Regional carbon emission accounting management system based on big data analysis

Abstract

The present application relates to a regional carbon emission accounting management system based on big data analysis. The system includes a central coordination server and multiple regional edge nodes. The regional edge node uses a feature decoupling encoder to decompose local industrial production process time series data and energy consumption monitoring data into regional exclusive working condition features and carbon emission common features. The carbon emission common features are used for local model forward propagation and gradient calculation. The gradient update quantity and carbon emission common feature representation are uploaded to the central coordination server. The central coordination server dynamically adjusts the weight coefficient according to the gradient variance to perform weighted aggregation and issue global model parameters. The system blocks the cross-domain transmission of original time series energy consumption data, reduces the communication bandwidth load, retains regional exclusive working condition features at the edge, and fuses them with global model parameters, eliminating the systematic drift of the accounting system caused by the single centralized global model deviating from the local physical working condition, and solving the cross-domain generalization failure problem.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, specifically relating to the technical field of regional energy consumption data processing and carbon emission accounting using a distributed collaborative computing model. Background Technology

[0002] Existing cross-regional industrial carbon emission accounting systems generally adopt a centralized data aggregation and unified model training architecture. Sub-factories in each region directly transmit time-series energy consumption data from their underlying energy meters and production equipment operating status data to a central cloud server via communication networks. After receiving the raw monitoring data from all regions, the central cloud server aggregates the multi-source heterogeneous data into a single data pool and uses the aggregated full sample to train a single centralized global carbon emission accounting model. Once the model training is complete, the central cloud server distributes this global model, containing fixed parameter weights, to edge computing nodes in each region. These edge computing nodes then load the global model to perform numerical carbon emission accounting for their local real-time industrial production processes.

[0003] Based on the conventional implementation of centralized data aggregation and unified model training, a core technical problem exists: a single centralized global model fails to generalize across regions when adapting to the differentiated industrial production conditions in various areas. The central cloud server forcibly uses a single model structure to fit the mixed data distribution of all regions, erasing unique local physical characteristics of each sub-region, such as equipment aging levels and specific raw material ratios, during training. When this global model, detached from local conditions, is distributed to specific regional edge computing nodes, the weight parameters fixed in the global model deviate from the actual physical production process of that region. This causes the calculation model to systematically drift its output when facing region-specific conditions, resulting in calculation distortion. Summary of the Invention

[0004] The purpose of this invention is to provide a solution to the problems described in the background section.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The regional carbon emission accounting management system based on big data analysis includes a central coordination server and multiple regional edge nodes. Each edge node locally stores time-series data of industrial production processes and energy consumption monitoring data for its corresponding region. Each edge node has a built-in feature decoupling encoder, which decomposes the time-series data of industrial production processes and the energy consumption monitoring data into region-specific operating condition features and common carbon emission features. The central coordination server has an adaptive aggregation weight allocation component. This component dynamically adjusts the weight coefficients of each edge node in the global model aggregation based on the gradient variance of the distribution of common carbon emission features uploaded by each edge node and the local data volume. Each edge node uses the decoupled common carbon emission features to perform forward propagation and gradient calculation of its local model, and uploads the gradient update and the common carbon emission feature representation to the central coordination server. The central coordination server performs weighted aggregation based on the weight coefficients and then issues global model parameters. The edge nodes fuse and combine the updated global model parameters with the locally retained region-specific operating condition features to output the regional carbon emission accounting results.

[0006] Preferably, the feature decoupling encoder includes a dual-branch feature extraction network and an orthogonal constraint layer. The dual-branch feature extraction network contains a shared bottom-level convolutional kernel and two independent fully connected layer branches. The industrial production process time-series data and the energy consumption monitoring data are input into the shared bottom-level convolutional kernel to extract basic time-series features. The two independent fully connected layer branches perform nonlinear mapping on the basic time-series features respectively. The orthogonal constraint layer is connected to the output of the two independent fully connected layer branches. The orthogonal constraint layer calculates the Gram matrix between the feature vectors output by the two branches and minimizes the off-diagonal elements, forcing the region-specific operating condition features and the common carbon emission features output by the two independent fully connected layer branches to maintain an orthogonal projection relationship in the feature space, thereby eliminating linear correlation interference between feature vectors.

[0007] Preferably, the adaptive aggregation weight allocation component includes a gradient similarity calculation component and a dynamic weight update component. The gradient similarity calculation component receives the gradient update amount uploaded by each region edge node and calculates the cosine similarity matrix between the gradient update amounts of any two region edge nodes. The dynamic weight update component calculates the deviation index of each region edge node relative to the global average gradient based on the cosine similarity matrix. The dynamic weight update component has a built-in decay factor mapping function. The deviation index is input into the decay factor mapping function and the weight suppression coefficient of the corresponding region edge node is output. When the deviation index of the region edge node exceeds a preset dynamic deviation threshold, the weight suppression coefficient decreases exponentially, reducing the dominant influence of abnormal edge nodes on the global model parameter update.

[0008] Preferably, before uploading the gradient update and the common carbon emission feature representation, the regional edge node performs local differential privacy noise addition processing. The regional edge node has a built-in adaptive gradient pruner and noise injector. The adaptive gradient pruner calculates the norm distribution of the gradient of each layer of the local model in the current communication round and prunes gradient components whose norm exceeds the moving average of the norm in the historical rounds to the moving average of the norm in the historical rounds. The noise injector generates a noise matrix that conforms to a Gaussian distribution based on the dimension of the pruned gradient components and superimposes the noise matrix into the pruned gradient components and the common carbon emission feature representation. The standard deviation of the Gaussian distribution is inversely proportional to the square root of the amount of time-series data of the industrial production process stored locally by the regional edge node.

[0009] Preferably, the regional edge node incorporates a feature fusion gating network. This network receives updated global model parameters from the central coordination server and locally retained regional-specific operating condition features. The feature fusion gating network includes a splicing layer, a fully connected mapping layer, and a Sigmoid activation layer. The splicing layer concatenates the hidden layer weight matrix of the global model parameters with the regional-specific operating condition features along the channel dimension to generate a fusion tensor. The fully connected mapping layer performs dimensionality reduction mapping on the fusion tensor. The Sigmoid activation layer transforms the dimensionality-reduced output into a gating weight coefficient matrix between 0 and 1. The regional edge node uses the gating weight coefficient matrix to perform element-wise weighted modulation on the hidden layer weight matrix of the global model parameters before inputting it into the calculation engine to output the regional carbon emission calculation result.

[0010] Preferably, before inputting the industrial production process time series data and the energy consumption monitoring data into the feature decoupling encoder, the region edge node performs a multi-scale spatiotemporal alignment operation through a data preprocessor. The data preprocessor obtains the production equipment operating cycle corresponding to the industrial production process time series data and the sensor sampling frequency corresponding to the energy consumption monitoring data. Using the production equipment operating cycle as the reference time window, the energy consumption monitoring data is downsampled and resampled to align the timestamps of the two. At the same time, the equipment number label and spatial location coordinates in the industrial production process time series data are extracted, and the spatial location coordinates are mapped to a pre-divided regional grid topology. The processed data is packaged into a three-dimensional structured tensor containing time dimension, spatial dimension and variable dimension and input into the feature decoupling encoder.

[0011] Preferably, the orthogonal constraint layer is further coupled with a physical mechanism discriminator, which is embedded between the two independent fully connected layer branches. The physical mechanism discriminator receives the region-specific operating condition characteristics and a pre-stored set of regional industrial material balance equations. The physical mechanism discriminator substitutes the material consumption rate parameter in the region-specific operating condition characteristics into the set of regional industrial material balance equations to calculate the material balance difference. The material balance difference is introduced as a penalty term into the Gram matrix loss function of the orthogonal constraint layer. While minimizing the off-diagonal elements of the Gram matrix, the material balance difference is constrained to approach 0, thus forcing the region-specific operating condition characteristics to conform to the regional physical mass conservation law under the condition of satisfying the orthogonal constraints.

[0012] Preferably, the dynamic weight update component is also connected to a historical state buffer queue, which stores the time series of deviation indexes of each regional edge node calculated by the central coordination server in the past preset communication rounds. The dynamic weight update component extracts the time series of deviation indexes to calculate the deviation change rate of each regional edge node. When the deviation index of the regional edge node does not exceed the dynamic deviation threshold, but the deviation change rate shows a monotonically increasing trend and the absolute value of the change rate exceeds the preset change rate surge threshold, the dynamic weight update component triggers a pre-warning mechanism to lock the weight suppression coefficient of the regional edge node to a fixed value of the current round, thereby preventing the continuous drift of model weights caused by gradual hardware failure.

[0013] Preferably, a gradient sparsity mask generator is provided between the adaptive gradient pruner and the noise injector. The gradient sparsity mask generator receives the pruned gradient components, arranges them in descending order according to the absolute value of the gradient components, extracts the gradient component indices at the first preset percentage positions to construct a sparse binary mask, and the noise injector performs element-wise multiplication of the Gaussian-distributed noise matrix with the sparse binary mask to filter out redundant noise superimposed at the zero-value gradient index positions. The filtered noise matrix is ​​then superimposed only at the retained gradient component index positions, reducing the communication load of the region edge nodes transmitting the gradient update amount and the carbon emission common feature representation to the central coordination server.

[0014] Preferably, the regional edge node is further connected in series with a residual feedback corrector after the feature fusion gating network. The residual feedback corrector obtains the regional carbon emission accounting result output by the regional edge node and the actual fuel consumption in the industrial production process time series data of the current communication round, calculates the absolute value of the residual between the regional carbon emission accounting result and the baseline carbon emission calculated based on the actual fuel consumption using the static emission factor method, and the residual feedback corrector has a built-in proportional-integral controller. The absolute value of the residual is input into the proportional-integral controller to generate a feedback bias term, and the feedback bias term is superimposed on the regional carbon emission accounting result to output the final regional carbon emission accounting result, thereby suppressing the systematic static bias introduced in the gating weighted modulation process.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention constructs an adaptive feature-decoupled federated learning architecture through a central coordination server and regional edge nodes. Regional edge nodes use a feature-decoupled encoder to decompose local time-series data into region-specific operating condition features and common carbon emission features, and only upload gradient updates calculated based on the common carbon emission features. The central coordination server dynamically adjusts the aggregation weights based on the gradient variance and distributes global model parameters. Regional edge nodes fuse the received global model parameters with the locally retained region-specific operating condition features to output the calculation result. This architecture blocks cross-domain transmission of the original time-series energy consumption data, reduces communication bandwidth load, and retains region-specific operating condition features at the edge, ensuring that the distributed global model parameters, after fusing local operating condition features, accurately reflect the actual physical production process of the region. This eliminates the systematic drift caused by a single centralized global model being detached from local physical operating conditions and solves the problem of cross-domain generalization failure.

[0016] 2. An orthogonal constraint layer, in conjunction with a physical mechanism discriminator, introduces minimization of off-diagonal elements of the Gram matrix and a penalty for material income / expenditure differences in the dual-branch feature extraction network. This forces region-specific operating condition features and common carbon emission features to conform to the law of mass conservation under the condition of satisfying orthogonal constraints, eliminating linear correlation interference between feature vectors. A dynamic weight update component, combined with a historical state buffer queue, triggers weight suppression or coefficient locking based on deviation indicators and deviation change rates, blocking parameter drift caused by abnormal nodes and hardware gradual failure nodes. A gradient sparsity mask generator, placed between the adaptive gradient pruner and the noise injector, filters redundant noise at zero-value index positions by constructing a sparse binary mask, reducing the communication load of gradient transmission. A residual feedback corrector, serially connected after the feature fusion gating network, uses a proportional-integral controller to convert the absolute value of the residual difference between the calculation result and the baseline carbon emission into a feedback bias term superimposed on the final output, suppressing the systematic static bias introduced by the gating weighted modulation process. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall operation of the regional carbon emission accounting and management system based on big data analysis according to the present invention. Figure 2 This is a flowchart of the dual-branch feature extraction and orthogonal constraint processing in the feature decoupling encoder of the present invention; Figure 3 This is a flowchart of gradient similarity calculation and dynamic weight update in the adaptive aggregation weight allocation component of the present invention; Figure 4 A flowchart illustrating the local differential privacy noise addition and gradient sparsification processing performed on the region edge nodes of this invention; Figure 5 This is a flowchart illustrating the fusion of global parameters and local features in the feature fusion gating network of the present invention. Figure 6 This is a flowchart illustrating the deviation suppression process of the residual feedback corrector of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0019] Please refer to Figure 1 This embodiment provides a regional carbon emission accounting and management system based on big data analytics, including a central coordination server and multiple regional edge nodes. The central coordination server is deployed on a physical cluster in the core data center, equipped with multiple high-performance central processing units, large-capacity error-correcting code memory, and a high-speed non-volatile storage array, responsible for executing weighted aggregation and scheduling tasks for global model parameters. Multiple regional edge nodes are deployed within industrial parks in different geographical locations, each equipped with an edge computing acceleration chip and local storage media. The regional edge nodes establish communication connections with the distributed control systems and energy consumption monitoring instruments in the local production workshops via industrial Ethernet switches, and simultaneously establish remote communication connections with the central coordination server via encrypted virtual private network tunnels. The regional edge nodes locally store the time-series data of the industrial production process and energy consumption monitoring data for their respective regions. Industrial production process time-series data includes the operating status words of production equipment at discrete time points, valve opening signals, spindle speed pulses, and sampling sequences of temperature and pressure sensors inside the reactor; energy consumption monitoring data includes the integrated values ​​of active and reactive power recorded by smart meters, the cumulative value of instantaneous flow rate of gas flow meters, and the instantaneous feed rate sequence of coal belt scales.

[0020] Existing cross-regional industrial carbon emission accounting systems generally adopt a centralized data aggregation and unified model training architecture. Sub-factories in each region directly transmit time-series energy consumption data from their underlying energy meters and production equipment operating status data to a central cloud server. The central cloud server forcibly uses a single model structure to fit the mixed data distribution of all regions, erasing local physical operating condition characteristics unique to each sub-region, such as equipment aging and specific raw material ratios, during the training process. When the global model, detached from local operating condition characteristics, is distributed to specific regional edge computing nodes, the weight parameters fixed in the global model deviate from the actual physical production process of that region. This causes the accounting model to systematically drift when facing region-specific operating conditions, resulting in accounting distortion. To solve this cross-domain generalization failure problem, this embodiment constructs an adaptive feature decoupling federated learning architecture. The regional edge nodes have built-in feature decoupling encoders. These encoders perform feature space mapping and decoupling operations on the time-series data of industrial production processes and energy consumption monitoring data locally, decomposing them into region-specific operating condition characteristics and common carbon emission characteristics. Regional-specific operating condition characteristics characterize the local physical laws unique to the region, such as the deviation in basic energy consumption caused by equipment wear and tear, and the incomplete combustion state caused by fluctuations in specific ore grades; common carbon emission characteristics characterize the cross-domain consistency laws, such as the exothermic law of carbon-oxygen atom combination and the general thermodynamic state transition process of fossil fuel combustion, which are followed in different regions.

[0021] Regional edge nodes utilize decoupled common carbon emission features locally for forward propagation and gradient calculation of their local models. The calculated gradient updates and common carbon emission feature representations are then uploaded to the central coordination server via an encrypted channel. Because only common feature representations and gradient information, decoupled from regional privacy attributes, are uploaded, this architecture physically blocks cross-domain transmission of raw time-series energy consumption data, reducing communication bandwidth load. The central coordination server incorporates an adaptive aggregation weight allocation component. This component dynamically adjusts the weight coefficients of each regional edge node in the global model aggregation based on the distribution of common carbon emission features uploaded by each regional edge node and the gradient variance of the local data volume. The central coordination server performs a weighted aggregation operation based on the calculated weight coefficients, generating updated global model parameters and distributing them to each regional edge node. After receiving the updated global model parameters, the regional edge nodes perform a fusion and splicing operation on the global model parameters and the regional-specific operating condition features that have not been uploaded locally. This allows the global model parameters to be consistent with the actual physical production process in the region after incorporating local operating condition features, eliminating the systematic drift caused by the single centralized global model being detached from local physical operating conditions, and outputting accurate regional carbon emission accounting results.

[0022] In one embodiment, before inputting industrial production process time-series data and energy consumption monitoring data into the feature decoupling encoder, the regional edge node performs a multi-scale spatiotemporal alignment operation through a data preprocessor. The industrial production process time-series data and energy consumption monitoring data originate from different underlying hardware acquisition terminals, exhibiting significant heterogeneity. The data preprocessor obtains the production equipment operating cycle corresponding to the industrial production process time-series data and the sensor sampling frequency corresponding to the energy consumption monitoring data. The production equipment operating cycle is obtained by parsing the encoder pulse signal of the equipment's main drive motor. Within a complete production cycle, the equipment undergoes stages such as preheating, steady-state operation, and deceleration shutdown. Using this production equipment operating cycle as a baseline time window, the data preprocessor performs downsampling and resampling processing on the energy consumption monitoring data. Assuming the baseline time window length for the industrial production process time-series data is... The sampling interval corresponding to the sensor sampling frequency of energy consumption monitoring data is The data preprocessor operates within the reference time window. Inside, for the included Each energy consumption monitoring sampling point undergoes either mean aggregation or piecewise integration to ensure strict alignment of their timestamps. The mathematical expression for mean aggregation is: ,in This is the raw energy consumption monitoring data. This provides energy consumption data aligned with timestamps. Simultaneously, the data preprocessor extracts equipment identification tags and spatial coordinates from the industrial production process time-series data. The spatial coordinates are obtained based on the Global Positioning System (GPS) or indoor positioning base stations within the factory. The data preprocessor maps these spatial coordinates to a pre-defined regional grid topology. The regional grid topology employs a quadtree-based spatial discretization method, dividing the factory's physical plane into sections with side lengths of... A set of square grids. For any spatial coordinates... Its mapped grid index The calculation formula is , The data preprocessor jointly encodes the timestamp-aligned energy consumption data, the cycle-aligned production process status data, and the corresponding grid index, packaging the processed data into a three-dimensional structured tensor containing time, space, and variable dimensions. In this tensor, Indicates the length of the time window sequence. Indicates the total number of spatial grids. This represents a multi-variable dimension including electrical power, gas flow rate, temperature, and pressure. The three-dimensional structured tensor is stored contiguously in memory using row-major order, and its internal elements have a linear addressing formula: This three-dimensional structured tensor is directly used as the input data for the subsequent feature decoupling encoder.

[0023] In one embodiment, reference Figure 2 The feature decoupling encoder consists of a two-branch feature extraction network and an orthogonal constraint layer. The two-branch feature extraction network contains a shared bottom-level convolutional kernel and two independent fully connected layer branches. Three-dimensional structured tensors. The input is fed into a shared underlying convolutional kernel, which is configured as a three-dimensional convolutional operation layer, containing... Each size is A three-dimensional convolutional filter. The sliding step size in the time dimension is... The sliding step size in the spatial dimension is The sliding step size in the variable dimension is For each element in the output base time series feature map The calculation method involves performing a three-dimensional dot product operation on the local receptive field region in the input tensor starting from the corresponding position, adding a bias term, and then activating it using a linear rectified function. The calculation formula is as follows: The shared underlying convolutional kernel extracts low-level temporal pattern features that are universal across different regions through a parameter-sharing mechanism, such as periodic load fluctuations and stepped start-stop power characteristics. The shared underlying convolutional kernel outputs the basic temporal features. After flattening, the inputs are fed into two independent fully connected layer branches. These two branches perform nonlinear mappings on the basic temporal features, with the first branch configured as the mapping function. Used to extract region-specific operating condition features The second independent fully connected layer branch is configured as a mapping function. Used to extract common characteristics of carbon emissions ,in is the dimension of the feature vector.

[0024] An orthogonal constraint layer connects to the outputs of two independent fully connected layer branches. To eliminate linear correlation interference between feature vectors, the orthogonal constraint layer calculates the Gram matrix between the output feature vectors of the two branches and minimizes the off-diagonal elements, forcing the region-specific operating condition features and common carbon emission features output by the two independent fully connected layer branches to maintain an orthogonal projection relationship in the feature space. Specifically, for batch processing... The region-specific working condition feature matrix is ​​represented as follows: (The matrix consists of samples.) The common characteristics matrix of carbon emissions is represented as follows: Gram matrix The calculation formula is Gram matrix The first in Line number Column elements The first characteristic representing the region-specific operating conditions The first dimension and common characteristics of carbon emissions The sum of inner products between each dimension. The orthogonal constraint layer constructs the orthogonal loss function. The loss function is defined as the Gram matrix. The sum of squares of the off-diagonal elements is expressed as follows: ,in To preserve the diagonal elements of the Gram matrix and set the off-diagonal elements to zero, This represents the Frobenius norm of the matrix. During the backpropagation phase of model training, the orthogonal loss function... Calculate the partial derivatives of the network weights for the two independent fully connected layer branches, and update the weight parameters using the gradient descent algorithm, so that the off-diagonal elements... Approaching 0. When When converged to the minimum value, the subspace spanned by the region-specific operating condition features and the subspace spanned by the common carbon emission features are orthogonal to each other, completely severing the linear information flow path between the two and preventing the implicit encoding of regional privacy information in the common carbon emission features through linear correlation.

[0025] In one embodiment, the orthogonal constraint layer is also coupled to a physical mechanism discriminator. The physical mechanism discriminator is embedded between two independent fully connected layer branches. The physical mechanism discriminator receives region-specific operating condition characteristics and a pre-stored set of regional industrial material balance equations. The regional industrial material balance equations are a set of algebraic equations based on the law of conservation of mass. For a specific industrial production region, the input raw material mass flow rate, the output product mass flow rate, the generated waste gas and waste residue mass flow rate, and the internal accumulation satisfy the following relationships: The physical mechanism discriminator substitutes the material consumption rate parameters from the region-specific operating conditions into the region's industrial material balance equations to calculate the material balance difference. Specifically, the physical mechanism discriminator internally incorporates a lightweight, fully connected decoding layer that uses region-specific operating conditions... As input, map the output to the predicted vector of key material consumption rates for that region within the current time window. Substitute the predicted vector into the pre-stored matrix form of the material balance equations. ,in This is the material balance coefficient matrix. This is a vector of constant terms. Material balance difference. The calculation formula is The physical mechanism discriminator incorporates the material cost difference as a penalty term into the Gram matrix loss function of the orthogonal constraint layer. While minimizing the off-diagonal elements of the Gram matrix, it constrains the material cost difference to approach zero, thus constructing a joint loss function. ,in The weight hyperparameters are penalized for physical mechanisms. By jointly optimizing orthogonal constraints and physical conservation constraints in the loss function, the region-specific operating condition features are forced to conform to the regional physical mass conservation law under the condition of satisfying orthogonal constraints. This prevents the feature decoupling process from getting trapped in mathematical local optima that do not conform to physical common sense, and improves the physical interpretability of feature representation.

[0026] In one embodiment, reference Figure 3 The adaptive aggregation weight allocation component includes a gradient similarity calculation component and a dynamic weight update component. Local edge nodes utilize the decoupled common carbon emission features for forward propagation of their local models, calculating the mean squared error loss between the local carbon emission accounting predictions and the actual labeled values. Based on this loss function, the local edge nodes calculate the partial derivatives of the network parameters of each layer of the local model with respect to the loss function using the backpropagation algorithm, forming the gradient update amount. The gradient similarity calculation component receives the gradient update amounts uploaded by each local edge node. It is assumed that the system contains... The edge node of the region, the first The gradient update values ​​of each node are flattened into a vector. ,in This represents the total dimension of the model parameters. The gradient similarity calculation component calculates the cosine similarity matrix between the gradient updates of edge nodes of any two regions. .matrix elements in The calculation formula is Cosine similarity matrix This reflects the degree of consistency of edge nodes in different regions in the direction of model parameter optimization.

[0027] The dynamic weight update component calculates the deviation index of each region's edge nodes relative to the global average gradient based on the cosine similarity matrix. (Global average gradient) The calculation formula is For the first The deviation index of each regional edge node Defined as the complement of the cosine similarity between the gradient of this node and the global average gradient, the formula is as follows: Deviation index The value of ranges from zero to two; a larger value indicates a greater difference between the gradient update direction of that node and the direction of global consensus. The dynamic weight update component has a built-in decay factor mapping function. The deviation index is input into the decay factor mapping function, which outputs the weight suppression coefficient of the corresponding edge node. The decay factor mapping function uses a piecewise exponential decay mechanism, and its mathematical expression is: ,in For decay rate control parameters, This is a preset dynamic deviation threshold. The deviation index of the region edge nodes... Not exceeding the dynamic deviation threshold hour, Weighted suppression coefficient This node participates in global aggregation with full weight. The deviation index of the region's edge nodes... Exceeding the preset dynamic deviation threshold At that time, the weighted suppression coefficient This decreases exponentially, reducing the dominant influence of anomalous edge nodes on global model parameter updates. The central coordination server uses the calculated weight suppression coefficients of each node. Perform a weighted aggregation operation on the gradient update amounts to update the global model parameters. The calculation formula is ,in The global learning rate, These are the global model parameters from the previous round.

[0028] In one embodiment, the dynamic weight update component is also connected to a historical state buffer queue. The historical state buffer queue is a circular queue data structure allocated in the memory of the central coordination server, storing the time series of deviation indicators for each region's edge nodes calculated by the central coordination server within a preset number of communication rounds. Assume the queue depth is... For the first Each region's edge node stores a time series in its historical state buffer queue. superscript This indicates the current communication round. The dynamic weight update component extracts the deviation index time series and calculates the deviation change rate of edge nodes in each region. The rate of change of deviation is calculated using the first-order difference approximation, and the formula is as follows: The dynamic weight update component has an internal pre-warning mechanism logic module that judges the deviation index of the region edge nodes. Not exceeding the dynamic deviation threshold At that time, the system further detects the rate of change of deviation. The temporal characteristics. The system traverses the continuous historical state buffer queue. The rate of change sequence of each round To determine whether the sequence exhibits a monotonically increasing trend, i.e., to check whether it satisfies... For all All of these hold true. If the rate of change of deviation shows a monotonically increasing trend, and the absolute value of the rate of change in the current round is... Exceeding the preset threshold for sudden increase in rate of change The dynamic weight update component triggers a pre-warning mechanism. After the pre-warning mechanism is triggered, the dynamic weight update component adjusts the weight suppression coefficient of the region edge nodes. Lock the value to a fixed value for the current round, i.e., perform the assignment operation. This mechanism prevents the continuous drift of model weights caused by gradual hardware failures. It addresses hidden faults such as gradual degradation of sensor accuracy in industrial settings or decreased computing power due to poor heat dissipation in edge computing units. Before node data deteriorates to the point of triggering a high deviation threshold, it maintains the stability of global model parameters by locking the weight coefficients.

[0029] In one embodiment, reference Figure 4 Before uploading gradient updates and common carbon emission feature representations, the regional edge nodes perform local differential privacy noise addition. Each regional edge node incorporates an adaptive gradient pruner and a noise injector. The adaptive gradient pruner calculates the norm distribution of the gradients of each layer of the local model in the current communication round. It is assumed that the local model contains... The network layer, for the _th _th_ ... Layer parameter gradient tensor The adaptive gradient pruner calculates its Frobenius norm. The adaptive gradient pruner internally maintains a historical epoch norm moving average register for each layer of the network. In the current communication round After completion, the update formula for the historical round norm moving average is: ,in This is the momentum decay coefficient, with a value between 0.9 and 0.99. Before uploading in the current round, the adaptive gradient pruner prunes gradient components whose norm exceeds the moving average of the norm in previous rounds to the moving average of the norm in previous rounds. The pruning operation is achieved through element-wise scaling, resulting in pruned gradient components. The calculation formula is ,in The constant is a very small constant to prevent the denominator from being zero. This adaptive pruning strategy dynamically adjusts the pruning threshold based on the historical gradient fluctuation amplitude of each layer, avoiding the loss of information in small gradient layers caused by a globally uniform fixed pruning threshold.

[0030] The noise injector generates a Gaussian-distributed noise matrix based on the dimensions of the clipped gradient components. For the clipped gradient components... Noise injector generation and Gaussian noise matrix with perfectly uniform shape Standard deviation of Gaussian distribution The amount of industrial production process time-series data stored locally at the regional edge nodes is inversely proportional to the square root of the data volume, and its mathematical expression is: ,in The clipping boundary constant (i.e., the moving average of the historical cycle norm) ), Differential privacy budget parameters configured for the system, For the first The number of data entries in the industrial production process time-series data locally stored at each edge node in a given region. Based on the mathematical theorem of differential privacy, by introducing Gaussian noise inversely proportional to the square root of the data volume, it can be strictly guaranteed that even in the extreme case where an attacker knows all the underlying energy consumption time-series data except for one record, they cannot infer the existence of a specific record by observing the uploaded gradient update amount. This provides a provable privacy security boundary from a mathematical perspective. The noise injector generates a noise matrix... Superimposed on the clipped gradient components Characterization of common features with carbon emissions In the process, noisy data to be uploaded is generated. and .

[0031] In one embodiment, a gradient sparsity mask generator is positioned between the adaptive gradient clipper and the noise injector. The gradient sparsity mask generator receives the clipped gradient components. To reduce the communication load of edge nodes transmitting gradient updates and carbon emission common feature representations to the central coordination server, the gradient sparsity mask generator performs a sparsity encoding operation. First, the gradient sparsity mask generator modifies the clipped gradient components... Flattened into a one-dimensional vector ,in This represents the total number of elements in the parameters of this layer. Sort the elements in descending order of their absolute values ​​to obtain the absolute value sequence. ,satisfy The system presets a sparsity percentage parameter. Extract and arrange first The gradient component indices at percentage positions are used to construct a sparse binary mask. Specifically, the gradient component indices at the previous position are selected. The original index set corresponding to the largest absolute value Initialize a one-dimensional vector filled with zeros. , index set The element at the corresponding position is set to one, that is... if ,otherwise This vector This is a sparse binary mask. The noise injector generates a Gaussian noise matrix. Then, the Gaussian-distributed noise matrix is ​​multiplied element-wise with the sparse binary mask, i.e., the Hadamard product operation is performed. This operation filters out redundant noise superimposed at the zero-value gradient index positions, and only superimposes the filtered noise matrix onto the retained gradient component index positions. The final uploaded gradient data is represented as follows: During the communication transmission phase, the system uses a compressed sparse row format or a coordinate list format to package non-zero indices and their corresponding values, reducing the actual amount of data transmitted from the original... Reduce to This significantly reduces communication bandwidth usage while maintaining the strength of differential privacy.

[0032] In one embodiment, reference Figure 5 The regional edge nodes incorporate a built-in feature fusion gating network. This network receives updated global model parameters from the central coordination server and locally retained region-specific operating condition features. The global model parameters from the central coordination server include the weight matrix of a multi-layer neural network; the feature fusion gating network primarily targets the weight matrix of the core hidden layers closest to the output layer in the model. Modulation is performed, in which This represents the number of neurons in the hidden layer. The input feature dimension is [insert dimension here]. The locally retained region-specific working condition features are [insert dimensions here]. The feature fusion gating network consists of a concatenation layer, a fully connected mapping layer, and a sigmoid activation layer. The concatenation layer stores the hidden layer weights of the global model parameters. Regional specific working condition characteristics A fused tensor is generated by stitching together the data along the channel dimension. Due to the dimension mismatch, the stitching layer first utilizes a linear projection layer to incorporate region-specific features. Mapped to a tensor with the same spatial size as the hidden layer weight matrix. The stitching layer performs a stacking operation along the channel dimension to generate a fused tensor. Its mathematical expression is The fully connected mapping layer performs dimensionality reduction mapping on the fused tensor, transforming the 3D fused tensor... Flatten and input to an output dimension of The fully connected layer is calculated using the following formula: ,in For mapping layer weights, For the bias term, the calculation result This is the tensor after dimensionality reduction mapping. The Sigmoid activation layer transforms the output of the dimensionality reduction mapping into a gating weight coefficient matrix between zero and one, calculated using the following formula: ,in The edge nodes of the region use a gated weight coefficient matrix to perform element-wise weighted modulation of the hidden layer weight matrix of the global model parameters, generating a modulated weight matrix that incorporates the region-specific physical conditions. Its calculation formula is ,in This represents the Hadamard element-wise product. The modulated weight matrix. The corresponding layer weights in the original global model are replaced and input into the calculation engine to perform forward propagation operations, outputting the regional carbon emission calculation results. The element values ​​in the gating weight coefficient matrix determine the contribution ratio of global general parameters and region-specific parameters in the final calculation decision, realizing a dynamic balance between general carbon emission laws and local operating condition deviations in the model parameters.

[0033] In one embodiment, reference Figure 6 In this system, a residual feedback corrector is connected in series after the feature fusion gating network at the regional edge nodes. Although the feature fusion gating network achieves adaptive parameter modulation through nonlinear mapping, the gating weighted modulation process is essentially a data-driven approximation, which can easily introduce systematic static biases under specific operating conditions. The residual feedback corrector obtains the regional carbon emission accounting results output by the regional edge nodes. Actual fuel consumption in the current communication round of industrial production process time-series data Actual fuel consumption includes the absolute cumulative amount from underlying instruments such as the mass of coal consumed and the volume of natural gas consumed. The residual feedback corrector has a built-in static emission factor method calculation module, which calculates the baseline carbon emissions based on actual fuel consumption using the static emission factor method. The calculation of baseline carbon emissions follows the standard accounting guidelines issued by the state or industry, and its general algebraic expression is: ,in For fuel type index, This represents the total number of fuel types. For the first The lower heating value of this fuel, Carbon content or carbon content per unit of calorific value. This represents the carbon oxidation rate. The static emission factor method is based on the principle of material balance and does not rely on complex model reasoning. It possesses absolute physical accuracy at the macroscopic statistical level, and therefore serves as a benchmark reference value.

[0034] The absolute value of the residual between the carbon emission accounting results of the residual feedback corrector calculation area and the baseline carbon emission amount The calculation formula is: superscript This indicates the current communication round. The residual feedback corrector incorporates a proportional-integral (PI) controller, using the absolute value of the residual as the input error signal. The PI controller consists of a proportional element and an integral element connected in parallel. The output of the proportional element is linearly related to the absolute value of the residual at the current moment, while the output of the integral element is the sum of the absolute values ​​of the residual over time. In the discrete-time domain, the feedback bias term generated by the PI controller... The incremental recursive formula is as follows ,in This is the proportional gain coefficient. This refers to the integral gain coefficient. The proportional gain coefficient is also mentioned. The integral gain coefficient is used to respond to sudden transient shifts in the calculation results relative to the baseline value. Used to eliminate steady-state static biases accumulated during long-term model operation. The residual feedback corrector calculates the feedback bias term. The final regional carbon emission accounting result is output by overlaying it onto the regional carbon emission accounting result. Its calculation formula is This residual feedback corrector forms a closed-loop negative feedback control loop at the system end. Utilizing the steady-state error-free characteristic of classical control theory, it forces the calculation results based on deep learning feature fusion output to be anchored near the benchmark carbon emissions based on physical material conservation. This completely suppresses the systematic static deviation introduced by feature space distribution drift or network weight initialization bias during the gated weighted modulation process, ensuring the absolute reliability and compliance of the regional carbon emission accounting system under long-term industrial operation conditions.

Claims

1. A regional carbon emission accounting and management system based on big data analysis, characterized in that, The system comprises a central coordination server and multiple regional edge nodes. Each regional edge node locally stores time-series data of industrial production processes and energy consumption monitoring data for its corresponding region. Each regional edge node has a built-in feature decoupling encoder, which decomposes the industrial production process time-series data and energy consumption monitoring data into region-specific operating condition features and common carbon emission features. The central coordination server is equipped with an adaptive aggregation weight allocation component. This component dynamically adjusts the weight coefficients of each regional edge node in the global model aggregation based on the gradient variance of the common carbon emission feature distribution uploaded by each regional edge node and the local data volume. Each regional edge node uses the decoupled common carbon emission features to perform forward propagation and gradient calculation of its local model, and uploads the gradient update and common carbon emission feature representation to the central coordination server. The central coordination server performs weighted aggregation based on the weight coefficients and then distributes global model parameters. Each regional edge node fuses and splices the updated global model parameters with the locally retained region-specific operating condition features and outputs the regional carbon emission calculation results.

2. The regional carbon emission accounting and management system based on big data analysis according to claim 1, characterized in that, The feature decoupling encoder includes a dual-branch feature extraction network and an orthogonal constraint layer. The dual-branch feature extraction network contains a shared bottom-level convolutional kernel and two independent fully connected layer branches. The industrial production process time-series data and the energy consumption monitoring data are input into the shared bottom-level convolutional kernel to extract basic time-series features. The two independent fully connected layer branches perform nonlinear mapping on the basic time-series features respectively. The orthogonal constraint layer is connected to the output of the two independent fully connected layer branches. The orthogonal constraint layer calculates the Gram matrix between the feature vectors output by the two branches and minimizes the off-diagonal elements, forcing the region-specific operating condition features and the common carbon emission features output by the two independent fully connected layer branches to maintain an orthogonal projection relationship in the feature space.

3. The regional carbon emission accounting and management system based on big data analysis according to claim 1, characterized in that, The adaptive aggregation weight allocation component includes a gradient similarity calculation component and a dynamic weight update component. The gradient similarity calculation component receives the gradient update amount uploaded by each region edge node and calculates the cosine similarity matrix between the gradient update amounts of any two region edge nodes. The dynamic weight update component calculates the deviation index of each region edge node relative to the global average gradient based on the cosine similarity matrix. The dynamic weight update component has a built-in decay factor mapping function. The deviation index is input into the decay factor mapping function and the weight suppression coefficient of the corresponding region edge node is output. When the deviation index of the region edge node exceeds the preset dynamic deviation threshold, the weight suppression coefficient decreases exponentially.

4. The regional carbon emission accounting and management system based on big data analysis according to claim 1, characterized in that, Before uploading the gradient update and the common carbon emission feature representation, the regional edge node performs local differential privacy noise addition processing. The regional edge node has a built-in adaptive gradient pruner and noise injector. The adaptive gradient pruner calculates the norm distribution of the gradient of each layer of the local model in the current communication round and prunes gradient components whose norm exceeds the moving average of the norm in the historical rounds to the moving average of the norm in the historical rounds. The noise injector generates a noise matrix that conforms to a Gaussian distribution based on the dimension of the pruned gradient components and superimposes the noise matrix into the pruned gradient components and the common carbon emission feature representation. The standard deviation of the Gaussian distribution is inversely proportional to the square root of the amount of time-series data of the industrial production process stored locally by the regional edge node.

5. The regional carbon emission accounting and management system based on big data analysis according to claim 1, characterized in that, The regional edge nodes have a built-in feature fusion gating network. The feature fusion gating network receives updated global model parameters from the central coordination server and locally retained regional-specific operating condition features. The feature fusion gating network includes a splicing layer, a fully connected mapping layer, and a sigmoid activation layer. The splicing layer splices the hidden layer weight matrix of the global model parameters and the regional-specific operating condition features in the channel dimension to generate a fusion tensor. The fully connected mapping layer performs dimensionality reduction mapping on the fusion tensor. The sigmoid activation layer converts the dimensionality reduction output into a gating weight coefficient matrix between 0 and 1. The regional edge nodes use the gating weight coefficient matrix to perform element-wise weighted modulation on the hidden layer weight matrix of the global model parameters and then input it into the calculation engine to output the regional carbon emission calculation result.

6. The regional carbon emission accounting and management system based on big data analysis according to claim 1, characterized in that, Before inputting the industrial production process time-series data and the energy consumption monitoring data into the feature decoupling encoder, the edge nodes of the region perform a multi-scale spatiotemporal alignment operation through a data preprocessor. The data preprocessor obtains the operating cycle of the production equipment corresponding to the industrial production process time-series data and the sensor sampling frequency corresponding to the energy consumption monitoring data. Using the operating cycle of the production equipment as the reference time window, the energy consumption monitoring data is downsampled and resampled to align the timestamps of the two. At the same time, the equipment number label and spatial location coordinates in the industrial production process time-series data are extracted, and the spatial location coordinates are mapped to a pre-divided regional grid topology. The processed data is packaged into a three-dimensional structured tensor containing time dimension, spatial dimension and variable dimension and input into the feature decoupling encoder.

7. The regional carbon emission accounting and management system based on big data analysis according to claim 2, characterized in that, The orthogonal constraint layer is also coupled with a physical mechanism discriminator, which is built between the two independent fully connected layer branches. The physical mechanism discriminator receives the region-specific operating condition characteristics and a pre-stored set of regional industrial material balance equations. The physical mechanism discriminator substitutes the material consumption rate parameter in the region-specific operating condition characteristics into the set of regional industrial material balance equations to calculate the material balance difference. The material balance difference is introduced as a penalty term into the Gram matrix loss function of the orthogonal constraint layer. While minimizing the off-diagonal elements of the Gram matrix, the material balance difference is constrained to approach 0, thus forcing the region-specific operating condition characteristics to conform to the regional physical mass conservation law under the condition of satisfying the orthogonal constraints.

8. The regional carbon emission accounting and management system based on big data analysis according to claim 3, characterized in that, The dynamic weight update component is also connected to a historical state buffer queue, which stores the time series of deviation indexes of each regional edge node calculated by the central coordination server in the past preset communication rounds. The dynamic weight update component extracts the deviation index time series to calculate the deviation change rate of each regional edge node. When the deviation index of the regional edge node does not exceed the dynamic deviation threshold, but the deviation change rate shows a monotonically increasing trend and the absolute value of the change rate exceeds the preset change rate surge threshold, the dynamic weight update component triggers a pre-warning mechanism to lock the weight suppression coefficient of the regional edge node to a fixed value of the current round.

9. The regional carbon emission accounting and management system based on big data analysis according to claim 4, characterized in that, A gradient sparsity mask generator is provided between the adaptive gradient clipper and the noise injector. The gradient sparsity mask generator receives the clipped gradient components, arranges them in descending order according to the absolute value of the gradient components, extracts the gradient component indices at the first preset percentage positions to construct a sparse binary mask, and the noise injector performs element-wise multiplication of the Gaussian-distributed noise matrix with the sparse binary mask to filter out redundant noise superimposed at the zero-value gradient index positions. The filtered noise matrix is ​​then superimposed only at the retained gradient component index positions.

10. The regional carbon emission accounting and management system based on big data analysis according to claim 5, characterized in that, The regional edge node is connected in series with a residual feedback corrector after the feature fusion gating network. The residual feedback corrector obtains the regional carbon emission accounting result output by the regional edge node and the actual fuel consumption in the industrial production process time series data of the current communication round. It calculates the absolute value of the residual between the regional carbon emission accounting result and the baseline carbon emission calculated based on the actual fuel consumption using the static emission factor method. The residual feedback corrector has a built-in proportional-integral controller. The absolute value of the residual is input into the proportional-integral controller to generate a feedback bias term. The feedback bias term is superimposed on the regional carbon emission accounting result to output the final regional carbon emission accounting result.

Images