Carbon production energy consumption and carbon emission collaborative prediction method and system

Abstract

The invention provides a carbon production energy consumption and carbon emission collaborative prediction method and system, and the method comprises the steps: respectively adding static or event feature data to short-term time sequence data, long-term time sequence data and spatial data according to the multi-source heterogeneous data of the whole process of carbon production, and correspondingly extracting different feature vectors; splicing the extracted different feature vectors to generate a fused feature vector; performing end-to-end training on the multi-task model by taking the fusion feature vector as an input of the multi-task model and taking a future energy consumption predicted value and a pollutant emission predicted value as an output of the multi-task model until the multi-task model converges and meets a preset multi-constraint condition, and obtaining a trained multi-task model; and inputting the multi-source heterogeneous data of the carbon production whole process in the to-be-predicted time period into the trained multi-task model, and outputting a collaborative prediction result of the energy consumption and the carbon emission, thereby improving the reliability of collaborative prediction of the energy consumption and the carbon emission in the carbon production.

Description

A method and system for synergistic prediction of energy consumption and carbon emissions in carbon production Technical Field

[0001] This invention relates to the field of energy management technology, and in particular to a method and system for coordinating prediction of energy consumption and carbon emissions in carbon production. Background Technology

[0002] As a core component in aluminum electrolysis, the quality of prebaked anodes directly affects the current efficiency and power consumption of the electrolytic aluminum production process. The production process of prebaked anodes is a typical complex industrial system, encompassing multiple physicochemically volatile steps such as petroleum coke calcination, pitch melting, medium crushing and screening, batching and mixing, forming, and roasting. This entire production process is characterized by significant high energy consumption, high emissions, and high coupling. High energy consumption is reflected in the large amounts of natural gas, electricity, and heavy oil consumed during production; high emissions refer to the large amounts of carbon dioxide, sulfur dioxide, nitrogen oxides, and pitch fumes generated during combustion and chemical reactions; high coupling refers to the close cascading relationship between upstream and downstream processes in terms of material flow, heat flow, and flue gas flow. For example, the high-temperature flue gas generated in the roasting furnace is often recycled to heat the heat transfer oil, which in turn heats the mixing pot in the forming process. This energy recycling means that the energy consumption of each process is no longer independent but exhibits a complex network of dependencies.

[0003] Against the backdrop of the national strategy to promote carbon peaking and carbon neutrality, carbon enterprises face immense pressure to conserve energy and reduce carbon emissions. Achieving refined energy management and carbon emission control relies first and foremost on accurate predictions of future energy consumption and emission levels. However, existing prediction technologies face significant technical bottlenecks when applied to carbon production processes.

[0004] First, existing technologies lack sufficient spatiotemporal coupling modeling capabilities. The carbon production process is not merely a time-series evolution of states, but also a spatial topological transfer of energy and matter. Traditional prediction methods are mostly based on single time-series models, such as autoregressive integral moving average models, support vector machines, or standard long short-term memory networks. While these methods can handle short-term fluctuations in individual equipment, they cannot explicitly capture the spatial dependencies between processes through physical connections. For example, when the discharge rate in the calcination process changes, its impact will be transmitted to the molding and roasting processes with a certain lag time. Based solely on single time-series data, the model struggles to understand this causal mechanism based on physical connections, thus limiting prediction accuracy.

[0005] Secondly, multi-scale temporal dynamics are difficult to decouple effectively. Energy consumption fluctuations in carbon production are influenced by multiple factors at different time scales. Equipment-level parameters such as current and pressure exhibit high-frequency fluctuations on the order of seconds or minutes; differences in shift operating habits lead to hourly energy consumption fluctuations; while adjustments to production plans, seasonal changes in ambient temperature, and equipment maintenance cycles bring about long-term trends on the order of months or even quarters. Existing deep learning models often struggle to simultaneously capture high-frequency dynamics and retain low-frequency trends over long periods. Standard Long Short-Term Memory (LSTM) networks suffer from the vanishing gradient problem when processing extremely long sequences, while the computational complexity of the self-attention mechanism in ordinary Transformer models increases quadratically with sequence length, making it difficult to efficiently handle long-term historical data containing thousands of time steps.

[0006] Furthermore, energy consumption forecasting and carbon emission forecasting are often separated. In current industrial practice, energy consumption forecasting and pollutant emission forecasting are usually completed independently by different systems or models. In reality, energy consumption and carbon emissions share a common origin; for example, the combustion of natural gas is both a major source of energy consumption and a major source of carbon dioxide and nitrogen oxide emissions. Forecasting them separately not only increases the consumption of computational resources but also ignores the inherent synergistic information between the indicators, potentially leading to logical contradictions in the forecast results and reducing overall consistency and reliability.

[0007] Finally, existing purely data-driven models lack the ability to recognize physical and real-world constraints. Neural network-based models are essentially probabilistic statistical fitting tools; when faced with unseen boundary conditions or abnormal data inputs, they may output predictions that defy physical common sense or industry standards. While this problem can be mitigated by adding a penalty term to the loss function, this approach often leads to a compromise between prediction accuracy and constraint compliance, making it impossible to avoid counterintuitive predictions. This unreliability makes it difficult for frontline engineers to truly trust and rely on models for production decisions; and it also results in low reliability of the coordinated prediction of energy consumption and carbon emissions in carbon production.

[0008] To address these issues, the present invention provides a method and system for synergistic prediction of energy consumption and carbon emissions in carbon production, thereby solving at least one of the aforementioned technical problems. Summary of the Invention

[0009] In order to solve the problems existing in the prior art, this invention innovatively proposes a method and system for predicting energy consumption and carbon emissions in carbon production, which effectively solves the problem of low reliability of predicting energy consumption and carbon emissions in carbon production due to the prior art, and effectively improves the reliability of predicting energy consumption and carbon emissions in carbon production.

[0010] The first aspect of this invention provides a method for synergistic prediction of energy consumption and carbon emissions in carbon production, comprising: acquiring and preprocessing multi-source heterogeneous data of the entire carbon production process, wherein the multi-source heterogeneous data includes short-term time series data, long-term time series data, spatial data, and static or event feature data; based on the multi-source heterogeneous data of the entire carbon production process, adding static or event feature data to the short-term time series data, long-term time series data, and spatial data respectively, and extracting short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors accordingly; concatenating the short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors to generate a fused feature vector; using the fused feature vector as the input of a multi-task model, and using future energy consumption predictions and pollutant emission predictions as the outputs of the multi-task model, performing end-to-end training on the multi-task model until the multi-task model converges and satisfies preset multi-constraint conditions, thereby obtaining the trained multi-task model; inputting the multi-source heterogeneous data of the entire carbon production process for the period to be predicted into the trained multi-task model, and outputting the synergistic prediction results of energy consumption and carbon emissions.

[0011] The second aspect of this invention provides a carbon production energy consumption and carbon emission collaborative prediction system, comprising: a data acquisition and preprocessing module for acquiring and preprocessing multi-source heterogeneous data of the entire carbon production process, wherein the multi-source heterogeneous data includes short-term time series data, long-term time series data, spatial data, and static or event feature data; a parallel feature extraction module for adding static or event feature data to the short-term time series data, long-term time series data, and spatial data respectively based on the multi-source heterogeneous data of the entire carbon production process, and extracting short-term time feature vectors, long-term time feature vectors, and spatially coupled feature vectors accordingly; and dynamic feature fusion. The module concatenates short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors to generate a fused feature vector. The collaborative regression and constraint optimization module uses the fused feature vector as input to the multi-task model and future energy consumption and pollutant emission predictions as outputs to train the multi-task model end-to-end until it converges and meets preset multi-constraint conditions, resulting in the trained multi-task model. The prediction service interaction module inputs multi-source heterogeneous data of the entire carbon production process for the period to be predicted into the trained multi-task model and outputs the collaborative prediction results of energy consumption and carbon emissions.

[0012] The technical solution adopted in this invention includes the following technical effects: 1. In this invention, based on multi-source heterogeneous data of the entire carbon production process, static or event feature data are added to short-term time series data, long-term time series data, and spatial data respectively, and short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are extracted accordingly; the short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are concatenated to generate a fused feature vector; the fused feature vector is used as the input of a multi-task model, and the predicted energy consumption and pollutant emissions are used as the output of the multi-task model. The multi-task model is trained end-to-end until it converges and meets the preset multi-constraint conditions, thus obtaining the trained multi-task model; the multi-source heterogeneous data of the entire carbon production process during the period to be predicted is input into the trained multi-task model, and the collaborative prediction results of energy consumption and carbon emissions are output, effectively solving the problem of low reliability of collaborative prediction of carbon production energy consumption and carbon emissions caused by existing technologies, and effectively improving the reliability of collaborative prediction of carbon production energy consumption and carbon emissions.

[0013] 2. In the technical solution of this invention, by combining graph neural networks, long short-term memory networks and Informer models, the complex carbon production process is finely deconstructed from three dimensions: spatial topological dependence, short-term operating condition fluctuations and long-term energy consumption rhythms. This improves the modeling ability for complex coupled operating conditions and ensures the reliability of the coordinated prediction of carbon production energy consumption and carbon emissions.

[0014] 3. In the technical solution of this invention, the short-term time feature vector, long-term time feature vector, and spatial coupling feature vector are mapped to a unified dimension. The fused feature vector is the weighted sum of the mapped short-term time feature vector, long-term time feature vector, and spatial coupling feature vector. The dimension of the fused feature vector is the same as the dimension of the mapped short-term time feature vector, long-term time feature vector, and spatial coupling feature vector. By comprehensively considering different types of feature vectors, the final fused feature vector used for prediction is obtained, which ensures the reliability of the coordinated prediction of carbon production energy consumption and carbon emissions.

[0015] 4. The multi-task learning architecture in the technical solution of this invention utilizes the homogeneity and strong correlation of energy consumption and emission data. By sharing the underlying feature extraction network, it realizes knowledge transfer and mutual promotion between different prediction tasks, ensuring the logical consistency of prediction results.

[0016] 5. The loss function in the technical solution of this invention includes a physical mechanism soft constraint loss term, and preset multiple constraint conditions include non-negativity constraints, statistical interval constraints, and material balance constraints. If the multi-task model violates any constraint, it is determined that the current multi-task model has fallen into a pseudo-convergence state, the multi-task model parameters are re-initialized, and a new round of training begins until the multi-task model converges and satisfies all preset constraint conditions, thereby enhancing the industrial cognition and reliability of the multi-task model. The proposed multi-dimensional constraint cognition iterative optimization training strategy adopts a verification-abandonment-remodeling hard constraint mechanism, internalizing physical conservation laws, industry energy efficiency standards, and environmental regulations into the model's constraint conditions, thereby avoiding the model outputting prediction values ​​that seriously violate common sense and improving the model's robustness in the face of unknown working conditions.

[0017] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

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

[0019] Figure 1 is a schematic diagram of the overall process of the method in Embodiment 1 of the technical solution of the present invention; Figure 2 is a schematic diagram of the network architecture for parallel spatiotemporal feature extraction and fusion in the method in Embodiment 1 of the technical solution of the present invention; Figure 3 is a schematic diagram of the topological structure of the carbon production process industrial map constructed in the method in Embodiment 1 of the technical solution of the present invention; Figure 4 is a schematic diagram of the principle of the ProbSparse self-attention mechanism in the Informer model in the method in Embodiment 1 of the technical solution of the present invention; Figure 5 is a schematic diagram of the logical process of the multi-dimensional constraint cognitive iterative optimization training strategy in the method in Embodiment 1 of the technical solution of the present invention; Figure 6 is a schematic diagram of the module structure of the system in Embodiment 2 of the technical solution of the present invention; Figure 7 is a schematic diagram of the hardware structure of the device in Embodiment 3 of the technical solution of the present invention. Detailed Implementation

[0020] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0021] This embodiment provides a method for the coordinated prediction of energy consumption and carbon emissions in carbon production. It can be applied to the production of prebaked anode carbon for aluminum, and also to the coordinated prediction of energy consumption and carbon emissions in other carbon production scenarios. Preferably, before formally starting data processing and model training, a global hyperparameter optimization step can be performed. This step constructs an outer loop optimization framework based on a Bayesian optimization algorithm, using key hyperparameters such as the number of graph neural network layers, the number of long short-term memory network units, the number of Informer multi-head attention heads, and the weights of the multi-task loss function as optimization variables. With the comprehensive fitness function containing constraint violation penalties on the validation set as the objective, it automatically searches for the globally optimal hyperparameter combination, providing the best configuration for subsequent model training.

[0022] The specific steps are shown in Figure 1. This invention provides a method for synergistic prediction of energy consumption and carbon emissions in carbon production, including: S100, acquisition and preprocessing of multi-source heterogeneous data, specifically: acquiring and preprocessing multi-source heterogeneous data of the entire carbon production process, including short-term time series data, long-term time series data, spatial data, and static or event feature data; S200, parallel spatiotemporal feature decoupling extraction, specifically: based on the multi-source heterogeneous data of the entire carbon production process, adding static or event feature data to the short-term time series data, long-term time series data, and spatial data respectively, and extracting short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors accordingly; S300, attention-weighted dynamic... Feature fusion, specifically, involves concatenating short-term time feature vectors, long-term time feature vectors, and spatially coupled feature vectors to generate a fused feature vector; S400, multi-dimensional constraint cognition collaborative regression and iterative optimization training, specifically, using the fused feature vector as input to the multi-task model and future energy consumption and pollutant emission predictions as outputs, performing end-to-end training on the multi-task model until it converges and satisfies preset multi-constraint conditions, resulting in the trained multi-task model; S500, outputting collaborative prediction results, specifically, inputting multi-source heterogeneous data of the entire carbon production process for the period to be predicted into the trained multi-task model, and outputting collaborative prediction results of energy consumption and carbon emissions.

[0023] In step S100, the quality and organization of the data determine the upper limit of the deep learning model's performance. In this embodiment, the data originates from the production site of a large prebaked anode carbon plant, encompassing multiple information silos such as distributed control systems (DCS), manufacturing execution systems (MES), energy management systems (EMS), and laboratory information management systems (LIMS).

[0024] First, multi-source data acquisition is performed. Using edge computing gateways or data acquisition servers deployed at the industrial site, and leveraging industrial communication protocols such as OPCUA and Modbus TCP, the following four types of core data are read in real time: short-term time series data, including real-time operating parameters of equipment in each process of carbon production; long-term time series data, including energy and medium consumption data for the entire plant and each process in carbon production; spatial data, including plant layout maps characterizing the physical connections and energy and material flows of each process in carbon production; and static or event-related data, including production plans, raw material batch information, and equipment maintenance records in the carbon production process. Specifically, the four types of core data are: 1. Short-term time series data (… This section of data primarily reflects the real-time operating status of the equipment, and the sampling frequency is typically high (e.g., 1 minute or 5 minutes). Specifically, it includes: Calcination process: rotary kiln head temperature, kiln tail temperature, kiln body negative pressure, exhaust fan frequency, feeding speed, and instantaneous natural gas flow rate of the burner; Molding process: mixing pot heat transfer oil temperature, mixing current, molding machine hydraulic system pressure, vibration molding time, and paste temperature; Roasting process: flue gas temperature, negative pressure, exhaust rack temperature, and natural gas main pipeline pressure and flow rate of each flue; Auxiliary system: desulfurization tower inlet flue gas temperature, spray liquid flow rate, and waste heat boiler steam pressure and temperature.

[0025] 2. Long-term time series data ( This section of data reflects the overall energy consumption and environmental changes of the entire plant and its various workshops. The sampling frequency is usually low (e.g., 1 hour or 1 day), but the historical span is long (e.g., the past 1-3 years). Specifically, it includes: active power readings from electricity meters for the entire plant and each process, cumulative flow readings from natural gas meters; and local historical meteorological data (ambient temperature, humidity, and air pressure), because ambient temperature directly affects the heat loss of the roasting furnace.

[0026] 3. Spatial data ( Digital analysis of the process piping and instrumentation diagram (PID) and equipment layout diagram of the factory, extracting the physical connection relationships between key equipment (processes), especially the connection topology of material transfer pipelines, high-temperature flue gas pipelines and heat transfer oil circulation pipelines.

[0027] 4. Static or event-related feature data ( ): Obtained from MES and LIMS, including: production schedule for the next 7 days (planned output, product specifications), physical and chemical properties of raw petroleum coke (sulfur content, ash content, volatile matter, true density), softening point of asphalt, and historical maintenance and failure records of equipment.

[0028] Secondly, the acquired data is cleaned, time-series aligned, and normalized.

[0029] Outlier handling: Employing a sliding window-based approach The criteria identify jump noise and outliers in sensor data. For identified outliers, if they are instantaneous impulse noise, they are replaced by linear interpolation of the data in their immediate vicinity; if they are data gaps over a long period, they are marked as invalid segments and not included in training or filled with the historical average of similar operating conditions.

[0030] Multi-scale temporal alignment: Due to inconsistent sampling frequencies from different data sources (from minutes to days), it is necessary to unify them to the model's prediction step size (e.g., 1 hour). For high-frequency... The data is downsampled to the hourly level using statistical aggregation methods. Specific aggregated statistics include the mean (representing the average level), standard deviation (representing the degree of fluctuation), and maximum and minimum values ​​(representing extreme operating conditions). For low-frequency data... (For example, production scheduling), upsampling to the hourly level is performed using forward filling or interpolation methods. Normalization: To eliminate differences in physical dimensions (such as temperature) With current To investigate the impact of gradient descent on neural networks, Z-score standardization was applied to all continuous variables, i.e., subtracting the mean and dividing by the standard deviation to approximate a standard normal distribution. For discrete variables (such as shift ID and raw material batch ID), One-Hot encoding or Entity Embedding was used to convert them into low-dimensional dense vectors.

[0031] Finally, a standardized multidimensional model input tensor is constructed, including a sequence tensor reflecting the time dimension and graph structure data reflecting the spatial dimension.

[0032] Preferably, before executing step S100, a global hyperparameter optimization step is also included: constructing an outer loop optimization framework based on the Bayesian optimization algorithm, taking the number of graph neural network layers, the number of long short-term memory network units, the number of Informer multi-head attention heads, and the weights of the multi-task loss function as optimization variables, and taking the comprehensive fitness function containing constraint violation penalty terms on the validation set as the objective, and automatically searching for the globally optimal hyperparameter combination.

[0033] In step S200, as shown in Figure 2, considering the complex spatiotemporal characteristics of the carbon production process, this invention designs three parallel neural network branches to extract spatial coupling features, short-term dynamic features (short-term time features), and long-term rhythmic features (long-term time features), respectively. Specifically, this includes: S210: Extracting spatial coupling feature vectors based on the spatial coupling features of the industrial graph, representing the energy and mass transfer relationships between processes throughout the plant: Carbon production is essentially a process of material and energy flow and transformation within a network of equipment. This step aims to utilize graph neural networks to enable the model to understand the physical mechanisms of this flow.

[0034] Industrial Mapping: Definition Map .

[0035] As shown in Figure 3, the node set :definition One key process unit is used as a node, for example:

[0036] Each node With initial eigenvectors The vector consists of static attributes (static or event-specific data) such as the rated power and designed capacity of the equipment, as well as the aggregated operating condition status at the current moment.

[0037] edge set Define a directed edge Indicates from node arrive The physical flow of edges. Edge types include: material flow (e.g., calcined coke flowing from a rotary kiln to a medium crusher), heat flow (e.g., high-temperature flue gas flowing from a roasting furnace to a waste heat boiler), and medium flow (e.g., thermal oil flowing from a waste heat boiler to a mixing pot). Edges have attribute weights. The initial value can be set to the magnitude of the flow rate or the reciprocal of the length of the connected pipe.

[0038] Graph Attention Network (GAT) Propagation and Feature Extraction: The constructed graph is processed using a multi-layer Graph Attention Network (GAT). The core advantage of GAT lies in its ability to dynamically assign different weights to different neighbor nodes, thereby simulating the changes in the degree of influence of upstream equipment on downstream equipment under different operating conditions.

[0039] For any node m, its neighbors are... Attention coefficients between (neighbor set or adjacent set of nodes) The calculation process is as follows: First, calculate the raw attention score. :

[0040] in, It is a shared linear transformation weight matrix used to increase feature dimensionality; It is the weight vector of a single-layer feedforward neural network; This represents a vector concatenation operation; LeakyReLU is a non-linear activation function. The physical meaning of this formula is to calculate the correlation between the features of node m and node n. Let m be the feature vector of node m. Let n be the feature vectors of the neighboring nodes. Then, the attention score is normalized using the Softmax function:

[0041] The weights of the state fluctuations of upstream process n transferred to downstream process m are dynamically quantified. For example, when the flue gas temperature of the calcining rotary kiln fluctuates drastically, during model training, the attention coefficient corresponding to the waste heat boiler node is... The automatic increase indicates that the state of the rotary kiln is the dominant factor affecting boiler energy consumption at this time.

[0042] Then, the attention coefficient is used to weight and aggregate the neighbor features to update the node features:

[0043] in, The features of the updated node m. This is the activation function.

[0044] go through After layer GAT propagation, each node in the graph incorporates information from its multi-hop neighbors.

[0045] Finally, the features of all nodes are aggregated using global average pooling or the Readout function to obtain a feature vector representing the spatial coupling state of the entire plant. .

[0046] S220: Short-term dynamic feature extraction based on operating condition sequence: Short-term time feature vectors representing equipment-level operation cycle and high-frequency operating condition fluctuations: For high-frequency fluctuations caused by equipment operation (such as periodic hydraulic loading of molding machines and PID adjustment of natural gas valves), a long short-term memory network (LSTM) is adopted.

[0047] 1. Input: Select past Operating condition characteristic sequence at each time step (e.g., the past 24 hours) .

[0048] 2. Network Structure: A bidirectional LSTM (Bi-LSTM) network is employed. Compared to a unidirectional LSTM network, the Bi-LSTM network contains a hidden layer that processes the sequence forward and a hidden layer that processes the sequence backward, enabling it to utilize both past and future contextual information (during training). Through the ingenious design of forget gates, input gates, and output gates, the LSTM units effectively address the vanishing gradient problem of traditional RNNs, thereby capturing short-term dependencies and perturbation propagation patterns in time series data.

[0049] 3. Output: The forward and backward hidden states of the last layer of the Bi-LSTM at the last time step are concatenated and mapped through a fully connected layer to obtain the short-term time feature vector. This vector highly summarizes the recent operating conditions and trends of the equipment.

[0050] S230: Extracting long-term time feature vectors based on the long-term dependence characteristics of energy consumption rhythm (long-term time features, i.e., long-term time series data and production plan sequences) to extract long-term time feature vectors that characterize the long-term energy consumption trend and rhythm under the influence of production plan adjustments, seasonal temperature changes and equipment maintenance cycles: In response to the impact of long-term factors such as production plan adjustments, seasonal climate changes and monthly maintenance cycles on energy consumption and carbon emissions, this invention adopts the advanced Informer model.

[0051] 1. Input: Select past Historical energy consumption sequences, meteorological data, and production plan sequences (static or event-specific data) for each time step (e.g., the past 30 days, 720 hours).

[0052] 2. ProbSparse Self-Attention Mechanism: The traditional Transformer model's self-attention mechanism requires calculating the dot product of all query vectors and key vectors, with a computational complexity of O(n log n). When the sequence length When the size is large, the computational resource consumption becomes unacceptable. The Informer model is based on the query sparsity assumption, which states that in long sequences, only a few key query vectors (such as the moment representing the point of change in operating conditions) contribute significantly to the attention computation, while most other queries exhibit uniformly distributed noise characteristics.

[0053] As shown in Figure 4, this embodiment uses KL divergence (KL divergence) to approximate the sparsity measure of each query vector. For a given query matrix... Its sparsity measure The larger the value, the more important the query. The Informer model only selects the top [values]. indivual( The active query vector with the highest sparsity metric is multiplied by all key vectors, and the attention values ​​of the remaining query vectors are directly set to the mean.

[0054] The parameter *c* is the Sparse Sampling Factor. It is a constant set by engineers before training the Informer model to control the density of key query vectors selected in the ProbSparse self-attention mechanism. It is a hyperparameter greater than 0 (typically between 3 and 10, depending on computational resource limitations and accuracy requirements). This parameter determines the density of key time steps that the Informer model focuses on when processing long sequences. In this embodiment, by setting *c*, the system can adaptively filter out active queries containing high information content from massive historical data, ignoring a large amount of redundant information in a stationary state, thereby achieving efficient extraction of long-period features.

[0055] In the production scenario of prebaked anodes for aluminum, the input is historical data spanning 30 days (720 hours).

[0056] Let's assume we set c=5. According to the formula: u=5×ln(720)≈5×6.58≈33. Physical interpretation: This means that in the past 720 hours of data, the algorithm doesn't need to calculate the pairwise relationships between all 720 moments, but automatically selects approximately 33 of the most representative key moments (e.g., raw material switching moments, equipment failure shutdown moments, sudden drops in ambient temperature, etc.). The Informer model focuses on calculating the correlation between these 33 key moments and the entire historical cycle, thus accurately capturing the key rhythmic characteristics affecting energy consumption and carbon emissions while ensuring extremely low computational latency.

[0057] This mechanism significantly reduces computational complexity to This enables the Informer model to efficiently process historical data up to one month in length, thereby capturing seasonal (such as increased natural gas consumption in winter) and periodic (such as weekly shutdowns for maintenance) energy consumption rhythms.

[0058] 3. Self-Attention Distillation: One-dimensional convolutional layers (Conv1D) and max pooling layers are introduced between the encoder layers of the Informer model, halving the temporal dimension of the feature sequence layer by layer. This operation is similar to a pyramid structure, enabling the extraction of more macroscopic trend features from micro-time steps.

[0059] 4. Output: The output of the Informer encoder is processed by a fully connected layer to obtain the long-term temporal feature vector. .

[0060] In the long-term time feature extraction based on energy consumption, the Informer model and its probabilistic sparse self-attention mechanism are introduced, which breaks through the computational efficiency bottleneck of traditional deep learning models when processing long-term data, enabling the model to effectively utilize longer historical information to assist in prediction.

[0061] In step S300, the operating conditions at the carbon production site are dynamically changing. For example, during stable equipment operation, historical periodic patterns (extracted by Informer) may exist. ) is most effective for prediction; however, in the event of sudden events such as raw material switching or equipment failure, real-time spectral correlation (extracted by GNN) is more effective. ) and short-term fluctuations (extracted by LSTM) This is even more crucial. To enable the model to possess this ability to assess situations, this embodiment designs a dynamic feature fusion module. Step S300 specifically includes: S310, mapping the short-term time feature vector, long-term time feature vector, and spatial coupling feature vector to a unified dimension; specifically, before feature concatenation, the feature vectors output by the three branches are first mapped to the same feature dimension through a fully connected layer. Because , , The original dimensions may be different (denoted as ). The system first utilizes three independent fully connected layers, mapping them to a unified hidden layer dimension. .

[0062]

[0063]

[0064]

[0065] At this point, the three eigenvectors have completely identical shapes, all being... .

[0066] in, This is the mapped spatial coupling feature vector. It is the mapping matrix of spatially coupled eigenvectors; This is the mapped short-term feature vector. The mapping matrix for short-term eigenvectors; This is the mapped long-term feature vector. It is the mapping matrix of long-term time eigenvectors.

[0067] S320, the fused feature vector is the weighted sum of the mapped short-term time feature vector, long-term time feature vector, and spatial coupling feature vector, and the dimension of the fused feature vector is the same as the dimension of the mapped short-term time feature vector, long-term time feature vector, and spatial coupling feature vector.

[0068] Specifically, the mapped feature vectors are concatenated and input into the attention module to calculate weights; finally, these weights are used to perform a weighted sum of the mapped feature vectors to obtain the final fused feature vector. To calculate the attention weights, the three aligned vectors are concatenated along the channel dimension to obtain the joint feature vector. .

[0069]

[0070] Joint eigenvectors The number of channels is the sum of the number of channels in the three feature vector branches, that is... .this The inputs are used solely as decision-making criteria in a multilayer perceptron (MLP) to calculate three scalar weights reflecting their respective importance. , , The spliced ​​H concat It is not directly used as the final feature output. Its role is to serve as input to the attention mechanism of the Multilayer Perceptron (MLP), using the MLP's output as the three scalar weights. , , .

[0071] The final fused feature vector is calculated using the calculated weights. The specific expression for the element-wise weighted summation of the aligned feature vector is as follows:

[0072] The final fused feature vector The dimension remains as (That is, consistent with the single branch dimension after alignment, rather than the sum of the three). The weights of the spatially coupled feature vectors. The weights of the short-term eigenvectors. These are the weights of the long-term eigenvectors.

[0073] In step S400, this step is the core innovation of the present invention, which aims to solve the problem that a purely data-driven model may output results that violate physical common sense (such as negative energy consumption or emission values ​​that violate the law of conservation of mass).

[0074] Multi-task regression network construction: fusing feature vectors The input is fed into an improved BPNN regressor (multi-task model). This regressor employs a multi-task learning (MTL) architecture, which includes shared layers and task-specific towers.

[0075] Energy Prediction Tower: Outputting the Future Total energy consumption (natural gas, electricity) forecast for each time step .

[0076] Carbon emission prediction tower: outputs the corresponding predicted CO2 emission values. .

[0077] Pollutant Prediction Tower: Outputs predicted SO2 and NOx emissions. Since energy consumption and carbon emissions share a common physical origin (primarily from fuel combustion), sharing a hidden layer allows the model to learn this common characteristic, thereby achieving collaborative prediction.

[0078] In the calcination and production process of prebaked anodes for aluminum, to balance data fitting accuracy and physicochemical logic, the expression for the total loss function of the multi-task model is as follows: ;in, For the total loss function, The basic fitting loss term (mean squared error, used to ensure the basic fitting accuracy) is used. The robustness loss term (Huber loss, which uses linear penalty instead of quadratic penalty for outliers (abnormal operating condition data) to improve the robustness of the model). This is the covariance consistency loss term (covariance regularization term, which forces the covariance matrix structure of the model predictions to converge with the covariance matrix of the training data, ensuring that the correlation between energy consumption and carbon emissions is not lost). This refers to the loss term based on physical mechanism soft constraints (soft physical constraint term). For example, penalty. The situation, or punishment (situation where the calorific value deviates significantly from the theoretical calorific value) The hyperparameters for the weights of the robustness loss term. The hyperparameter for the weights of the covariance consistency loss term. The hyperparameters are the weights of the soft constraint loss term based on the physical mechanism (in this embodiment, the values ​​are 0.5, 0.1, and 0.2, respectively).

[0079] Specifically, the basic fit loss The specific calculation method is as follows:

[0080] in, 2 represents the number of training batches. For the first The actual energy consumption values ​​of each sample (from EMS flow meters). For the first The predicted energy consumption value for each sample. For the first The true value (calculated or monitored value) of carbon emissions for each sample. For the first Predicted carbon emissions for each sample; basic fit loss This is the most fundamental constraint, requiring that the values ​​predicted by the model be as close as possible to the actual instrument values ​​read by the DCS and EMS systems. It ensures the accuracy of the model under normal operating conditions.

[0081] Robustness loss The calculation expression is:

[0082] in, It is a piecewise function. For the first The true value of energy consumption or carbon emissions for each sample. For the first The predicted energy consumption or carbon emissions for each sample, where the piecewise function... The definition is as follows:

[0083] To predict the residuals (i.e., the error). This represents the tolerance threshold for abnormal operating conditions. In this embodiment, It was set to three times the standard noise level of the sensor.

[0084] In carbon production sites, sensor signals are often accompanied by transient noise (such as electromagnetic interference caused by the start-up of high-power fans) or non-steady operating conditions (such as instantaneous pressure fluctuations when the furnace is switched on).

[0085] When the error is very small ( When the error is extremely large (), it is considered a normal fluctuation, and the squared error is used to accelerate convergence; when the error is extremely large (), it is considered a normal fluctuation, and the squared error is used to accelerate convergence. When a sensor malfunction occurs, the system identifies it as an "outlier" and uses linear error to prevent the multi-task model from sacrificing the overall generalization law in order to force a fit to a particular "sensor malfunction" or "abnormal jump".

[0086] Covariance Consistency Loss The calculation expression is:

[0087] in, This is the covariance matrix of the predicted energy consumption vector and carbon emission vector within the current training batch. This is the covariance matrix of the actual energy consumption vector and the actual carbon emission vector within the current training batch. Let Frobenius norm be used to measure the distance between two matrices; physically, energy consumption (natural gas combustion) and carbon emissions are highly positively correlated coupled variables. This loss function forces the multi-task model to learn correlations between variables that are consistent with the real world. That is, the multi-task model cannot predict a significant decrease in energy consumption while predicting that carbon emissions remain unchanged, which violates physical logic. It locks in the linkage between energy consumption and carbon emissions.

[0088] Physical mechanism soft constraint loss The calculation expression is:

[0089] ReLU is a linear rectified function (which produces a large penalty gradient when the predicted value is negative; it is 0 when the predicted value is positive). The theoretical combustion emission factor (for natural gas (methane CH4), per 1 The generated The quantity has a theoretical baseline value in chemistry (approximately 1.9 - 2.2 kg / L). (Depending on purity) : To prevent the minimum value of division by zero, These are the weighting coefficients for stoichiometric constraints.

[0090] In the physical world, the energy consumption and emissions of a roasting furnace cannot be negative. This ensures that the multi-task model does not produce negative values ​​that defy common sense. Stoichiometry constraints embed chemical reaction equations into the neural network. Although actual emission coefficients may fluctuate in industrial settings due to combustion efficiency issues, they cannot deviate too far from theoretical values. Even in the absence of label data, this constraint guides the model to output predictions that conform to chemical principles, preventing the generation of "pseudo-scientific" predictions.

[0091] As shown in Figure 5, the training process of the multi-task model adopts a multi-dimensional constraint cognitive iterative optimization training strategy: although the loss function includes soft constraint terms, the gradient descent method may still converge to a local minimum, leading to failure under extreme conditions. Therefore, this embodiment introduces an outer hard constraint verification and reset mechanism.

[0092] Phase 1: Main Training Loop. Minimize the training set using the AdamW optimizer. And monitor performance on the validation set.

[0093] Phase Two: Constraint Testing and Cognitive Verification. When the multi-task model reaches its optimum on the validation set (triggered by EarlyStopping), the multi-task model is not saved immediately. Instead, the weights are loaded, and inference is performed using a pre-set constraint test set (including non-negativity constraints, statistical interval constraints, and material balance constraints).

[0094] Phase 3: Hard Constraint Judgment. Check whether the inference results meet the following hard conditions: C1 Non-negativity Constraint: All energy consumption and emission predictions must... .

[0095] C2 Statistical Interval Constraint: The energy consumption per unit of product (unit consumption) must be within the historical statistical range. Within the range or within the physical limits specified by industry standards.

[0096] C3 Material Balance Constraints: Taking SO2 prediction as an example, the sulfur balance formula must be satisfied. The total sulfur content of the input system (sulfur content of petroleum coke + sulfur content of asphalt) should be approximately equal to the total sulfur content of the output system (residual sulfur in the finished anode + SO2 emission from flue gas + collection amount of desulfurized gypsum). If the predicted SO2 emission causes the material balance error to exceed the tolerance threshold (preset difference threshold, such as 5%), it is considered a violation.

[0097] Phase Four: Decision Making and Reset.

[0098] If all constraints are met: the multi-task model is deemed cognitively competent, the weights are saved, and training ends.

[0099] If any constraint is violated, the current multi-task model is determined to have fallen into a spurious convergence trap (i.e., the data fits well but the physical logic is flawed). In this case, instead of fine-tuning, all current weights are forcibly discarded, the random seed is changed, the model parameters are reinitialized, and training is restarted from stage one. This strategy is similar to Monte Carlo search, forcing the model to find a solution region in the parameter space that satisfies both the data distribution and the physical laws.

[0100] In step S500, the collaborative prediction results and applications are output: the multi-task model, which has been trained and passed constraint verification, is deployed on the inference server. The system receives real-time DCS data, MES production scheduling plans, and LIMS test data, and inputs them into the model for forward inference. The output results include: hourly energy consumption trend curves for the next 24 hours, total carbon emission predictions, and key pollutant concentration predictions (i.e., outputting hourly total energy consumption, energy consumption of each component, and predicted emissions of carbon dioxide, sulfur dioxide, and nitrogen oxides within a preset future time period). Applications of the results: 1. Energy dispatch: If a surge in natural gas consumption is predicted for a certain period in the future, the system automatically sends dispatch suggestions to the energy center to adjust pipeline pressure in advance.

[0101] 2. Environmental compliance: If SO2 emissions are predicted to exceed the standard, the system will suggest in advance to increase the flow rate of the desulfurization tower spray liquid or reduce the proportion of high-sulfur petroleum coke.

[0102] 3. Anomaly Diagnosis: If the deviation between the real-time observed value and the predicted value continues to widen, it indicates that a hidden equipment fault may have occurred (such as the furnace body insulation layer falling off, leading to increased heat dissipation), triggering an alarm.

[0103] In this invention, based on multi-source heterogeneous data from the entire carbon production process, static or event-related feature data are added to short-term time series data, long-term time series data, and spatial data, respectively, and corresponding short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are extracted. These vectors are then concatenated to generate a fused feature vector. This fused feature vector serves as the input to a multi-task model, while future energy consumption and pollutant emission predictions are used as the output. The multi-task model is trained end-to-end until it converges and satisfies preset multi-constraint conditions, resulting in the trained model. The multi-source heterogeneous data from the entire carbon production process for the period to be predicted is input into the trained multi-task model, outputting the collaborative prediction results of energy consumption and carbon emissions. This effectively solves the problem of low reliability in the collaborative prediction of carbon production energy consumption and carbon emissions caused by existing technologies, and significantly improves the reliability of the collaborative prediction of carbon production energy consumption and carbon emissions.

[0104] The technical solution of this invention combines graph neural networks, long short-term memory networks and Informer models to finely deconstruct the complex carbon production process from three dimensions: spatial topological dependence, short-term operating condition fluctuations and long-term energy consumption rhythms. This improves the modeling ability for complex coupled operating conditions and ensures the reliability of the coordinated prediction of carbon production energy consumption and carbon emissions.

[0105] In this invention, short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are mapped to a unified dimension. The fused feature vector is the weighted sum of the mapped short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors, and the dimension of the fused feature vector is the same as that of the mapped short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors. By comprehensively considering different types of feature vectors, the final fused feature vector used for prediction is obtained, ensuring the reliability of the coordinated prediction of carbon production energy consumption and carbon emissions.

[0106] The multi-task learning architecture in this invention utilizes the homogeneity and strong correlation between energy consumption and emission data. By sharing the underlying feature extraction network, it achieves knowledge transfer and mutual promotion between different prediction tasks, ensuring the logical consistency of the prediction results.

[0107] The loss function in this invention includes a physical mechanism soft constraint loss term, and preset multiple constraints include non-negativity constraints, statistical interval constraints, and material balance constraints. If the multi-task model violates any constraint, it is determined that the current multi-task model has fallen into a pseudo-convergence state. The parameters of the multi-task model are re-initialized and a new round of training begins until the multi-task model converges and satisfies all preset constraints, thereby enhancing the industrial cognition and reliability of the multi-task model. The proposed multi-dimensional constraint cognition iterative optimization training strategy adopts a verification-abandonment-remodeling hard constraint mechanism, internalizing physical conservation laws, industry energy efficiency standards, and environmental regulations as constraints of the model, thereby avoiding the model outputting prediction values ​​that seriously violate common sense and improving the robustness of the model when facing unknown working conditions.

[0108] As shown in Figure 6, the second embodiment of the present invention also provides a carbon production energy consumption and carbon emission collaborative prediction system, including: a data acquisition and preprocessing module, which acquires and preprocesses multi-source heterogeneous data of the entire carbon production process, wherein the multi-source heterogeneous data includes short-term time series data, long-term time series data, spatial data, and static or event feature data; a parallel feature extraction module, which, based on the multi-source heterogeneous data of the entire carbon production process, adds static or event feature data to the short-term time series data, long-term time series data, and spatial data respectively, and extracts short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors accordingly; dynamic... The state feature fusion module concatenates short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors to generate a fused feature vector. The collaborative regression and constraint optimization module uses the fused feature vector as input to the multi-task model and future energy consumption and pollutant emission predictions as outputs to train the multi-task model end-to-end until it converges and meets preset multi-constraint conditions, resulting in the trained multi-task model. The prediction service interaction module inputs multi-source heterogeneous data of the entire carbon production process for the period to be predicted into the trained multi-task model and outputs the collaborative prediction results of energy consumption and carbon emissions.

[0109] The data acquisition and processing module is responsible for connecting various information systems in the factory, performing data cleaning, alignment and normalization in step S100, and has a built-in real-time database for data storage.

[0110] The parallel feature extraction module comprises three independent units. The industrial map processing unit runs a graph neural network algorithm to process spatial data, obtaining spatially coupled feature vectors. The short-term sequence processing unit deploys a bidirectional long short-term memory network to process high-frequency sensor data streams (short-term time series data), obtaining short-term time feature vectors. The long-term sequence processing unit deploys an Informer model to analyze long-period historical data (long-term time series data), obtaining long-term time feature vectors.

[0111] The dynamic feature fusion module implements the attention mechanism in step S300, dynamically adjusting the weights of each feature channel according to real-time operating conditions.

[0112] The collaborative regression and constraint optimization module is the core of the system. It includes a model training engine for performing multi-task regression and backpropagation in step S400, and a constraint verification engine with an embedded physical formula library and industry standard database, responsible for performing hard constraint verification logic before the model is released.

[0113] The predictive service interaction module provides an application programming interface (API) to push predictive results to downstream energy management systems or advanced control systems, and provides a visualization interface to display predictive curves, feature importance analysis, and abnormal alarm information.

[0114] In this invention, based on multi-source heterogeneous data from the entire carbon production process, static or event-related feature data are added to short-term time series data, long-term time series data, and spatial data, respectively, and corresponding short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are extracted. These vectors are then concatenated to generate a fused feature vector. This fused feature vector serves as the input to a multi-task model, while future energy consumption and pollutant emission predictions are used as the output. The multi-task model is trained end-to-end until it converges and satisfies preset multi-constraint conditions, resulting in the trained model. The multi-source heterogeneous data from the entire carbon production process for the period to be predicted is input into the trained multi-task model, outputting the collaborative prediction results of energy consumption and carbon emissions. This effectively solves the problem of low reliability in the collaborative prediction of carbon production energy consumption and carbon emissions caused by existing technologies, and significantly improves the reliability of the collaborative prediction of carbon production energy consumption and carbon emissions.

[0115] The technical solution of this invention combines graph neural networks, long short-term memory networks and Informer models to finely deconstruct the complex carbon production process from three dimensions: spatial topological dependence, short-term operating condition fluctuations and long-term energy consumption rhythms. This improves the modeling ability for complex coupled operating conditions and ensures the reliability of the coordinated prediction of carbon production energy consumption and carbon emissions.

[0116] In this invention, short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are mapped to a unified dimension. The fused feature vector is the weighted sum of the mapped short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors, and the dimension of the fused feature vector is the same as that of the mapped short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors. By comprehensively considering different types of feature vectors, the final fused feature vector used for prediction is obtained, ensuring the reliability of the coordinated prediction of carbon production energy consumption and carbon emissions.

[0117] The multi-task learning architecture in this invention utilizes the homogeneity and strong correlation between energy consumption and emission data. By sharing the underlying feature extraction network, it achieves knowledge transfer and mutual promotion between different prediction tasks, ensuring the logical consistency of the prediction results.

[0118] The loss function in this invention includes a physical mechanism soft constraint loss term, and preset multiple constraints include non-negativity constraints, statistical interval constraints, and material balance constraints. If the multi-task model violates any constraint, it is determined that the current multi-task model has fallen into a pseudo-convergence state. The parameters of the multi-task model are re-initialized and a new round of training begins until the multi-task model converges and satisfies all preset constraints, thereby enhancing the industrial cognition and reliability of the multi-task model. The proposed multi-dimensional constraint cognition iterative optimization training strategy adopts a verification-abandonment-remodeling hard constraint mechanism, internalizing physical conservation laws, industry energy efficiency standards, and environmental regulations as constraints of the model, thereby avoiding the model outputting prediction values ​​that seriously violate common sense and improving the robustness of the model when facing unknown working conditions.

[0119] As shown in Figure 7, Example 3 provides a computer device for implementing the above method. This device can be an industrial-grade server or an edge computing workstation. Its hardware architecture includes a processor, a deep learning accelerator card, memory, a storage device, and a network interface. The storage device contains a computer program based on a deep learning framework and a graph neural network library. When the processor executes the computer program, it sequentially performs data preprocessing, parallel feature extraction, dynamic fusion, constraint verification, and prediction output according to the logical order described in Example 1, thereby achieving coordinated prediction of energy consumption and carbon emissions in the production process of prebaked anodes for aluminum.

[0120] In this invention, based on multi-source heterogeneous data from the entire carbon production process, static or event-related feature data are added to short-term time series data, long-term time series data, and spatial data, respectively, and corresponding short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are extracted. These vectors are then concatenated to generate a fused feature vector. This fused feature vector serves as the input to a multi-task model, while future energy consumption and pollutant emission predictions are used as the output. The multi-task model is trained end-to-end until it converges and satisfies preset multi-constraint conditions, resulting in the trained model. The multi-source heterogeneous data from the entire carbon production process for the period to be predicted is input into the trained multi-task model, outputting the collaborative prediction results of energy consumption and carbon emissions. This effectively solves the problem of low reliability in the collaborative prediction of carbon production energy consumption and carbon emissions caused by existing technologies, and significantly improves the reliability of the collaborative prediction of carbon production energy consumption and carbon emissions.

[0121] The technical solution of this invention combines graph neural networks, long short-term memory networks and Informer models to finely deconstruct the complex carbon production process from three dimensions: spatial topological dependence, short-term operating condition fluctuations and long-term energy consumption rhythms. This improves the modeling ability for complex coupled operating conditions and ensures the reliability of the coordinated prediction of carbon production energy consumption and carbon emissions.

[0122] In this invention, short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are mapped to a unified dimension. The fused feature vector is the weighted sum of the mapped short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors, and the dimension of the fused feature vector is the same as that of the mapped short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors. By comprehensively considering different types of feature vectors, the final fused feature vector used for prediction is obtained, ensuring the reliability of the coordinated prediction of carbon production energy consumption and carbon emissions.

[0123] The multi-task learning architecture in this invention utilizes the homogeneity and strong correlation between energy consumption and emission data. By sharing the underlying feature extraction network, it achieves knowledge transfer and mutual promotion between different prediction tasks, ensuring the logical consistency of the prediction results.

[0124] The loss function in this invention includes a physical mechanism soft constraint loss term, and preset multiple constraints include non-negativity constraints, statistical interval constraints, and material balance constraints. If the multi-task model violates any constraint, it is determined that the current multi-task model has fallen into a pseudo-convergence state. The parameters of the multi-task model are re-initialized and a new round of training begins until the multi-task model converges and satisfies all preset constraints, thereby enhancing the industrial cognition and reliability of the multi-task model. The proposed multi-dimensional constraint cognition iterative optimization training strategy adopts a verification-abandonment-remodeling hard constraint mechanism, internalizing physical conservation laws, industry energy efficiency standards, and environmental regulations as constraints of the model, thereby avoiding the model outputting prediction values ​​that seriously violate common sense and improving the robustness of the model when facing unknown working conditions.

[0125] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for synergistic prediction of energy consumption and carbon emissions in carbon production, characterized in that, include: Acquire and preprocess multi-source heterogeneous data of the entire carbon production process, including short-term time series data, long-term time series data, spatial data, and static or event feature data; Based on multi-source heterogeneous data from the entire carbon production process, static or event-related feature data are added to short-term time series data, long-term time series data, and spatial data, respectively, and corresponding short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are extracted. The short-term time feature vector, long-term time feature vector, and spatial coupling feature vector are concatenated to generate a fused feature vector. The fused feature vector is used as the input to a multi-task model, and the predicted energy consumption and pollutant emissions are used as the output to train the multi-task model end-to-end until the multi-task model converges and meets the preset multi-constraint conditions, resulting in the trained multi-task model. The multi-source heterogeneous data of the entire carbon production process for the period to be predicted is input into the trained multi-task model, and the collaborative prediction results of energy consumption and carbon emissions are output.

2. The method for synergistic prediction of energy consumption and carbon emissions in carbon production according to claim 1, characterized in that, Short-term time series data includes real-time operating parameters of equipment in each process of carbon production; long-term time series data includes energy and medium consumption data of the whole plant and each process in carbon production; spatial data includes plant layout maps that characterize the physical connection and energy and material flow of each process in carbon production; and static or event-related data includes production plans, raw and auxiliary material batch information, and equipment maintenance records in carbon production.

3. The method for synergistic prediction of energy consumption and carbon emissions in carbon production according to claim 2, characterized in that, Based on multi-source heterogeneous data from the entire carbon production process, static or event-related feature data are added to short-term time series data, long-term time series data, and spatial data, respectively. Correspondingly, short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors are extracted. Specifically, this includes: using a graph neural network to construct each process unit of carbon production as graph nodes, and constructing directed edges for material flow, heat flow, and flue gas flow between each process unit; learning and extracting spatial coupling feature vectors representing the energy and mass transfer relationships between processes throughout the plant through a graph attention network mechanism; short-term time feature extraction based on operating condition sequences, using a long short-term memory network to process short-term time series data and extract short-term time feature vectors representing equipment-level operating cycles and high-frequency operating condition fluctuations; and long-term time feature extraction based on energy consumption, using an improved Informer model to process long-term time series data and production plan sequences to extract long-term time feature vectors representing the long-term energy consumption trends and rhythms under the influence of production plan adjustments, seasonal temperature changes, and equipment maintenance cycles.

4. The method for synergistic prediction of energy consumption and carbon emissions in carbon production according to claim 3, characterized in that, Using a graph neural network, each process unit in carbon production is constructed as a graph node, and the material flow, heat flow, and flue gas flow between each process unit are constructed as directed edges. Through a graph attention network mechanism, the spatial coupling feature vector representing the energy and mass transfer relationship between processes in the whole plant is learned and extracted. Specifically, an industrial graph is constructed, process units are defined as nodes, and each node includes an initial feature vector, which includes the rated power, design capacity, and current aggregate operating conditions of the equipment in the node. A directed edge is defined to represent the physical flow between adjacent nodes, including material flow, heat flow, and medium flow. Different weights are assigned to different adjacent nodes, and the initial attention score of the directed edge is calculated and normalized. The features of adjacent nodes are weighted and aggregated using the attention coefficient to update the node features. The features of all nodes after aggregation and updating are used to obtain a feature vector representing the spatial coupling state of the entire plant.

5. The method for synergistic prediction of energy consumption and carbon emissions in carbon production according to claim 1, characterized in that, The process of concatenating short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors to generate a fused feature vector involves: mapping the short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors to a unified dimension; the fused feature vector is the weighted sum of the mapped short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors, and the dimension of the fused feature vector is the same as the dimension of the mapped short-term time feature vectors, long-term time feature vectors, and spatial coupling feature vectors.

6. The method for synergistic prediction of energy consumption and carbon emissions in carbon production according to claim 1, characterized in that, The expression for the total loss function of the multi-task model is: ;in, For the total loss function, Based on the fitting loss term, For robustness loss, This is the covariance consistency loss term. For the physical mechanism soft constraint loss term, The hyperparameters for the weights of the robustness loss term. The hyperparameter for the weights of the covariance consistency loss term. This is a hyperparameter for the weights of the soft constraint loss term based on the physical mechanism.

7. The method for synergistic prediction of energy consumption and carbon emissions in carbon production according to claim 6, characterized in that, Basic Fit Loss The specific calculation method is as follows: in, 2 represents the number of training batches. For the first The true energy consumption value of each sample. For the first The predicted energy consumption value for each sample. For the first The true value of carbon emissions for each sample. For the first Predicted carbon emissions for each sample; robustness loss The calculation expression is: in, It is a piecewise function. For the first The true value of energy consumption or carbon emissions for each sample. For the first The predicted energy consumption or carbon emissions for each sample, and the covariance consistency loss. The calculation expression is: in, This is the covariance matrix of the predicted energy consumption vector and carbon emission vector within the current training batch. This is the covariance matrix of the actual energy consumption vector and the actual carbon emission vector within the current training batch. Frobenius norm used to measure the distance between two matrices; physical mechanism soft constraint loss. The calculation expression is: Where ReLU is a linear rectified function. The theoretical combustion emission coefficient, : To prevent the minimum value of division by zero, These are the weighting coefficients for stoichiometric constraints.

8. The method for synergistic prediction of energy consumption and carbon emissions in carbon production according to claim 6, characterized in that, The preset constraints include nonnegativity constraints, statistical interval constraints, and material balance constraints. If the multi-task model violates any of these constraints, it is determined that the current multi-task model has fallen into a pseudo-convergence state. The parameters of the multi-task model are re-initialized and a new round of training begins until the multi-task model converges and satisfies all preset constraints.

9. The method for synergistic prediction of energy consumption and carbon emissions in carbon production according to claim 8, characterized in that, The non-negativity constraint requires that all predicted energy consumption and carbon emission values ​​must be non-negative; the statistical interval constraint requires that the energy consumption per unit product must be within the historical statistical interval or within the physical limit range specified by industry standards; the material balance constraint requires that the difference between the total content of a certain element in the material input system and the total content of that element in the material output system does not exceed a preset difference threshold.

10. A carbon production energy consumption and carbon emission co-prediction system, characterized in that, include: The data acquisition and preprocessing module acquires and preprocesses multi-source heterogeneous data from the entire carbon production process. The multi-source heterogeneous data includes short-term time series data, long-term time series data, spatial data, and static or event feature data. The parallel feature extraction module adds static or event feature data to short-term time series data, long-term time series data and spatial data respectively based on multi-source heterogeneous data of the entire carbon production process, and extracts short-term time feature vectors, long-term time feature vectors and spatial coupling feature vectors accordingly. The dynamic feature fusion module concatenates short-term, long-term, and spatially coupled feature vectors to generate a fused feature vector. The collaborative regression and constraint optimization module uses the fused feature vector as input to the multi-task model and future energy consumption and pollutant emission predictions as outputs to train the multi-task model end-to-end until it converges and meets preset constraints, resulting in the trained multi-task model. The prediction service interaction module inputs multi-source heterogeneous data of the entire carbon production process for the period to be predicted into the trained multi-task model and outputs the collaborative prediction results of energy consumption and carbon emissions.

Images