Industrial energy-saving monitoring method based on Copula graph model

Abstract

The invention discloses an industrial energy-saving monitoring method based on a Copula graph model, and relates to the technical field of industrial energy efficiency monitoring and energy-saving monitoring, and the method comprises the steps: collecting multi-class energy consumption original data, and carrying out the preprocessing of the data, and obtaining a multi-class energy consumption time sequence; fitting an edge probability distribution function for each energy consumption time sequence by adopting a nonparametric density estimation method, and obtaining a corresponding uniformly distributed sequence through probability integral transformation; taking the uniformly distributed sequence as a node variable, and adopting a graph structure learning algorithm to construct a Copula graph model; converting the real-time energy consumption data in the to-be-monitored time period into a real-time uniform distribution value, inputting the real-time uniform distribution value into the Copula graph model, and calculating a joint probability density value; and comparing the joint probability density value with a preset energy-saving abnormal threshold value, positioning a dominant energy type causing energy consumption abnormity based on the contribution degree of each node variable, and outputting energy-saving monitoring early warning information. According to the invention, the accuracy and interpretability of industrial energy-saving monitoring can be improved.

Description

Technical Field

[0001] This invention relates to the field of industrial energy efficiency monitoring and intelligent energy-saving management technology, and in particular to an industrial energy-saving monitoring method based on the Copula graph model. Background Technology

[0002] With the continuous advancement of industrial digitalization and intelligentization, industrial energy systems are gradually exhibiting complex characteristics such as parallel supply of multiple energy sources, coupled operation of multiple devices, and dynamic changes in multiple operating conditions. To achieve refined energy management under the "dual carbon" target, industrial energy conservation monitoring technology is gradually evolving from traditional manual inspections and single-index statistical analysis to a model based on IoT sensing, data-driven modeling, and intelligent analysis. Currently, industrial sites commonly deploy various energy consumption metering devices for electricity, steam, gas, and water, achieving energy efficiency assessment and anomaly identification through centralized collection and analysis of energy consumption data. However, existing energy conservation monitoring methods often rely on threshold judgments based on a single energy dimension or multivariate analysis models based on linear correlation assumptions, making it difficult to accurately characterize the nonlinear dependencies between different energy types in the time-series dimension. This is especially problematic in scenarios involving complex production load changes and multi-energy coordinated consumption, easily leading to false alarms or missed alarms. Furthermore, traditional methods often only focus on whether an anomaly has occurred, lacking the ability to finely trace the causes of anomalies, making it difficult to identify the dominant energy type causing energy consumption anomalies, thus limiting the practical application value of energy conservation monitoring results in production control and energy-saving decision-making.

[0003] CN114185960B discloses an optimization decision-making and management method for urban water, energy, and environmental systems based on Copula functions. This method constructs a joint Copula distribution function among water quantity, energy quantity, and environmental indicators to achieve optimal allocation and joint risk management of multi-resource systems under uncertain conditions. This scheme fully utilizes the Copula function's ability to characterize the correlation structure of variables, achieving synergistic optimization of the water-energy-environment system at the macro level. However, this method primarily addresses urban-level resource planning and long-term decision-making problems, focusing on system-level resource allocation and risk trade-offs. It does not address the real-time energy consumption monitoring needs of industrial production sites, nor does it involve anomaly detection mechanisms based on time-series energy consumption data. Furthermore, it lacks quantitative analysis of the contribution of various energy types in specific abnormal events, making it difficult to meet the requirements of real-time performance, location accuracy, and operability for industrial energy conservation monitoring.

[0004] The paper "A Multidimensional Data Spatial Scanning Monitoring Method Based on the Copula Model" proposes a method for anomaly monitoring by constructing a joint distribution of multidimensional variables using the Copula model and combining it with spatial scanning statistics. This research overcomes the limitations of traditional multivariate statistical process control on distribution assumptions, and can improve the monitoring sensitivity of multidimensional data out-of-control states to a certain extent, and detect data trend changes earlier. However, this method is mainly applied in the field of quality monitoring, focusing on the construction of monitoring statistics and out-of-control judgment. Its variable relationship modeling does not introduce graph structure constraints, making it difficult to reflect the conditional dependencies between multidimensional variables. At the same time, this method aims at anomaly detection and does not provide a structured analysis path for the source of anomalies, making it impossible to locate the dominant anomaly variable in complex energy consumption systems, which is not conducive to the division of energy consumption responsibilities and targeted regulation in energy conservation monitoring scenarios.

[0005] In summary, while existing techniques based on Copula functions or Copula models possess certain theoretical advantages in multivariate correlation modeling and joint distribution characterization, they primarily focus on macro-level resource optimization, quality process monitoring, or statistical runaway determination. They generally suffer from limitations such as difficulty in directly adapting to multi-source energy consumption time-series data in industrial settings, a lack of refined modeling of the conditional dependence structure of energy variables, and the inability to pinpoint the dominant energy type after energy consumption anomalies occur. Consequently, their practicality and guidance in industrial energy conservation monitoring scenarios remain limited. To address these technical shortcomings, this invention proposes an industrial energy conservation monitoring method based on a Copula graphical model, effectively improving the accuracy, interpretability, and decision support capabilities of industrial energy conservation monitoring in complex energy consumption coupling scenarios. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of the embodiments of the present invention and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section, as well as in the abstract and title of the present application, to avoid obscuring the purpose of this section, the abstract and title of the invention. Such simplifications or omissions shall not be used to limit the scope of the present invention.

[0007] In view of the shortcomings of existing industrial energy conservation monitoring technologies, such as insufficient ability to model the correlation of multi-source energy consumption data, difficulty in characterizing the nonlinear dependence between different energy types, and lack of effective location of the dominant abnormal energy type when energy consumption anomalies occur, this invention is proposed.

[0008] Therefore, the problem to be solved by this invention is how to construct a joint modeling mechanism that can reflect the coupling relationship of multiple energy sources in complex industrial energy consumption scenarios, realize the effective assessment and anomaly identification of energy consumption status, and provide interpretable technical support for energy conservation monitoring decisions.

[0009] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, embodiments of the present invention provide an industrial energy conservation monitoring method based on a Copula graph model, comprising, By deploying IoT smart monitoring terminals at industrial production sites, multiple types of raw energy consumption data are collected, and the raw energy consumption data is preprocessed to obtain multiple types of energy consumption time series sequences. The non-parametric density estimation method is used to fit the corresponding marginal probability distribution function of the multiple energy consumption time series, and the probability integral transformation is used to convert each type of energy consumption time series into the corresponding uniform distribution sequence to obtain multiple uniform distribution sequences. Using the aforementioned uniformly distributed sequences as node variables, a graph structure learning algorithm is employed to construct the topology of the Copula graph model, thereby obtaining the trained Copula graph model. The real-time energy consumption data collected during the monitoring period is converted into real-time uniform distribution values ​​and input into the trained Copula graph model to calculate the joint probability density value. The joint probability density value is compared with a preset energy-saving anomaly threshold, and the dominant energy type that causes the energy consumption anomaly is located based on the contribution of each node variable in the Copula graph model to the joint probability density value, and energy-saving monitoring and early warning information is output.

[0010] Secondly, embodiments of the present invention provide a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any step of the above-described industrial energy-saving monitoring method based on the Copula graph model.

[0011] Thirdly, embodiments of the present invention provide a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the above-described industrial energy-saving monitoring method based on the Copula graph model.

[0012] Compared with existing technologies, the advantages of this invention are as follows: By deploying IoT intelligent monitoring terminals to collect and preprocess multiple types of energy consumption time series, efficient integration and standardization of multi-source heterogeneous energy consumption data from industrial sites are achieved, overcoming the modeling difficulties caused by data dispersion and inconsistent formats in traditional monitoring; by using non-parametric density estimation to fit marginal distributions and performing probability integral transformation on various energy consumption time series, uniform distribution sequences are obtained, which not only avoids model biases that may be caused by parameterized distribution assumptions, but also eliminates the differences in the dimensions and scales of various energy consumption variables through normalization processing, enabling the dependence between different energy types to be measured in the same probability space, thereby enhancing the model's adaptability and robustness to complex industrial energy consumption scenarios; using uniform distribution sequences as node variables, a graph structure learning algorithm is used to construct the topology of the Copula graph model, this step is data-driven. Automatically identifying the conditional dependencies between various energy consumption variables and forming a sparse graph representation not only reduces model complexity but also reveals the implicit correlation network among multiple energy consumptions, giving the model good interpretability and structural generalization ability. Real-time energy consumption data is converted into uniformly distributed values ​​and input into the trained Copula graph model to calculate the joint probability density value. Leveraging the advantage of the Copula function in separating and modeling marginal distributions and dependency structures, accurate assessment of the joint probability of multi-dimensional energy consumption is achieved, enabling sensitive detection of subtle anomalies in the overall energy consumption pattern. By comparing the joint probability density value with a preset threshold and identifying the dominant abnormal energy type based on the contribution of each node variable to the joint probability, quantitative anomaly attribution analysis is provided. This shifts energy conservation monitoring from passive monitoring to proactive early warning and precise intervention, significantly improving the intelligence level and execution efficiency of industrial energy management. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of an industrial energy conservation monitoring method based on the Copula graph model. Detailed Implementation

[0014] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0015] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without inventive effort should fall within the scope of protection of this invention.

[0016] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0017] As mentioned in the background section, while existing techniques based on Copula functions or Copula models possess certain theoretical advantages in multivariate correlation modeling and joint distribution characterization, they primarily focus on macro-level resource optimization, quality process monitoring, or statistical runaway determination. They generally suffer from difficulties in directly adapting to multi-source energy consumption time-series data in industrial settings, a lack of refined modeling of the conditional dependence structure of energy variables, and the inability to pinpoint the dominant energy type after energy consumption anomalies occur. To address these issues, this invention provides an industrial energy conservation monitoring method based on a Copula graphical model.

[0018] Reference Figure 1 , Figure 1 This is a flowchart illustrating an industrial energy conservation monitoring method based on a Copula graph model according to an embodiment of the present invention. Figure 1 As shown, an industrial energy conservation monitoring method based on a Copula graph model includes: S1: Collect various types of raw energy consumption data through IoT smart monitoring terminals deployed at industrial production sites, and preprocess the various types of raw energy consumption data to obtain various types of energy consumption time series sequences. Specifically, IoT smart monitoring terminals are deployed at the power distribution cabinets, gas pipeline inlets, steam pipeline inlets, compressed air pipeline inlets, and hot and cold water pipeline inlets in industrial production sites, and a unified sampling period is set for the IoT smart monitoring terminals. The IoT smart monitoring terminals include electricity metering modules, gas flow metering modules, steam flow metering modules, compressed air flow metering modules, and hot and cold water flow metering modules. The value range of the unified sampling period is 1 minute to 15 minutes. The IoT smart monitoring terminals synchronously collect electricity consumption data, gas consumption data, steam consumption data, compressed air consumption data, and hot and cold water consumption data according to the unified sampling period, and collectively refer to the electricity consumption data, gas consumption data, steam consumption data, compressed air consumption data, and hot and cold water consumption data as multi-type energy consumption raw data.

[0019] Furthermore, missing value detection is performed on each type of raw energy consumption data in the multi-type raw energy consumption data, and an outlier detection method based on interquartile range is used to detect outliers. Specifically, when the number of consecutive missing data points detected is less than or equal to a preset missing threshold, a linear interpolation method is used to fill in the missing positions; when the number of consecutive missing data points detected is greater than the preset missing threshold, the corresponding data segment is marked as an invalid data segment and removed.

[0020] It should be noted that the preset missing threshold is set based on the continuity and stability requirements of specific industrial production processes. Typically, determining the preset missing threshold requires comprehensive consideration of equipment sampling cycles, the duration of process steps, and the autocorrelation characteristics of energy data. For example, in continuous process industries, if the stabilization time of a process segment is short, the threshold should be set relatively small (e.g., ≤3 sampling points) to avoid excessive interpolation masking of true fluctuations. For long-cycle, high-inertia production processes, the threshold can be appropriately relaxed (e.g., ≤10 sampling points). In practical applications, it is recommended to determine the threshold jointly through historical data analysis and domain expert experience to ensure data integrity while minimizing model bias introduced by data imputation.

[0021] Furthermore, the interquartile range is calculated based on the first and third quartiles; data points whose values ​​exceed the first quartile minus 1.5 times the interquartile range or exceed the third quartile plus 1.5 times the interquartile range are identified as outliers; these outliers are replaced by the mean of their adjacent valid data points to obtain the original energy consumption data for each category after outlier processing; the specific replacement rules are as follows: for a data point identified as an outlier... First, search forward and backward along time to find the first non-anomaly and valid observation, denoted as . and If valid values ​​are found in both directions, the arithmetic mean of the two values ​​is used for replacement. If only a valid value is found in one direction, that value is used directly for replacement. If no valid values ​​are found in either direction within the preset search window (e.g., 5 sampling points before and after), the data segment containing the outlier will be marked as invalid and removed, and the processing method is the same as for consecutive missing data segments.

[0022] Specifically, after outlier processing, the raw energy consumption data for each type is time-aligned according to the collection timestamp, forming an electricity consumption time series, a gas consumption time series, a steam consumption time series, a compressed air consumption time series, and a hot and cold water consumption time series.

[0023] Furthermore, the time series sequences of electricity consumption, gas consumption, steam consumption, compressed air consumption, and hot and cold water consumption are collectively referred to as multi-class energy consumption time series sequences. Each type of energy consumption time series sequence has the same time length and the same sampling time.

[0024] S2: For multiple types of energy consumption time series, the non-parametric density estimation method is used to fit the corresponding marginal probability distribution function. The probability integral transformation is used to convert each type of energy consumption time series into the corresponding uniform distribution sequence to obtain multiple types of uniform distribution sequences. Specifically, each type of energy consumption time series is extracted sequentially from multiple types of energy consumption time series as the target energy consumption time series, where the target energy consumption time series contains N energy consumption observations arranged in chronological order. .

[0025] Furthermore, a nonparametric density estimation method is used to fit the marginal probability density function to the target energy consumption time series, and the marginal probability density function is numerically integrated to obtain the marginal cumulative distribution function, specifically including: The Gaussian kernel function is selected as the kernel function, and the specific formula is as follows: ; in, The improved adaptive industrial energy consumption Gaussian kernel function takes the value at the standardized variable v, where v is the variable after bandwidth standardization. Let be the local adaptive bandwidth adjustment factor corresponding to the j-th energy consumption observation, and π be the mathematical constant pi with a value of approximately 3.14159. To use an initial fixed bandwidth for the j-th energy consumption observation The trial density value obtained by kernel density estimation. The geometric mean of all tested density values ​​is used for normalization. The sensitivity parameter has a value range of [0,1] and is set to 0.5 in this invention. n is the total number of energy consumption observations in the target energy consumption time series. Let j be the j-th energy consumption observation value arranged in chronological order in the target energy consumption time series; It should be noted that, The range of is (0, +∞), and it reaches its maximum value when v = 0. As |v| increases, the function value monotonically decreases and approaches 0. A larger function value indicates a higher contribution of the data point density at that position, while a smaller function value indicates a lower contribution of the data point density at that position.

[0026] The optimal bandwidth parameter for the kernel density estimation method is calculated using the Silverman empirical rule, and the specific formula is as follows: ; in, For optimal bandwidth parameters, Let be the sample standard deviation of the target energy consumption time series. The third quartile, or 75th percentile, of the target energy consumption time series is used. 1.34 represents the first quartile (25th percentile) of the target energy consumption time series, and 1.34 is the correction factor for converting the interquartile range into a standard deviation estimator. The first-order autocorrelation coefficient of the target energy consumption time series is used to measure the degree of linear correlation between energy consumption observations at adjacent time points. It should be noted that the optimal bandwidth parameter The range of is (0, +∞), and the actual value depends on the dispersion and temporal correlation of the energy consumption data; A larger value indicates a higher degree of smoothness in kernel density estimation, making it suitable for situations with large data fluctuations. A smaller value indicates that the kernel density estimation can capture more local details and is suitable for situations where the data distribution is relatively concentrated; The range of values ​​is [-1, 1], when When the value is close to 1, it indicates a strong positive correlation between adjacent energy consumption observations. When the value is close to -1, it indicates a strong negative correlation between adjacent energy consumption observations. When the value is close to 0, it indicates that there is no significant linear correlation between adjacent energy consumption observations.

[0027] Based on the Gaussian kernel function and the optimal bandwidth parameter, a kernel density estimation function for the target energy consumption time series is constructed, and the specific formula is as follows: ; in, This represents the probability density estimate of the time-weighted adaptive kernel density estimation function at point x. The energy consumption value is the probability density to be estimated. This is the time-series decay weight corresponding to the j-th energy consumption observation.

[0028] It should be noted that, The range of is (0,+∞), and its integral over the entire domain is 1 as a probability density function. A larger value indicates a higher probability density that the energy consumption observation falls near x, meaning that the energy consumption level is more likely to occur. The smaller the value, the lower the probability density of the energy consumption observation value falling near x, that is, the less likely that the energy consumption level will occur. The range of is (0,1). When j is close to n, A value close to 1 indicates that recent data is given higher weight; when j is much smaller than n, A value close to 0 indicates that the weight of early data gradually decreases.

[0029] Numerical integration of the kernel density estimation function from negative infinity to X yields the following formula for the marginal cumulative distribution function corresponding to the target energy consumption time series: ; in, The value of the time-weighted marginal cumulative distribution function at x represents the cumulative probability that the energy consumption observation is less than or equal to x. The cumulative distribution function of the standard normal distribution. Let the independent variable be a standard normal distribution. Let t be the error function and t be the integration variable.

[0030] It should be noted that, The range of x is (0,1), and as x approaches negative infinity, A value approaching 0 indicates that the current energy consumption value x is at the low end of the historical energy consumption distribution, meaning the energy consumption level is relatively low; when x approaches positive infinity, ... A value close to 1 indicates that the current energy consumption value x is at the high end of the historical energy consumption distribution, meaning that the energy consumption level is relatively high. A value close to 0.5 indicates that the current energy consumption value x is near the median of the historical energy consumption distribution, i.e., the energy consumption level is in a moderate state; this value range characteristic ensures that the subsequent probability integral transformation converts the energy consumption time series into a uniformly distributed series on the (0,1) interval.

[0031] Furthermore, based on each energy consumption observation in the target energy consumption time series, the energy consumption observation is substituted into the marginal cumulative distribution function to calculate the corresponding probability integral transformation value; all energy consumption observations in the target energy consumption time series are subjected to probability integral transformation to obtain all probability integral transformation values, and these are arranged in the original time order to form a uniform distribution sequence corresponding to the target energy consumption time series; the above operation is performed on each type of energy consumption time series in the multi-type energy consumption time series to obtain multi-type uniform distribution sequences, specifically including: repeating the above operation on the electricity energy consumption time series, gas energy consumption time series, steam energy consumption time series, compressed air energy consumption time series, and hot and cold water energy consumption time series in the multi-type energy consumption time series to obtain the electricity uniform distribution sequence, gas uniform distribution sequence, steam uniform distribution sequence, compressed air uniform distribution sequence, and hot and cold water uniform distribution sequence, and collectively referred to as the multi-type uniform distribution sequence.

[0032] S3: Using multiple uniformly distributed sequences as node variables, a graph structure learning algorithm is used to construct the topology of the Copula graph model, resulting in a trained Copula graph model. Specifically, each uniformly distributed sequence in the multi-class uniformly distributed sequence is defined as a node variable. And combine all node variables into a node variable set. ; set of node variables The mutual information value between any two node variables is calculated using a rank-correlation-based mutual information estimation method, as shown in the following formula: ; in, For the p-th uniformly distributed sequence With the qth class uniformly distributed sequence Mutual information value between them Let be the numerical stability constant, and let its value be . This is to prevent division by zero errors. The differential entropy estimate of the p-th uniformly distributed sequence is given. This represents the total number of data points in a uniformly distributed sequence. For the i-th observation of a uniformly distributed sequence of class p... The kernel density estimate at that location, Let be the observation value of the uniformly distributed sequence of class p at time i.

[0033] Furthermore, traverse the set of node variables. By pairing and combining all node variables, we obtain the mutual information matrix M, as shown in the following formula: ; Where M is the mutual information matrix, Let be the element value in the p-th row and q-th column of the normalized mutual information matrix, and D be the total number of node variables. In this invention, D is 5, corresponding to five types of energy consumption: electricity, gas, steam, compressed air, and hot and cold water. Let be the self-mutual information, or information entropy, of the p-th uniformly distributed sequence. Let be the self-mutual information of the uniformly distributed sequence of class q.

[0034] It should be noted that, The range of is [0,1]. When mpq=0, it indicates that the p-th type of energy consumption and the q-th type of energy consumption are completely independent and have no information sharing. ... When =1, it indicates that there is a complete dependency relationship between the p-th type of energy consumption and the q-th type of energy consumption or that they are the same variable; A larger value indicates a stronger dependence between the two types of energy consumption, and these should be prioritized for connection when constructing a Copula graph model; the mutual information matrix M is a 5×5 symmetric matrix, and the element in the i-th row and j-th column of the mutual information matrix M represents the variables of the two nodes. The mutual information values ​​between them.

[0035] Preferably, the maximum weight spanning tree algorithm is the Chow-Liu algorithm, which uses the mutual information value as the weight value of the edge connecting the corresponding two node variables. The edge with the largest weight value is selected from the mutual information matrix M and added to the topology in turn. When the newly added edge forms a loop with the existing edge, the edge is skipped until the topology connects all node variables. Furthermore, based on the mutual information matrix M, the topology of the Copula graph model is constructed using the maximum weight spanning tree algorithm. Considering the prior domain knowledge constraints in industrial energy consumption systems, a domain constraint penalty term and edge stability evaluation are introduced to construct a constrained maximum weight spanning tree. ; in, This represents the optimal constraint tree structure, i.e., the topological structure of the Copula graph model. The set of all candidate topologies that satisfy the tree structure constraints. Let T be the set of all edges in the tree structure T. Let be the stability coefficient of the edge connecting the p-th node and the q-th node. This is the domain constraint penalty coefficient, with a value ranging from [0,1]. In this invention, it is set to 0.3. As an indicator function, when the edge If the edge belongs to the forbidden edge set, the value is 1; otherwise, the value is 0. This is a set of forbidden edges defined based on domain knowledge, containing pairs of energy-consuming nodes that should not be directly connected.

[0036] It should be noted that the stability coefficient of an edge can be obtained by constructing a mutual information matrix for different time periods (e.g., by day, by week) of historical data, calculating the reciprocal of the coefficient of variation of the mutual information value of the edge on these subsets, and then normalizing it to the interval [0, 1]. The closer the value is to 1, the more stable the dependency. (Note: The last sentence appears to be incomplete and possibly refers to a separate section about edge sets.) : Based on predefined industrial process knowledge; for example, energy pairs that are physically unrelated and whose historical data shows extremely low correlation (such as long-distance lighting electricity and a certain air compressor) can be included, and the algorithm is prohibited from directly connecting edges between them; domain constraint penalty coefficient. The strength of the domain knowledge constraint needs to be determined by cross-validation within the range of [0, 1]. The example value of 0.3 in the original text can be used as the starting point.

[0037] Specifically, for each edge in the topology, based on the variables of the two nodes connected by the edge... For the tail dependency features of the corresponding uniformly distributed sequence data, a binary Copula function is selected, and the dependency parameters of the binary Copula function are determined using the maximum likelihood estimation method. The binary Copula functions include the Gumbel Copula function, the Clayton Copula function, and the Frank Copula function, and the specific formulas are as follows: ; in, Let be the adaptive hybrid binary Copula function between the node variables of the uniformly distributed sequences of class p and class q. For the observed values ​​of the uniformly distributed sequence of class p, For the observations of the q-th uniformly distributed sequence, For the parameter set of the mixed Copula function, For Gumbel Copula's adaptive blending weights, For the GumbelCopula function, This is a dependency parameter of Gumbel Copula, and its value range is (1, +∞). For Clayton Copula functions, This is a dependency parameter of Clayton Copula and its value ranges from (-1, +∞) to {0}. For Frank Copula functions, This is a dependency parameter of Frank Copula and its value range is (-∞, +∞). It should be noted that, The range of is [0,1], representing the value of the joint cumulative distribution function of two uniformly distributed variables. =0 indicates that at least one variable takes the minimum value of 0. =1 indicates that both variables take the maximum value of 1; the range of φ is [0,1]. When the value is close to 1, it indicates that the two types of energy consumption mainly exhibit upper-tail dependence, meaning that the correlation is stronger under high-energy-consumption conditions. When the value is close to 0, it indicates that the two types of energy consumption mainly exhibit tail dependence or symmetric dependence.

[0038] Furthermore, if two node variables If the edge exhibits an upper-tail dependency, then the Gumbel Copula function is chosen as the binary Copula function for the edge, as shown in the following formula: ; If two node variables exhibit a lower-tail dependency, then the Clayton Copula function is selected as the binary Copula function for the edge, as shown in the following formula: ; If two node variables exhibit symmetric dependency, then the Frank Copula function is chosen as the binary Copula function for the edge, as shown in the following formula: .

[0039] It should be noted that the maximum likelihood estimate of the dependent parameters is obtained by maximizing the log-likelihood function, as shown in the following formula: ; in, This is the value of the regularized time-weighted log-likelihood function. Let be the temporal weight corresponding to the observation at time i. For the mixed Copula density function, Let be the observation value of the uniformly distributed sequence of class p at time i. Let be the observation value of the uniformly distributed sequence of class q at time i. For Gumbel Copula's dependency parameters, For Clayton Copula's dependency parameters, These are the dependency parameters for Frank Copula.

[0040] Preferred, The range of is (-∞, +∞), and the larger the value, the better the model fits the data under the current parameters.

[0041] Specifically, the Akaike Information Criterion is used to evaluate the goodness of fit of the selected binary Copula function for each edge in the topology, and the initial topology that passes the goodness of fit evaluation is determined as the final topology of the Copula graph model.

[0042] Furthermore, the specific formula for the Akaike Information Criterion is as follows: ; in, Let Akaike information criterion value be the edge corresponding to the uniformly distributed sequence of class p and class q. To achieve optimal parameters The regularized time-weighted log-likelihood function value at [location]. denoted as the number of effective parameters in the hybrid Copula model between nodes of class p and class q, where p is the parameter deviation penalty coefficient, used to control the intensity of the penalty when the parameter deviates from the prior expectation, and its value ranges from [0,1]. In this invention, it is set to 0.1. For the dominant dependency parameter based on mixed weights, The reference dependency parameter value is calculated based on the rank correlation coefficient and is used as a prior expectation.

[0043] Preferred, The value range is (-∞, +∞). The smaller the value, the better the model achieves a balance between goodness of fit and complexity, i.e., the better the model selection. When the value of the Akaike Information Criterion exceeds the preset fitting threshold, it indicates that the currently selected Copula function type is not suitable for the dependency structure of this edge, and it is necessary to reselect other types of binary Copula functions and estimate the dependency parameters. The preset fitting threshold TAIC is set based on the statistical characteristics of historical energy consumption data, and is usually taken as the 75th percentile of all edge AIC values ​​plus 1.5 times the interquartile range.

[0044] Furthermore, the final topology, the type of the binary Copula function corresponding to each edge in the final topology, and the combination of dependent parameters of the binary Copula function are encapsulated to obtain the trained Copula graph model.

[0045] S4: Convert the real-time energy consumption data collected during the monitoring period into real-time uniform distribution values, input them into the trained Copula graph model, and calculate the joint probability density value. Specifically, during the monitoring period, real-time energy consumption data is collected through IoT smart monitoring terminals, and the validity of the real-time energy consumption data is verified. The validity verification includes: determining whether the values ​​of various data in the real-time energy consumption data are within the reasonable range of the corresponding energy consumption type; if there is data in the real-time energy consumption data that exceeds the reasonable range, such data is marked as invalid data and replaced with valid data from the previous sampling time.

[0046] It should be noted that real-time energy consumption data includes real-time electricity consumption data, real-time gas consumption data, real-time steam consumption data, real-time compressed air consumption data, and real-time hot and cold water consumption data.

[0047] Furthermore, the real-time energy consumption data after verification is transformed by probability integral transformation using the marginal probability distribution function to obtain the real-time uniform distribution values ​​corresponding to various types of real-time energy consumption data, forming a real-time uniform distribution vector.

[0048] Furthermore, the real-time uniformly distributed vector is input into the trained Copula graph model. Based on the final topological structure of the Copula graph model, the entire edge set is extracted. Based on the tree-like topological decomposition characteristics of the Copula graph model, the joint probability density value corresponding to the real-time uniformly distributed vector is calculated, as shown in the following formula: ; in, Let be the joint probability density value corresponding to the real-time uniform distribution vector t at time t, and D be the total number of node variables. The improved time-weighted kernel density estimation function for the p-th type of energy consumption at real-time effective value The probability density estimate at that location, The set of all edges in the final topology T of the trained Copula graph model. For the edge connecting the p-th node and the q-th node, Let be the improved adaptive hybrid binary Copula density function corresponding to edge (p,q). Let be the real-time uniform distribution value of the p-th type of real-time energy consumption at time t. Let be the real-time uniform distribution value of the q-th type of real-time energy consumption at time t. is the confidence enhancement factor for edge (p,q), used to strengthen the contribution of high-confidence edges.

[0049] It should be noted that, The value range is (0, +∞). This value reflects the probability density level of the current real-time energy consumption status in the historical joint distribution. The larger the value, the more the current energy consumption combination mode conforms to the joint distribution characteristics of the historical normal operation status, that is, the energy consumption status is normal. The smaller the value, the more the current energy consumption combination mode deviates from the joint distribution characteristics of the historical normal operation status, that is, there is a risk of abnormal energy consumption.

[0050] S5: Compare the joint probability density value with the preset energy-saving anomaly threshold, and based on the contribution of each node variable in the Copula graph model to the joint probability density value, locate the dominant energy type that causes energy consumption anomalies and output energy-saving monitoring and early warning information.

[0051] Furthermore, the local anomaly contribution of each node variable in the Copula graph model is normalized to obtain the anomaly contribution ratio of each node variable, specifically including: Based on the trained Copula graph model, an empirical distribution is constructed using the joint probability density values ​​corresponding to all samples in the historical training dataset. The α percentile of the empirical distribution is selected as the preset energy-saving anomaly threshold, where the value of α ranges from 1 to 5. The preset energy-saving anomaly threshold represents the lower bound of the joint probability density value under normal energy consumption conditions. When the joint probability density value is greater than or equal to the preset energy-saving anomaly threshold, the current energy consumption status is determined to be normal, and routine monitoring and recording operations are performed. Specifically, key information such as the real-time energy consumption data, the calculated joint probability density value, and the contribution analysis results of each energy type will be archived and stored in the historical database for updating the model training sample set or conducting long-term energy efficiency trend analysis. Simultaneously, the system maintains real-time monitoring without actively intervening in the production process, but can provide operators with feedback through the energy management system's monitoring interface indicating that the current operation is within the normal energy efficiency range, thus maintaining monitoring transparency. When the joint probability density value is less than the preset energy-saving anomaly threshold, the current energy consumption status is determined to be abnormal, and a multi-level early warning and diagnostic response mechanism is immediately activated. Specifically, based on the identified dominant energy types and their abnormal contribution ratios, structured energy-saving monitoring and early warning information is generated. This information not only includes basic data such as early warning timestamps, abnormal joint probability density values, and preset thresholds, but also highlights the suspected abnormal dominant energy sources (such as electricity and gas) and their real-time energy consumption values. This early warning information is sent in real time to the monitoring terminals of energy management personnel, workshop operation supervisors, and relevant monitoring departments through preset channels (such as audible and visual alarms, SMS, and platform message push) to ensure timely delivery of alerts.

[0052] Specifically, regarding the final topology of the Copula graph model and the node variables... For all connected edges, extract the corresponding binary Copula density function values, calculate the geometric mean, and use the negative logarithm of the geometric mean as the node variable. Local anomaly contribution The abnormal contribution percentages are sorted in descending order of numerical value. The node variable corresponding to the highest abnormal contribution percentage is selected as the dominant abnormal node variable. The dominant energy type is then determined, specifically: if the dominant abnormal node variable is V1, the dominant energy type is electricity; if the dominant abnormal node variable is V2, the dominant energy type is natural gas; if the dominant abnormal node variable is V3, the dominant energy type is steam; if the dominant abnormal node variable is V4, the dominant energy type is compressed air; and if the dominant abnormal node variable is V5, the dominant energy type is hot or cold water.

[0053] Furthermore, after determining the dominant energy type, the corresponding specific anomaly diagnosis and handling logic is executed, including: If the dominant energy type is determined to be electricity, a special power analysis is triggered: extract the real-time power consumption time series and calculate its dynamic time warping distance with the historical benchmark curve for the same period or under the same operating conditions; simultaneously retrieve the monitoring data of the power distribution system, including the power factor, three-phase imbalance, and harmonic distortion rate of each circuit; based on the dependency structure and parameters of the edges connected to the power node V1 in the Copula graph model, analyze whether the real-time data of the associated energy type (such as compressed air, hot and cold water) shows a coordinated abnormal mode; based on the above information, generate a potential root cause diagnosis report for abnormal motor load, unplanned lighting system activation, or power compensation device failure, and push it to the electrical maintenance terminal; If the dominant energy type is determined to be natural gas, a special analysis of natural gas is triggered: extract the real-time natural gas energy consumption time series, analyze its instantaneous flow fluctuation characteristics and combustion efficiency estimates; combine the steam energy consumption time series and the tail dependency characteristics of the edge connecting natural gas node V2 and steam node V3 in the Copula graph model to determine whether the fuel-output coupling relationship of the boiler or heating furnace deviates from the historical normal pattern; correlate with process monitoring data such as flue gas oxygen content and furnace temperature (if available) to generate potential root cause diagnostic reports for natural gas valve failure, burner air distribution imbalance or heat exchange efficiency decline, and push them to the thermal workshop operation terminal; If steam is determined to be the dominant energy source, a steam-specific analysis is triggered: extract the real-time steam energy consumption time series, distinguish its flow and pressure data, and calculate the pipeline pressure drop anomaly index; based on the dependency relationship between steam node V3 and power node V1 and gas node V2 in the Copula graph model, check whether the electrically driven feedwater pump or the gas-driven steam generation equipment is operating in coordination; combined with the process temperature parameters at the steam user end, generate potential root cause diagnostic reports for steam leakage, steam trap failure, or scaling of heat exchange equipment, and push them to the equipment management terminal; If the dominant energy type is determined to be compressed air, a special compressed air analysis is triggered: extract the real-time compressed air energy consumption time series and analyze the dynamic relationship between its unit time flow rate and pipeline pressure; based on the strong dependency edge between compressed air node V4 and power node V1 in the Copula graph model, associate the total current and load rate data of the air compressor unit; calculate the estimated value of system leakage rate and associate it with the start and stop signals of the main air-consuming equipment to generate a potential root cause diagnosis report for pipeline leakage, insufficient dryer efficiency or unreasonable pressure setting, and push it to the power workshop monitoring terminal; If the dominant energy type is determined to be hot and cold water, a special analysis of hot and cold water is triggered: extract the real-time hot and cold water energy consumption time series, and distinguish between supply and return water temperature difference and flow data; based on the dependency relationship between hot and cold water node V5 and power node V1 in the Copula graph model, associate the energy consumption data of the chiller unit or circulating pump; combine the ambient temperature data and process cooling requirements to generate a potential root cause diagnosis report for low cooling tower efficiency, uneven chilled water flow distribution, or temperature control valve failure, and push it to the HVAC system management terminal.

[0054] Furthermore, energy conservation monitoring and early warning information is generated and pushed to the monitoring terminal of the energy management system for display and stored in the energy conservation monitoring database. The energy conservation monitoring and early warning information includes early warning timestamp, joint probability density value, preset energy conservation anomaly threshold, dominant energy type, the anomaly contribution ratio corresponding to the dominant energy type, and real-time energy consumption data corresponding to the dominant energy type.

[0055] In summary, this invention achieves efficient integration and standardization of multi-source heterogeneous energy consumption data from industrial sites by deploying IoT intelligent monitoring terminals to collect and preprocess various types of energy consumption time series. This overcomes the modeling difficulties caused by data dispersion and inconsistent formats in traditional monitoring. By using nonparametric density estimation to fit marginal distributions and performing probability integral transformation on various energy consumption time series, a uniform distribution sequence is obtained. This not only avoids model biases that may be caused by parameterized distribution assumptions but also eliminates differences in the dimensions and scales of various energy consumption variables through normalization processing. This allows the dependencies between different energy types to be measured in the same probability space, thereby enhancing the model's adaptability and robustness to complex industrial energy consumption scenarios. Using the uniform distribution sequence as node variables, a graph structure learning algorithm is used to construct the topology of the Copula graph model. This step automatically identifies each node in a data-driven manner. The conditional dependencies between energy consumption variables are represented in a sparse graph, which not only reduces model complexity but also reveals the implicit network of relationships among multiple energy consumptions, giving the model good interpretability and structural generalization ability. Real-time energy consumption data is converted into uniformly distributed values ​​and input into the trained Copula graph model to calculate the joint probability density value. By leveraging the advantage of the Copula function in separating and modeling marginal distributions and dependency structures, accurate assessment of the joint probability of multidimensional energy consumption is achieved, thus enabling sensitive detection of subtle anomalies in the overall energy consumption pattern. By comparing the joint probability density value with a preset threshold and identifying the dominant abnormal energy type based on the contribution of each node variable to the joint probability, quantitative anomaly attribution analysis is provided, enabling energy conservation monitoring to shift from passive monitoring to proactive early warning and precise intervention, significantly improving the intelligence level and execution efficiency of industrial energy management.

[0056] This embodiment also provides a computer device applicable to the industrial energy-saving monitoring method based on the Copula graph model, including a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the industrial energy-saving monitoring method based on the Copula graph model proposed in the above embodiment.

[0057] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0058] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the industrial energy-saving monitoring method based on the Copula graph model as proposed in the above embodiments.

[0059] The storage medium proposed in this embodiment and the data storage method proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0060] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An industrial energy conservation monitoring method based on a Copula graph model, characterized in that: include, By deploying IoT smart monitoring terminals at industrial production sites, multiple types of raw energy consumption data are collected, and the raw energy consumption data is preprocessed to obtain multiple types of energy consumption time series sequences. The non-parametric density estimation method is used to fit the corresponding marginal probability distribution function of the multiple energy consumption time series, and the probability integral transformation is used to convert each type of energy consumption time series into the corresponding uniform distribution sequence to obtain multiple uniform distribution sequences. Using the aforementioned uniformly distributed sequences as node variables, a graph structure learning algorithm is employed to construct the topology of the Copula graph model, thereby obtaining the trained Copula graph model. The real-time energy consumption data collected during the monitoring period is converted into real-time uniform distribution values ​​and input into the trained Copula graph model to calculate the joint probability density value. The joint probability density value is compared with a preset energy-saving anomaly threshold, and the dominant energy type that causes the energy consumption anomaly is located based on the contribution of each node variable in the Copula graph model to the joint probability density value, and energy-saving monitoring and early warning information is output.

2. The industrial energy conservation monitoring method based on the Copula graph model as described in claim 1, characterized in that: The method for outputting the energy-saving monitoring and early warning information is as follows: The local anomaly contribution of each node variable in the Copula graph model is normalized to obtain the anomaly contribution ratio of each node variable. The abnormal contribution percentages are sorted in descending order of numerical value. The node variable corresponding to the abnormal contribution percentage at the top of the sorted list is selected as the dominant abnormal node variable to determine the dominant energy type. Energy conservation monitoring and early warning information is generated and pushed to the monitoring terminal of the energy management system for display and stored in the energy conservation monitoring database.

3. The industrial energy conservation monitoring method based on the Copula graph model as described in claim 1, characterized in that: The joint probability density value is compared with a preset energy-saving anomaly threshold, including: Based on the trained Copula graph model, an empirical distribution is constructed using the joint probability density values ​​of all samples in the historical training dataset, and the α percentile of the empirical distribution is selected as the preset energy-saving anomaly threshold. When the joint probability density value is greater than or equal to the preset energy-saving anomaly threshold, the energy consumption status at the current moment is determined to be normal, and routine monitoring and recording operations are performed. When the joint probability density value is less than the preset energy-saving anomaly threshold, the energy consumption status at the current moment is determined to be abnormal, and a multi-level early warning and diagnostic response mechanism is immediately activated.

4. The industrial energy conservation monitoring method based on the Copula graph model as described in claim 3, characterized in that: The method for obtaining the joint probability density value is as follows: During the monitoring period, real-time energy consumption data is collected through IoT smart monitoring terminals, and the validity of the real-time energy consumption data is verified. By performing probability integral transformation on the verified real-time energy consumption data using the marginal probability distribution function, the real-time uniform distribution values ​​corresponding to various types of real-time energy consumption data are obtained, forming a real-time uniform distribution vector. The real-time uniformly distributed vector is input into the trained Copula graph model, and the entire edge set is extracted based on the final topological structure of the Copula graph model. Based on the tree-like topological decomposition characteristics of the Copula graph model, the joint probability density value corresponding to the real-time uniform distribution vector is calculated.

5. The industrial energy conservation monitoring method based on the Copula graph model as described in claim 4, characterized in that: The trained Copula graph model includes: Each uniformly distributed sequence in the multi-class uniformly distributed sequence is defined as a node variable. And all the node variables are combined into a node variable set. ; For the set of node variables The mutual information value between any two node variables is calculated using a rank-correlation-based mutual information estimation method. Traverse the set of node variables All node variables are paired and combined to obtain the mutual information matrix M; Based on the mutual information matrix M, the topology of the Copula graph model is constructed using the maximum weight spanning tree algorithm; For each edge in the topology, based on the variables of the two nodes connected by the edge... Based on the tail dependency features of the corresponding uniformly distributed sequence data, a binary Copula function is selected, and the dependency parameters of the binary Copula function are determined using the maximum likelihood estimation method. The Akaike Information Criterion is used to evaluate the goodness of fit of the binary Copula function selected for each edge in the topology, and the initial topology that passes the goodness of fit evaluation is determined as the final topology of the Copula graph model. The final topology, the binary Copula function type corresponding to each edge in the final topology, and the dependent parameters of the binary Copula function are encapsulated to obtain the trained Copula graph model.

6. The industrial energy conservation monitoring method based on the Copula graph model as described in claim 5, characterized in that: The maximum weight spanning tree algorithm is the Chow-Liu algorithm, which uses the mutual information value as the weight value of the edge connecting the corresponding two node variables. The edge with the largest weight value is selected from the mutual information matrix M and added to the topology in turn. When the newly added edge forms a loop with the existing edge, the edge is skipped until the topology connects all node variables.

7. The industrial energy conservation monitoring method based on the Copula graph model as described in claim 5, characterized in that: The method for obtaining the multiple uniformly distributed sequences is as follows: Each type of energy consumption time series is extracted sequentially from multiple types of energy consumption time series as a target energy consumption time series, wherein the target energy consumption time series contains N energy consumption observations arranged in chronological order. ; The marginal probability density function is fitted to the target energy consumption time series using a nonparametric density estimation method, and the marginal probability density function is numerically integrated to obtain the marginal cumulative distribution function; Based on each energy consumption observation in the target energy consumption time series, the energy consumption observation is substituted into the marginal cumulative distribution function to calculate the corresponding probability integral transform value; All energy consumption observations in the target energy consumption time series are subjected to probability integral transformation to obtain all probability integral transformation values, and then formed into a uniform distribution sequence corresponding to the target energy consumption time series according to the original time order. The above operation is performed on each of the multiple energy consumption time series sequences to obtain multiple uniformly distributed sequences.

8. The industrial energy conservation monitoring method based on the Copula graph model as described in claim 7, characterized in that: The method for obtaining the marginal cumulative distribution function is as follows: The Gaussian kernel function is selected as the kernel function; The optimal bandwidth parameter for the kernel density estimation method is calculated using the Silverman empirical rule. Based on the Gaussian kernel function and the optimal bandwidth parameter, a kernel density estimation function for the target energy consumption time series is constructed. The kernel density estimation function is numerically integrated from negative infinity to X to obtain the marginal cumulative distribution function corresponding to the target energy consumption time series.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the industrial energy-saving monitoring method based on the Copula graph model as described in any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the industrial energy-saving monitoring method based on the Copula graph model as described in any one of claims 1 to 8.

Images