Carbon emission accurate measurement system based on multi-scale variational bayesian inference

Abstract

The present application relates to carbon emission technical field, and further relates to a carbon emission accurate measurement system based on multi-scale variational Bayesian inference, which comprises: a data acquisition unit, configured to acquire basic observation parameters at the entrance of each carbon emission monitoring unit, forming a raw basic data sequence arranged in chronological order; a time alignment and channel classification unit, configured to align the raw basic data sequence into basic time slices according to the minimum adjacent collection interval; and organize multi-scale observation structures of short windows, medium windows and long windows for each physical channel; a Bayesian inference unit, configured to proportionally align the fuel channel and the output channel with the exhaust channel as the reference channel; output the emission estimation sequence organized according to the monitoring unit and the basic time slice and the cumulative emission of each window when the preset convergence condition is met, otherwise continue iteration. The present application can significantly improve the accuracy, robustness and applicability of the measurement.

Description

Technical Field

[0001] This invention belongs to the field of carbon emission technology, specifically relating to a precise carbon emission measurement system based on multi-scale variational Bayesian inference. Background Technology

[0002] Accurate carbon emission measurement is a crucial element in energy management, environmental regulation, and carbon trading systems. As global carbon reduction policies deepen, carbon emission data not only needs to meet statistical aggregation requirements but also must achieve high accuracy in terms of temporal resolution, spatial resolution, and data traceability to support process optimization, real-time monitoring, and compliance verification. Existing technologies for calculating carbon emissions mainly fall into three categories: indirect estimation methods based on emission factors, direct measurement methods based on continuous sensor monitoring, and hybrid methods combining both.

[0003] Indirect estimation methods based on emission factors typically collect data on fuel consumption, product output, or other quantifiable production activities, and then perform calculations using pre-determined emission factors. These methods offer advantages such as low data acquisition costs and simplicity of implementation. However, because emission factors often originate from statistical averages or laboratory measurements, they cannot fully reflect the differences in actual operating conditions across different production units, leading to insufficient measurement accuracy. Furthermore, these methods exhibit lag in responding to short-term fluctuations and process anomalies, making them unsuitable for real-time management scenarios requiring rapid adjustment and precise control.

[0004] Direct measurement methods based on continuous sensor monitoring typically rely on instruments such as flue gas flow meters, gas analyzers, and fuel metering devices to collect emission-related physical quantity data in real time. While these methods theoretically offer high accuracy and real-time performance, practical applications face challenges such as sensor drift, inconsistent sampling frequencies, communication delays, data loss, and outlier interference. Especially in multi-sensor collaborative scenarios, the different sampling time bases and channel response characteristics of various devices often lead to misalignment of cross-channel data and inconsistent data quality, thus reducing the reliability of the calculation results.

[0005] Hybrid methods attempt to combine the advantages of indirect estimation and direct measurement by simultaneously incorporating fuel consumption data, exhaust flow data, and production output data. This allows for the supplementation of missing information and correction of biases through multi-source data. While these methods theoretically improve estimation stability under conditions of missing data, they generally lack a systematic multi-scale consistency processing mechanism in practical applications. Most current implementations only perform data fusion at a single time scale, failing to achieve a balance between short-term fluctuation suppression and long-term trend consistency, easily leading to a mismatch between local corrections and global conservation. Furthermore, there is still a lack of unified and repeatable technical pathways for handling outliers, rationally allocating discrepancies under missing data conditions, and ensuring information backhaul and numerical convergence across different time scales in multi-source data fusion. Summary of the Invention

[0006] The main objective of this invention is to provide a precise carbon emission measurement system based on multi-scale variational Bayesian inference. Under the premise of ensuring upper and lower bound constraints for each time slice, it achieves the unity of local conservation and cross-scale consistency. Through the limited completion of missing time slices and the iterative convergence mechanism, it can still output stable, traceable carbon emission estimation results that conform to physical constraints under the conditions of asynchronous acquisition, missing data and abnormal interference, thereby significantly improving the accuracy, robustness and applicability of the measurement.

[0007] To solve the above problems, the technical solution of the present invention is implemented as follows:

[0008] A carbon emission precision measurement system based on multi-scale variational Bayesian inference, comprising: a data acquisition unit, used to acquire basic observation parameters at the inlet of each carbon emission monitoring unit, including real-time consumption readings at a single fuel consumption metering point, exhaust flow readings at a single emission outlet, and real-time output readings at a single output metering point; adding timestamps to the basic observation parameters according to the acquisition time and labeling them with the corresponding monitoring unit number, forming a raw basic data sequence arranged in chronological order; a time alignment and channel classification unit, used to align the raw basic data sequence to basic time slices according to the minimum adjacent acquisition interval; and dividing each monitoring unit into fuel channels, exhaust channels, and output channels according to physical quantities. For each physical channel, a multi-scale observation structure of short, medium, and long windows is organized. A Bayesian inference unit is used to proportionally align the fuel and production channels with the exhaust channel as a reference channel. Sliding checks are performed on the three types of channels on the base time slice, and abnormal readings are replaced. Temporary emission estimates are generated based on available readings according to the time slice. Within the short window, the temporary emission estimates are balanced to ensure they are consistent with the total readings of the three types of channels within that window. The results from the short window are aggregated into the medium and long windows and adjusted by backpropagation of the difference to achieve cross-scale consistency. When the preset convergence conditions are met, the emission estimation sequence organized by the monitoring unit and the base time slice, as well as the cumulative emissions for each window, are output; otherwise, the iteration continues.

[0009] Furthermore, the lengths of the short window, medium window, and long window are 20, 100, and 500 base time slices, respectively.

[0010] Furthermore, the Bayesian inference unit includes: a preprocessing unit, a short-window conservation balancing unit, a medium-window and long-window consistency feedback unit, a missing time-slot detection unit, and a result output unit. The preprocessing unit performs a sliding check on the three types of channels and replaces abnormal readings on the base time-slot. The short-window conservation balancing unit generates initial emission estimates based on available readings within the time-slot; it then performs conservation balancing on the initial emission estimates within the short window to ensure consistency with the total readings of the three types of channels within that window. The medium-window and long-window consistency feedback unit aggregates the short-window results into the medium-window and long-window results and adjusts them according to the difference to achieve cross-scale consistency. The result output unit outputs the emission estimate sequence organized by the monitoring unit and the base time-slot, along with the cumulative values ​​for each window, when the preset convergence conditions are met; otherwise, it continues iterating.

[0011] Furthermore, the execution process of the preprocessing unit specifically includes: within each monitoring unit, time-slice scanning is performed on each of the three types of channels; when the absolute value of the difference between the reading of a certain channel in a certain time slice and the median of the readings of that channel in the next nine time slices (including the previous four time slices, the current time slice, and the next four time slices) exceeds half of the difference between the maximum and minimum values ​​of the readings in the next nine time slices, the reading of that channel in that time slice is replaced with the median of the readings in the next nine time slices; within each monitoring unit, for each basic time slice, the median of the available readings of the three types of channels in that time slice is taken as the temporary emission estimate for that time slice; if only two channels have readings, the larger of the two is taken; if only one channel has a reading, that reading is taken; if all three channels are missing, the temporary emission estimate is not generated for that time slice and is determined uniformly when the window is returned.

[0012] Furthermore, the execution process of the short-window conservation balancing unit specifically includes: within each monitoring unit, for each short window, calculating the total emission of the fuel channel, exhaust channel, and production channel within that short window, which is the sum of the readings of that channel within that short window, and taking the arithmetic mean of the three as the target total for that short window; simultaneously calculating the sum of all temporary emission estimates within that short window as the current total; when the target total is not equal to the current total, adjusting the temporary emission estimates of each time slot within the short window according to the following conservation rules until they are equal: if an increase is needed, prioritizing those time slots within this short window whose temporary emission estimates are lower than the maximum readings of the three types of channels in the same time slot, increasing them one by one in order from the time slot with the largest difference to the smallest, increasing each time... The magnitude of the reduction shall not exceed the total or remaining difference between the maximum value of the three channel readings in the current time slot and the current provisional emission estimate, prioritizing the equality of total emissions. If a reduction is required, priority shall be given to selecting time slots within the short window whose provisional emission estimates are higher than the minimum value of the three channel readings in the same time slot, and reducing them one by one in order from the time slot with the largest difference to the smallest. Each reduction shall not exceed the total or remaining difference between the current provisional emission estimate and the minimum value of the three channel readings in the current time slot, prioritizing the equality of total emissions. If a time slot has missing measurement markers that make the upper or lower limit unavailable, then the time slot is only allowed to be adjusted in the known direction. When only the upper limit is available, only an increase is allowed; when only the lower limit is available, only a decrease is allowed; when neither the upper nor the lower limit is available, the time slot is skipped.

[0013] Furthermore, the execution process of the consistency feedback unit for the medium and long windows specifically includes: within each monitoring unit, the estimated temporary emissions after balancing within the short window are first aggregated into the medium and long windows by summing over time to obtain the current total for each window; the average of the sum of the three types of channel readings within each medium and long window is calculated as the target total for the corresponding window; when there is a difference between the target total and the current total for a certain medium or long window, the difference is allocated according to the adjustable space within the included short windows: the adjustable space is defined as the sum of all time slices within each short window that can still be increased or decreased without exceeding the upper and lower limits of the three types of channel readings obtainable in the corresponding time slice; the difference is allocated to the short windows according to the proportion of the adjustable space of each short window in the total adjustable space of the window, and then processed within the short windows according to the conservation rules of the short window conservation balancing unit.

[0014] Furthermore, the execution process of the missing time slot detection unit specifically includes: when there is still an unallocated portion of the difference allocated by a short window or its upper-level window through the consistency feedback unit between the middle window and the long window, and there are time slots within the short window where the preprocessing unit has not generated temporary emission estimates, the remaining difference is divided equally among these time slots; for each equally divided time slot, if there is at least one channel reading in the same time slot, the average value of the time slot shall not exceed the upper and lower limits that the time slot can take; if all three channels are missing, the time slot is allowed to take a value between the minimum and maximum values ​​of the temporary emission estimates already determined within the short window, in order to complete the difference allocation.

[0015] Furthermore, the complete process executed by the preprocessing unit, the short window conservation balancing unit, the medium window and long window consistency feedback unit, and the missing time slot detection unit is regarded as an iteration. After completing one iteration, the sum of the absolute values ​​of the differences between the temporary emission estimates of all monitoring units and the corresponding values ​​of the previous iteration is calculated as the proportion of the total temporary emission estimates of the previous iteration. When this proportion does not exceed 1%, or when the number of update iterations reaches 200, the process stops. Otherwise, the process returns to the preprocessing unit to enter the next iteration. After stopping, the emission estimation sequence organized by monitoring unit and basic time slot, as well as the cumulative emissions of the short window, medium window, and long window are output.

[0016] The carbon emission accurate measurement system based on multi-scale variational Bayesian inference of this invention has the following beneficial effects: it achieves accurate measurement of carbon emissions and maintains the stability and traceability of emission estimates even in the event of asynchronous, missing, and abnormal interference in data acquisition. The method first performs time alignment and classification of fuel, exhaust, and production channels at the basic time-slice level, and uses a sliding window midpoint replacement strategy to suppress local extreme values, ensuring the robustness of the input data. Within the short window, a conservation balancing mechanism adjusts the temporary emission estimate to be consistent with the total readings of multiple channels, strictly adhering to the feasible upper and lower bound constraints for each time slice, thereby achieving a balance between local conservation and minimum disturbance. At the medium and long window levels, a consistent backpropagation mechanism is adopted to propagate the total constraints from larger time scales down to the short window, and the difference is allocated according to an adjustable spatial proportion, effectively eliminating statistical biases between different time scales. For time slices where emission estimates were not generated during the preprocessing stage, this invention utilizes equal distribution of the difference and boundary constraints for limited completion, making the missing locations a buffer for cross-scale difference settlement, thus improving the system's flexibility and convergence under incomplete information. The entire process is executed iteratively, using the global change ratio and the maximum number of rounds as convergence criteria to ensure stable final results within a finite number of steps. This invention achieves high-precision estimation while also considering multi-scale conservation, consistency, robustness, and reproducibility, making it suitable for various application scenarios such as continuous monitoring, process optimization, and compliance verification. It has significant implications for improving the scientific rigor and practicality of carbon emission measurement. Attached Figure Description

[0017] Figure 1 A schematic diagram of the system structure of the carbon emission accurate measurement system based on multi-scale variational Bayesian inference provided in an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram illustrating the principle of multi-scale window consistency backhaul adjustment provided in an embodiment of the present invention.

[0019] Figure 3 The iterative convergence process curve provided in the embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of 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 skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0021] refer to Figure 1A carbon emission accurate measurement system based on multi-scale variational Bayesian inference, the system comprising: a data acquisition unit, a time alignment and channel classification unit, and a Bayesian inference unit;

[0022] The data acquisition unit is located at the entrance of each carbon emission monitoring unit. Following the overall requirements of a carbon emission precision measurement system based on multi-scale variational Bayesian inference, it only collects basic observation parameters and ensures boundary consistency and traceability. Specifically, it binds a single fuel consumption metering point to the same monitoring unit to obtain real-time consumption readings, a single exhaust outlet to obtain exhaust flow readings, and a single output metering point to obtain real-time output readings. These three types of readings are read synchronously by the acquisition execution unit at the same acquisition time or sequentially within a very short interval to reduce timing deviations. After each reading, the data acquisition unit immediately adds a timestamp to each reading and attaches the monitoring unit number and acquisition source identifier. It standardizes the measurement units and decimal precision of the readings, checks for out-of-bounds, backtracking, or obviously illogical abrupt changes, and replaces any anomalies with a missing measurement marker while retaining the original bitstream for traceability. To support the subsequent alignment of the original basic data sequence by the time alignment and channel classification unit according to the minimum adjacent acquisition interval as the base time slice, the data acquisition unit records the timestamp of the most recent successful acquisition for each monitoring unit and maintains rolling statistics of adjacent acquisition intervals to indicate sampling stability and packet loss risk.

[0023] To ensure that the three channels can be accurately classified into fuel, exhaust, and output channels, the data acquisition unit explicitly labels the physical quantity category and metering point role for each record when writing the original basic data sequence. Readings from different monitoring units are not allowed to be mixed into the same sequence. At the network transmission level, the data acquisition unit sets up a local buffer queue and sequential numbering strategy to ensure that the original basic data sequence is written in chronological order after communication interruption or delayed recovery. The acquired content is not smoothed, interpolated, or summarized, maintaining its originality so that the Bayesian inference unit can perform sliding checks and replace abnormal readings, generate temporary emission estimates based on available readings according to time slices, perform conservation balancing within short windows, and provide consistent backhaul to medium and long windows. At the operation and maintenance level, the data acquisition unit supports periodic calibration and zero-point verification for single fuel consumption metering points, single exhaust ports, and single output metering points. Readings generated during calibration are marked as missing and continue to be written into the original basic data sequence to ensure that they can be identified and participate in the averaging and constraint of differences in subsequent missing time slice detection.

[0024] The time alignment and channel classification unit takes the original basic data sequence as input and outputs the three types of channel sequences aligned with the basic time slice and the corresponding multi-scale observation structures of short, medium, and long windows. First, the timestamps are consistent within each monitoring unit. A copy of the original basic data sequence containing only the records of that monitoring unit is established using the timestamps attached to the data acquisition unit and the monitoring unit number. Any readings across monitoring units are deleted, while the arrival order and original values ​​remain unchanged. Then, the acquisition time difference of adjacent records is scanned on this copy, and the smallest positive adjacent acquisition interval is selected as the alignment benchmark for that monitoring unit. This interval is defined as the length of the basic time slice, and a continuous time slice sequence is generated using the first valid timestamp of that monitoring unit as the starting boundary. The time alignment and channel classification unit calculates the basic time slice number to which each record in the original basic data sequence belongs and uses stable entry rules to handle the case of multiple records appearing in the same basic time slice. The reading with the latest acquisition time is retained as the final record of the basic time slice, while the overwritten readings are archived with redundant markers for traceability.

[0025] If two adjacent records span multiple base time slices, missing data markers are generated sequentially within the spanning intervals to ensure that the base time slices are continuous on the time axis and arranged in chronological order. After completing the base time slice mapping, the time alignment and channel classification unit assigns the readings within each base time slice to the corresponding positions in the fuel channel, exhaust channel, and production channel based on the physical quantity category and measurement point role in the original records. When a channel in a base time slice has no reading, it is explicitly marked as missing data without any interpolation or smoothing, to ensure the originality and interpretability of the input when performing sliding checks and replacing abnormal readings in the preprocessing unit, and when generating temporary emission estimates based on available readings in the base time slices. To facilitate the subsequent conservation balancing of temporary emission estimates within short windows and the consistent backhaul to medium and long windows, the time alignment and channel classification unit... Within each monitoring unit, the class unit generates multi-scale observation structures of short, medium, and long windows for the fuel channel, exhaust channel, and production channel, respectively. The short window consists of 20 consecutive basic time slices, the medium window consists of 100 consecutive basic time slices, and the long window consists of 500 consecutive basic time slices. All three types of windows cover the entire basic time slice sequence using a sliding method with a step size of 1. During the generation of the multi-scale observation structure, each window saves the start and end times of the basic time slices it covers, the number of basic time slices it includes, the available and missing status of each basic time slice in the three types of channels, and the index information required by the subsequent Bayesian inference unit to calculate the total number of channels and the target total number, but does not perform any summation, averaging, or other statistical operations, thus leaving the calculation and consistency responsibilities to the Bayesian inference unit and the consistency feedback unit of the medium and long windows.

[0026] The Yesian inference unit treats the readings of the fuel channel, exhaust channel, and output channel on the base time slice as three biased observations of the same potential emission process. Using variational Bayesian ideas, it jointly estimates the potential emission sequence and the channel ratio in an approximate posterior manner without changing the original base data sequence, and simultaneously applies conservation and consistency constraints on the multi-scale observation structure of short window, medium window, and long window.

[0027] The proportional alignment with the exhaust channel as a reference channel is not a simple ratio replacement, but rather stems from identifiability constraints:

[0028] To eliminate uncertainty at the overall scale, the reference channel provides this anchor point. The proportional relationship between the fuel and production channels relative to the reference channel is treated as a potential quantity to be estimated, and its feasible interval is jointly defined by the available upper and lower limits of the three types of channels on the base time slice. The sliding check and median replacement of the nine adjacent time slices in the preprocessing unit correspond to the robust likelihood approximation using the heavy-tailed assumption in the observation noise model. The robust statistic of the median reduces the impact of extreme readings on the posterior, and the missing data marker explicitly represents the information gap, ensuring that subsequent inferences are based on real and incomplete information. The generation of provisional emission estimates based on available readings according to the time slice projects the instantaneous constraints of the three types of channels on the same base time slice onto the point estimate of the potential emission sequence: when all three types of channels are available, the median is taken, reflecting symmetry, unbiasedness, and resistance to anomalies; when only two channels are available, the larger one is taken, reflecting a cautious estimate of the conservation lower bound to avoid systematic underestimation; when only one channel is available, the reading is taken to maintain traceability; when all three channels are missing, no generation is performed, and information feedback from the upper window is awaited.

[0029] The short-window conservation balancing embodies the equivalent constraint projection of the local conservation principle in variational inference: for each short window, the total amount of the three types of channels within that window is calculated separately, and the arithmetic mean of the three is used as the target total amount, which is equivalent to a symmetric fusion of the three types of evidence for posterior point estimation without the introduction of explicit weights; the target total amount is compared with the current total amount of all temporary emission estimates within the short window, and the temporary emission estimates are monotonically adjusted according to the gap priority principle without exceeding the upper and lower limits of the three types of channel readings in the same time slice, which is equivalent to a coordinate-based local update of the lower bound of the evidence within the feasible region; if there are missing measurement markers in the time slice that cause unilateral constraints, adjustments are only allowed in the known direction to ensure feasibility. The consistent backpropagation of the medium and long windows corresponds to the consistent propagation of hierarchical priors across multiple scales: first, the short window results are aggregated over time to obtain the current totals of the medium and long windows; then, the average of the sums of the three types of channel readings is used as the target total for the corresponding window. This choice reflects the symmetrical fusion of three-source evidence on a larger time scale. When there is a difference between the target total and the current total, the adjustable space that each short window can increase or decrease without exceeding the upper and lower limits of the time slice is used as the allocation weight. The difference is assigned back to the short windows, and the short window conservation balancing is called again to complete the local feasible projection. This is equivalent to performing consistency calibration between parent and child nodes in a deterministic manner in the hierarchical graph model. The averaging mechanism of the missing time slice detection unit is not arbitrary filling, but rather a constrained posterior quality allocation of the unallocated difference at the missing position of the short window: when at least one channel is available in the time slice, the average score is constrained by the upper and lower limits of the time slice; when all three channels are missing, the average score is limited to the minimum and maximum values ​​already determined in this short window, maintaining consistency with known information.

[0030] This set of projection-backpropagation-reprojection processes, from the base time slice to the short window, then to the medium and long windows, corresponds to coordinate ascent and hierarchical consistency in variational Bayesian methods, forming a convergence mechanism under conditions of multi-scale, missing data, and upper and lower bound constraints. To ensure global stability, the Bayesian inference unit defines the continuous execution of the preprocessing unit, the short window conservation balancing unit, the medium and long window consistency backpropagation unit, and the missing time slice detection unit as one iteration. The normalized total difference between two adjacent iterations of the temporary emission estimates of all base time slices for all monitoring units is used as the stopping criterion. When this proportion does not exceed 1% or the number of update iterations reaches 200, it is considered to have reached a fixed point that approximates the posterior. Based on this, the output unit provides the emission estimation sequence organized by monitoring unit and base time slice, as well as the cumulative emissions for the short, medium, and long windows. At the same time, it archives the feasible intervals obtained by proportionally aligning the fuel channel and the production channel with the exhaust channel as the reference channel, and the adjustable space of each window, which is convenient for auditing and recalculation.

[0031] Furthermore, the Bayesian inference unit includes: a preprocessing unit, a short-window conservation balancing unit, a medium-window and long-window consistency feedback unit, a missing time-slot detection unit, and a result output unit. The preprocessing unit performs a sliding check on the three types of channels and replaces abnormal readings on the base time-slot. The short-window conservation balancing unit generates initial emission estimates based on available readings within the time-slot; it then performs conservation balancing on the initial emission estimates within the short window to ensure consistency with the total readings of the three types of channels within that window. The medium-window and long-window consistency feedback unit aggregates the short-window results into the medium-window and long-window results and adjusts them according to the difference to achieve cross-scale consistency. The result output unit outputs the emission estimate sequence organized by the monitoring unit and the base time-slot, along with the cumulative values ​​for each window, when the preset convergence conditions are met; otherwise, it continues iterating.

[0032] Furthermore, the preprocessing unit performs minimal intervention on the anomaly suppression and information extraction of the readings of the three types of channels in the basic time slice. This provides stable, interpretable, and physically consistent input to the subsequent generation of provisional emission estimates, short-window conservation balancing, and consistent backhaul of medium and long windows in the carbon emission accurate measurement system based on multi-scale variational Bayesian inference. Its core idea is first reflected in the time slice scanning and the setting of a symmetrical neighborhood of nine adjacent time slices: by scanning the three types of channels one by one in each monitoring unit and evaluating local behavior through a symmetrical window consisting of the previous four time slices, the current time slice, and the next four time slices, the drift bias caused by using global statistics across windows is avoided. The symmetrical window, combined with the robust statistic of the median, makes the discrimination insensitive to single-point spikes and short-term pulses, thus providing stable local representative values ​​even when equipment jitter, metering transients, or communication jitter exist.

[0033] The threshold is half the difference between the maximum and minimum values ​​of the nine adjacent time-slice readings. It does not rely on a fixed constant but adaptively adjusts with the local dynamic range. In principle, this is equivalent to using half the width of the interval scale as the anomaly criterion. This covers common narrow-range fluctuations and triggers replacement when significant deviations occur, ensuring that replacement is limited to readings that are significantly inconsistent with the overall behavior of the neighborhood. Out-of-bounds readings are replaced with the median of the nine adjacent time-slice readings, reflecting the robustness principle of "replacing extreme values ​​with local representation": the median does not change the center position of the neighborhood, does not introduce new phase shifts, and is not pulled by adjacent extreme values ​​like the average. Therefore, it restores feasible upper and lower limits that can be used for subsequent conservation balancing without changing the original trend.

[0034] Secondly, the preprocessing unit generates provisional emission estimates based on available readings in each base time slice. The principle is to achieve immediate fusion of three-source evidence with minimal prior assumptions: when all three channels have readings, the median is taken, which is equivalent to symmetrical and anomaly-resistant central aggregation of the three-source evidence, ensuring that the provisional emission estimate is not dominated by noise from any single path; when only two channels have readings, the larger of the two is taken, out of conservative considerations of emission conservation and safety, to avoid systematic underestimation when evidence is incomplete and forced to make significant upward corrections in subsequent short-window conservation balancing; when only one channel has a reading, that reading is used directly, following the principle of "traceability priority", to avoid introducing inferences without evidence support when information is insufficient; when all three channels are missing, provisional emission estimates are not generated temporarily and are determined uniformly when the window is back, reflecting the dependence on multi-scale structure: the missing time slices will be filled by medium and long windows under the constraints of differential backhaul and adjustable space, thereby postponing the resolution of uncertainties to a scale with more complete information and avoiding unconstrained speculation on the base time slice. Through the coupling of the two mechanisms mentioned above, the preprocessing unit, on the one hand, uses the median replacement strategy of nine adjacent time slices to suppress outliers, maintain trend patterns and feasible constraints at the local scale; on the other hand, it uses the hierarchical rule of "median / larger / original value / not generated for now" to provide a consistent temporary emission estimate generation logic under different information completeness conditions. This provides a stable initial value for the conservation balancing of the current total amount with the target total amount in the subsequent short window, and reserves the necessary adjustable space for the consistent backhaul of the medium window and long window.

[0035] Furthermore, the short-window conservation balancing algorithm, without altering the original underlying data sequence, focuses on local conservation and feasible region constraints to project temporary emission estimates onto a target state consistent with the evidence over a short timescale. Its working principle involves calculating the total emissions from the fuel, exhaust, and production channels within each short window of each monitoring unit, and using the arithmetic mean of these three totals as the target total for that short window. This approach reflects the principle of symmetrical fusion of evidence from the three channels, avoiding any single path biasing the total emissions within the short window, while also providing an anchor point for subsequent cross-scale consistency. The sum of all temporary emission estimates within the short window is defined as the current total emissions. The conservation problem is then reduced to a constrained balancing problem: "adjusting the current total emissions to the target total emissions while ensuring that each time slice does not exceed the upper and lower limits."

[0036] To minimize disturbance to the original state, the short-window conservatism balancing unit employs a gap-priority and direction-restricted monotonic adjustment strategy: when an increase is needed, it is only implemented in time slots where the temporary emission estimate is lower than the maximum value of the three channel readings within the same time slot, and is advanced sequentially according to the gap with the upper limit of that time slot from largest to smallest. This is equivalent to prioritizing the consumption of "increasable margin," minimizing the number of time slots and adjustments involved before the total gap is filled. When a decrease is needed, it is only implemented in time slots where the temporary emission estimate is higher than the minimum value of the three channel readings within the same time slot, and is advanced sequentially according to the gap with the lower limit of that time slot from largest to smallest. Similarly, "reducable margin" is released first. The magnitude of each increase or decrease does not exceed the remaining gap between the corresponding time slot and its upper or lower limit, and does not exceed the total difference that has not yet been eliminated. This magnitude control avoids local overshoot leading to subsequent reverse corrections, thereby ensuring that the target total is reached within a finite number of steps and reducing unnecessary oscillations. If a time slice has missing measurement markers that make the upper or lower limit unavailable, then the time slice can only be adjusted in the known direction; if only the upper limit is available, it can only be increased, and if only the lower limit is available, it can only be decreased. When neither the upper nor the lower limit is available, the time slice is skipped. This ensures that the balancing process will not be pushed outward into the uncertain region, and maintains consistency and traceability with known evidence.

[0037] The above rules are, in principle, equivalent to finding the feasible point with the smallest disturbance from the current solution in the intersection of the "total conservation hyperplane" and the "time-slice box constraint set". The difference-first sorting is approximating a greedy solution for the smallest overall change, so that the balancing satisfies conservation and closely matches the temporary emission estimate given by the preprocessing unit in terms of time distribution. Since the time span of the short window is limited, the total channel volume can better reflect the local income and expenditure of the equipment periodicity and process steady state at this scale. Therefore, balancing the current total volume with the target total volume can absorb the consistent information of the three types of channels and suppress the cumulative impact of single-point deviations on the total volume without relying on external priors. Furthermore, the short window conservation balancing unit, by strictly following the upper and lower limit constraints, naturally reflects the feasible interval implied by proportionally aligning the fuel channel and the production channel with the exhaust channel as the reference channel in the adjustable space of each time slice, thus providing a calibrated and still flexible basis for the subsequent consistency backhaul of the medium and long windows. Ultimately, the balanced temporary emission estimate output by the short-window conservation balancing unit satisfies the target total for this short window while maintaining point-by-point respect for the upper and lower limits at the time slice level. This achieves the triple principle of "local conservation, minimum disturbance, and directional constraint," providing a stable starting point for the consistency of the carbon emission accurate measurement system based on multi-scale variational Bayesian inference on a larger time scale and for the constrained completion of missing time slices.

[0038] Furthermore, the medium-window and long-window consistent backhaul units implement cross-scale consistent propagation that integrates conservation and symmetry at hierarchical time scales. This ensures that the local results based on the short window remain consistent with the readings of the three channels over a larger time span, while minimizing unnecessary distortion of the temporal form of the balanced temporary emission estimates within the short window. The work begins by aggregating the balanced temporary emission estimates within the short window to the medium-window and long window through time summation, obtaining the current total for each window. This step elevates the feasible solution formed by the short-window conservation balancing unit within a local time span to a larger-scale statistical level. Subsequently, the average of the sums of the three channel readings within each medium-window and long window is calculated as the target total for the corresponding window. This choice reflects the symmetrical treatment of the fuel channel, exhaust channel, and production channel, avoiding any single channel imposing a systematic bias on a larger scale. It is also compatible with the setting of proportionally aligning the fuel and production channels with the exhaust channel as a reference channel, because the target total comes from the common information of the three channels rather than a single channel's unilateral scale. When there is a difference between the target total and the current total in a medium or long window, the key is how to distribute the difference unbiasedly and in a limited manner among the multiple short windows.

[0039] To address this, an adjustable margin is introduced as the basis for weight allocation. The adjustable margin is defined as the sum of all time slices within each short window that can be increased or decreased without exceeding the upper and lower limits of the three types of channel readings available for that time slice. This definition encodes the feasible interval formed by proportional alignment with the exhaust channel as the reference channel, the upper and lower limits after robust replacement of abnormal readings by the preprocessing unit, and the local elasticity retained by the short window conservation balancing unit in previous steps into a measurable "feasible margin." Allocating the difference to the short windows according to their proportion of the total adjustable margin within that window is essentially a box-constrained weight allocation strategy: short windows with larger adjustable margins bear more of the difference, and short windows with smaller adjustable margins bear less of the difference. This ensures that cross-scale consistency proceeds primarily in the safest and least costly direction without exceeding any time slice's upper and lower limits. After the difference is fed back to the short window, the conservation rules of the short window conservation balancing unit are called again in each short window with the allocated difference for processing. This closed loop ensures that the cross-scale adjustment still follows the local criteria of "total amount conservation, minimum disturbance, and direction restriction". Instead of directly rewriting the time slice point by point, it achieves smooth implementation through the established temporary emission estimate adjustment sequence.

[0040] The above mechanism ensures that three objectives are met simultaneously: First, conservation, meaning that the total target generated at the medium and long window levels can be consistent with the current total at the short window level through differential backpropagation; second, feasibility, meaning that all adjustments are subject to the hard constraint of "not exceeding the upper and lower limits of the three types of channel readings in the corresponding time slice," ensuring that no time slice is pushed into the infeasible region; and third, stability, meaning that by using the adjustable space as a weight and adopting a difference-first conservation rule within the short window, unnecessary repeated shifts and oscillations between multiple short windows are avoided, which is conducive to quickly approaching the convergence criterion in the overall iterative process. Since the medium and long windows converge three types of channel readings over a longer time range, their total target is less sensitive to occasional noise. Therefore, the uniform backpropagation unit distributes this more robust statistical information to the short window in the form of differentials, so that the short window results not only meet the local total requirements, but also remain consistent with the average evidence of the three types of channel readings on a larger time scale, thereby improving the cross-scale consistency and interpretability of the provisional emission estimates without changing the original basic data sequence. Ultimately, through a loop structure of "aggregating the current total amount - calculating the target total amount - allocating the difference according to the adjustable space - calling the conservation rule for processing within the short window", the consistent backhaul unit of the medium window and the long window achieves top-down constrained consistency propagation in the multi-scale structure, providing stable boundary conditions and executable paths for the subsequent processing of missing time slices and the convergence of global iteration.

[0041] Furthermore, the missing time slice detection unit, without altering the original basic data sequence, prioritizes the digestion of discrepancies generated by cross-scale consistency that have not yet been implemented in the basic time slice where the information is most uncertain but the constraints are still clear. This achieves conservation closure within the short window with minimal perturbation and creates monotonic convergence conditions for the overall iteration. Specifically, if the discrepancy allocated by a short window or its upper-level window through the consistency feedback unit of the middle and long windows still has an unallocated portion after calling the short window conservation balancing unit, it indicates that the increases or decreases made within the short window based on known upper and lower limits have reached the boundary of the adjustable space or the allocation path is limited. In this case, a basic time slice where the preprocessing unit has not generated temporary emission estimates is introduced as a backup adjustable space.

[0042] Distributing the remaining difference equally among these base time slices reflects the principles of unbiasedness and form preservation in scenarios with insufficient evidence: the unbiased principle means not artificially constructing time trends or phased structures within the missing data set, while the form preservation principle means avoiding secondary disturbances to the established sequence of provisional emission estimates. Equal distribution is not unconstrained filling, but rather controlled by the upper and lower limits of the same base time slice; when at least one channel reading exists in the base time slice, the average value must not exceed the upper and lower limits that the base time slice can take. This constraint ensures that the feasible range formed by proportionally aligning the fuel and production channels with the exhaust channel as a reference channel is strictly adhered to during the missing data filling process, and avoids allocating the difference to a value range that contradicts the evidence. When all three channels are missing, the base time slice is allowed to take values ​​between the minimum and maximum values ​​of the established provisional emission estimates within this short window. Essentially, this provides a feasible region based on the empirical boundaries of similar environments within the short window, thus ensuring that the missing data filling neither jumps out of the known range of the short window nor is forcibly pushed to extremes in the absence of evidence.

[0043] The above rules link the "difference - adjustable space - upper and lower limits - equal distribution" into a unidirectional constraint chain, ensuring that the remaining difference is always preferentially absorbed at positions that do not affect the established points, do not exceed the feasible region, and have the least impact on the morphology. The linkage with the medium and long window consistency backhaul units is reflected in two aspects: First, the missing time slice detection unit is only activated when the difference after consistency backhaul has not yet been fully settled, thus undertaking the last stage of constrained settlement task and avoiding competition with the preceding short window conservation balancing unit in the same adjustable space; Second, by setting upper and lower limits for the equal distribution value, the missing time slice detection unit refines the global constraints issued across scales into point-level feasible regions, ensuring that the top-down difference transmission and the bottom-up conservation closure are completed in the same feasible set, without introducing new out-of-bounds risks. To ensure the overall process convergence and measurability, the missing time slice detection unit, preprocessing unit, short window conservation balancing unit, and medium and long window consistency feedback unit together constitute a complete link for one iteration. After the iteration, the sum of the absolute values ​​of the differences between the temporary emission estimates of all basic time slices of all monitoring units and the corresponding values ​​of the previous round is used as the proportion of the total sum of all temporary emission estimates of the previous round to the total sum of all temporary emission estimates of the previous round, as the global change measure.

[0044] This metric offers two fundamental advantages: first, additivity, enabling it to cumulatively reflect the overall scale of disturbances across monitoring units and time slices; and second, non-directionality, using absolute values ​​to avoid the offsetting of positive and negative values ​​that mask local oscillations. Setting the threshold to no more than 1% as a stopping condition means that when the total change in the new round of provisional emission estimates is sufficiently small relative to the total of the previous round, the system can be considered to be near a near-fixed point, and further iteration has a very low marginal impact on the final result. At the same time, setting an upper limit of 200 update rounds provides iterative protection for extreme data patterns or severely constrained scenarios, preventing prolonged small-amplitude round trips when the adjustable space is locked by multiple constraints. The stopping criterion, together with the aforementioned unidirectional constraint chain, forms a "contraction-truncation" convergence structure: in each round, the short-window conservation balancing unit reduces the difference between the target total and the current total in a gap-first manner within the feasible region; the medium-window and long-window consistency feedback unit continues to reduce the cross-scale difference on a larger time scale; and the missing time slice detection unit provides a restricted settlement channel when the remaining difference is extremely small and often blocked by the boundary. The net effect of the three superimposed is that the global change measurement is non-incremental, until the proportion threshold is met or the round limit is reached.

[0045] After stopping, the output shows the emission estimation sequence organized by monitoring unit and base time slice, as well as the cumulative emissions for short, medium, and long windows. This reflects the hierarchical constraints from point-level feasible region to window-level conservation and then to multi-scale consistency, ultimately summarizing into an auditable time series and cumulative indicators. Through the above design principle, the missing time slice detection unit not only avoids arbitrary interpolation at missing locations but also transforms the missing locations into a buffer pool to receive residual differences. Under strict adherence to upper and lower limits, it completes the difference settling, thus working together with the short-window conservation balancing unit and the medium and long-window consistency backhaul unit to ensure that the carbon emission accurate measurement system based on multi-scale variational Bayesian inference achieves robust, interpretable, and convergent emission estimates under the conditions of multiple channels, missing data, and cross-scale constraints.

[0046] In one embodiment, a single monitoring unit is selected, a basic time-slice sequence with a covered window length is constructed, and the entire implementation process from data acquisition and time alignment to preprocessing, short-window conservation balancing, medium-window consistent backhaul, iteration, and output is fully demonstrated. Let the basic time-slice index be... .set up Indicates a single fuel consumption metering point in time slices Real-time consumption readings Indicates a single emission outlet in time slices The exhaust flow rate reading, Indicates the single output measurement point in time slice The instantaneous output readings. After the data acquisition unit writes the original basic data sequence, the time alignment and channel classification unit establishes a basic time slice within the monitoring unit with the minimum adjacent acquisition interval, resulting in the aligned three-category channel sequences. (This example) To test the robustness of the pretreatment and its ability to maintain conservation balance, the following observation characteristics were set: The fuel passage experienced a high-amplitude anomaly. ;exist Production channel lack of testing, in Fuel passage missing measurement, in The exhaust channel was missing a measurement, but all three channels had readings in the other time slices and fluctuated slowly around a stable baseline.

[0047] The preprocessing unit performs time-slice scanning on each channel one by one, using adjacent... A time slice (before) +Current +after The sliding check replaces abnormal readings. The preprocessing output is defined as... When a certain channel is in time slice When the reading is complete, record the channel in the window. The median within is The maximum and minimum values ​​are respectively ,like Then use Replace that reading. Taking the fuel channel as an example, the window Inside The median is (Rounded to) (number of decimal places, the same below), the maximum and minimum values ​​are respectively , ,have Therefore, it is replaced with When a channel is in time slice When a measurement is missing, it is neither included in the discrimination nor a replacement value is generated; the missing measurement marker is retained. Subsequently, at each time slice, a preliminary temporary emission estimate is generated based on available readings, denoted as... .

[0048] When all three channels have readings, take the median. If only two channels have readings, the larger of the two values ​​is used; if only one channel has a reading, that reading is used; if all three channels are missing, no reading is generated. For example, ,but ;by For example, ,but Define the upper and lower bounds of the feasible box-type constraints for each time slice for subsequent conservation balancing. and ,in , Missing channels are not included in the extreme value calculation; if all three channels are missing, then the time slice is... Undefined and cannot be directly adjusted within the short window. Short window conservation balancing is based on the window length. The step size is The sliding method covers the base time slice; in this example, the window set is... , For any short window Define the total number of channels , , Total target amount of the window Current total ,difference .when At that time, according to the conservation rules, without exceeding the upper and lower bounds of time-slice feasibility, Perform monotonic adjustments to obtain the balanced values. In this example, the total number of channels in all short windows, the target total, the current total, and the difference are as follows: (, , , ; have , , , ; have , , , ; have , , , ; have , , , .

[0049] by Taking the balancing of equations as an example, because The total amount needs to be reduced, and only those that meet the requirements should be selected. Time slices, according to intervals Sort by size from largest to smallest, and apply reductions one by one. And each time no more than ,until The actual time slice in this window where adjustments occurred is... For example, in Depend on Down to (reduce ), corresponding to the lower bound Upper Realm ;exist Depend on Down to (reduce ),correspond , After processing according to this rule, the result is... Using the same method, of A time slice, of A time slice, of A time slice, of Each time slice was adjusted to meet the requirements. Then proceed to the consistent postback in the middle window: This embodiment covers A medium window that defines the total amount of the three channels. , , Total target volume in the middle window Current total .

[0050] Calculated , ,and Therefore Since the current total is equal to the target total, there is no difference in the middle window that needs to be returned for allocation; the cumulative value of the long window within this coverage area is consistent with that of the middle window. If other data conditions occur... Then, the difference is allocated according to the proportion of "adjustable space" of each short window in the current direction: when an increase is needed, the short windows... The adjustable space is defined as When a reduction is required, it is defined as... The difference will be allocated to each short window according to their proportion of the total adjustable space. and in each The short-window conservation balancing rule is invoked again to complete the point-level landing. For missing time slot detection, the trigger condition is "after the above allocation, there is still an unallocated portion of the difference and there is a time slot within the short window where the preprocessing unit has not generated a temporary emission estimate". In this case, the remaining difference is equally distributed among these time slots. If a time slot to which the difference is distributed has at least one channel reading in the same time slot, the average value is affected. Constraints: If all three channels are missing, then the values ​​are allowed to be between the minimum and maximum values ​​of the temporary emission estimates already determined in this short window, in order to complete the difference allocation.

[0051] In this example, since at least one type of channel is available for all time slices and no discrepancy occurs in the middle window, the missing time slice detection is not triggered. The complete process of preprocessing, short window conservation balancing, middle window consistency backhaul, and missing time slice detection is considered as one iteration, denoted as the... Post-wheel temporary emission estimation sequence is Define the global change ratio This example is the first one. Rotation arrive The total change is ,Right now ;No. If the round is executed again with the input unchanged, the same result will be obtained. Therefore The process stops when the convergence condition is met. The final output is an emission estimation sequence organized by monitoring unit and base time slice. Short window cumulative emissions Cumulative emissions with the middle window Within this coverage area, the cumulative amount of the long window is consistent with that of the medium window.

[0052] Figure 2This paper demonstrates the key principle of the window-consistent backhaul adjustment technique in the multi-scale variational Bayesian inference system. As shown in the figure, the horizontal axis represents the short window number, with each short window containing 20 base time slices. The vertical axis represents the cumulative emissions in tons of CO2. The figure illustrates the processing of nine short windows, where short windows 1-5 constitute a medium window, and short windows 1-9 constitute a long window. Specifically, the solid black bars in the figure represent the cumulative emissions of each short window before backhaul adjustment, with values ​​of 750, 770, 630, 890, 740, 780, 640, 950, and 770 tons of CO2, respectively. The target total emission line for the medium window, presented as a dashed line, spans short windows 1-5, indicating that the average value of the sum of the three channel readings within this medium window is 1250 tons of CO2. The target total emission line for the long window, presented as an even longer dashed line, spans short windows 1-9, with a target total emission of 1250 tons of CO2. The core of this consistent backhaul adjustment process lies in the differential allocation mechanism. When there is a difference between the target total and the current total in a medium-sized or long window, the system first calculates the adjustable space of each short window. As shown in the figure, short windows 1-5 have adjustable spaces that can reduce CO2 by 80, 75, 60, 90, and 85 tons respectively, while short windows 6-9 have adjustable spaces that can increase CO2 by 95, 70, 110, and 75 tons respectively. The system allocates the difference according to the proportion of each short window's adjustable space in the total adjustable space of the window. As can be seen from the figure, the medium window generates a difference of -120 tons of CO2, which needs to be reduced and allocated among short windows 1-5; the long window generates a difference of +85 tons of CO2, which needs to be increased and allocated among short windows 6-9. After the feedback adjustment, the cumulative emissions of each short window are shown in the dashed bar chart in the figure, achieving cross-scale numerical consistency and ensuring the conservation and balance relationship between windows of different scales.

[0053] Figure 3The iterative convergence characteristics of the carbon emission accurate measurement system based on multi-scale variational Bayesian inference are illustrated in detail. The horizontal axis represents the number of iterations, ranging from 0 to 100; the vertical axis represents the rate of change, ranging from 0.5% to 25.0%. The horizontal dashed line in the figure indicates the preset convergence threshold of 1.0%, which is a key parameter for determining whether the system has reached the convergence condition. The iterative convergence curve clearly shows three typical stages. The first stage is a rapid decline stage, starting from an initial rate of change of 25.6%, rapidly decreasing to about 10% in the first 15 iterations. This stage reflects the system's ability to quickly correct initial abnormal data and balancing discrepancies. The second stage is a gradual decline stage, from the 15th to the 40th iteration, the rate of change gradually decreases from 10% to about 2%. This stage reflects the process of the system making fine-grained consistency adjustments across multi-scale windows. The third stage is the convergence stage, starting from the 40th iteration, the rate of change gradually approaches the 1% convergence threshold. The critical convergence point occurred in the 47th iteration, when the rate of change first dropped below 1.0%, reaching 0.93%, thus satisfying the preset convergence condition. This important node is clearly marked by a circle in the figure. The rate of change is calculated as the ratio of the absolute values ​​of the differences between the estimated temporary emissions of all monitoring units across all base time slices and their corresponding values ​​from the previous iteration, to the total sum of all estimated temporary emissions from the previous iteration. The system is designed to have a maximum of 200 iterations, as shown by the vertical dotted line on the right side of the figure, ensuring that the system can complete the calculation within a finite time even under extreme conditions. This convergence characteristic verifies the stability and computational efficiency of the multi-scale variational Bayesian inference algorithm, providing reliable technical support for practical engineering applications.

[0054] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A system for accurate measurement of carbon emissions based on multi-scale variational Bayesian inference, characterized in that, The system comprises: a data acquisition unit for acquiring basic observation parameters at the inlet of each carbon emission monitoring unit, including instantaneous consumption readings of single fuel consumption metering points, exhaust flow readings of single exhaust ports, and instantaneous output readings of single output metering points; adding time stamps to the basic observation parameters according to the acquisition time, and marking the corresponding monitoring unit number to form a raw basic data sequence arranged in chronological order; a time alignment and channel classification unit for aligning the raw basic data sequence into basic time slices according to the minimum adjacent acquisition interval; dividing each monitoring unit into fuel channels, exhaust channels, and output channels according to physical quantities, and organizing multi-scale observation structures of short windows, medium windows, and long windows for each physical channel respectively; a Bayesian inference unit for proportionally aligning the fuel channels and the output channels with reference to the exhaust channels; performing sliding inspection on the three types of channels on the basic time slices and replacing abnormal readings; generating temporary emission estimates according to available readings by time slice; implementing conservation balancing on the temporary emission estimates in the short window to make them consistent with the total amount of readings in the three types of channels in the short window; aggregating the short window results to the medium window and the long window and adjusting them by difference to achieve cross-scale consistency; outputting the emission estimate sequence organized by monitoring units and basic time slices and the cumulative emissions of each window when the preset convergence condition is met, otherwise continue iteration.

2. The multi-scale variational Bayesian inference based carbon emission precise measurement system of claim 1, wherein, The lengths of the short window, the medium window, and the long window are 20, 100, and 500 basic time slices respectively.

3. The multi-scale variational Bayesian inference based carbon emission precise measurement system of claim 2, wherein, The Bayesian inference unit comprises: a preprocessing unit, a short window conservation balancing unit, a medium window and long window consistency feedback unit, a missing time slice detection unit, and a result output unit; wherein the preprocessing unit is used to perform sliding inspection on the three types of channels on the basic time slices and replace abnormal readings; the short window conservation balancing unit is used to generate emission estimate initial values according to available readings by time slice; implement conservation balancing on the emission estimate initial values in the short window to make them consistent with the total amount of readings in the three types of channels in the short window; the medium window and long window consistency feedback unit is used to aggregate the short window results to the medium window and the long window and adjust them by difference to achieve cross-scale consistency; the result output unit is used to output the emission estimate sequence organized by monitoring units and basic time slices and the cumulative values of each window when the preset convergence condition is met, otherwise continue iteration.

4. The multi-scale variational Bayesian inference-based carbon emission precise measurement system of claim 3, wherein, The execution process of the preprocessing unit specifically comprises: in each monitoring unit, time slice scanning is performed on the three types of channels one by one; when the absolute value of the difference between the reading of a channel in a time slice and the median of the readings of the channel in adjacent 9 time slices, including the previous 4 time slices, the current time slice and the next 4 time slices, exceeds half of the difference between the maximum and minimum of the readings of the adjacent 9 time slices, the reading of the channel in the time slice is replaced by the median of the readings of the adjacent 9 time slices; in each monitoring unit, for each basic time slice, the median of the available readings of the three types of channels in the time slice is taken as the temporary emission estimation value of the time slice; if only two channels have readings, the larger one is taken; if only one channel has a reading, the reading is taken; if all three channels are missing, the temporary emission estimation value of the time slice is not generated, and is determined uniformly when the window is returned.

5. The multi-scale variational Bayesian inference-based carbon emission precise measurement system of claim 4, wherein, The execution process of the short window conservation balancing unit specifically comprises: in each monitoring unit, for each short window, the channel total quantity of the fuel channel, the exhaust channel and the output channel in the short window is calculated respectively, which is the sum of the readings of the channel in the short window, and the arithmetic average of the three is taken as the target total quantity of the short window; the sum of all the temporary emission estimation values in the short window is calculated as the current total quantity; when the target total quantity and the current total quantity are not equal, the temporary emission estimation values of each time slice are adjusted in the short window according to the following conservation rules until they are equal: if it needs to be increased, the time slices whose temporary emission estimation values are lower than the maximum value of the readings of the three types of channels in the same time slice are selected in the short window first, and are increased one by one in the order from the largest gap to the smallest, and the amplitude of each increase does not exceed the whole or the remaining difference between the maximum value of the readings of the three types of channels in the time slice and the current temporary emission estimation value, and the total quantity is equal first; if it needs to be reduced, the time slices whose temporary emission estimation values are higher than the minimum value of the readings of the three types of channels in the same time slice are selected in the short window first, and are decreased one by one in the order from the largest gap to the smallest, and the amplitude of each decrease does not exceed the whole or the remaining difference between the current temporary emission estimation value and the minimum value of the readings of the three types of channels in the time slice, and the total quantity is equal first; if there is a missing mark in a time slice, the upper limit or the lower limit is unavailable, the time slice is only allowed to be adjusted in the known direction; only the upper limit is available, only increase is allowed, only the lower limit is available, only decrease is allowed, and the time slice is skipped when both the upper limit and the lower limit are unavailable.

6. The multi-scale variational Bayesian inference-based carbon emission precise measurement system of claim 5, wherein, The execution process of the medium window and long window uniformization feedback unit specifically comprises: in each monitoring unit, first aggregate the temporary emission estimates in the short window after being balanced to the medium window and the long window by time summation to obtain the current total of each window; calculate the average of the sum of the three types of channel readings in each medium window and long window, as the target total of the corresponding window; when there is a difference between the target total and the current total of a certain medium window or long window, allocate the difference according to the adjustable space in the contained short window: the adjustable space is defined as the total sum of all time slices in each short window that can still be increased or decreased without breaking through the upper and lower limits of the three types of channel readings of the corresponding time slice; allocate the difference to the short window according to the proportion of the adjustable space of each short window in the total adjustable space of the short window, and then process it in the short window through the conservation rule of the short window conservation balancing unit.

7. The multi-scale variational Bayesian inference-based carbon emission precise measurement system of claim 6, wherein, The execution process of the missing time slice detection unit specifically comprises: when there is still an unallocated part of the difference allocated by the medium window and long window uniformization feedback unit to a certain short window or its upper window, and there are time slices in the short window for which the temporary emission estimate is not generated in the preprocessing unit, then the remaining difference is evenly divided among these time slices; for each divided time slice, if there is at least one channel reading in the same time slice, the divided value of the time slice cannot exceed the upper and lower limit range that can be taken by the time slice; if all three channels are missing, the time slice is allowed to take a value between the minimum and maximum values of the temporary emission estimate determined in the short window to complete the difference allocation.

8. The multi-scale variational Bayesian inference-based carbon emission precise measurement system of claim 7, wherein, The complete process of the preprocessing unit, the short window conservation balancing unit, the medium window and long window uniformization feedback unit, and the missing time slice detection unit is regarded as one round of iteration. After completing one round, calculate the sum of the absolute values of the difference between the temporary emission estimates of all basic time slices of all monitoring units and the corresponding values in the last round, and the proportion of the sum of all temporary emission estimates in the last round; when the proportion is not more than 1%, or the update round reaches 200 rounds, stop; Otherwise, return to the preprocessing unit for the next round; After stopping, output the emission estimate sequence organized by monitoring units and basic time slices, and the cumulative emission of short windows, medium windows and long windows.

Images