A zero-carbon equipment failure early warning method and system based on ESG scheduling

Abstract

The present invention relates to the technical field of zero-carbon equipment fault warning, and discloses a zero-carbon equipment fault warning method and system based on ESG scheduling. The method collects multi-source ESG parameter data such as carbon footprint intensity, clean energy consumption, and energy storage system charging and discharging efficiency, and uses a graph convolutional network combined with a gated cyclic unit to extract equipment state characteristics, calculates ESG correlation and time-segment cross-correlation matrix, and generates fault warning weights through a ReLU activation function after dimensionality reduction through principal component analysis. The benchmark threshold is dynamically adjusted based on the weight and the deviation is corrected with the help of an extended Kalman filter to generate warning parameters to trigger the warning process. At the same time, a multi-dimensional equipment state space is constructed to determine the fault state, and a joint optimization model of ESG parameters and warning thresholds is solved by a non-dominated sorting genetic algorithm to achieve coordinated optimization of ESG benefits and fault warnings, thereby improving warning accuracy and dynamic adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of zero-carbon equipment failure early warning, and in particular to a zero-carbon equipment failure early warning method and system based on ESG scheduling. Background Art

[0002] As the world strives to achieve carbon neutrality, the safe and stable operation of zero-carbon equipment, a key enabler for energy transformation and sustainable development, is crucial. With the widespread application of zero-carbon equipment in energy systems, traditional fault warning methods face numerous challenges.

[0003] Traditional fault warning systems often focus solely on the device's operating parameters, such as physical indicators like temperature, pressure, and vibration, while overlooking the environmental, social, and governance implications of zero-carbon equipment operation. The operation of zero-carbon equipment not only involves its own performance but is also closely linked to ESG parameters such as carbon footprint intensity, clean energy consumption, and energy storage system charge and discharge efficiency. These ESG parameters reflect the device's performance in terms of environmental sustainability, social responsibility, and governance structure, providing valuable insights into equipment fault warnings.

[0004] Furthermore, zero-carbon devices are often embedded in complex energy networks, where the interconnections and impacts between devices are even more pronounced. Traditional methods struggle to effectively capture these interconnections, resulting in inaccurate and timely early warnings. Furthermore, the operating conditions of zero-carbon devices are complex and volatile, influenced by a variety of factors, including energy supply, load demand, and weather conditions. Traditional fixed-threshold early warning methods are unable to adapt to these dynamic demands.

[0005] With the advancement of intelligent and digital energy systems, the operational data generated by zero-carbon equipment is multi-source, massive, and highly dimensional. Extracting effective feature information from this complex data to accurately predict equipment failures has become a pressing challenge. Existing data processing and analysis methods struggle with incomplete feature extraction and low information utilization when dealing with high-dimensional, multi-source, and heterogeneous data, making them inadequate for zero-carbon equipment failure warnings.

[0006] Furthermore, during the operation of zero-carbon equipment, ESG benefits and equipment failure warnings need to be synergistically optimized. Traditional approaches often separate the two, failing to maximize ESG benefits while ensuring safe equipment operation. Therefore, a fault warning method that comprehensively considers ESG parameters, equipment dependencies, and dynamic operating conditions is urgently needed to improve the safety and sustainability of zero-carbon equipment operations. Summary of the Invention

[0007] The purpose of the present invention is to provide a zero-carbon equipment failure early warning method and system based on ESG scheduling to solve the problems raised in the above background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a zero-carbon equipment failure early warning method based on ESG scheduling, the method comprising:

[0009] Collect multi-source ESG parameter data during the operation of zero-carbon equipment, where the ESG parameter data includes carbon footprint intensity, clean energy consumption, and energy storage system charging and discharging efficiency;

[0010] A graph convolutional network is used to extract device status features from ESG parameter data to obtain a state feature matrix. The ESG correlation is then calculated based on the covariance change of the state feature matrix.

[0011] The cross-correlation matrix of ESG correlation in different time periods is subjected to feature dimensionality reduction to extract the core feature vector. The core feature vector is then nonlinearly mapped using the ReLU activation function to obtain the fault warning weight for each time period.

[0012] The baseline threshold of the state feature matrix is ​​dynamically adjusted according to the fault warning weight to generate warning parameters, and the equipment fault warning process is triggered based on the warning parameters.

[0013] Preferably, the device status feature extraction includes:

[0014] Intercept ESG parameter data at a set sampling interval, construct a state tensor and perform normalization;

[0015] A multi-head graph convolutional layer is used to aggregate cross-device features of the state tensor and output the dot product sum of the adjacency matrix and the feature vector;

[0016] The aggregation results are input into the gated recurrent unit, multi-time scale features are extracted through the update gate and reset gate, and the feature dimensions are filtered through the attention mask to generate a state feature matrix.

[0017] Preferably, the calculating of ESG relevance includes:

[0018] Select the state feature matrix of any time period as the benchmark threshold and calculate the covariance value of the feature vectors of its adjacent k time periods;

[0019] The ratio of the mean of the covariance value to the Euclidean distance of the benchmark threshold is used as the correlation quantity within the time period;

[0020] Calculate the correlation coefficient between the benchmark threshold and the characteristics of the associated time period, and take its absolute value as the correlation coefficient between time periods;

[0021] The harmonic mean of the intra-period correlation and the inter-period correlation is taken as the ESG correlation.

[0022] Preferably, the extracting of the core feature vector includes:

[0023] Construct a cross-correlation matrix of ESG correlations in different time periods and perform principal component analysis on it to obtain orthogonal eigenvectors;

[0024] The eigenvectors whose principal component variance contribution rate exceeds the set threshold are selected to form the core feature subspace;

[0025] The cross-correlation matrix is ​​mapped to the core feature subspace to obtain the core feature vector after dimensionality reduction.

[0026] Preferably, obtaining the fault warning weight of each time period includes:

[0027] Normalize the core feature vector and calculate its cosine similarity with the preset reference vector;

[0028] The similarity value is input into the bidirectional long short-term memory network, and the initial weight is generated after transformation by the forward layer and the backward layer;

[0029] The initial weights are smoothed by moving average filtering, and the fault warning weights for each time period are output.

[0030] Preferably, the generating of warning parameters includes:

[0031] Perform matrix multiplication of the fault warning weight and the benchmark threshold to obtain the weight adjustment parameter;

[0032] Calculate the deviation between the weight adjustment parameter and the reference threshold, and dynamically correct the deviation through the extended Kalman filter;

[0033] The corrected deviation is added to the baseline threshold to generate the warning parameter.

[0034] Preferably, the system further comprises:

[0035] Construct a multi-dimensional equipment state space based on warning parameters and extract the extreme points and covariance mutation areas of the space;

[0036] When the density of extreme points exceeds the set threshold or the span of the covariance mutation area is greater than the limit value, it is determined to be a fault warning state and a multi-level warning instruction is generated.

[0037] Preferably, the multi-dimensional device state space construction includes:

[0038] Map the warning parameters to a multi-dimensional coordinate system according to the operating time sequence to generate a state distribution point set;

[0039] The kernel density estimation method is used to reconstruct the state distribution point set spatially, and the covariance gradient and kurtosis coefficient of the reconstructed space are calculated;

[0040] The noise of gradient distribution is suppressed by Gaussian filtering algorithm to eliminate random interference.

[0041] Preferably, the system further comprises:

[0042] A joint optimization model of ESG parameter data and equipment warning thresholds is established, and a non-dominated sorting genetic algorithm is used to solve the optimal warning strategy;

[0043] The optimal early warning strategy and early warning parameters are coupled in real time to generate a fault early warning plan that maximizes ESG benefits;

[0044] The joint optimization model solution includes:

[0045] The objective function is defined as the sum of the weighted absolute values ​​of the ESG indicator deviation and the early warning missed rate, and the constraint condition is the equipment operation safety range;

[0046] Perform Pareto optimal transformation on the objective function and decompose it into the warning threshold sub-problem and the ESG scheduling sub-problem;

[0047] The two sub-problems are solved alternately and iteratively until convergence, and the optimal early warning strategy that meets the constraints is output.

[0048] Preferably, the present invention further includes a zero-carbon equipment failure early warning system based on ESG scheduling, the system comprising:

[0049] A multi-source data acquisition module is used to obtain real-time data on carbon footprint intensity, clean energy consumption, and energy storage system charge and discharge efficiency during operation, and to construct a multi-dimensional ESG parameter time series matrix;

[0050] The state feature extraction module uses a graph convolutional network to aggregate cross-device features of the multi-dimensional ESG parameter time series matrix, captures multi-timescale device state features through a gated recurrent unit, and outputs a state feature matrix;

[0051] The ESG correlation calculation module generates intra-period and inter-period correlation quantities based on the covariance changes of the state characteristic matrix, and calculates their harmonic mean as the ESG correlation degree;

[0052] The cross-correlation matrix decomposition module constructs a cross-correlation matrix for ESG correlations in different time periods and extracts core eigenvectors through principal component analysis;

[0053] The weight mapping module uses the ReLU activation function to perform nonlinear mapping on the core feature vector to generate the fault warning weight for each time period;

[0054] The parameter optimization module dynamically adjusts the benchmark threshold based on the fault warning weight, corrects the parameter deviation through the extended Kalman filter, and generates warning parameters;

[0055] The real-time early warning module converts the early warning parameters into equipment failure early warning signals, which are transmitted to the monitoring platform through the industrial Internet to trigger the early warning instructions.

[0056] Compared with the prior art, the present invention has the following beneficial effects:

[0057] This method collects multi-source ESG parameter data such as carbon footprint intensity, clean energy consumption, and energy storage system charging and discharging efficiency during the operation of zero-carbon equipment. It breaks through the limitation of traditional fault warning that only focuses on the physical parameters of the equipment, and incorporates environmental, social and governance factors into the early warning system, making the early warning more comprehensive and scientific, and able to more accurately reflect the actual operating status and potential risks of zero-carbon equipment.

[0058] In terms of data processing, a graph convolutional network is used to extract device status features from ESG parameter data. This effectively captures the correlations between devices and, through cross-device feature aggregation, provides more comprehensive device status information. This is combined with a gated recurrent unit to extract multi-timescale features, and an attention mask is used to filter feature dimensions to generate a state feature matrix. This significantly improves the efficiency and accuracy of feature extraction, laying a solid foundation for subsequent fault warnings.

[0059] By calculating ESG correlation and comprehensively considering both intra-period and inter-period correlations, we can deeply analyze the inherent connections between ESG parameters across different time periods and better understand the changing trends in equipment operating status. We perform feature dimensionality reduction on the cross-correlation matrix, extract the core eigenvectors, and use the ReLU activation function for nonlinear mapping to derive fault warning weights. This allows for dynamic adjustment of warning weights, ensuring that warnings are more consistent with the actual equipment operation.

[0060] Dynamically adjusting the baseline threshold based on fault warning weights and using an extended Kalman filter to correct deviations and generate warning parameters can adapt to the dynamic changes in zero-carbon equipment operating conditions, improving the accuracy and timeliness of warnings. Constructing a multidimensional equipment state space and determining fault warning status by analyzing extreme point density and covariance mutation areas can more comprehensively capture equipment abnormalities and reduce missed and false alarms.

[0061] A joint optimization model of ESG parameter data and equipment warning thresholds was established, and the non-dominated sorting genetic algorithm was used to solve the optimal warning strategy, achieving the coordinated optimization of ESG benefits and equipment failure warnings. While ensuring the safe operation of equipment, it maximized ESG benefits and promoted the sustainable development of zero-carbon equipment.

[0062] The system's multi-source data acquisition module can obtain multi-dimensional ESG parameter time series matrix in real time, providing rich data support for the entire early warning system; various functional modules such as the state feature extraction module and the ESG correlation calculation module cooperate with each other to realize the automation of the entire process from data acquisition, feature extraction, correlation calculation to early warning generation, thereby improving the efficiency and reliability of early warning, and transmitting the early warning signal to the monitoring platform through the industrial Internet, realizing real-time early warning and rapid response to faults, and providing strong guarantees for the safe and stable operation of zero-carbon equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a working principle diagram of the zero-carbon equipment failure early warning method based on ESG scheduling according to the present invention;

[0064] Figure 2 Process diagram for device status feature extraction;

[0065] Figure 3 A flowchart for ESG relevance calculation;

[0066] Figure 4 Flowchart for core feature vector extraction. DETAILED DESCRIPTION

[0067] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0068] See also Figure 1-Figure 4 The present invention relates to a zero-carbon equipment failure early warning method based on ESG scheduling, and the specific implementation steps are as follows:

[0069] Collect multi-source ESG parameter data during the operation of zero-carbon equipment, where the ESG parameter data includes carbon footprint intensity, clean energy consumption, and energy storage system charging and discharging efficiency.

[0070] A graph convolutional network is used to extract device status features from ESG parameter data to obtain the status feature matrix, and the ESG correlation is calculated based on the covariance change of the status feature matrix.

[0071] The cross-correlation matrix of ESG correlation in different time periods is subjected to feature dimensionality reduction to extract the core feature vector, which is then nonlinearly mapped using the ReLU activation function to obtain the fault warning weight for each time period.

[0072] The baseline threshold of the state feature matrix is ​​dynamically adjusted according to the fault warning weight to generate warning parameters, and the equipment fault warning process is triggered based on the warning parameters.

[0073] Example 1: When extracting the device status features, the ESG parameter data is intercepted at a set sampling interval. The sampling interval here needs to take into account factors such as the type of equipment, operating conditions, and actual application scenarios. For example, for zero-carbon equipment such as wind turbines that are greatly affected by environmental factors and have more obvious fluctuations in operating status, in order to be able to capture subtle changes in their operating status in a timely manner, the sampling interval can be set to a shorter time, such as collecting data every minute or every five minutes; and for some energy storage system equipment with relatively stable operation, the sampling interval can be appropriately extended, such as collecting data every ten minutes or fifteen minutes. This can ensure that sufficient data reflecting the device status is obtained, and reduce the amount of data collected and the burden of subsequent processing. After intercepting the data, it is necessary to construct a state tensor. The construction method of the state tensor is related to the dimension of the collected ESG parameter data. The three types of parameter data, carbon footprint intensity, clean energy consumption, and energy storage system charge and discharge efficiency, will be organized into a tensor structure of a specific dimension to facilitate subsequent processing and analysis.

[0074] After the state tensor is constructed, it must be standardized. The purpose of standardization is to eliminate the differences in dimensions and numerical ranges between different parameters, so that each parameter has equal importance in the subsequent feature extraction process. Specifically, standardization usually subtracts the mean of each parameter data and then divides it by its standard deviation. In this way, the mean of the processed data is 0 and the standard deviation is 1, thereby achieving data normalization. For example, the unit of carbon footprint intensity may be kilograms of carbon dioxide per kilowatt-hour, while the unit of clean energy consumption may be kilowatt-hour. The dimensions and numerical ranges of the two are quite different. Through standardization, they can be converted into comparable values.

[0075] After normalization, a multi-head graph convolutional layer is used to aggregate cross-device features on the state tensor. The design of this multi-head graph convolutional layer takes into account the complex connections and interactions between zero-carbon devices. In practical applications, a zero-carbon device system may consist of multiple different types of devices, such as wind turbines, solar panels, and energy storage batteries. These devices are interconnected through power grids or other physical connections, forming a complex network structure. Each "head" in the multi-head graph convolutional layer can be viewed as extracting correlation features between devices from a different perspective. By processing multiple "heads" in parallel, a more comprehensive picture of the various characteristic relationships within the device network can be captured. When performing cross-device feature aggregation, an adjacency matrix must first be constructed. The adjacency matrix represents the connection relationships between devices, and the element values ​​in the matrix represent the strength of the association between the devices. The state tensor is then operated on with the adjacency matrix, outputting the dot product of the adjacency matrix and the eigenvector. This dot product sum incorporates both the device's own characteristics and the characteristics of the interactions between devices.

[0076] The aggregated results of the multi-head graph convolutional layer are input into a gated recurrent unit (GRU). A GRU is a special type of recurrent neural network that effectively processes time series data and captures temporal dependencies within the data. A GRU contains two important structures: an update gate and a reset gate. The update gate controls how much state information from the previous moment is passed to the current moment, while the reset gate controls the extent to which state information from the previous moment is ignored. The synergistic effect of these two gates enables the extraction of features at multiple time scales. For example, for short-term device state changes, the update gate allows more current state information to pass through, while for long-term trend features, the update gate retains more historical state information.

[0077] After extracting multi-timescale features through the gated recurrent unit, the feature dimensions need to be filtered through the attention mask. The role of the attention mask is to focus on those feature dimensions that are more important for device status analysis and suppress unimportant feature dimensions. In actual device operation data, different feature dimensions contribute differently to device fault warnings. Some dimensions may contain a large amount of noise information, while some dimensions directly reflect the critical status of the device. Through the attention mask, a weight can be assigned to each feature dimension. Dimensions with larger weights indicate higher importance and will be retained, while dimensions with smaller weights will be weakened or ignored. After filtering through the attention mask, the state feature matrix is ​​finally generated. This matrix contains the device status feature information after cross-device feature aggregation, multi-timescale feature extraction, and feature dimension screening. It can more accurately reflect the actual operating status of the device and provide a reliable basis for subsequent ESG correlation calculations and fault warnings.

[0078] Throughout the entire device state feature extraction process, each step is closely linked, with the results of the previous step serving as the input for the next. By rationally designing the processing methods and parameters for each step, key features reflecting the device state can be effectively extracted from multi-source ESG parameter data, laying a solid foundation for subsequent fault warning analysis. For example, the rational setting of the sampling interval ensures the timeliness and validity of the data, while standardization ensures data comparability. The multi-head graph convolutional layer achieves comprehensive aggregation of cross-device features, the gated recurrent unit captures multi-timescale features in the time series, and the attention mask selects important feature dimensions. The collaborative work of these steps enables the generated state feature matrix to accurately and comprehensively describe the operating status of the device.

[0079] Example 2: When calculating ESG relevance, select the state feature matrix of any time period as the baseline threshold. The selection of this time period should be based on the equipment's operating cycle and historical data patterns. For example, a period of continuous and stable operation under normal equipment conditions can be selected, such as the period after equipment startup when operating parameters tend to stabilize, or a typical operating period based on the equipment maintenance cycle. The baseline threshold should be representative and reflect the distribution of equipment state features under normal operating conditions.

[0080] Calculate the covariance value of the feature vectors of the k adjacent time periods of the benchmark threshold. The selection of k adjacent time periods needs to consider the temporal correlation of the device state changes. The k value can be adjusted according to the device type and operating characteristics. For example, for energy storage system equipment with a faster response speed, the k value can be set to 5-10 time periods to capture feature changes in a short period of time. For equipment such as wind turbines that are greatly affected by the environment and have relatively slow state changes, the k value can be set to 15-20 time periods to cover longer-term feature fluctuations. The calculation of the covariance value targets the correlation between feature vectors and reflects the degree of coordinated changes of different features in adjacent time periods. For example, whether the fluctuations of carbon footprint intensity and energy storage system charging and discharging efficiency in adjacent time periods are consistent.

[0081] The ratio of the mean covariance value to the Euclidean distance from the baseline threshold is used as the intra-period correlation measure. The mean covariance value reflects the average level of correlation between the feature vectors of k adjacent time periods, and the Euclidean distance is used to measure the magnitude of the difference between this mean and the baseline threshold. Specifically, the arithmetic mean of the k covariance values ​​is first calculated. Then, the Euclidean distance between this mean and the corresponding feature vector in the baseline threshold is calculated, which is the square root of the sum of the squares of the differences in each dimension. Finally, the Euclidean distance is divided by the mean covariance value to obtain a dimensionless intra-period correlation measure. This correlation measure reflects the degree of deviation of the correlation between the feature vectors in the current period and the baseline state. A larger value indicates a more significant difference between the feature changes in the period and the baseline threshold.

[0082] The cross-correlation coefficient between the baseline threshold and the associated time period characteristics is then calculated, and its absolute value is taken as the correlation between the time periods. The selection of associated time period characteristics can include feature vectors from the previous time period, the next time period, or multiple time periods before and after, determined based on the time lag in device state transmission. The cross-correlation coefficient is used to measure the similarity between the baseline threshold and the associated time period characteristics in the time series. It is calculated by calculating the ratio of the covariance of the two feature vectors to the product of their respective standard deviations. The value range is [-1, 1]. Taking the absolute value eliminates directional differences in positive and negative correlations, retaining only the quantitative value of the degree of similarity. For example, if a feature vector of the baseline threshold and a feature vector of the next time period show the same trend, the absolute value of the cross-correlation coefficient is close to 1, indicating a close correlation between the time periods.

[0083] The harmonic mean of the intra-period correlation and the inter-period correlation is used as the ESG correlation. The harmonic mean is calculated by multiplying the two numbers by 2 times, divided by their sum. This method prevents the results from being dominated by a single correlation value, which is too large or too small, and provides a more balanced representation of the degree of correlation within and between periods. For example, when the intra-period correlation is large but the inter-period correlation is small, the harmonic mean will tend to be smaller, providing a comprehensive reflection of the overall correlation level.

[0084] In practice, attention must be paid to the time alignment and feature dimensionality consistency of data. Collected ESG parameter data must be arranged in strict chronological order to ensure temporal continuity of feature vectors in adjacent time periods and avoid bias in correlation calculations due to timestamp errors. Furthermore, the baseline threshold and the dimensionality of the feature vectors for each time period must be consistent. The feature dimensions corresponding to parameters such as carbon footprint intensity, clean energy consumption, and energy storage system charge and discharge efficiency must correspond one-to-one to ensure the accuracy of covariance, Euclidean distance, and correlation coefficient calculations.

[0085] Furthermore, the calculation of ESG relevance must account for dynamic changes in the equipment's operating scenario. When a device switches from normal operation to a variable load condition or is subject to external environmental interference, the baseline threshold must be dynamically updated based on real-time operating data to adapt to the new operating state. For example, if grid load demand suddenly increases and the energy storage system transitions from a charging state to a discharging state, the state characteristic matrix for that condition must be reselected as the baseline threshold to avoid distortion in the correlation calculation due to a mismatch between the baseline threshold and the current operating state.

[0086] The entire ESG correlation calculation process, through the fusion of multi-dimensional correlation quantities, achieves a quantitative analysis of the correlation of equipment status characteristics in the time series. The intra-period correlation quantity captures the difference between the characteristic changes of the current period and the baseline state, and the inter-period correlation quantity reflects the transmission relationship of the characteristics of different periods. The harmonic mean of the two comprehensively reflects the overall correlation degree of ESG parameter data, providing a quantitative basis for subsequent feature dimensionality reduction and fault warning weight calculation. In specific implementation, it is necessary to fine-tune the parameters of each step according to the equipment type and application scenario, such as the setting of the k value and the selection of the correlation period, to ensure that the ESG correlation can accurately reflect the actual correlation characteristics of the equipment status.

[0087] Example 3: When extracting the core feature vector, it is necessary to construct a cross-correlation matrix of the ESG correlation of different time periods. The different time periods here can be multiple continuous time periods arranged in chronological order, for example, a day is divided into 24 hours, each hour as a time period, or the time periods are divided according to the operating cycle of the equipment. Taking the energy storage system of a zero-carbon park as an example, assuming that the system runs for 20 hours a day and is shut down for 4 hours for maintenance, the daily operating time can be divided into 20 time periods, each of which is 1 hour. After calculating the ESG correlation in each time period, these correlation values ​​are arranged in order of the time periods to construct a cross-correlation matrix. The elements in the cross-correlation matrix represent the correlation between the ESG correlations of different time periods. For example, the element in the i-th row and j-th column of the matrix represents the degree of correlation between the ESG correlations of the i-th time period and the j-th time period.

[0088] After constructing the cross-correlation matrix, principal component analysis (PCA) is performed on it to obtain orthogonal eigenvectors. PCA is a dimensionality reduction technique that converts high-dimensional data into low-dimensional principal components while preserving as much information as possible from the original data. This process first calculates the eigenvalues ​​and eigenvectors of the cross-correlation matrix. The eigenvectors are then orthogonalized to make them perpendicular to each other. These orthogonal eigenvectors represent the primary direction of variation in the data. Each eigenvector corresponds to an eigenvalue, and the magnitude of the eigenvalue represents the amount of information contained in that eigenvector.

[0089] It is necessary to select eigenvectors whose principal component variance contribution rate exceeds a set threshold to form the core feature subspace. The choice of the threshold needs to be determined based on the actual application scenario and the requirements for data dimensionality reduction. For example, the threshold can be set to 80%, which means that we only select eigenvectors whose sum of eigenvalues ​​accounts for 80% of the sum of the total eigenvalues. The subspace formed by these eigenvectors is called the core feature subspace. Taking the energy storage system mentioned above as an example, assuming that the cross-correlation matrix is ​​a 20×20 matrix, after principal component analysis, 20 eigenvectors and corresponding eigenvalues ​​are obtained. If the sum of the variance contribution rates of the first five eigenvectors exceeds 80%, then these five eigenvectors are selected to form the core feature subspace.

[0090] The cross-correlation matrix is ​​mapped to the core feature subspace to obtain the core feature vector after dimensionality reduction. The mapping process is actually to project the original cross-correlation matrix onto the core feature subspace, so that the high-dimensional data is converted into low-dimensional data. Specifically, for the ESG correlation vector of each time period, it is dot-producted with each feature vector in the core feature subspace to obtain the coordinate value of the time period on the core feature subspace. These coordinate values ​​constitute the core feature vector after dimensionality reduction. For example, the original ESG correlation vector of each time period is 20-dimensional. After mapping to the 5-dimensional core feature subspace, the core feature vector of each time period becomes 5-dimensional.

[0091] In practice, careful attention must be paid to data preprocessing and eigenvector selection. First, before constructing the cross-correlation matrix, ESG correlation data must be standardized to eliminate differences in data dimensions and numerical ranges across different time periods. Second, when selecting eigenvectors, a comprehensive consideration must be given to both the variance contribution rate and the physical significance of the eigenvectors. Some eigenvectors, despite their high variance contribution rates, may not have practical physical significance. This requires careful consideration and selection based on the specific circumstances.

[0092] Furthermore, the extraction of core eigenvectors also needs to consider the impact of the device's operating status and environmental factors. For example, the distribution of ESG correlations may change when the device is in different operating modes. In this case, it is necessary to reconstruct the cross-correlation matrix and perform principal component analysis to ensure that the core eigenvector accurately reflects the device's current operating status. Furthermore, changes in environmental factors such as temperature and humidity may affect the calculation of ESG correlations and, in turn, the extraction of core eigenvectors. Therefore, in practical applications, these environmental factors must be properly processed and compensated for.

[0093] The entire core feature vector extraction process, through steps such as constructing a correlation matrix, principal component analysis, feature vector selection and mapping, achieves dimensionality reduction processing of ESG correlation data, retains the main features of the data, and at the same time reduces the dimension of the data, improving the efficiency and accuracy of subsequent processing. In specific implementation, it is necessary to reasonably select parameters such as time period division and threshold setting based on the type of equipment, operating characteristics and application scenarios to ensure that the core feature vector can accurately reflect the status characteristics of the equipment and provide a reliable basis for subsequent fault warning weight calculation and fault warning. For example, for zero-carbon equipment such as wind turbines that are more affected by weather, the time period division can be more refined and the threshold setting can be appropriately increased to capture more feature changes; while for energy storage systems with relatively stable operation, the time period division can be relatively coarse and the threshold setting can be appropriately lowered to reduce the amount of calculation.

[0094] Example 4: When obtaining the fault warning weights for each time period, the core feature vector must first be normalized and its cosine similarity with the preset reference vector must be calculated. Taking the energy storage system of a photovoltaic power station as an example, assume that the core feature vector of a certain time period extracted through the steps of Example 3 is a numerical sequence containing 5 dimensions, such as [0.23, 0.41, -0.17, 0.35, 0.12]. During normalization, the vector must be converted into a unit vector, that is, by calculating the square root of the sum of the squares of the values ​​of each dimension and then dividing each dimension by the square root to make the vector length 1. Assume that after normalization, [0.37, 0.66, -0.27, 0.56, 0.19] is obtained. The preset reference vector can be determined based on the statistical mean of the historical core feature vectors during normal operation of the equipment. For example, by analyzing the core feature vectors of the energy storage system under normal operating conditions in the past month, the reference vector calculated is [0.30, 0.50, 0.10, 0.40, 0.20]. At this time, the calculation of cosine similarity needs to be based on the dot product of the two unit vectors, that is, the corresponding values ​​of each dimension are multiplied and summed. The closer the result is to 1, the more consistent the directions of the two are. For example, the cosine similarity between the core feature vector and the reference vector in this period is calculated to be 0.37×0.30+0.66×0.50+(-0.27)×0.10+0.56×0.40+0.19×0.20. The specific value needs to be obtained based on actual calculations. This value reflects the degree of deviation of the current feature from the normal state.

[0095] Next, the similarity values ​​are fed into a bidirectional long short-term memory (Bi-LSTM) network, where they are transformed by the forward and backward layers to generate initial weights. The Bi-LSTM network's architecture must account for the bidirectional nature of time series. For example, two hidden layers, each containing 50 neurons, are used to process similarity sequences across consecutive time periods. Using the example of a photovoltaic power plant energy storage system, assuming the similarity value for the 10th time period is currently being calculated, the network simultaneously receives similarity sequences from time periods 1 to 10. The forward layer processes from time period 1 to 10, capturing the characteristic trends of past time periods. The backward layer processes from time period 10 to 1, capturing the characteristic dependencies of future time periods (based on reverse inference from historical data). The forget gate, input gate, and output gate within the network adjust weights based on the input data. For example, if the similarity value for a certain time period suddenly decreases (indicating a deviation from normal characteristics), the forward layer will increase the influence of that time period on subsequent weights. The backward layer will verify whether this deviation is an abnormal fluctuation using historical data. After the forward and backward calculations of the network, an initial weight value between 0 and 1 is output, which integrates the dynamic feature changes in the time series.

[0096] The initial weights are then smoothed using a moving average filter to output the fault warning weights for each time period. The size of the moving average window should be set based on the frequency of device status changes. For example, for energy storage systems, a window size of 5 can be set. This means that the arithmetic mean of the initial weights for the current time period and the previous four time periods is calculated. Assuming the initial weight for the 10th time period is 0.82, and the initial weights for the previous four time periods are 0.75, 0.78, 0.80, and 0.81, respectively, the moving average weight is (0.82 + 0.75 + 0.78 + 0.80 + 0.81) ÷ 5. The specific value is determined based on actual calculations. Smoothing can eliminate sudden changes in weights caused by short-term data fluctuations. For example, if a brief sensor anomaly causes a sudden change in the similarity value during a certain period, the moving average filter can reduce the impact of this anomaly by averaging the historical weights, making the warning weight more consistent with the actual device status trend.

[0097] In practical applications, the update mechanism of the preset reference vector is particularly important. For example, when the energy storage system of a photovoltaic power station switches from daytime charging mode to nighttime discharging mode, the ESG parameter characteristics of the equipment will change significantly. At this time, the reference vector needs to be recalculated based on the historical core feature vector under the new working conditions. In specific operations, the reference vector can be set to be updated every hour, and adjusted based on the sliding window mean of the core feature vector in the past hour to ensure that the reference vector matches the current operating mode. In addition, the parameter training of the Bi-LSTM network needs to be based on the historical operating data of the equipment. For example, the core feature vectors and corresponding fault labels of the energy storage system in different states such as normal, minor faults, and serious faults are collected, and the network weights are optimized through the backpropagation algorithm so that the network can more accurately convert the similarity sequence into the initial weight reflecting the probability of failure.

[0098] The length of the input sequence for the bidirectional long-short-term memory network must also be set based on the equipment's fault latency. For slowly developing faults, such as energy storage battery capacity decay, the latency can last for several days, so the input sequence can be set to include the daily core feature vector similarity values ​​for the past seven days. For rapidly developing faults, such as inverter power module overheating, the latency may only be a few hours, so the input sequence can be shortened to the past 12 hours. This dynamic adjustment mechanism enables the network to better capture the temporal characteristics of different fault types.

[0099] The window size of the moving average filter must also be considered in conjunction with the required alert sensitivity. To detect sudden failures promptly, the window size can be reduced (e.g., set to 3) to make the weight more sensitive to recent changes. To prioritize long-term trends and reduce false alarms, the window size can be increased (e.g., set to 7). For example, during high summer temperatures, the energy storage system's charging and discharging efficiency may experience a short-term decrease due to ambient temperature fluctuations. A larger window size can prevent these normal fluctuations from being misinterpreted as fault warning signals.

[0100] The entire fault warning weight generation process achieves dynamic quantification of device status characteristics through similarity calculation, bidirectional time series feature extraction, and smoothing. Taking the photovoltaic power station energy storage system as an example, when the cosine similarity between the core feature vector and the reference vector in a certain period decreases, it indicates that the ESG parameter characteristics deviate from the normal state. The Bi-LSTM network will combine historical and future feature trends to determine whether the deviation is persistent. If it is confirmed that the deviation is persistent and exceeds the smoothed weight threshold, a higher warning weight is generated, providing a basis for subsequent warning parameter adjustments. In specific implementation, it is necessary to fine-tune the parameters of each step according to the device type, such as the update frequency of the reference vector, the number of hidden layer neurons in the Bi-LSTM, the window size of the moving average, etc., to ensure that the warning weight can accurately reflect the fault risk and adapt to the dynamic operating conditions of the equipment.

[0101] Example 5: When generating warning parameters, it is necessary to perform matrix multiplication of the fault warning weight and the benchmark threshold to obtain the weight adjustment parameter. Taking the energy storage converter of a certain wind farm as an example, it is assumed that the fault warning weight of a certain period obtained through the steps of Example 4 is a 3×1 matrix, such as [0.85, 1.2, 0.9], which corresponds to the warning weights of carbon footprint intensity, clean energy consumption and energy storage system charging and discharging efficiency respectively. The benchmark threshold is set according to the historical data of the normal operation of the equipment. For example, the benchmark threshold of carbon footprint intensity is The benchmark threshold for clean energy consumption is 800kWh, and the benchmark threshold for energy storage system charging and discharging efficiency is 0.9. When performing matrix multiplication, the weight is multiplied by each benchmark threshold to obtain the weight adjustment parameter, that is, the adjustment parameter for carbon footprint intensity is The adjustment parameter for clean energy consumption is 800 × 1.2 = 960 kWh, and the adjustment parameter for energy storage system charge and discharge efficiency is 0.9 × 0.9 = 0.81. This process dynamically scales the baseline threshold using weights, allowing the threshold to adjust based on the degree of deviation from the current device status characteristics.

[0102] Calculate the deviation between the weight adjustment parameter and the benchmark threshold, and dynamically correct the deviation through the extended Kalman filter. The deviation is calculated as the weight adjustment parameter minus the benchmark threshold. For example, the deviation of carbon footprint intensity is The deviation in clean energy consumption is 960-800=160kWh, and the deviation in the energy storage system's charge and discharge efficiency is 0.81-0.9=-0.09. The extended Kalman filter handles state estimation in nonlinear systems, requiring the establishment of the system's state equation and observation equation. Taking the operating state of an energy storage converter as an example, the state equation describes the dynamic changes in the device's internal parameters, such as capacitor aging and power device temperature, while the observation equation relates these internal states to the deviations in measurable ESG parameters. The filter calculates the optimal estimate of the current state based on the previous state estimate and the current observations, thereby correcting the deviation. For example, when a sudden increase in the deviation in clean energy consumption is detected, the filter combines the device's historical operating data with its current operating conditions to determine whether the deviation is caused by a device failure or external factors such as grid load fluctuations, and corrects the deviation accordingly.

[0103] The corrected deviation is added to the baseline threshold to generate the warning parameter. For example, the warning parameter of carbon footprint intensity is 0.5 + (-0.075 corrected value). Assuming the corrected deviation is -0.05, the warning parameter is The warning parameter for clean energy consumption is 800 + 160 after correction. Assuming a corrected deviation of 120, the warning parameter is 920 kWh. The warning parameter for energy storage system charge and discharge efficiency is 0.9 + (-0.09 after correction). Assuming a corrected deviation of -0.07, the warning parameter is 0.83. These warning parameters are derived based on baseline thresholds, combined with the fault warning weight and deviation correction for the current device state, to more accurately reflect the current level of equipment failure risk.

[0104] In practice, setting benchmark thresholds requires extensive historical equipment operating data. For example, for wind farm energy storage converters, it's necessary to collect normal operating data across different seasons and load conditions. Statistical analysis is then used to determine the benchmark threshold ranges for each ESG parameter. Equipment aging also needs to be considered. As equipment ages, benchmark thresholds may drift, necessitating regular updates.

[0105] Adjusting the parameters of the extended Kalman filter is also crucial. The filter's process noise covariance and measurement noise covariance need to be set based on the device's actual operating conditions. For example, for energy storage converters with relatively stable operating conditions, the process noise covariance can be set small. However, for devices that are more susceptible to external environmental influences, such as wind turbines, the process noise covariance needs to be set larger to accommodate their large state fluctuations. Furthermore, the filter's initial state estimate must be as accurate as possible. Typically, the device's initial operating state can be used as the filter's initial state.

[0106] After the warning parameters are generated, they need to be compared with the actual operating data of the equipment in real time. When the actual operating data exceeds the warning parameters, the fault warning process is triggered. For example, when the actual measured value of the carbon footprint intensity of the energy storage converter reaches When the above, the system will issue a corresponding early warning signal.

[0107] The entire warning parameter generation process achieves dynamic adjustment of equipment fault warning thresholds through steps such as weight adjustment, deviation correction, and threshold superposition. Taking a wind farm's energy storage converter as an example, when a minor fault occurs in the equipment, the fault warning weight increases accordingly. A matrix multiplication is performed to change the baseline threshold, and the deviation is corrected using an extended Kalman filter. The resulting warning parameters can promptly reflect the equipment's fault status and provide a reliable basis for fault warning. During implementation, the parameters of each step, such as the calculation method for the fault warning weight and the parameters of the extended Kalman filter, need to be appropriately set according to the equipment type and operating characteristics to ensure that the generated warning parameters accurately and promptly reflect the equipment's failure risk.

[0108] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0109] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A zero-carbon equipment failure early warning method based on ESG scheduling, characterized in that: include: Collect multi-source ESG parameter data during the operation of zero-carbon equipment, where the ESG parameter data includes carbon footprint intensity, clean energy consumption, and energy storage system charging and discharging efficiency; A graph convolutional network is used to extract device status features from ESG parameter data to obtain a state feature matrix. The ESG correlation is then calculated based on the covariance change of the state feature matrix. The cross-correlation matrix of ESG correlation in different time periods is subjected to feature dimensionality reduction to extract the core feature vector. The core feature vector is then nonlinearly mapped using the ReLU activation function to obtain the fault warning weight for each time period. The baseline threshold of the state feature matrix is ​​dynamically adjusted according to the fault warning weight to generate warning parameters, and the equipment fault warning process is triggered based on the warning parameters.

2. A zero-carbon equipment failure early warning method based on ESG scheduling according to claim 1, characterized in that: The device status feature extraction includes: Intercept ESG parameter data at a set sampling interval, construct a state tensor and perform normalization; A multi-head graph convolutional layer is used to aggregate cross-device features of the state tensor and output the dot product sum of the adjacency matrix and the feature vector; The aggregation results are input into the gated recurrent unit, multi-time scale features are extracted through the update gate and reset gate, and the feature dimensions are filtered through the attention mask to generate a state feature matrix.

3. The zero-carbon equipment failure early warning method based on ESG scheduling according to claim 1 is characterized in that: Calculating ESG relevance includes: Select the state feature matrix of any time period as the benchmark threshold and calculate the covariance value of the feature vectors of its adjacent k time periods; The ratio of the mean of the covariance value to the Euclidean distance of the benchmark threshold is used as the correlation quantity within the time period; Calculate the correlation coefficient between the benchmark threshold and the characteristics of the associated time period, and take its absolute value as the correlation coefficient between time periods; The harmonic mean of the intra-period correlation and the inter-period correlation is taken as the ESG correlation.

4. The zero-carbon equipment failure early warning method based on ESG scheduling according to claim 1 is characterized in that: The extracting of the core feature vector comprises: Construct a cross-correlation matrix of ESG correlations in different time periods and perform principal component analysis on it to obtain orthogonal eigenvectors; The eigenvectors whose principal component variance contribution rate exceeds the set threshold are selected to form the core feature subspace; The cross-correlation matrix is ​​mapped to the core feature subspace to obtain the core feature vector after dimensionality reduction.

5. The zero-carbon equipment failure early warning method based on ESG scheduling according to claim 1 is characterized in that: Obtaining the fault warning weights for each time period includes: Normalize the core feature vector and calculate its cosine similarity with the preset reference vector; The similarity value is input into the bidirectional long short-term memory network, and the initial weight is generated after transformation by the forward layer and the backward layer; The initial weights are smoothed by moving average filtering, and the fault warning weights for each time period are output.

6. The zero-carbon equipment failure early warning method based on ESG scheduling according to claim 1 is characterized in that: The generating of early warning parameters includes: Perform matrix multiplication of the fault warning weight and the benchmark threshold to obtain the weight adjustment parameter; Calculate the deviation between the weight adjustment parameter and the reference threshold, and dynamically correct the deviation through the extended Kalman filter; The corrected deviation is added to the baseline threshold to generate the warning parameter.

7. The zero-carbon equipment failure early warning method based on ESG scheduling according to claim 1 is characterized in that: Also includes: Construct a multi-dimensional equipment state space based on warning parameters and extract the extreme points and covariance mutation areas of the space; When the density of extreme points exceeds the set threshold or the span of the covariance mutation area is greater than the limit value, it is determined to be a fault warning state and a multi-level warning instruction is generated.

8. The zero-carbon equipment failure early warning method based on ESG scheduling according to claim 7 is characterized in that: The multi-dimensional device state space construction includes: Map the warning parameters to a multi-dimensional coordinate system according to the operating time sequence to generate a state distribution point set; The kernel density estimation method is used to reconstruct the state distribution point set spatially, and the covariance gradient and kurtosis coefficient of the reconstructed space are calculated; The noise of gradient distribution is suppressed by Gaussian filtering algorithm to eliminate random interference.

9. The zero-carbon equipment failure early warning method based on ESG scheduling according to claim 1 is characterized in that: Also includes: A joint optimization model of ESG parameter data and equipment warning thresholds is established, and a non-dominated sorting genetic algorithm is used to solve the optimal warning strategy; The optimal early warning strategy and early warning parameters are coupled in real time to generate a fault early warning plan that maximizes ESG benefits; The joint optimization model solution includes: The objective function is defined as the sum of the weighted absolute values ​​of the ESG indicator deviation and the early warning missed rate, and the constraint condition is the equipment operation safety range; Perform Pareto optimal transformation on the objective function and decompose it into the warning threshold sub-problem and the ESG scheduling sub-problem; The two sub-problems are solved alternately and iteratively until convergence, and the optimal early warning strategy that meets the constraints is output.

10. A zero-carbon equipment failure early warning system based on ESG scheduling, characterized in that: include: A multi-source data acquisition module is used to obtain real-time data on carbon footprint intensity, clean energy consumption, and energy storage system charge and discharge efficiency during operation, and to construct a multi-dimensional ESG parameter time series matrix; The state feature extraction module uses a graph convolutional network to aggregate cross-device features of the multi-dimensional ESG parameter time series matrix, captures multi-timescale device state features through a gated recurrent unit, and outputs a state feature matrix; The ESG correlation calculation module generates intra-period and inter-period correlation quantities based on the covariance changes of the state characteristic matrix, and calculates their harmonic mean as the ESG correlation degree; The cross-correlation matrix decomposition module constructs a cross-correlation matrix for ESG correlations in different time periods and extracts core eigenvectors through principal component analysis; The weight mapping module uses the ReLU activation function to perform nonlinear mapping on the core feature vector to generate the fault warning weight for each time period; The parameter optimization module dynamically adjusts the benchmark threshold based on the fault warning weight, corrects the parameter deviation through the extended Kalman filter, and generates warning parameters; The real-time early warning module converts the early warning parameters into equipment failure early warning signals, which are transmitted to the monitoring platform through the industrial Internet to trigger the early warning instructions.

Images