Intelligent building operation analysis method and system based on AI

Abstract

The invention discloses an AI-based intelligent building operation analysis method and system, and the method comprises the steps: collecting environment system data, energy consumption system data and people flow system data of a building through a sensor, and carrying out the time alignment of the three types of system data through time axis mapping, and obtaining an alignment matrix; performing adaptive wavelet packet decomposition on the alignment matrix to obtain extended features, and constructing a feature tensor based on the extended features; the features in the feature tensor serve as observation variables to construct a hidden variable model, the hidden variable model calculates the influence weight of the hidden variables on the observation variables through a measurement equation, and dynamic coupling degree calculation is conducted based on the influence weight; features in the feature tensor serve as nodes, the product of the transfer entropy and the dynamic coupling degree serves as an edge weight, a causal network is constructed, and root cause analysis is conducted on the causal network through a PrefixSpan algorithm. According to the method, the problem that dynamic interaction among multi-system data cannot be comprehensively considered in an existing building operation analysis technology is solved.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence, and particularly relates to a method and system for analyzing the operation of a smart building based on AI. Background Art

[0002] In the field of smart buildings, existing building operation analysis technologies mainly focus on data monitoring and analysis of a single system, such as environmental monitoring, energy consumption assessment, or pedestrian flow analysis. Most of these methods are based on traditional data collection means and lack the ability to comprehensively analyze multiple systems, resulting in difficulty in capturing the overall dynamic associations of building operations, and thus the phenomenon of information islands often occurs. In addition, most existing technologies rely on static models, lacking real-time and accuracy in analyzing the building operation status, and it is difficult to comprehensively reflect the operation efficiency of the building under different conditions. With the development of artificial intelligence technology, there is an urgent need for a method that integrates data from multiple systems and realizes dynamic analysis to more effectively support the intelligent management and optimization decision-making of buildings.

[0003] Chinese Patent with Publication No. CN118643955B discloses a building energy consumption optimization management system and method based on big data analysis. The system includes: a data collection module configured to collect real-time data; a data processing module configured to identify abnormal data and repair the abnormal data; a data analysis module configured to predict the equipment energy consumption at a future moment to obtain an equipment energy consumption trend, and form an initial energy consumption allocation plan according to the equipment energy consumption trend and a mapping relationship network; an energy consumption optimization module configured to obtain an optimal energy consumption allocation plan by using an optimization algorithm; a control management module configured to form a control instruction based on the optimal energy consumption allocation plan and send the control instruction to each device; and a visualization interface module configured to visualize various types of data and logs. This invention only relies on simple timestamp matching, which is prone to misjudging the correlation between the environment and energy consumption; at the same time, the time series modeling network of this invention only relies on historical energy consumption data and does not integrate building physical characteristics (such as thermal inertia, equipment efficiency) and multi-scale features, which may lead to long-term energy consumption prediction deviating from reality. Summary of the Invention

[0004] The purpose of the present invention is to provide a method and system for analyzing the operation of a smart building based on AI to solve the problem that existing building operation analysis technologies cannot comprehensively consider the dynamic interaction between multi-system data and meet the refined operation and maintenance requirements of smart buildings.

[0005] The technical solution of the present invention is as follows: On the one hand, the present invention provides a method for analyzing the operation of a smart building based on AI, including the following steps: Collect the environmental system data, energy consumption system data, and human flow system data of the building through sensors. Based on the three types of system data, construct three types of system pairs: environment - energy consumption, energy consumption - human flow, and human flow - environment. And perform time alignment on the three types of system data through time axis mapping to obtain an alignment matrix.

[0006] Perform adaptive wavelet packet decomposition on the alignment matrix to obtain extended features, and construct a feature tensor based on the extended features.

[0007] Use the features in the feature tensor as observed variables to construct a latent variable model. The latent variable model calculates the influence weight of the latent variable on the observed variable through a measurement equation, and performs dynamic coupling degree calculation based on the influence weight to quantify the real - time interaction strength between different system pairs.

[0008] Use the features in the feature tensor as nodes, calculate transfer entropy based on the feature tensor, and take the product of the transfer entropy and the dynamic coupling degree as the edge weight to construct a causal network. Perform root cause analysis on the causal network through the PrefixSpan algorithm to obtain the causal chain result.

[0009] Preferably, performing time alignment on the three types of system data through time axis mapping to obtain an alignment matrix is specifically as follows: Calculate the average time difference of the three types of system data:

[0010] In the formula, is the average time difference of the th type of system data, ; is the number of sensors for collecting the th type of sensor, ; is the collection time of the th sensor in the th type of system data; is the standard time.

[0011] Construct a time axis mapping function based on the average time difference as the optimization objective function. The time axis mapping function is expressed as:

[0012] In the formula, is the time axis mapping function; is the time; is the ownership weight of the th sensor in the th type of system data; is the standard deviation of the th sensor in the th type of system data; is the acquisition data of the th sensor in the class system data; is the cubic spline interpolation function on the target time axis of the th sensor in the class system data.

[0013] The LM algorithm is used to perform nonlinear optimization on the time axis mapping function to find the optimal alignment time, and the three types of system data are time-aligned to output an alignment matrix.

[0014] Preferably, the alignment matrix is adaptively decomposed by wavelet packet to obtain extended features, and a feature tensor is constructed based on the extended features specifically as follows: The alignment matrix is Fourier-transformed to obtain the power spectral density.

[0015] The spectrum of the candidate wavelet basis function is calculated, and the wavelet basis function with the largest integral value is selected based on the spectrum and the power spectral density to construct a wavelet packet decomposition tree. The selected wavelet basis function is expressed as:

[0016] In the formula, is the selected wavelet basis function; is the candidate wavelet basis function; is the Nyquist frequency; is the candidate wavelet basis function at the frequency spectrum energy; is the power spectral density, indicating the energy intensity of the alignment matrix at the frequency location.

[0017] Each eigenvector of the alignment matrix is subjected to a first-level wavelet packet decomposition using the selected wavelet basis function to generate high-frequency and low-frequency subbands.

[0018] For each subband node, the Shannon entropy is calculated. If the entropy value of the parent node and the entropy value of the child node satisfy:

[0019] In the formula, is the entropy value of the parent node; is the entropy value of the child node; is the entropy reduction rate threshold; then continue to decompose the node until the preset maximum decomposition level is reached or the entropy value of the parent node and the entropy value of the child node do not satisfy the above rules, and the construction of the wavelet packet decomposition tree is completed.

[0020] The energy entropy is calculated for each leaf node of the wavelet packet decomposition tree to obtain the energy entropy class feature.

[0021] Calculate the skewness feature, kurtosis feature, and Hurst exponent feature of the time-domain signal based on the features in the alignment matrix.

[0022] Perform dimensionality expansion on the alignment matrix based on energy entropy class features, skewness features, kurtosis features, and Hurst exponent features to obtain a feature tensor.

[0023] Preferably, use the features in the feature tensor as observed variables to construct a latent variable model. The latent variable model calculates the influence weight of the latent variable on the observed variable through the measurement equation, specifically: Use the features in the feature tensor as observed variables, and each observed variable corresponds to a certain feature dimension in the feature tensor.

[0024] The latent variable model defines that the latent variables include a thermal inertia factor, a human flow dynamic coefficient, and a device efficiency index. Based on the observed variables and the latent variables, establish a measurement equation:

[0025] In the formula, is the observed variable; is the factor loading matrix, which is used to describe the influence weight of the latent variable on the observed variable; is the latent variable; is the observation error.

[0026] Calculate the estimated value of the latent variable and the estimated value of the factor loading matrix .

[0027] Based on and perform a goodness-of-fit test. If the goodness-of-fit is greater than the set threshold, then calculate the dynamic coupling degree.

[0028] Preferably, calculate the dynamic coupling degree based on the influence weight to quantify the real-time interaction strength between different system pairs, specifically: Based on the estimated value of the factor loading matrix, calculate the dynamic Jacobian matrix:

[0029] In the formula, is the dynamic Jacobian matrix; ; is the rate of change of the latent variable with time, which is obtained by calculating the sliding window difference.

[0030] Calculate the dynamic coupling degree according to the dynamic Jacobian matrix:

[0031] In the formula, is the real-time coupling index between two types of system pairs; is at time , the Jacobian matrix element of the data pair of the type system data; , are respectively , the diagonal elements of the

[0032] Preferably, taking the features in the feature tensor as nodes, calculating the transfer entropy based on the feature tensor, and taking the product of the transfer entropy and the dynamic coupling degree as the edge weight to construct a causal network specifically as follows: Taking the features in the feature tensor as nodes, discretizing the features into three states of -1, 0, and 1, and calculating the transfer entropy between node pairs as the original edge weight. The specific formula for the transfer entropy is:

[0033] In the formula, is the original edge weight from node to node , is the probability function; is the state of node at time ; is the discretized historical sequence of node in the time window , used to control the influence length of the own history of the node; is the discretized historical sequence of node in the time window , used to control the influence length of the own history of the node; is the joint probability, , are conditional probabilities, estimated by statistical frequency through a sliding window.

[0034] Dynamically update the product of the transfer entropy between node pairs and the dynamic coupling degree of the corresponding system pair of node pairs as the edge weight to construct a causal network.

[0035] Preferably, perform root cause analysis on the causal network through the PrefixSpan algorithm, and the specific causal chain result is: Set the activation threshold, and encode the activation status of the edges with edge weights exceeding the activation threshold in the causal network as a symbol sequence within a preset time window.

[0036] Scan the symbol sequence, count the frequency of each symbol, and retain the symbols whose symbol support exceeds the set threshold as frequent items.

[0037] Recursively construct a projection database to expand the frequent items to generate longer frequent patterns.

[0038] Select the frequent patterns whose support and confidence both exceed the threshold as causal chains for output, and locate the earliest activated node in the causal chain as the root cause to complete the causal analysis.

[0039] On the other hand, the present invention provides an AI-based intelligent building operation analysis system, including a data acquisition and time alignment module, a tensor construction module, a dynamic coupling degree calculation module, and a root cause analysis module.

[0040] The data acquisition and time alignment module is used to collect the environmental system data, energy consumption system data, and pedestrian flow system data of the building through sensors, construct three types of system pairs of environment-energy consumption, energy consumption-pedestrian flow, and pedestrian flow-environment based on the three types of system data, and perform time alignment on the three types of system data through time axis mapping to obtain an alignment matrix.

[0041] The tensor construction module is used to perform adaptive wavelet packet decomposition on the alignment matrix to obtain extended features, and construct a feature tensor based on the extended features.

[0042] The dynamic coupling degree calculation module is used to construct a latent variable model with the features in the feature tensor as observed variables. The latent variable model calculates the influence weight of the latent variable on the observed variable through a measurement equation, and performs dynamic coupling degree calculation based on the influence weight to quantify the real-time interaction strength between different system pairs.

[0043] The root cause analysis module is used to use the features in the feature tensor as nodes, calculate the transfer entropy based on the feature tensor, and use the product of the transfer entropy and the dynamic coupling degree as the edge weight to construct a causal network, and perform root cause analysis on the causal network through the PrefixSpan algorithm to obtain the causal chain result.

[0044] On yet another aspect, the present invention further provides an electronic device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the AI-based intelligent building operation analysis method as described in any embodiment of the present invention.

[0045] On yet another aspect, the present invention further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the AI-based intelligent building operation analysis method as described in any embodiment of the present invention.

[0046] Compared with the prior art, the present invention has the following technical effects: 1. By introducing a weight factor for sensor characteristic adaptation and a non-linear optimization objective function, the present invention realizes the time alignment of multi-source data, effectively eliminates the pseudo-correlation caused by device clock drift, and ensures the temporal consistency of subsequent analysis.

[0047] 2. By constructing a latent variable model and calculating the dynamic coupling degree using the influence weights of observed variables, the present invention realizes the quantification of real-time interaction between each system pair. This innovative method ensures that during the operation of a building, the mutual influence between the environment, energy consumption, and human flow can be monitored and analyzed in real time, thereby providing data-driven decision support for managers and significantly improving the flexibility and response speed of building operation analysis.

[0048] 3. By constructing a causal network and using the product of calculated transfer entropy and dynamic coupling degree as the edge weight for root cause analysis, the present invention realizes the in-depth exploration of causal relationships within a complex system. It can clearly reveal the causal chain between different variables, provide a more targeted basis for optimizing building management, and thus improve resource allocation. Description of the Drawings

[0049] Figure 1 is the overall flowchart of the AI-based intelligent building operation analysis method described in the present invention. Detailed Embodiments

[0050] To make the objectives, technical solutions, and advantages of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with specific embodiments of the present application and with reference to the accompanying drawings.

[0051] Embodiment 1 This embodiment provides an AI-based intelligent building operation analysis method. Referring to Figure 1 as shown, it includes the following steps: Collect environmental system data, energy consumption system data, and human flow system data of the building through sensors, construct three types of system pairs: environment-energy consumption, energy consumption-human flow, and human flow-environment based on the three types of system data, and perform time alignment on the three types of system data through time axis mapping to obtain an alignment matrix. In addition, conventional data preprocessing steps such as abnormal cleaning need to be performed on the collected data.

[0052] Specifically, collect environmental system data such as temperature, humidity, and CO 2 , and illumination through environmental sensors, collect energy consumption system data such as HVAC power, lighting power, and device current through energy consumption meters, and collect human flow system data such as density, speed, and residence time through human flow cameras.

[0053] As a preferred implementation manner of this embodiment, the specific method for obtaining the alignment matrix by time-aligning the three types of system data through time-axis mapping is as follows: Calculate the average time difference of the three types of system data:

[0054] In the formula, is the average time difference of the th type of system data, ; is the number of sensors for collecting the th type of sensor, ; is the collection time of the rd sensor in the th type of system data; is the standard time.

[0055] Construct a time-axis mapping function based on the average time difference as the optimization objective function, and the time-axis mapping function is expressed as:

[0056] In the formula, is the time-axis mapping function; is the time; is the weight of the rd sensor in the th type of system data, , is the response time constant of the sensor (set according to the sensor type), and the weight is dynamically adjusted for alignment according to the accuracy and stability of the sensor; is the standard deviation of the rd sensor in the th type of system data, indicating the degree of data fluctuation; is the collected data of the rd sensor in the th type of system data; is the cubic spline interpolation function of the rd sensor on the target time axis in the th type of system data.

[0057] Use the LM algorithm to perform nonlinear optimization on the time-axis mapping function, find the optimal alignment time, and perform time alignment on the three types of system data to output the alignment matrix.

[0058] Specifically, the algorithm steps of the LM algorithm are as follows: randomly initialize a time alignment scheme; calculate the current error through the objective function; adjust the time alignment scheme according to the error and gradually reduce the error; stop the algorithm when the error change is less than the preset percentage threshold after two consecutive adjustments.

[0059] The alignment matrix is subjected to adaptive wavelet packet decomposition to obtain extended features, and a feature tensor is constructed based on the extended features.

[0060] As a preferred implementation manner of this embodiment, the alignment matrix is subjected to adaptive wavelet packet decomposition to obtain extended features, and constructing a feature tensor based on the extended features is specifically as follows: The alignment matrix is subjected to Fourier transform to obtain the power spectral density. The specific formula of the power spectral density is:

[0061] In the formula, is the power spectral density; is the total number of time points, ; is the frequency; is the time series data of the features in the alignment matrix; is the imaginary unit.

[0062] Different data are suitable for different analysis waveforms, and it is necessary to select the wavelet basis function that best matches the data spectrum. Therefore, the spectra of the candidate wavelet basis functions are calculated, and the wavelet basis function with the largest integral value is selected based on the spectrum and the power spectral density to construct the wavelet packet decomposition tree. The selected wavelet basis function is expressed as:

[0063] In the formula, is the selected wavelet basis function; is the candidate wavelet basis function (such as Daubechies wavelet, Symlets wavelet, etc.); is the Nyquist frequency, , is the sampling frequency; is the candidate wavelet basis function at the frequency the spectral energy at; is the power spectral density, indicating the energy intensity of the alignment matrix at the frequency at.

[0064] Each feature column of the alignment matrix is subjected to first-level wavelet packet decomposition using the selected wavelet basis function to generate high-frequency and low-frequency subbands.

[0065] For each subband node, the Shannon entropy is calculated. The larger the entropy, the more complex the signal, and further decomposition is required. The entropy value formula is expressed as:

[0066] In the formula, is the Shannon entropy of the node, describing the degree of disorder of the energy in each frequency band within a single node; For the child node The proportion of the energy of the to the total energy of the parent node.

[0067] In this embodiment, the depth of the wavelet packet decomposition tree is controlled by a dynamic threshold to balance the feature dimension and the amount of information. If the entropy value of the parent node and the entropy value of the child node satisfy:

[0068] In the formula, is the entropy value of the parent node; is the entropy value of the child node; is the preset entropy reduction rate threshold; Then continue to decompose this node until the preset maximum decomposition level is reached or the entropy value of the parent node and the entropy value of the child node do not satisfy the above rules, and the construction of the wavelet packet decomposition tree is completed.

[0069] Calculate the energy entropy for the frequency band corresponding to each leaf node of the wavelet packet decomposition tree to obtain the energy entropy class features. Usually, the operation analysis of intelligent buildings focuses on time scales above the minute level, and the second-level fluctuations are regarded as noise. Therefore, the energy entropy generally only extracts the hourly energy entropy, the daily energy entropy, and the weekly energy entropy to explain the physical meaning of the subsequent latent variables. In this embodiment, it is preferably to expand the features of the alignment matrix to the hourly energy entropy (capturing the fluctuations within an hour) and the daily energy entropy (analyzing the daily cycle trend). Its calculation formula is expressed as:

[0070] In the formula, is the energy entropy; is the th proportion of the total energy of the

[0071] Based on the time-domain signal of the features in the alignment matrix, calculate the skewness feature, the kurtosis feature, and the Hurst exponent feature. Specifically, the skewness feature is calculated through the mean and the standard deviation, which is used to represent whether the data distribution is symmetric; the kurtosis feature is used to represent whether the data is steep; the Hurst exponent feature is calculated through the rescaled range analysis (R / S analysis), which reflects the long-range correlation of the time series.

[0072] This embodiment synthesizes the frequency-domain energy distribution and the time-domain statistical characteristics to form multi-scale features, and expands the alignment matrix based on the energy entropy class features, the skewness feature, the kurtosis feature, and the Hurst exponent feature to obtain the feature tensor. Specifically, the feature tensor structure of this embodiment is preferably: time dimension * original features (4 types of environmental system data, 3 types of energy consumption system data, and 3 types of people flow system data) * expanded features (hourly energy entropy, daily energy entropy, skewness, kurtosis, and Hurst exponent). For example, the temperature feature is further expanded into 5 types of dimension features.

[0073] Construct a latent variable model with the features in the feature tensor as the observed variables. The latent variable model calculates the influence weights of the latent variables on the observed variables through a measurement equation, and calculates the dynamic coupling degree based on the influence weights to quantify the real-time interaction strength between different system pairs.

[0074] As a preferred implementation manner of this embodiment, the physical characteristics of the building system (including thermal inertia, people flow dynamics, equipment efficiency) are abstracted through latent variables, and the mapping relationship between the observed variables and the latent variables is established. Construct a latent variable model with the features in the feature tensor as the observed variables. The specific method for calculating the influence weights of the latent variables on the observed variables by the latent variable model through the measurement equation is as follows: Take the features in the feature tensor as the observed variables, and each observed variable corresponds to a certain feature dimension in the feature tensor.

[0075] The latent variable model defines that the latent variables include a thermal inertia factor, a people flow dynamics coefficient, and an equipment efficiency index. The thermal inertia factor reflects the heat storage capacity of the building envelope structure; the people flow dynamics coefficient quantifies the comprehensive influence of the people flow density and the movement pattern; the equipment efficiency index characterizes the operation efficiency of equipment such as HVAC and lighting.

[0076] Based on the observed variables and the latent variables, establish a measurement equation:

[0077] In the formula, is the observed variable; is the factor loading matrix, which is used to describe the influence weights of the latent variables on the observed variables, , where is the observed variable dimension; is the latent variable; is the observation error, which follows a Gaussian distribution.

[0078] Calculate the estimated value of the latent variable and the estimated value of the factor loading matrix through the two-stage least squares method. The specific solution process of the two-stage least squares method is as follows: The first stage: Use principal component analysis (PCA) to extract the estimated value of the latent variable , specifically: perform PCA analysis on the observed vector to extract the first 3 principal components (representing thermal inertia, people flow dynamics, and equipment efficiency respectively) as the estimated value of the latent variable.

[0079] The second stage: Calculate the estimated value of the factor loading matrix , specifically: taking the measurement equation as the regression model, the estimated values of the latent variables as the independent variables, and the observed variables as the dependent variables, and solving by ordinary least squares of the estimated values.

[0080] Based on and perform a goodness-of-fit test. If the goodness-of-fit is greater than the set threshold, calculate the dynamic coupling degree. By ensuring that the goodness-of-fit is greater than the set threshold, it can be ensured that the latent variable model fits well, thus guaranteeing the reliability of subsequent coupling analysis. Specifically, the closer the goodness-of-fit index is to 1, the better the latent variable model matches the actual data. Its formula is specifically:

[0081] In the formula, is the goodness-of-fit index; is the trace of the matrix; is the covariance matrix predicted by the latent variable model, calculated based on and ; is the covariance matrix of the actual observed data, calculated from the observed variables in the feature tensor; is the identity matrix.

[0082] As a preferred implementation manner of this embodiment, calculate the dynamic coupling degree based on the influence weights to quantify the real-time interaction strength between different system pairs. Specifically: Based on the estimated value of the factor loading matrix calculate the dynamic Jacobian matrix:

[0083] In the formula, is the dynamic Jacobian matrix; ; is the change rate of the latent variable over time, calculated by sliding window differencing (such as the change rate of the latent variable in the past 1 hour).

[0084] Calculate the dynamic coupling degree according to the dynamic Jacobian matrix. The higher the dynamic coupling degree, the stronger the real-time influence between the two types of systems:

[0085] In the formula, is the real-time coupling index between the two types of system pairs, ; is at time , the data pair of the type of system and the element of the Jacobian matrix of the data of the type of system; are respectively , The diagonal elements of the class system data.

[0086] Furthermore, the three types of data are divided into three system pairs: environment - energy consumption, people flow - energy consumption, and environment - people flow. By taking time as the X - axis, system pairs as the Y - axis, and real - time coupling index as the Z - axis, the real - time coupling index of all system pairs is calculated for each time point to form a spatio - temporal coupling matrix, so as to generate a three - dimensional surface graph of coupling degree. And through the persistence detection of the three - dimensional surface graph of coupling degree (such as identifying continuous time periods when the coupling index exceeds a set threshold), and ranking the contribution degrees of all persistence events, a critical path analysis report can be generated to further explain the correlation between events and data.

[0087] Taking the features in the feature tensor as nodes, calculating transfer entropy based on the feature tensor, and taking the product of the transfer entropy and the dynamic coupling degree as the edge weight to construct a causal network, and performing root - cause analysis on the causal network through the PrefixSpan algorithm to obtain the causal chain result.

[0088] As a preferred implementation manner of this embodiment, taking the features in the feature tensor as nodes, calculating transfer entropy based on the feature tensor, and taking the product of the transfer entropy and the dynamic coupling degree as the edge weight to construct a causal network specifically as follows: Taking the features in the feature tensor as nodes, discretizing the features into three states: - 1, 0, and 1. Specifically, the feature data is standardized and multiplied by a preset value to obtain a standardized value, and the standardized value is discretized into - 1, 0, and 1 through a threshold, discretizing the continuous features into finite states for facilitating the calculation of probability distribution. Furthermore, the random forest algorithm can be used to calculate the feature importance of the features to retain the features with higher importance as nodes.

[0089] Calculating the transfer entropy between node pairs as the original edge weight, where the transfer entropy is used to measure the improvement of the predictive ability of the historical information of one type of data on the future value of another type of data, and its formula is specifically:

[0090] In the formula, is the original edge weight from node to node , is the probability function; is the state of node at time ; is the discretized historical sequence of node in the time window , is used to control the influence length of the own history of node ; For a node In the time window Of the discretized historical sequence, Used to control the node The influence length of its own history; Is the joint probability, 、 Is the conditional probability, estimated by counting frequencies through a sliding window.

[0091] Multiply the transfer entropy between node pairs by the dynamic coupling degree of the corresponding system pair of the node pairs as the edge weight for dynamic update to construct a causal network. Specifically, each node pair belongs to a system pair, and the transfer entropy of the node pair is multiplied by the dynamic coupling degree of its affiliated system pair, highlighting the causal relationship in important system pairs. The dynamic coupling index is updated in real time through a sliding window (such as data in the past 1 hour), reflecting the current coupling strength of the system pair. The transfer entropy needs to count probabilities based on a sufficiently long period of data (such as data in the past 24 hours) to ensure stability. Therefore, an alignment strategy needs to be formulated for the dynamic update of the edge weight, and the alignment strategy can be: sliding window alignment or real-time dynamic weight update.

[0092] The specific sliding window alignment is: the transfer entropy calculation is based on long-term data (such as a 24-hour window), while the dynamic coupling degree can be based on a sliding average version of the same time window.

[0093] The specific real-time dynamic weight update is: use the currently calculated dynamic coupling index as the weight factor to adjust the transfer entropy value:

[0094] In the formula, Is the updated transfer entropy value; Is the initial transfer entropy value; Is the dynamic coupling index of the system pair at the current time.

[0095] Furthermore, a permutation test based on a sliding window can be performed on the causal network edges to filter out non-significant edges. The specific implementation steps of the permutation test are: randomly shuffle the time order of the nodes To destroy its potential causal relationship with Calculate the transfer entropy after shuffling; repeat multiple times (such as 1000 times) to generate the null distribution of the transfer entropy; calculate the percentile p of the true TE value in the null distribution; if p is less than the set threshold, it is considered that the causal relationship is significant and the edge is retained; otherwise, it is filtered.

[0096] As a preferred implementation manner of this embodiment, perform root cause analysis on the causal network through the PrefixSpan algorithm, and the specific causal chain result is: Set an activation threshold, and encode the activation status of the edges in the causal network whose edge weights exceed the activation threshold into a symbol sequence within a preset time window. The activation condition can be expressed as: , it is marked as activated (1), otherwise it is not activated (0).

[0097] Scan the symbol sequence, count the frequency of each symbol, and retain the symbols whose symbol support exceeds the set threshold as frequent items. The frequent items are the symbols that frequently appear in the time window sequence.

[0098] Recursively construct a projection database to expand the frequent items to generate longer frequent patterns.

[0099] Select the frequent patterns whose support and confidence both exceed the threshold as causal chains for output, and locate the earliest activated node in the causal chain as the root cause to complete the causal analysis.

[0100] To verify the effectiveness and superiority of the method provided in this embodiment, the following provides some simple cases: Based on the Ubuntu 22.04 LTS operating system, configure a virtual machine to simulate a building environment: Collect data for the central office building, where the environmental system data is collected through temperature, humidity, CO 2 as well as light sensors. In the energy consumption system data, the HVAC power is collected through the central air conditioning system, the lighting power and equipment current are collected through the energy consumption meter, and the people flow system data is collected through a binocular camera. The sampling duration is set to 24 hours and the frequency is at 10s intervals.

[0101] Obtain an alignment matrix through time alignment ; output a feature tensor through adaptive wavelet packet decomposition (select hourly energy entropy, daily energy entropy, skewness, kurtosis, Hurst index); construct a causal network to output a causal graph, and obtain a causal chain such as "people flow density↑→CO 2 concentration↑→HVAC power↑". In this causal chain, "people flow density↑" is the root cause. The root cause analysis for the above causal chain is: The sharp increase in people flow density (morning rush hour) is the root cause of the HVAC energy consumption exceeding the limit (peak value 350kW) at noon. And the daily energy entropy shows that the HVAC power has periodic fluctuations (once every 2 hours).

[0102] Embodiment 2 Correspondingly, this embodiment provides an AI-based intelligent building operation analysis system, which is used to implement the AI-based intelligent building operation analysis method as described in Embodiment 1 of the present invention, including a data collection and time alignment module, a tensor construction module, a dynamic coupling degree calculation module, and a root cause analysis module.

[0103] A data acquisition and time alignment module, which is used to collect environmental system data, energy consumption system data, and people flow system data of a building through sensors, construct three types of system pairs, namely environment-energy consumption, energy consumption-people flow, and people flow-environment, based on the three types of system data, and perform time alignment on the three types of system data through time axis mapping to obtain an alignment matrix.

[0104] A tensor construction module, which is used to perform adaptive wavelet packet decomposition on the alignment matrix to obtain extended features, and construct a feature tensor based on the extended features.

[0105] A dynamic coupling degree calculation module, which is used to construct a latent variable model with the features in the feature tensor as observation variables. The latent variable model calculates the influence weight of the latent variable on the observation variable through a measurement equation, and performs dynamic coupling degree calculation based on the influence weight to quantify the real-time interaction strength between different system pairs.

[0106] A root cause analysis module, which is used to use the features in the feature tensor as nodes, calculate transfer entropy based on the feature tensor, and use the product of the transfer entropy and the dynamic coupling degree as edge weights to construct a causal network. The causal network is subjected to root cause analysis through the PrefixSpan algorithm to obtain a causal chain result.

[0107] Embodiment 3 This embodiment provides an electronic device, which includes: a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the computer program, it implements the AI-based intelligent building operation analysis method as described in Embodiment 1 of the present invention.

[0108] Embodiment 4 This embodiment provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, it implements the AI-based intelligent building operation analysis method as described in Embodiment 1 of the present invention.

[0109] In the embodiments of the present application, "at least one" means one or more, and "a plurality" means two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent the situation where A exists alone, A and B exist simultaneously, or B exists alone. Where A and B may be singular or plural. The character " / " generally represents an "or" relationship between the associated objects before and after. "At least one of the following" and its similar expressions refer to any combination of these items, including any combination of single items or plural items. For example, at least one of a, b, and c may represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c may be single or multiple.

[0110] Those of ordinary skill in the art will realize that the various units and algorithm steps described in the embodiments disclosed herein can be implemented by electronic hardware, computer software, or a combination of electronic hardware. Whether these functions are executed in hardware or software depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered to exceed the scope of this application.

[0111] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the systems, devices, and units described above can refer to the corresponding processes in the foregoing method embodiments and will not be elaborated herein.

[0112] In several embodiments provided in this application, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (hereinafter referred to as ROM), random access memories (hereinafter referred to as RAM), magnetic disks, or optical discs that can store program codes.

[0113] The above are only the embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made by using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present invention.

Claims

1. A smart building operation analysis method based on AI, characterized in that: The following steps are involved: The environmental system data, energy consumption system data and human flow system data of the building are collected through sensors, and three types of system pairs, namely environment-energy consumption, energy consumption-human flow and human flow-environment, are constructed based on the three types of system data. The three types of system data are time-aligned through time axis mapping to obtain an alignment matrix; Adaptively decompose the alignment matrix with wavelet packets to obtain extended features, and construct feature tensors based on the extended features; The features in the feature tensor are used as observed variables to construct a latent variable model, wherein the latent variable model calculates the influence weight of the latent variable on the observed variable through a measurement equation, and calculates the dynamic coupling degree based on the influence weight to quantify the real-time interaction strength between different system pairs; The features in the feature tensor are used as nodes, the transfer entropy is calculated based on the feature tensor, and the product of the transfer entropy and the dynamic coupling degree is used as the edge weight to construct a causal network. The root cause analysis of the causal network is performed using the PrefixSpan algorithm to obtain the causal chain result.

2. The AI-based smart building operation analysis method according to claim 1 is characterized in that: The three types of system data are time-aligned through time axis mapping to obtain an alignment matrix: Calculate the average time difference of the three types of system data: In the formula, For the The average time difference of class system data, ; To collect The number of sensors of this type, ; For the Class system data The acquisition time of each sensor; is standard time; A time axis mapping function is constructed based on the average time difference as the objective function of optimization, and the time axis mapping function is expressed as: In the formula, is the time axis mapping function; For time; For the Class system data The weight of each sensor; For the Class system data The standard deviation of each sensor; For the Class system data The data collected by each sensor; For the Class system data Cubic spline interpolation function on the time axis of sensor target; The LM algorithm is used to perform nonlinear optimization on the time axis mapping function to find the optimal alignment time, and the three types of system data are time-aligned to output an alignment matrix.

3. The AI-based smart building operation analysis method according to claim 1 is characterized in that: The alignment matrix is ​​adaptively decomposed by wavelet packets to obtain extended features, and the feature tensor is constructed based on the extended features as follows: Perform Fourier transform on the alignment matrix to obtain the power spectral density; The spectrum of the candidate wavelet basis function is calculated, and the wavelet basis function with the largest integral value is selected based on the spectrum and power spectrum density to construct the wavelet packet decomposition tree. The selected wavelet basis function is expressed as: In the formula, is the selected wavelet basis function; is a candidate wavelet basis function; is the Nyquist frequency; is the candidate wavelet basis function In frequency The spectrum energy at is the power spectral density, which indicates the alignment matrix at frequency The energy intensity at Performing a first-level wavelet packet decomposition on each characteristic column of the alignment matrix using the selected wavelet basis function to generate a high-frequency subband and a low-frequency subband; For each sub-band node, the Shannon entropy is calculated. If the entropy value of the parent node and the entropy value of the child node satisfy: In the formula, is the entropy value of the parent node; is the entropy value of the child node; is the entropy reduction rate threshold; Then continue to decompose the node until the preset maximum number of decomposition layers is reached or the entropy value of the parent node and the entropy value of the child node do not meet the above rules, and the construction of the wavelet packet decomposition tree is completed; Calculate the energy entropy of each leaf node of the wavelet packet decomposition tree to obtain the energy entropy class feature; Calculate the skewness feature, kurtosis feature and Hurst exponent feature based on the time domain signal of the features in the alignment matrix; The alignment matrix is ​​dimensionally expanded based on energy entropy features, skewness features, kurtosis features, and Hurst exponent features to obtain a feature tensor.

4. The AI-based smart building operation analysis method according to claim 1 is characterized in that: The features in the feature tensor are used as observed variables to construct a latent variable model. The latent variable model calculates the influence weight of the latent variable on the observed variable through the measurement equation: The features in the feature tensor are used as observation variables, and each observation variable corresponds to a feature dimension in the feature tensor; The latent variable model defines latent variables including thermal inertia factor, passenger flow dynamic coefficient and equipment efficiency index, and establishes a measurement equation based on observed variables and latent variables: In the formula, is the observed variable; is the factor loading matrix, which is used to describe the influence weight of latent variables on observed variables; is a hidden variable; is the observation error; Computing latent variables by two-stage least squares method Estimated value of With the factor loading matrix Estimated value of ; based on and Perform a fitness test, and if the fitness is greater than the set threshold, perform a dynamic coupling calculation.

5. The AI-based smart building operation analysis method according to claim 4 is characterized in that: The dynamic coupling degree is calculated based on the influence weights to quantify the real-time interaction strength between different system pairs: Estimates based on the factor loading matrix Compute the dynamic Jacobian matrix: In the formula, is the dynamic Jacobian matrix; ; is the rate of change of the latent variable over time, calculated by sliding window difference; Calculate the dynamic coupling based on the dynamic Jacobian matrix: In the formula, is the real-time coupling index between two types of system pairs; For in time , Class System Data Pair Jacobian matrix elements of class system data; , They are , Diagonal elements of class system data.

6. The AI-based smart building operation analysis method according to claim 1 is characterized in that: The features in the feature tensor are used as nodes, the transfer entropy is calculated based on the feature tensor, and the product of the transfer entropy and the dynamic coupling degree is used as the edge weight to construct the causal network as follows: The features in the feature tensor are used as nodes, the features are discretized into three states: -1, 0, and 1, and the transfer entropy between node pairs is calculated as the original edge weight. The transfer entropy formula is specifically as follows: In the formula, For Node To Node The original edge weights of is the probability function; For Node In time Status; For Node In the time window The discretized history sequence of For controlling nodes the length of the impact of one’s own history; For Node In the time window The discretized history sequence of For controlling nodes the length of the impact of one’s own history; is the joint probability, , is the conditional probability, estimated by sliding window statistical frequency; The product of the transfer entropy between node pairs and the dynamic coupling degree of the system pair corresponding to the node pairs is used as the edge weight for dynamic update to construct a causal network.

7. The AI-based smart building operation analysis method according to claim 6 is characterized in that: The root cause analysis of the causal network is performed using the PrefixSpan algorithm, and the causal chain results are as follows: An activation threshold is set, and the edge activation states of the edge weights in the causal network exceeding the activation threshold are encoded as symbol sequences within a preset time window; Scan the symbol sequence, count the frequency of each symbol, and retain the symbols whose support exceeds the set threshold as frequent items; Recursively construct the projection database to expand frequent items to generate longer frequent patterns; Frequent patterns whose support and confidence both exceed the threshold are selected as causal chains for output, and the earliest activated node in the causal chain is located as the root cause to complete the causal analysis.

8. An AI-based smart building operation analysis system, characterized in that: The system is used to implement the AI-based smart building operation method as described in any one of claims 1 to 7, including a data acquisition and time alignment module, a tensor construction module, a dynamic coupling degree calculation module, and a root cause analysis module; The data acquisition and time alignment module is used to collect the building's environmental system data, energy consumption system data, and human flow system data through sensors, build three types of system pairs based on the three types of system data: environment-energy consumption, energy consumption-human flow, and human flow-environment, and perform time alignment on the three types of system data through time axis mapping to obtain an alignment matrix; A tensor construction module is used to perform adaptive wavelet packet decomposition on the alignment matrix to obtain extended features, and construct a feature tensor based on the extended features; A dynamic coupling degree calculation module is used to construct a latent variable model using the features in the feature tensor as observed variables. The latent variable model calculates the influence weight of the latent variable on the observed variable through a measurement equation, and performs dynamic coupling degree calculation based on the influence weight to quantify the real-time interaction strength between different system pairs; The root cause analysis module is used to use the features in the feature tensor as nodes, calculate the transfer entropy based on the feature tensor, and use the product of the transfer entropy and the dynamic coupling degree as the edge weight to construct a causal network. The root cause analysis of the causal network is performed through the PrefixSpan algorithm to obtain the causal chain result.

9. An electronic device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the AI-based smart building operation analysis method as described in any one of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the AI-based smart building operation analysis method as described in any one of claims 1 to 7 is implemented.

Images