Double-carbon situation awareness method suitable for zero-carbon building digital twinning

Abstract

The application belongs to the technical field of double-carbon situation awareness, and particularly relates to a double-carbon situation awareness method suitable for digital twinning of zero-carbon buildings, which comprises respectively constructing a measured state feature matrix and a virtual state feature matrix; determining a current working condition according to a building external environment and an internal equipment state, combining a parameter confidence matrix to generate a non-uniform quantization deviation boundary model dynamically modulated by the parameter confidence, calculating a deviation matrix and checking an overrun; when the overrun occurs, reversely estimating key parameters under an extended Kalman filtering framework, updating a parameter expectation value, a parameter error covariance matrix and a parameter confidence matrix, and generating a double-carbon situation matrix and a double-domain compensation vector of an adjusted control parameter based on a corrected digital twinning model. The application can correct virtual-real state deviation, improve the reliability of the awareness result and the low-carbon operation regulation and control capability.

Description

Technical Field

[0001] This invention relates to the field of dual-carbon situational awareness technology. More specifically, this invention relates to a dual-carbon situational awareness method applicable to digital twins of zero-carbon buildings. Background Technology

[0002] Real-time monitoring of carbon flux, prediction of carbon emission trajectories, and assessment of system carbon neutrality resilience during the actual operation of zero-carbon buildings are crucial foundations for achieving low-carbon management throughout the building's lifecycle. Digital twin technology can establish a mapping model of the physical building in virtual space, dynamically expressing the operating conditions of the building's energy system by fusing physical sensor data with virtual simulation results, and providing necessary decision-making basis for equipment scheduling and energy-carbon synergistic optimization. However, the complex types of equipment inside zero-carbon buildings and the frequent changes in external meteorological environment and energy supply and demand conditions place higher demands on the long-term accuracy and adaptability of digital twin models.

[0003] To address the complex dynamic environment and maintain long-term dual-carbon situational awareness, existing technologies have developed monitoring schemes based on digital twin models established using initial parameters or ideal conditions. These schemes primarily utilize fixed, preset parameters for forward simulation, attempting to continuously track energy flow patterns during building operation under idealized assumptions. The aim is to output predictive data over a longer time span, thereby helping to improve building energy efficiency and zero-carbon operation capabilities.

[0004] The existing methods relying on ideal states lack a consistency verification mechanism between virtual and real states and error boundary constraints that adapt to changes in operating conditions during long-term operation. When zero-carbon buildings experience equipment aging, system performance degradation, or changes in operating conditions, the physical building state and the virtual model output will evolve in different ways. Existing mechanisms struggle to identify model mismatches in virtual-real interactions in a timely manner, and the system lacks effective reverse parameter correction and system compensation methods to dynamically trace and update the underlying digital mapping. This type of model easily causes carbon emission predictions to gradually deviate from the actual trajectory, which not only reduces the reliability of dual-carbon situational awareness results but also directly weakens or even destroys the effectiveness of equipment control strategies. Summary of the Invention

[0005] To address the technical problem that existing dual-carbon situational awareness methods are prone to mismatch with the actual building status during long-term operation, resulting in inaccurate sensing results and insufficient equipment control effectiveness, this invention provides a dual-carbon situational awareness method suitable for zero-carbon building digital twins.

[0006] A dual-carbon situational awareness method applicable to digital twins of zero-carbon buildings includes: S1: acquiring sensor data streams of the zero-carbon building entity and virtual sensor data streams of the digital twin model under the same control commands, and constructing measured state feature matrices and virtual state feature matrices respectively; S2: determining the current operating condition based on external environmental parameters and internal equipment state parameters, and generating a non-uniform quantization deviation boundary model dynamically modulated by parameter confidence matrices; calculating the deviation matrix between the measured state feature matrix and the virtual state feature matrix, and verifying whether the deviation matrix is... S3: When the matrix elements of the deviation matrix exceed the limit, the deviation matrix is ​​expanded under the extended Kalman filter framework to obtain the measurement residual vector, and the key parameters of the digital twin model are back-estimated based on the measurement residual vector, and the expected values ​​of the parameters and the parameter error covariance matrix are updated to generate the updated parameter confidence matrix; Based on the corrected digital twin model, a dual carbon situation matrix containing real-time carbon flux values, carbon emission trajectory prediction curves and carbon neutrality resilience index is generated, and a dual-domain compensation vector for adjusting the equipment control parameters is calculated.

[0007] This invention obtains measured and virtual state feature matrices and dynamically generates a non-uniform quantization deviation boundary model based on parameter confidence to verify whether the state features exceed limits. When the deviation exceeds limits, the key parameters are inversely estimated using the measurement residual vector within the extended Kalman filter framework to update the digital twin model, generating a dual-carbon situation matrix and calculating and issuing a dual-domain compensation vector for equipment control. A dynamic deviation boundary verification mechanism modulated by parameter confidence is established, which can accurately identify virtual-real mismatch states when the building microenvironment or equipment performance degrades. The extended Kalman filter is used to achieve inverse calibration of mismatch parameters, ensuring the physical reliability of the digital twin model during long-term operation, thereby improving the effectiveness and anti-interference capability of the low-carbon operation control strategy of the microgrid system.

[0008] Preferably, the construction of the measured state feature matrix and the virtual state feature matrix includes: acquiring the current time measurement values ​​of multiple sensor channels and converting them into state column vectors; extracting the state measurement values ​​of the previous sampling period, subtracting the current time measurement value from the state measurement value of the previous sampling period, and dividing by a constant sampling period to calculate the first-order rate of change of each sensor data stream; for a preset combination of sensor data streams within a preset time window, extracting the sample distribution of the observation sequence of the data stream within the time window and calculating the cross-covariance value; arranging the acquired real-time measurement values, first-order rates of change, and cross-covariance values ​​according to a preset correspondence to generate the measured state feature matrix; and generating the virtual state feature matrix using the virtual sensor data stream according to the same data structure.

[0009] This invention uses constant period state measurement values ​​to calculate the first-order rate of change and extracts the sample distribution within the time window to calculate the cross-covariance value to complete the dimensions of the state feature matrix. It enriches the underlying state features from two dimensions: time gradient and cross-source correlation of devices. This enables the digital twin model to keenly capture the correlation patterns of small dynamic fluctuations during consistency verification and improves the system's ability to identify abnormal environmental noise disturbances.

[0010] Preferably, determining the current operating condition based on external environmental parameters and internal equipment status parameters includes: acquiring external environmental parameters and internal equipment status parameters and merging them into a multi-dimensional feature vector; inputting the multi-dimensional feature vector into a linear classifier model and calculating the signed boundary distance of the multi-dimensional feature vector relative to the classification boundary of each preset operating condition category; performing probability normalization processing on the signed boundary distances corresponding to each preset operating condition category to generate a conditional probability distribution, and determining the preset operating condition category corresponding to the maximum conditional probability as the current operating condition.

[0011] Preferably, the step of generating a non-uniform quantization deviation boundary model dynamically modulated by the parameter confidence matrix includes: querying a preset operating condition matching database to obtain an initial deviation boundary upper limit matrix and an initial deviation boundary lower limit matrix that correspond to the current operating condition and have the same dimensional structure; extracting the current confidence of each key parameter in the parameter confidence matrix and mapping it to a confidence scaling matrix that has the same dimension and structure as the measured state feature matrix; and multiplying the initial deviation boundary upper limit matrix and the initial deviation boundary lower limit matrix element-wise with the confidence scaling matrix to generate a non-uniform quantization deviation boundary model.

[0012] This invention utilizes the confidence scaling matrix obtained by mapping to multiply corresponding elements with the initial deviation boundary upper limit matrix and the initial deviation boundary lower limit matrix to construct a non-uniform quantization deviation boundary model. This realizes a dynamic and non-uniform threshold reduction mechanism, assigning trigger boundary thresholds to key parameters with high uncertainty and low confidence, thereby prioritizing the activation of correction logic for low-confidence parameters, intercepting and filtering out invalid residual calculations caused by conventional background noise, and accelerating the global convergence speed of parameter correction.

[0013] Preferably, the step of calculating the deviation matrix between the measured state feature matrix and the virtual state feature matrix, and verifying whether the deviation matrix exceeds the constraints of the non-uniform quantization deviation boundary model, includes: verifying the consistency of the dimension and structure between the measured state feature matrix and the virtual state feature matrix; if the structure is consistent, obtaining the initial deviation matrix by subtracting corresponding elements of the matrices; comparing each element in the initial deviation matrix with the corresponding upper and lower limit thresholds in the non-uniform quantization deviation boundary model item by item; when an element exceeds the corresponding upper and lower limit thresholds, marking the corresponding state feature as exceeding the limit and generating an over-limit identification matrix; setting the deviations corresponding to the non-over-limit elements to zero according to the over-limit identification matrix, and generating the deviation matrix while maintaining the dimension of the initial deviation matrix.

[0014] Preferably, the updating of the expected value of the parameters and the parameter error covariance matrix includes: applying a preset small perturbation to the key parameters to be estimated to run a digital twin model to obtain virtual state feature output; expanding the virtual state feature output in the same element order as the measurement residual vector; using the ratio of the expanded output change to the corresponding parameter perturbation as the corresponding element in the measurement Jacobian matrix; calculating the Kalman gain matrix using the measurement Jacobian matrix, the parameter error covariance matrix, and the measurement noise covariance matrix; multiplying the Kalman gain matrix with the measurement residual vector to obtain the key parameter correction amount; and superimposing the key parameter correction amount onto the expected value of the parameters at the previous time step to obtain the updated expected value of the parameters; and updating the parameter error covariance matrix based on the Kalman gain matrix and the measurement Jacobian matrix.

[0015] This invention obtains the measurement Jacobian matrix by applying small perturbations to key parameters and calculating the output changes. It then calculates the Kalman gain matrix by combining the measurement noise covariance matrix and performs cyclic updates on the parameter expectation value and the error covariance matrix. This provides a stable and effective numerical solution for the differential Jacobian for complex digital twin systems that lack direct analytical partial derivative functions. Even in nonlinear abrupt change scenarios, it can still ensure the computational smoothness of the reverse tracking process of key parameters and prevent gradient oscillation divergence in the reverse correction model.

[0016] Preferably, the real-time carbon flux value is obtained through the following steps: obtaining the real-time power of the building's total electrical load and the grid emission factor for the corresponding time period, and calculating the baseline cumulative carbon emissions without considering renewable energy offsets by integration; obtaining the carbon offset of the building photovoltaic system, and subtracting the carbon offset from the baseline cumulative carbon emissions to obtain the net carbon accumulation; and calculating the time derivative of the net carbon accumulation to obtain the real-time carbon flux value.

[0017] Preferably, the generation of a dual-carbon situation matrix based on the corrected digital twin model, which includes real-time carbon flux values, carbon emission trajectory prediction curves, and a carbon neutrality resilience index, includes: determining the maximum offset of carbon flux relative to the normal baseline trajectory when a building is subjected to extreme environmental disturbances; using the reciprocal of the product of the maximum offset and the disturbance recovery time as the carbon neutrality resilience index; and merging the real-time carbon flux values ​​and the carbon neutrality resilience index after expanding and aligning them according to the time series length of the carbon emission trajectory prediction curves generated based on long short-term memory networks to generate a dual-carbon situation matrix.

[0018] Preferably, the calculation of the dual-domain compensation vector used to adjust the equipment control parameters includes: taking the positive difference between the predicted carbon emissions and the set carbon emission benchmark limit as the carbon emission over-limit penalty; adding the square of the carbon emission over-limit penalty to the carbon emission penalty coefficient and the building operation energy purchase cost to obtain the joint objective function; using the parameters corrected by the digital twin model as the basic input, and using the Newton iteration method to solve for the minimum value of the joint objective function within the feasible domain of the control parameters.

[0019] Preferably, the Newton iteration method is used to solve for the minimum value of the joint objective function within the feasible domain of the control parameters, including: calculating the gradient vector and Hessian matrix of the joint objective function with respect to the control decision variables; using the gradient vector and Hessian matrix to calculate the Newton update direction to update the control decision variables, and performing boundary projection on the updated control decision variables to keep them within the physical safety limits of the equipment; when the absolute value of the difference between the objective function values ​​of two adjacent iterations is not greater than the preset convergence accuracy, the iteration is terminated and the current equipment control parameters are combined and output as a two-domain compensation vector.

[0020] This invention constructs a joint objective function by weighting and merging the penalty for exceeding carbon emission limits with the cost of purchasing energy. With the condition of minimizing the convergence of the objective function, it uses the Newton-Raphson iteration method to output a dual-domain compensation vector in the feasible domain. This integrates the environmental constraints on carbon emissions and energy consumption indicators of the system into a single optimization space for dimensionality reduction, enabling the control system to coordinate and adjust the control parameters within the physical safety boundary. This reduces the building's energy purchase cost while enhancing the carbon emission control resilience of zero-carbon buildings in the face of extreme disturbances.

[0021] The technical solution of the present invention has the following beneficial technical effects: This invention acquires sensor data streams from a zero-carbon building entity and virtual sensor data streams from a digital twin model under the same control commands, constructing measured state feature matrices and virtual state feature matrices respectively. Based on the building's external environment and internal equipment state parameters, the current operating condition is determined. A non-uniform quantization deviation boundary model dynamically modulated by parameter confidence is generated using a parameter confidence matrix. The deviation matrix between the measured and virtual state feature matrices is calculated, and it is verified whether the deviation matrix exceeds the constraints of the non-uniform quantization deviation boundary model. By constructing measured and virtual state feature matrices for the zero-carbon building entity and the digital twin model respectively, the problem of deviation exceeding limits in the measured and virtual state feature matrices is solved. The non-uniform quantization deviation boundary model, dynamically modulated by parameter confidence, is generated based on the building's external environment and internal equipment state parameters, with deviations corresponding to elements that do not exceed limits set to zero, maintaining the constraints of the non-uniform quantization deviation boundary model and solving the problem of matrix element exceeding limits in the deviation matrix.

[0022] When the number of elements in the bias matrix exceeds the limit, the bias matrix is ​​expanded according to a preset element order within the extended Kalman filter framework to obtain a measurement residual vector. Based on this measurement residual vector, the key parameters of the digital twin model are inversely estimated, updating the expected values ​​of the parameters and the parameter error covariance matrix. A parameter confidence matrix is ​​then generated based on the updated parameter error covariance matrix. This inverse estimation of the key parameters of the digital twin model using the measurement residual vector offsets the maximum offset of the out-of-limit bias matrix elements, thus resolving the problem of exceeding the limit in inverse estimation of key parameters.

[0023] Based on a digital twin model with completed parameter calibration, a dual-carbon situational awareness matrix is ​​generated, containing real-time carbon flux values, carbon emission trajectory prediction curves, and a carbon neutrality resilience index. A dual-domain compensation vector is then calculated to adjust equipment control parameters. The digital twin model generates the dual-domain compensation vector based on the minimum value of the joint objective function. Based on this vector, the equipment control parameters are adjusted. The real-time carbon flux values ​​and carbon neutrality resilience index are expanded and aligned according to the time series length of the carbon emission trajectory prediction curves generated by the prediction algorithm, and then the matrices are merged to generate the dual-carbon situational awareness matrix. This dual-domain compensation vector adjustment based on the minimum value of the joint objective function solves the problem of real-time carbon flux values ​​and carbon emission trajectory prediction curves exceeding limits, allowing for the merging of these two curves. The optimal compensation control quantity is output after merging the dual-carbon situational awareness matrix and the dual-domain compensation vector, maintaining the convergence accuracy of both the dual-carbon situational awareness method and the digital twin model. Attached Figure Description

[0024] Figure 1 This is a flowchart of a dual-carbon situational awareness method applicable to digital twins of zero-carbon buildings; Figure 2 This is a schematic diagram of the fluctuation curve of sensor data between indoor temperature and host power; Figure 3 This is a schematic diagram illustrating the probability distribution of multidimensional feature vectors under different preset working condition categories; Figure 4 This is a schematic diagram comparing the overall performance of the experimental group and the control group. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] This invention discloses a dual-carbon situational awareness method applicable to digital twins of zero-carbon buildings, referring to... Figure 1 This includes the following steps: S1. Constructing the state feature matrix. Specifically, acquire the sensor data stream of the zero-carbon building entity, and simultaneously acquire the virtual sensor data stream of the digital twin model under the same control commands. Based on the sensor data stream and the virtual sensor data stream, construct the measured state feature matrix and the virtual state feature matrix, respectively.

[0027] By subscribing to sensor topics on the zero-carbon building gateway via the MQTT IoT communication protocol, we continuously receive time-series data streams from measured sensors, including temperature, humidity, energy consumption, and carbon dioxide concentration. Timestamp alignment is performed using the NTP network time protocol.

[0028] The system calls the API interface provided by the EnergyPlus building energy consumption simulation engine to input the control commands for the current HVAC set temperature, lighting switch, and renewable energy access status into the pre-established building digital twin model. It then performs stepwise time integration calculations and outputs a virtual sensor data stream at the same time resolution.

[0029] For both sensor data streams and virtual sensor data streams, the first-order rate of change of each sensor physical quantity with respect to time is calculated by dividing the difference between the current sampled value and the previous sampled value by the sampling period. Within a fixed-time sliding window, the cross-covariance value is calculated according to a preset combination of sensor data streams. The real-time measurement value of each physical quantity, its first-order rate of change, and the cross-covariance value between each physical quantity and a preset associated physical quantity are concatenated row by row. An N-row, F-column measured state feature matrix and a virtual state feature matrix are instantiated in memory, where N represents the total number of sensor types and F represents the total number of state feature dimensions. When the current sampled value is the first measurement value after system startup, due to the lack of state measurement values ​​from the previous sampling period, the prior estimate of the corresponding rate of change in the state vector is directly used as the initial value during the extended Kalman filter initialization phase.

[0030] In an optional embodiment, a constant sampling period is set. The value range is from 0.5s to 5s. For example, a typical value of 1s is taken. The current time measurement values ​​of N independent sensor channels in the building, including indoor temperature, CO2 concentration and air conditioning unit power, are read simultaneously to construct a current state column vector of length N.

[0031] To calculate the first-order rate of change, assuming that the temperature of a certain area is measured to be 24.5℃ at timestamp 1s, and the temperature value at the previous sampling period of 0s is 24.3℃, the first-order rate of change can be calculated as 0.2℃ / s by difference operation and dividing by 1s.

[0032] For a preset combination of sensor data streams within a predefined time window, the distribution of the observed data series samples within the time window is extracted, and the cross-covariance value is calculated. For the calculation of the cross-covariance value, the preset time window length is determined to be 60 seconds based on the building's thermal inertia characteristics, and the number of sampling points L within the window is based on a constant sampling period. It is confirmed that L = 60 / The number of sampling points after rounding. When the constant sampling period is 1 second, then L is 60. For example, if indoor temperature is selected as the sequence x and chiller power as the sequence y, the mean value of the indoor temperature samples within the time window can be calculated respectively. At 24℃, the average power of the chiller unit (sample value) The power consumption is 5000W. This invention improves upon the cross-covariance method by using the product of the differences in sample means within a discrete time window to obtain a scalar value for the cross-covariance of two data series, representing the correlation between temperature fluctuations and device power consumption. The specific formula for calculating the cross-covariance value is as follows: ; in, The cross-variance value; This represents the number of sampling points in the window. This represents the i-th sampled measurement value of the first sensor within the time window; This represents the average value of the samples from the first sensor within the time window. This represents the i-th sampled measurement value from the second sensor within the time window. This represents the sample mean of the second sensor within the time window. The relationship between temperature fluctuations and device power consumption is numerically represented by the cross-covariance scalar value.

[0033] Reference Figure 2 Within a time window of 0s to 60s, the solid line of indoor temperature and the dashed line of main unit power show highly consistent fluctuation frequency and amplitude changes. When the chiller main unit power reaches its peak, the indoor temperature curve also shows a corresponding peak shape. The trend of the two curves intuitively reflects the physical fact that indoor temperature is driven by the power consumption of the equipment, proving that the cross-variance value can capture the underlying energy coupling relationship between different physical quantities.

[0034] For each sensor channel, the real-time measurement value, first-order rate of change, and cross-covariance value with preset associated key sensors are arranged column-wise and concatenated to generate an N×3 measured state feature matrix. The virtual sensor data stream output from the digital twin platform is then processed synchronously with the same dimensional structure and timestamps to generate a corresponding virtual state feature matrix. The preset sensor associations are determined based on the influence of building energy flow on equipment control and the correlation of historical operating data. For example, an indoor temperature sensor is associated with an air conditioning unit power sensor, a carbon dioxide concentration sensor with a fresh air unit airflow sensor, and a photovoltaic output power sensor with a solar radiation intensity sensor. In an alternative implementation, multiple preset associated sensors can be configured for each sensor channel to form an N×F dimensional state feature matrix.

[0035] S2. Determine the operating conditions and verify the deviation matrix. Specifically, determine the current operating conditions based on the external environment of the building and the state parameters of the internal equipment. Combined with the parameter confidence matrix, generate a non-uniform quantization deviation boundary model dynamically modulated by the parameter confidence. Calculate the deviation matrix between the measured state feature matrix and the virtual state feature matrix, and verify whether the deviation matrix exceeds the constraints of the non-uniform quantization deviation boundary model.

[0036] Extract outdoor temperature and humidity and solar radiation intensity obtained from the weather station, as well as the operating frequency of the HVAC main unit and the opening status parameters of the fresh air unit damper obtained from the building automation system.

[0037] The K-Means clustering algorithm is invoked to map the extracted state parameters to preset standard operating condition clusters such as seasonal, transitional, extreme cold, peak, and nighttime low conditions to determine the current operating condition.

[0038] The parameter confidence matrix in the form of a diagonal matrix is ​​retrieved from memory. The diagonal elements of the parameter confidence matrix correspond to the current confidence of key parameters such as building wall thermal resistance and equipment operating efficiency. The current confidence is obtained by performing a reciprocal transformation and normalization on the diagonal elements of the parameter error covariance matrix.

[0039] The system queries a pre-defined operating condition matching database. For the determined current operating condition, it extracts an upper and lower bound matrix of initial deviation boundaries that matches the dimension and structure of the measured state feature matrix. Based on the pre-defined correspondence between key parameters and state features, it maps the current confidence level in the parameter confidence matrix to a confidence scaling matrix that matches the dimension and structure of the measured state feature matrix. The upper and lower bound matrices of the initial deviation boundaries are then multiplied element-wise with the confidence scaling matrix to generate a non-uniform quantization deviation boundary model with dynamically modulated parameter confidence. This non-uniform quantization deviation boundary model is represented by upper and lower bound matrices that assign independent dynamic thresholds to each state feature.

[0040] The initial deviation matrix is ​​obtained by subtracting the virtual state feature matrix from the measured state feature matrix. Each element of the initial deviation matrix is ​​compared with the corresponding threshold values ​​of the initial deviation boundary upper and lower bound matrices. When an element exceeds the corresponding upper or lower threshold, its corresponding state feature is marked as out of bounds. The out-of-bounds judgment results are then arranged according to their original positions in the initial deviation matrix, outputting an out-of-bounds flag matrix containing Boolean elements, thus completing the out-of-bounds verification. The deviations corresponding to elements that do not exceed the limits are set to zero, maintaining the dimensions of the initial deviation matrix to generate the final deviation matrix.

[0041] In an optional embodiment, during the parameter acquisition phase, outdoor air temperature (e.g., 35°C) and solar radiation intensity (e.g., 800 W / m²) are read in real time from an environmental weather station. 2 The photovoltaic inverter output power (e.g., 50kW) and the charge / discharge status of the energy storage battery (e.g., SOC at 80% and discharge power at 10kW) are obtained in real time through the energy management system.

[0042] The collected numerical values ​​are normalized using Min-Max and then merged to generate a 5×1 multidimensional feature vector. This multidimensional feature vector is then input into a linear classifier model pre-trained based on historical operational data. Using historical building operational data as samples, a linear support vector machine is used for training to solve for the optimal hyperplane weight vector and bias term. Taking the determination of high-load energy storage discharge conditions in summer as an example, the elements of the hyperplane weight vector corresponding to high-load energy storage discharge conditions in summer are 0.4, 0.5, 0.2, -0.3, and 0.1, respectively, with a bias term of -1.5.

[0043] Based on the signed boundary distance calculation method, the inner product of the multidimensional feature vector and the hyperplane weight vector is added with a bias term, and then divided by the Euclidean norm of the hyperplane weight vector to obtain the signed boundary distance of the multidimensional feature vector relative to the preset working condition classification boundary.

[0044] The signed boundary distances of the multidimensional feature vector V relative to the classification boundaries of each preset working condition category are calculated in the same manner, and the signed boundary distances are subjected to Softmax probability normalization to generate a multi-class working condition condition probability distribution sequence with values ​​ranging from 0 to 1.

[0045] For example, after normalization, the probability that the multidimensional feature vector belongs to the summer high-load energy storage and discharge condition is 0.92, the probability that it belongs to the conventional cooling condition is 0.06, and the probability of other condition categories is lower than 0.06. Since 0.92 is the maximum value in the conditional probability distribution sequence, the current operating condition of the building is determined to be the summer high-load energy storage and discharge condition.

[0046] Reference Figure 3 Among the four bar graphs, the bar corresponding to the high-load energy storage and discharge category has the highest height, reaching 0.92, while the bar corresponding to the conventional cooling category has a lower height, at 0.06. The bar heights for the natural ventilation and mild heating categories are close to 0. This significant difference in bar height clearly reflects the probability distribution weights of the multi-dimensional feature vectors in each preset operating condition category. This indicates that through the linear classifier model and Softmax probability normalization, the system can accurately converge environmental parameters and equipment states to specific operating conditions, identifying the current complex operating state.

[0047] In an optional embodiment, a preset operating condition matching database is queried to determine the current operating condition, such as the high-load energy storage discharge condition in summer, as an index. The initial deviation boundaries of the upper and lower limits of the corresponding state feature positions of various key equipment under the current operating condition are retrieved from the preset operating condition matching database. An upper and lower limit matrix of the initial deviation boundaries, corresponding to the current operating condition and consistent with the dimensions and structure of the measured state feature matrix, is obtained. For example, the initial deviation boundaries of the photovoltaic system's photoelectric conversion efficiency-related state features are extracted to be ±5%, and the initial deviation boundaries of the chiller's performance coefficient-related state features are ±0.3.

[0048] The parameter error covariance matrix obtained through recursive updates within the extended Kalman filter framework is read and parsed. The diagonal elements of the parameter error covariance matrix are extracted as the variance values ​​of each key parameter. These variance values ​​are then subjected to reciprocal transformation and normalization to generate a parameter confidence matrix, representing the current accuracy of each key parameter. Assuming the confidence level of the current photovoltaic conversion efficiency is 0.8, while the confidence level of the chiller performance coefficient is lower, at only 0.4, the confidence scaling factor for the state features corresponding to the cross-covariance values ​​can be determined based on the current confidence levels of the key parameters associated with the two sensor channels constituting the cross-covariance values, preferably the smaller of the two current confidence levels.

[0049] Using the extracted confidence level as a scaling factor, the initial deviation threshold of the chiller's performance coefficient (±0.3) is multiplied by a scaling factor of 0.4, reducing it to ±0.12. The photovoltaic conversion efficiency boundary, with higher confidence levels, is reduced to ±4%. This non-uniform scaling mechanism lowers the threshold for parameters with greater uncertainty to trigger out-of-limit states, thus prioritizing the parameter correction mechanism. All weighted and scaled upper and lower limit thresholds are then reorganized to obtain the non-uniform quantized deviation boundary model constraint.

[0050] In an optional embodiment, the consistency of dimension and structure between the measured state feature matrix and the virtual state feature matrix is ​​verified. After confirming structural consistency, difference calculation is performed element by element. For example, if the measured temperature element is 24.5°C and the corresponding virtual simulation temperature element is 24°C, then a deviation value of 0.5°C is generated at the corresponding position in the initial deviation matrix.

[0051] The previously generated non-uniform quantization deviation boundary model is invoked for position-mapping-based mask comparison. Assuming the allowable upper limit threshold for the temperature state after confidence modulation in the boundary model is 0.4°C, and since the initial deviation value of 0.5°C is higher than the corresponding upper limit threshold, the location of the temperature state is marked as a Boolean value of "true," indicating an out-of-limit state. For elements that do not exceed the limit, such as a deviation of 2% between measured and virtual humidity that falls within the reduced upper and lower limit thresholds under the current operating conditions, the location of these elements is marked as a Boolean value of "false." The Boolean values ​​at each location are arranged according to the row and column positions of the initial deviation matrix to generate an out-of-limit identification matrix. The deviation values ​​at out-of-limit locations are retained based on this matrix, while the deviation values ​​at non-out-of-limit locations are set to zero. Thus, after mask filtering and threshold selection, a sparse matrix that filters out normal environmental noise and minor fluctuations is constructed, and this sparse matrix with a dimension of 15×3 is output as the deviation matrix.

[0052] S3. Update the digital twin model and generate a dual-carbon situation. Specifically, when the number of elements in the deviation matrix exceeds the limit, under the extended Kalman filter framework, the deviation matrix is ​​expanded according to a preset element order to obtain a measurement residual vector. Based on the measurement residual vector, the key parameters of the digital twin model are inversely estimated, the expected values ​​of the parameters and the parameter error covariance matrix of the digital twin model are updated, and a parameter confidence matrix is ​​generated based on the updated parameter error covariance matrix. Based on the digital twin model with completed parameter correction, a dual-carbon situation matrix containing real-time carbon flux values, carbon emission trajectory prediction curves, and a carbon neutrality resilience index is generated, and a dual-domain compensation vector for adjusting equipment control parameters is calculated.

[0053] The system iterates through the limit-crossing indicator matrix. If a true element exists, it determines that a state feature has exceeded the limit, and initializes the expected values ​​of the parameters, the parameter error covariance matrix, the process noise covariance matrix, and the measurement noise covariance matrix required for the extended Kalman filter algorithm. The deviation matrix is ​​flattened into a one-dimensional measurement residual vector according to a preset element order. Based on the input-output response relationship of the digital twin model, the measurement Jacobian matrix is ​​obtained using the finite difference sensitivity calculation method. Specifically, preset small perturbations are applied to each of the key parameters to be estimated, and the digital twin model is run to obtain the virtual state feature outputs before and after the perturbation. The virtual state feature outputs are expanded according to the same element order as the measurement residual vector, and the ratio of the expanded output change to the corresponding parameter perturbation is used as the corresponding element in the measurement Jacobian matrix. The small perturbation is taken as 0.1% to 1% of the current value of the key parameter to ensure that the change in the model output after the perturbation is greater than 10 times the floating-point error of the numerical calculation.

[0054] The Kalman gain matrix for the current step is calculated using the measurement Jacobian matrix, the parameter error covariance matrix, and the measurement noise covariance matrix. Multiplying the Kalman gain matrix by the measurement residual vector yields the key parameter correction. This key parameter correction is then added to the expected parameter values ​​from the previous time step, completing the inverse estimation and update of the key parameters. Simultaneously, the parameter error covariance matrix is ​​updated according to the Kalman covariance update equation, and a parameter confidence matrix is ​​generated based on the updated parameter error covariance matrix. The smaller the variance of the corresponding key parameter in the parameter error covariance matrix, the higher the confidence level of the key parameter. The updated expected parameter values ​​are then substituted back into the digital twin model for forward simulation.

[0055] Extract the total building electricity load from the simulation results, multiply it by the grid emission factor to calculate the baseline carbon emissions without considering renewable energy offsets, and then subtract the carbon offsets calculated from photovoltaic power generation to calculate the real-time carbon flux value.

[0056] The Long Short-Term Memory (LSTM) time series prediction algorithm is invoked, and the updated historical carbon flux is used as the input sequence to infer the predicted carbon emission trajectory curve for the next 24 hours.

[0057] By combining extreme environmental disturbance simulations, the maximum offset of carbon flux relative to the normal baseline trajectory and the disturbance recovery time are determined, and the carbon neutrality resilience index is calculated based on the maximum offset and the disturbance recovery time. The real-time carbon flux values, carbon emission trajectory prediction curves, and carbon neutrality resilience index are assembled and merged to generate a dual-carbon situation matrix.

[0058] Based on optimal control theory, the CasADi nonlinear optimization solver is invoked. Using a joint objective function comprising building operation energy costs and carbon emission exceedance penalties as the optimization objective, a dual-domain compensation vector for adjusting equipment control parameters is solved within the feasible domain of the equipment control parameters. This dual-domain solution includes the energy cost domain and the carbon emission constraint domain. By distributing the dual-domain compensation vector to the programmable logic controller (PLC), adaptive adjustment of the building system can be achieved.

[0059] In an optional embodiment, when calculating the current cumulative carbon emissions, the real-time power P of the building's total electrical load is continuously collected with second-level precision, for example, P is maintained at 200kW, while the marginal grid emission factor for the corresponding time period is retrieved, for example, a baseline value of 0.581 kgCO2 / kWh. A Riemann integral is performed within the time interval 0 to t. If the time window is 1 hour, based on the integral result, the baseline cumulative carbon emissions without considering renewable energy offsets are 116.2 kgCO2. Simultaneously, building-integrated photovoltaic (BIPV) system power generation data is collected, and the carbon offset is calculated to be 29.05 kgCO2. Subtracting this from the net carbon offset yields a net carbon accumulation of 87.15 kgCO2. The backward difference method is used to derive the time derivative of the net carbon accumulation, thereby calculating the instantaneous real-time carbon flux value of 87.15 kgCO2 / h.

[0060] The prediction period was set to 24 hours, and extreme disturbance tests were incorporated in conjunction with weather forecasts. Simulations recorded a maximum deviation of 150 kg CO2 / h in carbon flux from the normal baseline trajectory caused by abrupt changes, with a recovery time of 2 hours for the building to return to its initial carbon emission trajectory. Within the set prediction period, the maximum deviation of the building's carbon flux from the normal baseline trajectory under extreme environmental disturbances was determined. The reciprocal of the product of the maximum carbon flux deviation and the disturbance recovery time was used as the carbon neutrality resilience index, resulting in a calculated carbon neutrality resilience index of 0.0033.

[0061] Resilience assessment typically involves solving for the performance loss area of ​​the integrated system. This invention improves upon this by calculating the carbon neutrality resilience index using the reciprocal of the product of the maximum offset and the recovery time. The carbon neutrality resilience index characterizes a building's ability to recover to a stable carbon emission state after a disturbance. A higher value indicates a smaller carbon flux offset or a shorter recovery time.

[0062] A time-series predictive neural network is used to predict future carbon emission trends. The input to the time-series predictive neural network is a multi-dimensional time series matrix containing historical real-time carbon flux values ​​and building power data. The output is a sequence of predicted carbon emission trajectories for the next 24 hours, with a length of 24.

[0063] The time-series prediction neural network employs a Long Short-Term Memory (LSTM) architecture. The input to the LSTM architecture is a multi-dimensional time-series matrix containing historical real-time carbon flux values ​​and building power data. The LSTM architecture includes input and forget gates, as well as output gates and cell state update units, used to recursively model the temporal dependencies in the historical carbon flux sequence. The hidden states output from the LSTM units are input to a fully connected output layer, and regression calculations are performed to obtain the carbon emission trajectory prediction curve sequence.

[0064] The real-time carbon flux value in scalar form and the carbon neutrality resilience index are expanded into a row vector containing 24 elements through sequence replication, and then stacked and merged with the prediction curve in the row dimension to output a 3×24 dimension dual carbon situation matrix.

[0065] In an optional embodiment, during the construction of the joint objective function, the building operation energy purchase cost is calculated to be RMB 80 based on the real-time time-of-use electricity price, such as the peak electricity price of RMB 0.8 per kilowatt-hour, and the predicted electricity consumption, such as 100 kWh.

[0066] For the carbon emission constraint, if the model predicts carbon emissions of 500 kg for the next cycle, while the baseline limit for building allocation is 400 kg, the penalty for exceeding the limit is calculated using nonlinear truncation logic as 100 kg. The carbon emission penalty coefficient is determined with reference to the historical average transaction price in the national carbon emission trading market, ranging from 1.5 to 3; an example uses 2 yuan / (kgCO2). 2 The product of the square of the carbon emission over-limit penalty and the carbon emission penalty coefficient is calculated and then added to the building operation and energy purchase cost to obtain the current joint objective function value of 20080 yuan. While the objective function is generally solved directly by adding costs, this invention improves upon this by introducing a quadratic term for the carbon emission over-limit penalty to obtain the current joint objective function value.

[0067] Using the corrected parameters of the digital twin model as the basic input, the Newton-Raphson iteration method is employed to find the minimum value of the joint objective function within the feasible region of the control parameters. Specifically, in the optimization phase, the air conditioning set temperature and energy storage charging / discharging power in the corrected and updated digital twin model are used as control decision variables. The gradient vector and Hessian matrix of the joint objective function relative to the control decision variables are calculated. Subsequently, a Newton-Raphson iteration algorithm with boundary projection is executed: first, the Newton update direction is calculated based on the gradient vector and Hessian matrix, and the control decision variables are updated accordingly; then, boundary projection is performed on the updated control decision variables to keep them within the physical safety limits of the equipment. The physical safety limits of the equipment include the air conditioning temperature set range of 22°C to 28°C and the energy storage charging / discharging power not exceeding 100kW.

[0068] Based on the engineering requirements for sensitivity to changes in the objective function value, i.e., the resolution of the actual impact on building energy purchase costs after the control command is issued, a preset convergence accuracy parameter of 0.01 yuan is set. After each update of the control parameters, a new round of objective function value is calculated. The iteration termination condition is met when the absolute value of the difference between the result of the nth iteration and the result of the (n-1)th iteration is less than or equal to 0.01.

[0069] The decision variables under the minimum convergence condition, such as setting the air conditioner temperature to 26.5°C and the energy storage discharge power to 80kW, are converted into specific electrical signal control sequences and issued as a dual-domain compensation vector that includes the coordinated adjustment of the energy consumption cost domain and the carbon emission constraint domain.

[0070] The experiment was conducted in a region with a total area of ​​15,000 m². 2 A controlled experiment was conducted in a typical commercial building, with a testing period of 30 consecutive days during the summer. The experimental hardware consisted of a 500kW chiller and a 100kWh energy storage battery pack within the building. The experiment was divided into two groups: the control group employed a traditional fixed deviation boundary and single energy cost optimization control strategy, while the experimental group adopted the proposed parameter confidence dynamic modulation non-uniform quantization deviation boundary model and a dual-domain compensation strategy combining a dual-carbon situational matrix and a joint objective function. Both systems were tested at an outdoor temperature of 35°C and a solar radiation intensity of 800W / m². 2 Parallel testing was conducted under the same initial environmental conditions.

[0071] After continuous operation, the control group had an average daily energy purchase cost of 2150 yuan, an average daily carbon emission exceeding the limit by 45 kg, an average convergence time of 45 s for Newton's iteration optimization of equipment control parameters, and a carbon neutrality resilience index of only 0.0015 in response to extreme load disturbances. The experimental group reduced its average daily energy purchase cost to 1890 yuan and its average daily carbon emission exceeding the limit to 2 kg. Due to the filtering out of minor environmental noise, the experimental group shortened the average convergence time for parameter iteration optimization to 12 s. The carbon flux recovery time after a disturbance was reduced from 3 h to 1 h, and the calculated carbon neutrality resilience index increased to 0.0035.

[0072] Reference Figure 4 As shown in the figure, the heights of the energy purchase cost bar chart and the carbon emission exceedance bar chart of the experimental group are significantly lower than those of the corresponding bar charts of the control group. Furthermore, the convergence time curve of the experimental group remains at a lower time level. The difference in the height of the two groups of bars clearly reflects that by constructing a joint objective function that includes a carbon emission penalty term and combining it with a dual carbon situation matrix for optimization, the system can reduce the building's energy purchase cost while reducing the risk of carbon emission exceedance. This demonstrates the superiority of the dual-domain compensation vector calculation method in achieving dual constraint control of microgrid energy consumption and carbon emissions.

[0073] This invention utilizes a confidence matrix to dynamically adjust the non-uniform quantization deviation boundary, implementing stricter mismatch detection and threshold screening for low-confidence parameters. This helps reduce the computational load of invalid residuals, thereby accelerating model convergence and improving the accuracy of control parameter solutions to some extent. By employing a dual-carbon situation matrix and constructing a joint objective function that includes a carbon emission penalty term, the system can comprehensively adjust the air conditioning setpoint temperature and energy storage charging / discharging power within the feasible region. This reduces energy purchase costs while enhancing the carbon emission reduction stability and operational resilience of building microgrids.

[0074] This invention also discloses a dual-carbon situational awareness system for zero-carbon building digital twins, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a dual-carbon situational awareness method for zero-carbon building digital twins according to the present invention.

[0075] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A dual-carbon situational awareness method suitable for zero-carbon building digital twins, characterized in that, include: S1: Obtain the sensor data stream of the zero-carbon building entity and the virtual sensor data stream of the digital twin model under the same control command, and construct the measured state feature matrix and the virtual state feature matrix respectively; S2: Determine the current operating conditions based on the building's external environmental parameters and internal equipment status parameters. Combine this with the parameter confidence matrix to generate a non-uniform quantization deviation boundary model dynamically modulated by the parameter confidence. Calculate the deviation matrix between the measured state feature matrix and the virtual state feature matrix, and verify whether the deviation matrix exceeds the constraints of the non-uniform quantization deviation boundary model. S3: When the number of matrix elements in the deviation matrix exceeds the limit, the deviation matrix is ​​expanded under the extended Kalman filter framework to obtain the measurement residual vector. Based on the measurement residual vector, the key parameters of the digital twin model are back-estimated, and the expected values ​​of the parameters and the parameter error covariance matrix are updated to generate the updated parameter confidence matrix. Based on the corrected digital twin model, a dual carbon situation matrix containing real-time carbon flux values, carbon emission trajectory prediction curves, and carbon neutrality resilience index is generated, and a dual-domain compensation vector for adjusting the equipment control parameters is calculated.

2. The dual-carbon situational awareness method for zero-carbon building digital twins according to claim 1, characterized in that, The steps of constructing the measured state feature matrix and the virtual state feature matrix respectively include: acquiring the current time measurement values ​​of multiple sensor channels and converting them into state column vectors; extracting the state measurement values ​​of the previous sampling period, subtracting the current time measurement value from the state measurement value of the previous sampling period, and dividing by a constant sampling period to calculate the first-order rate of change of each sensor data stream; for a preset combination of sensor data streams within a preset time window, extracting the sample distribution of the observation series of the data stream within the time window and calculating the cross-covariance value; arranging the acquired real-time measurement values, first-order rate of change, and cross-covariance values ​​according to a preset correspondence to generate the measured state feature matrix; and generating the virtual state feature matrix using the virtual sensor data stream according to the same data structure.

3. The dual-carbon situational awareness method for zero-carbon building digital twins according to claim 1, characterized in that, The step of determining the current operating condition based on external environmental parameters and internal equipment status parameters includes: acquiring external environmental parameters and internal equipment status parameters and merging them into a multi-dimensional feature vector; inputting the multi-dimensional feature vector into a linear classifier model and calculating the signed boundary distance of the multi-dimensional feature vector relative to the classification boundary of each preset operating condition category; performing probability normalization processing on the signed boundary distance corresponding to each preset operating condition category to generate a conditional probability distribution, and determining the preset operating condition category corresponding to the maximum conditional probability as the current operating condition.

4. The dual-carbon situational awareness method for zero-carbon building digital twins according to claim 1, characterized in that, The step of generating a non-uniform quantization deviation boundary model dynamically modulated by the parameter confidence matrix includes: querying a preset operating condition matching database to obtain an initial deviation boundary upper limit matrix and an initial deviation boundary lower limit matrix that correspond to the current operating condition and have the same dimensional structure; extracting the current confidence of each key parameter in the parameter confidence matrix and mapping it to a confidence scaling matrix that has the same dimension and structure as the measured state feature matrix; and multiplying the initial deviation boundary upper limit matrix and the initial deviation boundary lower limit matrix element-wise with the confidence scaling matrix to generate a non-uniform quantization deviation boundary model.

5. The dual-carbon situational awareness method for zero-carbon building digital twins according to claim 1, characterized in that, The calculation of the deviation matrix between the measured state feature matrix and the virtual state feature matrix, and the verification of whether the deviation matrix exceeds the constraints of the non-uniform quantization deviation boundary model, includes: verifying the consistency of the dimension and structure between the measured state feature matrix and the virtual state feature matrix; if the structure is consistent, performing a full matrix operation by subtracting corresponding elements of the matrices to obtain the initial deviation matrix; comparing each element in the initial deviation matrix with the corresponding upper and lower limit thresholds in the non-uniform quantization deviation boundary model item by item; when an element exceeds the corresponding upper and lower limit thresholds, marking the corresponding state feature as exceeding the limit and generating an over-limit identification matrix; setting the deviations corresponding to the non-over-limit elements to zero according to the over-limit identification matrix, and generating the deviation matrix while maintaining the dimension of the initial deviation matrix.

6. The dual-carbon situational awareness method for zero-carbon building digital twins according to claim 1, characterized in that, The process of updating the expected value of the parameters and the parameter error covariance matrix includes: applying a preset small perturbation to the key parameters to be estimated to run a digital twin model to obtain virtual state feature output; expanding the virtual state feature output in the same element order as the measurement residual vector; using the ratio of the expanded output change to the corresponding parameter perturbation as the corresponding element in the measurement Jacobian matrix; calculating the Kalman gain matrix using the measurement Jacobian matrix, the parameter error covariance matrix, and the measurement noise covariance matrix; multiplying the Kalman gain matrix by the measurement residual vector to obtain the key parameter correction; and superimposing the key parameter correction onto the expected value of the parameters at the previous time step to obtain the updated expected value of the parameters; and updating the parameter error covariance matrix based on the Kalman gain matrix and the measurement Jacobian matrix.

7. The dual-carbon situational awareness method for zero-carbon building digital twins according to claim 1, characterized in that, The real-time carbon flux value is obtained through the following steps: obtaining the real-time power of the building's total electrical load and the grid emission factor for the corresponding time period, and calculating the baseline cumulative carbon emissions without considering renewable energy offsets by integration; Obtain the carbon offset of the building photovoltaic system, subtract the carbon offset from the baseline cumulative carbon emissions to obtain the net carbon accumulation; calculate the time derivative of the net carbon accumulation to obtain the real-time carbon flux value.

8. The dual-carbon situational awareness method for zero-carbon building digital twins according to claim 7, characterized in that, The generation of a dual-carbon situation matrix based on the corrected digital twin model includes real-time carbon flux values, carbon emission trajectory prediction curves, and a carbon neutrality resilience index. This includes: determining the maximum offset of carbon flux relative to the normal baseline trajectory when a building is subjected to extreme environmental disturbances; using the reciprocal of the product of the maximum offset and the disturbance recovery time as the carbon neutrality resilience index; and merging the real-time carbon flux values ​​and the carbon neutrality resilience index after expanding and aligning them according to the time series length of the carbon emission trajectory prediction curve generated based on a long short-term memory network to generate the dual-carbon situation matrix.

9. The dual-carbon situational awareness method for zero-carbon building digital twins according to claim 1, characterized in that, The calculation of the dual-domain compensation vector used to adjust the equipment control parameters includes: taking the positive difference between the predicted carbon emissions and the set carbon emission benchmark limit as the carbon emission over-limit penalty; adding the square of the carbon emission over-limit penalty to the carbon emission penalty coefficient and the building operation energy purchase cost to obtain the joint objective function; using the parameters corrected by the digital twin model as the basic input, and using the Newton iteration method to solve for the minimum value of the joint objective function in the feasible domain of the control parameters.

10. The dual-carbon situational awareness method for zero-carbon building digital twins according to claim 9, characterized in that, The Newton-Raphson iteration method is used to find the minimum value of the joint objective function within the feasible region of the control parameters. This includes: calculating the gradient vector and Hessian matrix of the joint objective function with respect to the control decision variables; using the gradient vector and Hessian matrix to calculate the Newton update direction to update the control decision variables, and performing boundary projection on the updated control decision variables to keep them within the physical safety limits of the equipment; when the absolute value of the difference between the objective function values ​​of two adjacent iterations is not greater than the preset convergence accuracy, the iteration is terminated and the current equipment control parameters are output as a two-domain compensation vector.

Images