A BIM-based energy consumption prediction monitoring system

Abstract

The application discloses a kind of energy consumption prediction monitoring systems based on BIM, it is related to intelligent building energy management technical field, including based on modal weight, definition optimization objective function, using OIALM algorithm is iterated and updated, obtain low rank matrix, through prediction model output initial predicted energy consumption, using golden jackal optimization algorithm is optimized to model parameter, generate final predicted energy consumption;Through three modal alignment and sparse inhibition tensor modeling method based on BIM and internet of things data, effectively fuse building physical semantics and dynamic environment characteristics, improve the accuracy and robustness of energy consumption feature expression, construct the prediction framework of fusion key time node rearrangement, attention-enhanced Tucker decomposition and OIALM low rank optimization, improve the modeling ability and energy consumption prediction accuracy to non-steady thermal process.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent building energy management, and particularly relates to a BIM-based energy consumption prediction monitoring system. BACKGROUND

[0002] The BIM technology can not only provide geometric topological information of building components, but also provide thermal performance, equipment parameters and operating conditions, and is an important data source for realizing fine analysis of building energy consumption. With the development of the Internet of Things technology, external meteorological environment data can be collected through high-frequency collection of multiple types of sensors and real-time data platforms to form a dynamic environment input sequence, so that building energy consumption modeling gradually evolves from a static empirical model to a data-driven model, and the timeliness and fine granularity of building energy consumption simulation are improved.

[0003] However, the prior art still has the following problems: the prior art directly inputs BIM geometry, thermal parameters and environment data into a model, lacks a time-space alignment mechanism, the physical properties of components cannot accurately correspond to the time dimension and real-time environment data, the prediction features have inter-modal asynchrony, the prior BIM energy consumption prediction lacks mining of low-rank structures, and the prior BIM energy consumption prediction is not optimized by combining an OIALM optimization algorithm, so that the model is easily affected by noise and has weak generalizability, and key dynamic phenomena such as power consumption peaks and wall heat storage cannot be accurately described. SUMMARY

[0004] In view of the above problems existing in the prior art, the present application is provided.

[0005] Therefore, the present application provides a BIM-based energy consumption prediction monitoring system, which solves the problems that the prior art still has the following problems: the prior art directly inputs BIM geometry, thermal parameters and environment data into a model, lacks a time-space alignment mechanism, the physical properties of components cannot accurately correspond to the time dimension and real-time environment data, the prediction features have inter-modal asynchrony, the prior BIM energy consumption prediction lacks mining of low-rank structures, and the prior BIM energy consumption prediction is not optimized by combining an OIALM optimization algorithm, so that the model is easily affected by noise and has weak generalizability, and key dynamic phenomena such as power consumption peaks and wall heat storage cannot be accurately described.

[0006] To solve the above technical problems, the present application provides the following technical scheme:

[0007] The present application provides a BIM-based energy consumption prediction monitoring system, which includes,

[0008] The unified space-time matrix and dynamic energy consumption driving sequence construction module is used for obtaining the BIM model through an API interface and obtaining real-time meteorological data through an Internet of Things platform, respectively pre-processing the BIM model and the Internet of Things platform, forming a unified space-time matrix, calculating component transient heat flux energy, room transient load and HVAC equivalent load based on the unified space-time matrix, and obtaining a dynamic energy consumption driving sequence;

[0009] The sparse inhibition and smooth tensor generation module is used for constructing a three-modal feature tensor based on the structured energy consumption parameter matrix, the extended environment sequence and the dynamic energy consumption driving sequence, calculating an energy value for each feature dimension based on the three-modal feature tensor, obtaining a sparse inhibition tensor through screening, performing linear interpolation on the time mode in the sparse inhibition tensor to obtain a local gradient, constructing an optimal node sequence, rearranging the time sequence based on the optimal node sequence to obtain a rearranged sparse inhibition tensor, adding random disturbance to each time point to generate a smooth signal, and obtaining a smooth sparse inhibition tensor.

[0010] The attention-enhanced feature reconstruction module is used for calculating a first-order gradient for the building physical feature mode in the three-modal feature tensor based on the smooth sparse inhibition tensor, generating a column uniformization matrix, and performing SVD decomposition to obtain a factor matrix, obtaining a core tensor through Tucker inverse mapping, extracting a row vector based on the factor matrix and calculating an attention weight, constructing a diagonal weighting matrix to weight and reconstruct the core tensor along the time direction, obtaining a final enhanced feature through Tucker reconstruction, and constructing a modal weight.

[0011] The low-rank optimization and energy consumption prediction output module is used for defining an optimization objective function based on the modal weight, performing iterative updating using an OIALM algorithm to obtain a low-rank matrix, outputting an initial predicted energy consumption through a prediction model, optimizing model parameters using a golden jackal optimization algorithm to generate a final predicted energy consumption.

[0012] As a preferred scheme of the BIM-based energy consumption prediction and monitoring system, wherein: the rearranging the time sequence based on the optimal node sequence to obtain the rearranged sparse inhibition tensor comprises:

[0013] The structured energy consumption parameter matrix, the extended environment sequence and the dynamic energy consumption driving sequence are horizontally spliced in the feature dimension to obtain a unified splicing tensor, the feature dimension of the unified splicing tensor is divided to obtain a three-modal feature tensor, an energy value is calculated for each feature dimension of the three-modal feature tensor, if the energy value is less than an energy value threshold, the corresponding feature dimension is deleted to obtain a sparse inhibition tensor, a local gradient is calculated and sorted in descending order, the time point corresponding to the maximum local gradient is selected to obtain an optimal node sequence.

[0014] The time sequence in the sparse inhibition tensor is rearranged based on the optimal node sequence to obtain a rearranged sparse inhibition tensor, a random disturbance is added to each time point in the rearranged sparse inhibition tensor, a smoothed signal is calculated, the smoothed signal is replaced with a row in a development matrix, and a smoothed sparse inhibition tensor is obtained.

[0015] As a preferred scheme of the BIM-based energy consumption prediction monitoring system, the generation of the column uniformization matrix and the SVD decomposition to obtain the factor matrix comprises:

[0016] Based on the smoothed sparse inhibition tensor, a first-order gradient is calculated for the building physical feature mode to generate a column uniformization matrix, the uniformization matrix is subjected to SVD decomposition to obtain a time mode factor matrix, and the three-mode tensor is subjected to mode-2 and mode-3 development to obtain a building physical feature matrix and an environment and operation feature matrix, and the two matrices are subjected to SVD decomposition respectively to obtain a building physical feature factor matrix and an environment and operation feature factor matrix.

[0017] As a preferred scheme of the BIM-based energy consumption prediction monitoring system, the construction of the diagonal weight matrix for the core tensor in the time direction comprises:

[0018] The time mode factor matrix, the building physical feature factor matrix, and the environment and operation feature factor matrix are subjected to Tucker inverse mapping to obtain a core tensor.

[0019] The row vector corresponding to each time point is extracted from the time mode factor matrix, the attention weight is calculated, the diagonal weight matrix is constructed by constructing a diagonal matrix, the core tensor is subjected to weighted reconstruction in the time direction to obtain an attention-enhanced core tensor, and the attention-enhanced core tensor is subjected to Tucker reconstruction to obtain the final enhanced feature.

[0020] As a preferred scheme of the BIM-based energy consumption prediction monitoring system, the construction of the modal weight based on the final enhanced feature comprises:

[0021] Based on the time mode, the building physical feature mode, and the environment and operation feature mode in the final enhanced feature, matrix development is performed respectively to obtain a modal matrix, the energy intensity of each modal matrix is calculated, and the modal weight is constructed.

[0022] As a preferred scheme of the BIM-based energy consumption prediction monitoring system, the definition of the optimization objective function and the iterative update using the OIALM algorithm to obtain a low-rank matrix comprise:

[0023] Based on the modal weight, an optimization objective function is defined, a modal matrix is subjected to a mode folding operation, an initialization matrix is obtained, and the initialization matrix is iteratively updated through an OIALM algorithm to obtain a low-rank matrix.

[0024] As a preferred scheme of the BIM-based energy consumption prediction monitoring system, the initial predicted energy consumption is output through the prediction model, including:

[0025] The low-rank matrix is subjected to a mode folding operation to obtain a stable time sequence, the sequence is right-padded by rows to obtain a TCN input sequence, the TCN network is input, a deep feature vector is obtained, the deep feature vector is input into a bidirectional GRU to obtain a hidden state, an importance score of each time point is obtained through an attention mechanism, an attention weight of the importance score is calculated through a Softmax function, the importance score and the attention weight are weighted and summed to obtain a context vector, and the context vector is input into a fully connected output layer to obtain the initial predicted energy consumption.

[0026] As a preferred scheme of the BIM-based energy consumption prediction monitoring system, the model parameters are optimized using a golden wolf optimization algorithm to generate a final predicted energy consumption, including:

[0027] The initialization of the to-be-optimized hyperparameter space includes the convolution kernel size, the expansion rate and the regularization coefficient of the TCN, the number of hidden units and the number of layers of the bidirectional GRU, and the key vector dimension of the attention mechanism, the golden wolf optimization algorithm is used to iteratively update the to-be-optimized hyperparameter space to obtain an optimal hyperparameter space, and the final predicted energy consumption is generated.

[0028] As a preferred scheme of the BIM-based energy consumption prediction monitoring system, the BIM model is obtained through an API interface and the real-time meteorological data is obtained through an Internet of Things platform, and the BIM model and the Internet of Things platform are preprocessed, including:

[0029] The BIM model is obtained through an API interface and the real-time meteorological data is obtained through an Internet of Things platform, and the BIM model and the Internet of Things platform are preprocessed to obtain a structured energy consumption parameter matrix and an extended environment sequence.

[0030] As a preferred scheme of the BIM-based energy consumption prediction monitoring system, the dynamic energy consumption driving sequence is obtained by calculating the transient heat flux energy of the component, the transient load of the room and the equivalent load of the HVAC, including:

[0031] The structured energy consumption parameter matrix and the extended environment sequence are spliced to obtain an alignment matrix, and all alignment matrices within a time t are arranged in chronological order The dimensions are stacked to obtain a unified space-time matrix, a dynamic energy consumption driving sequence is constructed, including calculating component transient heat flux energy, room transient load and HVAC equivalent load, combining solar heat, performing normalization processing and vector transverse splicing operation, and obtaining the dynamic energy consumption driving sequence.

[0032] The application has the advantages that: the application defines an optimization objective function based on modal weights, iteratively updates using an OIALM algorithm to obtain a low-rank matrix, outputs an initial predicted energy consumption through a prediction model, optimizes model parameters using a golden wolf optimization algorithm, and generates a final predicted energy consumption; effectively fuses building physical semantics and dynamic environmental characteristics, improves the accuracy and robustness of energy consumption feature expression, and improves the modeling capability of non-steady-state thermal processes and energy consumption prediction accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0033] In order to more clearly illustrate the technical solutions of the embodiments of the application, the following will briefly introduce the drawings needed to be used in the embodiment description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor on the basis of these drawings.

[0034] Figure 1 The flowchart of the BIM-based energy consumption prediction monitoring system in embodiment 1.

[0035] Figure 2 The structural diagram of the BIM-based energy consumption prediction monitoring system in embodiment 1. DETAILED DESCRIPTION

[0036] In order to make the above-mentioned purposes, features and advantages of the application more obvious and easy to understand, the specific embodiments of the application will be described in detail below with reference to the drawings of the specification.

[0037] In the following description, many specific details are set forth in order to provide a thorough understanding of the application, but the application can also be implemented in other ways different from those described herein, and those skilled in the art can make similar generalizations without departing from the connotation of the application, therefore the application is not limited to the specific embodiments disclosed below.

[0038] Secondly, the "one embodiment" or "embodiment" referred to herein means that the specific features, structures or characteristics can be included in at least one implementation of the application. "In one embodiment" appearing in different places in the specification does not mean the same embodiment, nor is it an independent or alternative embodiment that excludes other embodiments.

[0039] Embodiment 1, refer to Figure 1 and Figure 2For the first embodiment of the present application, the embodiment provides a BIM-based energy consumption prediction monitoring system, comprising:

[0040] S1, a unified space-time matrix and dynamic energy consumption driving sequence construction module, configured to obtain a BIM model through an API interface and obtain real-time meteorological data through an Internet of Things platform, preprocess the BIM model and the Internet of Things platform respectively, form a unified space-time matrix, calculate component transient heat flux energy, room transient load and HVAC equivalent load based on the unified space-time matrix, and obtain a dynamic energy consumption driving sequence;

[0041] Specifically, the BIM model is obtained through the API interface and the real-time meteorological data is obtained through the Internet of Things platform, and the BIM model and the Internet of Things platform are preprocessed respectively, including:

[0042] The BIM model is obtained through the API interface and the real-time meteorological data is obtained through the Internet of Things platform, and the BIM model and the Internet of Things platform are preprocessed respectively, to obtain a structured energy consumption parameter matrix and an extended environment sequence;

[0043] The preprocessing of the BIM model includes extracting the space classification in the BIM model, such as the space to which the building, floor, room and component (such as wall, roof, floor, window and door) belong, and arranging the hierarchical sequence horizontally;

[0044] Based on the hierarchical sequence, the geometric parameters (synchronous data acquisition when obtaining the BIM model data through the API) of each component are extracted, such as area, thickness, orientation and inclination, the thermal parameters of each component are extracted, such as thermal conductivity, specific heat capacity, total thermal resistance, density and solar heat gain, and the HVAC (such as VRV, fan coil and fresh air unit) equipment parameters of the space to which each component belongs are extracted, such as rated power, air volume, supply air temperature and COP value;

[0045] The geometric parameters, thermal parameters and equipment parameters are constructed into a structured energy consumption parameter matrix, and the formula is:

[0046] ,

[0047] Among them, and are data in the geometric parameters, thermal parameters and equipment parameters of the i-th component, is the number of components;

[0048] The real-time meteorological data, such as temperature, humidity, wind speed, radiation and cloud cover;

[0049] ​The preprocessing of the real-time meteorological data includes filling in missing values of the real-time meteorological data by using bidirectional spline interpolation, and extracting meteorological variables such as radiation and wind speed from the filled real-time meteorological data, normalizing the meteorological components to obtain preprocessed real-time meteorological data;

[0050] Based on the preprocessed real-time meteorological data, the temperature difference between time t and time t-1 is calculated to obtain the temperature change rate, the radiation difference between time t and time t-1 is calculated to obtain the radiation increment, and the radiation increment in time t is accumulated to obtain the cumulative amount of sunshine;

[0051] The cumulative amount of sunshine, the temperature change rate, the radiation variable and the preprocessed real-time meteorological data are arranged horizontally to obtain an extended environment sequence.

[0052] By performing independent preprocessing on the BIM model and real-time meteorological data and constructing a structured energy consumption parameter matrix and an extended environment sequence, deep coupling of building static properties and external dynamic environment is realized. The extraction of space levels, component geometry, thermal properties and HVAC equipment performance in the preprocessing process can significantly enhance the expression ability of the model to the multi-source heterogeneity of building energy systems, so that the energy consumption characteristics present clear physical meaning in the input stage. By using bidirectional interpolation, normalization and increment calculation on meteorological data, abnormal fluctuations caused by sensor noise can be significantly reduced, and the environment input has stronger time sequence sensitivity through the construction of the cumulative amount of sunshine and the temperature change rate. The structured parameter matrix and the extended environment sequence formed ultimately lay a stable and highly interpretable data foundation for subsequent sequence energy consumption deduction, realize the unified expression of building static properties and external dynamic factors, and significantly improve the discrimination accuracy of the subsequent prediction model.

[0053] Further, the component transient heat flux energy, the room transient load and the HVAC equivalent load are calculated to obtain a dynamic energy consumption driving sequence, including:

[0054] The structured energy consumption parameter matrix and the extended environment sequence are spliced to obtain an alignment matrix, and the formula is:

[0055] ,

[0056] Wherein, is the alignment matrix, is the extended environment sequence, is the structured energy consumption parameter matrix;

[0057] All alignment matrices in time t are stacked according to the time dimension to obtain a unified space-time matrix;

[0058] Based on the unified space-time matrix, a dynamic energy consumption driving sequence is constructed, including calculation of component transient heat flux energy, room transient load, and HVAC equivalent load, combined with solar heat, normalized and vector transverse splicing operation to obtain the dynamic energy consumption driving sequence, the formula is:

[0059] ,

[0060] wherein, is the component transient heat flux energy, is the component transient heat flux energy of the i-th building component (such as wall), is the time, is the outdoor temperature, is the component temperature, is the total thermal resistance of the component, is the solar heat of the i-th light-transmitting component (such as window), is the room transient load, is the HVAC equivalent load, is the equipment energy efficiency ratio, which is obtained from the HVAC equipment parameters, and are the number of building components and the number of light-transmitting components. By constructing the dynamic energy consumption driving sequence, the transient thermal behavior of the building component, the room thermal load and the HVAC operation response are integrated in the form of a unified space-time matrix, which has significant systematic value. On the one hand, the structured energy consumption parameter matrix and the extended environment sequence are spliced by aligning the matrix, realizing the precise alignment of the building physical characteristics and the external disturbance in the time dimension, effectively eliminating the time offset problem between modes. On the other hand, by dynamically deriving the transient heat flux energy, the transient load and the equivalent HVAC load, the energy consumption driving sequence has a clear physical causal logic, especially in the case of high temperature sudden rise, wind speed mutation and strong radiation, which can present real thermal inertia effect and equipment response law. The unified driving sequence after vector splicing has both continuity and physical consistency, providing a deep dynamic driving force expression for energy consumption prediction, significantly improving the sensitivity and generalizability of the model to cold and heat load fluctuations.

[0061] By constructing the dynamic energy consumption driving sequence, the transient thermal behavior of the building component, the room thermal load and the HVAC operation response are integrated in the form of a unified space-time matrix, which has significant systematic value. On the one hand, the structured energy consumption parameter matrix and the extended environment sequence are spliced by aligning the matrix, realizing the precise alignment of the building physical characteristics and the external disturbance in the time dimension, effectively eliminating the time offset problem between modes. On the other hand, by dynamically deriving the transient heat flux energy, the transient load and the equivalent HVAC load, the energy consumption driving sequence has a clear physical causal logic, especially in the case of high temperature sudden rise, wind speed mutation and strong radiation, which can present real thermal inertia effect and equipment response law. The unified driving sequence after vector splicing has both continuity and physical consistency, providing a deep dynamic driving force expression for energy consumption prediction, significantly improving the sensitivity and generalizability of the model to cold and heat load fluctuations.

[0062] ​S2, the sparse suppression and smoothing tensor generation module, is used to construct a three-modal feature tensor based on the structured energy consumption parameter matrix, the extended environment sequence, and the dynamic energy consumption driving sequence. Based on the three-modal feature tensor, the energy value is calculated for each feature dimension. The sparse suppression tensor is obtained by filtering. The time modes in the sparse suppression tensor are linearly interpolated to obtain the local gradient. The optimal node sequence is constructed. The time order is rearranged based on the optimal node sequence to obtain the rearranged sparse suppression tensor. Random perturbation is added to each time point to generate a smooth signal and obtain the smoothed sparse suppression tensor.

[0063] Specifically, the temporal order is rearranged based on the optimal node sequence to obtain the rearranged sparse suppression tensor, including:

[0064] The structured energy consumption parameter matrix (with time-broadcast extension), extended environmental sequence, and dynamic energy consumption driving sequence are normalized respectively. The normalized data are then horizontally concatenated along the feature dimensions to obtain a unified concatenated tensor. The feature dimensions of the unified concatenated tensor are then divided to obtain a three-modal feature tensor, including time mode, building physical feature mode (e.g., wall thermal resistance, window-to-wall ratio, equipment rated power, etc.), and environmental and operational feature mode (e.g., external temperature, humidity, solar radiation, operating conditions).

[0065] Calculate the energy value for each feature dimension of the trimodal feature tensor. If the energy value is less than the energy value threshold (set based on empirical rules), delete the corresponding feature dimension to obtain the sparse suppression tensor, as shown in the formula:

[0066] ,

[0067] in, For the first Energy values ​​for each feature dimension For the number of time points, For the first Each feature dimension;

[0068] Linear interpolation is performed on the time modes in the sparse suppression tensor to obtain the time mode interpolation. The difference between the time mode interpolation at time t and time t-1 is calculated to obtain the local gradient.

[0069] Divide the timestamps into equal intervals to obtain R equal intervals. Sort the local gradients of each equal time interval in descending order, select the time point corresponding to the maximum local gradient, set it as the optimal node, and arrange the optimal nodes horizontally to obtain the optimal node sequence.

[0070] The time sequence in the sparse inhibition tensor is rearranged based on the optimal node sequence (the sparse inhibition tensor is arranged in time t, and the rearranged sparse inhibition tensor is arranged in the time point corresponding to the optimal node sequence) to obtain a rearranged sparse inhibition tensor;

[0071] A random disturbance is added to each time point in the rearranged sparse inhibition tensor to calculate a smoothed signal, and the formula is:

[0072] ,

[0073] wherein, is the smoothed signal, is an expected value, is the rearranged sparse inhibition tensor, is a random disturbance;

[0074] The smoothed signal is replaced with a row in the unfolding matrix to obtain a smoothed unfolding matrix, that is, a smoothed sparse inhibition tensor is obtained.

[0075] The unfolding matrix refers to mode-1 unfolding (indicating unfolding along the time mode) of the rearranged sparse inhibition tensor to obtain the unfolding matrix.

[0076] By constructing the sparse inhibition tensor and the optimal node sequence, the key problems of multi-modal feature dimension redundancy, time sequence information noise and dynamic mutation point difficult to capture are solved. Dimension sparsification is realized through energy value screening, which can effectively eliminate low-contribution features, so that the model has higher computational efficiency while maintaining expressiveness. The optimal node is screened using local gradient, which can accurately identify key disturbance points in long time sequence, such as heat load sudden rise time, wind speed mutation point or radiation steep rise point, so that the sequence after time rearrangement presents a more structured dynamic distribution. The smoothed signal is generated by random disturbance, which can retain dynamic change characteristics while suppressing environmental noise, thereby constructing a stable sparse inhibition tensor. This process effectively improves the structural clarity of time sequence expression, enabling subsequent tensor decomposition to run in a purer space, significantly enhancing the model's ability to capture mutation behavior.

[0077] S3, an attention-enhanced feature reconstruction module, is configured to calculate a first-order gradient of a building physical feature mode in a three-modal feature tensor based on the smoothed sparse inhibition tensor, generate a column uniformization matrix, perform SVD decomposition to obtain a factor matrix, obtain a core tensor through Tucker inverse mapping, extract a row vector based on the factor matrix and calculate an attention weight, construct a diagonal weighting matrix to weight and reconstruct the core tensor along the time direction, obtain a final enhanced feature through Tucker reconstruction, and construct a modal weight.

[0078] Specifically, the column uniformization matrix is generated, and the factor matrix is obtained by performing SVD decomposition, including:

[0079] Based on the smooth sparse inhibition tensor, the first-order gradient of the building physical feature mode is calculated, and a column consistent matrix is generated, and the formula is:

[0080]

[0081] Among them, is the first-order gradient, is the smooth sparse inhibition tensor, is the element of the column consistent matrix, is the smooth sparse inhibition tensor, is the time index;

[0082] The SVD decomposition is performed on the consistent matrix to obtain the time mode factor matrix, and the mode-2 and mode-3 expansion is performed on the three-mode tensor (building physical feature mode and environment and operation feature mode) to obtain the building physical feature matrix and the environment and operation feature matrix. The SVD decomposition is performed on the two matrices respectively to obtain the building physical feature factor matrix and the environment and operation feature factor matrix.

[0083] By generating the column consistent matrix and performing singular value decomposition, the highly representative latent factor structure can be extracted from the multi-modal tensor, and the building physical feature, the environment and operation feature and the time feature are deeply decoupled. By performing consistent processing on the first-order gradient of the building physical feature mode, the differences of different components in feature changes are uniformly mapped, thereby reducing the distribution bias problem of the structure feature. By performing singular value decomposition on the three modes respectively, the factor matrix with low rank structure can be obtained, so that the building attributes, time changes and environmental disturbances can be represented in a compact manner. This decomposition mechanism not only improves the stability of data representation, but also provides a standardized basis for subsequent Tucker reconstruction, so that the building operation feature can be expressed independently according to the mode and maintain the overall correlation, which significantly improves the structured explainability of the overall model.

[0084] Further, a diagonal weighting matrix is constructed to weight and reconstruct the core tensor along the time direction, and the final enhanced feature is obtained by Tucker reconstruction, including:

[0085] The Tucker inverse mapping is performed on the time mode factor matrix, the building physical feature factor matrix and the environment and operation feature factor matrix to obtain the core tensor;

[0086] The row vector corresponding to each time point is extracted from the time mode factor matrix, and the attention weight is calculated, and the formula is:

[0087]

[0088] Among them, is the significance score of the time ,​​ is an attention weight vector, is a transpose, is a time is a row vector, is a time is a saliency score, is a time is an attention weight;

[0089] Based on the attention weight, a diagonal weighting matrix is constructed by constructing an operation on a diagonal matrix, and the core tensor is weighted and reconstructed in the time direction to obtain an attention-enhanced core tensor, and the formula is:

[0090] ,

[0091] wherein, is an attention-enhanced core tensor, is a core tensor, indicates a matrix-tensor multiplication along the time modal (1st modal) direction, is a diagonal weighting matrix;

[0092] The attention-enhanced core tensor is Tucker reconstructed to obtain a final enhanced feature.

[0093] By constructing a diagonal weighting matrix and performing Tucker reconstruction, the time feature saliency is emphasized, and the core tensor obtains a dynamic attention enhancement effect. By extracting the saliency score from the time factor, the contribution of the key moment in the energy consumption change can be effectively highlighted, so that the model has stronger perception ability in dealing with key stages such as peak load, cold and heat source switching, and strong radiation disturbance. The weighted core tensor has higher discrimination ability in the time dimension, so that the final enhanced feature after Tucker reconstruction exhibits significant time sequence sensitivity and more sufficient modal expression ability. This process strengthens the dynamic coupling between the three modalities, so that the enhanced feature has a more distinct pattern structure and time recognition degree, providing high-quality input for subsequent optimization and prediction, and effectively improving the robustness of energy consumption prediction.

[0094] Further, based on the final enhanced feature, a modal weight is constructed, including:

[0095] The time modal, building physical feature modal, and environment and operation feature modal in the final enhanced feature are extracted, respectively, matrix expansion is performed to obtain a modal matrix, the energy intensity of each modal matrix is calculated, and a modal weight is constructed, and the formula is:

[0096] ,

[0097] wherein, is the energy intensity of the modal , modal matrix of modal , Frobenius norm, modal weight of modal , energy value of modal , modal number.

[0098] By calculating the modal weight of the final enhanced features, the importance of each modal is quantitatively expressed, providing a target-oriented weighted basis for subsequent low-rank optimization. By calculating the energy intensity of the time modal, building physical modal and environmental operation modal respectively, the contribution degree of each modal to energy consumption change can be accurately reflected. The introduction of modal weight can strengthen the key modal, so that the low-rank decomposition process is more focused on the decisive features under the guidance of weight, while the weak contribution modal is naturally weakened, thereby improving the efficiency and accuracy of the overall decomposition. This mechanism can significantly improve the structural stability of the model, making the further OIALM optimization step more targeted and robust.

[0099] S4, a low-rank optimization and energy consumption prediction output module, is configured to define an optimization objective function based on the modal weight, perform iterative updates using an OIALM algorithm, obtain a low-rank matrix, output an initial predicted energy consumption through a prediction model, optimize the model parameters using a golden jackal optimization algorithm, and generate a final predicted energy consumption.

[0100] Specifically, the optimization objective function is defined, and the OIALM algorithm is used for iterative updates to obtain a low-rank matrix, including:

[0101] Based on the modal weight, the optimization objective function is defined, the modal matrix is subjected to mode folding operation to obtain an initialization matrix, and the initialization matrix is iteratively updated through the OIALM (Orthogonal Inexact Augmented Lagrange Multiplier Method) algorithm, including singular value thresholding of the low-rank term, soft thresholding update of the sparse abnormal term, and adjusting the weight of the kernel norm of the modal matrix based on the modal weight, to obtain a low-rank matrix. The formula is:

[0102] ,

[0103] wherein, the optimization objective function, the low-rank term of modal , the sparse regularization parameter, the sparse abnormal term, the low-rank term of modal , the kernel norm of the low-rank term of modal.

[0104] By constructing an optimization target and using the OIALM algorithm for weighted low-rank decomposition, the sparse abnormal items can be effectively suppressed while the global structure of the data is maintained, so that the low-rank matrix has high stability and anti-noise capability. The iterative optimization framework of OIALM enables the low-rank items and sparse expression to be updated synchronously, realizing global smoothing and local anomaly repair of the energy consumption sequence. Through mode weight guidance, the decomposition process can be closer to the real contribution structure of building energy consumption, so as to obtain a low-rank matrix with higher interpretability, which effectively removes invalid disturbances and has good robustness to local anomalies caused by extreme weather and abnormal equipment behavior, providing a very stable input basis for the prediction model.

[0105] Further, the initial predicted energy consumption is output by the prediction model, including:

[0106] Based on the low-rank matrix, the initial predicted energy consumption is output by the prediction model, including performing a mode folding operation on the low-rank matrix to obtain a stable time series, performing row right padding on the sequence to obtain a TCN input sequence, inputting the TCN input sequence into the TCN network to obtain a deep feature vector;

[0107] The deep feature vector is input into the bidirectional GRU to obtain a hidden state, the importance score of each time point is obtained through the attention mechanism, the importance score is calculated through the Softmax function to obtain the attention weight, and the importance score and the attention weight are weighted and summed to obtain a context vector. The context vector is input into the fully connected output layer to obtain the initial predicted energy consumption.

[0108] By using the low-rank matrix as a high-quality input, the initial energy consumption prediction is realized through a multi-layer time series model, which can significantly improve the prediction accuracy. The TCN can capture local patterns in long sequences, the bidirectional GRU can learn the context relationship, and the attention mechanism can further identify the key time window in energy consumption changes, thereby realizing differentiated modeling of energy consumption contribution in different stages. Through the context vector and the fully connected layer, the initial predicted energy consumption is obtained, so that the model presents stronger generalization ability after multi-modal deep fusion, and has stable capture ability for peak energy consumption and low-frequency trend.

[0109] Further, the golden wolf optimization algorithm is used to optimize the model parameters to generate the final predicted energy consumption, including:

[0110] The space of the to-be-optimized hyperparameters is initialized, including the convolution kernel size, the dilation rate and the regularization coefficient of the TCN, the number of hidden units and the number of layers of the bidirectional GRU, and the key vector dimension of the attention mechanism. The golden wolf optimization algorithm is used to iteratively update the to-be-optimized hyperparameter space (minimize the root mean square error) to obtain the optimal hyperparameter space, and the final predicted energy consumption is generated.

[0111] The global optimization of the prediction model super parameter is carried out through the golden wolf optimization algorithm, so that the final prediction result keeps high precision under different environmental conditions, the algorithm has strong global search ability and high ability of jumping out of local optimum, can quickly find the optimal combination in the complex super parameter space, makes the model structure get the optimal configuration, minimizes the error driving optimization process, and finally generates the prediction result with lower deviation and higher stability, which provides strong engineering practical value for building energy consumption prediction.

[0112] To sum up, the application defines the optimization objective function based on the modal weight, iteratively updates by using the OIALM algorithm, obtains the low-rank matrix, outputs the initial predicted energy consumption by the prediction model, optimizes the model parameters by using the golden wolf optimization algorithm, and generates the final predicted energy consumption; effectively fuses the building physical semantics and dynamic environmental characteristics, improves the accuracy and robustness of the energy consumption feature expression, and improves the modeling capability for non-steady-state thermal process and the energy consumption prediction precision.

[0113] It should be noted that the above examples are only used to illustrate the technical solutions of the application and are not limiting. Although the application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the application can be modified or replaced equivalently without departing from the spirit and scope of the technical solutions of the application, which should be covered in the scope of the claims of the application.

Claims

1. A BIM-based energy consumption prediction and monitoring system, characterized in that: include, The unified spatiotemporal matrix and dynamic energy consumption driving sequence construction module is used to obtain BIM models through API interfaces and real-time meteorological data through IoT platforms. The BIM models and IoT platforms are preprocessed to form a unified spatiotemporal matrix. Based on the unified spatiotemporal matrix, the transient heat flux energy of components, transient load of rooms and HVAC equivalent load are calculated to obtain the dynamic energy consumption driving sequence. The sparse suppression and smoothing tensor generation module is used to construct a three-modal feature tensor based on a structured energy consumption parameter matrix, an extended environment sequence, and a dynamic energy consumption driving sequence. Based on the three-modal feature tensor, the energy value is calculated for each feature dimension. The sparse suppression tensor is obtained by filtering. The time modes in the sparse suppression tensor are linearly interpolated to obtain local gradients. The local gradients of each equal time interval are sorted in descending order. The time point corresponding to the maximum local gradient is selected as the optimal node and a sequence of optimal nodes is constructed. The time order is rearranged based on the optimal node sequence to obtain the rearranged sparse suppression tensor. Random perturbation is added to each time point to generate a smoothing signal and obtain the smoothed sparse suppression tensor. The attention-enhanced feature reconstruction module is used to calculate the first-order gradient of the building physics feature mode in the three-modal feature tensor based on the smooth sparse suppression tensor, generate a column-uniform matrix, and perform SVD decomposition to obtain the factor matrix. The core tensor is obtained through Tucker inverse mapping. Row vectors are extracted based on the factor matrix and attention weights are calculated. A diagonal weighted matrix is ​​constructed to reconstruct the core tensor along the time direction. The final enhanced features are obtained through Tucker reconstruction. Based on the time mode, building physics feature mode, and environment and operation feature mode in the final enhanced features, matrix expansion is performed to obtain the modality matrix. The energy intensity is calculated for each modality matrix, and modality weights are constructed. The low-rank optimization and energy consumption prediction output module is used to define an optimization objective function based on modal weights, perform mode folding operations on the modal matrix to obtain an initialization matrix, and iteratively update the initialization matrix using the OIALM algorithm. This includes thresholding singular values ​​for low-rank terms, soft-thresholding updates for sparse outliers, and adjusting the kernel norm weights of the modal matrix based on the modal weights using a weighted kernel norm to obtain a low-rank matrix. The module then outputs the initial predicted energy consumption through the prediction model, including performing mode folding operations on the low-rank matrix to obtain a stable time series, performing row right padding on the series to obtain the TCN input sequence, inputting it into the TCN network to obtain a deep feature vector, inputting the deep feature vector into a bidirectional GRU to obtain the hidden state, obtaining the importance score for each time point through an attention mechanism, calculating the attention weights of the importance scores using the Softmax function, and performing a weighted sum of the importance scores and attention weights to obtain a context vector. This context vector is input into the fully connected output layer to obtain the initial predicted energy consumption. Finally, the Golden Jackal optimization algorithm is used to optimize the parameters of the prediction model to generate the final predicted energy consumption.

2. The BIM-based energy consumption prediction and monitoring system as described in claim 1, characterized in that: The process of rearranging the temporal order based on the optimal node sequence to obtain the rearranged sparse suppression tensor includes: The structured energy consumption parameter matrix, extended environment sequence, and dynamic energy consumption driving sequence are horizontally concatenated along the feature dimension to obtain a unified concatenated tensor. The feature dimensions of the unified concatenated tensor are then divided to obtain a three-modal feature tensor. The energy value of each feature dimension of the three-modal feature tensor is calculated. If the energy value is less than the energy value threshold, the corresponding feature dimension is deleted to obtain a sparse suppression tensor. The local gradient is calculated, and the local gradients of each equal time interval are sorted in descending order. The time point corresponding to the maximum local gradient is selected and set as the optimal node. The optimal nodes are then horizontally arranged to obtain the optimal node sequence. The temporal order in the sparse suppression tensor is rearranged based on the optimal node sequence to obtain the rearranged sparse suppression tensor. Random perturbation is added to each time point in the rearranged sparse suppression tensor, and a smoothing signal is calculated. The smoothing signal is then used to replace the rows in the expanded matrix to obtain the smoothed sparse suppression tensor.

3. The BIM-based energy consumption prediction and monitoring system as described in claim 2, characterized in that: The generation of the column-consistent matrix and the subsequent SVD decomposition to obtain the factor matrix include: Based on the smooth sparse suppression tensor, the first-order gradient of the building physical feature modes is calculated to generate a column-uniform matrix. The uniform matrix is ​​then decomposed by SVD to obtain the time mode factor matrix. The three-mode tensor is expanded by mode-2 and mode-3 to obtain the building physical feature matrix and the environmental and operational feature matrix. The two matrices are then decomposed by SVD to obtain the building physical feature factor matrix and the environmental and operational feature factor matrix.

4. The BIM-based energy consumption prediction and monitoring system as described in claim 3, characterized in that: The constructed diagonal weighted matrix is ​​used to reconstruct the core tensor along the time direction, and the final enhanced features are obtained through Tucker reconstruction, including: The core tensor is obtained by performing Tucker inverse mapping on the time modality factor matrix, the building physics feature factor matrix, and the environmental and operational feature factor matrix. The row vectors corresponding to each time point are extracted from the temporal modality factor matrix, the attention weights are calculated, a diagonal weighted matrix is ​​constructed by constructing a diagonal matrix, the core tensor is reconstructed in the temporal direction to obtain the attention-enhanced core tensor, and the attention-enhanced core tensor is reconstructed by Tucker to obtain the final enhanced features.

5. The BIM-based energy consumption prediction and monitoring system as described in claim 4, characterized in that: The step of optimizing the parameters of the prediction model using the Golden Jackal optimization algorithm to generate the final predicted energy consumption includes: Initialize the hyperparameter space to be optimized, including the kernel size, dilation rate and regularization coefficient of TCN, the number of hidden units and layers of bidirectional GRU, and the key vector dimension of the attention mechanism. Use the Golden Jackal optimization algorithm to iteratively update the hyperparameter space to be optimized to obtain the optimal hyperparameter space and generate the final predicted energy consumption.

6. The BIM-based energy consumption prediction and monitoring system as described in claim 5, characterized in that: The process involves obtaining a BIM model via an API interface and acquiring real-time meteorological data through an IoT platform, followed by preprocessing of both the BIM model and the IoT platform, including: By acquiring BIM models through API interfaces and real-time meteorological data through IoT platforms, the BIM models and IoT platforms are preprocessed to obtain structured energy consumption parameter matrices and extended environmental sequences.

7. The BIM-based energy consumption prediction and monitoring system as described in claim 6, characterized in that: The calculation of transient heat flux energy of the component, transient room load, and HVAC equivalent load yields a dynamic energy consumption drive sequence, including: The structured energy consumption parameter matrix and the extended environment sequence are concatenated to obtain an alignment matrix. All alignment matrices within time t are then sorted according to time. The dimensions are stacked to obtain a unified spatiotemporal matrix, and a dynamic energy consumption driving sequence is constructed. This includes calculating the transient heat flux energy of components, the transient load of the room, and the equivalent load of HVAC. Combined with the solar heat gain, the normalization process is performed and a vector horizontal splicing operation is carried out to obtain the dynamic energy consumption driving sequence.

Images