Building energy consumption dynamic prediction method and system based on digital twin calibration

Abstract

The application discloses a building energy consumption dynamic prediction method and system based on digital twin calibration, relates to the technical field of digital twin, and comprises the following steps: constructing a building digital twin multi-level model, collecting data and performing multi-time scale benchmark prediction, extracting the space-time distribution characteristics of prediction errors, performing hierarchical calibration decision based on the space-time distribution characteristics of errors, performing hierarchical calibration and updating the digital twin model, calculating a calibration utility function and evaluating calibration effect, performing weighted dynamic prediction, and iterative circulation; the system comprises a data collection and preprocessing module, a digital twin model management module, a multi-time scale prediction engine, an error space-time feature extraction module, a hierarchical calibration decision module, a gradual parameter calibration module, a calibration utility evaluation module, a confidence atlas management module, a weighted prediction fusion module, and a closed-loop iterative control module. The application can accurately calibrate the digital twin model in layers, and continuously optimize the prediction accuracy through closed-loop iteration.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to a method and system for dynamic prediction of building energy consumption based on digital twin calibration. Background Technology

[0002] As building energy consumption continues to rise as a proportion of final energy consumption, accurate prediction of building energy consumption has become a key aspect of building energy-efficient operation and intelligent management. Digital twin technology, by constructing a virtual mirror image of the physical building, provides a new technological approach for building energy consumption prediction.

[0003] In recent years, several technical solutions have been proposed for building energy consumption prediction based on digital twins. For example, some existing technologies generate the start time of step excitation by identifying equipment state transitions, calculate the lag time of the physical and virtual environmental thermal responses, and correct thermal inertia parameters based on the time lag deviation to establish a dynamic calibration model. Other solutions update the digital twin model when the difference between the predicted and actual electricity consumption at the target time is greater than or equal to a preset difference threshold. Still other solutions predict the system state based on a state-space model, calculate the residuals between the prediction and observation, and dynamically generate a confidence coefficient for weighted fusion. Furthermore, some existing technologies employ physical information neural networks to embed the building heat conduction equation into a neural network loss function, constructing a physical-data fusion digital twin model.

[0004] However, the aforementioned existing technologies all have significant limitations. First, the calibration triggering mechanisms of existing technologies are too simplistic, relying either on deviations in a single physical quantity or on whether the prediction error exceeds a preset threshold. They lack the ability to identify the multidimensional sources of prediction errors and cannot accurately determine whether the deviation stems from geometric model distortion, physical parameter shifts, or changes in operational behavior patterns. Second, the calibration objects are homogeneous. Existing solutions typically perform overall parameter correction on the digital twin model, failing to differentiate calibration for different levels of the model, leading to a "ripple effect"—correcting a deviation in one aspect introduces errors in other areas. Third, the relationship between calibration and prediction is a unidirectional, sequential process, i.e., "calibrate first, then predict," lacking a bidirectional closed-loop mechanism where "prediction results guide calibration in reverse," thus failing to achieve continuous self-optimization of the model. Finally, existing technologies do not consider the differentiated impact of calibration operations on prediction accuracy at different time scales. A single calibration may improve ultra-short-term prediction accuracy but impair medium-term prediction accuracy; there is a lack of quantitative assessment and adaptive control of this trade-off.

[0005] To address the above problems, this invention proposes a method and system for dynamic prediction of building energy consumption based on digital twin calibration. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for dynamic prediction of building energy consumption based on digital twin calibration, so as to solve the technical problems in the prior art, such as a single calibration trigger mechanism, homogeneity of calibration objects, lack of two-way closed loop between calibration and prediction, and lack of quantitative evaluation of calibration utility across multiple time scales.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] On the one hand, the method for dynamic prediction of building energy consumption based on digital twin calibration includes the following steps:

[0009] Step S1: Construct a multi-level digital twin model of the building. The multi-level model includes three levels: geometric layer, physical layer, and behavioral layer. The geometric layer contains the geometric dimension information and spatial topology relationship extracted from the building BIM model. The physical layer contains the thermal parameters of the building envelope and the performance parameters of the equipment. The behavioral layer contains the parameters of personnel work and rest patterns and equipment usage patterns.

[0010] Step S2: Collect real-time building operation data, including indoor and outdoor environmental data, equipment operation status data, and actual energy consumption data. The sampling frequency of real-time data is pre-configured during system initialization based on the physical response characteristics of the sensors. Meteorological forecast data is updated at a fixed period. Data from different sources are input into the prediction engine after time alignment. Based on the current digital twin model, perform benchmark energy consumption predictions for three time scales: ultra-short-term, short-term, and medium-term.

[0011] Step S3: Calculate the prediction error for each time scale and each spatial region, and construct the error spatiotemporal distribution feature vector. Each element in the error spatiotemporal distribution feature vector corresponds to the error feature value of a combination of a time scale and a spatial region. Based on the distribution pattern of the error spatiotemporal distribution feature vector, determine the source level of the prediction deviation.

[0012] Step S4: Based on the deviation source level determined in Step S3, perform a hierarchical calibration decision: if the error is mainly concentrated in a specific spatial region, it is determined to be a geometric layer or local physical parameter deviation and a spatial local calibration is triggered; if the error systematically increases with the extension of the prediction time, it is determined to be a physical parameter offset and a physical layer global calibration is triggered; if the error shows a periodic fluctuation pattern, it is determined to be a behavioral layer parameter change and a behavioral layer calibration is triggered; if the error distribution has no obvious pattern and deviations exist at all levels, a multi-level joint calibration is triggered.

[0013] Step S5: Based on the calibration decision in Step S4, only the model levels that need calibration are adjusted, and the target parameters are gradually adjusted using a progressive parameter adjustment strategy. The progressive coefficient increases linearly with the number of iterations, and smoothly approaches 1 from an initial small value. When a sudden and sharp increase in the prediction error is detected, the calibration operation is suspended, the model parameters remain unchanged, and the prediction confidence interval is widened accordingly in Step S7 to reflect the increased uncertainty. After the error remains high for several consecutive calibration cycles, the hierarchical calibration decision in Step S4 is triggered again. After confirming the source of the deviation, the calibration is performed according to the normal progressive parameter adjustment process.

[0014] Step S6: Calculate the calibration utility function and evaluate the changes in ultra-short-term, short-term, and medium-term prediction accuracy after calibration; if the overall calibration utility function value is lower than the preset threshold, roll back the calibration operation or reduce the calibration intensity.

[0015] Step S7: Update the confidence scores of each sub-module of the digital twin model and construct a digital twin confidence map; when performing the next round of dynamic energy consumption prediction, weight and fuse the prediction results according to the confidence scores of each sub-module, and output the final energy consumption prediction results and the corresponding confidence intervals.

[0016] Step S8: Compare the prediction results of step S7 with the actual energy consumption data, and return to step S3 to form a continuous closed-loop iteration that sequentially performs prediction, error spatiotemporal feature extraction, hierarchical calibration decision, model update and re-prediction.

[0017] Further, in step S3, the method for constructing the error spatiotemporal distribution feature vector is as follows: dividing the building space into... The region is divided into three scales: ultra-short-term, short-term, and medium-term. For each scale... ,area Combined calculation of error eigenvalues ,constitute dimensional error spatiotemporal distribution eigenvector:

[0018]

[0019] Among them, error eigenvalues This represents the normalized mean absolute percentage error.

[0020] Furthermore, in step S4, the hierarchical calibration decision uses a classifier based on the error spatiotemporal distribution feature vector to determine the source of deviation. The classifier is trained using historical calibration records or offline simulation data, and its input is the error spatiotemporal distribution feature vector, and its output is the deviation source hierarchical label.

[0021] Furthermore, in step S5, the progressive parameter adjustment strategy is as follows: in the first... The actual adjustment amount of the target parameter during the second calibration. Defined as:

[0022]

[0023] in, Adjustment amount for theoretical target; Let be the asymptotic coefficient, satisfying ,and It monotonically approaches 1 as the number of calibration iterations increases.

[0024] Further, in step S6, the expression for the calibration utility function is:

[0025]

[0026] in, , , These represent the improvement in prediction accuracy for the ultra-short term, short term, and medium term after calibration, respectively. , , The corresponding weight coefficients, and satisfying ;when If this occurs, the calibration is deemed invalid. This is a preset utility threshold.

[0027] Further, in step S7, the method for updating the digital twin confidence map is as follows: for the first... Each submodule has a confidence score calculated based on the moving average of prediction errors over several recent time windows:

[0028]

[0029] in This is the normalized average prediction error for the region corresponding to this submodule over the most recent preset number of time windows, with a value range of [0,1]. It is a very small positive number, preferably This is used to avoid numerical calculation anomalies rather than to reflect the model's reliability; the confidence scores of each submodule are normalized and used as the weighted fusion weights for the next round of predictions, while the average confidence score of each submodule is used as the overall confidence score. When outputting the final energy consumption prediction results, the standard deviation of the prediction results of each submodule is used as the base interval radius, and the final confidence interval is:

[0030]

[0031] Among them, scaling function This linear scaling mechanism maintains the core intent of "widening the interval with low confidence" while avoiding reciprocal scaling. The extreme interval expansion that occurs at moderately low confidence levels ensures that the prediction interval has practical decision-making reference value across the full confidence range.

[0032] On the other hand, a building energy consumption dynamic prediction system based on digital twin calibration includes:

[0033] The data acquisition and preprocessing module is used to collect real-time data on the building's indoor and outdoor environment, equipment operating status, and actual energy consumption, and to perform data cleaning and time alignment.

[0034] The digital twin model management module is used to store and manage digital twin models with a three-layer architecture, including a geometric layer, a physical layer, and a behavioral layer, and supports reading and modifying layer parameters.

[0035] A multi-timescale forecasting engine is used to perform energy consumption forecasts at three timescales: ultra-short-term, short-term, and medium-term, based on the current digital twin model.

[0036] The error spatiotemporal feature extraction module is used to calculate the prediction error at each time scale and in each spatial region, construct the error spatiotemporal distribution feature vector, and determine the source level of the prediction deviation.

[0037] The hierarchical calibration decision module is used to perform hierarchical calibration decisions based on the error spatiotemporal distribution feature vector to determine the model level and calibration intensity that need to be calibrated.

[0038] The progressive parameter calibration module is used to perform progressive parameter corrections at specified model levels.

[0039] The calibration utility evaluation module is used to calculate the calibration utility function, evaluate the comprehensive impact of calibration operations on the prediction accuracy at various time scales, and decide whether to perform or roll back the calibration.

[0040] The confidence graph management module is used to maintain and dynamically update the confidence scores of each sub-module of the digital twin model.

[0041] The weighted prediction fusion module is used to weight and fuse the prediction results according to the confidence level of each sub-module, and output the final energy consumption prediction result and confidence level range.

[0042] The closed-loop iterative control module is used to coordinate the runtime sequence of the above modules, forming a continuous closed-loop iteration that sequentially executes prediction, error spatiotemporal feature extraction, hierarchical calibration decision, model update, and re-prediction.

[0043] In the digital twin model management module, the geometry layer contains geometric dimension information and spatial topology relationships extracted from the building BIM model, the physical layer contains thermal parameters of the building envelope and equipment performance parameters, and the behavior layer contains personnel work and rest patterns and equipment usage pattern parameters. Parameters are transferred between each layer through standardized interfaces, and modifications to parameters at any layer do not affect the model structure of other layers.

[0044] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0045] This invention proposes a hierarchical calibration mechanism driven by the spatiotemporal distribution characteristics of errors. Unlike the calibration triggering method in the prior art that relies on the deviation of a single physical quantity or the prediction error threshold, this invention constructs a spatiotemporal distribution feature vector of errors to finely identify the source of prediction deviation from both time and space dimensions, thereby achieving accurate hierarchical calibration oriented towards the source of errors. This overcomes the defects of parameter coupling interference and neglecting one aspect in the overall unified calibration.

[0046] This invention constructs a closed-loop iterative framework based on alternating prediction and calibration cycles. Unlike the unidirectional serial relationship between calibration and prediction in existing technologies, this invention continuously feeds the prediction results back to the calibration decision-making stage, forming a positive feedback loop for model self-optimization, enabling the digital twin model to continuously approach the real state of the physical building during continuous operation.

[0047] This invention introduces a multi-timescale quantitative evaluation mechanism for calibration utility. By constructing a calibration utility function, it simultaneously evaluates the comprehensive impact of calibration operations on the accuracy of ultra-short-term, short-term, and medium-term predictions. When calibration has a negative impact on a certain timescale, it automatically adjusts or rolls back to ensure that the calibration behavior serves the global optimum rather than the local optimum.

[0048] This invention proposes a digital twin confidence graph and weighted prediction method. By independently evaluating the confidence of each sub-module of the digital twin model and then weighting and fusing them according to the confidence during prediction, it realizes differentiated trust and intelligent utilization of the digital twin model, significantly improving the reliability and robustness of the prediction results. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the process steps of the building energy consumption dynamic prediction method based on digital twin calibration of the present invention;

[0050] Figure 2 This is a schematic diagram of the structure of the digital twin multi-level model in this invention;

[0051] Figure 3 This is a flowchart illustrating the construction of the spatiotemporal distribution feature vector of errors and the hierarchical calibration decision-making process in this invention;

[0052] Figure 4This is a schematic diagram of the overall architecture of the building energy consumption dynamic prediction system based on digital twin calibration of the present invention. Detailed Implementation

[0053] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0054] Please refer to Figure 1 , Figure 1 This is a schematic diagram of the process steps of the building energy consumption dynamic prediction method based on digital twin calibration of the present invention, including step S1 constructing a multi-level digital twin model of the building, step S2 collecting real-time building operation data and performing multi-timescale benchmark prediction, step S3 extracting the spatiotemporal distribution characteristics of prediction errors, step S4 performing hierarchical calibration decision based on the spatiotemporal distribution characteristics of errors, step S5 performing hierarchical calibration and updating the digital twin model, step S6 calculating the calibration utility function and evaluating the calibration effect, step S7 performing weighted dynamic prediction based on confidence spectrum, and step S8 iterative loop.

[0055] Step S1 involves constructing a multi-level digital twin model of the building, comprising three layers: a geometric layer, a physical layer, and a behavioral layer. Specifically, the geometric layer extracts geometric dimensions and spatial topological relationships from the building's BIM model, including geometric parameters such as building area, floor height, room layout, window location and size, and wall thickness for each floor. The physical layer contains thermal parameters of the building envelope and equipment performance parameters. Thermal parameters of the building envelope include heat transfer coefficients, heat capacity, and solar radiation absorption coefficients of components such as walls, roofs, floors, and windows; equipment performance parameters include the rated power, energy efficiency ratio, and part-load performance curves of the HVAC system. The behavioral layer contains parameters related to occupant activity patterns and equipment usage patterns, including the variation of occupancy rates in each area over time, lighting and equipment usage schedules, and variations in indoor temperature setpoints.

[0056] In a preferred embodiment of the present invention, the digital twin multi-level model adopts a three-layer decoupled architecture design, with parameters being passed between each layer through a standardized interface. Modification of parameters at any layer does not affect the model structure of other layers, thereby providing a technical basis for subsequent layered calibration.

[0057] Please refer to Figure 2 , Figure 2This is a schematic diagram of the structure of the multi-level digital twin model in this invention. For example... Figure 2 As shown, the geometry layer is connected to the physical layer through a data interface, and the physical layer is connected to the behavior layer through a data interface. The three-layer model is jointly output to the energy consumption simulation engine, which performs dynamic simulation calculations of building energy consumption based on the input parameters of the three layers.

[0058] Step S2 involves collecting real-time building operation data and performing multi-timescale benchmark prediction. The real-time collected data includes: environmental data such as indoor and outdoor temperature, humidity, and light intensity; operating status and power data of equipment such as HVAC systems, lighting systems, and power equipment; and actual energy consumption data for each area of ​​the building. The collected data undergoes data cleaning to remove outliers and noise, and is then time-aligned. In a preferred embodiment of the invention, the sampling frequency of real-time data is 15 minutes / time, and the update frequency of meteorological forecast data is 1 hour / time. Cubic spline interpolation is used for time alignment between the two to ensure time consistency of data from different sources. It should be noted that the 15 minutes / time refers to the sampling frequency of the temperature sensor; the power sensor samples independently at 1 minute / time; and each type of sensor uses its own preset sampling frequency. These specific values ​​are merely illustrative and can be configured during system initialization according to the building size and prediction accuracy requirements in actual deployment, and do not constitute a limitation of the invention.

[0059] As an optional implementation method, the specific approach to adaptively setting the sampling frequency based on sensor type is as follows: For slowly changing sensors such as temperature and humidity, a fixed-period sampling frequency of 15 minutes / time is configured; for rapidly changing sensors such as power and illuminance, a fixed-period sampling frequency of 1 minute / time is configured; for switch-type sensors related to equipment start / stop status, event-triggered acquisition is configured, meaning the current status and timestamp are recorded only when the status changes. The sampling frequency configuration for each of the above sensors is completed once during system initialization and remains constant throughout the operating cycle, without being adjusted online according to changes in operating conditions; its upper limit is constrained by the transmission bandwidth and storage capacity of the data acquisition system, and its lower limit is constrained by the prediction accuracy requirements. "Adaptive" means that, based on the sensor's own physical response characteristics—namely, slowly changing, rapidly changing, and switch-type—the corresponding preset sampling configuration scheme is automatically selected during the system deployment phase, rather than uniformly using the same frequency.

[0060] Based on the current digital twin model, energy consumption baseline forecasts are performed at three time scales: ultra-short-term, short-term, and medium-term. The ultra-short-term energy consumption baseline forecast covers the next 15 minutes to 1 hour, the short-term forecast covers the next 1 hour to 24 hours, and the medium-term forecast covers the next 1 day to 7 days. The forecasts at all three time scales use the same digital twin model but with different input conditions: the ultra-short-term forecast uses the most recent real-time data as initial conditions, the short-term forecast integrates numerical weather prediction data, and the medium-term forecast uses statistical values ​​from historical meteorological data for the same period.

[0061] In a preferred embodiment of the present invention, to balance computational speed and prediction accuracy, ultra-short-term prediction employs a simplified resistance-capacitance lumped parameter model, i.e., the RC model, for rapid simulation, while short-term and medium-term predictions employ detailed physical models, such as multi-region models based on the thermal balance method, for simulation. Model switching between prediction engines at different time scales is automatically managed by the multi-time-scale prediction engine.

[0062] Step S3: Extract the spatiotemporal distribution characteristics of the prediction error. Divide the building space into N areas according to functional zones, such as by floor or by room function type, where room function types include office areas, meeting rooms, corridors, and equipment rooms, etc. Divide the prediction time into three scales: ultra-short-term, short-term, and medium-term. For each scale... ,area Combine the data and calculate the prediction error for that combination. The error index uses Mean Absolute Percentage Error (MAPE), which is calculated using the following formula:

[0063]

[0064] in, To predict the number of sampling points within the time window, For the region In scale Next moment The actual energy consumption value, These are the corresponding predicted values. After normalizing each error value, the following is formed: dimensional error spatiotemporal distribution eigenvector .

[0065] Based on the distribution pattern of the spatiotemporal distribution feature vector of the error, the source level of the prediction deviation is determined. Specifically, if the error value shows a significant regional clustering in the spatial dimension, that is, the error in a specific region is significantly higher than that in other regions, then the deviation is mainly due to the deviation of the geometric layer or local physical parameters; if the error value shows a systematic increasing trend with the extension of the prediction time scale, that is, medium-term error > short-term error > ultra-short-term error, then the deviation is mainly due to the offset of physical layer parameters such as thermal inertia parameters and heat transfer coefficients; if the error value shows a periodic fluctuation pattern in the time dimension, such as the error pattern being significantly different between working days and non-working days, then the deviation is mainly due to the change of behavioral layer parameters; if the error distribution has no obvious pattern and there are significant deviations at each level, then multi-level joint calibration is required.

[0066] Please refer to Figure 3 , Figure 3 This is a flowchart illustrating the construction of the spatiotemporal distribution feature vector of errors and the hierarchical calibration decision-making process in this invention. Predicted error data is input into the error spatiotemporal feature extraction module. The extraction module outputs the error spatiotemporal distribution feature vector to the hierarchical calibration decision-making module. The decision-making module outputs a calibration decision based on preset deviation source determination rules, where the rule-based calibration decision includes the calibration level and calibration intensity.

[0067] Step S4: Perform hierarchical calibration decisions based on the spatiotemporal distribution characteristics of the error. Based on the judgment result of step S3, perform the following hierarchical calibration decisions: If the judgment deviation mainly originates from geometric layer or local physical parameter deviations, trigger spatial local calibration. In this case, only the geometric parameters or local physical parameters corresponding to the error cluster area are corrected, while the parameters of other areas and levels remain unchanged. The specific method of spatial local calibration is: taking the error cluster area as the center, and using the measured energy consumption data of that area, reversely correct the heat transfer coefficient or thermal inertia parameter of that area, so that the model output of that area approximates the measured value. If the judgment deviation mainly originates from the global parameter offset of the physical layer, trigger physical layer global calibration. In this case, systematically correct the thermal parameter system of the entire building, including the overall adjustment of parameters such as heat transfer coefficient, heat capacity, and solar radiation absorption coefficient. The specific method of physical layer global calibration is: using the multi-timescale measured energy consumption data of all areas as the target, and employing parameter estimation methods such as least squares or Kalman filtering to globally optimize the key parameters of the physical layer. If the judgment deviation mainly originates from changes in behavioral layer parameters, trigger behavioral layer calibration. At this point, parameters related to personnel work and rest patterns or equipment usage patterns are adjusted, including the time distribution of personnel in each area, equipment usage schedules, and the variation patterns of indoor temperature setpoints. The specific method for behavioral layer calibration is as follows: behavioral pattern parameters are identified in reverse based on recent measured data, and current behavioral pattern features are extracted using pattern recognition or time series decomposition methods. If multi-level joint calibration is deemed necessary, calibration is performed sequentially according to the priority order of the geometric layer, physical layer, and behavioral layer. After each level of calibration, the error distribution is reassessed, and if the error meets the requirements, the calibration of subsequent levels is terminated.

[0068] In a preferred embodiment of the present invention, the hierarchical calibration decision uses a classifier based on the spatiotemporal distribution feature vector of the error to automatically determine the source of the deviation. This classifier is a support vector machine or random forest classifier trained on historical calibration records or offline simulation data, and its input is the spatiotemporal distribution feature vector of the error. The output is a hierarchical label of the deviation source, where the hierarchical label includes geometric layer, physical layer, behavioral layer, and multi-level joint labels. The training samples for the classifier are obtained by using offline high-precision simulation as a "virtual real value," artificially setting geometric parameter deviations, physical parameter deviations, and behavioral parameter deviations under different working conditions to generate massive simulation data and their corresponding spatiotemporal error distribution feature vectors, thereby training the classifier; during actual operation, actual calibration records are collected to continuously optimize the classifier parameters. The offline high-precision simulation is, for example, a full-year dynamic simulation of the building based on EnergyPlus. In this invention, the "calibration intensity" is used to control the aggressiveness of correction in a single calibration operation, and it is related to the asymptotic coefficient in step S5. The initial value is related to this. Specifically, it is based on the overall magnitude of the error spatiotemporal distribution characteristic vector. The calibration intensity is divided into three levels: low, medium, and high. For example, but not limited to, the root mean square or average value of the error characteristics of each region. Because this scheme employs a linearly increasing gradual strategy, the calibration intensity is adjusted by setting a preset gradual step size. Right now In Value to achieve: when When the value is in the lower range, indicating that the model's current bias is small and mainly fluctuates within normal limits, a low-intensity calibration should be used, with a larger setting. Values ​​such as , corresponding to the initial The rise is extremely gradual, with minor adjustments made to avoid unnecessary disturbances to the already nearly stable model; when When within the medium range, use medium intensity calibration and set an appropriate value. Values ​​such as , corresponding to the initial Perform parameter corrections for normal amplitude; when When the value is in a high range, indicating significant model bias or substantial changes in building operation, a high-intensity calibration should be used, with a smaller value set. Values ​​such as , corresponding to the initial Furthermore, it only requires 5 iterations to reach 1, enabling rapid response to deviations with a larger step size and shortening the convergence period. If a classifier is used for hierarchical calibration decisions, the classifier can output the corresponding intensity level label along with the deviation source hierarchy label; this intensity level label can be obtained by comparing the error amplitude recorded in historical calibration records with the optimal initial... The mapping relationship of values ​​is obtained through training, or it can be calculated and determined in real time based on the current error amplitude. This intensity level serves as the starting basis for the progressive parameter adjustment in step S5, while the utility evaluation result in step S6 is used to dynamically correct the subsequent execution of this intensity, such as downgrading or reducing the intensity when necessary. The baseline value is used. This method effectively solves the problem of difficulty in obtaining the labels that are the source of bias in actual operation, while ensuring the adaptive ability of the classifier in practical applications.

[0069] Step S5: Perform tiered calibration and update the digital twin model. Based on the calibration decision in step S4, parameter adjustments are made only for the model tiers requiring calibration. The calibration intensity level determined in step S4 corresponds to... The initial value is the starting coefficient for the incremental parameter adjustment in this step. An incremental parameter adjustment strategy is used during calibration to avoid model instability caused by abrupt parameter changes. Specifically, the actual adjustment amount of the target parameter at the k-th calibration is... ,in The adjustment amount for the theoretical target is the correction amount for the ideal parameters calculated based on the difference between the measured data and the model output. For asymptotic coefficients, ,and The asymptotic coefficient monotonically increases from a preset small initial value to 1 with each calibration iteration. The update rules are as follows: ,in For a preset asymptotic step number and K>1, preferably This linear growth strategy enables... From initial value The value is initially increased evenly, such as by 0.1, which ensures both a conservative initial adjustment to avoid model oscillation and a full response to the current deviation in the later stages.

[0070] It should be noted that when a sudden and sharp increase in prediction error is detected, such as, but not limited to, an increase exceeding the preset mutation threshold compared to the previous calibration cycle, it indicates a significant change in the building's operational status, such as extreme weather events, equipment failures, or sudden changes in usage patterns. In this case, the surge in error may not be due to a systematic bias in the model parameters themselves, but rather to an occasional external disturbance. If the model parameters are adjusted hastily in this situation, the transient disturbance may be mistakenly interpreted as model inaccuracy and incorrectly corrected, thus compromising the model's accuracy under normal operating conditions. Therefore, the current calibration operation should be suspended, keeping the model parameters unchanged. Simultaneously, step S7 should be used to widen the prediction confidence interval accordingly, ensuring the prediction results accurately reflect the current high level of uncertainty and avoiding providing building managers with artificially inflated prediction accuracy. Only after several consecutive calibration cycles, such as 3-5 cycles (the specific number can be preset by those skilled in the art based on building operational characteristics and response sensitivity), and after the error remains persistently high, can it be determined as a systematic bias rather than a transient disturbance. At this point, the stratified calibration decision in step S4 should be retried. After confirming the level of the bias source, calibration should then be performed according to the normal progressive parameter adjustment process. This processing mechanism effectively avoids model parameter errors caused by occasional external interference, ensuring the robustness of the calibration operation. The specific value of the preset mutation threshold can be flexibly set by those skilled in the art based on the actual building scale, climate zone, and prediction accuracy requirements; this invention does not impose such limitations. After each calibration operation, information such as the calibration level, parameter adjustment amount, and error changes before and after calibration are recorded and stored in the calibration knowledge base for continuous optimization of the subsequent calibration decision classifier.

[0071] Step S6: Calculate the calibration utility function and evaluate the calibration effect. After calibration, re-execute multi-timescale predictions and calculate the improvement in ultra-short-term, short-term, and medium-term prediction accuracy after calibration. , , The expression for the calibration utility function is:

[0072]

[0073] in , , For the corresponding weight coefficients, and The weighting coefficient is determined based on the actual management needs of the building: if the building has high requirements for real-time control response, the weighting coefficient should be increased. If the requirements for daily energy consumption management are high, then increase If the requirements for weekly energy consumption planning are high, then increase... In a preferred embodiment of the present invention, , , .

[0074] when The calibration was deemed invalid at that time. To preset the utility threshold, preferably At this point, the calibration operation can be rolled back, restoring the model parameters to their pre-calibration state, or the calibration intensity can be reduced and the calibration can be re-executed. Reducing the calibration intensity means increasing the K value determined in step S4, which is equivalent to reducing the initial α1 and extending the increment period.

[0075] Furthermore, this invention also monitors the independent changes in prediction accuracy at each time scale. If the improvement in prediction accuracy at a certain time scale is negative, meaning the calibration operation has impaired the prediction accuracy at that time scale, then the calibration operation is marked as "invalid for that time scale," and similar calibrations will be preferentially avoided in subsequent calibration decisions. In particular, considering that the error sources of medium-term forecasts are largely influenced by the uncertainties of numerical weather predictions rather than the building physics parameters themselves, the assessment of the improvement in medium-term forecast accuracy is crucial. When making a comprehensive judgment, the uncertainty range of the weather forecast should be considered to avoid misjudging the calibration effect due to the deviation of the weather forecast itself.

[0076] Step S7: Perform weighted dynamic prediction based on the confidence profile. Divide the digital twin model into several sub-modules according to functional modules, including but not limited to: thermal envelope sub-module, HVAC system sub-module, lighting and equipment sub-module, and personnel behavior sub-module. Each sub-module corresponds to one or more spatial regions.

[0077] Each submodule is independently evaluated for its current confidence level. To avoid a sluggish response to mutations due to long-term accumulation of confidence scores, this invention employs a sliding window-based confidence update method. Each submodule at time Confidence score The calculation formula is:

[0078]

[0079] in The normalized average prediction error for the region corresponding to this submodule over the most recent L time windows, with a value ranging from 0 to 1; L is the sliding window length, representing the number of most recent time windows participating in the confidence assessment, preferably... It corresponds to 10 calibration cycles. It is a very small positive number, preferably The purpose of setting this lower limit is to avoid prediction errors exceeding 100%. This results in a negative confidence score, preventing numerical computational anomalies in subsequent weighted fusion and interval scaling. This lower limit is extremely small, designed to ensure computational safety rather than artificially inflating model confidence. Under most normal operating conditions, the confidence score is determined by its actual calculated value. The decision is made to only reach this lower limit when the prediction error is abnormally high. At this point, it indicates that the prediction results of this submodule are completely unreliable and should be assigned an extremely low fusion weight.

[0080] The confidence scores of each submodule constitute a digital twin confidence map. When performing the next round of dynamic energy consumption prediction, each submodule independently outputs its prediction results, which are then weighted and fused according to their confidence scores.

[0081]

[0082] in For the first The prediction results of each submodule Rate its confidence level. The confidence score is the sum of the confidence scores for all submodules. The average confidence score for each submodule is then used as the overall confidence score. This is used to characterize the overall credibility of the digital twin model. The final output is the prediction result. The confidence interval is included and is calculated as follows: First, calculate the standard deviation of the prediction results for each submodule. As the base interval width; then using the overall confidence level The interval width is adaptively scaled, and the final confidence interval is: Among them, the scaling function The function as a whole acts as a multiplier on the width of the base interval. β is the preset maximum expansion coefficient, preferably This allows the confidence interval width to smoothly expand to 1+β times the base width as the overall confidence level drops from 1 to 0, preferably not exceeding 3 times; The preset confidence coefficient is preferably... This corresponds to a 95% confidence level; this linear scaling mechanism, while maintaining the core intent of "widening the interval at low confidence levels," avoids inverse scaling. Extreme interval expansion at moderately low confidence levels ensures that the prediction interval has practical decision-making reference value across the entire confidence range. This scaling mechanism ensures that: when the confidence levels of all submodules are high, the output confidence interval is narrow, reflecting the reliability of the prediction results; when the confidence levels of all submodules are low, such as in the event of sudden operating conditions or severe model mismatch, the confidence interval automatically widens, truly reflecting the high degree of uncertainty in the prediction, and providing effective decision-making reference for building managers.

[0083] In a preferred embodiment of the present invention, the system is in a cold start phase when it is first deployed, and there is no historical operating data. Under the premise of maintaining the complete operation of the closed-loop iterative framework of steps S1 to S8, the hierarchical calibration decision module temporarily fixes the output of the "physical layer global calibration" strategy due to the lack of sufficient spatiotemporal error data to support the classifier's judgment. The rated parameters of the building design are used as the initial values, and the measured data of the first week are used as the benchmark for preliminary calibration. The confidence interval is relatively wide at this stage due to insufficient data, which accurately reflects the uncertainty of the initial prediction. As the operating data continues to accumulate, when the error spatiotemporal distribution feature vector reaches the preset data accumulation requirement, it usually takes 2 to 3 weeks for the hierarchical calibration decision module to naturally transition to the normal operation mode of intelligent hierarchical calibration based on the error spatiotemporal distribution features. This transition is naturally achieved by the continuous accumulation of error data within the sliding window, without involving parameter mutations or prediction jumps caused by strategy switching.

[0084] Step S8, Iterative Loop. The weighted prediction result from step S7 is compared with the actual energy consumption data to calculate the prediction error for the new round. The process returns to step S3, forming a continuous closed-loop iteration that sequentially performs prediction, extracts the spatiotemporal features of the error, performs hierarchical calibration decisions, updates the model, and re-predicts. Through continuous iterative loops, the digital twin model achieves self-optimization during the building's continuous operation, gradually approximating the actual thermodynamic behavior and energy consumption patterns of the physical building.

[0085] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the overall architecture of the building energy consumption dynamic prediction system based on digital twin calibration of the present invention. The system includes a data acquisition and preprocessing module, a digital twin model management module, a multi-timescale prediction engine, an error spatiotemporal feature extraction module, a hierarchical calibration decision module, a progressive parameter calibration module, a calibration utility evaluation module, a confidence map management module, a weighted prediction fusion module, and a closed-loop iterative control module.

[0086] The data acquisition and preprocessing module connects to various sensors and smart meters deployed in the building to collect real-time indoor and outdoor environmental data, equipment operating status data, and actual energy consumption data, and performs data cleaning and time alignment processing.

[0087] The digital twin model management module is used to store and manage digital twin models with a three-layer architecture, including a geometric layer, a physical layer, and a behavioral layer. It provides functions for reading, modifying, and versioning parameters at each layer.

[0088] The multi-timescale prediction engine connects to the data acquisition and preprocessing module and the digital twin model management module, respectively, and is used to perform energy consumption predictions at three time scales: ultra-short-term, short-term, and medium-term, based on the current digital twin model and real-time / predictive input data.

[0089] The error spatiotemporal feature extraction module connects to the multi-timescale prediction engine. It is used to receive prediction results and compare them with actual energy consumption data, calculate the prediction error at each time scale and in each spatial region, and construct the error spatiotemporal distribution feature vector.

[0090] The hierarchical calibration decision module, connected to the error spatiotemporal feature extraction module, is used to perform hierarchical calibration decisions based on the error spatiotemporal distribution feature vector to determine the model level and calibration intensity that need to be calibrated.

[0091] The progressive parameter calibration module is connected to the hierarchical calibration decision module and the digital twin model management module, respectively. It is used to perform progressive parameter correction on a specified model level according to the calibration decision and write the corrected parameters back to the digital twin model management module.

[0092] The calibration utility evaluation module connects to the multi-timescale prediction engine and is used to reacquire multi-timescale prediction results after calibration, calculate the calibration utility function, and evaluate the overall effect of the calibration operation.

[0093] The confidence graph management module, connected to the error spatiotemporal feature extraction module, is used to maintain the confidence scores of each sub-module of the digital twin model and dynamically update them based on the prediction error.

[0094] The weighted prediction fusion module is connected to the multi-timescale prediction engine and the confidence map management module, respectively. It is used to weight and fuse the prediction results according to the confidence of each sub-module, and output the final energy consumption prediction result and confidence range.

[0095] The closed-loop iterative control module, as the central control unit of the system, connects to all the above modules and coordinates the runtime sequence and data flow of each module to form a continuous closed-loop iteration that sequentially performs prediction, error spatiotemporal feature extraction, hierarchical calibration decision, model update and re-prediction.

[0096] To verify the effectiveness of this invention and the contribution of its innovative technical features, this specific embodiment further provides experimental verification and effect comparison after step S8, using backtracking simulation verification based on actual operating data from a typical office building. This test building is located in a hot-summer, cold-winter region, with a total building area of ​​approximately 12,000 square meters, 8 floors above ground, and a standard floor area of ​​1,500 square meters. The building functions include office areas, conference rooms, a data center server room, a catering area, and equipment rooms. The building envelope meets national energy-saving design standards. The air conditioning system uses a variable air volume system with chiller units, and the lighting system is LED intelligent lighting. The equipment power density is set according to typical values ​​for office buildings. The experimental data comes from the building's actual operating data for 15 minutes over two years, from January 2023 to December 2024. This includes total building electricity consumption, sub-item electricity consumption (HVAC, lighting sockets, power equipment), indoor and outdoor temperature and humidity, and hourly meteorological parameters from a weather station, specifically including temperature, humidity, solar radiation, and wind speed. The data from the first 12 months is used for initial model calibration and training, while the data from the latter 12 months is used for backtracking verification of the prediction effect. This dataset fully covers the summer cooling season, the winter heating season, and the spring and autumn transition season, including various operating conditions such as weekdays and holidays, and has sufficient representativeness and diversity.

[0097] The prediction accuracy evaluation uses the mean absolute percentage error (MAPE) as the core indicator. Its calculation formula is consistent with the formula given in step S3, that is, calculating the actual value for each time scale and each spatial region. Compared with the predicted value The absolute value of the difference is divided by the actual value, multiplied by 100%, and then averaged over all sampling points within the prediction time window. The quality of the confidence interval is comprehensively evaluated using two indicators: prediction interval coverage and prediction interval average width. Prediction interval coverage measures the proportion of the predicted confidence interval that actually contains the true value. Specifically, it is calculated by checking each point to see if the actual energy consumption value falls within the prediction confidence interval, and then counting the percentage of points falling within the predicted confidence interval. The closer this value is to the nominal confidence level (95%), the more reliable the interval estimate. Prediction interval average width measures the tightness of the confidence interval. Specifically, it is calculated by averaging the confidence interval widths over all prediction times and expressing it as a percentage of the actual energy consumption value for the same period. Under the same coverage, a smaller average interval width indicates a more accurate interval estimate and higher decision-making reference value. The number of model convergence iterations is defined as the total number of calibration iterations required from the initial calibration until the absolute value of the error change rate is less than 2% for three consecutive calibration cycles. Fewer iterations indicate faster model convergence and a more stable calibration process. The precision balance across time scales is represented by the utility function value U in step S6. The expression for the utility function is completely consistent with step S6 in the original document, i.e. The improvement in prediction accuracy across the ultra-short-term, short-term, and medium-term timescales is weighted and summed. A higher U value indicates a more significant overall positive effect of the calibration operation across all timescales and less obvious accuracy imbalance.

[0098] In Example 1, a rolling prediction-calibration closed loop was implemented for 12 months using the complete scheme of steps S1 to S8 of the present invention. The ultra-short-term prediction time was 1 hour into the future, the short-term prediction time was 24 hours into the future, and the medium-term prediction time was 7 days into the future. The spatial area was divided into 12 zones based on a mix of floors and functional areas. The confidence assessment sliding window length L was set to 10 calibration cycles, and a preset utility threshold was used. Set to 0.05, the number of steps for progressive parameter adjustment. Set to 10, multi-timescale utility weight , , The maximum expansion coefficients for the confidence intervals are set to 0.3, 0.4, and 0.3 respectively. The confidence coefficient is set to 2. Setting it to 1.96 corresponds to a 95% confidence level, which is the lower limit of the submodule confidence level. The value is set to 0.01. During system startup, in the cold start phase, due to the lack of historical error data to support the hierarchical calibration decision classifier, the output is temporarily fixed at a "physical layer global calibration" strategy. Initial calibration is performed using the design parameters as initial values ​​and the measured data from the first week as a benchmark. During this phase, the confidence interval is relatively wide, accurately reflecting the uncertainty of the initial predictions. After approximately three weeks of operation, the spatiotemporal distribution feature vector of the error has accumulated sufficiently, and the system automatically transitions to the classifier-based intelligent hierarchical calibration mode.

[0099] To systematically verify the technical effects of each innovation of the present invention, four comparative examples were set up. Each comparative example changed only one technical feature, and the remaining steps were exactly the same as in Example 1, so as to eliminate the interference of variables.

[0100] Comparative Example 1 employs a unified overall calibration, eliminating the hierarchical calibration decision in step S4 and the hierarchical selective calibration in step S5. Instead, when the prediction error in any region or at any time scale exceeds a preset threshold (MAPE greater than 10%), all parameters in the geometric, physical, and behavioral layers are uniformly corrected. The correction magnitude is calculated based on the global average error. The remaining steps, including closed-loop iteration, utility evaluation, and confidence-weighted prediction, are the same as in Example 1. The purpose of this comparative example is to address the "homogenization of calibration objects" defect pointed out in the background art and to verify the advantages of the hierarchical calibration mechanism of this invention in avoiding parameter coupling interference.

[0101] Comparative Example 2 employs open-loop calibration, meaning the closed-loop iterative mechanism of step S8 is eliminated. After completing the weighted prediction and outputting the result in step S7, it does not return to step S3. Each prediction is based on the same set of initial model parameters, and only one initial calibration is performed. All subsequent predictions use this calibrated fixed model, and the remaining steps are the same as in Example 1. The purpose of this comparative example is to address the deficiency pointed out in the background art that "calibration and prediction are a unidirectional serial process," and to verify the advantages of the closed-loop iterative mechanism of this invention in continuous self-optimization.

[0102] Comparative Example 3 eliminates the calibration utility evaluation and rollback mechanism. That is, it retains the calculation process of step S6 but does not perform utility judgment and rollback operations. Even if the calibration utility function value is lower than a preset threshold, the calibration is still performed. Any calibration that reduces the prediction error is forcibly executed. The remaining steps are the same as in Example 1. The purpose of this comparative example is to address the deficiency of "lack of multi-timescale quantitative evaluation of calibration utility" pointed out in the background art, and to verify the role of the utility evaluation mechanism of this invention in ensuring the balance of prediction accuracy across multiple time scales.

[0103] Comparative Example 4 uses the traditional inverse scaling method to place the confidence interval, which is the scaling function in step S7. Replace with the reciprocal scaling function commonly used in traditional methods. The remaining steps are the same as in Example 1. The purpose of this comparative example is to verify the advantage of the linear scaling mechanism of the present invention in avoiding excessive interval expansion and maintaining decision reference value when the confidence level is moderately low.

[0104] The combined results of the rolling forecast validation for the following 12 months are shown in the table below:

[0105] Ultra-short-term MAPE (1h) 4.1% 6.8% 7.5% 4.3% 4.1% Short-term MAPE (24h) 5.8% 8.2% 11.3% 7.1% 5.8% Mid-term MAPE (7 days) 8.2% 11.5% 18.6% 14.8% 8.2% Predicted Interval Coverage 94.2% 88.6% 76.3% 92.8% 96.1% Average width of prediction interval 22.5% 18.3% 15.6% 24.1% 41.7% Model convergence iterations 4 times 12 oscillations Non-convergence 8 times 4 times Precision uniformity across time scales (represented by utility function value U) 0.072 0.031 0.019 -0.015 0.072

[0106] As can be seen from the experimental results in the table above, Comparative Example 1, using a unified overall calibration, exhibits significantly higher prediction errors across all three time scales compared to this invention. The ultra-short-term MAPE increases by 2.7 percentage points, the short-term by 2.4 percentage points, and the medium-term by 3.3 percentage points. More importantly, Comparative Example 1 requires 12 calibration iterations to converge, and exhibits significant error oscillations during the iteration process. That is, while correcting parameters at a certain level reduces errors in some regions, parameter coupling leads to a reverse increase in errors in other regions or at other time scales. This phenomenon directly verifies the "one-size-fits-all" technical defect pointed out in the background section. In contrast, this invention, through hierarchical calibration driven by the spatiotemporal distribution characteristics of errors, precisely corrects only the level of error source, achieving stable convergence in just 4 iterations. This verifies the significant advantage of the hierarchical calibration mechanism in avoiding parameter coupling interference.

[0107] The experimental results of Comparative Example 2 further validate the crucial role of the closed-loop iteration mechanism. After canceling the closed-loop iteration, the mid-term prediction error of the open-loop calibration reached 18.6%, an increase of 10.4 percentage points compared to the present invention; the short-term error reached 11.3%, an increase of 5.5 percentage points compared to the present invention; and the ultra-short-term error reached 7.5%, an increase of 3.4 percentage points compared to the present invention. The errors at all three time scales deteriorated sharply with the extension of the prediction time window, indicating that without continuous prediction-calibration closed-loop feedback, the digital twin model cannot track the slow drift of the building structure caused by factors such as equipment aging, envelope performance degradation, and changes in usage patterns, resulting in a continuous degradation of model accuracy over time. Furthermore, the prediction interval coverage of Comparative Example 2 was only 76.3%, far below the nominal confidence level of 95%, indicating that its confidence interval was severely inaccurate and could not provide reliable decision-making references for building managers. This invention enables the digital twin model to continuously approximate the real state of the physical building through continuous closed-loop iteration. The errors on all three time scales are significantly lower than those of the open-loop scheme, and the prediction interval coverage reaches 94.2%, which is close to the nominal confidence level, thus verifying the core value of the closed-loop iteration mechanism.

[0108] The experimental results of Comparative Example 3 reveal the necessity of the calibration utility evaluation mechanism. After removing the utility backoff mechanism, although the ultra-short-term prediction accuracy was slightly better than the present invention (MAPE of 4.3%), the medium-term prediction accuracy deteriorated significantly, with a MAPE as high as 14.8%, an increase of 6.6 percentage points compared to the present invention. From the utility function values, the U value of Example 1 was 0.072, indicating that the calibration operation produced a positive overall effect across all three time scales; while the U value of Comparative Example 3 was -0.015, a negative value, indicating that after removing the utility evaluation and backoff mechanism, excessively pursuing ultra-short-term accuracy actually damaged the medium-term prediction accuracy, resulting in a negative overall effect of the calibration operation. The fundamental reason for this phenomenon is that the sensitivity of different levels of parameters in the building energy consumption prediction model to the prediction time scale varies significantly. While the correction of certain physical parameters can make the ultra-short-term prediction more closely match the measured values, it may severely distort the cumulative trend of the medium-term prediction by changing the model's thermal inertia response characteristics. Without utility constraints, the system tends to frequently perform overcalibrations that are "beneficial in the ultra-short term but detrimental in the medium term," leading to a severe imbalance in accuracy across multiple time scales. This invention constructs a calibration utility function to quantify the combined impact of calibration operations on the three time scales in the form of a weighted sum, and backs down or reduces the intensity when the utility value is below a preset threshold, effectively ensuring the global balance of prediction accuracy across all time scales.

[0109] The experimental results of Comparative Example 4 verify the engineering practical advantages of the linear scaling confidence interval of this invention. Regarding the accuracy of prediction points, since the scaling strategy does not affect the predicted point values, the MAPE index of Example 1 and Comparative Example 4 is completely consistent, which is also in line with theoretical expectations. However, in terms of interval estimation, although Comparative Example 4, which uses traditional reciprocal scaling, has a slightly higher prediction interval coverage of 96.1%, its average prediction interval width is as high as 41.7%, approximately 1.85 times that of Example 1. The fundamental reason for this huge difference is that when the overall confidence level is at a moderately low level, reciprocal scaling will produce an extreme expansion of more than 2 times, resulting in an excessively wide confidence interval that loses its practical decision-making reference value. In contrast, the linear scaling of this invention maintains a controllable expansion range across the entire confidence range. While maintaining a high coverage of 94.2%, the interval width is only 54% of that of the reciprocal scaling scheme, providing building managers with a reliable and compact energy consumption prediction interval, greatly improving the operability of the prediction results in actual energy-saving control.

[0110] The above four sets of comparative experiments systematically verified the effectiveness and necessity of the key technical features of this invention from different dimensions. The hierarchical calibration mechanism of this invention solves the problem of homogenization of calibration objects, the closed-loop iterative mechanism enables continuous self-optimization of the model, the calibration utility evaluation mechanism ensures the balance of prediction accuracy across multiple time scales, and the linearly shortened confidence interval balances reliability and practicality. The synergistic effect of these four innovative features enables this invention to achieve significantly superior overall performance compared to existing technologies in terms of accuracy, stability, reliability, and practicality in dynamic prediction of building energy consumption.

[0111] It should be noted that the above experimental data were obtained based on specific test buildings and climatic conditions. The actual prediction accuracy may vary for different building types and different climate zones. However, the relative advantages and disadvantages of the hierarchical calibration mechanism, closed-loop iterative framework, utility evaluation method, and linear shrinkage confidence interval strategy proposed in this invention are universally applicable. Those skilled in the art can adaptively adjust the parameter settings in the above embodiments according to the specific characteristics of actual application scenarios, and such adjustments do not depart from the protection scope of this invention. The actual operating data used in this experiment comes from real office buildings. The 12-month verification period fully covers four typical operating conditions: summer cooling season, winter heating season, and spring and autumn transition season. It also includes different behavioral patterns during long holidays such as Spring Festival and National Day, as well as ordinary working days, making the data comprehensive. All comparative experiments use the same hardware platform, software environment, and input dataset. Only the technical feature variables to be verified are changed in each comparative example, while the other parameters are completely consistent, ensuring the comparability and repeatability of the experimental results.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for dynamic prediction of building energy consumption based on digital twin calibration, characterized in that, Includes the following steps: Step S1: Construct a multi-level digital twin model of the building, which includes a geometric layer, a physical layer, and a behavioral layer; Step S2: Collect real-time building operation data. The sampling frequency of the real-time operation data is based on the sensor type and a preset fixed sampling period. Meteorological forecast data is updated according to a fixed period. Data from different sources are time-aligned and then input into the prediction engine. Based on the current digital twin model, perform benchmark energy consumption predictions for three time scales: ultra-short-term, short-term, and medium-term. Step S3: Calculate the prediction error for each time scale and each spatial region, construct the spatiotemporal distribution feature vector of the error, and determine the source level of the prediction deviation based on the distribution pattern of the spatiotemporal distribution feature vector of the error; Step S4: Based on the deviation source level determined in Step S3, perform a stratified calibration decision to determine the model level and calibration intensity that need to be calibrated; Step S5: Based on the calibration decision in Step S4, only the model levels that need calibration are adjusted, and the target parameters are gradually adjusted using a progressive parameter adjustment strategy. When a sudden and sharp increase in prediction error is detected, the calibration operation is suspended, the model parameters are kept unchanged, and the prediction confidence interval is widened accordingly in Step S7 to reflect the increased uncertainty. After the error remains high for several consecutive calibration cycles, the hierarchical calibration decision in Step S4 is triggered again. After confirming the source of the deviation, the calibration is performed according to the normal progressive parameter adjustment process. Step S6: Calculate the calibration utility function and evaluate the changes in ultra-short-term, short-term, and medium-term prediction accuracy after calibration. If the overall calibration utility function value is lower than the preset threshold, roll back the calibration operation or reduce the calibration intensity. Step S7: Update the confidence scores of each sub-module of the digital twin model, construct the digital twin confidence map, weight and fuse the prediction results according to the confidence of each sub-module, and output the final energy consumption prediction results and confidence intervals. Step S8: Compare the prediction results of step S7 with the actual energy consumption data, and return to step S3 to form a continuous closed-loop iteration that sequentially performs prediction, error spatiotemporal feature extraction, hierarchical calibration decision, model update and re-prediction.

2. The method for dynamic prediction of building energy consumption based on digital twin calibration according to claim 1, characterized in that: In step S3, the method for constructing the spatiotemporal distribution feature vector of the error is as follows: the building space is divided into N regions, the prediction time is divided into three scales: ultra-short-term, short-term, and medium-term, and the error feature value is calculated for each (scale i, region j) combination. This constitutes an N×3 dimensional error spatiotemporal distribution feature vector. Among them, the error eigenvalues This represents the normalized mean absolute percentage error.

3. The method for dynamic prediction of building energy consumption based on digital twin calibration according to claim 1, characterized in that: In step S4, the specific method for the hierarchical calibration decision is as follows: if the error is mainly concentrated in a specific spatial region, it is determined to be a deviation of geometric layer or local physical parameters and spatial local calibration is triggered; if the error systematically increases with the extension of prediction time, it is determined to be a deviation of physical parameters and physical layer global calibration is triggered. If the error exhibits a periodic fluctuation pattern, it is determined to be a change in the behavioral layer parameters and behavioral layer calibration is triggered; if the error distribution has no obvious pattern and there are deviations in each level, multi-level joint calibration is triggered.

4. The method for dynamic prediction of building energy consumption based on digital twin calibration according to claim 3, characterized in that: In step S4, the hierarchical calibration decision uses a classifier based on the spatiotemporal distribution feature vector of error to determine the source of deviation. The classifier is trained using historical calibration records or offline simulation data. Its input is the spatiotemporal distribution feature vector of error, and its output is the hierarchical label of the source of deviation.

5. The method for dynamic prediction of building energy consumption based on digital twin calibration according to claim 1, characterized in that: In step S5, the progressive parameter adjustment strategy is: the actual adjustment amount of the target parameter during the k-th calibration. ,in Adjustment amount for theoretical target For asymptotic coefficients, ,and The value increases monotonically from the preset initial value as the number of calibration iterations increases, eventually approaching 1.

6. The method for dynamic prediction of building energy consumption based on digital twin calibration according to claim 1, characterized in that: In step S6, the expression for the calibration utility function is: in, , , These represent the improvement in prediction accuracy for the ultra-short-term, short-term, and medium-term periods after calibration. , , For the corresponding weight coefficients, and ;when The calibration was deemed invalid at that time. This is a preset utility threshold.

7. The method for dynamic prediction of building energy consumption based on digital twin calibration according to claim 1, characterized in that: In step S7, the method for updating the digital twin confidence map is as follows: For the first... Each submodule has a confidence score calculated based on the moving average of prediction errors over several recent time windows: in This is the normalized average prediction error for the region corresponding to this submodule over the most recent preset number of time windows, with a value range of [0,1]. The confidence score is a very small positive number, used to avoid numerical calculation anomalies rather than reflecting the model's reliability; the normalized confidence scores of each submodule are used as the weighted fusion weights for the next round of predictions, while the average confidence score of each submodule is used as the overall confidence score. ; When outputting the final energy consumption prediction results, the standard deviation of the prediction results of each submodule is used as the base interval radius, and the overall confidence level is calculated. The width of the final confidence interval after adaptive scaling: when the overall confidence level... When the overall confidence level approaches 1, the interval width is close to the base width; when the overall confidence level... When the value approaches 0, the range is appropriately widened to truly reflect the high degree of uncertainty in the overall forecast.

8. A building energy consumption dynamic prediction system based on digital twin calibration, applied to the building energy consumption dynamic prediction method based on digital twin calibration as described in any one of claims 1 to 7, characterized in that, include: The data acquisition and preprocessing module is used to collect real-time data on the building's indoor and outdoor environment, equipment operating status, and actual energy consumption, and to perform data cleaning and time alignment. The digital twin model management module is used to store and manage digital twin models with a three-layer architecture, including a geometric layer, a physical layer, and a behavioral layer, and supports reading and modifying layer parameters. A multi-timescale forecasting engine is used to perform energy consumption forecasts at three timescales: ultra-short-term, short-term, and medium-term, based on the current digital twin model. The error spatiotemporal feature extraction module is used to calculate the prediction error at each time scale and in each spatial region, construct the error spatiotemporal distribution feature vector, and determine the source level of the prediction deviation. The hierarchical calibration decision module is used to perform hierarchical calibration decisions based on the error spatiotemporal distribution feature vector to determine the model level and calibration intensity that need to be calibrated. The progressive parameter calibration module is used to perform progressive parameter corrections at specified model levels. The calibration utility evaluation module is used to calculate the calibration utility function, evaluate the comprehensive impact of calibration operations on the prediction accuracy at various time scales, and decide whether to perform or roll back the calibration. The confidence graph management module is used to maintain and dynamically update the confidence scores of each sub-module of the digital twin model. The weighted prediction fusion module is used to weight and fuse the prediction results according to the confidence level of each sub-module, and output the final energy consumption prediction result and confidence level range. The closed-loop iterative control module is used to coordinate the runtime sequence of the above modules, forming a continuous closed-loop iteration that sequentially executes prediction, error spatiotemporal feature extraction, hierarchical calibration decision, model update, and re-prediction.

9. The building energy consumption dynamic prediction system based on digital twin calibration according to claim 8, characterized in that: In the digital twin model management module, the geometry layer contains geometric dimension information and spatial topology relationships extracted from the building BIM model, the physical layer contains thermal parameters of the building envelope and equipment performance parameters, and the behavior layer contains personnel work and rest patterns and equipment usage pattern parameters. Parameters are transferred between each layer through standardized interfaces, and modifications to parameters at any layer do not affect the model structure of other layers.

Images