Building energy efficiency dynamic simulation method and system based on digital twinning

Abstract

The invention relates to the technical field of simulation optimization and simulation, in particular to a building energy efficiency dynamic simulation method and system based on digital twinning, a building energy efficiency digital twinning model is designed through a neural symbol method, and an energy efficiency optimization strategy can be autonomously generated through interactive simulation by utilizing reinforcement learning according to real-time data and external changes. In order to ensure the accuracy of the model, all data are endowed with a unified semantic framework and knowledge association, and a machine learning algorithm is utilized to realize automatic synchronous updating and full-dimensional adaptive correction of the model, including correction of state variables, model parameters and structural information. In addition, future energy consumption prediction can be carried out, a future control action sequence can be generated, the event-driven MPC algorithm generates high-level instructions for key events to achieve intelligent control, and the purposes that the building energy consumption is lowest and various constraint conditions are met are achieved. According to the method, the advantages of data driving and model driving are fully utilized, and the intelligent level of building energy efficiency management is improved.

Description

Technical Field

[0001] The present invention relates to the field of simulation optimization and simulation technology, and specifically to a building energy efficiency dynamic simulation method and system based on digital twins. Background Art

[0002] In the field of digital twins and energy efficiency optimization of industrial buildings, computer-aided simulation, modeling, and optimization provide key technical support. However, the lack of unified semantic standards for multi-source data within industrial buildings makes it difficult to achieve the semantic mapping and standardization required for simulation. Existing models are unable to effectively integrate heterogeneous data through feature extraction and cross-modal alignment through machine learning, which restricts the construction and optimization of digital twin models. Existing energy efficiency models are mostly based on static historical data and lack a dynamic update mechanism. They cannot respond to physical changes in buildings, equipment status, or production plan adjustments in real time, causing simulation results to deviate from actual operating conditions, affecting energy consumption prediction and optimization decisions. Existing models are mostly based on rule-based or PID control models, which are difficult to simulate collaboratively. They cannot achieve balanced optimization under dynamic operating conditions or multi-objective constraints, such as energy consumption, comfort, and carbon emissions. Control lags lead to failure of global optimization. Existing models lack an event detection and response framework based on machine learning, and cannot trigger real-time simulation and strategy optimization for emergencies.

[0003] To this end, a dynamic simulation method and system for building energy efficiency based on digital twins is proposed. Summary of the Invention

[0004] This invention aims to address the current pain points of industrial building energy efficiency management, such as data isolation, static models, delayed responses, and insufficient optimization. By building a dynamic, adaptive, and intelligent digital twin model, it enables accurate prediction and real-time optimization of building energy consumption, significantly improving the building's energy efficiency and overall operational benefits.

[0005] A dynamic simulation method for building energy efficiency based on digital twins, including the following steps: A neural symbolic approach is used to design a digital twin model of building energy efficiency. By combining building physics with a multi-dimensional comprehensive dataset, a dynamic hybrid self-evolving network is constructed. This dynamic hybrid self-evolving network uses reinforcement learning to interactively simulate and autonomously generate energy efficiency optimization strategies based on real-time sensor data, external environmental changes, and production equipment aging. Assign a unified semantic framework and knowledge association to all data in the building energy efficiency digital twin model, and use machine learning algorithms to automatically synchronize and update the model and perform full-dimensional adaptive correction based on the building's multi-dimensional comprehensive data set. The correction includes correction of state variables, model parameters, and structural information. The updated and revised digital twin model is used to predict future energy consumption, and the future control action sequence is obtained by following the constrained rolling optimization control instructions. The time domain behavior of the model is simulated based on the future control action sequence, and future scenarios are predicted. The event-driven MPC algorithm is used to optimize the future scenarios and generate high-level optimization instructions to guide the EVK energy efficiency cabinet to achieve intelligent control.

[0006] Preferably, the building physics mechanism is to obtain static and semi-static data on the geometric structure of industrial buildings, material properties of enclosure structures, production equipment models and design parameters, production plans and shift patterns from the historical database of the EVK energy efficiency cabinet; the sources and types of the multi-dimensional comprehensive data set include: using various sensors and instruments deployed inside the industrial building to obtain environmental parameters in real time, including production workshop temperature, humidity, CO2 concentration, and light intensity; the real-time operating status of the factory includes the start and stop status, power, current, and electrical energy and thermal energy consumption data of production equipment, pumps, fans, and chillers; using external information sources to obtain weather forecast data, including outdoor temperature, solar radiation, wind speed, power grid signals and calendar information, the power grid signals include time-of-use electricity prices, peak and valley electricity prices, and demand response instructions, and the calendar information includes holidays and working days.

[0007] Preferably, the design of the building energy efficiency digital twin model using the neural symbolic method specifically includes: a neural network module, used for pattern recognition, feature extraction and behavior prediction of the multi-dimensional comprehensive data set; a symbolic reasoning module, used for encoding and logical representation of the building physical mechanisms, engineering rules and domain knowledge; the engineering rules and the domain knowledge are derived from industry standards, expert experience, historical data analysis and basic scientific principles; a knowledge graph module, used to construct a unified semantic framework and knowledge association of the building energy efficiency digital twin model.

[0008] Preferably, the dynamic hybrid self-evolutionary network performs interactive simulation and autonomously generates energy efficiency optimization strategies, specifically including: using the building energy efficiency digital twin model as a simulation environment, defining the state space as the internal environmental state of the building, external environmental parameters, and real-time operating status of the factory; defining the action space as HVAC model parameter setting points, lighting switches and brightness adjustments, and production equipment start and stop control instructions; using preset energy consumption, production workshop comfort, production equipment operating life and carbon emission indicators to construct the reinforcement learning reward function; and obtaining energy efficiency optimization strategies for dynamic operating conditions and external environmental changes through iterative interaction and trial-and-error learning by reinforcement learning agents in the simulation environment of the model.

[0009] Preferably, the specific steps of the automatic synchronous update and the full-dimensional adaptive correction include: using semantic standards and knowledge graphs to semantically map, integrate and standardize the internal heterogeneous sensor data, production equipment information, spatial topology and model configuration data of the building; using online machine learning algorithms to continuously monitor the real-time deviation between the physical entity of the building and the model, and automatically calibrate, estimate and adjust the state variables, model parameters and structural information of the model; the state variables include real-time temperature, humidity, and CO2 concentration; the model parameters include thermal conductivity, production equipment efficiency, and sensor drift; and the structural information includes model connections and topology changes.

[0010] Preferably, the specific steps of obtaining the future control action sequence include: using the updated and revised digital twin model as a prediction model, integrating historical operation data and real-time sensor data, and combining external environment predictions, including weather forecasts, production scheduling plans and energy price information, to predict the electricity, heating and cooling load energy consumption of the building; using mixed integer linear programming, taking the production workshop temperature comfort range, production equipment operating power range, and grid demand response signal within the prediction time domain as constraints, and performing rolling optimization solution to obtain the future control action sequence that can minimize energy consumption and meet the constraints within the prediction time domain.

[0011] Preferably, the event-driven MPC is used to optimize the future scenario and generate high-level optimization instructions, specifically including: continuously monitoring the real-time operating status of the building and changes in the external environment, identifying key events that are preset and / or learned through machine learning models, including sudden changes in production load, process flow adjustments, sudden extreme weather events, production equipment failures and significant performance degradation; the model will simulate and generate simulation results after the key events are triggered; and using the event-driven MPC algorithm, based on the future control action sequence and the simulation results, restart the optimization calculation to obtain high-level optimization instructions after the key events are triggered.

[0012] Preferably, a building energy efficiency dynamic simulation system based on digital twin is integrated, specifically including: The digital twin model construction and self-evolution module is used to design a digital twin model of building energy efficiency using a neural symbolic approach. By combining building physics with a multi-dimensional integrated data set, a dynamic hybrid self-evolving network is constructed. This dynamic hybrid self-evolving network uses reinforcement learning to interactively simulate and autonomously generate energy efficiency optimization strategies based on real-time sensor data, external environmental changes, and production equipment aging. A data semantics and model adaptive correction module is used to give a unified semantic framework and knowledge association to all data in the building energy efficiency digital twin model. Based on the multi-dimensional comprehensive data set of the building, the model is automatically and synchronously updated and fully adaptively corrected using a machine learning algorithm. The correction includes correction of state variables, model parameters and structural information. The energy consumption prediction and intelligent optimization control module is used to use the updated and revised digital twin model to predict future energy consumption, follow the constrained rolling optimization control instructions, and obtain the future control action sequence; based on the future control action sequence, it simulates the time domain behavior of the model, predicts future scenarios, and uses the event-driven MPC algorithm to optimize the future scenarios, generate high-level optimization instructions, and guide the EVK energy efficiency cabinet to achieve intelligent control.

[0013] Compared with the prior art, the present invention has the following beneficial effects: 1. This method combines the neural symbolic method with the physical mechanism of buildings to construct a dynamic hybrid self-evolving network, and uses reinforcement learning for interactive simulation to optimize energy efficiency strategies in real time. Compared with traditional models that rely on static models, this method can dynamically perceive production load, environmental changes and equipment aging, and adjust control strategies such as optimizing HVAC set points according to operating status. The introduction of reinforcement learning enables the model to autonomously learn the optimal solution under the constraints of multiple objectives. For example, the lighting model will autonomously balance the lighting intensity and shading model to take into account both comfort and energy consumption. In addition, by predicting the energy efficiency decline trend of equipment, the model can plan maintenance in advance to reduce additional energy consumption, and reduce peak load based on rolling optimization to improve energy elasticity.

[0014] 2. By unifying the semantic framework and knowledge graph, this method integrates multi-source heterogeneous data, such as multi-source sensor information, to resolve data silos and normalize different types of data. Combined with online machine learning algorithms, the model can calibrate the state variables, parameters, and structure of the digital twin in real time. When significant deviations are detected, it can autonomously correct them, ensuring synchronization and alignment between the building entity and the digital twin. For example, efficiency deviations caused by equipment wear can be automatically corrected. This automated mechanism directly enables real-time monitoring and control of internal building equipment, significantly reducing the need for manual intervention. Control commands only need to be submitted to the system for confirmation, reducing manual operation and maintenance costs.

[0015] 3. Energy consumption prediction based on the digital twin model, combined with an event-driven MPC algorithm, allows the constructed digital twin model to generate forward-looking optimization instructions, such as those for responding to large fluctuations in electricity prices. When a critical event, such as extreme weather or equipment failure, is detected, the optimization calculation is immediately restarted and the control strategy is adjusted, such as to quickly respond to a sudden temperature drop. The optimization instructions are sent to the EVK energy efficiency cabinet via a standard protocol and converted into executable signals that the equipment can accept and understand, achieving closed-loop precise control from virtual simulation instructions to physical layer execution. Leveraging event-driven and intelligent response mechanisms, the system continuously expands its knowledge of identifying and responding to emergencies through pre-set or machine learning models. This makes it impossible to adjust strategies in the face of emergencies in a timely manner, resulting in high energy consumption and low operational efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of a dynamic simulation method for building energy efficiency based on digital twins proposed in this invention application; Figure 2 This is a workflow diagram of the reinforcement learning agent in Example 1 of the present invention; Figure 3 This is a flowchart of the steps for identifying key events in Example 1 of the present invention; Figure 4 This is an architectural diagram of a digital twin-based building energy efficiency dynamic simulation system proposed in this invention application. DETAILED DESCRIPTION

[0017] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0018] See also Figures 1 to 4 The present invention provides a building energy efficiency dynamic simulation method and system based on digital twins. The technical solution is as follows. Figure 1 , which is a flow chart of the method applied for the present invention: A neural symbolic approach is used to design a digital twin model of building energy efficiency. By combining building physics with a multi-dimensional comprehensive dataset, a dynamic hybrid self-evolving network is constructed. This dynamic hybrid self-evolving network uses reinforcement learning to interactively simulate and autonomously generate energy efficiency optimization strategies based on real-time sensor data, external environmental changes, and production equipment aging. Assign a unified semantic framework and knowledge association to all data in the building energy efficiency digital twin model, and use machine learning algorithms to automatically synchronize and update the model and perform full-dimensional adaptive correction based on the building's multi-dimensional comprehensive data set. The correction includes correction of state variables, model parameters, and structural information. The building energy efficiency digital twin model is used to predict future energy consumption, and the future control action sequence is obtained by following the constrained rolling optimization control instructions. The time domain behavior of the model is simulated based on the future control action sequence to predict future scenarios. The event-driven MPC algorithm is used to optimize the future scenarios and generate high-level optimization instructions to guide the EVK energy efficiency cabinet to achieve intelligent control. Figure 4 , which is an architecture diagram of a building energy efficiency dynamic simulation system based on digital twins proposed in the present invention application; Example 1: This implementation is applied to a precision processing plant. With a floor area of ​​5,000 square meters, the plant primarily produces high-precision automotive parts, requiring stringent temperature and humidity requirements. The plant's primary energy-consuming equipment includes CNC machine tools, central air conditioning models, and chillers, water pumps, fans, air compressors, and lighting models.

[0019] This method aims to achieve dynamic simulation and intelligent optimization control of energy efficiency by building a digital twin model of the plant's energy efficiency. The physical execution terminal of this control is the EVK series energy efficiency cabinet deployed in the plant's power distribution room.

[0020] Data collection and model foundation construction, specifically covering physical mechanism data and multi-dimensional comprehensive data sets, including: Static data is obtained from the historical database of the EVK energy efficiency cabinet by consulting the design drawings and historical database; The geometric structure defines the three-dimensional dimensions of the factory building and the division of its internal functional areas, such as production area and office area; The material properties of the building envelope include the material composition and thermal insulation performance of the exterior walls, roof, windows, etc. The exterior walls are made of concrete and rock wool insulation, and their comprehensive heat transfer performance indicators are quantified. The production equipment models and design parameters are registered in detail, including key parameters such as the model, rated power, and design efficiency of each production equipment. This includes specific model and energy consumption information. For example, in Area A, 50 CNC machine tools with a rated power of 15 kW are deployed, and the cooling capacity, rated power, and design energy efficiency ratio of the central air conditioning model are recorded. Semi-static data includes production plans and shift patterns, which record the factory's production and planning arrangements, personnel number information, daily two-shift working hours, and production load patterns on weekends or holidays.

[0021] Furthermore, acquiring a multi-dimensional, comprehensive data set of internal environmental parameters primarily involves deploying a network of sensors within the production workshop. These sensors collect real-time information on parameters such as temperature, humidity, carbon dioxide concentration, and light intensity in each area every minute. The factory has clear control targets for these parameters. For example, the temperature must be maintained within a range of 22 degrees Celsius, plus or minus 1 degree. These parameters constantly influence the stability of product quality.

[0022] Through the collection module in the EVK energy efficiency cabinet, the start and stop status, real-time power and cumulative energy consumption of each major equipment are monitored every 15 seconds to obtain the real-time operating status of the production equipment.

[0023] Through the network interface, it automatically obtains external information sources such as hourly weather forecast data for the next few days, including outdoor temperature, solar radiation intensity, time-of-use electricity price information released by the power grid, including specific electricity prices during peak, flat and valley periods, and load control instructions, and combines calendar information to distinguish between weekdays and holidays.

[0024] By acquiring static and semi-static building physics data from the EVK energy efficiency cabinet historical database, as well as real-time, multi-dimensional integrated data sets from sensors, meters, and external information sources, a comprehensive and accurate data foundation is provided for the digital twin model. This enables the model to accurately reflect the building's inherent characteristics, real-time operating conditions, and external environmental influences, including economic signals, thereby ensuring the accuracy and practicality of energy efficiency prediction and optimization strategies, and achieving more refined energy management.

[0025] Furthermore, based on the above data, a convolutional long short-term memory network is used to learn massive, dynamically changing data and identify complex patterns in the data. For example, based on historical and future meteorological data and equipment operation data, the spatial and temporal distribution trend of temperature and humidity inside the factory is predicted.

[0026] The symbolic reasoning module encodes explicit, rule-based knowledge. Physical mechanisms are the underlying physical laws. For example, the total heat change in a space is the result of the combined effects of heat generated by the sun, internal equipment, and occupants, and heat lost through walls, roofs, and other structures. The heat generated by equipment is calculated based on its operating power and efficiency.

[0027] Engineering rules and domain knowledge are rules embedded in the model, derived from industry standards, expert operational experience, and historical data analysis. For example, "When outdoor weather conditions are favorable, prioritize natural cooling, which consumes less energy," or "During nighttime hours when electricity prices are low, pre-generate and store some cooling capacity to meet peak energy demand the following day."

[0028] Furthermore, a unified semantic framework is constructed to connect all isolated data and entities in the model. By defining entities such as devices, spaces, and sensors, as well as the relationships between them, such as location, control, and measurement, a vast knowledge network is formed. Each entity and relationship in this knowledge network is timestamped, connecting isolated data points into an intelligent knowledge network, imbuing the data with context and logical relationships. The model no longer merely displays data; it also displays corresponding relationships. This allows users to understand not only the present but also the past, enabling in-depth review of complex events and accurate predictions of future trends. This knowledge-driven approach will significantly improve operational efficiency, decision-making capabilities, and the system's autonomous intelligence.

[0029] The knowledge graph is stored in a graph database, whose graph structure is ideal for efficiently storing entities and relationships and supporting fast queries. The knowledge graph is queried using the SPARQL query language. By training a regression model, the nonlinear mapping relationship between device operating parameters and actual performance is learned. The corresponding parameter values ​​in the knowledge graph are inferred and corrected based on real-time operating data. Furthermore, during structural updates, the reasoning capabilities of the knowledge graph can be leveraged for consistency checks. For example, this can check whether newly added devices have valid connection points or whether removing a device will disrupt energy flow, thereby preventing the introduction of logical errors.

[0030] The neural symbolic approach is designed with a neural network module for data pattern recognition and behavior prediction; a symbolic reasoning module for encoding and logically representing building physics mechanisms, engineering rules, and domain knowledge; and a knowledge graph module for constructing a unified semantic framework. This combination enables the model to combine data learning and logical reasoning capabilities, improving its predictive accuracy, decision interpretability, and knowledge management efficiency, ensuring that energy efficiency optimization strategies are both intelligent and reliable.

[0031] Furthermore, the constructed digital twin model is used as a virtual testing ground to train an intelligent agent to autonomously learn and generate the optimal energy efficiency control strategy. The simulation environment is the digital twin model constructed above that can highly simulate the physical behavior of the real plant. Figure 2 , which is a workflow diagram of the reinforcement learning agent in Example 1 of the present invention; Specifically, the state space defines the information that the intelligent agent needs to perceive, and contains 150 normalized vectors with specific values. These include: temperature, humidity, and carbon dioxide concentration readings for each of the 10 key areas inside the factory; the start and stop status and real-time power of each of the 80 CNC machine tools; the operating power and frequency of two chillers, four water pumps, and 10 air-conditioning box fans; the current outdoor temperature, solar radiation intensity, wind speed, and the current grid time-of-use electricity price.

[0032] The action space defines all the control operations that the agent can perform, including: the outlet water temperature set point of the central air-conditioning chiller, which can be set between 7.0 degrees Celsius and 12.0 degrees Celsius; the inverter operating frequency of all chilled water pumps, cooling water pumps and air-conditioning box fans, which can be set between 30% of the rated minimum frequency and 100% of the maximum frequency; the brightness of the lighting model can be continuously adjusted between 0% (off) and 100% (brightest); Reward function construction: In order to guide the agent to learn the desired behavior, a comprehensive reward function is designed. At each control decision point, the model calculates a reward value. The calculation method of this reward value is: First, set a basic reward of zero. Then, deduct points based on performance in three aspects: Energy cost penalty: multiply the total power of all equipment at the current moment by the electricity price at that time to obtain the instantaneous electricity cost. The larger the cost value, the more points will be deducted. Comfort deviation penalty: calculate the absolute difference between the actual temperature of all measuring points in the production workshop and the target temperature of 22 degrees Celsius. The larger the difference, the more uncomfortable the environment, and the more points will be deducted. Equipment loss penalty: calculate the absolute value of the difference between the operating frequency of key equipment at the current moment and its frequency at the previous moment. The larger the difference, the more drastic the equipment adjustment, which may accelerate equipment wear and fatigue and affect the operating life. Therefore, the model tries its best to avoid actively avoiding operations that will cause high real-time pressure in generating decisions, and will automatically find a running path that meets business needs and makes the equipment most comfortable. In essence, it is performing preventive maintenance every minute and every second.

[0033] Finally, these three deduction items are weighted together, with energy cost weighted at 0.6, comfort weighted at 0.3, and equipment loss weighted at 0.1. The inverse of this sum is the final reward. The agent's goal is to maximize this reward.

[0034] The learning rate was set to 0.0001 for the actor network and 0.0005 for the critic network; the discount factor γ = 0.99; the experience replay buffer size was 1,000,000; the batch size was 64; the soft update coefficient τ = 0.001; and the exploration noise used was the Ornstein-Uhlenbeck process with parameters μ = 0, σ = 0.1, and θ = 0.15. Model training was considered converged when the standard deviation of the average reward over 100,000 consecutive time steps was less than 0.005, or the average reward growth rate was less than 0.01%.

[0035] During the first 500,000 steps of the initialization phase, the actor network's strategy was very basic. Simulating a summer workday, it set the chiller outlet water temperature to a minimum of 7.0°C during the peak electricity price period at 2:00 PM. This caused a sharp increase in instantaneous power consumption, resulting in an extremely low reward. Due to high electricity costs and equipment wear and tear penalties, the single-step reward was -20. Furthermore, it was unable to effectively cope with fluctuations in production load, causing workshop temperatures to frequently exceed the comfort zone of 21-23°C, further reducing the reward. During the learning phase, from 500,000 to 1.5 million steps, the actor network gradually learned basic energy-saving strategies. It began to schedule more energy-intensive cooling operations during the nighttime electricity price trough and learned to moderately increase the temperature setpoint while meeting the minimum comfort level. During this phase, the control strategy it output was able to increase the average single-step reward from -20 to around -5, with the reward curve showing a clear upward trend. After 1.5 million steps in the convergence phase, the actor-critic mechanism began to discover more refined optimization strategies. The network not only learned to "smooth out" peaks but also fine-tuned the frequency of pumps and fans based on the next hour's weather forecast, such as cloud cover reducing solar radiation, and production schedules, such as equipment downtime for maintenance in a certain area. For example, it learned to dynamically adjust the frequency of a chilled water pump from a constant 80% to a range of 65%-85%, which alone could result in additional energy savings. At this point, the average reward per step stabilized at around -2, and over the subsequent 500,000 steps of training, the reward curve remained largely flat, with minimal fluctuations.

[0036] Furthermore, training is completed when the average reward value no longer shows a statistically significant increase over 100,000 consecutive time steps. At this point, the optimized actor network is saved as a separate file, which contains the optimal strategy knowledge for various working conditions and is the final energy efficiency optimization strategy.

[0037] During the actual deployment phase, the DDPG algorithm was integrated into the plant's central control model. Here's an example of the actual workflow at 2:00 PM on a summer weekday: At 2:00 PM, the control model collected and integrated the current plant status from various data sources to form a standardized input vector. Specifically, it included: Internal environmental conditions: average temperature of area A is 22.8°C, average temperature of area B is 22.5°C, humidity is 55%RH, etc. Production equipment status: All 50 CNC machine tools in Area A are running, with an average load rate of 90%; 25 of the 30 CNC machine tools in Area B are running, with an average load rate of 80%; chiller_01 operates at 160kW, etc. External environment information: outdoor real-time temperature 34°C, total solar radiation 850W / m², etc. Grid and calendar information: The current peak electricity price is 1.2 yuan / kWh, on a weekday.

[0038] Further, model inference and action output: This input vector is fed into the deployed actor network. The network completes a forward propagation calculation within milliseconds and outputs a deterministic optimal control action vector that is free of exploration noise. Specifically, it is: [8.5, 0.75, 0.82, 0.90, 0.88]. This vector is parsed and converted into a specific, readable sequence of control actions: The first value, 8.5, corresponds to adjusting the chiller outlet water temperature setpoint to 8.5°C. The model balances the near-limit room temperature with the high peak electricity price, selecting a setpoint that is more energy-efficient than the conventional 7°C. The second value, 0.75, corresponds to setting the chilled water pump variable frequency frequency to 75%; the third value, 0.82, corresponds to setting the cooling water pump variable frequency frequency to 82%; the fourth value, 0.90, corresponds to setting the fan frequency of the air conditioners in Area A to 90%. Because Area A has a higher temperature, increased cooling air delivery is required; and the fifth value, 0.88, corresponds to setting the fan frequency of the air conditioners in Area B to 88%.

[0039] Furthermore, a high-level optimization instruction is issued: This action sequence is encapsulated into a high-level optimization instruction and sent to the EVK energy efficiency cabinet via the BACnet protocol. The instruction content is as follows: "Chiller outlet water temperature setpoint = 8.5 degrees Celsius, chilled water pump frequency = 75%, cooling water pump frequency = 82%, area A air conditioning box fan frequency = 90%, area B air conditioning box fan frequency = 88%." Furthermore, the controller of the EVK energy efficiency cabinet receives and parses the instruction, converting it into a physical control signal for the underlying equipment. For example, it outputs a 7.5V analog signal to the inverter of the chilled water pump to correspond to a frequency of 75%.

[0040] Afterwards, at 14:15:00, the model will collect new real-time status data again and repeat the above process to achieve continuous, dynamic and intelligent rolling optimization control of building energy efficiency.

[0041] Reinforcement learning is used to autonomously generate energy efficiency optimization strategies. By using the digital twin model as a simulation environment, a clear state space and action space are defined, and a reward function is constructed that incorporates energy consumption, comfort, equipment lifespan, and carbon emissions. Through iterative interaction and trial-and-error learning, the reinforcement learning agent can autonomously explore and generate optimal energy efficiency strategies for dynamic operating conditions and external environmental changes, significantly improving the system's adaptability and optimization capabilities.

[0042] Furthermore, the model continuously compares the actual measurements of the physical plant's sensors with the predicted values ​​from the digital twin model at a fixed five-minute interval. When discrepancies are detected, the model uses Kalman filtering to immediately correct the state variables within the model to ensure synchronization with the actual state of the physical world. The model then tracks the accuracy of its predictions over time. If the model's predicted energy consumption consistently falls short of actual consumption, this indicates that the efficiency of a particular piece of equipment is declining due to aging. At this point, the model automatically triggers a parameter estimation process to fine-tune the efficiency parameters corresponding to that piece of equipment in the knowledge graph, ensuring that the model more accurately reflects its actual performance.

[0043] Furthermore, structural information can be modified. When the physical plant changes, such as adding a new production line, operations and maintenance personnel only need to enter the new equipment information into the model. The model automatically updates its knowledge graph and neural network structure, incorporating the thermal load characteristics and energy consumption patterns of the new equipment into future simulations and predictions. Before the updated model is officially put into use, it undergoes rigorous closed-loop testing using preset test cases and simulated data. This is then compared to real data. If there is a mismatch, the knowledge is marked as "pending verification" and operations and maintenance personnel are prompted to perform proofreading. This establishes a self-verifying closed-loop to prevent catastrophic prediction or control errors caused by erroneous structural updates, greatly enhancing the robustness of the system.

[0044] The model's automated, synchronous updates and fully adaptive correction process integrates heterogeneous data through semantic mapping and knowledge graphs, and utilizes online machine learning algorithms to continuously monitor deviations between the physical entity and the model. This mechanism automatically calibrates, estimates, and adjusts the model's state variables, parameters, and structural information, ensuring the logical consistency of the digital twin model with actual building operations and the continued accuracy of simulation and optimization results.

[0045] Furthermore, the load forecasting model uses the digital twin model in the early morning of each day, combined with the weather forecast, production plan and electricity price information for the next day, to predict in detail the cooling, heating and electricity load demand curves every 15 minutes in the next 24 hours.

[0046] The rolling optimization solution is based on the predicted load, and the model is constructed and solves the following optimization problem.

[0047] The objective function is to find a control solution that minimizes total operating costs over the next 24 hours while meeting all production and environmental requirements. Constraints are imposed by the strict adherence to a series of boundary conditions during the solution process. These define the normal operating range of building functions, such as ensuring that workshop temperatures remain within the required comfort range and limiting the frequency of production equipment starts and stops to prevent overload, in order to comply with power restrictions issued by the power grid.

[0048] The final output of the optimization calculation is a detailed set of control action sequences for the next 24 hours, which clearly states what adjustments to make at what time. For example, the sequence will clearly indicate in which mode the air conditioning model should be set to operate to "store cold" during the early morning electricity price valley period, and how the settings should be adjusted to save costs during the afternoon electricity price peak period.

[0049] The digital twin model integrates historical and real-time data with external forecasts, such as weather, production plans, and energy prices, to accurately predict electricity, heating, and cooling loads. Using mixed-integer linear programming, a rolling optimization solution is employed to generate the control action sequence that minimizes energy consumption within the forecast horizon, while meeting constraints such as comfort, equipment operating limits, and grid demand response. This enables forward-looking and cost-effective energy scheduling.

[0050] Furthermore, the specific steps for event identification include: The model continuously monitors unexpected critical events that impact energy efficiency during real-time operation. During the afternoon production peak, the factory suddenly received an urgent order, causing the equipment load in Area B to far exceed the original plan. The model detected a sharp increase in power consumption in this area within a short period of time and identified it as a "sudden change in production load" event.

[0051] Reference Figure 3 , which is a flow chart of the steps for identifying key events in the first embodiment of the present invention. Once an event is triggered, the digital twin model will immediately simulate the impact of the event on the future environment. For example, it predicts that the temperature in area B will exceed the standard within half an hour. The event-driven MPC will be activated immediately. The originally planned control sequence will be abandoned, and based on the current sudden actual situation and the latest simulation results of the model, the short-term optimization calculation will be completed within a few seconds to one minute, and a high-level optimization instruction will be generated. The result of this emergency re-optimization is a new, well-defined high-level optimization instruction, specifically "In order to quickly suppress the temperature rise in area B, immediately start the standby chiller and at the same time adjust the air supply volume of the air conditioner in area B to the maximum."

[0052] Furthermore, for example, a sudden change in production load is defined as a 5-minute increase in total power consumption of any major production area exceeding 30% or a decrease exceeding 20% ​​compared to the previous period. If total power consumption decreases after an equipment upgrade, an LSTM-based sequence anomaly detection model can be used to learn patterns in historical normal operation data and identify deviations from normal trajectory as increases of more than 20% or decreases of more than 10% compared to the previous period, thus identifying these as potential critical events.

[0053] By continuously monitoring and identifying critical events such as sudden load changes, extreme weather, and equipment failures, the model simulates the consequences of these events and rapidly restarts optimization calculations based on future control action sequences. This mechanism enables the system to respond quickly and accurately to emergencies, generating high-level optimization instructions tailored to the event. This avoids the lag inherent in traditional periodic control, ensuring the real-time and robust nature of energy efficiency optimization.

[0054] Furthermore, this high-level optimization instruction is securely and rapidly transmitted from the upper-level server to the on-site EVK energy efficiency cabinet via a standardized building automation communication protocol. The controller within the EVK energy efficiency cabinet receives and understands the intent of this high-level instruction. It then decomposes this abstract instruction into specific, quantified control signals that can be directly executed by lower-level devices. For example, "Start the standby chiller" is converted into an electrical signal that closes a relay, while "Adjust air flow to maximum" is converted into a specific voltage signal that causes the fan inverter to operate at maximum speed.

[0055] The EVK energy efficiency cabinet transmits these low-level control signals to on-site actuators through its output modules. Once the equipment is activated, sensors within the plant transmit the latest environmental and equipment status data back to the digital twin platform, forming a complete intelligent control closed loop of "perception-analysis-decision-execution-feedback" to ensure effective control.

[0056] High-level optimization instructions guide the EVK energy efficiency cabinet to achieve intelligent control. These high-level instructions are transmitted to the EVK energy efficiency cabinet via standard communication protocols. The cabinet controller converts them into low-level control signals that can be directly executed by the HVAC, lighting, and shading models, such as set point adjustment, start and stop, and variable frequency control. Ultimately, the actuators receive and execute these signals, achieving a seamless connection from intelligent decision-making to physical execution, ensuring the effective implementation of optimization strategies, thereby improving the automation and intelligence level of building operations and maximizing energy savings.

[0057] This example details the application of digital twins in the dynamic simulation of energy efficiency in precision machining industrial plants. First, a highly realistic digital twin model is constructed by integrating the plant's static physical properties, semi-static production plans, real-time environmental sensor data, equipment operating status, and external meteorological and power grid information. Second, a neural symbolic approach is employed to design the model. The neural network module is responsible for data pattern recognition and behavior prediction, while the symbolic reasoning module incorporates building physics laws and engineering rules to ensure strategy rationality. The knowledge graph module unifies data semantics to enhance energy management capabilities. Then, using the digital twin model as a virtual environment, an intelligent agent is trained through reinforcement learning, iteratively generating optimal energy efficiency strategies guided by a comprehensive reward function. Ultimately, an intelligent control closed loop is formed. The model features automatic synchronous updates and full-dimensional correction capabilities, predicting energy consumption and continuously optimizing energy scheduling. An event-driven multi-processor (MPC) rapidly responds to emergencies and issues resolution instructions, which are then transmitted to the EVK energy efficiency cabinet via a standard protocol for execution. This enables refined, intelligent, and optimized control of the building energy model, significantly reducing energy consumption while ensuring a stable production environment.

[0058] Example 2: This implementation is applied to a large-scale biopharmaceutical cleanroom. With a floor area of ​​20,000 square meters, the facility has extremely stringent requirements for cleanliness, temperature, humidity, and pressure differentials in the production environment. The facility's energy system is highly complex, involving four major energy sources: electricity, process cooling water, HVAC chilled water, and steam. A self-contained gas-fired combined heat and power (CHP) unit is also used as a supplemental energy source.

[0059] This example achieves the goal of dynamic simulation and intelligent optimization control of the multi-energy system within a biopharmaceutical plant by constructing an energy efficiency digital twin model of the plant. First, the model is obtained from the historical database and design documents of the EVK energy efficiency cabinet. A three-dimensional plant model is accurately mapped to the layout of each clean area (such as Class A / B / C / D areas), buffer rooms, power stations, and storage areas. Specifically, the pressure gradient requirements between clean areas are included, with Class B areas requiring a +10 Pa relative to Class C areas. The enclosure structures and materials include heat transfer and airtightness parameters for special enclosures such as cleanroom panels and high-efficiency filters. Production equipment and process parameters include the model, design power, and heat load characteristics of bioreactors, chromatography columns, purified water systems, and online cleaning / online sterilization stations. Specifically, the production process flow for different drug batches is obtained from the manufacturing execution system, such as the duration and target temperature of the fermentation phase, and the steam pressure, flow rate, and duration required for the SIP phase.

[0060] Energy equipment parameters include the central air-conditioning system, including chillers, water pumps, large modular air-conditioning units, steam boilers, and rated capacity, efficiency curves, and start-stop constraints of CHP units.

[0061] Furthermore, in addition to temperature, humidity, and CO2 concentration, it also includes suspended particle concentrations in each clean area, sedimentation bacteria data, and micro-pressure differences between key areas. The collection frequency is increased to every second to strictly test the sterile conditions of the internal environment of the plant; real-time monitoring of the stirring power and temperature of the bioreactor; the power generation, heat production, and gas consumption of the CHP unit; the outlet pressure and flow of the steam boiler; and the variable frequency and current of water pumps and fans at all levels.

[0062] A graph neural network was used to model the plant's energy-process network. Nodes represented reactors, chillers, and CHPs; edges represented physical connections, such as pipes and circuits; and logical associations represented that the cooling needs of a reactor were met by a chiller. Good Manufacturing Practice regulations and standard operating procedures (SOPs) were encoded as logical rules and used as constraints in the optimization process to ensure the rationality and compliance of the optimization strategy.

[0063] Build a deeper knowledge network, connecting not only "equipment-space" but also "batch-process-equipment-energy." For example, the knowledge graph clearly shows that "Batch BA202405" is in the "fermentation stage," which requires "reactor R-101" to be maintained at "37±0.5°C." This stage's cooling is provided by "chilled water loop L1," and the cooling capacity of L1 can ultimately be traced back to "chiller CH-02" or "CHP waste heat cooling unit."

[0064] Furthermore, the highly coupled digital twin model constructed based on the above-mentioned neural symbolic method sets the state space to include the environmental parameters of all clean areas, the pressure difference status, the current process stage of the production batch, the operating mode of the CHP unit and boiler, and the real-time pressure and temperature of the chilled water or steam pipeline network; the action space includes adjusting the heat-to-electricity ratio of the CHP, deciding to start / stop the standby boiler, adjusting the AHU air supply volume to finely control the pressure difference, and pre-preparing or storing cold or heat to cope with the upcoming process peak.

[0065] The reward function adopts a multi-objective optimization design, coordinating the minimization goals of energy costs and carbon emissions through weighted coefficients, and setting two types of rigid constraints. First, for key parameters of pharmaceutical production quality management specifications, such as pressure difference and temperature, if they deviate from their allowable range, the system will impose a very large negative penalty value to ensure that energy optimization is absolutely subject to pharmaceutical quality and safety requirements; second, a progressive penalty mechanism is set for delays in the production process rhythm, ensuring that production efficiency is not affected by energy efficiency optimization through gradient penalties.

[0066] Furthermore, the model continuously monitored that the differential pressure across a particular HEPA filter was increasing at a rate exceeding the theoretical attenuation curve. The online machine learning algorithm identified this as an early sign of filter clogging. The model automatically corrected the filter's air resistance parameters and issued an early warning to the operations and maintenance system. When a new antibody-drug conjugate production line was added to the factory, operations and maintenance personnel simply entered the new line's equipment list and process flow into the system. The knowledge graph and graph neural network models automatically expanded to incorporate the new line's energy requirements and process constraints into the overall factory optimization model, enabling plug-and-play model expansion.

[0067] Based on the production plan provided by the MES, the digital twin model performs a seven-day multi-energy load forecast. A mixed-integer linear programming solver integrates fluctuations in electricity and gas prices to develop hourly start and shutdown plans for equipment such as CHPs and boilers. When a process control system signal triggers an abnormal process event, such as "urgent draining of the bioreactor," the model immediately simulates a surge in demand for steam and water for injection. This triggers the event-driven MPC algorithm, which abandons the original plan and completes a two-hour short-term optimization within minutes, generating new instructions: "Switch the CHP to full load based on heat-to-power, start the backup boiler, and reduce the air conditioning load in non-core areas." The instructions are transmitted to the EVK energy efficiency cabinet, where they are converted into underlying control signals, which are then used as communication protocol instructions to control the CHP, dry contacts to start the boiler, and analog signals to adjust the AHU inverter. The execution results are fed back through sensors, forming a closed-loop control system.

[0068] This embodiment is applied in high-end manufacturing industries such as biopharmaceuticals. Due to the sterile environment of the production plant and real-time supervision of each link in the drug production process, the plant building is simulated through the constructed digital twin model, and optimal adjustment instructions are given based on the predicted information related to the future environment. The complete introduction of this embodiment not only achieves energy conservation and consumption reduction, but also ensures production safety and quality through the two-way coordination of energy efficiency and process, demonstrating the outstanding value of intelligent energy management.

[0069] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A dynamic simulation method for building energy efficiency based on digital twins, characterized in that: include: A neural symbolic approach is used to design a digital twin model for building energy efficiency. By combining building physics with multi-dimensional comprehensive datasets, a dynamic hybrid self-evolving network is constructed. The dynamic hybrid self-evolving network uses reinforcement learning to conduct interactive simulation based on real-time sensor data, external environment changes, and production equipment aging to autonomously generate energy efficiency optimization strategies; Assign a unified semantic framework and knowledge association to all data in the building energy efficiency digital twin model, and use machine learning algorithms to automatically synchronize and update the model and perform full-dimensional adaptive correction based on the building's multi-dimensional comprehensive data set. The correction includes correction of state variables, model parameters, and structural information. The updated and revised digital twin model is used to predict future energy consumption, and the future control action sequence is obtained by following the constrained rolling optimization control instructions. The time domain behavior of the model is simulated based on the future control action sequence, and future scenarios are predicted. The event-driven MPC algorithm is used to optimize the future scenarios and generate high-level optimization instructions to guide the EVK energy efficiency cabinet to achieve intelligent control.

2. A building energy efficiency dynamic simulation method based on digital twins according to claim 1, characterized in that: The building physics mechanism is to obtain static and semi-static data on industrial building geometry, enclosure material properties, production equipment models and design parameters, production plans and shift patterns from the historical database of EVK energy efficiency cabinets; The sources and types of the multi-dimensional comprehensive dataset include: using various sensors and instruments deployed inside industrial buildings to obtain real-time environmental parameters, including production workshop temperature, humidity, CO2 concentration, and light intensity; The real-time operating status of the factory includes the start and stop status, power, current, and electricity and heat energy consumption data of production equipment, pumps, fans, and chillers; external information sources are used to obtain weather forecast data, including outdoor temperature, solar radiation, wind speed, power grid signals and calendar information. The power grid signals include time-of-use electricity prices, peak and valley electricity prices, and demand response instructions. The calendar information includes holidays and weekdays.

3. The method for dynamic simulation of building energy efficiency based on digital twin according to claim 1, characterized in that: The neural symbolic method is used to design a digital twin model of building energy efficiency, which specifically includes: a neural network module for pattern recognition, feature extraction and behavior prediction of the multi-dimensional comprehensive data set; a symbolic reasoning module for encoding and logically representing the building physical mechanisms, engineering rules and domain knowledge; the engineering rules and the domain knowledge are derived from industry standards, expert experience, historical data analysis and basic scientific principles; a knowledge graph module is used to construct a unified semantic framework and knowledge association of the digital twin model of building energy efficiency.

4. The method for dynamic simulation of building energy efficiency based on digital twin according to claim 1, characterized in that: The dynamic hybrid self-evolutionary network performs interactive simulation and autonomously generates energy efficiency optimization strategies, specifically including: using the building energy efficiency digital twin model as the simulation environment, defining the state space as the internal environmental state of the building, external environmental parameters, and real-time operating status of the plant; defining the action space as HVAC model parameter set points, lighting switches and brightness adjustments, and production equipment start and stop control instructions; using preset energy consumption, production workshop comfort, production equipment operating life and carbon emission indicators to construct the reinforcement learning reward function; and obtaining energy efficiency optimization strategies for dynamic operating conditions and external environmental changes through iterative interaction and trial-and-error learning by reinforcement learning agents in the simulation environment of the model.

5. The method for dynamic simulation of building energy efficiency based on digital twin according to claim 1, characterized in that: The specific steps of the automatic synchronous update and the full-dimensional adaptive correction include: using semantic standards and knowledge graphs to semantically map, integrate and standardize the internal heterogeneous sensor data, production equipment information, spatial topology and model configuration data of the building; using online machine learning algorithms to continuously monitor the real-time deviation between the physical entity of the building and the model, and automatically calibrate, estimate and adjust the state variables, model parameters and structural information of the model.

6. The method for dynamic simulation of building energy efficiency based on digital twin according to claim 1, characterized in that: The specific steps for obtaining the future control action sequence include: using the updated and revised digital twin model as a prediction model, integrating historical operating data and real-time sensor data, and combining external environment predictions, including weather forecasts, production scheduling plans and energy price information, to predict the electricity, heating and cooling load energy consumption of the building; using mixed integer linear programming, taking the production workshop temperature comfort range, production equipment operating power range, and grid demand response signal within the prediction time domain as constraints, and performing rolling optimization solution to obtain the future control action sequence that can minimize energy consumption and meet the constraints within the prediction time domain.

7. The method for dynamic simulation of building energy efficiency based on digital twin according to claim 1, characterized in that: The event-driven MPC is used to optimize the future scenario and generate high-level optimization instructions, specifically including: continuously monitoring the real-time operating status of the building and changes in the external environment, identifying key events that are preset and / or learned through machine learning models, including sudden changes in production load, process flow adjustments, extreme weather emergencies, production equipment failures and significant performance degradation; the model will simulate and generate simulation results after the key events are triggered; and using the event-driven MPC algorithm, based on the future control action sequence and the simulation results, restarting the optimization calculation to obtain high-level optimization instructions after the key events are triggered.

8. A building energy efficiency dynamic simulation system based on digital twins, characterized by: include: The digital twin model construction and self-evolution module is used to design a digital twin model of building energy efficiency using a neural symbolic approach. By combining building physics mechanisms with multi-dimensional comprehensive data sets, a dynamic hybrid self-evolutionary network is constructed. The dynamic hybrid self-evolving network uses reinforcement learning to conduct interactive simulation based on real-time sensor data, external environment changes, and production equipment aging to autonomously generate energy efficiency optimization strategies; The data semantics and model adaptive correction module is used to give a unified semantic framework and knowledge association to all data in the building energy efficiency digital twin model. According to the multi-dimensional comprehensive data set of the building, the machine learning algorithm is used to automatically and synchronously update and fully adaptively correct the model. The correction includes correction of state variables, model parameters and structural information; the energy consumption prediction and intelligent optimization control module is used to use the updated and corrected digital twin model to predict future energy consumption, follow the constrained rolling optimization control instructions, and obtain the future control action sequence; based on the future control action sequence, the time domain behavior of the model is simulated, future scenarios are predicted, and the future scenarios are optimized with the help of event-driven MPC algorithms. High-level optimization instructions are generated to guide the EVK energy efficiency cabinet to achieve intelligent control.

Images