A method for carbon emission tracing based on the synergy between flue gas energy consumption and copper smelting process

Abstract

This invention relates to a carbon emission tracing method based on coordinated energy consumption in copper smelting flue gas, belonging to the field of copper smelting technology. Addressing the bottleneck of nonlinear time lag in material and energy flows during long processes, this method extracts the physical parameters of the smelting furnace outlet and calculates them into flue gas physical parameters, generating a time-stamped source-end feature sequence. It also extracts the power consumption of downstream flue gas treatment stages and converts it into an end-stage indirect carbon emission rate sequence. A dynamic time warping algorithm is used to calculate the shortest warped path between the two sets of sequences. According to the time distortion relationship indicated by the path, the end-stage sequence is nonlinearly mapped, stretched, or compressed on the time axis. Finally, the mapped and aligned end-stage data nodes are bound to the source-end data nodes for tracing. This invention effectively eliminates the fluid thermal inertia and dynamic delay effects caused by long-distance pipeline transmission, avoids mismatched carbon emission data, and achieves accurate and coordinated tracing of cross-process responsible carbon emissions.

Description

Technical Field

[0001] This invention belongs to the field of copper smelting technology, specifically, it relates to a method for carbon emission traceability based on the coordinated energy consumption of flue gas in copper smelting processes. Background Technology

[0002] Copper smelting is a typical long-process industry characterized by high energy consumption and high carbon emissions. In the integrated copper smelting production process, carbon emissions mainly consist of two parts: first, direct carbon emissions from fossil fuel combustion and concentrate oxidation desulfurization in upstream smelting furnaces (such as flash smelters and converters); and second, indirect carbon emissions from the large amount of electricity consumed in downstream flue gas treatment processes (including waste heat boilers, electrostatic precipitators, dynamic wave scrubbers, and main exhaust fans in acid production workshops, etc.) to treat and transport high-temperature flue gas. To achieve refined carbon emission control and operational optimization, it is essential to accurately trace and allocate downstream indirect electrical carbon emissions to specific upstream smelting operation batches, thus achieving carbon emission source tracing in conjunction with flue gas energy consumption.

[0003] However, in actual industrial production sites, due to objective reasons for equipment layout, there is usually a physical spatial span of hundreds or even thousands of meters between the upstream smelting core area and the downstream flue gas acid production area, which are connected by a large-diameter flue gas pipeline network. This long physical transmission path results in extremely significant fluid thermal inertia and time lag effects between the material flow (high-temperature flue gas) and the energy flow (power consumption of downstream equipment). For example, if the upstream smelting furnace experiences a sudden increase in flue gas enthalpy due to an increase in the amount of feed, this energy shock wave needs to travel through pipelines for several minutes to more than ten minutes, experiencing eddy current obstruction and heat dissipation along the way, before finally forcing the downstream acid production main exhaust fan frequency converter to increase its operating frequency to cope with the load change.

[0004] Existing carbon emission tracing and accounting methods commonly suffer from distortion due to spatiotemporal misalignment when processing data from long-process industrial systems. Current conventional accounting methods often employ a "static time section method," directly aligning and superimposing upstream direct carbon emission data with downstream indirect power consumption data within the same clock cycle. This method, which ignores the physical delays in fluid transmission, incorrectly matches downstream power consumption peaks caused by processing earlier flue gas with upstream operation records that may already be in a stable holding period, resulting in severe causal inconsistencies. Some optimized existing technologies attempt to introduce a "fixed time constant translation method" to compensate for pipeline delays. However, because smelting operations require frequent switching between the blowing period (high flow rate) and the holding period (low flow rate), the flue gas velocity, density, and heat transfer rate within the pipeline exhibit highly nonlinear dynamic changes. A fixed time translation constant is fundamentally unsuitable for the fluctuating fluid transmission conditions and struggles to eliminate dynamically changing phase differences. Therefore, how to overcome the nonlinear fluid time delay caused by long-distance pipelines by relying on data processing mechanisms without increasing the high cost of on-site hardware modification, and accurately trace the indirect carbon emissions of downstream dynamic fluctuations back to the real physical triggering source upstream, has become a technical bottleneck that urgently needs to be overcome in the field of refined management of carbon emissions in copper smelting. Summary of the Invention

[0005] The purpose of this invention is to provide a carbon emission tracing method that coordinates energy consumption in copper smelting flue gas. This method solves the technical problem in existing long-process industrial systems where, due to the nonlinear thermal inertia delay caused by long-distance fluid transport, downstream indirect electrical carbon emissions cannot be accurately aligned and traced back to the actual upstream process triggering source. Furthermore, it easily leads to the misattribution of energy consumption from equipment malfunctions to upstream production operations.

[0006] The objective of this invention can be achieved through the following technical solutions: A carbon emission tracing method for the coordinated energy consumption of flue gas in copper smelting process includes extracting the source-end physical parameter sequence of flue gas at the outlet of the smelting furnace, and extracting the end-end power consumption of the downstream flue gas treatment process and converting it into an end-end indirect carbon emission rate sequence. The method further includes the following steps: Based on the source-end physical parameter sequence, the flue gas physical odor is calculated in real time, and a source-end odor value fluctuation characteristic sequence with timestamps is generated. The shortest regularized path between the source-end carbon emission rate fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence is calculated using a dynamic time warping algorithm. The end-indirect carbon emission rate sequence is mapped and aligned on the time axis according to the time warp relationship indicated by the shortest regularization path. The data nodes corresponding to the mapped and aligned end-point indirect carbon emission rate sequence are linked to the data nodes corresponding to the source-point value fluctuation characteristic sequence for source tracing.

[0007] Furthermore, after calculating the shortest warped path using the dynamic time warping algorithm, a frequency domain phase synchronization calibration step is performed using cross wavelet transform, specifically including: performing continuous wavelet transform on the source-end sine wave fluctuation characteristic sequence and the terminal indirect carbon emission rate sequence to extract the high condensation energy region of the two sets of sequences in the multi-scale frequency domain.

[0008] Furthermore, after extracting the high condensation energy region, the phase angle difference between the source-end carbon emission rate fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence within the high condensation energy region is calculated.

[0009] Furthermore, the phase angle difference is converted into dynamic hysteresis data, and the data nodes bound to the source are time-axis shifted and matched based on the dynamic hysteresis data.

[0010] Furthermore, the method also includes a step of constructing a directed graph topology network for the entire system, wherein the directed graph topology network uses carbon emission devices as network nodes and physical pipes connecting the devices as network edges; and uses a spatiotemporal graph convolutional neural network to process the source-end carbon emission fluctuation feature sequence and the end-end indirect carbon emission rate sequence.

[0011] Furthermore, an attention mechanism network layer is configured in the spatiotemporal graph convolutional neural network, and a dynamic attention weight allocation matrix that maps downstream network nodes to multiple upstream network nodes is extracted from the output of the attention mechanism network layer. The weight ratio of the source binding is allocated according to the dynamic attention weight allocation matrix.

[0012] Furthermore, in the process of calculating the shortest warped path using the dynamic time warping algorithm, a thermal inertia energy dissipation damping model of the downstream flue gas treatment equipment is constructed, and the damping coefficient output based on the thermal inertia energy dissipation damping model is extracted. The damping coefficient is then input as a dynamic penalty weight value into the cost function of the dynamic time warping algorithm.

[0013] Furthermore, the numerical values ​​of adjacent time nodes in the end-of-line indirect carbon emission rate sequence are extracted to calculate the ramp slope. When the ramp slope is greater than the maximum flue gas energy ramp rate threshold set by the equipment, a set extreme value penalty term is applied to the current calculation node in the cost function to block the regular matching of the current time series path.

[0014] Furthermore, during the continuous wavelet transform of the source-end carbon emission fluctuation characteristic sequence and the terminal indirect carbon emission rate sequence, real-time flue gas velocity data measured by sensors is acquired. Based on the flue gas velocity data, a mapping relationship function between the wavelet basis function window width and the reciprocal of the velocity is established. The time window width for performing the continuous wavelet transform is adjusted in real time according to the mapping relationship function.

[0015] The beneficial effects of this invention are: 1. This invention does not employ the traditional fixed-constant translation method. Instead, it extracts source-end physical parameters to calculate a physical sequence characterizing the actual energy pulsation and innovatively uses the Dynamic Time Warping (DTW) algorithm to calculate the shortest warped path between this sequence and the end-of-pipe carbon emission sequence. This scheme can elastically stretch or compress the end-of-pipe response sequence on the time axis according to the fluctuating fluid transport conditions in the actual pipeline network, fundamentally eliminating the fluid thermal inertia and nonlinear time delay caused by long-flow pipeline transmission, and achieving misaligned and precise binding of dynamic carbon emissions across processes.

[0016] 2. To address the blind matching defects that are prone to occur in purely mathematical alignment algorithms, this invention pre-constructs a thermal inertia energy consumption damping model for downstream equipment and sets a maximum flue gas energy ramp rate threshold when calculating the regularized path. When the ramp rate of terminal power consumption is detected to exceed the limit of this fluid dynamics theory, an extreme value penalty term is applied to the algorithm's cost function. This mechanism successfully separates abnormal power consumption peaks caused by mechanical failures of downstream equipment such as fan jamming and short circuits from normal process fluctuations, effectively preventing the algorithm from erroneously transferring abnormal physical losses to the upstream smelting process, and greatly improving the engineering authenticity and anti-interference capability of the traceability logic.

[0017] 3. For complex and interconnected operating conditions involving the convergence of flue gas from multiple furnaces, this invention further utilizes a Spatiotemporal Graph Convolutional Neural Network (STGCN) to process the feature sequences of multiple nodes and configures an attention mechanism network layer to output a dynamic attention weight allocation matrix. This scheme overcomes the limitations of single time-series tracing, dynamically analyzing the energy causal transfer network hidden beneath complex physical pipelines. It accurately cuts and distributes the total carbon emissions of a single downstream common node to multiple different upstream smelting carbon source nodes according to a scientific weight ratio, solving the problem of unquantifiable decoupling of carbon emission responsibility under mixed multi-flow conditions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is the basic flowchart for time-domain regularized carbon emission source tracing in this invention; Figure 2 This is a flowchart of the frequency domain cross-wavelet phase synchronization process of the present invention; Figure 3 This is a flowchart of the spatiotemporal graph network topology carbon emission tracing process of this invention; Figure 4 This is a flowchart illustrating the process of preventing misjudgment through physical damping constraint optimization in this invention. Figure 5 This is a flowchart of the flow velocity adaptive wavelet frequency domain analysis of the present invention. Detailed Implementation

[0020] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1 like Figure 1 As shown, a method for carbon emission tracing based on the coordinated energy consumption of flue gas in copper smelting includes extracting the source-end physical parameter sequence of the flue gas exiting the smelting furnace, and extracting the end-stage power consumption of the downstream flue gas treatment process and converting it into an end-stage indirect carbon emission rate sequence; the method further includes the following steps: Based on the source-end physical parameter sequence, the flue gas physical odor is calculated in real time, and a source-end odor value fluctuation characteristic sequence with timestamps is generated. The shortest regularized path between the source-end carbon emission rate fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence is calculated using a dynamic time warping algorithm. The end-indirect carbon emission rate sequence is mapped and aligned on the time axis according to the time warp relationship indicated by the shortest regularization path. The data nodes corresponding to the mapped and aligned end-point indirect carbon emission rate sequence are linked to the data nodes corresponding to the source-point value fluctuation characteristic sequence for source tracing.

[0022] The 1200℃ flue gas discharged from the flash smelting furnace must pass through a settling chamber, a 350-meter waste heat boiler cooling section, an electrostatic precipitator, and a dynamic wave scrubber before reaching the main exhaust fan in the acid production workshop. This physical transmission involves a nonlinear delay of 3 to 15 minutes. If the power consumption of the main exhaust fan increases by 200kW instantaneously, the attribution of the increase in indirect carbon emissions becomes spatiotemporally misaligned, making it difficult to determine whether it originates from the current furnace condition fluctuations or a specific feeding operation 10 minutes prior. In the early stages of the project, an attempt was made to inject a radioactive isotope tracer into the main pipeline to calibrate the flow rate delay. The on-site flue gas volumetric flow rate oscillates between 100,000 and 300,000 Nm³ / h with the blowing process cycle, and the static delay constant deviates from the true value in less than 24 hours. A sliding window cross-correlation analysis was then used; for this type of unsteady fluid, the calculated lag time exhibits a step-like jump on the time axis.

[0023] The system ultimately selected a dynamic time warping algorithm with physical pipeline constraints for time-domain alignment. An edge computing gateway based on a heterogeneous architecture of ARM and FPGA was deployed within the workshop electrical control cabinet, with a sampling period locked at 10ms. In the underlying data acquisition stage, the system extracted the source-end physical parameter sequence of the flue gas exiting the smelting furnace. The smelting furnace outlet is equipped with a platinum-rhodium thermocouple, an absolute pressure transmitter, and a tunable semiconductor laser absorption spectrometer (TDLAS). This source-end physical parameter sequence specifically includes on-site measured temperature, absolute pressure, and multi-component volume concentration data. Simultaneously, the system extracted the end-of-pipe power consumption of the downstream flue gas treatment stage and converted it into an end-of-pipe indirect carbon emission rate sequence.

[0024] Edge gateway directly reads the three-phase active power of the low-voltage switchgear for the main exhaust fan and conversion blower in the acid production workshop. Calling the real-time emission factor issued by the power grid dispatch center (Taking 0.57 kgCO2 / kWh), perform multiply-add matrix operations:

[0025] The computational output includes a sequence of indirect carbon emission rates at the end of the flow path with consecutive timestamps. A single temperature scalar of 1150°C cannot characterize the true work potential of the fluid. After receiving the data, the GPU-accelerated cluster calculates the flue gas physical energy in real time based on the source-end physical parameter sequence, generating a time-stamped sequence of source-end energy value fluctuation characteristics. The state engine invokes the steady-state flow energy equation:

[0026] for Physical power at source at time (kW); Mass flow rate (kg / s) measured for Pitot tubes; and These are the real-time specific enthalpy and specific entropy obtained by interpolating from the NIST standard property library based on the temperature, pressure, and concentration values ​​in the source-end physical parameter sequence. The reference state is 25℃ and 101.3kPa. A continuous source-end carbon emission rate fluctuation characteristic sequence is generated. The algorithm incorporates a spatiotemporal alignment core module, employing a dynamic time warping algorithm to calculate the shortest warped path between the source-end carbon emission rate fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence.

[0027] Basic algorithm for constructing Euclidean distance matrix And based on the state transition equation, the minimum cumulative cost is sought:

[0028] Model testing using 30 days of historical on-site operating data revealed logical divergences at the fluid dynamics level in the basic algorithm's output. Constrained by the single mathematical requirement of minimizing global distance, the algorithm forcibly matched the peak power consumption surge of the acid-producing blower caused by grid connection switching at 2 AM on a certain day to the thermal fluctuation peak of the smelting furnace at 5 PM the previous day. This mapping spanned a period of 9 hours.

[0029] In a real physical environment, high-temperature flue gas remaining in a pipeline for 9 hours would cause the temperature to drop below the acid dew point, resulting in the condensation of a large amount of sulfuric acid vapor into liquid acid. The gas volume would collapse drastically, and flow transmission would cease. Pure mathematical optimization probes, lacking industrial common-sense boundary constraints, generate unrealistic spatiotemporal topological connections to fit local numerical features. The system forcibly injects pipeline fluid dynamics boundary constraints at the matrix operation level. The entire plant's pipeline and instrumentation flow is extracted. Figure 3 Based on the data, the total internal volume of the main flue gas conveying pipeline network is determined to be 15,200 m³. Combining the minimum allowable flow velocity for preventing ash accumulation of 5.5 m / s under the frequency conversion control of the centrifugal fan and the full-load design flow velocity of 24.8 m / s, the shortest fluid transmission time in the pipeline network is calculated to be approximately 105 seconds, and the longest dead zone time is approximately 980 seconds.

[0030] This 105-980 second physical effective time interval, defined by the mechanical structure of the physical pipeline network, is converted into the dynamic index deviation threshold in the distance matrix using a 10ms sampling step. (The range spans from 10,500 to 98,000 matrix index bits).

[0031] The GPU iteratively calculates each coordinate grid node. When the cumulative cost is reached, the control bus forcibly embeds absolute critical value judgment conditions in the optimization logic: Index Difference The algorithm represents the time span it covers in the digital space. If the physical time corresponding to this span exceeds the 980-second limit, the current optimization path is deemed to deviate from the laws of fluid dynamics transport.

[0032] When an out-of-limit condition is triggered, the GPU controller initiates a hardware interrupt, directly severing the current state transition pipeline process, and incurring cumulative costs for that node. The saturation extremum is forcibly overwritten. This extremum is set as the upper limit for single-precision floating-point numbers in the IEEE 754 standard. This low-level register write operation constructs a numerical insulating wall within the optimization cost matrix. The algorithm probe, driven by the goal of minimizing the global cost, reads... At that time, its accumulated gradient becomes invalid, forcing the algorithm to actively abandon illegal paths that exceed the limit, and suppressing the search trajectory within the envelope defined by the fluid dynamics laws from 105 seconds to 980 seconds.

[0033] Under strict fluid boundary constraints, the GPU ultimately calculates the shortest regularized path that conforms to thermodynamic causality. The system maps and aligns the end-stage indirect carbon emission rate sequence on the time axis according to the time distortion relationship indicated by the shortest regularized path. Under extremely low flow rate conditions during the furnace holding period, the algorithm stretches the downstream energy consumption response sequence in digital space to compensate for physical transport delays; under high flow rate conditions with multiple nozzles operating at full load, the algorithm dynamically compresses the time axis to match transient energy impacts. After unifying the time axis benchmark, the system performs a clearing operation, binding the data nodes corresponding to the mapped and aligned end-stage indirect carbon emission rate sequence with the data nodes corresponding to the source-end sludge fluctuation characteristic sequence. The additional 540 kWh of electricity consumed by the acid production workshop to handle specific flue gas fluctuations is precisely converted into equivalent carbon emissions and, as a responsibility label, is rigidly written into the database record of the 300 Nm³ / h oxygen concentration adjustment operation performed by a specific shift at 14:30.

[0034] Example 2 like Figure 2 As shown, after calculating the shortest warped path using the dynamic time warping algorithm, a frequency domain phase synchronization calibration step is performed using cross wavelet transform. Specifically, this includes: performing continuous wavelet transform on the source-end carbon emission fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence to extract the high condensation energy region of the two sequences in the multi-scale frequency domain; after extracting the high condensation energy region, calculating the phase angle difference between the source-end carbon emission fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence within the high condensation energy region; converting the phase angle difference into dynamic hysteresis data, and performing time axis translation matching on the data nodes bound to the source tracing based on the dynamic hysteresis data.

[0035] This embodiment provides a frequency domain phase synchronization calibration scheme based on cross-wavelet transform. As a core parallel replacement and deep calibration strategy for the pure time-domain warping algorithm, its core engineering purpose is to completely overcome the complex signal phase distortion problem caused by the extreme variable operating conditions of copper smelting systems. In the actual heavy metal pyrometallurgical process, the thermodynamic operating state of the system is not kept absolutely constant. When the smelting system switches frequently between the blowing period and the holding period, the intensity of the physicochemical reactions inside the furnace can change by orders of magnitude.

[0036] This sudden change in furnace conditions directly leads to an extremely steep nonlinear change in flue gas velocity and enthalpy density within the physical pipes. When the underlying sensor network detects that the flue gas velocity rise rate per unit time exceeds the conventional steady-state boundary of fluid dynamics, the alignment algorithm, which is purely based on time-domain amplitude approximation, lacks a global view in the frequency dimension. This leads to severe waveform distortion and misalignment / overlap due to excessive and frequent stretching or compression of the time axis, thereby generating a false traceability link that violates the causal law of thermodynamic mass transport. To ensure the absolute rigidity and matching continuity of the carbon emission traceability chain under extreme operating conditions, the system's underlying data flow scheduling bus will quickly trigger a hardware interrupt mechanism, seamlessly switching the computing power channel to the frequency domain transformation analysis link constructed in this embodiment.

[0037] In processing this industrial data stream carrying massive amounts of high-frequency environmental noise and process abrupt changes, this system employs a heterogeneous computing cluster deployed in the core data center at the plant level. This computing cluster uses a front-end field-programmable gate array (FPGA) as the high-frequency throughput and feature buffering hub for the underlying raw discrete signals, and relies on a high-performance digital signal processor (DSP) equipped with a multi-core tensor processing unit to fully execute the core algorithm code involving high-dimensional complex function integration and frequency domain feature matrix extraction. This robust computing power allocation architecture ensures that the real-time solution capability of the entire frequency domain transformation and phase synchronization model can perfectly cover any harsh nonlinear flow field abrupt changes in the industrial environment. After receiving the discrete-time series pushed at high speed by the front-end FPGA through direct memory access technology, the DSP's primary action is to perform continuous wavelet transform on the source-end variability fluctuation feature sequence and the terminal indirect carbon emission rate sequence.

[0038] The flue gas energy fluctuation signal in industrial sites exhibits typical non-stationary random characteristics. Traditional Fourier transforms can only reveal the macroscopic frequency components contained in the global signal, but completely lose the specific physical time node information of the occurrence of the particular frequency during the global integration process. The continuous wavelet transform model, by introducing a mother wavelet function with infinite translation and nonlinear scaling capabilities, achieves joint localization and accurate analysis on the two-dimensional topological plane interwoven with the time and frequency axes.

[0039] The mathematical formula for the continuous wavelet transform invoked by the digital signal processor for any one-dimensional discrete physical sequence of input is constructed as follows:

[0040] In this mathematical formula for continuous wavelet transform, the variables Represents the original one-dimensional time series The wavelet coefficients obtained under the dual constraints of a specific scale and a specific time shift are mathematically a high-dimensional complex matrix containing real and imaginary parts. The square of the absolute value of the complex coefficient is directly proportional to the local energy concentration intensity of the original physical signal at that frequency band and at that time.

[0041] variable The scaling factor represents the dynamic control of the scaling and broadening of the wavelet basis function. Numerically, the scaling factor is strictly inversely proportional to the actual physical equivalent frequency of the signal. A larger scaling factor results in a proportionally wider basis function, specifically designed to match and capture low-frequency, long-period trend signals at the system's core, such as the slow heat capacity drift during furnace insulation. Conversely, a smaller scaling factor results in a proportionally narrower basis function, specifically designed to capture steep, high-frequency abrupt changes caused by instantaneous high-pressure pulverized coal injection.

[0042] variable Represents the translation factor, indicating the point-by-point sliding step size coordinate of the mother wavelet function on the time scale. Variable This represents the complex conjugate form of the mother wavelet function. To achieve optimal physical isomorphism between the purely mathematical model and the complex dissipation laws of industrial physics, the digital signal processor is forced to use the Moray wavelet as the underlying mother wavelet function. The internal mathematical structure of the Moray wavelet is represented by the product of a Gaussian envelope and a complex sine wave. Its combined topological characteristics of time-domain decay and frequency-domain oscillation have a very high degree of fit with the diffusion and dissipation characteristics of thermodynamic energy shock waves in long-distance physical pipelines. It can perfectly map a one-dimensional time-domain fluctuation sequence to a two-dimensional complex matrix surface with extremely low energy leakage cost.

[0043] After completing the two-dimensional frequency domain expansion and reshaping of the independent sequences, the digital signal processor then drives the algorithm engine to enter the cross-wavelet analysis stage to explore the causal relationship between upstream and downstream energy pulsations, in order to extract the high-condensation energy regions of the two sets of sequences in the multi-scale frequency domain. The core engineering function of the cross-wavelet transform model is to directly expose and accurately locate the high-energy overlapping topological parts of two independent physical time series at the same frequency and the same physical time.

[0044] Assume the wavelet transform calculation result matrix of the source-end variability characteristic sequence is as follows: The wavelet transform matrix of the terminal indirect carbon emission rate sequence is as follows: The digital signal processor constructs the mathematical formula for calculating the cross spectral density based on these two independent matrices as follows:

[0045] In this cross-spectral mathematical formula, the variables For the final generated cross-wavelet spectral density matrix, the variables are... This represents the complex conjugate transpose of the wavelet coefficient matrix of the end-of-pipe indirect carbon emission rate sequence. The absolute value of the elements of the cross-spectral density matrix is ​​the cross-wavelet power, which intuitively and quantitatively represents the macroscopic synergistic resonance intensity of the abnormal energy fluctuations in the upstream smelting furnace flue gas and the surge in indirect power consumption of the downstream main exhaust fan at the same spatiotemporal scale. Real-world industrial sensor data streams are always filled with stray background noise caused by electromagnetic interference from high-power frequency converters or eccentric vibrations from heavy fluid machinery. To accurately extract the thermodynamic causal resonance region truly caused by the physical high-temperature flue gas fluid transport from this vast ocean of disordered noise, the system introduces a rigorous statistical hypothesis testing mechanism when performing feature extraction to determine whether the cross-wavelet power truly reaches the significance threshold condition for process causal correlation.

[0046] Considering that time series of industrial thermal parameters generally exhibit strong first-order autoregressive memory characteristics, a joint test mathematical formula based on a red noise background random spectrum is established by the digital signal processor:

[0047] In this hypothesis testing mathematical formula, the variable These are parameters obtained at a given 95% confidence level based on multiple Monte Carlo stochastic simulations and iterations. Variables With variables These represent the source and terminal sequences at specific scales, respectively. The theoretical red noise background power characteristic spectrum under constraints. Only when the calculated cross-wavelet power is strictly greater than the upper limit of the theoretical red noise threshold will the digital signal processor (DSP) mark and extract the high-value connected component in the two-dimensional matrix as a high-condensation-energy region. This successfully selected high-condensation-energy region in the real physical world corresponds precisely to a sustained large-scale thermal shock wave caused by a sudden change in high-intensity oxygen concentration or concentrated feeding. This shock wave contains real energy density to penetrate the damping of long-distance complex pipeline networks, ultimately forcing downstream power equipment to generate fluctuating electrical loads. After successfully extracting and locating the high-condensation-energy region, the DSP's computing resources penetrate deeply into this specific region to calculate the phase angle difference between the source-end carbon emission fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence within the high-condensation-energy region. In the multi-scale complex frequency domain, this core indicator, the phase angle difference, directly quantifies the relative time lead or lag physical mapping relationship between two physical wave signals at a specific resonant dominant frequency.

[0048] The digital signal processor extracts all complex cross-spectral density data points within the boundary of the high-condensation energy region, and performs local cross-phase angle calculations one by one. The mathematical formula is as follows:

[0049] In the mathematical formula for calculating the phase angle, the variables This represents the local cross-phase angle calculated by the computational model under the dual constraints of a specific time-shift reference and a specific frequency domain scale. Its value is strictly distributed within a closed interval from negative to positive pi. Operator Operator specifically designed for extracting the imaginary part of elements in a complex cross-spectral density matrix. It is specifically designed to extract the real part values ​​of the corresponding complex matrix elements. Within this highly condensed energy region, which has been rigorously verified to have causal relationships, the system extracts and calculates the average phase angle difference. To completely eliminate the numerical truncation jump phenomenon that inevitably occurs when the phase angle crosses the physical boundary of positive and negative pi, the digital signal processor incorporates a continuous phase dewinding filtering algorithm to ensure that the evolution of the phase angle difference on the time axis presents an absolutely smooth and continuous physical transition state.

[0050] After obtaining the smooth, continuous phase angle difference curve, the system must map it back from a relative angular dimension to an absolute physical second time dimension, converting the phase angle difference into dynamic hysteresis data. The mathematical formula for phase angle conversion dynamic hysteresis called by the digital signal processor is defined as follows: In this core transformation formula, the variables This refers to the dynamic lag data ultimately acquired by the system, which accurately represents the absolute number of seconds that a specific form of physical energy is actually consumed in as it is transmitted and diffused from the high-temperature outlet of the smelting furnace to the downstream flue gas treatment equipment.

[0051] variable This refers to the dominant physical equivalent frequency of resonance within the currently locked high-condensation energy region. This dynamic hysteresis data is a nonlinear dynamic parameter that evolves in real time with the process, completely eliminating the engineering fallacy of static constant delay in traditional algorithms. The system performs time-axis translation matching on the data nodes bound to the source tracing based on this dynamic hysteresis data. The digital signal processor shifts the downstream end-point indirect carbon emission rate sequence as a whole on the digital time axis. The absolute time step ensures that the peak of energy consumption at the end and the peak of physical energy fluctuation at the source end that caused the peak are absolutely diagonally coincident on the same time scale. Through the above-mentioned hardware and software co-processing architecture that deeply integrates continuous wavelet frequency domain expansion and cross-spectral density phase extraction, this embodiment perfectly constructs a parallel physical verification link independent of pure time-domain regularity.

[0052] The system completely avoids the algorithmic trap of time-domain amplitude distortion under extreme variable operating conditions, transforming the complex physical problem of carbon emission tracing into a mathematical and geometric solution process of finding high-energy resonance regions and their inherent phase angle differences in a two-dimensional frequency domain topological space. This alignment mechanism based on frequency domain phase synchronization is perfectly compatible with the nonlinear fluid impact caused by the high-frequency switching between blowing and holding states in large-scale heavy metal smelting facilities. Low-level control engineers with expertise in digital signal processing, based on a deep understanding of the principles of cross-wavelet transform and hardware tensor acceleration architecture, and relying on mainstream high-performance computing nodes in industry, have put the frequency domain transformation and time axis translation matching scheme described in this embodiment into engineering practice. This enables the establishment of a robust cross-process collaborative carbon emission responsibility tracking system in any high-energy-consuming industrial scenario with large-span fluid condition changes.

[0053] Example 3 like Figure 3 As shown, the method further includes a step of constructing a directed graph topology network for the entire system. This network uses carbon emission equipment as network nodes and physical pipelines connecting the equipment as network edges. A spatiotemporal graph convolutional neural network is used to process the source-end carbon emission fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence. An attention mechanism network layer is configured in the spatiotemporal graph convolutional neural network, and a dynamic attention weight allocation matrix, output by the attention mechanism network layer and mapping from downstream network nodes to multiple upstream network nodes, is extracted. The weight ratio for source tracing is allocated according to this dynamic attention weight allocation matrix. The production architecture of modern large-scale heavy non-ferrous metal smelting and chemical production bases is not a single, linear, parallel, independent process, but rather exhibits an extremely complex cross-network topology.

[0054] The extremely high-temperature flue gas containing high concentrations of sulfur dioxide emitted from multiple flash smelting furnaces and converters during different smelting cycles is uniformly drawn and collected into a large-volume common flue gas manifold. Subsequently, according to the plant's overall production scheduling instructions and pipeline pressure balance requirements, it is diverted in parallel to multiple independent waste heat boiler cooling systems, dynamic wave scrubbing dust removal systems, and multiple parallel acid-producing converter branches. In this many-to-many physical topology with multiple parallel flue gas branches and complex cross-energy-consuming equipment, a drastic fluctuation in the load of a specific downstream acid-producing main exhaust fan or a sharp increase in its electrical active power results in hundreds of kilowatts of additional electrical energy being converted into indirect carbon emissions. This is often the combined result of the complex physicochemical reactions occurring in multiple upstream smelting furnaces. Single time-series alignment algorithms completely fail in the face of this highly nonlinear multi-source mixing and multi-path diffusion scenario, and are fundamentally unable to define the initial mapping responsibility under multi-stream mixing conditions. The project team had tried to install orifice plate flow meters or insertion Annubar flow meters in each branch flue to rigidly allocate downstream electricity carbon emissions according to the physical flow rate. However, the smelting flue gas, which was as high as 1,100 degrees Celsius and contained a large amount of liquid metal dust, could corrode the stainless steel probe in less than two weeks. The huge pressure drop in the pipeline caused by the physical measuring device was an engineering obstacle that the smelting process could not tolerate.

[0055] An attempt was made to introduce a blind source separation algorithm to decouple mixed signals at the pure software level. However, this statistical model strongly relies on the assumption of linear independent mixing, and its operational logic completely ignores the spatial orientation of the physical pipeline network and the thermal inertia of the fluid. The calculated carbon emission responsibility weights deviate significantly from the actual field operation records. To overcome the challenge of collaborative carbon emission traceability under this many-to-many network topology, this embodiment has undergone a deep reconstruction of the underlying hardware architecture of information flow processing. The system deploys a dedicated graph computing acceleration server cluster at the backbone node of the plant-level industrial internet. This cluster is equipped with a tensor processing unit array specifically optimized for instruction set optimization of high-dimensional sparse matrix operations. Industrial-grade intelligent gateways with edge inference capabilities are distributed and deployed in the field control cabinets of each workshop to preprocess massive amounts of high-frequency industrial sensor data in real time with millisecond-level low latency, providing a solid physical computing foundation for complex graph neural network models.

[0056] After receiving the basic ledger and real-time operating status data of all equipment in the plant, the dedicated graph computing acceleration server cluster initiates the process of constructing a directed graph topology network for the entire system. The server abstracts and reconstructs the entire complex physical system of copper smelting and flue gas treatment into a high-dimensional mathematical directed graph structure. The server uniformly and rigorously defines all carbon emission source smelting equipment and downstream energy-consuming key fluid machinery within the system as network nodes, and uniformly abstracts the complex flues and valves connecting these specific physical devices as directed network edges connecting the network nodes. To accurately replicate the fluid dynamics characteristics and pipeline resistance distribution of the physical world in the digital world, the server must assign precise connection weights to each directed network edge. These weights directly characterize the smoothness of energy transmission and the strength of thermodynamic coupling between two specific nodes. When the physical resistance within the pipeline increases or the transmission distance lengthens, the attenuation of fluid energy transfer exhibits an exponential trend.

[0057] The server has a built-in mathematical formula for calculating the elements of the topological adjacency matrix:

[0058] In this topological weighting mathematical formula, the variables Represents the first term in a directed graph topology network. The physical node points to the first The directed edge weights of each physical node, i.e., the specific element values ​​in the global static adjacency matrix. (Variable) This parameter represents the actual physical centerline length of the flue connecting these two physical devices. (Variable) This represents the equivalent hydraulic diameter parameter of the connecting flue section. (Variable) It is a comprehensive friction damping coefficient dynamically calibrated based on the kinematic viscosity of flue gas and the absolute roughness of the pipe wall.

[0059] The energy transfer connectivity between nodes is proportional to the square of the equivalent hydraulic diameter and inversely proportional to the physical pipe length, exhibiting a strict nonlinear evolution law. If the system determines that there is no direct pipe connection between two devices in physical space, the system forcibly sets this critical value to zero, i.e., element... Strictly equal to zero. The server traverses the entire plant's process flow diagrams and equipment asset database, instantiating a global static adjacency matrix in massive video memory that perfectly maps the physical connection impedance characteristics of the real plant. This lays an absolutely reliable spatial topology foundation for subsequent high-order graph convolution operations. After completing the digital reconstruction of the spatial topology network, the dedicated graph computing acceleration server cluster retrieves real-time sensor data pushed by the edge gateway and uses a spatiotemporal graph convolutional neural network to process the source-end sludge fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence on all nodes.

[0060] The spatiotemporal graph convolutional neural network is designed to simultaneously capture the material flow and diffusion relationships in the physical spatial topology of industrial data, as well as the thermodynamic hysteresis relationships on the time series axis. The server performs graph convolution matrix operations in the spatial dimension, enabling each energy-consuming or carbon-emitting node to perceive the energy state changes of its direct or indirect neighboring nodes through mathematical aggregation. The mathematical formula for spatial graph convolution aggregation calculation called by the server is constructed as follows:

[0061] In this core graph calculation formula, variables Indicates the neural network's first... The node feature state matrix of the layer, in the input layer, loads the source-end physical emission sequence of each upstream node and the terminal indirect carbon emission rate sequence of each downstream node, which are collected from the field and preprocessed. Variables This represents the global static adjacency matrix with an added unit self-loop matrix. The engineering purpose of introducing the unit self-loop is to ensure that each physical node can retain its historical device operating characteristics during iterative state updates. (Variable) It is the degree matrix, used to perform strict symmetric normalization on the adjacency matrix to prevent gradient explosion of node feature values ​​due to the convergence of multiple branches during the forward propagation operation of deep networks.

[0062] variable These are the trainable weight matrix parameters in this network layer, iteratively optimized by the tensor processing unit based on massive amounts of historical data. (Variables) The activation function is nonlinear, and the system forces the use of a modified linear unit function to ensure that the neural network can fit complex industrial thermodynamic nonlinear surfaces. Through the high-frequency brushing operation of this spatial convolutional layer, the downstream acid converter node seamlessly receives and integrates the energy mixing state information from multiple upstream flash smelting furnace nodes in the mathematical feature dimension. In the time dimension processing, the server applies a one-dimensional standard convolutional kernel to perform a fixed-step sliding scan along the time series axis on the aggregated node feature matrix, extracting the evolutionary cycle law of energy pulsation in the time dimension, completely breaking down the two-dimensional computational barrier between the spatial topology matrix and the time delay sequence. While spatiotemporal graph convolution operations alone can extract the spatiotemporal coupling features of the entire network, they still cannot accurately quantify which upstream node's abnormal operating condition fluctuation caused the current surge in energy consumption of a specific downstream node. In order to find the optimal carbon emission responsibility matching relationship from the complex many-to-many entangled network, the system configures and trains a core attention mechanism network layer in the high-level architecture of this neural network. The essential attribute of the attention mechanism is to dynamically score and weight the mixed information from multiple sources. The server extracts high-dimensional features from the deep output of the model and constructs a dynamic attention weight allocation matrix. When the system needs to determine the proportion by which the increase in electricity consumption and carbon emissions of a specific downstream acid-cooling main exhaust fan node should be allocated to multiple upstream smelting furnace nodes, the attention layer uses the current energy consumption surge value of the downstream node as the query vector and the historical abnormal fluctuation value value of each upstream node as the key vector.

[0063] The attention scoring and normalization mathematical formula executed by the server are set as follows:

[0064] In this crucial formula for calculating attention, the variable This refers to the dynamic attention weight coefficients ultimately extracted by the system. These coefficients quantify the attention weights at a specific physical time. At that moment, the upstream The energy state of the flue gas at the first smelting node affects the downstream... The percentage of carbon emission responsibility contribution from changes in the energy consumption status of each defense node. (Variable) Indicates that the downstream node is The hidden layer feature vector at time step 1. Variables This represents the historical hidden layer feature vector of the upstream node after fully considering the lag time of fluid transmission in the pipeline. Variables It is a linear transformation parameter matrix shared by the network. The double vertical bar symbol indicates that the upstream and downstream high-dimensional feature vectors are concatenated into a matrix.

[0065] variable This is the weight vector of a single-layer feedforward neural network, used to map the concatenated high-dimensional feature space into a scalar evaluation score representing the coupling strength. The denominator performs a standard normalized exponential function operation, with set notation. Representing all downstream [items] The total set of upstream kiln nodes for which each node has a physically connected topological path.

[0066] Through this forced normalization mechanism, the system ensures that the sum of all upstream responsibility weight coefficients for a given downstream energy-consuming node is strictly equal to 100%, achieving an absolute balanced distribution of carbon emission responsibility in the topological mathematical space. The dynamic attention weight allocation matrix output by this formula accurately reveals the distribution law of the energy causal transfer network hidden beneath the intricate physical steel pipe shell. After the tensor processing unit outputs a smooth and accurate dynamic attention weight allocation matrix, the system proceeds to the final carbon emission ledger settlement stage. The dedicated graph computing acceleration server cluster allocates the weight ratios of the source binding according to the dynamic attention weight allocation matrix, strictly dividing and allocating the actual physical indirect carbon emissions generated by downstream nodes according to the calculated weight ratios and binding them to the corresponding upstream smelting carbon source nodes. In the real pipeline fluid transmission environment, due to local damage to the insulation layer of the massive pipeline or unavoidable environmental temperature differences, the fluid energy will inevitably generate heat loss to the environment during the transmission process of hundreds of meters. This lost fluid energy is also a carbon emission cost that the system must pay to maintain the stable operation of the overall chemical process.

[0067] Without compensation for heat loss, the total amount of electricity carbon recorded through reverse carbon emission tracing will not be able to match the total electricity consumption of the entire plant. The server will forcibly introduce a weighted allocation and compensation correction mathematical formula in the final carbon emission reverse carbon tracing binding stage:

[0068] In this final settlement formula for carbon emission source tracing, the variables are... This represents the upstream [number] [unit] within the set monthly or team statistical assessment period. Each core smelting node should independently bear the total indirect carbon emission traceability amount derived from the power consumption of multiple downstream network devices, calculated in tons of CO2 equivalent. This represents the total set of high-energy-consuming nodes in all downstream flue gas treatment processes throughout the plant. (Variable) Indicates the downstream first Each node Instantaneous carbon emissions from electricity are measured and calculated at all times using underlying smart meters. (Variable) Assign a percentage to the dynamic attention weights output by the pre-attention mechanism layer. (Variable) It is a pre-set environmental heat loss compensation coefficient, which is specifically used to correct and compensate for energy loss from nodes. Transport to node The system experiences additional power load due to convection cooling in the pipeline network during the long physical process. Through this complete tensor formula derivation and calculation allocation, the huge electricity consumption of the acid production workshop, which was originally regarded as a shared indicator for the entire plant at the physical workshop level, is neatly segmented and broken down, and back-linked to a specific flash smelting furnace or a specific batch of converter blowing operations with absolutely precise data granularity. This traceability scheme, deeply embedded in the pipeline network topology and jointly driven by spatiotemporal graph networks and attention mechanisms, eliminates the engineering dead end of carbon emission data not being decoupled under complex multi-stream mixed conditions, and establishes a technical foundation with extreme physical robustness and theoretical rigor for the refined carbon emission assessment and control of heavy non-ferrous metal smelting processes.

[0069] Example 4 like Figure 4 As shown, in the process of calculating the shortest warp path using the dynamic time warping algorithm, a thermal inertia energy consumption damping model of the downstream flue gas treatment equipment is constructed. The damping coefficient output by the thermal inertia energy consumption damping model is extracted, and this damping coefficient is input as a dynamic penalty weight value into the cost function of the dynamic time warping algorithm. The numerical values ​​of adjacent time nodes in the end-of-pipe indirect carbon emission rate sequence are extracted to calculate the ramp-up slope. When the ramp-up slope exceeds the maximum flue gas energy ramp-up rate threshold set by the equipment, a set extreme value penalty term is applied to the current calculation node in the cost function to block the warping matching of the current time series path. This embodiment provides a time-domain warping optimization scheme based on physical damping constraints. As a deep exploration and underlying reconstruction of the purely mathematical time series alignment algorithm, its core engineering purpose is to completely eradicate the fatal flaw of abnormal energy consumption spikes caused by mechanical failures or abnormal physical losses of downstream flue gas treatment equipment in extremely harsh and complex industrial heavy-load environments, which are incorrectly traced back to and transferred to upstream smelting process fluctuations by the basic algorithm.

[0070] In the real copper smelting production chain, the basic dynamic time warping algorithm, as a mathematical tool purely for exploring numerical approximation and morphological matching, completely ignores the physical laws of the industrial environment when searching for the shortest warping path. When core power equipment in the downstream defense line, such as giant main exhaust fans used to extract high-temperature flue gas or heavy-duty centrifugal pumps used to transport high-concentration sulfuric acid solutions, suffers from severe rotor bearing lubrication problems due to long-term continuous operation, thick scale buildup on the impeller surface, or partial short circuits in the motor stator windings, the mechanical transmission resistance of the equipment will instantly change exponentially and increase sharply. In order to maintain the constant flow and absolute pressure setpoint required by the process system, the high-power frequency converter at the bottom will automatically and forcibly increase the output frequency and stator voltage, which will directly cause an extremely drastic sudden increase in the instantaneous electrical active power of the equipment. This power spike, purely caused by internal mechanical friction or electrical faults, will leave a huge and steep data peak on the indirect carbon emission rate sequence at the end. If relying solely on a basic algorithm without physical constraints, the system, in pursuit of minimizing the global distance matrix, will forcibly distort and stretch the mapping relationship on the digital timeline, rigidly matching this false fault peak to the energy of a normally fluctuating flue gas discharged from the upstream furnace a few minutes earlier, thus generating an extremely absurd and seriously erroneous chain of evidence for tracing the source of the problem, which seriously deviates from the thermodynamic causal law.

[0071] To fundamentally prevent such erroneous matching that defies common sense in physics, this embodiment constructs a hardware-software co-constraint architecture that deeply integrates stringent physical boundary conditions into the underlying matrix operation core. At the hardware foundation level of system information flow processing, this embodiment deploys industrial-grade edge computing gateways integrating high-performance field-programmable gate arrays in the electrical control room close to the industrial production line. These edge gateways are directly coupled to the three-phase power quality analyzers and high-frequency vibration sensors of each core power equipment via high-frequency data buses, processing massive amounts of underlying equipment operation status awareness data in real time with extremely high time resolution. All discrete time series cleaned and preprocessed by the edge gateways are then continuously pushed to an independent graphics processor acceleration server cluster deployed in the central control data center through an industrial backbone network with deterministic low-latency characteristics.

[0072] This graphics processing unit (GPU) server cluster, equipped with massive high-bandwidth video memory and tens of thousands of concurrent tensor computing cores, will be fully responsible for executing the massive and complex physical state equations and high-dimensional constrained matrix optimization operations in this embodiment. The primary core action of the GPU server cluster upon receiving the startup command is to pre-build thermal inertia energy dissipation damping models for all critical flue gas treatment equipment in the downstream defense line within its high-speed video memory. In the complex industrial scenario where macroscopic fluid mechanics and heat transfer are intertwined, when the high-temperature, high-pressure flue gas emitted from the upstream smelting furnace, carrying enormous thermodynamic energy, flows through a complex network of pipes into downstream power equipment such as centrifugal fans, this massive input energy cannot be instantly and completely converted into the load power on the motor stator side under an ideal state of zero time delay. The massive cast iron casing, the heavy rotating impeller assembly, and the large volume of the internal fluid medium together constitute an energy buffer pool with enormous heat capacity and mechanical rotational inertia. The power consumption response curve of the equipment must strictly follow a gradual upward trajectory bound by physical laws. The rate of increase of its power value is proportional to the difference between the energy of the flue gas input at the front end and the current mechanical work state of the equipment, and is strictly inversely proportional to the comprehensive inertial resistance determined by the structural mass and specific heat capacity of the equipment itself.

[0073] To translate this physical damping effect into digital rules that computers can understand and execute, the graphics processing unit (GPU) server cluster constructs a first-order thermo-mechanical coupling dynamic damping differential equation for each individual device. The mathematical formula for this model is precisely defined as:

[0074] In the differential mathematical formula of this thermal inertia energy dissipation damping model, the core dependent variable is... This represents the instantaneous electrical active power of a specific downstream power device, collected in real time by the underlying smart meter and transmitted to the server. Its physical dimension is kilowatt, and this absolute value directly determines the system's current indirect carbon emission intensity. Time is the independent variable. A physical time scale representing continuous passage. The parameter on the right side of the equation. The steady-state gain coefficient, representing the energy conversion of the equipment under stable operating conditions, reflects the inherent efficiency of fluid machinery in converting fluid energy into mechanical work.

[0075] Core variables This represents the dynamic thermodynamic physical properties of flue gas actually flowing into the inlet section of the equipment from the upstream physical pipeline network at the same continuous time point. It is the most critical parameter on the left side of the equation. This refers to the pre-extracted and fixed comprehensive thermomechanical damping coefficient of the equipment. This damping coefficient is a physical parameter precisely measured and entered into a relational database through rigorous step response identification experiments under new and rated operating conditions. It represents the physical resilience of the equipment to resist external energy impacts and maintain its original operating state. Extracting the damping coefficient of the equipment under rated operating conditions is the cornerstone for establishing subsequent physical boundaries. This system of differential equations mathematically declares that any transient power jump attempting to exceed the equipment's inherent damping coefficient is a fault manifestation that violates basic physical principles. Based on the established differential model, the graphics processing unit server cluster further extrapolates to calculate the theoretical limit of power ramp-up that the equipment may experience under the most extreme process allowable boundaries.

[0076] The server retrieves the maximum allowable fluid load from the pipeline engineering design and derives the mathematical formula for the maximum flue gas energy ramp rate threshold:

[0077] In this core threshold definition formula, the parameters This refers to the maximum flue gas energy ramp rate threshold, which is strictly calculated, set, and dynamically updated by the system based on the objective physical attributes of the equipment. (Variable) This represents the theoretical maximum physical density of flue gas that can be transported at full load from the current pipeline network's physical structure and the upstream smelting furnace's designed capacity. (Variable) This refers to the real-time instantaneous load power of the power equipment in the current millisecond-level computing cycle. This maximum flue gas energy ramp rate threshold acts like an insurmountable moat, precisely defining the absolute physical upper limit of the legitimate carbon emission rate increase driven by real fluid heat transfer. After completing rigorous digital modeling of the physical boundaries, the computing resources of the graphics processor server cluster are fully shifted to the high-frequency monitoring and dynamic comparison interception defense line of real-time data streams. The server memory array continuously receives the end-indirect carbon emission rate sequence pushed by the underlying edge gateway and activates the high-frequency numerical differentiation logic unit to extract the numerical calculation ramp rate of adjacent time nodes in the end-indirect carbon emission rate sequence in real time.

[0078] To completely eliminate high-frequency noise spikes caused by electromagnetic interference from high-power frequency converters in industrial settings, which can lead to divergence in differential calculations, the server performs necessary smoothing corrections on the original time series using a low-pass filter before executing high-frequency differentiation. The mathematical formula for real-time extraction of the discrete slope is then used:

[0079] In this real-time slope calculation formula, the variable Represents the system at the current physical time node Numerical calculation of the ramp-up slope of the precisely calculated end-point indirect carbon emission rate sequence. Variables It is the latest core data of indirect carbon emission rate measurement points generated at the current time point, while the variables... This refers to the historical data of indirect carbon emission rates from the previous consecutive sampling period, stored in the graphics processor's cache. (Denominator parameter) It strictly corresponds to the fixed high-frequency sampling period of the analog-to-digital converter in the edge computing gateway.

[0080] After accurately calculating the current numerical ramp-up slope, the server's high-speed logic unit immediately executes a Boolean judgment instruction with the highest system priority, rigorously comparing the ramp-up slope with the maximum flue gas energy ramp-up rate threshold set based on the equipment's physical attributes. When the actual extracted numerical ramp-up slope is within or below the safe dead zone envelope of the maximum flue gas energy ramp-up rate threshold, the system logic determines that the current energy consumption increase trend fully conforms to the hysteresis law of fluid heat and mass transfer, belonging to the legitimate thermodynamic fluctuation characteristics allowed by the process. The underlying dynamic time warping basic alignment algorithm can then continue to execute the matrix filling operation smoothly.

[0081] However, once the graphics processor detects a drastic change in the actual ramp rate that exceeds the safety dead zone threshold—that is, when the actual ramp rate exceeds this threshold—it indicates that the surge in energy consumption exceeds the physical limits of fluid heat transfer and is caused by the equipment itself. The system control logic will then instantly trigger the highest level of intervention and blocking mechanism. The only explanation for this mathematical phenomenon of the ramp rate exceeding the physical threshold in the underlying physical world is that the energy source causing the core motor load to surge is not the upstream flue gas heat energy slowly transported through long-distance pipelines, but rather, very likely, abnormal mechanical energy consumption behavior of the equipment itself, such as fan bearing seizure and wear, severe slippage of the drive belt, or foreign object blockage inside the pump.

[0082] This surge beyond physical limits indicates the system has entered an extremely dangerous saturation anomaly zone. To prevent the blindly mathematical alignment algorithm from wrongly blaming this massive carbon emission surge caused by mechanical failure on the innocent front-end smelting operation team, the server immediately activates a defensive penalty mechanism, converting the damping coefficient into an extreme value penalty term, which is then input into the cost function of the dynamic time warping algorithm. The mathematical formula for generating and mapping this extreme value penalty term is rigorously constructed as follows:

[0083] In this dynamic penalty generation formula, the variable... This is the extremum penalty term generated by the system, which represents the time point in the dynamic time-warped global spatial distance matrix corresponding to the occurrence of mechanical anomalies in the downstream carbon emission sequence. With any upstream time node The penalty weight constant assigned to the combined grid coordinate points. When an out-of-limit judgment occurs, this penalty value will exhibit an explosive exponential nonlinear change trend. Parameters in the formula. This is a positive adjustment weighted gain constant manually set by the system algorithm engineer to enhance the matrix interception effect.

[0084] parameter This is the damping coefficient extracted beforehand, used here for correction and compensation calculations to maintain the absolute consistency of the internal dimensions of the exponential function. (Constant) This is the maximum floating-point physical limit that the graphics processing unit's hardware registers can safely support. This enormous extreme value penalty term will be directly and forcibly input into and superimposed on the core state transition cost function of the underlying dynamic time warping algorithm, thereby completely reconstructing the optimization terrain of the basic algorithm at the computational level.

[0085] The mathematical formula for the modified cost transfer with physical constraint damping penalty executed by the graphics processor server under this intervention evolves as follows:

[0086] In the final revised core cost function mathematical formula, matrix elements This represents the total global cost accumulated by the algorithm from the starting point of the digital matrix to the current specific grid coordinate point; this value is also the core output of the cost function. Parameters The absolute Euclidean distance represents the basic state characteristics of the upstream and downstream sequences at this point. The minimization function is responsible for backtracking to examine the historical cumulative costs of the three surrounding legal predecessor nodes and selecting the best matching path. The most crucial interception and blocking force comes from the extremum penalty term newly added to the end of the formula.

[0087] During the execution of thousands of high-concurrency matrix filling operations by the graphics processing unit (GPU), by imposing an extremum penalty term, once the algorithm's internal optimization probe attempts to traverse time nodes representing mechanical failures in downstream equipment along an aligned path, the superimposed massive extremum penalty term instantly raises the cumulative total value along that specific path direction to an astronomical number that the hardware registers can hold. Since the highest optimization criterion of the dynamic time warping algorithm is to achieve absolute minimization of global cost, this robust digital wall, derived from a physical damping model, forces the optimization algorithm to automatically abandon this false matching association, thus precisely and thoroughly blocking the warping matching of the current abnormal time series path at the algorithm's microscopic level. Through this comprehensive hardware and software co-constraint scheme that deeply integrates the fundamental laws of heat transfer and fluid mechanics with high-dimensional matrix dynamic programming algorithms, the system completes its underlying evolution, transforming into a highly intelligent industrial discriminator with the ability to make judgments based on physical common sense.

[0088] It cleverly utilizes the inherent thermal inertia and energy-dampening physical properties of downstream flue gas treatment equipment to establish an uncompromising physical security gate for every energy consumption data stream within the plant. Any abnormal energy consumption attempting to disguise itself as a mechanical failure of the equipment itself due to fluctuations in the upstream process will be exposed through a rigorous dynamic comparison between the numerical rise rate and the physical theoretical threshold, and will be completely intercepted and removed from the carbon emission traceability link before the digital barrier constructed by the extreme value penalty term. This highly integrated hardware and software design maximizes the protection of the physical purity and engineering authenticity of the traceability logic, ensuring that only carbon emissions caused by real fluctuations in flue gas are traced. This fundamentally guarantees that the intelligent control system can maintain absolutely clear causal discrimination even during the extremely frequent and complex equipment degradation and deterioration operation cycles in industrial sites. This makes the system automatically finding the optimal solution a stable and deliverable technical means in real-world engineering, building a truly robust and optimal anti-interference capability and an absolutely closed-loop physical logic carbon emission responsibility accurate identification system for the high-energy-consuming heavy non-ferrous metal smelting industry. Automation control and underlying algorithm engineers with professional qualifications in this field, based on a deep understanding of the core principles of the above-mentioned first-order differential equations and constrained dynamic programming, can fully utilize mainstream parallel accelerated computing hardware platforms in the industry to implement the complete technical solution described in this embodiment into field engineering practice without reservation.

[0089] Example 5 like Figure 5 As shown, during the continuous wavelet transform of the source-end carbon emission fluctuation characteristic sequence and the terminal indirect carbon emission rate sequence, real-time flue gas velocity data measured by sensors is acquired. Based on the flue gas velocity data, a mapping function between the wavelet basis function window width and the reciprocal of the velocity is established. The time window width for performing the continuous wavelet transform is adjusted in real time according to the mapping function. This embodiment provides a cross-wavelet high-frequency feature tracing scheme based on adaptive variable window for flue gas velocity. As the underlying adaptive evolution of the frequency domain transform tracing architecture, the core engineering purpose of this scheme is to completely overcome the dual destruction of signal processing frequency domain resolution and time domain resolution caused by the extreme variable operating conditions of heavy metal smelting systems. In the actual pyrometallurgical process, the operation of the smelting furnace is by no means in an ideal constant steady state. The process system needs to switch between operating conditions with a large span between the blowing period and the holding period according to the metallurgical phase transformation requirements. When the system is in the pure oxygen blowing period with multiple concentrate nozzles fully open, the intense exothermic oxidation reaction inside the furnace causes the instantaneous flue gas volume to change exponentially and non-linearly. The physical flue gas velocity in the main flue will climb to an extremely high level of over 25 meters per second in a very short time. When the system enters the heat preservation period or copper tapping period, the reaction inside the furnace almost stops, the main exhaust fan operates at a reduced frequency, and the flue gas velocity in the pipeline network will drop precipitously to an extremely low level of about 3 meters per second.

[0090] Faced with drastic conditions where flue gas velocity changes by nearly tenfold, the basic cross-wavelet transform algorithm with a fixed time window inevitably falls into the uncertainty principle trap inherent in signal processing. When the flue gas velocity is extremely high, the energy pulsation propagation time in the pipeline is extremely short, and transient process characteristics are fleeting. In this case, an excessively wide wavelet window will result in severely insufficient time-domain resolution, blurring multiple different transient operation peaks into a broad envelope and completely losing the precise time coordinates for carbon emission tracing. When the flue gas velocity is extremely slow, energy transfer exhibits a slow, low-frequency evolution trend. In this case, an excessively narrow wavelet window will lead to a lack of frequency-domain resolution, failing to capture the long-period resonant coupling relationship between upstream and downstream data. To fundamentally resolve this physical contradiction of frequency-domain resolution, the system constructs an adaptive variable-window control architecture that directly drives the underlying frequency-domain transform mathematical model using on-site fluid dynamics parameters.

[0091] The system is configured in the core data center of the plant with a heterogeneous edge computing micro-cluster equipped with a large-scale field-programmable gate array (FPGA) and a multi-channel parallel floating-point digital signal processor (DSP). The FPGA is specifically responsible for throughput and alignment of high-frequency pulse signals from multiple heterogeneous sensors in the field pipeline network with nanosecond-level extremely low latency. The multi-channel parallel DSP is fully responsible for performing the massive and complex adaptive window mapping calculation and high-dimensional complex function integration operations, ensuring that the computing power bandwidth of the entire adaptive tracing system can perfectly cover any extreme changes in the smelting conditions. In the process of performing continuous wavelet transforms on the source-end sludge fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence, the primary task of the heterogeneous edge computing micro-cluster is to synchronously acquire and analyze the flue gas velocity data measured in real time by the physical pipeline sensors. To ensure the absolute reliability of the velocity data and extremely high dynamic response frequency, the system directly interfaces with the array-type Pitot tubes and high-precision differential pressure transmitters installed on the straight section of the main flue gas duct at the outlet of the smelting furnace.

[0092] A field-programmable gate array (FPGA) acquires the total pressure and static pressure electrical signals output by the Pitot tube in real time at an extremely high sampling rate. A built-in digital Kalman filter removes nonlinear aerodynamic noise spikes caused by airflow oscillations in the Pitot tube backflush system in real time. The purified differential pressure sequence is then sent to a digital signal processor (DSP) for hydrodynamic state calculation. Based on Bernoulli's energy conservation equation, the DSP calculates the instantaneous absolute velocity of the flue gas at the current physical time point in real time. The mathematical formula for velocity measurement is defined as follows:

[0093] In this fluid dynamics calculation formula, the variables Represents the system at the current physical time node The actual flue gas flow rate, calculated instantaneously and accurately, has the physical dimension of meters per second. (Variable) Indicates the differential pressure transmitter in The absolute value of the dynamic differential pressure between the total pressure and static pressure at the pipe cross-section, measured at any given time, in Pascals. Variable The system calculates the current actual operating density of the flue gas, expressed in kilograms per cubic meter, by combining synchronously collected real-time flue gas temperature, absolute pressure, and real-time volume concentration of multiple components, and by interpolating using the industrial standard thermodynamic equations of state. Parameters This is the inherent instrument correction factor for the Pitot tube probe, used to compensate for the minor local drag disturbances caused by the probe geometry to the originally smooth laminar or turbulent flow field.

[0094] After continuously acquiring the instantaneous flow velocity sequence, the digital signal processor calculates the first-order time derivative of the flow velocity in real time to accurately capture the critical evolution point of the operating condition switch. When the absolute value of this first-order derivative exceeds the critical value for drastic changes in operating conditions preset by the system based on the fluid dynamics of the pipeline network capacity, it indicates that a violent oscillation has occurred in the flow field environment. The system data flow scheduling bus immediately blocks the traditional fixed-window wavelet transform logic pipeline, forcibly guiding the underlying computing resources to fully switch to the adaptive variable-window model processing link.

[0095] After entering the adaptive control link, the digital signal processor (DSP) activates and constructs the core dynamic adjustment control logic, establishing a dynamic mapping function between the wavelet basis function window width and the reciprocal of the flow velocity based on the flue gas velocity data. This mapping function constitutes the mathematical core of the entire adaptive source tracing optimization. Its design principle strictly follows the spatial transmission laws of solid fluid mechanics, namely, the faster the energy carrier moves in a pipe of fixed physical length, the shorter the absolute physical time required for it to pass through a specific cross-section, and the observation shutter time required by the algorithm must be shortened proportionally. The slower the energy carrier moves, the longer the observation exposure time required by the algorithm must be extended proportionally. The width parameter of the wavelet time window must be strictly inversely proportional to the value of the instantaneous flue gas velocity.

[0096] To precisely quantify this macroscopic physical law into algorithmic control parameters executable by the digital signal processor, the system invokes the adaptive window mapping mathematical formula as follows:

[0097] In this core mapping formula, the variables Represents a specific physical time point The theoretical wavelet time window broadening coefficient is preliminarily calculated based on the actual flow field conditions. This coefficient is a dimensionless mathematical scalar that directly determines the degree of expansion of the mother wavelet function on the digital time axis. Variables This refers to the instantaneous physical velocity of the flue gas calculated in real time during the preceding steps.

[0098] parameter It is a dynamically adaptive scaling gain coefficient pre-calibrated by system algorithm engineers based on the total physical centerline length of the entire plant's flue and the average resistance coefficient of the pipe wall. Its physical meaning is to smoothly map and convert the meter-per-second dimension from the field of fluid mechanics into a dimensionless scale spatial scaling parameter from the field of signal processing. Parameter This is the minimum intrinsic basic window width reference value that the mother wavelet function must retain to maintain its integrability and admissibility conditions at the applied mathematical level. In extremely harsh industrial environments, differential pressure sensors may occasionally output false minimum differential pressure signals close to absolute zero due to blockage of the pressure tap by sticky soot in the pipes or strong electromagnetic interference from high-power frequency converters, or output extremely high peak flow noise at the moment of abnormal pressure release from explosion-proof doors.

[0099] Directly inputting the theoretical window width value into matrix integration operations can easily lead to a memory overflow error caused by division by zero in the underlying algorithm engine, or result in an infinitely large window, thus completely paralyzing the cluster's computing resources. The digital signal processor forcibly connects a physical boundary constraint model with dead-zone and saturation characteristics in series at the output of the mapping function. The mathematical formula for this saturation cutoff protection is defined as:

[0100] In this boundary constraint formula, the variable These are the actual optimal time window widening coefficients that the system ultimately approves and distributes to the underlying continuous wavelet transform calculation engine after rigorous filtering by anti-disturbance logic. Parameters This is the upper limit threshold for wavelet window width saturation, set based on the capacity limit of the high-speed static random access memory within the digital signal processor and the minimum acceptable time-domain resolution. When the calculated value derived from the formula attempts to exceed this physical limit, the system forcibly truncates the value, rigidly maintaining the output within the maximum allowable range. Parameters This represents the hard lower limit of the window width and the dead zone threshold set by the system to avoid a complete collapse of the frequency domain resolution. Through this rigorous upper and lower limit saturation constraint control mechanism, the system ensures that the adaptive mapping function can output absolutely safe and physically consistent algorithm control parameters in any extreme and noisy heavy industrial pipeline network environment.

[0101] After the dynamic mapping function continuously outputs the optimal time window widening coefficients with strict constraints, the digital signal processor cluster performs a true flow velocity adaptive continuous wavelet transform on the source-end sludge fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence based on this mapping function. In this high-concurrency, high-dimensional complex matrix operation step, the algorithm engine dynamically reshapes the mother wavelet basis function at the millisecond level according to the actual measured flow velocity state at the current moment. When the flue gas velocity increases, the mapping function outputs a smaller scale scalar, and the computing engine automatically narrows the time window width of the continuous wavelet transform in real time. The mother wavelet basis function is drastically compressed on the time axis, thereby accurately depicting the exact absolute number of seconds of the short energy pulse peak with extremely high time domain resolution.

[0102] As the flue gas velocity decreases, the mapping function outputs a larger scale scalar. The computing engine automatically widens the time window in real time, and the mother wavelet basis function slowly expands along the time axis, thereby capturing extremely weak energy ripple correlations over long periods with extremely high frequency domain resolution. The mathematical formula for the digital signal processor to perform adaptive wavelet transform with time-varying window dynamic constraints is as follows:

[0103] In this adaptive integral formula, the variable Indicates a specific frequency domain scale Translation step size at a specific time and the measured flow velocity at the moment of translation. The adaptive wavelet complex coefficient matrix obtained under comprehensive physical constraints. Original one-dimensional time series. Complex conjugate form with mother wavelet This involves deep convolution integral operations. The core technological breakthrough lies in the fact that the broadening denominator inside the integral kernel function is no longer a single static scaling factor as in traditional algorithms. Instead, it introduces an optimal time window broadening factor determined in real time by a fluid dynamics sensor. The product term. This means that for every millisecond of industrial sensor data slice, the algorithm engine is using a basis function form tailored to the current flow field for feature extraction. After completing the adaptive frequency domain expansion of the source and end sequences respectively, the digital signal processor cross-multiplies the complex matrices of the two to calculate the cross spectral density, and extracts the core phase angle difference in the high condensation energy region from the complex matrix. The previous basis function form has perfectly matched the change law of pipeline flow velocity. At this time, the extracted phase angle difference completely gets rid of the mathematical distortion caused by the time-frequency resolution mismatch due to the fixed window, and obtains a phase angle difference with absolute physical fidelity. From the extracted phase angle difference to the final carbon emission physical tracing and binding, the system must overcome the last nonlinear fluid dynamics correction threshold. The process energy pulsation needs to go through a long physical span from the upstream furnace to the downstream exhaust fan. During this delay period of several minutes, the average gas flow velocity of the entire complex pipeline system is often in a continuous nonlinear acceleration or deceleration state.

[0104] Directly converting physical time using a purely static phase angle would result in cumulative drift errors on the order of milliseconds or even seconds for the source tracing coordinates. To completely eliminate this systematic error caused by flow field non-stationarity and ensure the accuracy of subsequent source tracing, the digital signal processor (DSP) forcibly introduces a feedforward correction compensation mechanism for the fluid acceleration state when calculating the final dynamic time delay. The mathematical formula for this time delay calculation, which includes the compensation factor, is constructed as follows:

[0105] In this final traceability time conversion formula, the variable Representing the corrected dynamic absolute lag time of the system's final output, this parameter clarifies the exact physical number of seconds it takes for the peak of end-of-pipe energy consumption to trace back to the trough of source-end carbon emissions. Variable The phase angle difference extracted by the adaptive wavelet engine. Variable This represents the dominant physical equivalent frequency in the current specific resonant energy region. The most crucial core correction multiplier is an exponential function based on the natural constant, where the differential term... This represents the rate of change of the average flue gas velocity in the pipeline network within the calculated delay time window, i.e., the macroscopic fluid acceleration. Parameter This refers to the pre-set flow field acceleration compensation weighting coefficients based on the pipeline resistance model. When the pipeline is in the blowing acceleration period, the fluid acceleration is positive, and the overall calculation result of the exponential compensation term is greater than one. The system automatically amplifies the lag time appropriately to accurately compensate for the extra transmission time consumed by the fluid in the first half of the pipeline due to the lower initial velocity. When the pipeline is in the shutdown deceleration period, the fluid acceleration is negative, and the system automatically reduces the lag time appropriately according to the exponential decay ratio.

[0106] After acquiring precise time lag data through adaptive optimization of sensor hardware and feedforward flow field compensation, the heterogeneous edge computing micro-cluster directly and rigidly shifts the indirect electrical carbon emission rate sequence of the downstream acid production workshop on the time axis. Each unit of electricity consumed in acid production at the end can be precisely locked and bound to the specific upstream feed or ventilation operation responsible for the process fluctuation based on the adjusted digital time axis. This adaptive variable window frequency domain tracing technology, deeply embedded in the underlying logic of fluid dynamics and driven by intelligent heterogeneous computing power, completely breaks through the time-frequency resolution bottleneck of traditional fixed mathematical algorithms under extreme smelting conditions. It provides a robust technical foundation with extreme physical robustness and theoretical rigor for the refined, full-cycle carbon emission collaborative management of the high-energy-consuming non-ferrous metallurgical industry. Low-level control engineers with hardware and software co-development capabilities can efficiently reproduce the full engineering performance of this technology and achieve accurate responsibility tracing in any long-process heavy chemical industry scenario by deploying matching on-site flow velocity sensing arrays and digital signal processing acceleration computing nodes.

[0107] In the description of this specification, the references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0108] The above description is merely an example and illustration of the concept of the present invention. Those skilled in the art can make various modifications or additions to the specific embodiments described or use similar methods to replace them, as long as they do not deviate from the concept of the invention or exceed the scope defined in the claims, they should all fall within the protection scope of the present invention.

Claims

1. A carbon emission source tracing method for the coordinated energy consumption of flue gas in copper smelting process, comprising extracting the source-end physical parameter sequence of flue gas at the outlet of the smelting furnace, and extracting the end-end power consumption of the downstream flue gas treatment process and converting it into an end-end indirect carbon emission rate sequence. characterized in that The method further includes the following steps: Based on the source-end physical parameter sequence, the flue gas physical odor is calculated in real time, and a source-end odor value fluctuation characteristic sequence with timestamps is generated. The shortest regularized path between the source-end carbon emission rate fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence is calculated using a dynamic time warping algorithm. The end-indirect carbon emission rate sequence is mapped and aligned on the time axis according to the time warp relationship indicated by the shortest regularization path. The data nodes corresponding to the mapped and aligned end-point indirect carbon emission rate sequence are linked to the data nodes corresponding to the source-point value fluctuation characteristic sequence for source tracing.

2. The method of claim 1, wherein the method is used in a copper smelting process flue gas energy consumption correlated carbon emission tracing method, characterized in that, After calculating the shortest warped path using the dynamic time warping algorithm, a frequency domain phase synchronization calibration step is performed using cross wavelet transform. Specifically, this includes performing continuous wavelet transform on the source-end carbon emission fluctuation characteristic sequence and the terminal indirect carbon emission rate sequence to extract the high condensation energy region of the two sequences in the multi-scale frequency domain.

3. The method of claim 2, wherein the method is used in a copper smelting process flue gas energy consumption correlated carbon emission tracing method, characterized in that, After extracting the high condensation energy region, the phase angle difference between the source-end carbon emission rate fluctuation characteristic sequence and the end-end indirect carbon emission rate sequence within the high condensation energy region is calculated.

4. The method of claim 3, wherein the method is used for copper smelting process flue gas energy consumption correlated carbon emission tracing. The phase angle difference is converted into dynamic hysteresis data, and the data nodes bound to the source are time-axis shifted and matched based on the dynamic hysteresis data.

5. The method of claim 1, wherein the method is used in a copper smelting process flue gas energy consumption correlated carbon emission tracing method, characterized in that, The method further includes a step of constructing a directed graph topology network for the entire system, wherein the directed graph topology network uses carbon emission devices as network nodes and physical pipes connecting the devices as network edges; and uses a spatiotemporal graph convolutional neural network to process the source-end carbon emission fluctuation feature sequence and the end-end indirect carbon emission rate sequence.

6. The carbon emission source tracing method for the coordinated energy consumption of flue gas in copper smelting process according to claim 5, characterized in that, An attention mechanism network layer is configured in the spatiotemporal graph convolutional neural network. The dynamic attention weight allocation matrix output by the attention mechanism network layer, which maps from downstream network nodes to multiple upstream network nodes, is extracted. The weight ratio of the source binding is allocated according to the dynamic attention weight allocation matrix.

7. The carbon emission source tracing method for the coordinated energy consumption of flue gas in copper smelting process according to claim 1, characterized in that, In the process of calculating the shortest warped path using the dynamic time warping algorithm, a thermal inertia energy dissipation damping model of the downstream flue gas treatment equipment is constructed, and the damping coefficient output based on the thermal inertia energy dissipation damping model is extracted. The damping coefficient is then used as a dynamic penalty weight value and input into the cost function of the dynamic time warping algorithm.

8. The carbon emission traceability method for the coordinated energy consumption of flue gas in copper smelting process according to claim 7, characterized in that, The numerical values ​​of adjacent time nodes in the terminal indirect carbon emission rate sequence are extracted to calculate the ramp slope. When the ramp slope is greater than the maximum flue gas energy ramp rate threshold set by the equipment, a set extreme value penalty term is applied to the current calculation node in the cost function to block the regular matching of the current time series path.

9. The carbon emission traceability method for the coordinated energy consumption of flue gas in copper smelting process according to claim 2, characterized in that, During the continuous wavelet transform of the source-end carbon emission fluctuation characteristic sequence and the terminal indirect carbon emission rate sequence, real-time flue gas velocity data measured by sensors is acquired. Based on the flue gas velocity data, a mapping relationship function between the wavelet basis function window width and the reciprocal of the velocity is established. The time window width for performing the continuous wavelet transform is adjusted in real time according to the mapping relationship function.

Images