An auxiliary calculation method for carbon emissions of industrial users based on a space-time correlation matrix

Abstract

The application provides an auxiliary calculation method for industrial user comprehensive energy carbon emission based on a space-time correlation matrix, a virtual sensor array is constructed by fully utilizing the power metering instrument which has been scaled and arranged in a factory, and non-electric energy metering instruments do not need to be purchased, installed and transformed; based on the space-time correlation matrix mathematical model which is bound to the physical prior constraint of a production process, the low-frequency non-electric energy cumulative consumption of an hour and a day is reconstructed into the minute-granularity instantaneous non-electric energy consumption time sequence which is synchronous with the high-frequency electric active power, the lag defect of traditional carbon emission accounting is solved, and real-time energy-saving intervention in a production process and accurate carbon emission abnormality tracing are realized.

Description

Technical Field

[0001] This invention relates to the field of industrial energy metering and enterprise greenhouse gas carbon emission accounting technology, specifically to an auxiliary method for calculating the comprehensive energy carbon emissions of industrial users based on a spatiotemporal correlation matrix. Background Technology

[0002] In the comprehensive and refined management of energy across the entire industrial production process and the optimization of energy conservation and carbon reduction in production processes, high-precision, high-time-granularity enterprise greenhouse gas carbon emission accounting results are core supporting data. Currently, the energy metering equipment deployed in various manufacturing plants in China suffers from spatiotemporal resolution heterogeneity. While large-scale enterprise-deployed power metering instruments can output minute-level high-frequency active power sequence data, non-electric energy metering equipment such as natural gas transmission flow meters, steam zone flow meters, and compressed air metering instruments, constrained by hardware sampling limits and industrial field communication bus bandwidth, can only output hourly and daily low-frequency non-electric energy cumulative consumption data. Existing industrial enterprise greenhouse gas carbon emission accounting technologies can only use the data reporting cycle of low-frequency non-electric energy metering equipment as a unified accounting benchmark, performing static and coarse-grained aggregation calculations on power active power sequence data and low-frequency non-electric energy cumulative consumption data at different time granularities. This technical approach has inherent limitations.

[0003] First, the results of corporate greenhouse gas carbon emission accounting are seriously lagging, making it impossible to capture instantaneous carbon emission fluctuations caused by minute-level production equipment start-up and shutdown, raw material fluctuations, and fluid pipeline leaks. It can only make corrections after the fact, making it difficult to achieve real-time energy-saving intervention and accurate source tracing of carbon emission anomalies during the production process.

[0004] Second, the correlation fitting between active power and non-electric energy consumption is not bound to the constraints of the factory's physical production process. It relies solely on statistical correlation modeling of time-series data, which can easily generate false correlations without causal relationships for material and energy transmission, significantly reducing the accuracy of enterprises' greenhouse gas carbon emission calculations. Summary of the Invention

[0005] The purpose of this invention is to provide an auxiliary method for calculating the integrated energy carbon emissions of industrial users based on a spatiotemporal correlation matrix, so as to solve the technical problems of spatiotemporal resolution imbalance and low calculation accuracy in existing industrial multi-energy flow carbon calculation methods.

[0006] To address the aforementioned problems, this invention discloses an auxiliary method for calculating the integrated energy carbon emissions of industrial users based on a spatiotemporal correlation matrix. This method comprises two independent modules: an offline modeling and training phase and an online real-time inversion calculation phase. Specifically, it includes the following steps:

[0007] S1: Acquisition of raw data on multi-source heterogeneous energy consumption and multi-dimensional spatiotemporal synchronous standardization and feature cleaning.

[0008] Collect high-frequency active power sequence data output from all power monitoring nodes in the industrial site, and simultaneously collect low-frequency non-electric energy cumulative consumption data output from all non-electric energy metering nodes in the industrial site; the observation period corresponding to the low-frequency non-electric energy cumulative consumption data is an integer multiple of the sampling interval of the high-frequency active power sequence data.

[0009] Lagrange polynomial interpolation is used to perform numerical completion processing on the missing data segments in the high-frequency power active power sequence data; an anomaly identification algorithm based on the absolute deviation of the median is used to identify abnormal peak pulse data points in the power active power sequence; and the abnormal data points are replaced by smoothing the effective power data inside the sliding window to generate a standardized fine-grained power active power load matrix.

[0010] An adaptive accumulation and mutation identification algorithm is used to continuously monitor the first-order difference sequence of the total power curve of all power metering instruments. The time nodes corresponding to the difference values ​​that exceed the preset sensitivity parameters in multiple consecutive segments are marked as production condition mutation boundary nodes. Based on all production condition mutation boundary nodes, the complete historical energy consumption time series is divided into two types of time series segments: steady-state continuous production segments and non-production shutdown segments. The constant consumption of each non-electric energy metering instrument within all non-production shutdown segments is extracted as the steady-state basic loss constant of the corresponding non-electric energy metering area.

[0011] S2: Constructing a spatial dependency adjacency matrix and a process time delay matrix based on industrial production process mechanisms.

[0012] Retrieve the original bill of materials information and the original topology data of the production process flow layout stored within the factory's manufacturing execution system; when the power-consuming production equipment controlled by the power monitoring instrument and the heat-consuming and gas-consuming production equipment controlled by the non-electric energy metering instrument are located in the same core reaction process area or are directly physically interconnected through a closed fluid transport pipeline, assign an initial physical connectivity value at the corresponding position in the spatial dependency adjacency matrix; when there is no direct material flow channel between the power-consuming production equipment controlled by the power monitoring instrument and the heat-consuming and gas-consuming production equipment controlled by the non-electric energy metering instrument, assign a value of zero at the corresponding position in the spatial dependency adjacency matrix;

[0013] The global minimum cost time alignment operation is performed only on the pairing of power-non-power nodes marked as physically connected within the spatially dependent adjacency matrix. The optimal time alignment path is solved for the high-frequency power active power sequence and the low-frequency non-power energy cumulative consumption sequence after cubic spline smoothing. The process transmission hysteresis time step corresponding to the optimal time alignment path is extracted, and the complete process time delay matrix is ​​constructed by filling in all process transmission hysteresis time steps.

[0014] S3: Construct a discretized industrial multi-energy flow spatiotemporal mapping mathematical model and perform convex optimization offline iterative training.

[0015] A discrete industrial multi-energy flow spatiotemporal mapping mathematical model is constructed to characterize the quantitative correlation between the instantaneous non-electric energy consumption in a single non-electric energy metering area and the historical active power data after backtracking through the process transmission lag time step. Taking the physical conservation of the total energy consumption within each complete observation cycle of low-frequency non-electric energy as the equality constraint, a joint optimization objective function is constructed, consisting of a macroscopic total conservation error term, a spatial topological mask L1 sparse regularization term, and a weighted stable L2 norm regularization term.

[0016] An alternating direction multiplier iteration algorithm with integrated coordinate descending sub-iteration logic is used to perform alternating iterative solution operations on the joint optimization objective function. After the iterative calculation process converges, the three core model parameters obtained by the solution, namely the optimal spatiotemporal correlation matrix, the optimal process time delay matrix, and the steady-state basic loss constant of each region, are completely and persistently stored in the industrial edge gateway memory database.

[0017] S4: Real-time online simulation of the plant's instantaneous non-electric energy consumption vector and calculation of the integrated instantaneous carbon emissions of industrial users.

[0018] According to the fixed sampling period of high-frequency active power sequence data, the real-time active power sampling values ​​output by all power metering instruments are continuously collected. The three core model parameters of optimal spatiotemporal correlation matrix, optimal process time delay matrix and steady-state basic loss constant of each region are retrieved from the persistent storage of the industrial edge gateway memory database. The instantaneous non-electric energy consumption at the current micro sampling moment is deduced for each region and spliced ​​to generate the instantaneous non-electric energy consumption vector of the whole plant.

[0019] Two types of standardized emission intensity diagonal matrices are constructed: a diagonal matrix of carbon emission factors for electricity and a diagonal matrix of carbon emission factors for non-electric energy. Based on the linear superposition operation of the complete matrix, the instantaneous carbon emissions of industrial users' comprehensive energy at any micro-sampling time are calculated in real time, and a dynamic real-time monitoring data stream of industrial users' comprehensive energy carbon emissions with minute-level granularity is continuously output.

[0020] S5: Execute the low-frequency metrology closed-loop residual verification process and the model parameter adaptive incremental update process.

[0021] Each time a new low-frequency cycle's actual cumulative non-electric energy consumption reading is received from an industrial site non-electric energy metering instrument, the relative numerical deviation between the total instantaneous non-electric energy consumption calculated by the model within the current complete low-frequency observation cycle and the actual cumulative consumption of the instrument is calculated. When the calculated relative numerical deviation exceeds a pre-set deviation tolerance threshold, the background asynchronously starts an incremental retraining thread for model parameters. A first-in-first-out time-series sample queue is used to complete the real-time update of the offline training dataset. The complete set of original model parameters fixed in the industrial edge gateway's memory database is updated using an incremental gradient descent algorithm for lightweight fine-tuning. After fine-tuning, the new optimal spatiotemporal correlation matrix and the new optimal process time delay matrix synchronously replace the original model parameters in the industrial edge gateway's memory database, which are then used for the calculation of instantaneous non-electric energy consumption in the next high-frequency power data sampling cycle.

[0022] Preferably, the Lagrange polynomial interpolation used for numerical completion of missing communication data segments in step S1 is either third-order or fourth-order Lagrange polynomial interpolation; the threshold for identifying abnormal spike pulse data points is a value between 3 and 5 times the absolute deviation of the median in the sliding window, and the sliding window completely includes 5 to 15 high-frequency power active power sequence data sampling steps.

[0023] Preferably, the sensitivity parameter of the adaptive accumulation and mutation identification algorithm in step S1 is in the range of 0.02 to 0.1; when three or more consecutive first-order difference values ​​exceed the preset sensitivity parameter, the corresponding time node is determined as the production condition mutation boundary node.

[0024] Preferably, in step S2, the global minimum cost time alignment operation uses Euclidean distance as the distance metric between the two sets of time series data, and a unidirectional time extension constraint is added to the time alignment path; the global minimum cost time alignment operation only performs the complete calculation operation on the pairing of power-non-power nodes marked as physically connected within the spatially dependent adjacency matrix.

[0025] Preferably, the complete operation logic of the L1 sparse regularization term of the internal spatial topology mask in step S3 is to perform Hadamard element-wise multiplication on the spatial dependency adjacency matrix and the spatiotemporal correlation matrix and then apply L1 norm constraints; the value range of the penalty hyperparameter for the L1 sparse regularization term of the spatial topology mask is 0.05 to 0.2; the value range of the penalty hyperparameter for the weighted stable L2 norm regularization term is 0.01 to 0.1; the optimal values ​​of the two types of penalty hyperparameters are determined by K-fold cross-validation.

[0026] Preferably, the discrete industrial multi-energy flow spatiotemporal mapping mathematical model in step S3 includes a local process nonlinearity correction term; the local process nonlinearity correction term is constructed using a second-order polynomial kernel function or a radial basis kernel function; the local process nonlinearity correction term is only enabled for special production process sections with phase change reactions and temperature trigger threshold effects, while the local process nonlinearity correction term can be omitted for conventional continuous linear production sections to simplify online real-time simulation calculations.

[0027] Preferably, the learning rate for the incremental gradient descent algorithm in step S5 is between 0.001 and 0.01; the total number of iterations for a single lightweight fine-tuning is controlled between 50 and 200; and the deviation tolerance threshold for the low-frequency measurement closed-loop residual verification process is a conventionally preset value of 3%.

[0028] Preferably, the indirect carbon emission intensity values ​​of electricity stored at each diagonal position of the diagonal matrix of electricity carbon emission factors in step S4 can be dynamically updated and adjusted according to the real-time power generation structure of the regional power grid and the written terms of the enterprise's green electricity procurement agreement.

[0029] Beneficial effects:

[0030] This invention discloses an auxiliary method for calculating the integrated energy carbon emissions of industrial users based on a spatiotemporal correlation matrix. By fully utilizing the large-scale power metering instruments already deployed in the factory to construct a virtual sensor array, there is no need to purchase, install, or modify non-electric energy metering instruments. Based on a spatiotemporal correlation matrix mathematical model bound to the physical prior constraints of the production process, the cumulative consumption of low-frequency non-electric energy at the hourly and daily granular levels is reconstructed into a time series of instantaneous non-electric energy consumption at the minute granular level, synchronized with the active power of high-frequency electricity. This solves the lag defect of traditional carbon emission accounting and realizes real-time energy-saving intervention in the production process and accurate source tracing of carbon emission anomalies. Attached Figure Description

[0031] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0032] Figure 1 The diagram shows a flowchart of an auxiliary method for calculating the integrated energy carbon emissions of industrial users based on a spatiotemporal correlation matrix, as disclosed in this invention. Detailed Implementation

[0033] To better understand the above-described objects, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Many specific details are set forth in the following description to provide a thorough understanding of the present invention; however, the present invention may also be practiced in other ways different from those described herein, and all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0034] like Figure 1 As shown, the purpose of this invention is to propose an auxiliary method for calculating the integrated energy carbon emissions of industrial users based on a spatiotemporal correlation matrix, in order to solve the technical defects of existing industrial enterprises’ greenhouse gas carbon emission accounting technology, such as asymmetric spatiotemporal resolution of multi-energy flow metering, easy generation of false statistical correlations without process causality in fitting models, and continuous decay of calculation accuracy after long-term operating condition drift.

[0035] This invention constructs a virtual sensor array using all existing power metering instruments in a factory. It builds a spatial dependency adjacency matrix and a process time delay matrix using production process topology data stored in the manufacturing execution system. A convex optimization algorithm with macroscopic total conservation constraints is used to solve the spatiotemporal correlation matrix offline. During online operation, it continuously collects real-time high-frequency active power sequence data, reverse-engineers the instantaneous non-electric energy consumption at the synchronous time granularity, and calculates the instantaneous carbon emissions of industrial users in real time using a standardized carbon emission factor diagonal matrix. Simultaneously, it establishes a low-frequency metering closed-loop residual verification and model parameter adaptive incremental update process. Without adding any physical energy metering hardware, it achieves high-precision, minute-level dynamic real-time monitoring of enterprise greenhouse gas carbon emissions throughout complex industrial scenarios.

[0036] This invention comprises two independent modules: an offline modeling and training phase and an online real-time inversion calculation phase.

[0037] The offline modeling and training phase involves retrieving historical multi-source heterogeneous energy consumption data from the factory and original information on the production process topology of the manufacturing execution system. It then constructs a spatial dependency adjacency matrix and a process time delay matrix with prior physical process constraints. Finally, it solves for the optimal spatiotemporal correlation matrix, the optimal process time delay matrix, and the steady-state basic loss constants of each region using a convex optimization iterative algorithm that integrates multiple types of regular constraints. The three types of core model parameters are then persistently stored in the industrial edge gateway's memory database.

[0038] The online real-time inversion calculation stage serves to continuously collect real-time active power sequence data from all power metering instruments according to a fixed high-frequency sampling period, retrieve all model parameters fixed in the memory database, deduce the instantaneous non-electric energy consumption at the same granularity as the power data, and calculate the instantaneous carbon emissions of industrial users' comprehensive energy on a minute-by-minute basis using a standardized carbon emission factor diagonal matrix. At the same time, it uses the real cumulative consumption uploaded by low-frequency non-electric energy metering instruments to conduct closed-loop residual verification. When the relative numerical deviation exceeds the preset deviation tolerance threshold, the background asynchronously executes lightweight incremental fine-tuning of model parameters to achieve long-term autonomous adaptation of the model to drift in industrial field conditions.

[0039] Specifically, the steps to implement this method are as follows:

[0040] Step 1: Acquisition of raw data on multi-source heterogeneous energy consumption and multi-dimensional spatiotemporal synchronous standardization feature cleaning.

[0041] Step 1.1 Complete Acquisition of Raw Data on Multi-Source Heterogeneous Energy Consumption

[0042] The system synchronously collects two types of heterogeneous raw energy consumption data through the factory data acquisition and monitoring platform and the industrial IoT data acquisition gateway:

[0043] The first type of fine-grained high-frequency energy consumption raw data consists of high-frequency active power sequence data output from all power monitoring nodes in the industrial site; N power metering instruments are deployed throughout the plant, forming a set of power monitoring nodes. Unified data sampling interval The value is 1 minute, and each sampling interval outputs a complete power active power time sequence record.

[0044] The second category of coarse-grained low-frequency energy consumption raw data includes: cumulative low-frequency non-electric energy consumption data output from all natural gas, steam, and compressed air metering nodes in the industrial site; and M non-electric energy metering instruments deployed throughout the plant, forming a set of non-electric monitoring nodes. Due to limitations imposed by the flow meter's hardware sampling limit and the on-site Modbus communication bandwidth, the sampling interval for non-electric energy data is... Much larger than the power data sampling interval, with values ​​of 1 hour and 1 day, satisfying... , where k is a positive integer.

[0045] Step 1.2 Processing missing values ​​in high-frequency power active power sequence data

[0046] Offline gateways and communication signal interruptions in industrial sites can cause continuous missing data segments in the high-frequency active power sequence. The system employs Lagrange polynomial interpolation to complete the missing values: third-order Lagrange polynomial interpolation is used when the missing duration is less than or equal to 5 high-frequency power data sampling steps. Fourth-order Lagrange polynomial interpolation is used when the missing duration is greater than 5 high-frequency power data sampling steps. If the missing window exceeds 30 high-frequency power data sampling steps, cubic spline interpolation is used as an auxiliary method for global smoothing completion, fully preserving the local variation trend of the power load time series.

[0047] Step 1.3 Identification and Smoothing Replacement of Abnormal Pulses in High-Frequency Power Sequence

[0048] Instantaneous motor startup surges, inverter switching impacts, and equipment short circuits can generate abnormal spike pulses with no practical production significance within the active power sequence, interfering with the convergence effect of subsequent offline modeling training. An anomaly identification algorithm based on median absolute deviation is used to identify and repair anomalies.

[0049] ① Set the sliding window length to 5 to 15 high-frequency power data sampling steps, and calculate the median of all active power values ​​within the window;

[0050] ② Calculate the absolute deviation between the active power value of a single power line and the median value in the window point by point, and set the anomaly judgment threshold to be 3 to 5 times the absolute deviation of the median value in the window.

[0051] ③ Data points exceeding the judgment threshold are marked as abnormal spike pulses, and smooth replacement is completed by using the average effective active power value within the sliding window;

[0052] After completing the missing data completion and abnormal pulse replacement processes, a standardized fine-grained power load matrix is ​​generated. , where L represents the total number of high-frequency power data sampling points within the selected historical training window.

[0053] Step 1.4 The adaptive accumulation and mutation identification algorithm divides the production operation time segment and extracts the steady-state basic loss constant of each region.

[0054] A complete industrial energy consumption time series includes steady-state continuous production segments, shift change shutdown segments, and equipment maintenance shutdown segments. During shutdown periods, there is no energy consumption coupled with production processes; only constant basic losses exist due to heat dissipation from fluid pipelines and minor leaks in the pipeline network. Directly incorporating these into model training would introduce significant noise and reduce the accuracy of the spatiotemporal correlation matrix fitting. This invention employs an adaptive accumulation and mutation identification algorithm to automatically segment the time series. The complete execution logic is as follows:

[0055] ① Obtain the first-order difference sequence of the total power curve of all power metering instruments in the entire plant;

[0056] ② The sensitivity parameters of the adaptive accumulation and mutation identification algorithm are configured, with a fixed value range of 0.02 to 0.1;

[0057] ③ When three or more consecutive first-order difference values ​​exceed the preset sensitivity parameter, the corresponding time node is marked as the production condition change boundary node;

[0058] ④ Using the boundary nodes of sudden changes in all production conditions as the dividing boundaries, the complete historical energy consumption time series is divided into multiple steady-state continuous production segments and multiple non-production shutdown segments.

[0059] ⑤ Extract the average consumption of each non-electric energy metering instrument within all non-production shutdown segments, and use it as the steady-state basic loss constant of the corresponding non-electric energy metering area. In the discrete industrial multi-energy flow spatiotemporal mapping mathematical model, constant basic losses are represented by independent constant terms.

[0060] Step 2: Construct a spatial dependency adjacency matrix and a process time delay matrix based on the production process mechanism.

[0061] This application abstracts the real physical production line and material and energy transfer topology of the factory into two types of quantitative constraint matrices, adding strong physical process prior constraints to the convex optimization iterative algorithm, and filtering out false temporal correlations without causal relationship between material and energy transfer from the root.

[0062] Step 2.1 Complete construction process of spatial dependency adjacency matrix.

[0063] Spatial Dependency Adjacency Matrix This is used to fully characterize whether there is a direct material and energy transmission process coupling relationship between each power monitoring node and each non-electric energy metering node, and to fully construct the execution process:

[0064] ① Retrieve the original information of the bill of materials and the original data of the production process flow plan layout topology stored in the factory's manufacturing execution system;

[0065] ② Traverse all power monitoring nodes and non-electric energy metering nodes for pairing and combination, and assign values ​​in two scenarios:

[0066] Scenario 1: If the power-consuming production equipment controlled by the power monitoring instrument and the heat-consuming and gas-consuming production equipment controlled by the non-electric energy metering instrument are located in the same core reaction, drying, or extraction process unit, or are directly physically interconnected through a closed fluid transport pipeline or a continuous material conveyor belt, and there is a continuous material and energy transmission process, then the corresponding element of the spatially dependent adjacency matrix is ​​assigned a value of 1.

[0067] Scenario 2: The power-consuming production equipment controlled by power monitoring instruments and the heat-consuming and gas-consuming production equipment controlled by non-electric energy metering instruments belong to independent workshops and there is no direct material, fluid, or energy transmission channel. In this case, the corresponding element of the spatial dependency adjacency matrix is ​​assigned a value of 0.

[0068] ③ After the matrix is ​​initialized, a sparse matrix structure is naturally formed, which greatly compresses the parameter search range of subsequent convex optimization iterative algorithms.

[0069] Taking a paper mill as an example, the power metering instrument for the pulping pump and the non-electric energy metering instrument for the steam main pipe of the drying section are directly interconnected through a closed pulp conveying pipeline, and the corresponding element of the spatial dependency adjacency matrix is ​​assigned a value of 1; the power metering instrument for the lighting of the administrative office building and the non-electric energy metering instrument for the steam main pipe of the drying section have no material or fluid transmission channels, and the corresponding element of the spatial dependency adjacency matrix is ​​assigned a value of 0.

[0070] Step 2.2 Global minimum cost timing alignment operation to construct the process time delay matrix

[0071] In industrial production processes, the transfer of materials and fluids within closed pipelines and intermediate silos involves a fixed physical lag time. When the electrical load collected by power metering instruments changes, the corresponding steam and natural gas consumption does not respond synchronously, resulting in a fixed process transfer lag. Accurately quantifying the process transfer lag time step is a key step in improving the accuracy of instantaneous non-electrical energy consumption estimation. The complete calculation process is as follows:

[0072] ① Only power-non-electric nodes that are marked as physically connected and have a value of 1 within the spatially dependent adjacency matrix are subjected to global minimum cost time alignment operation; for nodes without process coupling, the process transmission delay time step is directly set to 0, reducing the overall computing power consumption.

[0073] ② Perform cubic spline smoothing interpolation on the original time series of low-frequency non-electric energy cumulative consumption to generate a pseudo-high-frequency non-electric energy consumption time series with the same time granularity as the high-frequency power active power sequence data.

[0074] ③ Perform global minimum cost time series alignment operation on the high-frequency power active power sequence and the smoothed pseudo-high-frequency non-electric energy consumption time series. Euclidean distance is used as the sequence distance metric. The time series alignment path is subject to a unidirectional time series extension constraint. The mathematical expression for the operation is as follows:

[0075]

[0076] In the formula: P represents the set of all legal time-aligned paths. This represents the pairing of the k-th time series points on the alignment path. This represents the Euclidean distance between two sets of time points.

[0077] ④ Solve the global minimum cost timing alignment operation to obtain the optimal timing alignment path, extract the timing offset steps corresponding to the optimal timing alignment path as the process transfer hysteresis time step, and fill it into the process time delay matrix. Corresponding position.

[0078] Step 3: Construction of a discrete industrial multi-energy flow spatiotemporal mapping mathematical model and offline iterative optimization training for convex optimization.

[0079] Step 3.1 Complete Expression of Discretized Industrial Multi-Energy Flow Spatiotemporal Mapping Mathematical Model

[0080] The instantaneous non-electric energy consumption of a single non-electric energy metering region at micro-sampling time t is composed of three parts: a linear superposition term of multi-channel retrospective hysteresis active power, a local process nonlinearity correction term, and a steady-state basic loss constant term. The discrete mathematical expression is as follows:

[0081]

[0082] in, The estimated value of instantaneous non-electric energy consumption of the m-th non-electric energy metering region at micro-sampling time t;

[0083] The core weight coefficients within the spatiotemporal correlation matrix have the physical meaning of the incremental non-electric energy consumption associated with the process, which is associated with the unit of active power consumption.

[0084] Retrospective analysis of the nth power meter Historical active power sampling values ​​after the transmission hysteresis time step of each process.

[0085] The process transfer hysteresis time step is stored internally in the process time delay matrix;

[0086] The local process nonlinearity correction term is used to capture the nonlinear energy consumption response characteristics caused by phase change reactions, temperature threshold triggering, and catalytic reactions. In the engineering implementation process, a second-order polynomial kernel function or a radial basis kernel function can be used to construct this term. In conventional continuous linear production sections, this term can be omitted to simplify online simulation calculations.

[0087] The steady-state basic loss constant corresponding to the m-th non-electric energy metering area corresponds to the constant energy consumption formed by fluid pipeline heat dissipation and minor leakage in the pipeline network during equipment shutdown.

[0088] Step 3.2 Complete Logic for Constructing the Joint Optimization Objective Function Integrating Multiple Constraints

[0089] Non-electric energy metering instruments can only output low-frequency periodic cumulative consumption and cannot provide instantaneous consumption calibration values ​​at each micro-sampling moment. Therefore, the physical conservation of the total energy consumption within each complete observation cycle of low-frequency non-electric energy is used as a hard equality constraint, and a joint objective function for minimizing convex optimization is constructed by integrating three types of loss terms:

[0090]

[0091] in,

[0092] (1) Macroeconomic aggregate conservation L2 error term, i.e., the first term of the formula: This represents the actual cumulative consumption of non-electric energy metering instruments uploaded during the kth low-frequency metering cycle. The representative model is used to extrapolate the instantaneous non-electric energy consumption at each high-frequency sampling time to obtain the periodic predicted total consumption; the L2 mean square error is used to measure the deviation between the predicted total and the actual measured total, and the model is forced to satisfy the physical conservation of energy on a macroscopic measurement scale.

[0093] (2) Spatial topology mask L1 sparse regularization term, i.e. the second term of the formula: the symbol ⊙ represents the Hadamard element-wise product operation, which applies L1 norm constraint after the spatial dependency adjacency matrix is ​​used as a numerical mask and the spatiotemporal correlation matrix is ​​multiplied element-wise; the function of this term is to force the weight coefficients corresponding to nodes without process coupling relationship to converge to close to 0, and filter out false associations that are only temporally accidental synchronizations and have no causal relationship of material and energy transfer; the L1 sparse regularization term is accompanied by a penalty hyperparameter. The value range is fixed at 0.05 to 0.2.

[0094] (3) Weight stabilization L2 norm regularization term, i.e. the third term of the formula: apply a second penalty constraint to all weight vectors of the spatiotemporal correlation matrix to solve the problem of multicollinearity of input features caused by the synchronous start and stop of multiple similar electrical devices, and avoid extreme expansion of the value of a single weight coefficient and model fitting instability; L2 norm regularization term is matched with penalty hyperparameter. The value range is fixed at 0.01 to 0.1.

[0095] The optimal values ​​for the two types of penalty hyperparameters were determined through K-fold cross-validation.

[0096] Step 3.3 Solving and persistently storing parameters using the alternating direction multiplier iterative algorithm with integrated coordinate descent iterative logic.

[0097] The joint optimization objective function includes an L1 norm nondifferentiable piecewise function and a high-dimensional sparse spatiotemporal correlation matrix. Conventional gradient descent iterative algorithms cannot converge stably. This invention employs an alternating direction multiplier iterative algorithm that integrates coordinate descent sub-iteration logic to complete the convex optimization solution.

[0098] ① Introduce auxiliary variables to split the original joint optimization objective function into multiple low-complexity independent subproblems, and alternately iterate and update the spatiotemporal correlation matrix weights, auxiliary variables, and Lagrange multipliers.

[0099] ② Each independent subproblem executes coordinate descent sub-iteration logic to update the weight values ​​of the spatiotemporal correlation matrix row by row and column by column, which greatly improves the iteration convergence speed in high-dimensional sparse matrix scenarios.

[0100] ③ Set the convergence criterion for iteration: When the change in the calculated value of the loss function is less than 1e-5 in two consecutive iterations, the iteration calculation process is considered complete.

[0101] ④ After iterative convergence, three types of core model parameters are output: the optimal spatiotemporal correlation matrix. Optimal process time delay matrix Steady-state fundamental loss constant vector for each region .

[0102] ⑤ The three types of core model parameters are completely and persistently written into the industrial edge gateway's memory database, supporting millisecond-level read and call during the online real-time inversion calculation stage.

[0103] Step 4: Online real-time instantaneous non-electric energy consumption projection and integrated calculation of instantaneous carbon emissions from comprehensive energy sources.

[0104] Step 4.1 Complete Process of Real-Time Vector Simulation of Instantaneous Non-Electric Energy Consumption for the Entire Plant

[0105] ① The system continuously collects real-time active power sampling values ​​from all power metering instruments in the plant at fixed high-frequency power data sampling intervals, and splices them to generate a real-time active power vector. .

[0106] ② Retrieve the optimal spatiotemporal correlation matrix, optimal process time delay matrix, and all model parameters of steady-state basic loss constants for each region from the industrial edge gateway memory database.

[0107] ③ Retrieve back data from each power metering instrument within the local ring-shaped time-series buffer of the industrial edge gateway. The historical active power sampling values ​​of the process transmission hysteresis time step.

[0108] ④ Substitute each non-electric energy metering region into the discrete industrial multi-energy flow spatiotemporal mapping mathematical model to calculate the instantaneous non-electric energy consumption at the current micro-sampling moment.

[0109] ⑤ By splicing together the instantaneous non-electric energy consumption data of all M non-electric energy metering areas, a vector of instantaneous non-electric energy consumption for the entire plant is generated. .

[0110] The entire simulation operation only includes time buffer lookup, linear weighting, and basic summation operations. Conventional industrial edge gateways can complete all calculation operations in milliseconds, meeting the computing power requirements for real-time monitoring at the minute level.

[0111] Step 4.2 Complete Process of Real-Time Fusion Calculation of Instantaneous Carbon Emissions of Integrated Energy for Industrial Users

[0112] The system incorporates a standardized carbon emission factor database that fully complies with the "Guidelines for the Compilation of Provincial Greenhouse Gas Inventories" and the "Guidelines for Enterprise Greenhouse Gas Emission Accounting Methods and Reporting," and constructs two standardized emission intensity diagonal matrices to distinguish between indirect carbon emissions from purchased electricity and direct carbon emissions from fossil fuels.

[0113] Diagonal matrix of carbon emission factors for electricity The diagonal position stores the indirect carbon emission intensity value of the power grid in the corresponding region. The diagonal element can be dynamically updated according to the time-of-use power generation structure of the regional power grid and the terms of the enterprise's green power procurement agreement.

[0114] Diagonal matrix of non-electric energy carbon emission factors The diagonal positions store the direct carbon emission intensity values ​​of various media, including natural gas combustion, purchased steam, compressed air, and refrigerant emissions.

[0115] Formula for calculating the instantaneous carbon emissions of all industrial users in the plant at the micro-sampling time t:

[0116]

[0117] in,

[0118] —Instantaneous carbon emissions of total energy consumption by all industrial users in the plant at the micro-sampling time t;

[0119] — An N-dimensional, all-one column vector used to summarize the indirect carbon emissions from electricity corresponding to all electricity metering instruments;

[0120] — An M-dimensional, all-one column vector used to summarize the direct carbon emissions from fossil fuels corresponding to all non-electric energy metering instruments.

[0121] The formula calculates the indirect greenhouse gas carbon emissions from the consumption of purchased electricity, as well as the direct greenhouse gas carbon emissions from the consumption of natural gas, steam, and compressed air. The system automatically performs a complete matrix operation at each high-frequency power data sampling interval, continuously outputting a real-time dynamic monitoring data stream of integrated energy carbon emissions for industrial users at minute-level granularity. This data can be pushed in real-time to enterprise energy management platforms and carbon emission monitoring visualization dashboards.

[0122] Step 5: Complete process for low-frequency metrology closed-loop residual verification and adaptive incremental update of model parameters

[0123] Industrial equipment ages annually, seasonal temperature and humidity changes, production scheduling adjustments, and raw material batch attributes continuously alter the coupling ratio between active power and instantaneous non-electric energy consumption. Fixed model parameters accumulate measurement errors over long-term operation. This invention employs a low-frequency metering closed-loop residual verification and lightweight incremental update process to achieve autonomous adaptation of model parameters to changing operating conditions throughout their entire lifecycle.

[0124] Step 5.1 Calculation process for relative numerical deviation of low-frequency period:

[0125] For each set of non-electric energy metering instruments that receive the complete low-frequency cycle real cumulative consumption reading, the corresponding high-frequency extrapolated instantaneous non-electric energy consumption within the current low-frequency observation window is accumulated, and the relative numerical deviation between the extrapolated cumulative total consumption and the instrument's real cumulative consumption is calculated.

[0126] Step 5.2 Deviation tolerance threshold determination branch logic:

[0127] A deviation tolerance threshold is preset, with the standard engineering preset value being 3%.

[0128] If the calculated relative numerical deviation is less than or equal to the deviation tolerance threshold, the model's calculation accuracy meets the preset standard, and the system directly proceeds to the next high-frequency power data sampling cycle to perform instantaneous non-electric energy consumption extrapolation calculations.

[0129] If the calculated relative numerical deviation is greater than the deviation tolerance threshold, the model calculation accuracy is not up to standard. The background asynchronously starts the model parameter incremental retraining thread. The retraining thread runs independently and does not occupy the computing power of the real-time inference main thread.

[0130] Step 5.3 Update the training dataset using the first-in-first-out time-series sample queue:

[0131] The offline training dataset is managed by a first-in-first-out time series sample queue. After the queue storage length reaches a preset threshold, the old historical working condition time series samples stored at the front of the queue are removed, and complete energy consumption time series samples corresponding to the current low-frequency cycle with excessive deviation are added to ensure that the training dataset continuously matches the latest industrial field working conditions.

[0132] Step 5.4 Lightweight fine-tuning of model parameters using incremental gradient descent algorithm:

[0133] There is no need to re-execute the complete convex optimization offline training from scratch. Local weight fine-tuning is carried out based on the optimal spatiotemporal correlation matrix and optimal process time delay matrix solidified in the industrial edge gateway memory database. The learning rate of the incremental gradient descent algorithm is fixed in the range of 0.001 to 0.01, and the total number of iterations of a single lightweight fine-tuning is controlled between 50 and 200.

[0134] Step 5.5 After the update, the model parameters are synchronized and overwritten in storage:

[0135] After all the lightweight fine-tuning operations are completed, the new optimal spatiotemporal correlation matrix and the new optimal process time delay matrix synchronously cover and replace the original model parameters in the industrial edge gateway's memory database. The instantaneous non-electric energy consumption extrapolation calculation for the next high-frequency power data sampling cycle directly calls the updated parameters, forming a complete adaptive iterative closed loop.

[0136] The applicant declares that the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention fall within the protection and disclosure scope of the present invention.

Claims

1. A method for auxiliary calculation of integrated energy carbon emissions of industrial users based on a spatiotemporal correlation matrix, characterized in that, It includes two independent modules: an offline modeling and training phase and an online real-time inversion calculation phase, specifically comprising the following steps: S1: Acquisition of raw data on multi-source heterogeneous energy consumption and multi-dimensional spatiotemporal synchronous standardization and feature cleaning. Collect high-frequency active power sequence data output from all power monitoring nodes in the industrial site, and simultaneously collect low-frequency non-electric energy cumulative consumption data output from all non-electric energy metering nodes in the industrial site; the observation period corresponding to the low-frequency non-electric energy cumulative consumption data is an integer multiple of the sampling interval of the high-frequency active power sequence data. Lagrange polynomial interpolation is used to perform numerical completion processing on the missing data segments in the high-frequency power active power sequence data; an anomaly identification algorithm based on the absolute deviation of the median is used to identify abnormal peak pulse data points in the power active power sequence; and the abnormal data points are replaced by smoothing the effective power data inside the sliding window to generate a standardized fine-grained power active power load matrix. An adaptive accumulation and mutation identification algorithm is used to continuously monitor the first-order difference sequence of the total power curve of all power metering instruments. The time nodes corresponding to the difference values ​​that exceed the preset sensitivity parameters in multiple consecutive segments are marked as production condition mutation boundary nodes. Based on all production condition mutation boundary nodes, the complete historical energy consumption time series is divided into two types of time series segments: steady-state continuous production segments and non-production shutdown segments. The constant consumption of each non-electric energy metering instrument within all non-production shutdown segments is extracted as the steady-state basic loss constant of the corresponding non-electric energy metering area. S2: Constructing a spatial dependency adjacency matrix and a process time delay matrix based on industrial production process mechanisms. Retrieve the original bill of materials information and the original topology data of the production process flow layout stored within the factory's manufacturing execution system; when the power-consuming production equipment controlled by the power monitoring instrument and the heat-consuming and gas-consuming production equipment controlled by the non-electric energy metering instrument are located in the same core reaction process area or are directly physically interconnected through a closed fluid transport pipeline, assign an initial physical connectivity value at the corresponding position in the spatial dependency adjacency matrix; when there is no direct material flow channel between the power-consuming production equipment controlled by the power monitoring instrument and the heat-consuming and gas-consuming production equipment controlled by the non-electric energy metering instrument, assign a value of zero at the corresponding position in the spatial dependency adjacency matrix; The global minimum cost time alignment operation is performed only on the pairing of power-non-power nodes marked as physically connected within the spatially dependent adjacency matrix. The optimal time alignment path is solved for the high-frequency power active power sequence and the low-frequency non-power energy cumulative consumption sequence after cubic spline smoothing. The process transmission hysteresis time step corresponding to the optimal time alignment path is extracted, and the complete process time delay matrix is ​​constructed by filling in all process transmission hysteresis time steps. S3: Construct a discretized industrial multi-energy flow spatiotemporal mapping mathematical model and perform convex optimization offline iterative training. A discrete industrial multi-energy flow spatiotemporal mapping mathematical model is constructed to characterize the quantitative correlation between the instantaneous non-electric energy consumption in a single non-electric energy metering area and the historical active power data after backtracking through the process transmission lag time step. Taking the physical conservation of the total energy consumption within each complete observation cycle of low-frequency non-electric energy as the equality constraint, a joint optimization objective function is constructed, consisting of a macroscopic total conservation error term, a spatial topological mask L1 sparse regularization term, and a weighted stable L2 norm regularization term. An alternating direction multiplier iteration algorithm with integrated coordinate descending sub-iteration logic is used to perform alternating iterative solution operations on the joint optimization objective function. After the iterative calculation process converges, the three core model parameters obtained by the solution, namely the optimal spatiotemporal correlation matrix, the optimal process time delay matrix, and the steady-state basic loss constant of each region, are completely and persistently stored in the industrial edge gateway memory database. S4: Real-time online simulation of the plant's instantaneous non-electric energy consumption vector and calculation of the integrated instantaneous carbon emissions of industrial users. According to the fixed sampling period of high-frequency active power sequence data, the real-time active power sampling values ​​output by all power metering instruments are continuously collected. The three core model parameters of optimal spatiotemporal correlation matrix, optimal process time delay matrix and steady-state basic loss constant of each region are retrieved from the persistent storage of the industrial edge gateway memory database. The instantaneous non-electric energy consumption at the current micro sampling moment is deduced for each region and spliced ​​to generate the instantaneous non-electric energy consumption vector of the whole plant. Two types of standardized emission intensity diagonal matrices are constructed: a diagonal matrix of carbon emission factors for electricity and a diagonal matrix of carbon emission factors for non-electric energy. Based on the linear superposition operation of the complete matrix, the instantaneous carbon emissions of industrial users' comprehensive energy at any micro-sampling time are calculated in real time, and a dynamic real-time monitoring data stream of industrial users' comprehensive energy carbon emissions with minute-level granularity is continuously output. S5: Low-frequency metrology closed-loop residual verification and adaptive incremental update of model parameters Each time a new low-frequency cycle's actual cumulative non-electric energy consumption reading is received from an industrial site non-electric energy metering instrument, the relative numerical deviation between the total instantaneous non-electric energy consumption calculated by the model within the current complete low-frequency observation cycle and the actual cumulative consumption of the instrument is calculated. When the calculated relative numerical deviation exceeds a pre-set deviation tolerance threshold, the background asynchronously starts an incremental retraining thread for model parameters. A first-in-first-out time-series sample queue is used to complete the real-time update of the offline training dataset. The complete set of original model parameters fixed in the industrial edge gateway's memory database is updated using an incremental gradient descent algorithm for lightweight fine-tuning. After fine-tuning, the new optimal spatiotemporal correlation matrix and the new optimal process time delay matrix synchronously replace the original model parameters in the industrial edge gateway's memory database, which are then used for the calculation of instantaneous non-electric energy consumption in the next high-frequency power data sampling cycle.

2. The method for auxiliary calculation of integrated energy carbon emissions of industrial users based on spatiotemporal correlation matrix according to claim 1, characterized in that: In step S1, the Lagrange polynomial interpolation used for filling in missing data segments in communication is either third-order or fourth-order Lagrange polynomial interpolation. The threshold for identifying abnormal spike pulse data points is a value between 3 and 5 times the absolute deviation of the median in the sliding window. The sliding window completely includes 5 to 15 high-frequency power active power sequence data sampling steps.

3. The method for auxiliary calculation of integrated energy carbon emissions of industrial users based on spatiotemporal correlation matrix according to claim 1, characterized in that: The sensitivity parameter for the internal adaptive accumulation and mutation identification algorithm in step S1 ranges from 0.02 to 0.

1. When three or more consecutive first-order difference values ​​exceed the preset sensitivity parameter, the corresponding time node is determined as the production condition mutation boundary node.

4. The method for auxiliary calculation of integrated energy carbon emissions of industrial users based on spatiotemporal correlation matrix according to claim 1, characterized in that: Step S2 uses Euclidean distance as the distance metric between the two sets of time series data for the global minimum cost time alignment operation. A one-way time extension constraint is added to the time alignment path. The global minimum cost time alignment operation only performs the complete calculation operation on the pairing of power-non-power nodes marked as physically connected within the spatially dependent adjacency matrix.

5. The method for auxiliary calculation of integrated energy carbon emissions of industrial users based on spatiotemporal correlation matrix according to claim 1, characterized in that: The complete computational logic of the L1 sparse regularization term for the internal spatial topology mask in step S3 is as follows: after performing Hadamard element-wise multiplication on the spatial dependency adjacency matrix and the spatiotemporal correlation matrix, an L1 norm constraint is applied; the value range of the penalty hyperparameter for the L1 sparse regularization term of the spatial topology mask is 0.05 to 0.2; the value range of the penalty hyperparameter for the weighted stable L2 norm regularization term is 0.01 to 0.1; the optimal values ​​of the two types of penalty hyperparameters are determined by K-fold cross-validation.

6. The method for auxiliary calculation of integrated energy carbon emissions of industrial users based on spatiotemporal correlation matrix according to claim 1, characterized in that: Step S3 internal discretized industrial multi-energy flow spatiotemporal mapping mathematical model includes local process nonlinearity correction terms; the local process nonlinearity correction terms are constructed using second-order polynomial kernel functions or radial basis kernel functions; the local process nonlinearity correction terms are only enabled for special production process sections with phase change reactions and temperature trigger threshold effects, while the local process nonlinearity correction terms can be omitted for conventional continuous linear production sections to simplify online real-time simulation calculations.

7. The method for auxiliary calculation of integrated energy carbon emissions of industrial users based on spatiotemporal correlation matrix according to claim 1, characterized in that: The learning rate for the incremental gradient descent algorithm in step S5 is between 0.001 and 0.01; the total number of iterations for a single lightweight fine-tuning is controlled between 50 and 200. The default value for the deviation tolerance threshold in the low-frequency metrology closed-loop residual verification process is 3%.

8. The method for auxiliary calculation of integrated energy carbon emissions of industrial users based on spatiotemporal correlation matrix according to claim 1, characterized in that: The indirect carbon emission intensity values ​​of electricity stored at each diagonal position of the internal electricity carbon emission factor diagonal matrix in step S4 can be dynamically updated and adjusted based on the real-time power generation structure of the regional power grid and the written terms of the enterprise's green electricity procurement agreement.

Images