction module and the digital twin module. It is used to compare the energy consumption prediction results with the preset energy consumption budget and generate a set of energy-saving control instructions for HVAC, lighting, energy storage and charging pile systems. The operation and maintenance linkage module is connected to the energy consumption prediction module and is used to automatically generate and dispatch inspection work orders to the operation and maintenance service module when abnormal energy consumption of equipment is predicted. Multiple edge acquisition and control nodes are deployed in each park to collect local data and upload it to the cloud server, as well as to receive and execute the energy-saving control instruction set. The microgrid monitoring module is integrated into the cloud server or the multiple edge acquisition and control nodes. It is used to monitor the operating status of photovoltaic, energy storage, charging piles and power distribution systems in real time, and provide the status data to the energy consumption prediction module and the prediction and scheduling decision module. The digital twin module, energy consumption prediction module, prediction scheduling decision module, operation and maintenance linkage module, microgrid monitoring module, and multiple edge acquisition and control nodes work together to form a cross-park data acquisition, prediction analysis, optimization decision and closed-loop execution control system.

10. The control system according to claim 9, characterized in that, The operation and maintenance linkage module is configured to: continuously monitor the device-level energy consumption anomaly prediction results and alarm events output by the prediction scheduling decision module; automatically call the work order creation function of the operation and maintenance service module through the internal API interface, convert the abnormal events into executable predictive inspection work orders, and automatically assign inspection priorities for inspection.

Images