Building digital twin data processing method based on artificial intelligence

Abstract

The invention relates to the field of electric digital data processing, in particular to a building digital twin data processing method based on artificial intelligence, which comprises the following steps of: acquiring real building digital twin data, training a data expansion model by adopting a generative adversarial network algorithm based on topology dual optimization to obtain training data, and calculating the real building digital twin data according to the training data. Obtaining a central building energy consumption evaluation model and a local building energy consumption evaluation model by adopting a distributed federated learning method, training the local building energy consumption evaluation model of each building by adopting a self-adaptive constraint-based extreme learning machine algorithm, continuously iterating the local building energy consumption evaluation model, and evaluating the energy consumption of each building. And evaluating new real building digital twin data by using the trained local building energy consumption evaluation model. An existing building digital twin data processing method has the problem that energy consumption grade evaluation is not accurate, and the method is accurate in energy consumption grade evaluation.

Description

Technical Field

[0001] The present invention relates to the field of electrical digital data processing, and in particular to an artificial intelligence-based building digital twin data processing method. Background Art

[0002] In order to achieve energy conservation and emission reduction in buildings, it is necessary to conduct building energy consumption assessment. During the assessment, usually not just a single building is assessed, but an entire building complex is assessed. For example, a community is about to deliver a new phase of housing, which is usually a building complex composed of multiple independent buildings. During the simulation, the operation status of the entire building complex will be simulated, and the overall building energy consumption assessment will also be conducted on the building complex. Because building energy consumption must be assessed based on data from building operation, traditional building energy consumption is always assessed after the actual operation of the building. If the energy consumption is found to be too high after the energy consumption assessment, it will be more troublesome to rectify.

[0003] Therefore, people combine digital twin technology with building energy consumption assessment to create a digital copy that is completely consistent with the actual building in a virtual environment to simulate the building. Specifically, taking the above-mentioned building complex as an example, people will deploy a variety of sensors in each building to collect real-time data of the building, such as indoor and outdoor temperature, humidity, light intensity, etc., and transmit the collected data to the digital copy. The digital copy simulates the operating status of the building based on these data, thereby obtaining the simulated operating energy consumption data of each building, namely the building digital twin data. People use the building digital twin data processing method to evaluate the building energy consumption based on the building digital twin data, obtain the energy consumption level, and take corresponding actions based on the energy consumption level. After the action, the newly collected data can be uploaded and simulated again, which is very convenient and fast.

[0004] The existing building digital twin data processing method mainly uses the building energy consumption assessment model to evaluate the energy consumption level of the building digital twin data after obtaining the building digital twin data, and obtains its energy consumption level. The building energy consumption assessment model is essentially a classifier. However, the number of existing building digital twin data is limited, and the insufficient amount of training data will limit the training effect and generalization ability of the classifier, thereby affecting the classification results of the classifier. At the same time, the existing classifiers mainly use traditional machine learning algorithms for classification, but these algorithms do not adequately handle the multidimensional characteristics and nonlinear relationships in the building digital twin data, and cannot accurately capture the complex collaborative relationship between features. The evaluation accuracy and adaptability of the classifier are restricted, and there is a lack of effective feature selection mechanism. It is difficult to screen out key features that have a significant impact on the energy consumption assessment task in high-dimensional data. There is feature redundancy in the classifier training process, which leads to reduced computational efficiency and insufficient reliability of the evaluation results.

[0005] Therefore, the existing building digital twin data processing methods have the problem of inaccurate assessment of energy consumption levels. Summary of the invention

[0006] The technical problem to be solved by the present invention is to overcome the shortcomings of the above-mentioned prior art and provide an artificial intelligence-based building digital twin data processing method that is more accurate in evaluating energy consumption levels.

[0007] In order to solve the above technical problems, the present invention provides an artificial intelligence-based building digital twin data processing method comprising: S1. Obtain the real building digital twin data of each building in the building complex, and mark the energy consumption level corresponding to the real building digital twin data of each building; S2. According to the real building digital twin data of each building, a data expansion model is trained using a generative adversarial network algorithm based on topological dual optimization to obtain a trained data expansion model. The trained data expansion model of each building generates synthetic building digital twin data of the building according to the real building digital twin data of the building, and the real building digital twin data and synthetic building digital twin data of each building are used as the training data of the building; S3. Use the distributed federated learning method to construct and initialize the central building energy consumption assessment model on the central server, send the central building energy consumption assessment model to each building to form a local building energy consumption assessment model, and use the training data of each building to train the local building energy consumption assessment model of each building using the extreme learning machine algorithm based on adaptive constraints to obtain the parameters of each local building energy consumption assessment model; S4. Each local building energy consumption assessment model sends the trained parameters to the central server, which aggregates the parameters and updates the central building energy consumption assessment model. The updated central building energy consumption assessment model is then sent to each building to form a new local building energy consumption assessment model. The new local building energy consumption assessment model repeats S3 to start a new iteration until the training converges, and the trained central building energy consumption assessment model and the trained local building energy consumption assessment model are obtained. S5. Obtain new real building digital twin data for each building in the building complex, input the new real building digital twin data of each building into the corresponding trained local building energy consumption assessment model, and obtain the energy consumption level corresponding to the new real building digital twin data of each building.

[0008] As a further improvement of the present invention: S2 uses a generative adversarial network algorithm based on topological dual optimization to generate synthetic building digital twin data for each building, including: S201. Initialize the generative adversarial network; S202. When training the generator, a random noise vector is received and samples with a distribution similar to that of the real building digital twin data are generated. The generator adopts topological dual optimization in the loss function, and the calculation method of the generator loss is expressed as: , In the formula, is the loss function of the generator, is the expected value of the input noise, is the discriminator function, is the generator function, Output building digital twin data to the generator, is the input noise, is the weight parameter of the generator, is the weight coefficient, is the topological dual loss function; S203. After the generator outputs the building digital twin data, the discriminator receives the real building digital twin data and the synthetic building digital twin data, and continuously improves the discrimination ability by distinguishing the two. The Riemann manifold regularization method is used in the training process of the discriminator, and the calculation method of the discriminator loss is expressed as: , In the formula, is the discriminator loss function, is the expected value of the real building digital twin data, is the discriminator function, For real building digital twin data, is the expected value of the input noise, Output building digital twin data to the generator, is the weight coefficient of Riemann manifold regularization; is the Riemann manifold regularization term; S204. It is necessary to iteratively update the parameters of the generator and discriminator based on the adversarial loss and its additional topological and Riemann manifold constraints. The parameter update method of the generator and discriminator is expressed as: , , In the formula, For the The generator parameters for the iterations, For the The generator parameters for the iterations, is the learning rate of the generator, is the gradient of the generator’s loss function with respect to its parameters, For the The discriminator parameters for the iteration, For the The discriminator parameters of the iteration, is the learning rate of the discriminator, is the gradient of the discriminator's loss function with respect to its parameters; S205. Repeat the above steps until a preset stop iteration condition is met.

[0009] Preferably, the topological dual loss function in S202 is The calculation method is expressed as: , In the formula, To constrain the topological distance of the overall distribution of real building digital twin data, To map the synthetic building digital twin data in the topological space, It is the mapping of real building digital twin data in topological space. For real building digital twin data, is the probability density of real building digital twin data.

[0010] Preferably, the mapping of synthetic building digital twin data in topological space Mapping with real building digital twin data in topological space The calculation method is expressed as: , , In the formula, To map the digital twin data of the synthetic building in the topological space; is a mapping function, specifically a variational autoencoder function; is the generator function, is the input noise, is the weight parameter of the generator; are the parameters of the mapping model, specifically the parameters of the variational autoencoder; It is the mapping of real building digital twin data in topological space. It is the digital twin data of real buildings.

[0011] Preferably, the Riemann manifold regularization term in S203 The calculation method is expressed as: , In the formula, is the Riemann manifold regularization term, For input and expectations, is the gradient of the discriminator with respect to the input building digital twin data; It is the Hessian matrix of the building digital twin data on the manifold, which describes the local curvature information of the building digital twin data manifold.

[0012] As a further improvement of the present invention: the local building energy consumption assessment model of each building trained by the extreme learning machine algorithm based on adaptive constraints in S3 includes: S301. Initialize the extreme learning machine model; S302. Adopt the dynamic feature weighting adjustment method of adaptive constraint optimization to dynamically weight each feature and adjust its influence on the final classification result. Define the weighting coefficient to represent the importance of each feature in the final classification. The calculation method is expressed as: , In the formula, For the The weighting coefficient of each feature; is the attenuation factor; For the Features, For all building digital twin data input to the extreme learning machine The mean of the features represents the position of the feature in the building digital twin data; is the weight factor of the gradient information, balancing the influence of the gradient size on the weight coefficient; is the loss function of the extreme learning machine with respect to the The gradient of a feature; S303. A dynamic penalty term is used to solve the constraint problem of the loss function in the extreme learning training process. The weight of the extreme learning machine is adaptively adjusted when facing non-uniform or dynamically changing building digital twin data streams. The loss function calculation method of the extreme learning machine is expressed as: , In the formula, is the loss function of the extreme learning machine, are all the parameters of the extreme learning machine model, is the L2 norm, is the target output matrix, is the hidden layer output of the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine; is the regularization coefficient of the extreme learning machine; is the dynamic penalty coefficient of the extreme learning machine; is the penalty function of the extreme learning machine, is the building digital twin data matrix input to the extreme learning machine; S304. A dynamic search strategy based on a constraint set is adopted. The initialized constraint set stores the correlation between each feature and the classification result. The constraint set is incrementally updated according to the classification error and feature contribution of each iteration, so that the model focuses on features with high contribution and dynamically removes features that are not beneficial to classification or bring interference. The calculation method is expressed as: , In the formula, is the constraint set, For collection update operations, is the hidden layer activation function, is the building digital twin data matrix input to the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine; S305. Considering the influence of the extreme learning machine to the hidden layer parameters on the classification performance and the joint effect of the dynamic penalty term, the hidden layer parameters are interactively updated, and the calculation method is expressed as: , In the formula, is the gradient of the extreme learning machine’s loss function with respect to the hidden layer weights, is the hidden layer output of the extreme learning machine, is the transpose of the output of the hidden layer of the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine, is the target output matrix; is the gradient regularization term coefficient; is the weight matrix of the hidden layer of the extreme learning machine; S306. Adopt adaptive parameters to update the dynamic penalty term coefficient of the extreme learning machine weight and perform model complexity control. The calculation method is expressed as: , In the formula, For the The dynamic penalty coefficient of the extreme learning machine weight of the iteration, For the The dynamic penalty coefficient of the extreme learning machine weights for the iteration; is the learning rate decay coefficient; is the current iteration number; S307. Dynamically adjust the loss function, using a dynamic adjustment item based on the gradient size so that the loss function can be flexibly adapted according to the current training progress and error state. The calculation method is expressed as: , In the formula, For the The loss function of the extreme learning machine at the iteration is, For the The loss function of the extreme learning machine at the iteration; is the dynamic adjustment coefficient of the loss function; For the The gradient of the extreme learning machine's loss function with respect to all parameters at the iteration; S308. Repeat iterations S302 to S307 until a preset stop iteration condition is met.

[0013] Preferably, initializing the extreme learning machine model in S301 includes: Initialize the parameters of the extreme learning machine, which include the weight and bias parameters of the extreme learning machine. The initialization method is random initialization and the initialized parameters obey the normal distribution with a mean of 0 and a variance of the unit matrix. The activation method of the hidden layer activation function of the extreme learning machine is defined so that the building digital twin data processing method can provide good expression ability for the nonlinear characteristic relationship of multi-dimensional building digital twin data. The calculation method is expressed as: , In the formula, is the hidden layer output of the extreme learning machine; is the hidden layer activation function, and the hidden layer activation function adopts the Sigmoid activation function; is the building digital twin data matrix input to the extreme learning machine, is the weight matrix of the hidden layer of the extreme learning machine, is the hidden layer bias of the extreme learning machine.

[0014] The beneficial effects of the present invention are as follows: The artificial intelligence-based building digital twin data processing method provided by the present invention is more accurate in evaluating energy consumption levels.

[0015] This method adds a step of using a data augmentation model to augment the real building digital twin data, and uses a generative adversarial network algorithm based on topological dual optimization to train the data augmentation model. The generator generates samples similar to the real data distribution through random noise, and combines the topological dual loss constraint to optimize the geometric consistency and authenticity of the generated data. The discriminator improves the recognition ability of complex data structures through manifold regularization. In addition, by using variational autoencoders for high-dimensional topological mapping, the high-dimensional characteristics of building digital twin data can be effectively captured, the distortion of generated data can be avoided, and the applicability of generated data in complex building structures can be enhanced.

[0016] At the same time, this method uses an extreme learning machine algorithm based on adaptive constraints to train each local building energy consumption assessment model. The extreme learning machine algorithm based on adaptive constraints optimizes the importance of features according to data distribution and gradient information through a dynamic feature weighted adjustment method, thereby improving the adaptability to complex multidimensional features. At the same time, the dynamic penalty term combines distribution offset and gradient information to optimize the loss function to adapt to non-uniform data flows, thereby enhancing model stability and prediction accuracy. In addition, in order to effectively screen key features, a dynamic search strategy based on constraint sets is adopted. By iteratively updating the relevance and contribution of features, the most useful features for energy consumption assessment are screened out, reducing feature redundancy and improving model efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is the influence diagram of the regularization parameter on the classification accuracy; Figure 2 Importance map of different features to the model output; Figure 3 This is the influence diagram of the dynamic penalty term on the model stability; Figure 4 The following are performance graphs of various algorithms on complex data sets. DETAILED DESCRIPTION

[0018] The specific implementation modes of the present invention are further described in detail below in conjunction with the accompanying drawings.

[0019] Definitions of relevant terms involved in this invention: (1) Building energy consumption: According to the definition in the China Building Energy Consumption Research Report (2016), building energy consumption refers to the energy consumption during the building operation phase, including the energy consumption of terminal equipment for maintaining the building environment (such as heating, cooling, ventilation, air conditioning and lighting, etc.) and the energy consumption of terminal equipment for various activities within the building (such as office, cooking, etc.).

[0020] (2) Digital Twin: Digital Twin is a comprehensive technology. It is the result of the definition of the materialized things in the objective world and their development laws by software. It can be understood as a 1:1 virtual version (i.e., twin) of a physical object in time and space. Simply put, a digital twin can be described as a highly complex digital model of an object. If the object is a real building, then the digital twin is a digital copy of the real building and its systems and environment (temperature, humidity, air quality index, etc.).

[0021] The present invention provides an artificial intelligence-based building digital twin data processing method comprising: S1. Obtain the real building digital twin data of each building in the building complex, manually annotate the energy consumption level corresponding to the real building digital twin data of each building and store it in JSON format; The real in the real building digital twin data is to distinguish it from the synthetic building digital twin data in the later stage. The attributes of the real building digital twin data include building area energy consumption, indoor temperature, outdoor temperature, water consumption, peak power consumption, power efficiency, lighting energy consumption, heating, ventilation and air conditioning system energy consumption, building physical status and building functional status.

[0022] Energy consumption levels include low energy consumption, medium energy consumption and high energy consumption.

[0023] S2. According to the real building digital twin data of each building, a data expansion model is trained using a generative adversarial network algorithm based on topological dual optimization to obtain a trained data expansion model. The trained data expansion model of each building generates synthetic building digital twin data of the building according to the real building digital twin data of the building, and the real building digital twin data and synthetic building digital twin data of each building are used as the training data of the building; The generative adversarial network algorithm based on topological dual optimization refers to the generator adopting a topological dual optimization strategy in the loss function to improve the authenticity and consistency of the topological structure of the synthetic building digital twin data in the high-dimensional space. The topological dual optimization strategy is to realize the synthetic building digital twin data through the constraints of the topological dual loss. Specifically, the training process is constrained by measuring the distance between the synthetic building digital twin data and the real building digital twin data in the topological space. The mapping of the synthetic building digital twin data in the topological space and the mapping of the real building digital twin data in the topological space are both implemented by a feature learning method based on variational autoencoders to effectively capture the high-dimensional characteristics and implicit topological structure of the data.

[0024] The synthetic building digital twin data of each building is generated using a generative adversarial network algorithm based on topological dual optimization, including: S201. Initializing the generative adversarial network, including: initializing the parameters of the generator and the discriminator, wherein the initialization method is random initialization and the initialized parameters obey a normal distribution with a mean of 0 and a variance of a unit matrix; S202. When training the generator, a random noise vector is received and samples with a distribution similar to that of the real building digital twin data are generated. In order to improve the authenticity and richness of the synthetic building digital twin data in high-noise and diverse data scenarios, the generator adopts topological dual optimization in the loss function to ensure that the topological structure of the synthetic building digital twin data in high-dimensional space can better match the real building digital twin data, solving the problem that the traditional generative adversarial network is insufficient in describing complex topological structures. The calculation method of the generator loss is expressed as: , In the formula, is the loss function of the generator, is the expected value of the input noise, is the discriminator function, is the generator function, Output building digital twin data to the generator, is the input noise, is the weight parameter of the generator; is the weight coefficient of the topological duality loss used to control the influence of the topological duality loss on the total loss, which is set to 0.3; is the topological dual loss function; Topological dual loss function in S202 Indicates the distance between the synthetic building digital twin data and the real building digital twin data in the topological space. In order to realize the topological space constraints of the synthetic building digital twin data and ensure the consistency of the synthetic building digital twin data in the topological structure, the topological loss function is expressed by measuring the distance between the synthetic building digital twin data and the real building digital twin data in the topological space. When generating a complex building digital twin model, the topological dual optimization can ensure that the generated model is consistent in spatial layout and connectivity. The calculation method is expressed as: , In the formula, To constrain the topological distance of the overall distribution of real building digital twin data, To map the synthetic building digital twin data in the topological space, It is the mapping of real building digital twin data in topological space. For real building digital twin data, is the probability density of real building digital twin data.

[0025] Mapping of synthetic building digital twin data in topological space Mapping with real building digital twin data in topological space The mapping of building digital twin data in topological space adopts the feature learning method based on variational autoencoder to realize topological mapping, which can effectively capture the high-dimensional characteristics and implicit topological structure of the data, so that the generated data not only conforms to the statistical distribution, but also maintains authenticity in geometric structure. For twin data containing complex curves or multi-story building layouts, the variational autoencoder mapping can avoid the distortion problem caused by insufficient feature extraction. The calculation method is expressed as: , , In the formula, To map the digital twin data of the synthetic building in the topological space; is a mapping function, specifically a variational autoencoder function; is the generator function, is the input noise, is the weight parameter of the generator; are the parameters of the mapping model, specifically the parameters of the variational autoencoder, and are the training parameters; It is the mapping of real building digital twin data in topological space. It is the digital twin data of real buildings.

[0026] S203. After the generator outputs the building digital twin data, the discriminator receives the real building digital twin data and the synthetic building digital twin data, and continuously improves the discrimination ability by distinguishing the two. In order to solve the discrimination difficulties caused by the high-noise and complexly distributed building digital twin data, the Riemann manifold regularization method is used in the training process of the discriminator, so that the discriminator can pay attention to the geometric structure of the building digital twin data distribution and improve its recognition ability of the complex building digital twin data manifold, so as to more accurately distinguish the real and synthetic building digital twin data. The calculation method of the discriminator loss is expressed as: , In the formula, is the discriminator loss function, is the expected value of the real building digital twin data, is the discriminator function, For real building digital twin data, is the expected value of the input noise, Output building digital twin data to the generator, is the weight coefficient of Riemann manifold regularization, set to 0.3; is a Riemann manifold regularization term used to adjust the matching degree of the discriminator to the manifold structure of the building digital twin data; Riemannian manifold regularization term Based on the curvature constraint of building digital twin data in high-dimensional space, the discriminator can better grasp the geometric structure. The manifold regularization term enables the discriminator to better understand the local curvature information of building digital twin data in high-dimensional space. For digital twin data containing complex polyhedral building designs, the manifold regularization term can help the discriminator more accurately distinguish between real and generated samples and improve the robustness of the discrimination results. The calculation method is expressed as: , In the formula, is the Riemann manifold regularization term used to adjust the matching degree of the discriminator to the manifold structure of the building digital twin data, For input and expectations, is the gradient of the discriminator with respect to the input building digital twin data; It is the Hessian matrix of the building digital twin data on the manifold, which describes the local curvature information of the building digital twin data manifold. It is used to enable the discriminator to better understand the manifold structure of the building digital twin data and avoid classification based only on the local linear structure.

[0027] Hessian matrix of building digital twin data on manifold The local curvature information of the building digital twin data manifold is described by the second-order derivative, which is expressed as: , In the formula, represents the second-order partial derivative of the discriminator with respect to the input building digital twin data, is the discriminator function, It is the digital twin data of real buildings; S204. It is necessary to iteratively update the parameters of the generator and the discriminator based on the adversarial loss and its additional topological and Riemann manifold constraints to strike a balance between the statistical distribution and geometric structure of the high-dimensional space, thereby improving the stability and accuracy of the generator in generating realistic building digital twin data and the discriminator in correctly identifying building digital twin data. The parameter update method of the generator and the discriminator is expressed as: , , In the formula, For the The generator parameters for the iterations, For the The generator parameters for the iterations, is the index of the iteration number, is the learning rate of the generator, Set to 0.001; is the gradient of the generator’s loss function with respect to its parameters, For the The discriminator parameters of the iteration, For the The discriminator parameters of the iteration; is the learning rate of the discriminator, Set to 0.0005; is the gradient of the discriminator's loss function with respect to its parameters; S205. Repeat the above steps until the preset stop iteration condition is met, indicating that the model training is completed. The preset stop iteration condition is reaching the preset maximum number of iterations, and the preset maximum number of iterations is set to 1000 times.

[0028] If there are 800 pieces of real building digital twin data collected and 200 pieces of synthetic building digital twin data are generated, there will be a total of 1,000 pieces of training data.

[0029] S3. Use the distributed federated learning method to construct and initialize the central building energy consumption assessment model on the central server, send the central building energy consumption assessment model to each building to form a local building energy consumption assessment model, and use the training data of each building to train the local building energy consumption assessment model of each building using the extreme learning machine algorithm based on adaptive constraints to obtain the parameters of each local building energy consumption assessment model; The extreme learning machine algorithm based on adaptive constraints refers to the use of a dynamic feature weighting method with adaptive constraint optimization to dynamically adjust multicollinearity and nonlinear feature relationships, thereby enhancing the model's adaptability under complex building digital twin data relationships. At the same time, a dynamic penalty term is used, combined with the building digital twin data distribution offset and gradient information, to achieve adaptive optimization of the loss function, thereby maintaining model stability in the non-uniform data stream of the building digital twin and improving classification accuracy and prediction reliability.

[0030] The local building energy consumption assessment model of each building is trained using the extreme learning machine algorithm based on adaptive constraints, including: S301. Initializing the extreme learning machine model, including: initializing the parameters of the extreme learning machine, the parameters of the extreme learning machine include the weight and bias parameters of the extreme learning machine, the initialization method is random initialization and the initialized parameters obey the normal distribution with a mean of 0 and a variance of the unit matrix; At the same time, the activation method of the hidden layer activation function of the extreme learning machine is defined, so that the building digital twin data processing method can provide good expression ability for the nonlinear characteristic relationship of multi-dimensional building digital twin data. The calculation method is expressed as: , In the formula, is the hidden layer output of the extreme learning machine; is the hidden layer activation function, and the hidden layer activation function adopts the Sigmoid activation function; is the building digital twin data matrix input to the extreme learning machine, is the weight matrix of the hidden layer of the extreme learning machine, is the hidden layer bias of the extreme learning machine.

[0031] S302. The dynamic feature weighting adjustment method of adaptive constraint optimization is used to enhance the adaptability of the model when processing building digital twin data with complex relationships. In particular, when facing features with multicollinearity or nonlinear relationships, each feature can be dynamically weighted to adjust its impact on the final classification result. The weighting coefficient is defined to represent the importance of each feature in the final classification, which is not only related to the relevance of the feature (determined by the gradient and data offset), but also related to the nonlinear relationship between the feature and the target category. The calculation method is expressed as: , In the formula, For the The weighting coefficient of the feature, is the feature index of the sample; The attenuation factor used to adjust the weighting coefficient controls the influence of the relationship between the characteristics and the building digital twin data distribution on the weighting coefficient; For the Features, For all building digital twin data input to the extreme learning machine The mean of the features represents the position of the feature in the building digital twin data; is the weight factor of the gradient information, balancing the influence of the gradient size on the weight coefficient; is the loss function of the extreme learning machine with respect to the The gradient of a feature; The feature vector of the building digital twin data input into the extreme learning machine is weighted according to the weighted coefficient of the feature. In view of the multidimensional, nonlinear and multicollinear characteristics of building digital twin data, in the building energy consumption assessment, there may be complex nonlinear relationships between features such as regional energy consumption, indoor temperature and HVAC system energy consumption. By adjusting the feature weights in real time, the contribution of each feature is dynamically optimized according to the data distribution and feature gradient relationship. For example, when the coordinated changes of indoor temperature and HVAC system energy consumption significantly affect the energy consumption level, the dynamic weighting mechanism can give these two higher weights, thereby improving the model's ability to identify high-energy-consuming buildings, expressed as: , In the formula, is the weighted feature vector; is the sample feature dimension input to the extreme learning machine, For the The weighting coefficient of the feature, For the Features.

[0032] S303. A dynamic penalty term is used to solve the constraint problem of the loss function in the extreme learning training process. The weight of the extreme learning machine is adaptively adjusted when facing non-uniform or dynamically changing building digital twin data streams. The loss function calculation method of the extreme learning machine is expressed as: , In the formula, is the loss function of the extreme learning machine, are all the parameters of the extreme learning machine model, is the L2 norm, is the target output matrix, is the hidden layer output of the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine; is the regularization coefficient of the extreme learning machine, set to 0.2; is the dynamic penalty coefficient of the extreme learning machine, set to 0.3; is the penalty function of the extreme learning machine, is the building digital twin data matrix input to the extreme learning machine; In order to quantify the distribution offset of building digital twin data and combine it with the gradient changes in the training process, the dynamic penalty term in the loss function can be adaptively corrected. For example, the energy consumption of the HVAC system increases significantly in winter. The dynamic penalty term dynamically corrects the loss function by combining the distribution offset and gradient information, so that the model can adapt to the changes of different data streams and ensure the prediction accuracy of the building energy consumption assessment model in the case of sudden changes in data distribution. The calculation method of the building digital twin data distribution offset is expressed as: , In the formula, is the building digital twin data distribution offset, is the building digital twin data matrix input to the extreme learning machine; is the mean vector calculated based on the sliding window, and the step size of the sliding window is set to 3; Based on the combination of building digital twin data offset and gradient information, the dynamic penalty function is calculated, which is expressed as: , In the formula, is a dynamic penalty function; is the building digital twin data matrix input to the extreme learning machine; is the first penalty function coefficient, set to 0.3; Distribute offsets for building digital twin data; is the second penalty function coefficient, set to 0.3; is the gradient of the extreme learning machine’s loss function with respect to the input building digital twin data; is the third penalty function coefficient, Set to 0.4; is the gradient of the weighted feature vector to the hidden layer weight of the extreme learning machine, is the weighted feature vector; S304. In order to solve the feature redundancy and filter out the key features in the high-dimensional building digital twin data, a dynamic search strategy based on constraint sets is adopted to solve how to continuously focus on the most useful features for classification and eliminate irrelevant or interfering features during the training process. The initialized constraint set stores the correlation between each feature and the classification result. According to the classification error and feature contribution of each iteration, the constraint set is incrementally updated to make the model focus on the features with high contribution and dynamically eliminate the features that are not helpful or interfere with the classification. For example, short-term water consumption has little effect on long-term energy consumption evaluation. By dynamically searching and updating the constraint set, the model can focus on key features (such as peak power consumption and HVAC energy consumption) and eliminate features that are not helpful for classification, thereby improving the evaluation efficiency and accuracy. The calculation method is expressed as: , In the formula, is the constraint set, For collection update operations, is the hidden layer activation function, is the building digital twin data matrix input to the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine; In order to use the constraint set for the correlation between the feature and the output error, the dynamic constraint set is expanded according to the measurement of the correlation between the feature and the target output. The calculation method is expressed as: , In the formula, For the The correlation measure of each feature reflects the degree of association between each feature and the final classification result; is the gradient of the extreme learning machine’s loss function with respect to the input building digital twin data; is the regularization factor used to balance the influence of gradient information and building digital twin data distribution, Set to 0.3; is the mean vector calculated based on the sliding window; According to the feature correlation metric, the features that meet the threshold requirements are added to the constraint set. The calculation method is expressed as: , In the formula, For the The constraint set for the iteration, For the The constraint set for the iteration, For the Features, For the The correlation measure of the features, The threshold for filtering out highly correlated features is set to 0.3.

[0033] S305. Considering the influence of extreme learning machine to hidden layer parameters on classification performance and the joint effect of dynamic penalty terms, the hidden layer parameters are interactively updated to solve the problem of how to combine dynamic penalty terms with hidden layer weight updates, and maintain the robustness of the model in noisy and high-dimensional building digital twin data scenarios, further suppress overfitting and make the training process adaptive to the distribution of building digital twin data. By calculating the gradient of the loss function relative to the hidden layer weights and defining the adaptive gradient correction coefficient, the dynamic penalty term is integrated to update the hidden layer weights. The calculation method is expressed as: , In the formula, is the gradient of the extreme learning machine’s loss function with respect to the hidden layer weights, is the hidden layer output of the extreme learning machine, is the transpose of the output of the hidden layer of the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine, is the target output matrix; is the gradient regularization term coefficient, set to 0.3; is the weight matrix of the hidden layer of the extreme learning machine; In order to adaptively correct the update step size according to the gradient size during the training process, an adaptive gradient correction coefficient is used to adjust the step size of each gradient update. The calculation method is expressed as: , In the formula, is the adaptive correction coefficient; is the correction coefficient adjustment factor, set to 0.3; is the gradient of the extreme learning machine's loss function with respect to the hidden layer weights; Under the condition of integrating the dynamic penalty gradient, the model can dynamically adjust the hidden layer weight update process according to the gradient size through the adaptive gradient correction coefficient, thereby avoiding overfitting noise data while maintaining the ability to capture the actual energy consumption trend. The update method of the hidden layer weight of the extreme learning machine is expressed as: , In the formula, For the The hidden layer weights of the extreme learning machine of the iteration, For the The hidden layer weights of the extreme learning machine of the iteration, is the adaptive correction coefficient, is the gradient of the extreme learning machine’s loss function with respect to the hidden layer weights, is the dynamic penalty coefficient of the extreme learning machine weight, is the gradient of the dynamic penalty function with respect to the hidden layer weights.

[0034] S306. In order to ensure that each stage of training can achieve a stable convergence effect and prevent overfitting, the dynamic penalty term coefficient of the extreme learning machine weight is updated by adaptive parameters to solve how to maintain the rationality of the training step under different batches of building digital twin data, and control the complexity of the model so that the model can maintain good convergence and avoid oscillation in different periods, so that the model can be more robust in the later stage. The calculation method is expressed as: , In the formula, For the The dynamic penalty coefficient of the extreme learning machine weight of the iteration, For the The dynamic penalty coefficient of the extreme learning machine weights for the iteration; is the learning rate decay coefficient, set to 0.3; is the current iteration number; S307. In order to ensure that the loss function can better cope with the complexity of building digital twin data when using dynamic penalty terms and constraint sets, the loss function is dynamically adjusted. The dynamic adjustment term based on the gradient size is used so that the loss function can be flexibly adapted according to the current training progress and error state. The calculation method is expressed as: , In the formula, For the The loss function of the extreme learning machine at the iteration is, For the The loss function of the extreme learning machine at the iteration; is the dynamic adjustment coefficient of the loss function, set to 0.3; For the The gradient of the extreme learning machine's loss function with respect to all parameters at the iteration; S308. Repeat iterations S302 to S307 until a preset stop iteration condition is met, indicating that the model training is completed. The preset stop iteration condition is reaching a preset maximum number of iterations, which is set to 1000 times.

[0035] For the final fine-tuning of the model parameters after training, additional adjustment coefficients are used to correct the classification output. Based on the existing weights, the training error information is used to make fine-grained adjustments to the category output to improve the final classification accuracy and make the prediction results more reliable, which is expressed as: , In the formula, is the fine-tuning coefficient, is the target output matrix, is the hidden layer output of the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine; Use the fine-tuning coefficient to correct the final classification result. The calculation method is expressed as: , In the formula, Output the final classification results, including low energy consumption, medium energy consumption and high energy consumption; is the fine-tuning coefficient, is the hidden layer output of the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine.

[0036] like Figure 1 As shown in , by analyzing the influence of regularization parameters on classification accuracy, it is shown that the change of regularization parameters has a positive impact on classification accuracy, and the extreme learning machine based on adaptive constraints has a significant improvement in classification accuracy. Figure 2 As shown in Figure 2, by comparing the importance of different features to the model output, the weighted effects of different features in the extreme learning machine method based on adaptive constraints and the traditional extreme learning machine are different. The extreme learning machine method based on adaptive constraints has a higher weight in key feature identification. Figure 3 As shown in Figure 1, by analyzing the impact of dynamic penalty terms on model stability and showing the comparison of loss changes over time, it is shown that the extreme learning machine based on adaptive constraints has higher stability and faster convergence speed in dynamic data. Figure 4 As shown in the figure, by comparing the overall performance of the extreme learning machine based on adaptive constraints with other algorithms (such as support vector machine, traditional extreme learning machine, random forest, etc.), it is shown that the extreme learning machine based on adaptive constraints has higher classification accuracy.

[0037] S4. Each local building energy consumption assessment model sends the trained parameters to the central server, which aggregates the parameters and updates the central building energy consumption assessment model. The updated central building energy consumption assessment model is then sent to each building to form a new local building energy consumption assessment model. The new local building energy consumption assessment model repeats S3 to start a new iteration until the training converges, and the trained central building energy consumption assessment model and the trained local building energy consumption assessment model are obtained. The interaction between the local building energy consumption assessment model and the central building energy consumption assessment model is the update flow of the building energy consumption assessment model parameters. represents the building energy consumption assessment model parameters updated after local training, represents the parameters received from the central building energy consumption assessment model; The central building energy consumption assessment model aggregates updates from local building energy consumption assessment models. The update of the central building energy consumption assessment model is a way for the federated learning process to aggregate individual updates and then create a new improved central building energy consumption assessment model. After the central building energy consumption assessment model is updated, the final trained building energy consumption assessment model is obtained. Building energy consumption assessment model parameters arrive Exchange between the local building energy consumption assessment model and the central building energy consumption assessment model in a manner that allows the central building energy consumption assessment model to be aggregated and updated, and for the local building energy consumption assessment model to receive the new, aggregated parameters; Through multiple iterations, the local building energy consumption assessment model is trained, the updates are sent to the central building energy consumption assessment model, a global update is performed, and then the updated parameters are sent back to the local building energy consumption assessment model; Based on this federated learning architecture, the privacy protection of building digital twin data can be achieved. During the training process, the original building digital twin data is not shared between clients or with the central server, but only the building energy consumption assessment model parameters are exchanged for update. That is, the parameters are synchronized to each client through the central building energy consumption assessment model. Assign a value to each client.

[0038] S5. Obtain new real building digital twin data for each building in the building complex, input the new real building digital twin data of each building into the corresponding trained local building energy consumption assessment model, and obtain the energy consumption level corresponding to the new real building digital twin data of each building.

Claims

1. A method for processing building digital twin data based on artificial intelligence, characterized in that: include: S1. Obtain the real building digital twin data of each building in the building complex, and mark the energy consumption level corresponding to the real building digital twin data of each building; S2. According to the real building digital twin data of each of the buildings, a data expansion model is trained using a generative adversarial network algorithm based on topological dual optimization to obtain a trained data expansion model. The trained data expansion model of each of the buildings generates synthetic building digital twin data of the building according to the real building digital twin data of the building, and the real building digital twin data of each of the buildings and the synthetic building digital twin data are used as training data of the building; S3. construct and initialize a central building energy consumption assessment model on a central server using a distributed federated learning method, send the central building energy consumption assessment model to each of the buildings to form a local building energy consumption assessment model, and use the training data of each building to train the local building energy consumption assessment model of each building using an extreme learning machine algorithm based on adaptive constraints to obtain parameters of each local building energy consumption assessment model; S4. Each of the local building energy consumption assessment models sends the parameters obtained through training to the central server, and the central server aggregates the parameters and updates the central building energy consumption assessment model, and then sends the updated central building energy consumption assessment model to each of the buildings to form a new local building energy consumption assessment model; the new local building energy consumption assessment model repeats S3 to start a new iteration until the training converges, and obtains a trained central building energy consumption assessment model and a trained local building energy consumption assessment model; S5. Obtain new real building digital twin data for each building in the building complex, input the new real building digital twin data of each building into the corresponding trained local building energy consumption assessment model, and obtain the energy consumption level corresponding to the new real building digital twin data of each building.

2. The method for processing building digital twin data based on artificial intelligence according to claim 1 is characterized in that: The method of generating synthetic building digital twin data of each building by using a generative adversarial network algorithm based on topological dual optimization in S2 includes: S201. Initialize the generative adversarial network; S202. When training the generator, a random noise vector is received and samples with a distribution similar to that of the real building digital twin data are generated. The generator adopts topological dual optimization in the loss function, and the calculation method of the generator loss is expressed as: , where is the loss function of the generator, is the expected value of the input noise, is the discriminator function, is the generator function, Output building digital twin data to the generator, is the input noise, is the weight parameter of the generator, is the weight coefficient, is the topological dual loss function; S203. After the generator outputs the building digital twin data, the discriminator receives the real building digital twin data and the synthetic building digital twin data, and continuously improves the discrimination ability by distinguishing the two. The Riemann manifold regularization method is used in the training process of the discriminator, and the calculation method of the discriminator loss is expressed as: ， In the formula, is the discriminator loss function, is the expected value of the real building digital twin data, is the discriminator function, For real building digital twin data, is the expected value of the input noise, Output building digital twin data to the generator, is the weight coefficient of Riemann manifold regularization; is the Riemann manifold regularization term; S204. It is necessary to iteratively update the parameters of the generator and discriminator based on the adversarial loss and its additional topological and Riemann manifold constraints. The parameter update method of the generator and discriminator is expressed as: ， ， In the formula, For the The generator parameters for the iterations, For the The generator parameters for the iterations, is the learning rate of the generator, is the gradient of the generator’s loss function with respect to its parameters, For the The discriminator parameters of the iteration, For the The discriminator parameters of the iteration, is the learning rate of the discriminator, is the gradient of the discriminator's loss function with respect to its parameters; S205. Repeat the above steps until a preset stop iteration condition is met.

3. The method for processing building digital twin data based on artificial intelligence according to claim 2 is characterized in that: The topological dual loss function in S202 The calculation method is expressed as: ， In the formula, To constrain the topological distance of the overall distribution of real building digital twin data, To map the synthetic building digital twin data in the topological space, It is the mapping of real building digital twin data in topological space. For real building digital twin data, is the probability density of real building digital twin data.

4. The method for processing building digital twin data based on artificial intelligence according to claim 3 is characterized in that: Mapping of the synthetic building digital twin data in topological space Mapping of the real building digital twin data in the topological space The calculation method is expressed as: ， ， In the formula, To map the digital twin data of the synthetic building in the topological space; is a mapping function, specifically a variational autoencoder function; is the generator function, is the input noise, is the weight parameter of the generator; are the parameters of the mapping model, specifically the parameters of the variational autoencoder; It is the mapping of real building digital twin data in topological space. It is the digital twin data of real buildings.

5. The method for processing building digital twin data based on artificial intelligence according to claim 2 is characterized in that: The Riemann manifold regularization term in S203 The calculation method is expressed as: ， In the formula, is the Riemann manifold regularization term, For input and expectations, is the gradient of the discriminator with respect to the input building digital twin data; It is the Hessian matrix of the building digital twin data on the manifold, which describes the local curvature information of the building digital twin data manifold.

6. The method for processing building digital twin data based on artificial intelligence according to claim 1 is characterized in that: The method of using the extreme learning machine algorithm based on adaptive constraints to train the local building energy consumption assessment model of each building in S3 includes: S301. Initialize the extreme learning machine model; S302. Adopt the dynamic feature weighting adjustment method of adaptive constraint optimization to dynamically weight each feature and adjust its influence on the final classification result. Define the weighting coefficient to represent the importance of each feature in the final classification. The calculation method is expressed as: ， In the formula, For the The weighting coefficient of each feature; is the attenuation factor; For the Features, For all building digital twin data input to the extreme learning machine The mean of the features represents the position of the feature in the building digital twin data; is the weight factor of the gradient information, balancing the influence of the gradient size on the weight coefficient; is the loss function of the extreme learning machine with respect to the The gradient of a feature; S303. A dynamic penalty term is used to solve the constraint problem of the loss function in the extreme learning training process. The weight of the extreme learning machine is adaptively adjusted when facing non-uniform or dynamically changing building digital twin data streams. The loss function calculation method of the extreme learning machine is expressed as: ， In the formula, is the loss function of the extreme learning machine, are all the parameters of the extreme learning machine model, is the L2 norm, is the target output matrix, is the hidden layer output of the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine; is the regularization coefficient of the extreme learning machine; is the dynamic penalty coefficient of the extreme learning machine; is the penalty function of the extreme learning machine, is the building digital twin data matrix input to the extreme learning machine; S304. A dynamic search strategy based on a constraint set is adopted. The initialized constraint set stores the correlation between each feature and the classification result. The constraint set is incrementally updated according to the classification error and feature contribution of each iteration, so that the model focuses on features with high contribution and dynamically removes features that are not beneficial to classification or bring interference. The calculation method is expressed as: ， In the formula, is the constraint set, For collection update operations, is the hidden layer activation function, is the building digital twin data matrix input to the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine; S305. Considering the influence of the extreme learning machine to the hidden layer parameters on the classification performance and the joint effect of the dynamic penalty term, the hidden layer parameters are interactively updated, and the calculation method is expressed as: ， In the formula, is the gradient of the extreme learning machine’s loss function with respect to the hidden layer weights, is the hidden layer output of the extreme learning machine, is the transpose of the output of the hidden layer of the extreme learning machine, is the weight matrix from the hidden layer to the output layer of the extreme learning machine, is the target output matrix; is the gradient regularization term coefficient; is the weight matrix of the hidden layer of the extreme learning machine; S306. Adopt adaptive parameters to update the dynamic penalty term coefficient of the extreme learning machine weight and perform model complexity control. The calculation method is expressed as: ， In the formula, For the The dynamic penalty coefficient of the extreme learning machine weight of the iteration, For the The dynamic penalty coefficient of the extreme learning machine weights for the iteration; is the learning rate decay coefficient; is the current iteration number; S307. Dynamically adjust the loss function, using a dynamic adjustment item based on the gradient size so that the loss function can be flexibly adapted according to the current training progress and error state. The calculation method is expressed as: ， In the formula, For the The loss function of the extreme learning machine at the iteration is, For the The loss function of the extreme learning machine at the iteration; is the dynamic adjustment coefficient of the loss function; For the The gradient of the extreme learning machine's loss function with respect to all parameters at the iteration; S308. Repeat iterations S302 to S307 until a preset stop iteration condition is met.

7. The method for processing building digital twin data based on artificial intelligence according to claim 6 is characterized in that: Initializing the extreme learning machine model in S301 includes: Initialize the parameters of the extreme learning machine, which include the weight and bias parameters of the extreme learning machine. The initialization method is random initialization and the initialized parameters obey the normal distribution with a mean of 0 and a variance of the unit matrix. The activation method of the hidden layer activation function of the extreme learning machine is defined so that the building digital twin data processing method can provide good expression ability for the nonlinear characteristic relationship of multi-dimensional building digital twin data. The calculation method is expressed as: ， In the formula, is the hidden layer output of the extreme learning machine; is the hidden layer activation function, and the hidden layer activation function adopts the Sigmoid activation function; is the building digital twin data matrix input to the extreme learning machine, is the weight matrix of the hidden layer of the extreme learning machine, is the hidden layer bias of the extreme learning machine.

Images