A carbon capture energy consumption dynamic prediction method and system based on imitation learning

Abstract

This invention provides a method and system for dynamic prediction of carbon capture energy consumption based on imitation learning, comprising: collecting real-time flue gas parameters, real-time operation parameters, and real-time utility parameters of a carbon capture system and fusing the data; inputting real-time multi-source fused time-series data into a pre-trained joint energy consumption prediction model; and realizing energy consumption prediction of the carbon capture system. The model is trained as follows: acquiring historical multi-source fused time-series data, identifying sub-data that meet preset conditions as a standard demonstration trajectory set, and using sub-data from the historical multi-source fused time-series data other than the standard demonstration trajectory set as a benchmark sample set; inputting samples from the benchmark sample set and the standard demonstration trajectory set into a preset basic model to obtain multi-task prediction loss and imitation learning loss; and optimizing the model parameters using the two losses to obtain the aforementioned model. This invention can effectively improve the accuracy of energy consumption prediction for carbon capture systems.

Description

Technical Field

[0001] This invention application relates to the field of energy consumption prediction for carbon capture systems, and more particularly to a method and system for dynamic prediction of carbon capture energy consumption based on imitation learning. Background Technology

[0002] With the advancement of relevant policies, carbon capture, utilization, and storage (CCUS) technology has become a key means to reduce carbon emissions from sources such as coal-fired power plants and industrial boilers. Among these, post-combustion chemical absorption is widely used due to its high technological maturity and strong adaptability. In the actual operation of carbon capture systems, energy consumption mainly consists of the heat consumption for amine regeneration and the electricity consumption for auxiliary equipment such as fans and pumps. Accurately predicting the dynamic changes in system energy consumption is a prerequisite for achieving operational optimization and reducing operating costs.

[0003] Currently, most existing carbon capture energy consumption prediction schemes are based on thermodynamic mechanism models or single data-driven models, and their data processing often focuses on single-dimensional parameter acquisition, such as fitting local patterns based solely on the amine circulation rate or specific thermodynamic parameters on the operating side. However, carbon capture systems involve complex heat and mass exchange processes, and their final energy consumption is deeply influenced by the coupling of multiple heterogeneous parameters from the flue gas side, the operating side, and the utility side. Because existing technologies mainly rely on a single data dimension, when faced with complex operating conditions where heterogeneous parameters are deeply intertwined, the lack of key information can easily lead to models failing to accurately represent the physical field state. This limitation of single-dimensional modeling makes it difficult for prediction models to maintain high-precision mapping relationships when multiple parameters fluctuate, thus affecting the accuracy of carbon capture system energy consumption prediction. Summary of the Invention

[0004] This invention application provides a method and system for dynamic prediction of carbon capture energy consumption based on imitation learning, in order to solve the technical problem of how to improve the accuracy of energy consumption prediction for carbon capture systems.

[0005] To address the aforementioned technical problems, this invention provides a method for dynamically predicting carbon capture energy consumption based on imitation learning, comprising:

[0006] Real-time flue gas parameters, real-time operation parameters, and real-time utility parameters of the carbon capture system are collected; the real-time flue gas parameters, real-time operation parameters, and real-time utility parameters are fused to obtain real-time multi-source fused time-series data; The real-time multi-source fused time-series data is input into a pre-trained joint energy consumption prediction model. The joint energy consumption prediction model includes a shared network, a heat consumption prediction branch, and an electricity consumption prediction branch. The shared network is used to extract time-series features from the real-time multi-source fused time-series data to obtain shared time-series features. The heat consumption prediction branch outputs the predicted unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, and the electricity consumption prediction branch outputs the predicted unit carbon dioxide electricity consumption under the current operating condition based on the shared time-series features. Based on the predicted unit carbon dioxide heat consumption and unit carbon dioxide electricity consumption, the energy consumption prediction of the carbon capture system is realized.

[0007] As a preferred embodiment, the joint energy consumption prediction model is trained in the following manner: Acquire historical multi-source fusion time-series data of the carbon capture system; Sub-data that meet preset conditions are identified from the historical multi-source fusion time series data as a standard demonstration trajectory set, and sub-data other than the standard demonstration trajectory set in the historical multi-source fusion time series data are used as a benchmark sample set. The samples in the benchmark sample set are input into the preset base model, and the difference between the corresponding output and the actual energy consumption value is calculated to obtain the multi-task prediction loss; the samples in the standard demonstration trajectory set are input into the base model, and the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set is calculated to obtain the imitation learning loss. The parameters of the joint energy consumption prediction model are optimized based on the multi-task prediction loss and the imitation learning loss to obtain the pre-trained joint energy consumption prediction model.

[0008] As a preferred embodiment, the step of identifying sub-data that meets preset conditions from the historical multi-source fused time-series data as a standard demonstration trajectory set includes: The historical multi-source fusion time series data is divided into multiple candidate trajectory segments by a sliding window according to a preset time window length. For each candidate trajectory segment, the fluctuation amplitude of the preset key parameters in the candidate trajectory segment within the time window is calculated, and candidate trajectory segments with fluctuation amplitude greater than the preset fluctuation threshold are removed to obtain a set of steady-state trajectory segments; Based on the operating condition characteristic parameters of each steady-state trajectory segment in the steady-state trajectory segment set, the steady-state trajectory segments in the steady-state trajectory segment set are divided into multiple operating condition groups; wherein, the operating condition characteristic parameters include at least the flue gas load rate and the flue gas carbon dioxide concentration; For each of the aforementioned operating condition groups, the unit carbon dioxide comprehensive energy consumption value corresponding to each steady-state trajectory segment within the operating condition group is sorted in ascending order, and the steady-state trajectory segments located within the preset quantile interval after sorting are determined as the preferred segment set corresponding to the operating condition group. The union of the preferred segment sets corresponding to each of the aforementioned operating condition groups is taken as the standard demonstration trajectory set; wherein, for operating condition groups where the preferred segment set is empty, the steady-state trajectory segment with the lowest unit carbon dioxide comprehensive energy consumption value is selected from the steady-state trajectory segment set corresponding to the operating condition group and added to the standard demonstration trajectory set.

[0009] As a preferred embodiment, the shared network includes a multi-scale temporal convolutional layer, a feature dimension self-attention module, and a temporal gating aggregation module. The multi-scale temporal convolutional layer includes at least two temporal convolutional channels with different time window lengths set in parallel. The shared network is used to extract time-series features from the real-time multi-source fused time-series data to obtain shared time-series features, including: The real-time multi-source fusion time-series data is input into the multi-scale temporal convolutional layer; wherein, each of the temporal convolutional channels is used to extract the temporal features of the real-time multi-source fusion time-series data under the corresponding time window length; The temporal features output from at least two of the temporal convolutional channels are concatenated along the feature dimension to obtain multi-scale fused temporal features; The multi-scale fused temporal features are input into the feature dimension self-attention module; wherein, the feature dimension self-attention module is used to divide the multi-scale fused temporal features into multiple feature channels along the feature dimension, and to perform feature similarity measurement on each feature channel and other feature channels respectively, and to calculate the coupling attention weight between each feature channel; The multi-scale fused temporal features are weighted according to the feature dimension based on the coupled attention weights to obtain coupled enhanced temporal features; The coupling-enhanced temporal features are input into the temporal gating aggregation module; wherein, the temporal gating aggregation module is used to generate corresponding temporal gating weights based on the features of each time step in the coupling-enhanced temporal features, and to perform weighted aggregation of the coupling-enhanced temporal features along the time dimension based on the temporal gating weights to obtain the shared temporal features.

[0010] As a preferred embodiment, the standard energy consumption value corresponding to the standard demonstration trajectory set includes the standard heat consumption value and the standard electricity consumption value of each sample in the standard demonstration trajectory set; The step of inputting samples from the standard demonstration trajectory set into the base model, calculating the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set, and obtaining the imitation learning loss includes: Input the samples from the standard demonstration trajectory set into the basic model to obtain the predicted value of unit carbon dioxide heat consumption and unit carbon dioxide electricity consumption for each sample. The imitation learning loss is obtained by summing the predicted unit carbon dioxide heat consumption value and the standard heat consumption value corresponding to all samples in the standard demonstration trajectory set, and the predicted unit carbon dioxide electricity consumption value and the standard electricity consumption value, according to a preset imitation loss weighting coefficient.

[0011] As a preferred embodiment, the heat dissipation prediction branch includes a heat dissipation feature mapping sub-network and a heat dissipation output layer connected in sequence; The heat consumption prediction branch outputs a predicted value of unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, including: The shared temporal features are nonlinearly mapped using the heat dissipation feature mapping subnetwork to obtain heat dissipation specific features. The heat consumption-specific characteristics are regressed through the heat consumption output layer to output the predicted value of unit carbon dioxide heat consumption under the current operating conditions.

[0012] As a preferred embodiment, the power consumption prediction branch includes a power consumption feature mapping sub-network and a power consumption output layer connected in sequence; The power consumption prediction branch outputs a predicted unit carbon dioxide power consumption value under the current operating condition based on the shared time-series features, including: The shared time-series features are nonlinearly mapped using the power consumption feature mapping subnetwork to obtain power consumption-specific features. The power consumption-specific characteristics are regressed through the power consumption output layer to output the predicted unit carbon dioxide power consumption under the current operating conditions.

[0013] As a preferred embodiment, the step of predicting the energy consumption of the carbon capture system based on the predicted unit carbon dioxide heat consumption and the predicted unit carbon dioxide electricity consumption includes: Based on the preset energy conversion factor, the predicted value of unit carbon dioxide electricity consumption is converted into the energy equivalent corresponding to the electricity consumption, and the predicted value of unit carbon dioxide heat consumption is summed with the energy equivalent corresponding to the electricity consumption to obtain the predicted value of unit carbon dioxide comprehensive energy consumption under the current working condition. The carbon dioxide capture flow rate of the carbon capture system under the current operating conditions is obtained, and the product of the carbon dioxide capture flow rate and the predicted value of the comprehensive energy consumption per unit of carbon dioxide is used as the energy consumption prediction result of the carbon capture system under the current operating conditions.

[0014] As a preferred embodiment, the real-time utility parameters include the reboiler input steam flow rate; the real-time operating parameters include the reboiler inlet amine liquid temperature, the reboiler outlet amine liquid temperature, the amine liquid circulation rate, and the lean liquid temperature; and the real-time flue gas parameters include the flue gas flow rate, the flue gas inlet carbon dioxide concentration, and the flue gas outlet carbon dioxide concentration. The process of fusing the real-time flue gas parameters, the real-time operation parameters, and the real-time utility parameters to obtain real-time multi-source fused time-series data includes: The real-time flue gas parameters, real-time operation parameters, and real-time utility parameters are time-aligned according to a preset reference time resolution to obtain aligned data. Based on the preset mass conservation relationship, the carbon dioxide absorption rate derivative sequence is calculated based on the flue gas flow rate, flue gas inlet carbon dioxide concentration, and flue gas outlet carbon dioxide concentration of the aligned data; wherein, the data point at each moment in the carbon dioxide absorption rate derivative sequence represents the mass of carbon dioxide absorbed by the amine liquid system from the flue gas per unit time. According to the preset energy conservation relationship, based on the reboiler input steam flow rate, reboiler inlet amine liquid temperature and reboiler outlet amine liquid temperature of the aligned data, the actual heat load derivation sequence of the reboiler is calculated; wherein, the data point at each moment in the actual heat load derivation sequence of the reboiler represents the actual thermal energy power input from the reboiler to the desorption tower. Based on the carbon dioxide absorption rate derived sequence and the reboiler actual heat load derived sequence, a unit carbon dioxide instantaneous heat consumption intensity derived sequence is calculated; wherein, the data point at each moment in the unit carbon dioxide instantaneous heat consumption intensity derived sequence is the ratio of the data point corresponding to the reboiler actual heat load derived sequence to the data point corresponding to the carbon dioxide absorption rate derived sequence. Based on the preset thermodynamic equilibrium relationship of amine liquid, and using the amine liquid circulation volume and lean liquid temperature of the aligned data, a derived sequence of amine liquid carbon dioxide absorption capacity is calculated; wherein, the data point at each moment in the derived sequence of amine liquid carbon dioxide absorption capacity represents the upper limit of the carbon dioxide absorption capacity of the amine liquid in the lean liquid state at that moment. The aligned data, the carbon dioxide absorption derivative sequence, the reboiler actual heat load derivative sequence, the unit carbon dioxide instantaneous heat consumption intensity derivative sequence, and the amine liquid carbon dioxide absorption capacity derivative sequence are stacked along the feature dimension, and the stacking result is divided according to a preset time window length to obtain the real-time multi-source fusion time series data.

[0015] Accordingly, this invention application also provides a dynamic prediction system for carbon capture energy consumption based on imitation learning, including a data fusion module, a carbon dioxide parameter prediction module, an energy consumption prediction module, and a model training module; wherein, The data fusion module is used to collect real-time flue gas side parameters, real-time operation side parameters, and real-time utility parameters of the carbon capture system; and to fuse the real-time flue gas side parameters, the real-time operation side parameters, and the real-time utility parameters to obtain real-time multi-source fused time-series data. The carbon dioxide parameter prediction module is used to input the real-time multi-source fusion time-series data into a pre-trained joint energy consumption prediction model. The joint energy consumption prediction model includes a shared network, a heat consumption prediction branch, and an electricity consumption prediction branch. The shared network is used to extract time-series features from the real-time multi-source fusion time-series data to obtain shared time-series features. The heat consumption prediction branch outputs the predicted unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, and the electricity consumption prediction branch outputs the predicted unit carbon dioxide electricity consumption under the current operating condition based on the shared time-series features. The energy consumption prediction module is used to predict the energy consumption of the carbon capture system based on the predicted unit carbon dioxide heat consumption and the predicted unit carbon dioxide electricity consumption.

[0016] As a preferred embodiment, the joint energy consumption prediction model is trained in the following manner: The model training module acquires historical multi-source fusion time-series data of the carbon capture system; Sub-data that meet preset conditions are identified from the historical multi-source fusion time series data as a standard demonstration trajectory set, and sub-data other than the standard demonstration trajectory set in the historical multi-source fusion time series data are used as a benchmark sample set. The samples in the benchmark sample set are input into the preset base model, and the difference between the corresponding output and the actual energy consumption value is calculated to obtain the multi-task prediction loss; the samples in the standard demonstration trajectory set are input into the base model, and the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set is calculated to obtain the imitation learning loss. The parameters of the joint energy consumption prediction model are optimized based on the multi-task prediction loss and the imitation learning loss to obtain the pre-trained joint energy consumption prediction model.

[0017] As a preferred embodiment, the model training module identifies sub-data that meets preset conditions from the historical multi-source fused time-series data as a standard demonstration trajectory set, including: The model training module performs sliding window segmentation on the historical multi-source fusion time series data according to a preset time window length to obtain multiple candidate trajectory segments; For each candidate trajectory segment, the fluctuation amplitude of the preset key parameters in the candidate trajectory segment within the time window is calculated, and candidate trajectory segments with fluctuation amplitude greater than the preset fluctuation threshold are removed to obtain a set of steady-state trajectory segments; Based on the operating condition characteristic parameters of each steady-state trajectory segment in the steady-state trajectory segment set, the steady-state trajectory segments in the steady-state trajectory segment set are divided into multiple operating condition groups; wherein, the operating condition characteristic parameters include at least the flue gas load rate and the flue gas carbon dioxide concentration; For each of the aforementioned operating condition groups, the unit carbon dioxide comprehensive energy consumption value corresponding to each steady-state trajectory segment within the operating condition group is sorted in ascending order, and the steady-state trajectory segments located within the preset quantile interval after sorting are determined as the preferred segment set corresponding to the operating condition group. The union of the preferred segment sets corresponding to each of the aforementioned operating condition groups is taken as the standard demonstration trajectory set; wherein, for operating condition groups where the preferred segment set is empty, the steady-state trajectory segment with the lowest unit carbon dioxide comprehensive energy consumption value is selected from the steady-state trajectory segment set corresponding to the operating condition group and added to the standard demonstration trajectory set.

[0018] As a preferred embodiment, the shared network includes a multi-scale temporal convolutional layer, a feature dimension self-attention module, and a temporal gating aggregation module. The multi-scale temporal convolutional layer includes at least two temporal convolutional channels with different time window lengths set in parallel. The shared network is used to extract time-series features from the real-time multi-source fused time-series data to obtain shared time-series features, including: The real-time multi-source fusion time-series data is input into the multi-scale temporal convolutional layer; wherein, each of the temporal convolutional channels is used to extract the temporal features of the real-time multi-source fusion time-series data under the corresponding time window length; The temporal features output from at least two of the temporal convolutional channels are concatenated along the feature dimension to obtain multi-scale fused temporal features; The multi-scale fused temporal features are input into the feature dimension self-attention module; wherein, the feature dimension self-attention module is used to divide the multi-scale fused temporal features into multiple feature channels along the feature dimension, and to perform feature similarity measurement on each feature channel and other feature channels respectively, and to calculate the coupling attention weight between each feature channel; The multi-scale fused temporal features are weighted according to the feature dimension based on the coupled attention weights to obtain coupled enhanced temporal features; The coupling-enhanced temporal features are input into the temporal gating aggregation module; wherein, the temporal gating aggregation module is used to generate corresponding temporal gating weights based on the features of each time step in the coupling-enhanced temporal features, and to perform weighted aggregation of the coupling-enhanced temporal features along the time dimension based on the temporal gating weights to obtain the shared temporal features.

[0019] As a preferred embodiment, the standard energy consumption value corresponding to the standard demonstration trajectory set includes the standard heat consumption value and the standard electricity consumption value of each sample in the standard demonstration trajectory set; The step of inputting samples from the standard demonstration trajectory set into the base model, calculating the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set, and obtaining the imitation learning loss includes: Input the samples from the standard demonstration trajectory set into the basic model to obtain the predicted value of unit carbon dioxide heat consumption and unit carbon dioxide electricity consumption for each sample. The imitation learning loss is obtained by summing the predicted unit carbon dioxide heat consumption value and the standard heat consumption value corresponding to all samples in the standard demonstration trajectory set, and the predicted unit carbon dioxide electricity consumption value and the standard electricity consumption value, according to a preset imitation loss weighting coefficient.

[0020] As a preferred embodiment, the heat dissipation prediction branch includes a heat dissipation feature mapping sub-network and a heat dissipation output layer connected in sequence; The heat consumption prediction branch outputs a predicted value of unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, including: The shared temporal features are nonlinearly mapped using the heat dissipation feature mapping subnetwork to obtain heat dissipation specific features. The heat consumption-specific characteristics are regressed through the heat consumption output layer to output the predicted value of unit carbon dioxide heat consumption under the current operating conditions.

[0021] As a preferred embodiment, the power consumption prediction branch includes a power consumption feature mapping sub-network and a power consumption output layer connected in sequence; The power consumption prediction branch outputs a predicted unit carbon dioxide power consumption value under the current operating condition based on the shared time-series features, including: The shared time-series features are nonlinearly mapped using the power consumption feature mapping subnetwork to obtain power consumption-specific features. The power consumption-specific characteristics are regressed through the power consumption output layer to output the predicted unit carbon dioxide power consumption under the current operating conditions.

[0022] As a preferred embodiment, the energy consumption prediction module, based on the predicted unit carbon dioxide heat consumption and the predicted unit carbon dioxide electricity consumption, realizes the energy consumption prediction of the carbon capture system, including: The energy consumption prediction module converts the predicted value of electricity consumption per unit of carbon dioxide into the energy equivalent corresponding to the electricity consumption according to the preset energy conversion coefficient, and sums the predicted value of heat consumption per unit of carbon dioxide with the energy equivalent corresponding to the electricity consumption to obtain the predicted value of comprehensive energy consumption per unit of carbon dioxide under the current working condition. The carbon dioxide capture flow rate of the carbon capture system under the current operating conditions is obtained, and the product of the carbon dioxide capture flow rate and the predicted value of the comprehensive energy consumption per unit of carbon dioxide is used as the energy consumption prediction result of the carbon capture system under the current operating conditions.

[0023] As a preferred embodiment, the real-time utility parameters include the reboiler input steam flow rate; the real-time operating parameters include the reboiler inlet amine liquid temperature, the reboiler outlet amine liquid temperature, the amine liquid circulation rate, and the lean liquid temperature; and the real-time flue gas parameters include the flue gas flow rate, the flue gas inlet carbon dioxide concentration, and the flue gas outlet carbon dioxide concentration. The data fusion module fuses the real-time flue gas parameters, the real-time operation parameters, and the real-time utility parameters to obtain real-time multi-source fused time-series data, including: The data fusion module performs time alignment on the real-time flue gas parameters, the real-time operation parameters, and the real-time utility parameters according to a preset reference time resolution to obtain aligned data. Based on the preset mass conservation relationship, the carbon dioxide absorption rate derivative sequence is calculated based on the flue gas flow rate, flue gas inlet carbon dioxide concentration, and flue gas outlet carbon dioxide concentration of the aligned data; wherein, the data point at each moment in the carbon dioxide absorption rate derivative sequence represents the mass of carbon dioxide absorbed by the amine liquid system from the flue gas per unit time. According to the preset energy conservation relationship, based on the reboiler input steam flow rate, reboiler inlet amine liquid temperature and reboiler outlet amine liquid temperature of the aligned data, the actual heat load derivation sequence of the reboiler is calculated; wherein, the data point at each moment in the actual heat load derivation sequence of the reboiler represents the actual thermal energy power input from the reboiler to the desorption tower. Based on the carbon dioxide absorption rate derived sequence and the reboiler actual heat load derived sequence, a unit carbon dioxide instantaneous heat consumption intensity derived sequence is calculated; wherein, the data point at each moment in the unit carbon dioxide instantaneous heat consumption intensity derived sequence is the ratio of the data point corresponding to the reboiler actual heat load derived sequence to the data point corresponding to the carbon dioxide absorption rate derived sequence. Based on the preset thermodynamic equilibrium relationship of amine liquid, and using the amine liquid circulation volume and lean liquid temperature of the aligned data, a derived sequence of amine liquid carbon dioxide absorption capacity is calculated; wherein, the data point at each moment in the derived sequence of amine liquid carbon dioxide absorption capacity represents the upper limit of the carbon dioxide absorption capacity of the amine liquid in the lean liquid state at that moment. The aligned data, the carbon dioxide absorption derivative sequence, the reboiler actual heat load derivative sequence, the unit carbon dioxide instantaneous heat consumption intensity derivative sequence, and the amine liquid carbon dioxide absorption capacity derivative sequence are stacked along the feature dimension, and the stacking result is divided according to a preset time window length to obtain the real-time multi-source fusion time series data.

[0024] Compared with the prior art, this invention application has the following beneficial effects: This invention provides a method and system for dynamic prediction of carbon capture energy consumption based on imitation learning. It collects real-time flue gas parameters, real-time operation parameters, and real-time utility parameters of the carbon capture system and fuses these data to obtain real-time multi-source fused time-series data. This fused data is then used as input to a prediction model. Compared to existing technologies that rely solely on single-dimensional parameters such as the amine circulation rate and specific thermodynamic parameters on the operation side for localized pattern fitting, the multi-source fused time-series data of this invention comprehensively covers the key coupling information affecting the energy consumption of the carbon capture system from the flue gas, operation, and utility sides. This avoids the loss of key information due to single-dimensional data collection, enabling the prediction model to accurately characterize the physical field state under complex operating conditions with deeply intertwined heterogeneous parameters, thereby significantly improving the accuracy of carbon capture system energy consumption prediction. Furthermore, in the joint energy consumption prediction model adopted in this invention application, the shared network performs unified time-series feature extraction on real-time multi-source fusion time-series data. After obtaining the shared time-series features, the heat consumption prediction branch and the electricity consumption prediction branch output the predicted value of heat consumption per unit of carbon dioxide and the predicted value of electricity consumption per unit of carbon dioxide, respectively. On the one hand, since heat consumption and electricity consumption share the same time-series feature extraction network, the repetitive calculation of the underlying feature extraction process when building independent models for heat consumption and electricity consumption is avoided, reducing the overall computational cost and parameter scale of the model. On the other hand, since the heat consumption prediction branch and the electricity consumption prediction branch output prediction results based on the same set of shared time-series features, the prediction results of the two branches are based on a consistent operating condition representation, avoiding the distortion of the coupling relationship between the predicted values ​​of heat consumption and electricity consumption caused by inconsistent feature extraction when modeling independently, and improving the consistency of the predicted results of heat consumption and electricity consumption under the same operating condition. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating an embodiment of the carbon capture energy consumption dynamic prediction method based on imitation learning provided in this invention application.

[0026] Figure 2 This is a schematic diagram of an embodiment of the carbon capture energy consumption dynamic prediction system based on imitation learning provided in this invention application. Detailed Implementation

[0027] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Example 1 Please refer to Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the carbon capture energy consumption dynamic prediction method based on imitation learning provided in this invention application.

[0029] Figure 1 The illustrated embodiment includes steps S101 to S103; each step is described in detail below: Step S101: Collect real-time flue gas side parameters, real-time operation side parameters, and real-time utility parameters of the carbon capture system; perform data fusion on the real-time flue gas side parameters, the real-time operation side parameters, and the real-time utility parameters to obtain real-time multi-source fused time-series data.

[0030] Specifically, in this embodiment, the carbon capture system continuously collects three types of operating parameters through field instruments and a data acquisition system (DCS): among which, the real-time flue gas side parameters include flue gas flow rate, flue gas inlet carbon dioxide concentration, and flue gas outlet carbon dioxide concentration. These parameters reflect the flue gas load status and carbon capture efficiency entering the absorption tower.

[0031] Real-time operating parameters include the amine solution temperature at the reboiler inlet, the amine solution temperature at the reboiler outlet, the amine solution circulation rate, and the lean solution temperature. These parameters reflect the circulation status of the amine solution system and the thermodynamic operating status of the desorption tower.

[0032] Real-time utility parameters include the reboiler input steam flow rate, which reflects the direct energy input source of the carbon capture system's heat consumption. Since these three types of parameters originate from different sensor subsystems and their original acquisition frequencies differ, data fusion processing is required to construct a unified and complete multi-source time-series feature space for use in subsequent energy consumption prediction models.

[0033] In some preferred embodiments, the step of fusing the real-time flue gas parameters, the real-time operation parameters, and the real-time utility parameters to obtain real-time multi-source fused time-series data includes: The real-time flue gas parameters, real-time operation parameters, and real-time utility parameters are time-aligned according to a preset reference time resolution to obtain aligned data. Based on the preset mass conservation relationship, the carbon dioxide absorption rate derivative sequence is calculated based on the flue gas flow rate, flue gas inlet carbon dioxide concentration, and flue gas outlet carbon dioxide concentration of the aligned data; wherein, the data point at each moment in the carbon dioxide absorption rate derivative sequence represents the mass of carbon dioxide absorbed by the amine liquid system from the flue gas per unit time. According to the preset energy conservation relationship, based on the reboiler input steam flow rate, reboiler inlet amine liquid temperature and reboiler outlet amine liquid temperature of the aligned data, the actual heat load derivation sequence of the reboiler is calculated; wherein, the data point at each moment in the actual heat load derivation sequence of the reboiler represents the actual thermal energy power input from the reboiler to the desorption tower. Based on the carbon dioxide absorption rate derived sequence and the reboiler actual heat load derived sequence, a unit carbon dioxide instantaneous heat consumption intensity derived sequence is calculated; wherein, the data point at each moment in the unit carbon dioxide instantaneous heat consumption intensity derived sequence is the ratio of the data point corresponding to the reboiler actual heat load derived sequence to the data point corresponding to the carbon dioxide absorption rate derived sequence. Based on the preset thermodynamic equilibrium relationship of amine liquid, and using the amine liquid circulation volume and lean liquid temperature of the aligned data, a derived sequence of amine liquid carbon dioxide absorption capacity is calculated; wherein, the data point at each moment in the derived sequence of amine liquid carbon dioxide absorption capacity represents the upper limit of the carbon dioxide absorption capacity of the amine liquid in the lean liquid state at that moment. The aligned data, the carbon dioxide absorption derivative sequence, the reboiler actual heat load derivative sequence, the unit carbon dioxide instantaneous heat consumption intensity derivative sequence, and the amine liquid carbon dioxide absorption capacity derivative sequence are stacked along the feature dimension, and the stacking result is divided according to a preset time window length to obtain the real-time multi-source fusion time series data.

[0034] In this embodiment, time alignment is first performed. The real-time flue gas parameters, real-time operational parameters, and real-time utility parameters are time-aligned according to a preset reference time resolution to obtain aligned data. Specifically, since the flue gas parameters, operational parameters, and utility parameters originate from different types of sensors with inconsistent original sampling intervals, direct splicing would cause misalignment of the parameters on the time axis, failing to accurately reflect the system state at the same moment. Therefore, this step uses a preset reference time resolution as a unified time reference to resample or interpolate and align the parameter sequences, ensuring that each time step in the aligned data corresponds to a complete observation of the three types of parameters at the same moment. In this embodiment, the preset reference time resolution can be 1 minute, meaning an aligned time step containing all parameters is generated every minute; those skilled in the art can also adjust the preset reference time resolution to other suitable values ​​based on the actual system's data acquisition frequency and prediction accuracy requirements.

[0035] Secondly, based on the preset mass conservation principle, a carbon dioxide absorption derivative sequence is calculated using the aligned data, including flue gas flow rate, flue gas inlet carbon dioxide concentration, and flue gas outlet carbon dioxide concentration. Each data point in the carbon dioxide absorption derivative sequence represents the mass of carbon dioxide absorbed by the amine liquid system from the flue gas per unit time. Specifically, at each aligned time step, based on the mass conservation principle and the difference between the current flue gas flow rate and the flue gas inlet and outlet carbon dioxide concentrations, the actual mass of carbon dioxide absorbed by the amine liquid system from the flue gas per unit time at that moment can be calculated, thus forming the carbon dioxide absorption derivative sequence. The introduction of this derivative sequence compresses multiple measurement parameters from the original flue gas side into a comprehensive quantity that directly characterizes the physical effectiveness of carbon capture, helping subsequent models to more accurately establish the mapping relationship between energy consumption and capture volume at the feature level.

[0036] Furthermore, based on the pre-defined energy conservation principle, and using the reboiler input steam flow rate, reboiler inlet amine liquid temperature, and reboiler outlet amine liquid temperature from the aligned data, a derived sequence of the actual heat load of the reboiler is calculated. Each data point in this derived sequence represents the actual thermal power input from the reboiler to the desorption tower. Specifically, at each aligned time step, based on the principle of energy conservation, and considering the temperature difference between the reboiler input steam flow rate and the inlet / outlet amine liquid at the current moment, the actual thermal power input from the reboiler to the desorption tower at that moment can be calculated, thus forming the derived sequence of the actual heat load of the reboiler. Compared to using steam flow rate or temperature parameters alone, this derived sequence more directly represents the actual energy consumption state of the desorption process at the thermodynamic level, helping to improve the model's ability to perceive dynamic changes in heat consumption.

[0037] Based on the carbon dioxide absorption derivative sequence and the reboiler actual heat load derivative sequence, a unit carbon dioxide instantaneous heat consumption intensity derivative sequence is calculated. Specifically, the data point at each moment in the unit carbon dioxide instantaneous heat consumption intensity derivative sequence is the ratio of the corresponding data point in the reboiler actual heat load derivative sequence to the corresponding data point in the carbon dioxide absorption derivative sequence. At each aligned time step, dividing the current reboiler actual heat load by the current carbon dioxide absorption yields the instantaneous heat energy consumed to capture a unit mass of carbon dioxide at that moment, thus forming the unit carbon dioxide instantaneous heat consumption intensity derivative sequence. This derivative sequence, in the form of normalized energy consumption intensity, establishes a real-time quantitative correspondence between heat energy supply and carbon capture effect, serving as an explicit feature reflecting the operating efficiency under current conditions and providing direct monitoring signals for subsequent models. It should be noted that when the carbon dioxide absorption data point at a certain moment approaches zero (e.g., during abnormal operating conditions such as system start-up or shutdown), the data point at that moment should be marked as abnormal or filtered to avoid numerical anomalies caused by division by zero contaminating subsequent model inputs.

[0038] Based on the preset thermodynamic equilibrium relationship of the amine liquid, and using the aligned data on the amine liquid circulation rate and lean liquid temperature, a derived sequence of amine liquid carbon dioxide absorption capacity is calculated. Each data point in this derived sequence represents the upper limit of the amine liquid's carbon dioxide absorption capacity under lean liquid conditions at that moment. Specifically, at each aligned time step, the carbon dioxide absorption capacity of the amine liquid under lean liquid conditions is significantly affected by the lean liquid temperature; the higher the temperature, the lower the upper limit of the amine liquid's absorption capacity. Based on the preset thermodynamic equilibrium relationship of the amine liquid, and combining the current amine liquid circulation rate and lean liquid temperature, the upper limit of the carbon dioxide absorption capacity that the amine liquid system can bear under lean liquid conditions at that moment can be estimated, thus forming the derived sequence of amine liquid carbon dioxide absorption capacity. This derived sequence provides a state characterization of the system's current absorption capacity from a thermodynamic perspective, helping the model to perceive the dynamic absorption margin of the amine liquid system under varying operating conditions, thereby improving prediction accuracy.

[0039] The aligned data, the carbon dioxide absorption derivative sequence, the reboiler actual heat load derivative sequence, the unit carbon dioxide instantaneous heat consumption intensity derivative sequence, and the amine liquid carbon dioxide absorption capacity derivative sequence are stacked along the feature dimension, and the stacking result is segmented according to a preset time window length to obtain the real-time multi-source fused time series data. Specifically, the stacking operation merges the original acquisition parameters (flue gas flow rate, flue gas inlet carbon dioxide concentration, flue gas outlet carbon dioxide concentration, reboiler inlet amine liquid temperature, reboiler outlet amine liquid temperature, amine liquid circulation rate, lean liquid temperature, and reboiler input steam flow rate) in the aligned data with the above four derivative sequences along the feature dimension, forming a complete multi-dimensional time series matrix in which each time step contains the original parameters and derived features. Subsequently, the multi-dimensional time series matrix is ​​segmented by a sliding window according to the preset time window length, so that each segmented sample contains complete feature observations of several consecutive time steps, thereby preserving the temporal context information of the system operating parameters for subsequent energy consumption joint prediction model to extract time series features.

[0040] This preferred embodiment eliminates the temporal misalignment of multi-source heterogeneous parameters by time alignment, and transforms the original measurement parameters into comprehensive characterization quantities with physical mechanism significance by introducing four derived sequences. Thus, the final real-time multi-source fusion time series data not only retains the complete information of various original parameters, but also adds high-level features reflecting the thermodynamic state and operating efficiency of the system, providing high-quality and higher information density input for the time series feature extraction of the energy consumption joint prediction model in the subsequent step S102.

[0041] Step S102: Input the real-time multi-source fusion time series data into the pre-trained joint energy consumption prediction model.

[0042] The energy consumption joint prediction model includes a shared network, a heat consumption prediction branch, and an electricity consumption prediction branch. The shared network is used to extract time-series features from the real-time multi-source fusion time-series data to obtain shared time-series features. The heat consumption prediction branch outputs the predicted value of unit carbon dioxide heat consumption under the current operating conditions based on the shared time-series features, and the electricity consumption prediction branch outputs the predicted value of unit carbon dioxide electricity consumption under the current operating conditions based on the shared time-series features.

[0043] In some implementations, the shared network includes a multi-scale temporal convolutional layer, a feature dimension self-attention module, and a temporal gating aggregation module. The multi-scale temporal convolutional layer includes at least two temporal convolutional channels with different time window lengths set in parallel. The shared network is used to extract time-series features from the real-time multi-source fused time-series data to obtain shared time-series features, including: The real-time multi-source fusion time-series data is input into the multi-scale temporal convolutional layer; wherein, each of the temporal convolutional channels is used to extract the temporal features of the real-time multi-source fusion time-series data under the corresponding time window length; The temporal features output from at least two of the temporal convolutional channels are concatenated along the feature dimension to obtain multi-scale fused temporal features; The multi-scale fused temporal features are input into the feature dimension self-attention module; wherein, the feature dimension self-attention module is used to divide the multi-scale fused temporal features into multiple feature channels along the feature dimension, and to perform feature similarity measurement on each feature channel and other feature channels respectively, and to calculate the coupling attention weight between each feature channel; The multi-scale fused temporal features are weighted according to the feature dimension based on the coupled attention weights to obtain coupled enhanced temporal features; The coupling-enhanced temporal features are input into the temporal gating aggregation module; wherein, the temporal gating aggregation module is used to generate corresponding temporal gating weights based on the features of each time step in the coupling-enhanced temporal features, and to perform weighted aggregation of the coupling-enhanced temporal features along the time dimension based on the temporal gating weights to obtain the shared temporal features.

[0044] In this embodiment, the real-time multi-source fused time-series data is input into the multi-scale temporal convolutional layer; wherein, each temporal convolutional channel is used to extract the temporal features of the real-time multi-source fused time-series data within a corresponding time window length. Specifically, the operating parameters of the carbon capture system exhibit dynamic changes at different scales on the time axis: parameters such as flue gas flow rate and reboiler input steam flow rate may fluctuate rapidly in a short period of time, while parameters such as amine liquid circulation volume and lean liquid temperature show a strong slow-change trend. If only a convolutional channel with a single time window length is used for feature extraction, it is difficult to simultaneously take into account both the rapid and slow-change dynamic modes mentioned above. Therefore, the multi-scale temporal convolutional layer sets up multiple parallel temporal convolutional channels, each channel using a different time window length, wherein the temporal convolutional channel with a shorter time window length focuses on the rapid change features of the sensing parameters within a local time range, while the temporal convolutional channel with a longer time window length focuses on the trend change features of the sensing parameters over a longer time span. In this embodiment, the multi-scale temporal convolutional layer may include three parallel temporal convolutional channels, corresponding to short, medium, and long time scales respectively; those skilled in the art may also set other numbers of parallel temporal convolutional channels and corresponding time window lengths according to the dynamic response characteristics of the actual system.

[0045] Subsequently, the temporal features output from at least two of the aforementioned temporal convolutional channels are concatenated along the feature dimension to obtain multi-scale fused temporal features. Specifically, the output of each parallel temporal convolutional channel contains local temporal pattern information at the corresponding time scale. By concatenating along the feature dimension, temporal features at different scales can be integrated into the same feature space, enabling the multi-scale fused temporal features to simultaneously contain fast-changing local features and slow-changing trend features, thereby providing a more complete temporal representation for subsequent modules.

[0046] The multi-scale fused temporal features are input into the feature dimension self-attention module. This module divides the multi-scale fused temporal features into multiple feature channels along the feature dimension, performs feature similarity measurement on each feature channel with other feature channels, and calculates the coupling attention weights between the feature channels. Specifically, complex physical coupling relationships exist between the multi-source operating parameters of a carbon capture system. For example, changes in the reboiler input steam flow rate synchronously affect the reboiler outlet amine liquid temperature, thereby affecting the carbon dioxide absorption capacity of the amine liquid, ultimately reflected in changes in the instantaneous heat consumption intensity per unit of carbon dioxide. Simple temporal convolution operations only extract local patterns in the time dimension and cannot explicitly model the coupling dependencies between the aforementioned feature channels. To this end, the feature dimension self-attention module divides the multi-scale fused temporal features into multiple feature channels in the feature dimension, performs feature similarity measurement on each feature channel and other feature channels respectively (or uses the representation of each feature channel as the query (Q), key (K) and value (V) of the attention mechanism respectively), calculates the coupling attention weight between any two feature channels, and thus explicitly captures the degree of mutual influence between each parameter channel in the feature dimension.

[0047] The multi-scale fused temporal features are weighted according to the feature dimension based on the coupling attention weight to obtain the coupling-enhanced temporal features. Specifically, the coupling attention weight reflects the degree of contribution of each feature channel to other feature channels. By reweighting the multi-scale fused temporal features according to this weight in the feature dimension, the expression of coupling feature channels that are strongly correlated with energy consumption prediction can be highlighted, while the interference of feature channels with low correlation to subsequent predictions can be suppressed. Thus, the coupling-enhanced temporal features more accurately reflect the multi-parameter coupled physical state of the carbon capture system in the feature dimension.

[0048] The coupled-enhanced time-series features are input into the time-series gating aggregation module. The time-series gating aggregation module generates corresponding time-series gating weights based on the features of each time step in the coupled-enhanced time-series features, and performs weighted aggregation of the coupled-enhanced time-series features along the time dimension based on the time-series gating weights to obtain the shared time-series features. Specifically, in the actual operation of a carbon capture system, the operating status of different time steps within a time window has different levels of reference value for judging the current operating condition: time steps in a stable operating condition phase usually contain more reliable energy consumption pattern information, while time steps in a phase of operating condition switching or parameter mutation may introduce noise interference. If the features of all time steps within the time window are aggregated equally, the above differences cannot be distinguished. To this end, the time-series gating aggregation module applies a learnable gating transformation to the feature vector of each time step in the coupled enhancement time-series features, generating a scalar gating weight corresponding to that time step. After normalization, this weight reflects the contribution of the corresponding time step to the final operating condition judgment within the current time window. Subsequently, based on the gating weights of each time step, the coupled enhancement time-series features are weighted and aggregated along the time dimension, compressing the variable-length time-step sequence into a fixed-dimensional feature vector, thus obtaining the shared time-series features. These shared time-series features, as a compact representation of the system's operating state within the current time window, will serve as the common input for subsequent heat consumption prediction and power consumption prediction branches, used to predict unit carbon dioxide heat consumption and power consumption under the current operating condition, respectively.

[0049] In some preferred embodiments, the heat dissipation prediction branch includes a heat dissipation feature mapping subnetwork and a heat dissipation output layer connected in sequence; The heat consumption prediction branch outputs a predicted value of unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, including: The shared temporal features are nonlinearly mapped using the heat dissipation feature mapping subnetwork to obtain heat dissipation specific features. The heat consumption-specific characteristics are regressed through the heat consumption output layer to output the predicted value of unit carbon dioxide heat consumption under the current operating conditions.

[0050] In this preferred embodiment, the shared time-series features are uniformly extracted by a shared network, and include operating condition characterization information related to both heat consumption prediction and power consumption prediction. If the shared time-series features are directly used as the regression input for heat consumption prediction, the power consumption-related components in the feature vector will be introduced as redundant information into the heat consumption prediction process, affecting the accuracy of heat consumption prediction. Therefore, the heat consumption feature mapping sub-network uses several nonlinear transformation layers to perform task-specific feature extraction on the shared time-series features, strengthening the feature components highly correlated with the heat consumption prediction task while suppressing components with low correlation to heat consumption prediction, thereby obtaining heat consumption-specific features semantically specific to the heat consumption prediction task. These heat consumption-specific features are closer to the decision boundary of the heat consumption prediction task in the feature space, helping the subsequent heat consumption output layer achieve more accurate regression prediction.

[0051] The heat consumption-specific characteristics are regressed through the heat consumption output layer to output the predicted heat consumption per unit of carbon dioxide under the current operating conditions. Specifically, the heat consumption output layer maps the heat consumption-specific characteristics to a single scalar output, which is the heat energy required to capture a unit mass of carbon dioxide under the current operating conditions. In this embodiment, the unit of the predicted heat consumption per unit of carbon dioxide can be GJ / tCO2; those skilled in the art can also use other equivalent units according to actual engineering needs. The output result of the heat consumption prediction branch will be jointly transmitted with the predicted electricity consumption per unit of carbon dioxide output from the electricity consumption prediction branch to the subsequent step S103 to realize the comprehensive energy consumption prediction of the carbon capture system.

[0052] The power consumption prediction branch includes a power consumption feature mapping sub-network and a power consumption output layer connected in sequence; The power consumption prediction branch outputs a predicted unit carbon dioxide power consumption value under the current operating condition based on the shared time-series features, including: The shared time-series features are nonlinearly mapped using the power consumption feature mapping subnetwork to obtain power consumption-specific features. The power consumption-specific characteristics are regressed through the power consumption output layer to output the predicted unit carbon dioxide power consumption under the current operating conditions.

[0053] In this preferred embodiment, the power consumption of the carbon capture system mainly originates from the operating power consumption of auxiliary equipment such as fans and pumps. Its dynamic changes are more directly correlated with parameters such as flue gas flow rate and amine liquid circulation volume, and its physical causes are fundamentally different from heat consumption, which is driven primarily by reboiler steam consumption. Although the shared time-series features have formed a unified and compact representation of the overall system operating state, they still contain feature components related to both heat and power consumption prediction. If the shared time-series features are directly used as the regression input for power consumption prediction, the feature components related to heat consumption will act as redundant information, interfering with the power consumption prediction process. Therefore, the power consumption feature mapping sub-network uses several nonlinear transformation layers to extract task-specific features from the shared time-series features, strengthening the feature components highly correlated with the power consumption prediction task while suppressing components with low correlation to power consumption prediction, thereby obtaining power consumption-specific features semantically specific to the power consumption prediction task. The power consumption-specific features and the heat consumption-specific features output by the heat consumption prediction branch together embody the design idea of ​​decoupling task-specific features under the multi-task learning framework. That is, on the basis of sharing the underlying temporal representation, each prediction branch completes feature refinement for its own prediction target, thereby effectively distinguishing the specific feature spaces of heat consumption and power consumption prediction while retaining the common information between tasks.

[0054] The power consumption output layer performs regression processing on the power consumption-specific characteristics to output the predicted power consumption per unit of carbon dioxide under the current operating conditions. Specifically, the power consumption output layer maps the power consumption-specific characteristics to a single scalar output, i.e., the electrical energy required to capture a unit mass of carbon dioxide under the current operating conditions. In this embodiment, the unit of the predicted power consumption per unit of carbon dioxide can be kWh / tCO2; those skilled in the art can also use other equivalent units according to actual engineering needs. The output result of the power consumption prediction branch, together with the predicted heat consumption per unit of carbon dioxide output from the heat consumption prediction branch, is passed to the subsequent step S103 to realize the comprehensive energy consumption prediction of the carbon capture system.

[0055] Step S103: Based on the predicted unit carbon dioxide heat consumption and the predicted unit carbon dioxide electricity consumption, the energy consumption prediction of the carbon capture system is realized.

[0056] The predicted unit carbon dioxide heat consumption output by the heat consumption prediction branch and the predicted unit carbon dioxide electricity consumption output by the electricity consumption prediction branch are not consistent in terms of dimensions: the former uses heat energy as the measurement basis directly, while the latter uses electricity energy as the measurement basis. If the two are directly added together, the overall energy consumption calculation result will lose its physical meaning due to the inconsistency in dimensions.

[0057] Therefore, in some preferred embodiments, step S103 includes: Based on the preset energy conversion factor, the predicted value of unit carbon dioxide electricity consumption is converted into the energy equivalent corresponding to the electricity consumption, and the predicted value of unit carbon dioxide heat consumption is summed with the energy equivalent corresponding to the electricity consumption to obtain the predicted value of unit carbon dioxide comprehensive energy consumption under the current working condition. The carbon dioxide capture flow rate of the carbon capture system under the current operating conditions is obtained, and the product of the carbon dioxide capture flow rate and the predicted value of the comprehensive energy consumption per unit of carbon dioxide is used as the energy consumption prediction result of the carbon capture system under the current operating conditions.

[0058] This preferred embodiment uses a thermoelectric equivalent conversion relationship to convert the predicted unit carbon dioxide electricity consumption from electrical energy units to thermal energy equivalent using a preset energy conversion factor. This ensures that the predicted electricity consumption and thermal consumption are consistent in dimensions, providing a physically clear calculation basis for subsequent summation operations. The energy conversion factor is preset based on the thermoelectric conversion efficiency or equivalent calorific value standard in the actual application scenario and remains fixed during system operation; those skilled in the art can adjust this factor according to specific engineering needs.

[0059] Furthermore, the predicted heat consumption per unit of carbon dioxide is summed with the energy equivalent corresponding to the power consumption to obtain the predicted comprehensive energy consumption per unit of carbon dioxide under the current operating conditions. Specifically, after unifying the dimensions, the predicted heat consumption per unit of carbon dioxide is directly added to the energy equivalent corresponding to the power consumption to obtain the total heat and power consumed in capturing a unit mass of carbon dioxide under the current operating conditions, i.e., the predicted comprehensive energy consumption per unit of carbon dioxide. This comprehensive energy consumption prediction is expressed in a unified heat energy benchmark unit, fully reflecting the overall energy intensity of the carbon capture system under the current operating conditions, and can serve as a core quantitative indicator for operational optimization decisions.

[0060] Finally, the carbon dioxide capture flow rate of the carbon capture system under the current operating conditions is obtained. The product of the carbon dioxide capture flow rate and the predicted energy consumption per unit of carbon dioxide is used as the energy consumption prediction result of the carbon capture system under the current operating conditions. Specifically, the carbon dioxide capture flow rate can be calculated in real time based on the flue gas side parameters collected in step S101, or directly read from the data acquisition system (DCS) of the carbon capture system. The predicted energy consumption per unit of carbon dioxide represents the comprehensive energy required to capture a unit mass of carbon dioxide, which is an intensity quantity; while in actual operation and management, the carbon capture system usually needs to know the absolute total energy consumption of the system under the current operating conditions, i.e., the extensive quantity. Therefore, by multiplying the carbon dioxide capture flow rate by the predicted energy consumption per unit of carbon dioxide, the absolute energy consumption at the system level can be calculated from the unit intensity quantity, thus obtaining the energy consumption prediction result of the carbon capture system under the current operating conditions. This prediction result can be directly used in application scenarios such as energy consumption management, operation scheduling optimization, and energy efficiency assessment of the carbon capture system.

[0061] In some preferred embodiments, the joint energy consumption prediction model is trained in the following manner: Acquire historical multi-source fusion time-series data of the carbon capture system; Sub-data that meet preset conditions are identified from the historical multi-source fusion time series data as a standard demonstration trajectory set, and sub-data other than the standard demonstration trajectory set in the historical multi-source fusion time series data are used as a benchmark sample set. The samples in the benchmark sample set are input into the preset base model, and the difference between the corresponding output and the actual energy consumption value is calculated to obtain the multi-task prediction loss; the samples in the standard demonstration trajectory set are input into the base model, and the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set is calculated to obtain the imitation learning loss. The parameters of the joint energy consumption prediction model are optimized based on the multi-task prediction loss and the imitation learning loss to obtain the pre-trained joint energy consumption prediction model.

[0062] In this embodiment, the historical multi-source fusion time series data can be obtained by fusing historical flue gas side parameters, historical operation side parameters and historical utility parameters using the same method as step S101.

[0063] In some implementations, identifying sub-data that meets preset conditions from the historical multi-source fused time-series data as a standard demonstration trajectory set includes: The historical multi-source fusion time series data is divided into multiple candidate trajectory segments by a sliding window according to a preset time window length. For each candidate trajectory segment, the fluctuation amplitude of the preset key parameters in the candidate trajectory segment within the time window is calculated, and candidate trajectory segments with fluctuation amplitude greater than the preset fluctuation threshold are removed to obtain a set of steady-state trajectory segments; Based on the operating condition characteristic parameters of each steady-state trajectory segment in the steady-state trajectory segment set, the steady-state trajectory segments in the steady-state trajectory segment set are divided into multiple operating condition groups; wherein, the operating condition characteristic parameters include at least the flue gas load rate and the flue gas carbon dioxide concentration; For each of the aforementioned operating condition groups, the unit carbon dioxide comprehensive energy consumption value corresponding to each steady-state trajectory segment within the operating condition group is sorted in ascending order, and the steady-state trajectory segments located within the preset quantile interval after sorting are determined as the preferred segment set corresponding to the operating condition group. The union of the preferred segment sets corresponding to each of the aforementioned operating condition groups is taken as the standard demonstration trajectory set; wherein, for operating condition groups where the preferred segment set is empty, the steady-state trajectory segment with the lowest unit carbon dioxide comprehensive energy consumption value is selected from the steady-state trajectory segment set corresponding to the operating condition group and added to the standard demonstration trajectory set.

[0064] In this preferred embodiment, the preset time window length can be consistent with the time window length used in the real-time inference stage of step S101, thereby ensuring that the input samples in the training and inference stages are completely consistent in terms of temporal length and feature structure, avoiding a decrease in model generalization ability due to mismatch between training and inference input formats. Through sliding window segmentation, historical multi-source fusion time-series data is divided into a large number of overlapping continuous time-series segments. Each candidate trajectory segment contains a complete multi-source feature observation sequence within the preset time window length, constituting the basic processing unit for subsequent steady-state screening and optimization evaluation.

[0065] Secondly, for each candidate trajectory segment, the fluctuation amplitude of a preset key parameter within the time window is calculated. Candidate trajectory segments with fluctuation amplitudes greater than a preset fluctuation threshold are removed to obtain a steady-state trajectory segment set. Specifically, the core idea of ​​imitation learning is to extract expert-level high-quality operational demonstrations from historical operating data for the model to imitate and learn from. However, if trajectory segments in non-steady-state transition stages such as operating condition switching or large parameter fluctuations are included in the demonstration set, the energy consumption performance corresponding to such segments often cannot truly reflect the operating rules of the system under stable operating conditions, and using them as demonstration trajectories will mislead the model's learning direction. Therefore, this embodiment uses core operating parameters that significantly affect energy consumption, such as flue gas flow rate and amine liquid circulation volume, as preset key parameters, and calculates the fluctuation amplitude of the above key parameters for each candidate trajectory segment within the time window. The preset fluctuation threshold is preset based on the statistical fluctuation distribution of the above key parameters in historical data. Candidate trajectory segments with fluctuation amplitudes exceeding the threshold are judged as non-steady-state segments and removed, thereby retaining steady-state trajectory segments with relatively stable operating states to form a steady-state trajectory segment set.

[0066] Furthermore, based on the operating condition characteristic parameters of each steady-state trajectory segment in the steady-state trajectory segment set, the steady-state trajectory segments in the set are divided into multiple operating condition groups; wherein, the operating condition characteristic parameters include at least the flue gas load rate and the flue gas carbon dioxide concentration. Specifically, the optimal operating strategies exhibited by the carbon capture system under different combinations of flue gas load rates and flue gas carbon dioxide concentrations vary significantly. If high-quality demonstrations are directly selected from the full set of steady-state trajectory segments without distinguishing operating conditions, segments with lower energy consumption under high-load conditions will lose their reference value under low-load conditions, failing to provide effective demonstrations covering the entire operating condition range for the model. Therefore, this embodiment groups the steady-state trajectory segments based on the flue gas load rate and the flue gas carbon dioxide concentration. Specifically, discrete binning or clustering methods can be used to group steady-state trajectory segments with similar operating conditions into the same operating condition group, thereby ensuring that subsequent optimization operations are carried out under the premise of consistent operating conditions, and ensuring that the selected demonstration trajectories have targeted representative significance for the optimal operation under each operating condition.

[0067] For each operating condition group, the steady-state trajectory segments within the group are sorted in ascending order based on their unit carbon dioxide energy consumption values. The steady-state trajectory segments within a preset quantile are then identified as the preferred segment set for that operating condition group. Specifically, within the same operating condition group, the unit carbon dioxide energy consumption values ​​of different steady-state trajectory segments still differ, reflecting the energy efficiency differences between different operating strategies under the same operating conditions. This step sorts the steady-state trajectory segments within each operating condition group in ascending order based on their unit carbon dioxide energy consumption values, identifying the steady-state trajectory segments with energy consumption levels in lower quantiles as the preferred segment set for that operating condition group. This represents an operational demonstration with relatively excellent energy efficiency in historical operation under that operating condition. The preset quantile is pre-set based on the energy consumption distribution of actual operating data; those skilled in the art can adjust this quantile according to specific application scenarios.

[0068] Finally, the union of the preferred segment sets corresponding to each operating condition group is taken as the standard demonstration trajectory set. For operating condition groups with empty preferred segment sets, the steady-state trajectory segment with the lowest unit carbon dioxide comprehensive energy consumption value is selected from the steady-state trajectory segment set corresponding to the operating condition group and added to the standard demonstration trajectory set. Specifically, by taking the union of the preferred segment sets of each operating condition group, a standard demonstration trajectory set covering the entire operating condition range can be obtained. Each trajectory segment in this set represents a relatively energy-efficient operation demonstration in historical operation under the corresponding operating condition, and can serve as a source of supervision signal for subsequent imitation learning loss calculation. It should be noted that, with limited historical data, the number of steady-state trajectory segments in some operating condition groups may be insufficient to cover the preset quantile interval, resulting in an empty preferred segment set for that operating condition group. To avoid coverage gaps in the standard demonstration trajectory set under this operating condition, this step sets up a fallback supplementation mechanism: for operating condition groups where the preferred segment set is empty, the steady-state trajectory segment with the lowest unit carbon dioxide comprehensive energy consumption value is selected from its corresponding steady-state trajectory segment set and supplemented into the standard demonstration trajectory set, thereby ensuring that the standard demonstration trajectory set has demonstration coverage for all operating condition groups and improving the adaptability of imitation learning under varying operating conditions.

[0069] Preferably, the actual energy consumption values ​​include the actual heat consumption value and the actual electricity consumption value corresponding to each sample in the benchmark sample set, which can be used as supervision labels for the heat consumption prediction task and the electricity consumption prediction task, respectively. For the heat consumption prediction task, the prediction loss of the heat consumption prediction task is calculated based on the difference between the predicted value and the actual heat consumption value per unit of carbon dioxide corresponding to all samples in the benchmark sample set; for the electricity consumption prediction task, the prediction loss of the electricity consumption prediction task is calculated based on the difference between the predicted value and the actual electricity consumption value per unit of carbon dioxide corresponding to all samples in the benchmark sample set. Subsequently, the prediction losses of the above two prediction tasks are weighted and summed according to a preset task loss weighting coefficient to obtain the multi-task prediction loss. The task loss weighting coefficient is preset according to the relative importance of the heat consumption prediction task and the electricity consumption prediction task in practical applications, and is used to balance the contribution ratio of the two prediction tasks to the optimization direction of model parameters during multi-task joint training; those skilled in the art can adjust the weighting coefficient according to specific engineering needs.

[0070] In some preferred embodiments, the standard energy consumption value corresponding to the standard demonstration trajectory set includes the standard heat consumption value and the standard electricity consumption value of each sample in the standard demonstration trajectory set; The step of inputting samples from the standard demonstration trajectory set into the base model, calculating the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set, and obtaining the imitation learning loss includes: Input the samples from the standard demonstration trajectory set into the basic model to obtain the predicted value of unit carbon dioxide heat consumption and unit carbon dioxide electricity consumption for each sample. The imitation learning loss is obtained by summing the predicted unit carbon dioxide heat consumption value and the standard heat consumption value corresponding to all samples in the standard demonstration trajectory set, and the predicted unit carbon dioxide electricity consumption value and the standard electricity consumption value, according to a preset imitation loss weighting coefficient.

[0071] In this embodiment, each sample in the standard demonstration trajectory set represents a relatively energy-efficient operating segment in the historical operation under the corresponding operating condition. The standard heat consumption value and standard electricity consumption value of each sample in the standard demonstration trajectory set are taken from the actual heat consumption and electricity consumption observation values ​​recorded in the historical operation of that trajectory segment. Compared with the actual energy consumption values ​​of the samples in the benchmark sample set, the standard heat consumption value and standard electricity consumption value of the samples in the standard demonstration trajectory set are in the quantile range of excellent energy efficiency within the same operating condition group, representing the expert-level low-energy consumption operating benchmark of the carbon capture system under the corresponding operating condition, and serving as the source of supervision signals for the simulation learning loss calculation.

[0072] The samples from the standard demonstration trajectory set are input into the base model to obtain the predicted unit carbon dioxide heat consumption and unit carbon dioxide electricity consumption for each sample. Specifically, the base model used here is the same model instance used in the multi-task prediction loss calculation step, i.e., an energy consumption joint prediction model with the same network structure and parameters currently in an unoptimized state. By inputting the samples from the standard demonstration trajectory set into this base model, the current prediction output of the model on expert-level high-quality demonstration samples can be obtained, providing a computational basis for subsequent comparison with standard energy consumption values.

[0073] The imitation learning loss is obtained by summing the differences between the predicted unit carbon dioxide heat consumption and the standard heat consumption value for all samples in the standard demonstration trajectory set, and the differences between the predicted unit carbon dioxide electricity consumption and the standard electricity consumption value, using a preset imitation loss weighting coefficient. Specifically, for the heat consumption prediction task, the imitation deviation in the heat consumption direction is calculated based on the differences between the predicted unit carbon dioxide heat consumption and the corresponding standard heat consumption value for all samples in the standard demonstration trajectory set; for the electricity consumption prediction task, the imitation deviation in the electricity consumption direction is calculated based on the differences between the predicted unit carbon dioxide electricity consumption and the corresponding standard electricity consumption value for all samples in the standard demonstration trajectory set. Subsequently, the imitation deviations in the above two directions are summed using a preset imitation loss weighting coefficient to obtain the imitation learning loss. The imitation loss weighting coefficient is preset according to the relative importance of heat consumption and electricity consumption in the imitation learning objective, and is independent of the task loss weighting coefficient in the multi-task prediction loss calculation step, controlling the contribution ratio of the imitation deviations in the heat consumption and electricity consumption directions to the optimization direction of model parameters respectively; those skilled in the art can adjust the weighting coefficient according to specific engineering needs.

[0074] The physical meaning of the imitation learning loss lies in measuring the gap between the current output of the base model and the expert-level high-quality operation demonstration. The smaller the value, the closer the model's predicted output is to the low-energy operation benchmark represented by the standard demonstration trajectory set. The imitation learning loss will be used together with the multi-task prediction loss to jointly optimize the parameters of the base model: the multi-task prediction loss ensures that the model maintains overall prediction accuracy on a large number of regular operation samples, while the imitation learning loss guides the model to converge towards low-energy operation on expert-level high-quality demonstration samples. The synergistic effect of the two makes the finally trained joint energy consumption prediction model have both prediction accuracy and low-energy operation orientation under varying operating conditions.

[0075] Accordingly, such as Figure 2 As shown, this invention application also provides a carbon capture energy consumption dynamic prediction system 200 based on imitation learning, including a data fusion module 201, a carbon dioxide parameter prediction module 202, an energy consumption prediction module 203, and a model training module; wherein, The data fusion module 201 is used to collect real-time flue gas side parameters, real-time operation side parameters, and real-time utility parameters of the carbon capture system; and to fuse the real-time flue gas side parameters, the real-time operation side parameters, and the real-time utility parameters to obtain real-time multi-source fused time-series data. The carbon dioxide parameter prediction module 202 is used to input the real-time multi-source fusion time-series data into a pre-trained energy consumption joint prediction model; wherein, the energy consumption joint prediction model includes a shared network, a heat consumption prediction branch, and an electricity consumption prediction branch; the shared network is used to extract time-series features from the real-time multi-source fusion time-series data to obtain shared time-series features; the heat consumption prediction branch outputs the predicted value of unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, and the electricity consumption prediction branch outputs the predicted value of unit carbon dioxide electricity consumption under the current operating condition based on the shared time-series features; The energy consumption prediction module 203 is used to predict the energy consumption of the carbon capture system based on the predicted unit carbon dioxide heat consumption and the predicted unit carbon dioxide electricity consumption.

[0076] As a preferred embodiment, the joint energy consumption prediction model is trained in the following manner: The model training module acquires historical multi-source fusion time-series data of the carbon capture system; Sub-data that meet preset conditions are identified from the historical multi-source fusion time series data as a standard demonstration trajectory set, and sub-data other than the standard demonstration trajectory set in the historical multi-source fusion time series data are used as a benchmark sample set. The samples in the benchmark sample set are input into the preset base model, and the difference between the corresponding output and the actual energy consumption value is calculated to obtain the multi-task prediction loss; the samples in the standard demonstration trajectory set are input into the base model, and the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set is calculated to obtain the imitation learning loss. The parameters of the joint energy consumption prediction model are optimized based on the multi-task prediction loss and the imitation learning loss to obtain the pre-trained joint energy consumption prediction model.

[0077] As a preferred embodiment, the model training module identifies sub-data that meets preset conditions from the historical multi-source fused time-series data as a standard demonstration trajectory set, including: The model training module performs sliding window segmentation on the historical multi-source fusion time series data according to a preset time window length to obtain multiple candidate trajectory segments; For each candidate trajectory segment, the fluctuation amplitude of the preset key parameters in the candidate trajectory segment within the time window is calculated, and candidate trajectory segments with fluctuation amplitude greater than the preset fluctuation threshold are removed to obtain a set of steady-state trajectory segments; Based on the operating condition characteristic parameters of each steady-state trajectory segment in the steady-state trajectory segment set, the steady-state trajectory segments in the steady-state trajectory segment set are divided into multiple operating condition groups; wherein, the operating condition characteristic parameters include at least the flue gas load rate and the flue gas carbon dioxide concentration; For each of the aforementioned operating condition groups, the unit carbon dioxide comprehensive energy consumption value corresponding to each steady-state trajectory segment within the operating condition group is sorted in ascending order, and the steady-state trajectory segments located within the preset quantile interval after sorting are determined as the preferred segment set corresponding to the operating condition group. The union of the preferred segment sets corresponding to each of the aforementioned operating condition groups is taken as the standard demonstration trajectory set; wherein, for operating condition groups where the preferred segment set is empty, the steady-state trajectory segment with the lowest unit carbon dioxide comprehensive energy consumption value is selected from the steady-state trajectory segment set corresponding to the operating condition group and added to the standard demonstration trajectory set.

[0078] As a preferred embodiment, the shared network includes a multi-scale temporal convolutional layer, a feature dimension self-attention module, and a temporal gating aggregation module. The multi-scale temporal convolutional layer includes at least two temporal convolutional channels with different time window lengths set in parallel. The shared network is used to extract time-series features from the real-time multi-source fused time-series data to obtain shared time-series features, including: The real-time multi-source fusion time-series data is input into the multi-scale temporal convolutional layer; wherein, each of the temporal convolutional channels is used to extract the temporal features of the real-time multi-source fusion time-series data under the corresponding time window length; The temporal features output from at least two of the temporal convolutional channels are concatenated along the feature dimension to obtain multi-scale fused temporal features; The multi-scale fused temporal features are input into the feature dimension self-attention module; wherein, the feature dimension self-attention module is used to divide the multi-scale fused temporal features into multiple feature channels along the feature dimension, and to perform feature similarity measurement on each feature channel and other feature channels respectively, and to calculate the coupling attention weight between each feature channel; The multi-scale fused temporal features are weighted according to the feature dimension based on the coupled attention weights to obtain coupled enhanced temporal features; The coupling-enhanced temporal features are input into the temporal gating aggregation module; wherein, the temporal gating aggregation module is used to generate corresponding temporal gating weights based on the features of each time step in the coupling-enhanced temporal features, and to perform weighted aggregation of the coupling-enhanced temporal features along the time dimension based on the temporal gating weights to obtain the shared temporal features.

[0079] As a preferred embodiment, the standard energy consumption value corresponding to the standard demonstration trajectory set includes the standard heat consumption value and the standard electricity consumption value of each sample in the standard demonstration trajectory set; The step of inputting samples from the standard demonstration trajectory set into the base model, calculating the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set, and obtaining the imitation learning loss includes: Input the samples from the standard demonstration trajectory set into the basic model to obtain the predicted value of unit carbon dioxide heat consumption and unit carbon dioxide electricity consumption for each sample. The imitation learning loss is obtained by summing the predicted unit carbon dioxide heat consumption value and the standard heat consumption value corresponding to all samples in the standard demonstration trajectory set, and the predicted unit carbon dioxide electricity consumption value and the standard electricity consumption value, according to a preset imitation loss weighting coefficient.

[0080] As a preferred embodiment, the heat dissipation prediction branch includes a heat dissipation feature mapping sub-network and a heat dissipation output layer connected in sequence; The heat consumption prediction branch outputs a predicted value of unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, including: The shared temporal features are nonlinearly mapped using the heat dissipation feature mapping subnetwork to obtain heat dissipation specific features. The heat consumption-specific characteristics are regressed through the heat consumption output layer to output the predicted value of unit carbon dioxide heat consumption under the current operating conditions.

[0081] As a preferred embodiment, the power consumption prediction branch includes a power consumption feature mapping sub-network and a power consumption output layer connected in sequence; The power consumption prediction branch outputs a predicted unit carbon dioxide power consumption value under the current operating condition based on the shared time-series features, including: The shared time-series features are nonlinearly mapped using the power consumption feature mapping subnetwork to obtain power consumption-specific features. The power consumption-specific characteristics are regressed through the power consumption output layer to output the predicted unit carbon dioxide power consumption under the current operating conditions.

[0082] As a preferred embodiment, the energy consumption prediction module 203, based on the predicted unit carbon dioxide heat consumption and the predicted unit carbon dioxide electricity consumption, realizes the energy consumption prediction of the carbon capture system, including: The energy consumption prediction module 203 converts the predicted value of unit carbon dioxide electricity consumption into the energy equivalent corresponding to the electricity consumption according to the preset energy conversion coefficient, and sums the predicted value of unit carbon dioxide heat consumption with the energy equivalent corresponding to the electricity consumption to obtain the predicted value of unit carbon dioxide comprehensive energy consumption under the current working condition. The carbon dioxide capture flow rate of the carbon capture system under the current operating conditions is obtained, and the product of the carbon dioxide capture flow rate and the predicted value of the comprehensive energy consumption per unit of carbon dioxide is used as the energy consumption prediction result of the carbon capture system under the current operating conditions.

[0083] As a preferred embodiment, the real-time utility parameters include the reboiler input steam flow rate; the real-time operating parameters include the reboiler inlet amine liquid temperature, the reboiler outlet amine liquid temperature, the amine liquid circulation rate, and the lean liquid temperature; and the real-time flue gas parameters include the flue gas flow rate, the flue gas inlet carbon dioxide concentration, and the flue gas outlet carbon dioxide concentration. The data fusion module 201 fuses the real-time flue gas parameters, the real-time operation parameters, and the real-time utility parameters to obtain real-time multi-source fused time-series data, including: The data fusion module 201 performs time alignment on the real-time flue gas side parameters, the real-time operation side parameters, and the real-time utility parameters according to a preset reference time resolution to obtain aligned data; Based on the preset mass conservation relationship, the carbon dioxide absorption rate derivative sequence is calculated based on the flue gas flow rate, flue gas inlet carbon dioxide concentration, and flue gas outlet carbon dioxide concentration of the aligned data; wherein, the data point at each moment in the carbon dioxide absorption rate derivative sequence represents the mass of carbon dioxide absorbed by the amine liquid system from the flue gas per unit time. According to the preset energy conservation relationship, based on the reboiler input steam flow rate, reboiler inlet amine liquid temperature and reboiler outlet amine liquid temperature of the aligned data, the actual heat load derivation sequence of the reboiler is calculated; wherein, the data point at each moment in the actual heat load derivation sequence of the reboiler represents the actual thermal energy power input from the reboiler to the desorption tower. Based on the carbon dioxide absorption rate derived sequence and the reboiler actual heat load derived sequence, a unit carbon dioxide instantaneous heat consumption intensity derived sequence is calculated; wherein, the data point at each moment in the unit carbon dioxide instantaneous heat consumption intensity derived sequence is the ratio of the data point corresponding to the reboiler actual heat load derived sequence to the data point corresponding to the carbon dioxide absorption rate derived sequence. Based on the preset thermodynamic equilibrium relationship of amine liquid, and using the amine liquid circulation volume and lean liquid temperature of the aligned data, a derived sequence of amine liquid carbon dioxide absorption capacity is calculated; wherein, the data point at each moment in the derived sequence of amine liquid carbon dioxide absorption capacity represents the upper limit of the carbon dioxide absorption capacity of the amine liquid in the lean liquid state at that moment. The aligned data, the carbon dioxide absorption derivative sequence, the reboiler actual heat load derivative sequence, the unit carbon dioxide instantaneous heat consumption intensity derivative sequence, and the amine liquid carbon dioxide absorption capacity derivative sequence are stacked along the feature dimension, and the stacking result is divided according to a preset time window length to obtain the real-time multi-source fusion time series data.

[0084] Compared with the prior art, this invention application has the following beneficial effects: This invention provides a method and system for dynamic prediction of carbon capture energy consumption based on imitation learning. It collects real-time flue gas parameters, real-time operation parameters, and real-time utility parameters of the carbon capture system and fuses these data to obtain real-time multi-source fused time-series data. This fused data is then used as input to a prediction model. Compared to existing technologies that rely solely on single-dimensional parameters such as the amine circulation rate and specific thermodynamic parameters on the operation side for localized pattern fitting, the multi-source fused time-series data of this invention comprehensively covers the key coupling information affecting the energy consumption of the carbon capture system from the flue gas, operation, and utility sides. This avoids the loss of key information due to single-dimensional data collection, enabling the prediction model to accurately characterize the physical field state under complex operating conditions with deeply intertwined heterogeneous parameters, thereby significantly improving the accuracy of carbon capture system energy consumption prediction. Furthermore, in the joint energy consumption prediction model adopted in this invention application, the shared network performs unified time-series feature extraction on real-time multi-source fusion time-series data. After obtaining the shared time-series features, the heat consumption prediction branch and the electricity consumption prediction branch output the predicted value of heat consumption per unit of carbon dioxide and the predicted value of electricity consumption per unit of carbon dioxide, respectively. On the one hand, since heat consumption and electricity consumption share the same time-series feature extraction network, the repetitive calculation of the underlying feature extraction process when building independent models for heat consumption and electricity consumption is avoided, reducing the overall computational cost and parameter scale of the model. On the other hand, since the heat consumption prediction branch and the electricity consumption prediction branch output prediction results based on the same set of shared time-series features, the prediction results of the two branches are based on a consistent operating condition representation, avoiding the distortion of the coupling relationship between the predicted values ​​of heat consumption and electricity consumption caused by inconsistent feature extraction when modeling independently, and improving the consistency of the predicted results of heat consumption and electricity consumption under the same operating condition.

[0085] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the scope of protection of the present invention. In particular, it should be noted that any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention for those skilled in the art.

Claims

1. A method for dynamic prediction of carbon capture energy consumption based on imitation learning, characterized in that, include: Collect real-time flue gas side parameters, real-time operation side parameters, and real-time utility parameters of the carbon capture system; The real-time flue gas side parameters, the real-time operation side parameters, and the real-time utility parameters are fused to obtain real-time multi-source fused time-series data. The real-time multi-source fused time-series data is input into a pre-trained joint energy consumption prediction model. The joint energy consumption prediction model includes a shared network, a heat consumption prediction branch, and an electricity consumption prediction branch. The shared network is used to extract time-series features from the real-time multi-source fused time-series data to obtain shared time-series features. The heat consumption prediction branch outputs the predicted unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, and the electricity consumption prediction branch outputs the predicted unit carbon dioxide electricity consumption under the current operating condition based on the shared time-series features. Based on the predicted unit carbon dioxide heat consumption and unit carbon dioxide electricity consumption, the energy consumption prediction of the carbon capture system is realized.

2. The method for dynamic prediction of carbon capture energy consumption based on imitation learning as described in claim 1, characterized in that, The joint energy consumption prediction model was trained in the following manner: Acquire historical multi-source fusion time-series data of the carbon capture system; Sub-data that meet preset conditions are identified from the historical multi-source fusion time series data as a standard demonstration trajectory set, and sub-data other than the standard demonstration trajectory set in the historical multi-source fusion time series data are used as a benchmark sample set. The samples in the benchmark sample set are input into the preset base model, and the difference between the corresponding output and the actual energy consumption value is calculated to obtain the multi-task prediction loss; the samples in the standard demonstration trajectory set are input into the base model, and the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set is calculated to obtain the imitation learning loss. The parameters of the joint energy consumption prediction model are optimized based on the multi-task prediction loss and the imitation learning loss to obtain the pre-trained joint energy consumption prediction model.

3. The method for dynamic prediction of carbon capture energy consumption based on imitation learning as described in claim 2, characterized in that, The step of identifying sub-data that meet preset conditions from the historical multi-source fused time-series data as a standard demonstration trajectory set includes: The historical multi-source fusion time series data is divided into multiple candidate trajectory segments by a sliding window according to a preset time window length. For each candidate trajectory segment, the fluctuation amplitude of the preset key parameters in the candidate trajectory segment within the time window is calculated, and candidate trajectory segments with fluctuation amplitude greater than the preset fluctuation threshold are removed to obtain a set of steady-state trajectory segments; Based on the operating condition characteristic parameters of each steady-state trajectory segment in the steady-state trajectory segment set, the steady-state trajectory segments in the steady-state trajectory segment set are divided into multiple operating condition groups; wherein, the operating condition characteristic parameters include at least the flue gas load rate and the flue gas carbon dioxide concentration; For each of the aforementioned operating condition groups, the unit carbon dioxide comprehensive energy consumption value corresponding to each steady-state trajectory segment within the operating condition group is sorted in ascending order, and the steady-state trajectory segments located within the preset quantile interval after sorting are determined as the preferred segment set corresponding to the operating condition group. The union of the preferred segment sets corresponding to each of the aforementioned operating condition groups is taken as the standard demonstration trajectory set; wherein, for operating condition groups where the preferred segment set is empty, the steady-state trajectory segment with the lowest unit carbon dioxide comprehensive energy consumption value is selected from the steady-state trajectory segment set corresponding to the operating condition group and added to the standard demonstration trajectory set.

4. The method for dynamic prediction of carbon capture energy consumption based on imitation learning as described in claim 2, characterized in that, The standard energy consumption value corresponding to the standard demonstration trajectory set includes the standard heat consumption value and standard electricity consumption value of each sample in the standard demonstration trajectory set; The step of inputting samples from the standard demonstration trajectory set into the base model, calculating the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set, and obtaining the imitation learning loss includes: Input the samples from the standard demonstration trajectory set into the basic model to obtain the predicted value of unit carbon dioxide heat consumption and unit carbon dioxide electricity consumption for each sample. The imitation learning loss is obtained by summing the predicted unit carbon dioxide heat consumption value and the standard heat consumption value corresponding to all samples in the standard demonstration trajectory set, and the predicted unit carbon dioxide electricity consumption value and the standard electricity consumption value, according to a preset imitation loss weighting coefficient.

5. The method for dynamic prediction of carbon capture energy consumption based on imitation learning as described in claim 1, characterized in that, The shared network includes a multi-scale temporal convolutional layer, a feature dimension self-attention module, and a temporal gated aggregation module. The multi-scale temporal convolutional layer includes at least two temporal convolutional channels with different time window lengths set in parallel. The shared network is used to extract time-series features from the real-time multi-source fused time-series data to obtain shared time-series features, including: The real-time multi-source fusion time-series data is input into the multi-scale temporal convolutional layer; wherein, each of the temporal convolutional channels is used to extract the temporal features of the real-time multi-source fusion time-series data under the corresponding time window length; The temporal features output from at least two of the temporal convolutional channels are concatenated along the feature dimension to obtain multi-scale fused temporal features; The multi-scale fused temporal features are input into the feature dimension self-attention module; wherein, the feature dimension self-attention module is used to divide the multi-scale fused temporal features into multiple feature channels along the feature dimension, and to perform feature similarity measurement on each feature channel and other feature channels respectively, and to calculate the coupling attention weight between each feature channel; The multi-scale fused temporal features are weighted according to the feature dimension based on the coupled attention weights to obtain coupled enhanced temporal features; The coupling-enhanced temporal features are input into the temporal gating aggregation module; wherein, the temporal gating aggregation module is used to generate corresponding temporal gating weights based on the features of each time step in the coupling-enhanced temporal features, and to perform weighted aggregation of the coupling-enhanced temporal features along the time dimension based on the temporal gating weights to obtain the shared temporal features.

6. The method for dynamic prediction of carbon capture energy consumption based on imitation learning as described in claim 1, characterized in that, The heat dissipation prediction branch includes a heat dissipation feature mapping subnetwork and a heat dissipation output layer connected in sequence; The heat consumption prediction branch outputs a predicted value of unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, including: The shared temporal features are nonlinearly mapped using the heat dissipation feature mapping subnetwork to obtain heat dissipation specific features. The heat consumption-specific characteristics are regressed through the heat consumption output layer to output the predicted value of unit carbon dioxide heat consumption under the current operating conditions.

7. The method for dynamic prediction of carbon capture energy consumption based on imitation learning as described in claim 1, characterized in that, The power consumption prediction branch includes a power consumption feature mapping sub-network and a power consumption output layer connected in sequence; The power consumption prediction branch outputs a predicted unit carbon dioxide power consumption value under the current operating condition based on the shared time-series features, including: The shared time-series features are nonlinearly mapped using the power consumption feature mapping subnetwork to obtain power consumption-specific features. The power consumption-specific characteristics are regressed through the power consumption output layer to output the predicted unit carbon dioxide power consumption under the current operating conditions.

8. A method for dynamic prediction of carbon capture energy consumption based on imitation learning as described in any one of claims 1 to 7, characterized in that, The method of predicting the energy consumption of the carbon capture system based on the predicted unit carbon dioxide heat consumption and the predicted unit carbon dioxide electricity consumption includes: Based on the preset energy conversion factor, the predicted value of unit carbon dioxide electricity consumption is converted into the energy equivalent corresponding to the electricity consumption, and the predicted value of unit carbon dioxide heat consumption is summed with the energy equivalent corresponding to the electricity consumption to obtain the predicted value of unit carbon dioxide comprehensive energy consumption under the current working condition. The carbon dioxide capture flow rate of the carbon capture system under the current operating conditions is obtained, and the product of the carbon dioxide capture flow rate and the predicted value of the comprehensive energy consumption per unit of carbon dioxide is used as the energy consumption prediction result of the carbon capture system under the current operating conditions.

9. A dynamic prediction system for carbon capture energy consumption based on imitation learning, characterized in that, It includes a data fusion module, a carbon dioxide parameter prediction module, an energy consumption prediction module, and a model training module; among which, The data fusion module is used to collect real-time flue gas side parameters, real-time operation side parameters, and real-time utility parameters of the carbon capture system; and to fuse the real-time flue gas side parameters, the real-time operation side parameters, and the real-time utility parameters to obtain real-time multi-source fused time-series data. The carbon dioxide parameter prediction module is used to input the real-time multi-source fusion time-series data into a pre-trained joint energy consumption prediction model. The joint energy consumption prediction model includes a shared network, a heat consumption prediction branch, and an electricity consumption prediction branch. The shared network is used to extract time-series features from the real-time multi-source fusion time-series data to obtain shared time-series features. The heat consumption prediction branch outputs the predicted unit carbon dioxide heat consumption under the current operating condition based on the shared time-series features, and the electricity consumption prediction branch outputs the predicted unit carbon dioxide electricity consumption under the current operating condition based on the shared time-series features. The energy consumption prediction module is used to predict the energy consumption of the carbon capture system based on the predicted unit carbon dioxide heat consumption and the predicted unit carbon dioxide electricity consumption.

10. A carbon capture energy consumption dynamic prediction system based on imitation learning as described in claim 9, characterized in that, The joint energy consumption prediction model was trained in the following manner: The model training module acquires historical multi-source fusion time-series data of the carbon capture system; Sub-data that meet preset conditions are identified from the historical multi-source fusion time series data as a standard demonstration trajectory set, and sub-data other than the standard demonstration trajectory set in the historical multi-source fusion time series data are used as a benchmark sample set. The samples in the benchmark sample set are input into the preset base model, and the difference between the corresponding output and the actual energy consumption value is calculated to obtain the multi-task prediction loss; the samples in the standard demonstration trajectory set are input into the base model, and the difference between the corresponding output and the standard energy consumption value corresponding to the standard demonstration trajectory set is calculated to obtain the imitation learning loss. The parameters of the joint energy consumption prediction model are optimized based on the multi-task prediction loss and the imitation learning loss to obtain the pre-trained joint energy consumption prediction model.

Images