Energy efficiency evaluation method, system, device and medium based on digital twinning

Abstract

This invention discloses a method, system, device, and medium for energy efficiency assessment based on digital twins, specifically relating to the fields of digital twins and energy efficiency assessment technology. The method includes: collecting multi-source data from energy-consuming systems, generating time-series datasets, and constructing equipment topology diagrams and operating condition labels; constructing a hybrid twin model pool comprising high-fidelity mechanisms, lightweight data-driven models, and reduced-order models; generating a fusion model and confidence sequence through weighted fusion using a dynamic scheduler; performing root cause diagnosis based on causal graphs and counterfactual reasoning; encapsulating the fusion model into a reinforcement learning environment, offline training a policy network to output the optimal action sequence and predicted energy efficiency gain curve; calculating a comprehensive energy efficiency assessment index to determine whether to issue control commands, and judging whether to update the model based on confidence and prediction errors. This invention achieves high-precision energy efficiency assessment, root cause diagnosis, and autonomous optimization control of energy-consuming systems through dynamic fusion of multiple models and causal inference.

Description

Technical Field

[0001] This invention relates to the field of digital twin and energy efficiency assessment technology, and more specifically, to energy efficiency assessment methods, systems, devices and media based on digital twins. Background Technology

[0002] With the rapid development of industrial IoT, big data analytics, and artificial intelligence technologies, energy efficiency assessment methods based on digital twins have gradually become a research hotspot in fields such as industrial energy systems and building energy systems. Currently, existing energy efficiency assessment methods can be mainly divided into three categories: First, physical mechanism-based modeling methods, which construct mechanistic models of equipment and systems through thermodynamic and fluid dynamic equations, providing high-precision steady-state simulation results, but with high computational complexity and difficulty adapting to rapidly changing dynamic operating conditions; second, data-driven modeling methods, which directly learn input-output mapping relationships from historical data using algorithms such as neural networks and support vector machines, offering advantages such as fast response speed and no need for in-depth understanding of physical mechanisms, but heavily relying on the coverage of training data and lacking generalization ability when extrapolating operating conditions; and third, single-model-based assessment methods, which rely solely on mechanistic models or data-driven models for energy efficiency calculation and diagnosis, making it difficult to balance accuracy and real-time performance across the entire operating condition range.

[0003] However, in practical use, it still has some shortcomings. For example, a single model cannot simultaneously meet the requirements of steady-state high accuracy and dynamic high response, and the evaluation accuracy decreases in scenarios with frequent changes in operating conditions. Most existing methods are based on correlation analysis for energy efficiency anomaly diagnosis, lacking causal inference ability, resulting in poor interpretability of diagnostic results, difficulty in accurately locating root causes and providing actionable optimization suggestions. Traditional energy efficiency optimization strategies mostly rely on manual experience or offline optimization, which cannot achieve an autonomous closed loop of "evaluation-diagnosis-optimization-control". Moreover, the optimization process requires trial and error on real systems, which poses safety risks and high implementation costs. Most existing digital twin models are statically constructed and lack an adaptive update mechanism based on model confidence and prediction error, making it difficult to cope with long-term changes such as equipment aging and operating condition drift, resulting in a gradual decrease in model accuracy over time. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a digital twin-based energy efficiency assessment method, system, device and medium, which solves the problems mentioned in the background art through the following solutions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an energy efficiency assessment method based on digital twins, characterized in that it includes:

[0006] S1: Deploy sensors and controllers at key energy efficiency nodes of the target energy system through an industrial IoT architecture to collect energy efficiency-related parameters, operating condition-related parameters, and equipment status parameters. Generate a standardized time-series dataset after preprocessing. Construct an equipment topology diagram based on the physical connection and control hierarchy of the equipment. Use a change point detection algorithm to identify operating condition switching events and generate a set of operating condition event tags with semantic labels.

[0007] S2: Based on standardized time-series datasets, equipment topology diagrams, and operating condition event label sets, high-fidelity mechanism models, lightweight data-driven models, and reduced-order models are constructed respectively to form a hybrid twin model pool;

[0008] S3: Construct a dynamic scheduler to calculate the fusion weights in real time based on the rate of change of operating conditions and the prediction residuals of each model. The high-fidelity mechanism model, the lightweight data-driven model, and the reduced-order model are weighted and fused to generate a fused twin model. The weighted residuals between the fused prediction value and the prediction values ​​of each sub-model are calculated, and a twin confidence sequence is generated based on the historical reference residuals.

[0009] S4: Construct a structural cause-effect graph using the equipment topology diagram as the skeleton. When the energy efficiency index is lower than the preset threshold, use the fusion twin model as the counterfactual reasoning engine to perform intervention analysis on each controllable variable, calculate the predicted energy efficiency gain after intervention, generate the counterfactual energy efficiency benchmark, and generate a root cause diagnosis report based on the maximum gain and twin confidence.

[0010] S5: The fusion twin model is encapsulated as a reinforcement learning environment simulator. The state space, action space and multi-objective reward function are defined. The policy network is trained offline using the proximal policy optimization algorithm. The optimal action sequence is output. The predicted energy efficiency gain curve is generated through forward simulation of the fusion twin model.

[0011] S6: Calculate the comprehensive energy efficiency assessment index and compare it with the set threshold to determine whether to issue a closed-loop control command; at the same time, based on the twin confidence level and the model prediction error, determine whether to trigger a model update. If triggered, feed back to S2 to update the hybrid twin model pool.

[0012] A digital twin-based energy efficiency assessment system is characterized by comprising:

[0013] The data acquisition and feature extraction module is configured to collect multi-source data from the energy-consuming system, generate a standardized time-series dataset after cleaning and alignment, construct a device topology diagram, and generate a set of operating condition event labels through change point detection.

[0014] The hybrid twin model pool construction module is configured to construct a high-fidelity mechanism model, a lightweight data-driven model, and a reduced-order model based on the standardized time-series dataset, equipment topology diagram, and operating condition event label set, forming a hybrid twin model pool;

[0015] The dynamic fusion and confidence generation module is configured to build a dynamic scheduler, calculate the fusion weight in real time based on the rate of change of operating conditions and the prediction residuals of each model, and generate a fusion twin model by weighted fusion of the high-fidelity mechanism model, the lightweight data-driven model and the reduced-order model, and generate a twin confidence sequence based on the weighted residuals of the fusion prediction value and the prediction values ​​of each sub-model.

[0016] The causal inference and root cause diagnosis module is configured to construct a structural causal graph with the device topology diagram as the skeleton. When the energy efficiency index is lower than the preset threshold, the fusion twin model is used to perform intervention analysis on each controllable variable, calculate and predict the energy efficiency gain and generate a root cause diagnosis report.

[0017] The optimized strategy generation module is configured to encapsulate the fusion twin model as a reinforcement learning environment simulator, define the state space, action space and multi-objective reward function, train the policy network offline to output the optimal action sequence, and generate a predicted energy efficiency gain curve through forward simulation.

[0018] The closed-loop evaluation and adaptive triggering module is configured to calculate the comprehensive energy efficiency evaluation index and compare it with a set threshold to determine whether to issue a closed-loop control command. At the same time, it determines whether to trigger a model update based on the twin confidence and model prediction error. If triggered, it feeds back to the hybrid twin model pool construction module to update the model pool, forming a continuous closed loop between the physical system and the digital twin.

[0019] This application provides an electronic device that adopts the following technical solution: it includes a processor, a memory, a user interface, and a network interface. The memory is used to store instructions, the user interface and the network interface are used to communicate with other devices, and the processor is used to execute the instructions stored in the memory to enable the electronic device to execute a computer program of any of the above-described digital twin-based energy efficiency assessment methods.

[0020] This application provides a computer-readable storage medium that stores a computer program capable of being loaded by a processor and executing any of the above-mentioned energy efficiency assessment methods based on digital twins.

[0021] The technical effects and advantages of this invention are as follows:

[0022] 1. This invention constructs a hybrid twin model pool that includes a high-fidelity mechanism model, a lightweight data-driven model, and a reduced-order model, and introduces a dynamic scheduler based on the rate of change of operating conditions and the prediction residual to achieve weighted fusion. It can ensure high-precision simulation under steady-state operating conditions using the mechanism model, achieve rapid response under transient operating conditions using the data-driven model, and smoothly connect in the transition region using the reduced-order model. Thus, it balances the accuracy and real-time performance of energy efficiency assessment across the entire operating condition range, solving the technical problem that a single model is difficult to adapt to dynamic operating conditions.

[0023] 2. This invention constructs a structural causal graph using the equipment topology diagram as the framework, and uses a fused twin model as a counterfactual reasoning engine to perform intervention analysis on each controllable variable and calculate energy efficiency gains. Combined with twin confidence, it generates a quantifiable root cause diagnosis report, realizing a technological leap from correlation analysis to causal inference, improving the interpretability of energy efficiency anomaly diagnosis and the accuracy of root cause location, and providing operation and maintenance personnel with clear optimization directions and expected benefits.

[0024] 3. This invention encapsulates the fusion twin model into a reinforcement learning environment simulator, uses a proximal policy optimization algorithm to train the policy network offline, outputs the optimal action sequence and generates a predicted energy efficiency gain curve, realizing an autonomous closed loop of "evaluation-diagnosis-optimization-control", avoiding trial-and-error control on real equipment, reducing the safety risks and costs of optimization implementation, and improving the efficiency and feasibility of energy efficiency optimization.

[0025] 4. This invention calculates the twin confidence sequence and constructs a model update triggering mechanism based on the twin confidence and model prediction error. It can automatically trigger the update of the hybrid twin model pool when the model confidence is insufficient or the prediction deviation exceeds the limit, so that the digital twin model has self-evolution capability, effectively copes with long-term changes such as equipment aging and operating condition drift, and ensures the continuous reliability of energy efficiency assessment and optimization control throughout the entire life cycle.

[0026] 5. This invention achieves quantitative assessment and visual presentation of energy efficiency status through the calculation of comprehensive energy efficiency evaluation index and graded early warning mechanism. Combined with the output of root cause diagnosis report and optimization action sequence, it provides complete data support for operation and maintenance decision-making, forming a continuous closed loop between physical system and digital twin, and improving the overall energy efficiency management level of energy-using system. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0028] Figure 2 This is a schematic diagram of the structure of the energy efficiency assessment method based on digital twins of the present invention.

[0029] Figure 3 This is a schematic diagram of the system module structure of the present invention.

[0030] Figure 4 This is a schematic diagram of the electronic device structure of the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] refer to Figures 1-2 The energy efficiency assessment method based on digital twins shown includes:

[0033] S1: Multi-source heterogeneous data acquisition and feature extraction

[0034] By deploying multiple types of sensors and controllers at key energy efficiency nodes, raw data in three categories—energy efficiency-related parameters, operating condition-related parameters, and equipment status parameters—are collected according to a preset sampling frequency. Then, a standardized time-series dataset is generated through outlier removal, missing value imputation, and time alignment. Simultaneously, a directed topology graph is constructed based on the physical connections and control hierarchy of the equipment. Finally, a change point detection algorithm combined with domain expert rules is used to identify operating condition switching events, generating event sequences with semantic labels. Specifically, this includes the following:

[0035] S101: Multi-sensor and controller deployment and data acquisition: Deploy measurement points on key energy efficiency nodes of the physical system through an industrial IoT architecture and perform data acquisition operations. The specific implementation is as follows:

[0036] Measurement point layout and parameter configuration: Based on the process flow diagram and energy flow topology analysis results of the energy-consuming system, the system deploys corresponding sensors and configures the acquisition parameters at the following locations:

[0037] Smart meters are installed in the power supply circuits of major energy-consuming equipment (chillers, water pumps, cooling towers, air compressors, fans, boilers, etc.) to collect active power data. (Unit: kW) Current (Unit: A) Voltage (Unit: V). During installation, ensure that the wiring of the current transformer and voltage transformer conforms to the equipment specifications, and complete the input of the transformation ratio parameters.

[0038] Electromagnetic or ultrasonic flow meters are installed at key pipeline nodes such as chilled water mains, cooling water mains, steam mains, and compressed air mains to collect media flow data. (Unit: m³ / h). During installation, select a sensor of appropriate size according to the pipe diameter, and complete the input of pipe parameters (material, wall thickness, inner diameter). For ultrasonic flow meters, transducer installation spacing calibration is also required.

[0039] Platinum resistance temperature sensors are installed at the inlet and outlet of heat exchangers, the inlet and outlet of evaporators and condensers in chillers, and the inlet and outlet of boilers to collect temperature data. (Unit: °C). During installation, ensure that the sensor probe is in full contact with the measured medium, and complete the range configuration and zero-point calibration of the temperature transmitter.

[0040] Pressure transmitters are installed at key pipeline nodes such as pump outlet, compressor exhaust port, and before and after filters to collect pressure data. (Unit: kPa). During installation, ensure that the pressure tap location meets the fluid dynamics requirements (avoiding areas prone to disturbance such as elbows and valves), and complete the range configuration and zero-point calibration.

[0041] Read the load rate through the communication interface of the device controller (PLC). (Value) ) and start / stop status (Boolean value). The system establishes periodic communication with the PLC via Modbus TCP / IP or OPC UA protocol, sets the reading cycle to 1 second, and parses the load rate value and start / stop status flag bit in the register address.

[0042] Install outdoor temperature and humidity sensors in an unobstructed, well-ventilated outdoor location to collect ambient temperature data. (Unit: °C) and ambient humidity (Unit: %). During installation, ensure that the sensor is kept away from interference locations such as air conditioner outdoor unit exhaust vents and roof heat radiation sources, and confirm the sensor's protection level.

[0043] Production cycle time is read from the production line PLC via industrial Ethernet. (Unit: pieces / hour) Set the data synchronization cycle to be consistent with the energy efficiency data acquisition cycle for subsequent load correlation analysis.

[0044] Read fault codes from the device controller or fault diagnosis system. (Integer) Set up an event-triggered reading mechanism to immediately record the time of change and the fault code value when the fault code changes.

[0045] Synchronize maintenance timestamps from the enterprise asset management system (Unix timestamp), set to synchronize once a day, recording the time of the most recent maintenance.

[0046] Data Acquisition and Transmission: Perform the following acquisition operations through the edge gateway:

[0047] Configure the corresponding channels of the edge gateway according to the physical interface type of each sensor (analog 4-20mA, RS485 Modbus, Ethernet, etc.) and complete the initialization of signal acquisition parameters (such as sampling frequency, range mapping, and data format).

[0048] Set a differentiated sampling frequency: for active power Current ,flow Start-stop status Configure sampling frequency for rapidly changing quantities. That is, a data acquisition is triggered every 0.1 seconds; for temperature... ,pressure Ambient temperature Ambient humidity Load rate Production cycle time Equal slow variables, configure sampling frequency That is, a data collection is triggered once per second; for fault codes Maintenance timestamp Quasi-static quantities are collected using event-triggered acquisition (actively reported when a fault occurs) or synchronized once daily.

[0049] The edge gateway periodically sends read commands to each sensor according to the configured sampling frequency and receives the returned measurement values. For analog sensors, the edge gateway converts the 4-20mA current signal into the corresponding physical quantity value according to the preset range; for digital sensors, it directly parses the data frames in the communication protocol.

[0050] Each data collection record is appended with a timestamp accurate to milliseconds and a device ID by the edge gateway, forming the original multimodal time-series dataset. The data is encapsulated in JSON format and written to a specified measurement table in the local time-series database in real time via the MQTT protocol. At the same time, a data cache queue is set up to cope with network interruptions and ensure that no data is lost.

[0051] S102: Data Cleaning and Time Alignment: For the collected raw data, outliers need to be removed, missing values ​​need to be filled, and data of different frequencies need to be aligned to a unified time base to ensure the accuracy and consistency of subsequent analysis. The specific implementation is as follows:

[0052] Outlier removal: For each parameter First, a physical feasibility check is performed: if the measured value exceeds the sensor's range or is physically impossible (e.g., active power). <0 and the device is in operation, or the temperature Values ​​exceeding 300℃ are directly marked as abnormal and discarded. For measurements within the physical range, the following applies: The principle is to calculate the mean and standard deviation within the moving window. Let the window length be... =60 sampling points (corresponding to a 6-second window for rapidly changing variables and a 60-second window for slowly changing variables), calculate:

[0053]

[0054] If the current value If a peak jump occurs, it is considered an outlier and removed. This statistical method can effectively identify spikes caused by occasional sensor malfunctions or electromagnetic interference.

[0055] Missing and Removed Value Handling: For missing values ​​caused by removal or transmission interruption, if the consecutive missing time... Linear interpolation is used for filling. Let the effective values ​​before and after the missing segment be... (time ), (time Then the interpolation formula is:

[0056]

[0057] If consecutive missing time If the time interval is greater than 30 seconds, the entire time interval will be marked as "invalid data" and excluded in subsequent modeling to avoid introducing false information.

[0058] Time alignment: Due to the different sampling frequencies of various parameters, all parameters need to be resampled to a unified time base to facilitate subsequent time series analysis and model training. A unified timestamp sequence is selected. Time interval (This can be adjusted to 0.5s or 2s depending on the system's dynamic characteristics). For each parameter The original discrete measurements are mapped to a unified timestamp through linear interpolation (for Boolean values ​​such as start / stop status). Nearest neighbor interpolation is used to prevent ambiguity in state transitions. After alignment, a standardized time-series dataset is generated. Its data structure is as follows:

[0059]

[0060] in For the first A vector of all parameter values ​​at each time step, where mm represents the total number of parameters. Each parameter also stores its physical unit to ensure that all subsequent calculations strictly adhere to dimensional consistency.

[0061] S103: Construction of Equipment Topology Diagram

[0062] Energy, material, and control flows exist between devices in an energy system. These relationships are crucial for subsequent causal analysis and graph neural network modeling. This step extracts all physical devices as graph nodes based on the system process flow diagram and electrical single-line diagram. Each node... Record the following attributes: Equipment ID (unique identifier), Equipment type (e.g., "centrifugal chiller", "fixed frequency water pump", "variable frequency cooling tower", "plate heat exchanger", etc.), Rated power. (Unit: kW), Rated Flow Rate (Unit: m³ / h), control level (e.g., "field level", "unit level", "system level").

[0063] Based on the direction of energy flow, material flow, or control signal transmission, directed edges are established between nodes, and the construction rules are as follows:

[0064] Energy flow: if the equipment To the equipment If electricity, heat, or mechanical energy is provided, then a directed edge is added. .

[0065] Material flow: If the medium (water, air, steam) flows from the equipment Flow to equipment Then add directed edges. .

[0066] Control flow: if the controller Send control commands to the actuator Then add directed edges. .

[0067] Each edge Record attributes: Medium type ("Electricity", "Water", "Wind", "Steam"), pipe diameter / conductor cross-sectional area, design flow rate / current value. The final product is a diagram of the equipment topology. It is stored in the form of a graph database (such as Neo4j) or an adjacency list for subsequent graph neural network modeling and causal graph construction.

[0068] S104: Generation of Operating Condition Event Tag Set

[0069] In order to identify the switching of operating conditions during system operation and provide a basis for subsequent model segmented training and dynamic scheduling, this step uses a change point detection algorithm combined with expert rules to generate a set of operating condition event labels.

[0070] Several parameters sensitive to changes in operating conditions are selected to form an feature vector. Typical sensitive parameters include:

[0071] Total power Load rate Start-up and shutdown status of main equipment (0 / 1) Total system traffic .from Extract these parameters in the time interval The sequence on is denoted as .

[0072] The Pruned Exact Linear Time (PELT) algorithm is used to identify mutation points in sequences. This algorithm minimizes the following objective function:

[0073]

[0074] in: For changing point positions; The cost function is the intra-segment cost function, and the negative log-likelihood is used for Gaussian distributed sequences:

[0075]

[0076] in For segment variance estimation;

[0077] The penalty coefficient controls the number of variables and is usually determined by the Bayesian Information Criterion (BIC) or cross-validation.

[0078] The algorithm outputs the set of all time points of change. For each detected change point and its corresponding interval, event type labeling is performed using domain knowledge:

[0079] Equipment startup: Before and after the change point From 0 to 1, and A step jump is performed, simultaneously confirming the associated device ID;

[0080] Equipment shutdown: before and after the change point From 1 to 0, and Step descent;

[0081] Load increase: before and after the change point The rise exceeds the threshold (e.g.) And the duration is >10 seconds;

[0082] Load Decrease: Before and After the Shift Point The drop exceeds the threshold (e.g.) );

[0083] Fault event: Corresponding time It can directly read the start and end times of the event from the fault code and associate them with the specific device.

[0084] After annotation is completed, a product operating condition event tag set is generated. Its data structure is as follows:

[0085]

[0086] Each event Includes the field: Event start time Event End Time (If it is a transient event, then) = ), event type (e.g., "Startup", "Stop", "Load Increase", "Fault"), associated device ID, event confidence level (range of values) ).

[0087] S2: Hybrid Twin Model Pool Construction

[0088] A hybrid twin model pool is constructed, comprising a high-fidelity mechanistic model, a lightweight data-driven model, and a reduced-order model, providing a multimodal model foundation for subsequent dynamic fusion and energy efficiency assessment. Through the complementary characteristics of the three models, both accuracy and real-time performance can be balanced under different operating conditions. Specifically, this includes the following:

[0089] S201: Construction of a High-Fidelity Mechanism Model

[0090] Based on the first law of thermodynamics, fluid mechanics equations, and equipment characteristic curves, a high-fidelity mechanism model is established for key energy-consuming equipment. Taking a chiller unit as an example, its coefficient of performance (COP) is defined as the ratio of cooling capacity to compressor input power:

[0091]

[0092] in For heat exchange in the evaporator, The compressor receives power input. Evaporator heat exchange is based on the energy balance between the refrigerant side and the chilled water side.

[0093]

[0094] here This refers to the mass flow rate of chilled water. This refers to the specific heat capacity of chilled water. , The inlet and outlet water temperatures of the evaporator are given. The compressor power model is based on the fitting of the compressor characteristic curve and is usually expressed as a function of refrigerant mass flow rate and pressure difference.

[0095]

[0096] in This is the refrigerant mass flow rate. Compressor pressure differential (unit: kPa). , , These are coefficients calibrated using equipment factory data or on-site testing.

[0097] use Parameter identification was performed using data from the intermediate steady-state range. The least squares method was used to determine the model parameters, minimizing the sum of squared errors between the model output and the measured values. This yielded a high-fidelity mechanism model for the product. The output is a system of differential equations and a parameter set. This model exhibits high accuracy under steady-state conditions, but its computation time is relatively long (on the order of seconds), making it suitable for offline analysis and benchmark calculations.

[0098] S202: Lightweight Data-Driven Model Building

[0099] To achieve rapid response in dynamic processes, select Dynamic process data, with The marked working conditions are used as the basis for segmentation. A graph neural network (GNN) model is trained, and... As a graph structure constraint, the model input consists of the key parameter vectors within the current time step and the historical window. The output is the predicted value of the energy efficiency index for the next time step.

[0100] Using a graph convolutional network with an attention mechanism, the node feature update formula is:

[0101]

[0102] in For nodes In the The feature vector of the layer, For nodes The set of neighboring nodes, Attention coefficient For learnable weight matrix, The activation function is ReLU. The network structure consists of three graph convolutional layers, with output dimensions of 64, 32, and 16 respectively, and finally outputs the energy efficiency prediction value through a fully connected layer.

[0103] The training objective is to minimize the mean squared error between the predicted and actual values.

[0104]

[0105] in Energy efficiency indicators (such as COP) predicted by the model. This is a measured energy efficiency indicator. The Adam optimizer was used with a learning rate of 0.001 and 500 training epochs, with an early stopping mechanism to prevent overfitting. The resulting lightweight data-driven model was then obtained. The output is a set of neural network weights, with an inference latency of <10ms, making it suitable for real-time online prediction.

[0106] S203: Construction of Reduced-Order Model

[0107] To improve computational speed while maintaining the accuracy of the high-fidelity mechanistic model, Intrinsic Orthogonal Decomposition (POD) is performed for dimensionality reduction. First, simulation results of the high-fidelity mechanism model under typical operating conditions (including equipment start-up and shutdown, load changes, and ambient temperature changes) are collected to form a snapshot matrix. ,in For state variables (such as temperature, pressure, flow rate, etc. of each node). This represents the number of snapshots. Perform singular value decomposition:

[0108]

[0109] Take before The projection matrix is ​​composed of left singular vectors. , Determined by energy retention rate:

[0110]

[0111] in It is a singular value. Set a preset energy retention threshold (0.99). Set the original state variables... Projected into a lower-dimensional space Establish state equations in low-dimensional space:

[0112]

[0113] in For the nonlinear function of the original high-fidelity model, Input variables (such as load rate, ambient temperature, etc.).

[0114] This low-dimensional model is obtained using the Galerkin projection method, with a computational complexity of […]. Downgraded to The product order reduction model is obtained. The output is a reduced-dimensional projection matrix and a low-dimensional state equation, which can provide simulation capabilities with accuracy close to the mechanism model and speed close to the data-driven model in steady-state and dynamic transition processes.

[0115] S3: Dynamic Fidelity Fusion and Twin Confidence Generation

[0116] A dynamic scheduler is constructed to adjust the fusion weights of the high-fidelity mechanism model, lightweight data-driven model, and reduced-order model in real time based on the rate of change of operating conditions and the prediction residuals of each model. This generates a fused twin model and a twin confidence sequence, providing a high-precision and high-reliability digital twin foundation for subsequent energy efficiency assessment and decision-making. Specifically, this includes the following:

[0117] S301: Dynamic Scheduler Construction

[0118] First, define the current time. Operating condition change rate:

[0119]

[0120] in For system load rate, The sampling interval is 1 second. This indicator reflects the drastic change in system operating conditions. When the value is small, the system is in a steady state; when the value is large, it is in a transient state.

[0121] Calculate the most recent The predicted residuals at each time point ( Take 60, corresponding to 60 seconds):

[0122]

[0123]

[0124]

[0125] in For actual measured energy efficiency indicators, , , These are the predicted values ​​from the three models.

[0126] Dynamic scheduler output fusion weights ,satisfy =1. The weight calculation adopts a strategy based on the inverse of the residual normalization and introduces a correction for the rate of change of operating conditions. First, calculate the basic weights:

[0127]

[0128]

[0129]

[0130] Introducing operating condition correction factors It increases monotonically with the rate of change of operating conditions:

[0131]

[0132] in The steepness coefficient of the S-curve is set to 5. The threshold for variable operating conditions is set to 0.05 / s. When When much smaller than the threshold, ≈0, the system mainly allocates weights based on the residuals; when When it is much greater than the threshold, ≈1, the system switches to transient preset weights. Data-driven model weights Set it to a higher value (e.g., 0.7) to prioritize response speed. The final weight is:

[0133]

[0134]

[0135]

[0136] S302: Generation of Fusion Twin Models

[0137] Based on the fusion weights output by the dynamic scheduler, a fusion twin model is generated:

[0138]

[0139] The output format is real-time energy efficiency index prediction values. This fusion model relies on a high-precision mechanistic model in steady state, a fast-response data-driven model in transient state, and a reduced-order model for smooth transition in the transition region, achieving a balance between high precision and high real-time performance across all operating conditions.

[0140] S303: Twin Confidence Calculation

[0141] To quantify the reliability of the current fusion model output, the twin confidence score is calculated. First, calculate the weighted residuals:

[0142] This value reflects the degree of discrepancy between the fused predictions and the predictions of each sub-model. A smaller discrepancy indicates higher consistency and reliability among the models. (Definition of reference residual) If the root mean square value of the weighted residuals over the past 24 hours is given, then the confidence level is calculated using the following formula:

[0143]

[0144] in A very small positive number (e.g.) ), to prevent division by zero. The value ranges from [0,1], where 1 indicates that the model output is highly reliable and 0 indicates that the model output is unreliable. This confidence level will be used for subsequent root cause diagnosis, optimization decisions, and model update triggering. Generate a twin confidence sequence for the generated products. .

[0145] S4: Causal Inference and Root Cause Diagnosis

[0146] Based on the device topology graph, a structural causal graph is constructed. A fused twin model is used as the counterfactual inference engine to perform intervention analysis and counterfactual inference on energy efficiency anomalies, generating counterfactual energy efficiency benchmarks and root cause diagnostic reports. This represents a technological leap from correlation analysis to causal inference, significantly improving the interpretability of the diagnosis. Specifically, this includes the following:

[0147] S401: Construction of Structural Cause-Effect Graph

[0148] by Using a framework and incorporating domain expert rules, a structural cause-effect graph (SCM) is constructed. Each node represents a key variable, and directed edges represent causal relationships, with their directions determined by the physical mechanism. Taking a central air conditioning system as an example, the cause-effect graph contains the following typical relationships:

[0149] Ambient temperature Affects the set temperature of cooling water This, in turn, affects the frequency of the cooling tower fan. Ultimately, this affects the return water temperature. and condensation pressure This changes the compressor's power consumption. .

[0150] Ambient temperature It also affects the set temperature of chilled water. This, in turn, affects the load rate of the chiller unit. and evaporation pressure This also ultimately affects the compressor's power consumption. .

[0151] In addition, fouling in the heat exchanger reduces the heat transfer coefficient, directly leading to increased evaporation and condensation temperatures, which in turn increases compressor power consumption.

[0152] Variables per node The causal function is defined as:

[0153]

[0154] in for The set of parent nodes, This represents exogenous noise variables (representing unobserved factors). Function Forms based on physical mechanisms or fusion twin models An approximate expression, for example, for condensation pressure. Its causal function can be approximated as:

[0155]

[0156] The fouling thermal resistance is a latent variable that can be obtained through inversion. The product structure causal graph (SCM) is obtained and stored in the form of a directed acyclic graph for subsequent intervention analysis.

[0157] S402: Counterfactual Energy Efficiency Benchmark Generation

[0158] When energy efficiency indicators Below the preset threshold (For example, when the design energy efficiency is 85%), a counterfactual reasoning process is triggered. For each controllable variable... (e.g., cooling water set temperature) Chilled water set temperature Water pump frequency (etc.), define the set of candidate intervention values. This typically includes ±5%, ±10%, etc., of the current value. For each intervention operation... ,use As a counterfactual reasoning engine, it recalculates the system state and energy efficiency while keeping other exogenous variables constant:

[0159]

[0160] Energy efficiency gains from computational intervention:

[0161]

[0162] Perform the above calculations on all controllable variables and their candidate intervention values ​​to generate a product counterfactual energy efficiency benchmark. The structure is [intervention variable, intervention value, predicted energy efficiency gain]. This benchmark not only provides optimization direction but also quantifies the potential benefits of adjusting each variable.

[0163] S403: Root Cause Diagnosis Report Generation

[0164] Will The intervention with the highest gain and twin confidence level Joint determination. First, define the energy efficiency gain threshold. (Taking 1%), only candidate root causes with gains greater than this threshold are retained. If multiple candidates exist, the one with the largest gain is selected as the primary root cause. Root cause confidence is also calculated.

[0165]

[0166] This formula considers the variable's contribution to energy efficiency and the reliability of the twin model, making the diagnostic results more robust. It generates a root cause diagnostic report. This includes: the root cause variable (e.g., "cooling water set temperature"); the recommended intervention value (the intervention value with the largest gain); and the expected energy-saving potential. (e.g., "3.2%"); Root cause confidence level (Value) (e.g., "0.87").

[0167] This report can be presented directly to operations and maintenance personnel, or used as input for subsequent optimization strategies.

[0168] S5: Multi-objective reinforcement learning and optimization policy generation

[0169] By using a fusion twin model as a high-fidelity simulator, a reinforcement learning environment is constructed. The policy network is trained offline to output the optimal action sequence, and the predicted energy efficiency gain curve is generated through simulation verification. This achieves an autonomous closed loop of "evaluation-optimization-control," avoiding trial-and-error control on real devices. Specifically, this includes the following:

[0170] S501: Building a Reinforcement Learning Environment

[0171] The fusion twin model It is encapsulated as a reinforcement learning environment simulator, which accepts the current state and action, and returns the next state and reward. State space. Defined as:

[0172]

[0173] in This refers to the current operating parameters (load rate, ambient temperature, outdoor humidity, etc.). For equipment health status characteristics (from The extracted quantitative values ​​of damage degree, such as scale thickness and wear amount, are used to measure the damage level. For twin confidence.

[0174] Action space Defined as the controllable parameter adjustment amount:

[0175]

[0176] in This represents the pump frequency adjustment amount, ranging from [-5, 5]. This represents the fan frequency adjustment amount, ranging from [-3, 3]. Set the temperature adjustment amount for chilled water, ranging from [-1, 1]; Set the temperature adjustment amount for the cooling water, ranging from [-2, 2]. This represents the valve opening adjustment amount, ranging from -10 to 10.

[0177] The reward function is defined as a multi-objective optimization form that takes into account energy efficiency improvement, comfort constraints, and equipment wear and tear.

[0178]

[0179] in:

[0180] Energy efficiency improvement rate (dimensionless), based on The simulation yielded the following results;

[0181] This is a comfort penalty (dimensionless). To predict indoor temperature, To set the temperature;

[0182] The incremental equipment wear (dimensionless) is calculated based on the equipment wear model fitted from historical data (such as the impact of start-stop times, speed changes, etc. on lifespan).

[0183] Let be the weighting coefficient, satisfying It can be adjusted according to the operating target (in this embodiment, it is taken as...). =0.6, =0.2, =0.2).

[0184] S502: Offline Training of Policy Networks

[0185] Using the Proximal Policy Optimization (PPO) algorithm in Offline training policy network. PPO is a reinforcement learning algorithm based on policy gradient, which improves training stability by limiting the magnitude of policy updates. Its objective function is:

[0186]

[0187] in For the comparison of the old and new strategies, The advantage function is estimated (by generalized advantage estimation, GAE). The pruning hyperparameter is set to 0.2. Both the policy network and the value network are 3-layer fully connected networks with 128 neurons per layer and ReLU activation function.

[0188] During the training process, The simulation ran for 10,000 rounds, each round consisting of 24 hours (8,640 time steps). The initial state was randomly sampled from historical data. The Adam optimizer was used with a learning rate of 0.0003 and a batch size of 256. The policy network was obtained after training converged. Utilize a policy network to analyze the current state. Perform reasoning and output the optimal action sequence:

[0189]

[0190] in To optimize the time domain length (12 steps, corresponding to 12 hours), a sequence of optimization actions for the product is generated. The format is [time step, device ID, action value].

[0191] S503: Generation of Predicted Energy Efficiency Gain Curve

[0192] use right Perform forward simulation, starting from the current state. Begin by applying the action sequence step by step, and calculating the energy efficiency index for each time step:

[0193]

[0194] Meanwhile, benchmark simulations were run without any optimization actions to obtain benchmark energy efficiency. Calculate the energy efficiency gain curve:

[0195]

[0196] Predicted energy efficiency gain curve of generated products The format is time series. This curve can be used to evaluate the expected effects of optimization strategies and provide a basis for subsequent control decisions.

[0197] S6: Closed-Loop Evaluation and Adaptive Triggering

[0198] The comprehensive energy efficiency assessment index is calculated and compared with a set threshold to determine whether to issue a control command. Simultaneously, based on the twin confidence level and model prediction error, it is determined whether the model needs to be updated, forming a continuous closed loop between the physical system and the twin, achieving self-evolutionary monitoring. Specifically, this includes the following:

[0199] S601: Calculation of Comprehensive Energy Efficiency Assessment Index

[0200] Calculate the current comprehensive energy efficiency assessment index:

[0201] in: The actual energy efficiency index of the current system (from (Real-time calculation)

[0202] To achieve the target energy efficiency value, take... Maximum expected energy efficiency and Weighted fusion of final gains:

[0203]

[0204] in for The maximum expected energy efficiency value in the range, To predict the final gain value of the energy efficiency gain curve.

[0205] Comprehensive Energy Efficiency Evaluation Index of Product It intuitively reflects the gap between the current energy efficiency level and the target, and a value less than 1 indicates that there is room for optimization.

[0206] S602: Closed-loop control command generation

[0207] judge With set threshold The relationship (with a value of 0.85):

[0208] like < And twin confidence Then, a closed-loop control command will be output, which will... The first action Converted into PLC executable instructions to achieve a closed-loop energy efficiency optimization;

[0209] like ≥ If so, no control is executed; only the evaluation results are recorded, and the current operating state is maintained.

[0210] like Even if < Instead of implementing controls, it triggers the model update process to avoid dangerous controls when the model is untrustworthy.

[0211] After the control command is generated, the execution result will be fed back to... This achieves a closed loop. It generates product closed-loop control commands. .

[0212] S603: Model update trigger judgment

[0213] Computational model update trigger flag:

[0214]

[0215] in: To integrate twin models The current predicted value; For the actual energy efficiency of the physical system (from calculate); The preset error threshold is 0.05, or 5%.

[0216] like This triggers the model update process, which is then fed back to S2 to utilize the latest version. Retraining (Incremental learning or full retraining), and update. The parameters (identified online via Kalman filtering or recursive least squares) enable the self-evolutionary closed loop of the twin model.

[0217] S604: Evaluation Results Output and Archiving

[0218] The comprehensive energy efficiency assessment index for this assessment period Root cause diagnosis report Optimize action sequences Predicted energy efficiency gain curve The results of control and execution are integrated into a standardized energy efficiency assessment report, which is simultaneously pushed to the cloud management platform and the terminals of operation and maintenance personnel. The value triggers the corresponding level of warning:

[0219] Green (normal), only record;

[0220] Yellow (Attention): A reminder to pay attention.

[0221] Orange (Abnormal), optimization recommended;

[0222] Red (Critical) - Immediately alert the police and recommend stopping the machine for inspection.

[0223] At the same time, all data from this assessment process (including raw data, preprocessing results, model parameters, assessment results, etc.) will be archived in the case library for subsequent model iteration and threshold optimization, so as to realize the self-evolution and continuous optimization of the monitoring system.

[0224] Based on the above method, this application also discloses an energy efficiency assessment system based on digital twins, with reference to Figure 3 The digital twin-based energy efficiency assessment system shown includes:

[0225] The data acquisition and feature extraction module is configured to collect multi-source data from the energy-consuming system, generate a standardized time-series dataset after cleaning and alignment, construct a device topology diagram, and generate a set of operating condition event labels through change point detection.

[0226] The hybrid twin model pool construction module is configured to construct a high-fidelity mechanism model, a lightweight data-driven model, and a reduced-order model based on the standardized time-series dataset, equipment topology diagram, and operating condition event label set, forming a hybrid twin model pool;

[0227] The dynamic fusion and confidence generation module is configured to build a dynamic scheduler, calculate the fusion weight in real time based on the rate of change of operating conditions and the prediction residuals of each model, and generate a fusion twin model by weighted fusion of the high-fidelity mechanism model, the lightweight data-driven model and the reduced-order model, and generate a twin confidence sequence based on the weighted residuals of the fusion prediction value and the prediction values ​​of each sub-model.

[0228] The causal inference and root cause diagnosis module is configured to construct a structural causal graph with the device topology diagram as the skeleton. When the energy efficiency index is lower than the preset threshold, the fusion twin model is used to perform intervention analysis on each controllable variable, calculate and predict the energy efficiency gain and generate a root cause diagnosis report.

[0229] The optimized strategy generation module is configured to encapsulate the fusion twin model as a reinforcement learning environment simulator, define the state space, action space and multi-objective reward function, train the policy network offline to output the optimal action sequence, and generate a predicted energy efficiency gain curve through forward simulation.

[0230] The closed-loop evaluation and adaptive triggering module is configured to calculate the comprehensive energy efficiency evaluation index and compare it with a set threshold to determine whether to issue a closed-loop control command. At the same time, it determines whether to trigger a model update based on the twin confidence and model prediction error. If triggered, it feeds back to the hybrid twin model pool construction module to update the model pool, forming a continuous closed loop between the physical system and the digital twin.

[0231] Please see Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Figure 4 As shown, the electronic device 1000 may include: at least one processor 1001, at least one communication bus 1002, user interface 1003, network interface 1004, and memory 1005.

[0232] The communication bus 1002 is used to realize the connection and communication between the above components, and to ensure the stable transmission of data and instructions.

[0233] The user interface 1003 may include a display screen and a touch interaction unit, and optionally also includes a standard wired debugging interface and a wireless Bluetooth interaction interface. The user interface 1003 is used to realize the operation interaction of on-site operation and maintenance personnel, the visualization of energy efficiency monitoring data, the viewing of energy efficiency damage assessment results, the pop-up prompts of early warning information, and the on-site configuration of equipment parameters. Optionally, it can also be connected to an audible and visual alarm unit to realize on-site early warning of energy efficiency risks.

[0234] The network interface 1004 may include a standard wired Ethernet interface, an industrial fieldbus interface, or a wireless interface (such as LoRa, 5G, Wi-Fi, or NB-IoT). The network interface 1004 is used to enable bidirectional communication between electronic devices and four levels of sensing devices, data acquisition gateways, cloud management platforms, and mobile terminals for maintenance personnel, and to complete core data interactions such as energy efficiency data acquisition, control command issuance, evaluation result uploading, and early warning information push.

[0235] The processor 1001 may include one or more processing cores. The processor 1001 connects to various functional units within the electronic device through various interfaces and lines. It executes all functions and data processing steps of the digital twin-based energy efficiency assessment method described in this invention by running or executing instructions, programs, code sets, or instruction sets stored in the memory 1005, and by calling data stored in the memory 1005. These functions include, but are not limited to: multi-source data acquisition and feature extraction, hybrid twin model pool construction, dynamic fidelity fusion and twin confidence generation, causal inference and root cause diagnosis, multi-objective reinforcement learning and optimization strategy generation, closed-loop evaluation and adaptive triggering, etc. Optionally, the processor 1001 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 1001 may integrate one or more combinations of a Central Processing Unit (CPU), a Graphics Processing Unit (GPU), and an embedded communication modem. The CPU primarily handles the operating system, monitoring business logic, user interaction programs, and core algorithm solving; the GPU is responsible for the calculation and rendering of energy efficiency data visualization, 3D rendering of twin models, and energy efficiency gain curve plotting; and the modem is used to handle the parsing and data transmission of various wireless communication protocols.

[0236] The memory 1005 may include random access memory (RAM) or read-only memory; optionally, the memory 1005 may include non-transitory computer-readable storage medium. The memory 1005 can be used to store instructions, programs, code, code sets, or instruction sets, and includes a program storage area and a data storage area.

[0237] The program storage area can store instructions for implementing an operating system, instructions for implementing at least one function, and instruction programs for implementing all the steps of the above-described method embodiments of the present invention.

[0238] The storage data area can store various types of data involved in the implementation of the method of this invention, including but not limited to: standardized time series datasets, equipment topology diagrams, operating condition event label sets, high-fidelity mechanism model parameters, lightweight data-driven model weights, reduced-order model projection matrices, fused twin model weight coefficients, twin confidence sequences, structural causal graphs, counterfactual energy efficiency benchmarks, root cause diagnosis reports, reinforcement learning policy network parameters, optimal action sequences, predicted energy efficiency gain curves, comprehensive energy efficiency evaluation indices, closed-loop control instructions, model update trigger flags, and historical archived case data.

[0239] In the electronic device disclosed in this embodiment, the processor 1001 can call the executable program of the energy efficiency assessment method based on digital twin stored in the memory 1005. When the executable program is executed by one or more processors 1001, the electronic device performs all the steps of the energy efficiency assessment method based on digital twin as described in any of the above embodiments, so as to realize low power consumption, high accuracy and high reliability of online energy efficiency monitoring and optimization control throughout the entire life cycle of the energy system.

[0240] An electronic device readable storage medium stores instructions that, when executed by one or more processors, cause the electronic device to perform one or more of the methods described in the above embodiments.

[0241] It should be noted that the computer-readable storage medium may include, but is not limited to, USB flash drives, portable hard drives, magnetic disks, optical disks, solid-state drives (SSDs), embedded read-only memory, random access memory, and various non-volatile memories, as well as other media capable of storing program code. When the computer program product runs on hardware carriers such as electronic devices, edge computing terminals, and cloud servers, it can fully execute the entire process of monitoring, judgment, location, evaluation, and optimization of the method embodiments of the present invention, realizing full-domain coverage, hierarchical linkage, and precise quantification of energy efficiency online monitoring and autonomous optimization of the energy-consuming system, adapting to the long-term unattended monitoring needs of the entire life cycle of the energy-consuming system.

[0242] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0243] In conclusion, 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, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for energy efficiency assessment based on digital twinning, characterized in that, include: S1: Deploy sensors and controllers at key energy efficiency nodes of the target energy system through an industrial IoT architecture to collect energy efficiency-related parameters, operating condition-related parameters and equipment status parameters, and generate standardized time-series datasets after preprocessing. A device topology diagram is constructed based on the physical connection and control hierarchy of the devices; a change point detection algorithm is used to identify operating condition switching events and generate a set of operating condition event tags with semantic labels. S2: Based on standardized time-series datasets, equipment topology diagrams, and operating condition event label sets, high-fidelity mechanism models, lightweight data-driven models, and reduced-order models are constructed respectively to form a hybrid twin model pool; S3: Construct a dynamic scheduler to calculate the fusion weights in real time based on the rate of change of operating conditions and the prediction residuals of each model. The high-fidelity mechanism model, the lightweight data-driven model, and the reduced-order model are weighted and fused to generate a fused twin model. The weighted residuals between the fused prediction value and the prediction values ​​of each sub-model are calculated, and a twin confidence sequence is generated based on the historical reference residuals. S4: Construct a structural cause-effect graph using the equipment topology diagram as the skeleton. When the energy efficiency index is lower than the preset threshold, use the fusion twin model as the counterfactual reasoning engine to perform intervention analysis on each controllable variable, calculate the predicted energy efficiency gain after intervention, generate the counterfactual energy efficiency benchmark, and generate a root cause diagnosis report based on the maximum gain and twin confidence. S5: The fusion twin model is encapsulated as a reinforcement learning environment simulator. The state space, action space and multi-objective reward function are defined. The policy network is trained offline using the proximal policy optimization algorithm. The optimal action sequence is output. The predicted energy efficiency gain curve is generated through forward simulation of the fusion twin model. S6: Calculate the comprehensive energy efficiency assessment index and compare it with the set threshold to determine whether to issue a closed-loop control command; at the same time, based on the twin confidence level and the model prediction error, determine whether to trigger a model update. If triggered, feed back to S2 to update the hybrid twin model pool.

2. The digital-twin-based energy efficiency assessment method according to claim 1, characterized in that, The operating condition event tag set includes: selecting parameters sensitive to changes in operating conditions to form a feature vector, including total power, load rate, start / stop status of major equipment, and total system flow; using a pruning precise linear time algorithm to identify abrupt change points in the time series, and determining the change point location by minimizing the objective function containing the cost function and penalty coefficient within the segment; labeling the intervals divided by each change point with event types using domain knowledge, the event types including equipment start-up, equipment shutdown, load increase, load decrease, and fault events; and generating an operating condition event tag set containing event start time, event end time, event type, associated equipment ID, and event confidence.

3. The digital-twin-based energy efficiency assessment method according to claim 1, characterized in that, The construction of the dynamic scheduler includes: defining the rate of change of operating conditions as the absolute value of the derivative of the load rate with respect to time, reflecting the severity of changes in system operating conditions; calculating the prediction residuals of the high-fidelity mechanism model, the lightweight data-driven model, and the reduced-order model within a preset time window; calculating the basic weights based on the reciprocal normalization of the prediction residuals of each model, and introducing a condition correction factor, which monotonically increases with the rate of change of operating conditions; when the rate of change of operating conditions is less than the variable operating condition threshold, the system allocates fusion weights according to the basic weights; when the rate of change of operating conditions is greater than the variable operating condition threshold, the system switches to the transient preset weights; finally, the fusion weights are obtained by weighting the basic weights and the transient preset weights using the condition correction factor.

4. The digital-twin-based energy efficiency assessment method of claim 1, wherein, The generation of the twin confidence sequence includes: calculating the weighted residuals between the fusion predicted values ​​and the predicted values ​​of the high-fidelity mechanism model, the lightweight data-driven model, and the reduced-order model. The weighted residuals are obtained by weighted summation of the prediction residuals of each sub-model based on the fusion weights output by the dynamic scheduler; defining the historical reference residual as the root mean square value of the weighted residuals within a preset historical period; normalizing the ratio of the weighted residuals to the historical reference residuals and mapping it to a preset confidence interval to generate a twin confidence sequence with values ​​ranging from 0 to 1. A confidence value close to 1 indicates that the model output is highly reliable, and a confidence value close to 0 indicates that the model output is unreliable.

5. The digital-twin-based energy efficiency assessment method according to claim 1, wherein, The generation of the root cause diagnosis report includes: constructing a structural causal graph based on the equipment topology diagram and physical mechanisms, where each node in the structural causal graph represents a key variable and directed edges represent causal relationships determined by physical mechanisms; when the energy efficiency index is lower than a preset threshold, using a fusion twin model as a counterfactual inference engine, performing intervention analysis on each controllable variable, defining a set of candidate intervention values, recalculating the system state and energy efficiency while keeping other exogenous variables unchanged, calculating the predicted energy efficiency gain after intervention, and generating a counterfactual energy efficiency benchmark containing intervention variables, intervention values, and predicted energy efficiency gains; jointly determining the intervention item with the largest gain in the counterfactual energy efficiency benchmark with the twin confidence level, calculating the root cause confidence level based on the product of the ratio of the largest energy efficiency gain to the sum of the energy efficiency gains of all controllable variables and the twin confidence level, and generating a root cause diagnosis report containing root cause variables, recommended intervention values, expected energy-saving potential, and root cause confidence levels.

6. The digital-twin-based energy efficiency assessment method according to claim 1, characterized in that, The generation of the optimal action sequence includes: encapsulating the fusion twin model as a reinforcement learning environment simulator, defining a state space, an action space, and a multi-objective reward function; the state space includes current operating parameters, equipment health status characteristics, and twin confidence; the action space is defined as controllable parameter adjustment amounts, including water pump frequency adjustment, fan frequency adjustment, chilled water set temperature adjustment, cooling water set temperature adjustment, and valve opening adjustment; the multi-objective reward function is a weighted sum of energy efficiency improvement rate, comfort penalty term, and equipment loss increment; a proximal policy optimization algorithm is used to train the policy network offline in the fusion twin model, and its objective function limits the policy update magnitude through a pruning operation of the new and old policy ratio and the advantage function estimation; after training convergence, the policy network is used to infer the current state and output the optimal action sequence containing action values ​​at each time step within the preset optimization time domain.

7. The digital-twin-based energy efficiency assessment method according to claim 1, wherein, The triggering of the model update includes: calculating a model update trigger flag, wherein the model update trigger flag is that the twin confidence level is lower than a preset lower threshold, or the relative error between the predicted value of the fused twin model and the actual energy efficiency value of the physical system is greater than a preset error threshold; when the model update trigger flag is true, the model update process is triggered and fed back to step S2, which uses the newly generated standardized time series dataset to update the hybrid twin model pool, including performing incremental learning or full retraining on the lightweight data-driven model, and online identification of the parameters of the high-fidelity mechanism model through Kalman filtering or recursive least squares.

8. An energy efficiency assessment system based on digital twinning, characterized in that, include: The data acquisition and feature extraction module is configured to collect multi-source data from the energy-consuming system, generate a standardized time-series dataset after cleaning and alignment, construct a device topology diagram, and generate a set of operating condition event labels through change point detection. The hybrid twin model pool construction module is configured to construct a high-fidelity mechanism model, a lightweight data-driven model, and a reduced-order model based on the standardized time-series dataset, equipment topology diagram, and operating condition event label set, forming a hybrid twin model pool; The dynamic fusion and confidence generation module is configured to build a dynamic scheduler, calculate the fusion weight in real time based on the rate of change of operating conditions and the prediction residuals of each model, and generate a fusion twin model by weighted fusion of the high-fidelity mechanism model, the lightweight data-driven model and the reduced-order model, and generate a twin confidence sequence based on the weighted residuals of the fusion prediction value and the prediction values ​​of each sub-model. The causal inference and root cause diagnosis module is configured to construct a structural causal graph with the device topology diagram as the skeleton. When the energy efficiency index is lower than the preset threshold, the fusion twin model is used to perform intervention analysis on each controllable variable, calculate and predict the energy efficiency gain and generate a root cause diagnosis report. The optimized strategy generation module is configured to encapsulate the fusion twin model as a reinforcement learning environment simulator, define the state space, action space and multi-objective reward function, train the policy network offline to output the optimal action sequence, and generate a predicted energy efficiency gain curve through forward simulation. The closed-loop evaluation and adaptive triggering module is configured to calculate the comprehensive energy efficiency evaluation index and compare it with a set threshold to determine whether to issue a closed-loop control command. At the same time, it determines whether to trigger a model update based on the twin confidence and model prediction error. If triggered, it feeds back to the hybrid twin model pool construction module to update the model pool, forming a continuous closed loop between the physical system and the digital twin.

9. An electronic device, comprising: It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method according to any one of claims 1 to 7.

Images