A power plant carbon emission prediction method and system

Abstract

The application discloses a power plant carbon emission prediction method and system, and belongs to the technical field of power plant carbon emission monitoring. The method is as follows: a static feature vector is obtained based on power plant operation data, and a dynamic feature matrix in a current sliding time window is obtained; a hybrid expert output vector is obtained based on the dynamic feature matrix and a hybrid expert network; an expert gate weight distribution is obtained based on the static feature vector, the dynamic feature matrix and a gate network; a fused dynamic feature vector is obtained based on the expert gate weight distribution and the hybrid expert output vector; a carbon emission prediction value is obtained based on the static feature vector, the fused dynamic feature vector and a prediction output network comprising a benchmark prediction branch and a fluctuation prediction branch; dynamic features in the dynamic feature matrix are sorted based on the contribution degree of the expert gate weight distribution, and target dynamic features meeting a preset screening condition are obtained; and a carbon emission prediction result is formed based on the target dynamic features and the carbon emission prediction value, so that real-time and accurate prediction of carbon emission is realized.

Description

Technical Field

[0001] This invention relates to the field of carbon emission monitoring technology for power plants, and in particular to a method and system for predicting carbon emissions from power plants. Background Technology

[0002] With the increasingly severe global climate change problem, the power industry, as a key sector consuming fossil energy, relies heavily on carbon emission monitoring and forecasting for carbon quota management, enterprise energy conservation and control, and combustion efficiency optimization. Currently, carbon emission data acquisition from power plants mainly relies on hardware-based continuous emission monitoring systems and data-driven soft measurement technologies (i.e., carbon emission prediction models). However, due to the susceptibility of hardware equipment to field environmental influences leading to data drift, resulting in data loss or large measurement errors, and the high cost of equipment maintenance, deep learning soft measurement methods based on unit operation data have been widely applied. Among these, hybrid expert models and residual structures have shown strong potential in time series prediction tasks and are gradually becoming an important research direction in the field of carbon emission prediction.

[0003] However, existing technical solutions have significant shortcomings in practical applications. Specifically, existing hybrid expert prediction models lack explicit physical phase space guidance mechanisms, making it impossible to distinguish between steady-state and transient operating conditions based on the thermodynamic operating state of the unit. Prediction lag is prone to occur when switching between different load conditions. At the same time, single prediction networks are difficult to handle the scale coupling problem of large base and small fluctuations in carbon emissions, and have poor ability to capture small fluctuations in emissions. In addition, the model training process lacks effective constraints on the diversity of experts, and hybrid expert networks are prone to pattern collapse, resulting in the failure of expert division of labor and the inability to fully utilize the advantages of ensemble learning. This makes it difficult to meet the real-time carbon emission prediction requirements of power plants for high accuracy and high robustness. Summary of the Invention

[0004] This invention provides a method and system for predicting carbon emissions from power plants, which can solve the technical problem that existing carbon emission prediction models cannot take into account both large-scale emissions and small fluctuations due to the use of a single network, resulting in insufficient prediction accuracy, and achieve real-time and accurate prediction of carbon emissions.

[0005] This invention provides a method for predicting carbon emissions from power plants, applied to a pre-built carbon emission prediction model. The carbon emission prediction model includes a pre-built hybrid expert network, a gated network, and a prediction output network. The method includes: Static feature vectors and dynamic feature matrices within the current sliding time window are obtained based on real-time collected power plant operation data. The hybrid expert output vector is obtained based on the dynamic feature matrix and the hybrid expert network. The expert gating weight distribution is obtained based on the static feature vector, dynamic feature matrix, and gating network. The hybrid expert output vector is weighted and summed based on the expert gating weight distribution to obtain a fused dynamic feature vector. Based on the static feature vector, the fused dynamic feature vector, and the prediction output network including the baseline prediction branch and the fluctuation prediction branch, the carbon emission prediction value is obtained. Specifically: a baseline carbon emission scalar is obtained based on the static feature vector and the baseline prediction branch; a carbon emission fluctuation scalar is obtained based on the fused dynamic feature vector and the fluctuation prediction branch; and the carbon emission prediction value is obtained based on the baseline carbon emission scalar and the carbon emission fluctuation scalar. Based on the expert gating weight distribution, the dynamic features in the dynamic feature matrix are sorted by contribution to obtain target dynamic features that meet the preset screening conditions. Carbon emission prediction results are generated based on the target's dynamic characteristics and the predicted carbon emission values.

[0006] This invention provides a method for predicting carbon emissions from power plants. Based on a pre-built carbon emission prediction model, it first obtains static feature vectors and dynamic feature matrices. The dynamic feature matrix is ​​input into a hybrid expert network to extract multi-dimensional features adapted to the power plant's operating state. The static feature vectors and dynamic feature matrix are then input into a gating network to generate an expert gating weight distribution that matches the unit's operating patterns. Based on this weight distribution, the hybrid expert output vector is fused to achieve accurate integration of features under different operating conditions. During carbon emission prediction, a baseline prediction branch and a fluctuation prediction branch are used to process the static feature vector and the fused dynamic feature vector respectively, separating the baseline and fluctuation components of carbon emissions. This avoids large base values ​​dominating the optimization process and ignoring small fluctuations, solving the technical problem of insufficient prediction accuracy caused by existing carbon emission prediction models using a single network that cannot simultaneously handle large base emissions and small fluctuations. This achieves real-time and accurate carbon emission prediction. Simultaneously, the contribution of dynamic features is ranked by the expert gating weight distribution to locate key dynamic features affecting carbon emissions. Finally, the target dynamic features are combined with the predicted carbon emission values ​​for output, providing an interpretable basis for carbon emission regulation while ensuring accurate prediction results.

[0007] Furthermore, the acquisition of static feature vectors and dynamic feature matrices within the current sliding time window based on real-time collected power plant operation data includes: Based on real-time collected power plant operation data, a set of static attribute features and an initial set of dynamic time-series features are obtained; Based on the static attribute feature set, an encoding process is performed to obtain a static feature vector; First-order difference calculation is performed based on the initial dynamic time series feature set to obtain the load differential feature sequence; Based on the load differential feature sequence, the initial dynamic time series feature set is expanded in dimension to obtain a dynamic time series feature set; The dynamic time-series feature set is subjected to sliding window truncation and standardization processing to obtain the dynamic feature matrix within the current sliding time window.

[0008] The above scheme converts discrete inherent unit attributes into continuous numerical vectors that the model can process by encoding the static attribute feature set, ensuring the effective representation of static features. By performing first-order difference calculation on the initial dynamic time series feature set, the load differential feature sequence is obtained and the dynamic feature dimension is expanded to capture the transient change trend of unit load, providing key physical features for subsequent operating condition discrimination and differentiated modeling. At the same time, by using sliding window truncation and standardization, the dimensional differences of dynamic time series features are eliminated, so that dynamic operating data of different dimensions can be adapted to model calculation. In addition, the sliding window processing method can effectively extract local features of time series data, improving the ability of the dynamic feature matrix to represent the real-time operating status of the unit.

[0009] Furthermore, it also includes: Based on the dynamic feature matrix, obtain the absolute value of the load and the differential characteristic of the load at the end of the current sliding time window; The current operating condition label is obtained based on the absolute value of the load, the differential characteristics of the load, and the preset operating condition judgment threshold, and the current operating condition label is used as an auxiliary result for carbon emission prediction.

[0010] The above scheme adds steps for acquiring and outputting operating condition tags. It extracts the absolute load value and differential load characteristics at the end of the current sliding time window from the dynamic feature matrix, and obtains the current operating condition tag by combining it with a preset operating condition judgment threshold. This tag is then used as an auxiliary result for carbon emission prediction. Since the absolute load value and differential load characteristics are the core basis for defining the thermodynamic operating conditions of the unit, the operating condition tag obtained from this can accurately reflect the current state of the unit, whether it is a low-load steady state, a high-load steady state, a rising load steady state, or a falling load steady state. This allows power plant operators to intuitively and quickly judge the real-time operating status of the unit. At the same time, the operating condition tag complements the carbon emission prediction value and target dynamic characteristics. Operators can combine the actual operating conditions of the unit to conduct a more comprehensive and physically accurate analysis of the causes of carbon emission changes, further enhancing the guiding value of the prediction results for the actual operation adjustment of the power plant.

[0011] Furthermore, the hybrid expert network includes a steady-state expert network and a transient expert network, and the step of obtaining the hybrid expert output vector based on the dynamic feature matrix and the hybrid expert network includes: The steady-state output result is obtained based on the dynamic feature matrix and the steady-state expert network in the hybrid expert network. The transient output result is obtained based on the dynamic feature matrix and the transient expert network in the hybrid expert network. The hybrid expert output vector is obtained based on the steady-state output result and the transient output result.

[0012] The above scheme sets up steady-state and transient expert networks in a hybrid expert network to perform differentiated modeling of the steady-state and transient operating characteristics of power plant carbon emissions. The steady-state expert network accurately fits the static mapping relationship of carbon emissions under steady-state conditions, while the transient expert network effectively captures the long-range time-series dependence features of carbon emissions under transient conditions. By combining the two types of output results, the fused hybrid expert output vector can comprehensively and accurately represent the carbon emission characteristics of the unit under different operating conditions, avoiding the problem of poor adaptability of a single network to different operating conditions. This makes the representation of dynamic features more consistent with the actual thermodynamic operating state of the unit, providing more accurate feature support for the subsequent weight allocation and feature fusion of the gating network, and further improving the accuracy of carbon emission prediction.

[0013] Further, the step of obtaining the steady-state output result based on the dynamic feature matrix and the steady-state expert network in the hybrid expert network includes: The steady-state expert network includes a low-load steady-state expert channel and a high-load steady-state expert channel; The dynamic feature matrix is ​​input into the low-load steady-state expert channel and the high-load steady-state expert channel in the steady-state expert network for parallel computation to obtain the low-load steady-state output result and the high-load steady-state output result, and then a steady-state output result is formed based on the low-load steady-state output result and the high-load steady-state output result.

[0014] The above-mentioned scheme splits the steady-state expert network into low-load steady-state expert channels and high-load steady-state expert channels. It can adaptively fit the corresponding static mapping relationship according to the carbon emission patterns of different load steady states. Compared with a single steady-state expert network, it can more accurately capture the differences in carbon emission characteristics under low and high load steady states, and make the steady-state output results more detailed and consistent in reflecting the carbon emission status of the unit in different steady-state load ranges, thus improving the accuracy of steady-state characteristic representation.

[0015] Further, the step of obtaining the transient output result based on the dynamic feature matrix and the transient expert network in the hybrid expert network includes: The hidden state of the dynamic feature matrix is ​​extracted based on the gated recurrent unit in the transient expert network to obtain the hidden state of each time step. For each time step: a linear transformation is performed based on the hidden layer state and preset transformation parameters to obtain linear hidden layer data, and a preset activation function is used to numerically map the linear hidden layer data to obtain an energy score. Then, an inner product is calculated based on the energy score and a preset context query vector to obtain the matching degree. The matching degree of all time steps is subjected to exponential transformation and normalization to obtain the attention weight distribution of time steps; The hidden layer states are weighted and summed based on the time-step attention weight distribution to obtain the transient output result.

[0016] The above scheme extracts the hidden state of the dynamic feature matrix through the gated recurrent unit in the transient expert network. Combining the dynamic features of the past and present, it fully captures the long-range temporal dependencies of carbon emission data under transient conditions. The hidden state of each time step is transformed linearly and mapped by the activation function to obtain the energy score. The inner product of the preset context query vector is used to calculate the initial determination of the importance of the features at each time step. Then, the matching degree of all time steps is subjected to exponential transformation and normalization to obtain the time step attention weight distribution that can accurately quantify the degree of influence of each time step on transient carbon emissions. Finally, the hidden state is weighted and summed based on the weight distribution to achieve focused extraction of key temporal node features in transient conditions. This effectively solves the prediction lag problem caused by thermal inertia and hysteresis effect in traditional models under transient conditions, so that the final transient output results can more accurately reflect the changing pattern of transient carbon emissions.

[0017] Further, the step of obtaining the transient output result based on the dynamic feature matrix and the transient expert network in the hybrid expert network includes: The transient expert network includes a load increase transient expert channel and a load decrease transient expert channel; The dynamic feature matrix is ​​input into the load increase transient expert channel and the load decrease transient expert channel in the transient expert network for parallel calculation to obtain the load increase transient output result and the load decrease transient output result, and then a transient output result is formed based on the load increase transient output result and the load decrease transient output result.

[0018] The above scheme sets up a transient expert network including two expert calculation channels: a load increase transient expert channel and a load decrease transient expert channel. The dynamic feature matrix is ​​input into the two channels respectively, and the transient output results of load increase and decrease are calculated in parallel and then fused to form the transient output result. The design of the two channels is in line with the asymmetric characteristics of thermal inertia and hysteresis effects during the actual unit load increase and decrease process. It can accurately capture the key temporal characteristics of carbon emissions under load increase and decrease transients. Compared with a single transient expert network, it can reflect the differences in carbon emission characteristics under different transient operating conditions in more detail and improve the accuracy of transient feature representation. At the same time, the parallel computing method ensures the computational efficiency of the model, so that the hybrid expert output vector can more accurately represent the carbon emission characteristics of various transient operating conditions of the unit.

[0019] Further, the step of ranking the dynamic features in the dynamic feature matrix by contribution based on the expert gating weight distribution to obtain target dynamic features that meet preset screening conditions includes: Based on the expert gating weight distribution, the time-step attention weight distribution of the load-increasing transient expert channel and the time-step attention weight distribution of the load-reducing transient expert channel are weighted and fused to obtain global attention; For each time step within the current sliding time window, a contribution sequence is obtained based on the dynamic feature matrix and global attention, and the target dynamic features in the dynamic feature matrix that meet the preset filtering conditions are output based on the contribution sequence.

[0020] The above scheme uses a weighted fusion of the attention weight distribution of the time steps in the transient expert channels for load increases and decreases based on the expert gating weight distribution to obtain global attention. By combining global attention with the weight allocation of transient experts by the gating network, it can accurately reflect the comprehensive impact of each time step on carbon emissions under different transient expert channels. Then, combined with the dynamic feature matrix within the current sliding time window, the contribution sequence of each dynamic feature is calculated based on global attention, and target dynamic features that meet the preset conditions are selected. This selection method is based on the actual prediction logic of the model, making the selected target dynamic features more correlated with carbon emission changes and more accurate. It can intuitively point out the key dynamic factors affecting carbon emissions for power plant operators, greatly improving the interpretability of the prediction results, and thus enabling the prediction results to more effectively guide the power plant's combustion strategy adjustment and energy-saving and emission-reduction operations.

[0021] Furthermore, the pre-construction process of the carbon emission prediction model includes: An initial carbon emission prediction model is constructed based on a pre-built hybrid expert network, gating network, and prediction output network. For any training sample in the preset power plant operation data training set, obtain the corresponding sample static feature vector and sample dynamic feature matrix; obtain the sample hybrid expert output vector based on the sample dynamic feature matrix and the hybrid expert network in the initial carbon emission prediction model; obtain the sample gating weight distribution based on the sample static feature vector, sample dynamic feature matrix and the gating network in the initial carbon emission prediction model; obtain the sample fusion dynamic feature vector based on the sample gating weight distribution and the sample hybrid expert output vector; and obtain the sample carbon emission prediction value based on the sample static feature vector, sample fusion dynamic feature vector and the prediction output network in the initial carbon emission prediction model. The mean square error is calculated based on the standardized real labels of the preset power plant operation data training set and the sample carbon emission prediction values ​​to obtain the single sample principal loss, and the principal loss is obtained based on the single sample principal loss of all the training samples. The single-sample gating entropy is calculated based on the sample gating weight distribution, and the single-sample gating entropy of all the training samples is summed to obtain the total entropy value. The total entropy value is inverted and normalized to obtain the entropy regularization loss; The joint loss function is obtained by weighted summation of the main loss and the entropy regularization loss. The initial carbon emission prediction model is updated by backpropagation based on the joint loss function until the preset parameter update stopping condition is met, thus obtaining the carbon emission prediction model.

[0022] The above scheme provides a pre-construction and training process for carbon emission prediction models. First, an initial carbon emission prediction model is built based on a hybrid expert network, a gating network, and a prediction output network. Then, samples from the power plant operation data training set are used to complete the model's forward computation to obtain sample carbon emission prediction values. The main loss is calculated by standardizing the mean square error between the real labels and the sample prediction values, ensuring the model can accurately fit the real data patterns of carbon emissions. Next, the single-sample gating entropy is calculated based on the sample gating weight distribution, and the summation and inversion normalization yields the entropy regularization loss. This encourages the gating network to fully utilize all expert networks among samples, avoiding the mode collapse problem commonly found in hybrid expert model training. Finally, the main loss and entropy regularization loss are weighted and summed to obtain a joint loss function. Based on this function, the initial model's parameters are backpropagated and updated until the stopping condition is met. This ensures that the trained model maintains carbon emission prediction accuracy while possessing good robustness and generalization ability, stably adapting to the complex and non-stationary actual operating data of power plants, ultimately achieving continuous and accurate prediction of carbon emissions.

[0023] This invention provides a power plant carbon emission prediction system, comprising a feature extraction module, a carbon emission model prediction module, a feature fusion module, and a result output module. The carbon emission model prediction module includes a hybrid expert output unit, a weight distribution calculation unit, and a carbon emission prediction unit, wherein: The feature extraction module is used to obtain static feature vectors and dynamic feature matrices within the current sliding time window based on real-time collected power plant operation data; The hybrid expert output unit is used to obtain the hybrid expert output vector based on the dynamic feature matrix and the hybrid expert network; The weight distribution calculation unit is used to obtain the expert gate weight distribution based on the static feature vector, dynamic feature matrix, and gated network; The feature fusion module is used to perform a weighted summation of the hybrid expert output vector based on the expert gating weight distribution to obtain a fused dynamic feature vector. The carbon emission prediction unit is used to obtain carbon emission prediction values ​​based on the static feature vector, the fused dynamic feature vector, and the prediction output network including a baseline prediction branch and a fluctuation prediction branch. Specifically: a baseline carbon emission scalar is obtained based on the static feature vector and the baseline prediction branch; a carbon emission fluctuation scalar is obtained based on the fused dynamic feature vector and the fluctuation prediction branch; and a carbon emission prediction value is obtained based on the baseline carbon emission scalar and the carbon emission fluctuation scalar. The result output module is used to sort the dynamic features in the dynamic feature matrix by contribution based on the expert gating weight distribution, obtain the target dynamic features that meet the preset screening conditions, and form a carbon emission prediction result based on the target dynamic features and the carbon emission prediction value.

[0024] This invention provides a power plant carbon emission prediction system. Through modular division of labor among a feature extraction module, a carbon emission model prediction module, a feature fusion module, and a result output module, it achieves standardized and regulated processing of the entire process, from real-time operation data processing and feature extraction to model prediction, feature fusion, key dynamic feature screening, and prediction result output. Furthermore, the carbon emission model prediction module is further subdivided into a hybrid expert output unit, a weight distribution calculation unit, and a carbon emission prediction unit, making the internal logic of the model prediction clearer. The collaborative cooperation between modules and units ensures the efficient and orderly execution of the carbon emission prediction process. Simultaneously, the modular architecture design gives the system good maintainability and scalability, allowing for functional adjustments and optimizations based on the actual operational needs of the power plant, achieving real-time, efficient monitoring and accurate prediction of power plant carbon emissions.

[0025] Another embodiment of the present invention provides a terminal device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein when the processor executes the computer program, it implements the steps of the power plant carbon emission prediction method of the present invention.

[0026] Another embodiment of the present invention provides a computer-readable storage medium item, including: a stored computer program, which, when the computer program is running, controls the device where the computer-readable storage medium is located to perform the steps of the power plant carbon emission prediction method of the present invention.

[0027] This invention provides a method and system for predicting carbon emissions from power plants. By performing refined static and dynamic decomposition and preprocessing on real-time collected power plant operation data, it extracts the load differential, a key physical feature reflecting changes in unit operating conditions. Combining this with the unit's thermodynamic operating laws, it constructs a heterogeneous hybrid expert network containing steady-state, transient, and variable load channels, enabling differentiated and refined modeling of carbon emission characteristics under different unit operating conditions. This effectively solves the prediction lag problem caused by thermal inertia and hysteresis effects in traditional models under transient conditions. Furthermore, it achieves adaptive weight allocation of the expert network through a gating network combined with static and dynamic features, and incorporates a joint loss with entropy regularization. The function-trained model ensures both prediction accuracy and avoids the mode collapse problem of hybrid expert models, thus improving the model's robustness and generalization ability. The dual-branch prediction architecture (baseline and fluctuation branches) decouples the prediction of large-scale carbon emission baselines from small fluctuation values, effectively solving the scale coupling problem of single networks and significantly improving the model's sensitivity to real-time operational adjustments and prediction accuracy. Simultaneously, this scheme combines attention weights and expert gating weights to complete the attribution analysis of key dynamic features and outputs real-time unit operating condition labels, making the prediction results both accurate and highly interpretable, providing decision-making guidance for power plant operators in line with the physical operation of the units. This invention not only provides accurate and reliable evidence for carbon quota verification but also provides comprehensive and effective data support for carbon trading markets, the formulation of energy-saving and emission-reduction strategies, and the optimization of unit combustion efficiency, comprehensively improving the intelligence and accuracy of power plant carbon emission monitoring and prediction. Attached Figure Description

[0028] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0029] Figure 1 This is a schematic diagram of a power plant carbon emission prediction method provided in this embodiment; Figure 2 This is a schematic diagram of a prediction output network provided in this embodiment; Figure 3 This is a schematic diagram of a carbon emission prediction method based on a phase space differential-guided hybrid expert and residual decoupling architecture provided in this embodiment; Among them: 01, benchmark prediction branch; 02, volatility prediction branch. Detailed Implementation

[0030] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0031] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the application; the terms “comprising” and “having”, and any variations thereof, in the specification, claims, and foregoing description of the drawings are intended to cover non-exclusive inclusion.

[0032] In the description of the embodiments of this application, technical terms such as "first" and "second" are used only to distinguish different objects and should not be construed as indicating or implying relative importance or implicitly specifying the number, specific order, or primary and secondary relationship of the indicated technical features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly defined.

[0033] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0034] In the description of the embodiments in this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship.

[0035] In the description of the embodiments of this application, the term "multiple" refers to two or more (including two), similarly, "multiple sets" refers to two or more (including two sets), and "multiple pieces" refers to two or more (including two pieces).

[0036] In the description of the embodiments of this application, unless otherwise expressly specified and limited, technical terms such as "installation," "connection," "joining," and "fixing" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. For those skilled in the art, the specific meaning of the above terms in the embodiments of this application can be understood according to the specific circumstances.

[0037] Example 1: This embodiment provides a method for predicting carbon emissions from power plants, such as... Figure 1 As shown, the method is applied to a pre-built carbon emission prediction model, which includes a pre-built hybrid expert network, a gating network, and a prediction output network. The method includes: S1. Obtain static feature vectors and dynamic feature matrices within the current sliding time window based on real-time collected power plant operation data; S2. Obtain the hybrid expert output vector based on the dynamic feature matrix and the hybrid expert network; S3. Obtain the expert gating weight distribution based on the static feature vector, dynamic feature matrix, and gating network; S4. Based on the expert gating weight distribution, the hybrid expert output vector is weighted and summed to obtain the fused dynamic feature vector; S5. Based on the static feature vector, the fused dynamic feature vector, and the prediction output network including the baseline prediction branch and the fluctuation prediction branch, the carbon emission prediction value is obtained. Specifically: a baseline carbon emission scalar is obtained based on the static feature vector and the baseline prediction branch; a carbon emission fluctuation scalar is obtained based on the fused dynamic feature vector and the fluctuation prediction branch; and the carbon emission prediction value is obtained based on the baseline carbon emission scalar and the carbon emission fluctuation scalar. S6. Based on the expert gating weight distribution, sort the dynamic features in the dynamic feature matrix by contribution to obtain target dynamic features that meet the preset screening conditions. S7. Based on the target dynamic characteristics and the carbon emission prediction value, a carbon emission prediction result is formed.

[0038] This embodiment provides a method for predicting carbon emissions from power plants. Based on a pre-built carbon emission prediction model, it first obtains static feature vectors and dynamic feature matrices. The dynamic feature matrix is ​​input into a hybrid expert network to extract multi-dimensional features adapted to the power plant's operating state. The static feature vectors and dynamic feature matrices are then input into a gating network to generate an expert gating weight distribution that matches the unit's operating patterns. Based on this weight distribution, the hybrid expert output vector is fused to achieve accurate integration of features under different operating conditions. During carbon emission prediction, the static feature vector and the fused dynamic feature vector are processed by a baseline prediction branch and a fluctuation prediction branch, respectively, separating the baseline and fluctuation components of carbon emissions. This avoids large base values ​​dominating the optimization process and ignoring small fluctuations, solving the technical problem that existing carbon emission prediction models using a single network cannot simultaneously account for large base emissions and small fluctuations, resulting in insufficient prediction accuracy. This achieves real-time and accurate carbon emission prediction. Simultaneously, the contribution of dynamic features is ranked by the expert gating weight distribution to locate key dynamic features affecting carbon emissions. Finally, the target dynamic features are combined with the predicted carbon emission values ​​for output, providing an interpretable basis for carbon emission regulation while ensuring the accuracy of the prediction results.

[0039] In the specific implementation process, this embodiment constructs a power plant carbon emission prediction method based on a phase space differential-guided hybrid expert and residual decoupling architecture. The method is built on the deep learning framework PyTorch 1.10, and the running environment is Ubuntu 20.04 operating system. The hardware platform configuration is a single NVIDIA GeForce RTX 3090 graphics card (24GB video memory) and an Intel Core i9-10900K processor.

[0040] Optionally, step S1 includes: Based on real-time collected power plant operation data, a set of static attribute features and an initial set of dynamic time-series features are obtained; Based on the static attribute feature set, an encoding process is performed to obtain a static feature vector; First-order difference calculation is performed based on the initial dynamic time series feature set to obtain the load differential feature sequence; Based on the load differential feature sequence, the initial dynamic time series feature set is expanded in dimension to obtain a dynamic time series feature set; The dynamic time-series feature set is subjected to sliding window truncation and standardization processing to obtain the dynamic feature matrix within the current sliding time window.

[0041] In the specific implementation process, the static attribute feature set includes unit type code, boiler design model, main fuel type code, and rated installed capacity, etc. One-Hot encoding is applied to the static attribute feature set to obtain the static feature vector. The initial dynamic time-series feature set includes real-time power generation load, main steam flow rate, inlet air temperature, flue gas temperature, and flue gas oxygen content, etc. The first-order difference of the real-time power generation load in the initial dynamic time-series feature set is calculated as the load differential feature. During sliding window truncation and standardization, the dynamic time-series feature set is first truncated using a sliding window, with the current time window length set to 60, and then standardized using the Z-Score method.

[0042] Specifically, this embodiment converts the raw sensor data of the power plant (i.e., the real-time collected power plant operation data) into a tensor format that the model can process, and extracts key physical guidance features. The physical partitioning and standardization preprocessing of this multimodal heterogeneous data specifically includes: First, real-time collection of all data from the power plant's historical operation database is used as the power plant operation data, and it is divided into a static attribute feature set and an initial dynamic time-series feature set. The static attribute feature set contains four fixed-dimensional physical attributes: unit type code, boiler design model, main fuel type code, and rated installed capacity; the initial dynamic time-series feature set contains five time-varying operation indicators: real-time power generation load, main steam flow rate, inlet air temperature, flue gas temperature, and flue gas oxygen content. Second, to capture the transient change trend of the unit, a first-order difference calculation is performed on the real-time power generation load sequence in the initial dynamic time-series feature set to obtain a load differential feature sequence. This feature reflects the rate and direction of load change and will serve as a key basis for subsequent steps in partitioning physical operating conditions. The load differential feature sequence is added to the initial dynamic time series feature set, and the final dynamic time series feature set is expanded to 6 dimensions.

[0043] Next, the static attribute feature set is encoded. One-hot encoding is used to convert discrete classification attributes into continuous numerical vectors, resulting in static feature vectors, which have a fixed dimension of 16 in this embodiment.

[0044] Finally, a sliding window truncation and standardization are performed on the dynamic time series feature set. The current time window length is set to 60, and the prediction step size is 1. For each time step within the window (… ) and each feature dimension ( The Z-Score standardization method is used to eliminate dimensional differences. The specific calculation formula is as follows: ; in, These are the standardized feature values ​​from the dynamic time-series feature set. For the original data in the dynamic time series feature set Time of the first The numerical value of the dimension; The first in the dynamic time series feature set The mean of the dimensional features; The first in the dynamic time series feature set The standard deviation of the dimensional features. After processing, a dynamic feature matrix is ​​obtained, with dimensions of... The standardized data is mapped using the hyperbolic tangent function (Tanh) to limit the numerical range. Between these intervals, to prevent gradient explosion.

[0045] Optional, also includes: Based on the dynamic feature matrix, obtain the absolute value of the load and the differential characteristic of the load at the end of the current sliding time window; The current operating condition label is obtained based on the absolute value of the load, the differential characteristics of the load, and the preset operating condition judgment threshold, and the current operating condition label is used as an auxiliary result for carbon emission prediction.

[0046] In the specific implementation process, this embodiment constructs a physical phase space state monitor, which generates soft labels using the load differential features and load absolute values ​​contained in the dynamic feature matrix to define physical operating condition standards. First, the physical operating condition discrimination logic is defined. Specifically, it is based on the load absolute value at the end of the current time window. and the load differential value (i.e., the load differential characteristic). The current power plant operating data being predicted is divided into four thermodynamic states, generating physical state labels. The four thermodynamic states specifically include: when... and When the unit is determined to be operating stably and in a low-load zone, combustion efficiency is mainly affected by the static load rate, and it is marked as a low-load equilibrium state (Label 0). and When the unit is at full capacity and its thermal parameters are stable, it is marked as being in a high-load equilibrium state (Label 1). During the load increase process, the boiler exhibits significant combustion hysteresis and thermal inertia, requiring close monitoring of its time history, and is marked as a positive dynamic transient (Label 2). The time period, i.e., the load reduction process, is marked as a negative dynamic transient (Label 3). The preset operating condition judgment thresholds include steady-state thresholds. And median load, median load = , Indicates the rated capacity, where, Set to 1% of rated capacity per minute. .

[0047] In practical applications, this embodiment employs an implicit physical guidance mechanism. By directly inputting the extracted load differential features as information into the dynamic feature matrix into the gating network, the gating network can adaptively discover physical laws and complete expert weight allocation under the optimization of the joint loss function. This is more consistent with the continuous characteristics of actual dynamic operating conditions than using discrete labels for forced supervision. The real role of these four types of discrete labels in the overall scheme of this embodiment is mainly reflected in the online deployment and real-time monitoring stages of the model. These labels will serve as discrete monitoring indicators reflecting the real-time thermodynamic state of the unit, and will be output synchronously with the feature attribution score, i.e., the carbon emission prediction result. This helps power plant operators intuitively and quickly determine the current physical operating condition of the unit, thereby comprehensively assisting in the adjustment of combustion strategies.

[0048] Optionally, the hybrid expert network includes a steady-state expert network and a transient expert network, and step S2 includes: The steady-state output result is obtained based on the dynamic feature matrix and the steady-state expert network in the hybrid expert network. The transient output result is obtained based on the dynamic feature matrix and the transient expert network in the hybrid expert network. The hybrid expert output vector is obtained based on the steady-state output result and the transient output result.

[0049] Optionally, obtaining the steady-state output result based on the dynamic feature matrix and the steady-state expert network in the hybrid expert network includes: The steady-state expert network includes a low-load steady-state expert channel and a high-load steady-state expert channel; The dynamic feature matrix is ​​input into the low-load steady-state expert channel and the high-load steady-state expert channel in the steady-state expert network for parallel computation to obtain the low-load steady-state output result and the high-load steady-state output result, and then a steady-state output result is formed based on the low-load steady-state output result and the high-load steady-state output result.

[0050] Optionally, obtaining the transient output result based on the dynamic feature matrix and the transient expert network in the hybrid expert network includes: The hidden state of the dynamic feature matrix is ​​extracted based on the gated recurrent unit in the transient expert network to obtain the hidden state of each time step. For each time step: a linear transformation is performed based on the hidden layer state and preset transformation parameters to obtain linear hidden layer data, and a preset activation function is used to numerically map the linear hidden layer data to obtain an energy score. Then, an inner product is calculated based on the energy score and a preset context query vector to obtain the matching degree. The matching degree of all time steps is subjected to exponential transformation and normalization to obtain the attention weight distribution of time steps; The hidden layer states are weighted and summed based on the time-step attention weight distribution to obtain the transient output result.

[0051] Optionally, obtaining the transient output result based on the dynamic feature matrix and the transient expert network in the hybrid expert network includes: The transient expert network includes a load increase transient expert channel and a load decrease transient expert channel; The dynamic feature matrix is ​​input into the load increase transient expert channel and the load decrease transient expert channel in the transient expert network for parallel calculation to obtain the load increase transient output result and the load decrease transient output result, and then a transient output result is formed based on the load increase transient output result and the load decrease transient output result.

[0052] In the specific implementation process, the physical state concept defined by the four types of discrete labels divided by the physical phase space state supervisor in this embodiment constitutes the physical theoretical basis for the targeted design of four heterogeneous expert network structures in this embodiment. The four heterogeneous expert network structures are the low-load steady-state expert channel, the high-load steady-state expert channel, the load-increasing transient expert channel, and the load-decreasing transient expert channel, including two steady-state models and two transient models, which respectively form steady-state expert networks and transient expert networks, and then form a hybrid expert network containing four expert network structures.

[0053] In the specific implementation process, this embodiment constructs heterogeneous expert networks and gating networks based on the above-mentioned physical state concept. In particular, an attention mechanism is introduced into the transient expert network to support subsequent attribution analysis. First, heterogeneous expert networks are constructed. This embodiment sets up four parallel expert networks: a low-load steady-state expert channel (expert 0), a high-load steady-state expert channel (expert 1), an increasing-load transient expert channel (expert 2), and a decreasing-load transient expert channel (expert 3). Among them, expert 0 and expert 1 (steady-state expert networks) adopt a lightweight multi-layer perceptron (MLP) structure, and the calculation process of the two steady-state expert network channels is shown in the following formula: ; In the formula, This represents the steady-state output of expert k in a steady-state expert network. This represents the lightweight multilayer perceptron of expert k. Represents the dynamic characteristic matrix; the low-load steady-state expert channel (expert 0) and the high-load steady-state expert channel (expert 1) are represented by the dynamic characteristic matrix within the current sliding time window, respectively. As input, the dynamic feature matrix is ​​first flattened, transforming it from a two-dimensional "time step × feature dimension" structure into a one-dimensional global feature vector. This eliminates the structural limitations of the time dimension, adapting to the one-dimensional input requirements of the MLP, allowing the MLP to directly extract global static features from time-series data. The MLP then autonomously learns and fits the static mapping relationship between dynamic features and carbon emissions within its corresponding steady-state load range (low and high load), performing parallel computation to obtain the outputs of two channels. Finally, each channel outputs its corresponding steady-state feature vector. (Low-load steady-state output results) and (High-load steady-state output results), together with the steady-state output results of the steady-state expert network.

[0054] In practice, Expert 0 and Expert 1 (steady-state experts) are completely identical in terms of network structure (both are MLPs) and processing flow for dynamic features. The difference lies in that they are two network branches with independently initialized parameters. Since the gating network generates preference assignments based on input features (such as different load levels) during training, these two independent MLPs will eventually converge to different parameter spaces. Therefore, even if the input is the same, their output feature representations will be significantly different, thus playing the role of adaptively fitting the static mapping relationship of different steady-state intervals (such as high-load steady-state and low-load steady-state).

[0055] Expert 2 and Expert 3 (Transient Expert Networks) employ a combination of Gated Recurrent Units (GRUs) and Temporal Attention mechanisms to not only capture long-range dependencies but also compute importance weights for each time step. The calculation process for both transient expert channels is as follows: First, hidden state extraction is performed: the hidden state sequence in the dynamic feature matrix is ​​extracted using GRU. Next, attention is calculated for each time step, and the specific calculation process is shown in the following formula: ; ; ; in, For the first The transient expert on the first Attention weights for each time step; For GRU in the The hidden state at each time step; The energy fraction of the hidden layer state; A pre-defined context query vector that is randomly initialized; The weight matrix and bias are learnable.

[0056] Specifically, the gated recurrent unit (GRU) reads the dynamic feature vectors from the dynamic feature matrix sequentially according to the time steps. The dynamic feature matrix contains 60 time steps, each containing 6 dimensions of dynamic features (real-time power generation load, main steam flow, inlet air temperature, flue gas temperature, flue gas oxygen content, and load differential features). At each time step t, the GRU combines the input features of the current moment... and the hidden state at the previous time step The internal update gate and reset gate mechanisms determine how much historical transient information to retain and how much new information from the current moment to introduce, ultimately iteratively calculating the hidden state at the current moment. .

[0057] Next, a single-layer feedforward neural network (or a single-layer fully connected network) with a hyperbolic tangent (tanh) activation function is used to calculate the energy fraction. Specifically, through a learnable weight matrix and bias terms For hidden state A linear spatial transformation is performed, followed by nonlinear compression using the tanh function to extract the temporal features from the GRU. This is mapped to a new importance assessment space. Because the weights of the carbon emission lag effect differ at different time steps (e.g., the instant the valve opens and the stage where the load stabilizes) during the transient process of unit load increases and decreases, the calculated... This is the absolute energy representation of the feature at that time step in this evaluation space, which will then be compared with the pre-defined context query vector learned independently by the experts. The inner product is then performed to calculate the relative importance of that time step within the entire time window, which is the final attention weight. Finally, based on the time-step attention weight distribution obtained for each transient expert channel, the hidden state obtained in that channel is summed in a weighted manner to obtain the transient output result of the current channel.

[0058] In practice, the basic network structures (GRU and time attention mechanism) of Expert 2 and Expert 3 (transient experts) are completely identical and their parameters are independent. The difference in their outputs depends on the fact that they learned different preset context query vectors during training, thus calculating different time step attention weights for the same time series. Because the thermal inertia and hysteresis effects of unit load increase (positive transient) and load decrease (negative transient) are significantly asymmetrical, these two transient experts will autonomously focus on different key nodes within the historical time window (i.e., the current sliding time window).

[0059] This embodiment also constructs a gated network (Router) in the carbon emission prediction model. The static feature vector... and dynamic feature matrix Statistical mean The concatenated data is input into a gating network, which consists of two fully connected layers and outputs the probability distribution of choices made by the four experts. (That is, the expert gating weight distribution), the specific calculation formula is shown in the following formula: ; Will and After splicing, the multilayer sensor passes through a gating network. After performing a linear transformation, the transformed result is normalized using the Softmax function, ultimately generating the expert gating weight distribution G for weighted fusion of the output feature vectors of each expert network.

[0060] Specifically, the gating network first concatenates the input static feature vector with the statistical mean of the dynamic feature matrix containing parameters such as load differential over time, thereby fusing the unit's basic inherent attributes with the global dynamic operating information within the current time window. Then, the concatenated fused features are input into the gating network, which consists of fully connected layers, to map the original evaluation scores of the four expert networks. Finally, the original scores are normalized using the Softmax activation function, and the expert selection probability distribution G, with a sum of 1, is calculated, thus completing the allocation of weights for each expert. In this embodiment, from the perspective of the underlying mathematical logic of end-to-end training, the generation of G does not directly introduce the previously generated operating condition labels as hard supervision signals to calculate the loss. However, at the physical feature guidance level, there is a deep implicit correlation between the two. Since the dynamic features received by the gating network explicitly include the load differential feature sequence, and the load differential features are the core discrimination criterion for classifying the four types of physical operating condition labels. Therefore, under the joint loss optimization of the entire model, the gated network can spontaneously capture the rate of change of operating conditions in the features, adaptively discover the thermodynamic laws of the system, and realize the soft routing allocation of weights. This design in this embodiment makes the expert selection probability distribution G highly consistent with the evolution of the four types of physical operating conditions in terms of business logic, while avoiding the gradient differentiation discontinuity problem caused by forced intervention using discrete hard labels. In summary, this embodiment relies on the soft routing mechanism of the gated network for implicit guidance. Since the input of the gated network incorporates dynamic features such as load differential, under the joint optimization of the mean square error main loss and sparse entropy regularization loss, the network will adaptively learn the allocation of expert weights according to the dynamic rate of change of the data, enabling the carbon emission prediction model to spontaneously discover and approximate the optimal solution of the four types of physical operating conditions. This soft guidance design based on feature input avoids the discontinuous differentiation problem caused by artificially setting hard thresholds to divide the data, and ensures the smoothness and stability of the output weights of the model in different operating condition transition stages (such as the gradual process from steady state to transient state).

[0061] Finally, weighted fusion is performed based on G, utilizing gating weights. Feature vectors output by each expert The weighted summation yields the final dynamic feature representation vector (i.e., the fused dynamic feature vector). The specific formula is as follows: .

[0062] In the specific implementation process, the prediction output network (i.e., the prediction output model) in the carbon emission prediction model constructed in this embodiment contains two independent branches, as shown in the example below. Figure 2As shown, a dual-path parallel architecture is used to predict the baseline and fluctuation values ​​of carbon emissions separately, and then physical fusion is performed to address the problem that a single network cannot simultaneously fit large baseline values ​​and small fluctuation values, achieving prediction output based on a physical residual decoupling architecture. Specifically, the baseline prediction branch 01 only receives static feature vectors. The baseline carbon emission scalar is directly mapped and output through the MLP network. The fluctuation prediction branch 02 only receives fused dynamic feature vectors. Carbon emission fluctuation scalar is output through MLP network mapping. The final prediction result is the algebraic sum of the two results, followed by inverse normalization to obtain the predicted carbon emissions. This involves performing residual synthesis and inverse normalization, adding the baseline value and the fluctuation value, and converting them back to the original physical dimensions to obtain the final prediction result. The specific calculation process is shown in the following formula: ; ; in, Standardized predicted values; The final output is a carbon emission forecast in "tons per hour"; and These are the standard deviation and mean of the label data, respectively.

[0063] Specifically, the baseline prediction branch 01 only receives static feature vectors. As input, it passes through an MLP network containing three fully connected layers (with 16-32-1 nodes respectively), such as Figure 2 As shown, the baseline prediction branch 01 specifically includes fully connected layer 11, fully connected layer 12, fully connected layer 13, and output layer 14, which directly maps to the output baseline carbon emission scalar. The physical meaning of this value is the theoretical average emission of this type of unit under standard operating conditions; the fluctuation prediction branch 02 only receives the fused dynamic feature vector. As input, it passes through another MLP network containing three fully connected layers (with 64-32-1 nodes respectively), such as Figure 2 As shown, the fluctuation prediction branch 02 specifically includes fully connected layers 21, 22, and 23, and an output layer 24, which maps and outputs a carbon emission fluctuation scalar. The physical meaning of this value is the increase or decrease in emissions caused by the current operating parameters relative to standard operating conditions.

[0064] Optionally, step S6 includes: Based on the expert gating weight distribution, the time-step attention weight distribution of the load-increasing transient expert channel and the time-step attention weight distribution of the load-reducing transient expert channel are weighted and fused to obtain global attention; For each time step within the current sliding time window, a contribution sequence is obtained based on the dynamic feature matrix and global attention, and the target dynamic features in the dynamic feature matrix that meet the preset filtering conditions are output based on the contribution sequence.

[0065] In the specific implementation process, this embodiment adds attribution analysis and inference to the real-time carbon emission monitoring process. It utilizes a trained carbon emission prediction model for online inference and outputs key influencing factors. Specifically, real-time data access and processing are performed. Every minute, the latest static configuration and dynamic operating window data are read from the distributed control system (DCS system) and processed according to standardized parameters (…). The data undergoes preprocessing to obtain static feature vectors and dynamic feature matrices within the current sliding time window. Next, forward propagation is performed. The processed data is input into the model, sequentially passing through a gating network, a dynamic hybrid expert network, and a residual decoupled prediction head to obtain the current carbon emission prediction value. .

[0066] Finally, key feature attribution analysis is performed. The expert gating weight distribution calculated by the transient expert network is extracted. Global attention is obtained by weighting the attention weights of all experts. : ; Then, the contribution of each feature at each time step within the current window to the result is calculated, resulting in a contribution sequence. The calculation process for each time step is shown in the following formula: ; in, For the first A dynamic feature at time step The standardized value. Normalized sorting is performed, and the target dynamic characteristics (such as "abnormal fluctuation of air inlet temperature") that meet the preset screening conditions (such as the top 3) are output as the basis for interpreting carbon emission changes. Combined with the carbon emission prediction value, the final carbon emission prediction result is formed to assist operators in making combustion adjustments.

[0067] Optionally, the pre-construction process of the carbon emission prediction model includes: An initial carbon emission prediction model is constructed based on a pre-built hybrid expert network, gating network, and prediction output network. For any training sample in the preset power plant operation data training set, obtain the corresponding sample static feature vector and sample dynamic feature matrix; obtain the sample hybrid expert output vector based on the sample dynamic feature matrix and the hybrid expert network in the initial carbon emission prediction model; obtain the sample gating weight distribution based on the sample static feature vector, sample dynamic feature matrix and the gating network in the initial carbon emission prediction model; obtain the sample fusion dynamic feature vector based on the sample gating weight distribution and the sample hybrid expert output vector; and obtain the sample carbon emission prediction value based on the sample static feature vector, sample fusion dynamic feature vector and the prediction output network in the initial carbon emission prediction model. The mean square error is calculated based on the standardized real labels of the preset power plant operation data training set and the sample carbon emission prediction values ​​to obtain the single sample principal loss, and the principal loss is obtained based on the single sample principal loss of all the training samples. The single-sample gating entropy is calculated based on the sample gating weight distribution, and the single-sample gating entropy of all the training samples is summed to obtain the total entropy value. The total entropy value is inverted and normalized to obtain the entropy regularization loss; The joint loss function is obtained by weighted summation of the main loss and the entropy regularization loss. The initial carbon emission prediction model is updated by backpropagation based on the joint loss function until the preset parameter update stopping condition is met, thus obtaining the carbon emission prediction model.

[0068] In the specific implementation process, during the model training phase, the preset power plant operation data training set is constructed using historical power plant operation data. A validation set is also constructed simultaneously, and its content is consistent with the real-time power plant operation data collected earlier. The data processing process in the initial carbon emission prediction model is completely identical to that of the training carbon emission prediction model, with differences only in model parameters (various preset hyperparameters, etc.). This embodiment constructs a joint loss function incorporating sparse entropy regularization to simultaneously optimize prediction accuracy and the clarity of expert division of labor. The joint loss function constructed in this embodiment... Main loss due to mean square error and entropy regularization loss Composition. Among them: Main loss function Entropy regularization loss is used to measure the difference between predicted and actual values. Used to encourage gating vectors Maintain diversity among samples to prevent the model from falling into pattern collapse.

[0069] First, the main loss function is used. The main loss of the prediction task is calculated, and the mean squared error is used to measure the difference between the predicted carbon emissions of the sample and the standardized true labels: ; in, Set the training batch size to 64; For the first The standardized prediction output of each sample is the predicted carbon emissions of the sample. For the first Each sample is standardized with a real label.

[0070] Secondly, the entropy regularization loss of the gated network is calculated using the entropy regularization loss function. To prevent the model from falling into "mode collapse" (i.e., all samples use only the same expert), an entropy regularization term is introduced to encourage the gating vectors to be regularized. To maintain diversity across samples but sparsity within a single sample (i.e., explicitly selecting a few experts for a single sample), the specific calculation formula is as follows: ; in, For the first The nth sample pair Gating weights for each expert; To prevent overflow during logarithmic calculations, take the minimum value. .

[0071] Finally, calculate the joint loss function. : ; in, To balance the hyperparameters, they are set to 0.01. This embodiment uses the AdamW optimizer, setting the initial learning rate to 0.001 and the weight decay coefficient to 0.01, based on... Backpropagation is performed on all parameters of the model until the validation set loss no longer decreases, which meets the preset parameter update stopping condition, thus obtaining the carbon emission prediction model.

[0072] This embodiment provides a method for predicting carbon emissions from power plants. It proposes a heterogeneous expert modeling mechanism guided by physical phase space differentials. By introducing load differential characteristics to divide steady-state and transient operating conditions, a heterogeneous expert structure combining steady-state MLP and transient gated cyclic units with a time attention mechanism is designed to form a hybrid expert network. This allows the model to automatically switch computational logic according to the unit's operating status. Furthermore, the time attention mechanism in the transient expert effectively solves the prediction lag problem caused by thermal inertia in traditional models during load changes. This embodiment also designs a residual decoupling architecture based on physical priors, decomposing the prediction task into a DC reference component determined by static attributes and an AC fluctuation component determined by dynamic parameters. Utilizing the physical law of large base values ​​and small fluctuations in carbon emission data, this avoids the problem of the model ignoring minute dynamic changes due to excessively large base values ​​during optimization, significantly improving the model's sensitivity to real-time operational adjustments. This embodiment also introduces a sparse entropy regularization training strategy. By adding a gated entropy regularization term to the loss function, the model is forced to make full use of all expert networks during training, avoiding the pattern collapse problem of "the strong get stronger and the weak get weaker" common in hybrid expert models, and ensuring the robustness and generalization ability of the model when dealing with complex non-stationary time series data.

[0073] Example 2: In practical applications, power plants currently rely primarily on two approaches to acquire carbon emission data: hardware-based Continuous Emission Monitoring Systems (CEMS) and data-driven soft measurement technologies (i.e., carbon emission prediction models). Although CEMS systems are widely deployed, in actual industrial settings, hardware often faces challenges such as corrosion from high-temperature and high-humidity environments, probe clogging, and standard gas drift, leading to data loss or large measurement errors, and incurring high equipment maintenance costs. Therefore, using artificial intelligence algorithms for soft measurement based on unit operating data (such as load, coal consumption, and air volume) has become a current research hotspot.

[0074] Existing carbon emission prediction methods primarily rely on time-series deep learning models. While Mixture of Experts (MoE) and Residual Network (ResNet) have shown promise in relevant time-series prediction tasks, their direct application to carbon emission data suffers from methodological flaws. First, existing MoE models lack explicit guidance mechanisms for the physical phase space. Traditional MoEs typically employ blindly gating networks to select experts, ignoring the physical nature of unit operating states. Specifically, the thermodynamic equilibrium (steady-state) and non-equilibrium kinetic (transient) states differ significantly in combustion mechanisms. Steady-state processes are primarily characterized by static mapping, while transient processes (such as load changes) are dominated by thermal inertia and hysteresis. If the model cannot differentiate between these two operating conditions based on the load change rate (differential characteristic) and assign them appropriately to experts with different structures (such as MLP or GRU), prediction lag occurs during drastic load transitions, and the model cannot reasonably explain the selection of a particular expert. Second, existing prediction architectures struggle to address the scale coupling problem of large base values ​​with small fluctuations. The total carbon emissions from power plants are typically enormous (hundreds of tons per hour), while emission fluctuations caused by operational adjustments are relatively small (several tons per hour). When a single network attempts to fit data of these two magnitudes simultaneously, it is often dominated by gradients with a large base, resulting in a loss of ability to capture small fluctuations. Although some studies have attempted to use residual connections, the lack of explicit decoupling designs based on physical meaning (baseline value vs. fluctuation value) makes it difficult to achieve truly decoupled predictions. Furthermore, the loss functions of existing models lack effective constraints on expert diversity. During MoE training, mode collapse often occurs, where gating networks tend to use only the strongest expert, leaving other experts idle and losing the advantages of ensemble learning. In summary, there is an urgent need for a carbon emission prediction method that can guide expert division of labor using load differential characteristics, model steady-state and transient states separately through heterogeneous networks, and achieve decoupling between baseline and fluctuation using a dual-path residual architecture.

[0075] Therefore, based on Example 1, this example provides a carbon emission prediction method based on a phase space differential-guided hybrid expert and residual decoupling architecture, such as... Figure 3 As shown, it includes the following steps: S21. Collect power plant operation data, divide it into a static attribute feature set and a dynamic time series feature set, calculate the load differential characteristics, and perform encoding and standardization preprocessing respectively. S22. Construct a physical phase space state monitor, generate physical condition labels using load differential characteristics and absolute load levels, and define physical discrimination criteria between steady state and transient state. S23. Construct a phase space differential-oriented physical hybrid expert network, set up heterogeneous steady-state experts and transient experts, introduce a time attention mechanism in the transient experts, use a gating network to generate expert weights based on static and dynamic features, and perform weighted fusion of the output features of each expert. S24. Construct a prediction output module based on a physical residual decoupling architecture, establish parallel baseline prediction branch 01 and fluctuation prediction branch 02 to predict the theoretical baseline value and emission fluctuation value respectively, and finally obtain the final prediction result through residual synthesis. S25. Construct a joint loss function that includes sparse entropy regularization. While minimizing the prediction error, use the entropy regularization term to prevent expert mode collapse and train the model end-to-end. S26. Real-time monitoring of carbon emissions is performed based on the trained model, and key feature attribution analysis is conducted using attention weights.

[0076] This embodiment relates to the fields of artificial intelligence and power plant carbon emission monitoring technology, and solves the technical problems of existing hybrid expert models lacking physical condition guidance in carbon emission prediction, resulting in poor ability to capture transient hysteresis effects, and single networks being unable to balance the prediction accuracy of large base emission values ​​and small fluctuation values, and being prone to model collapse.

[0077] Example 3: This embodiment provides a power plant carbon emission prediction system, including a feature extraction module, a carbon emission model prediction module, a feature fusion module, and a result output module. The carbon emission model prediction module includes a hybrid expert output unit, a weight distribution calculation unit, and a carbon emission prediction unit, wherein: The feature extraction module is used to obtain static feature vectors and dynamic feature matrices within the current sliding time window based on real-time collected power plant operation data; The hybrid expert output unit is used to obtain the hybrid expert output vector based on the dynamic feature matrix and the hybrid expert network; The weight distribution calculation unit is used to obtain the expert gate weight distribution based on the static feature vector, dynamic feature matrix, and gated network; The feature fusion module is used to perform a weighted summation of the hybrid expert output vector based on the expert gating weight distribution to obtain a fused dynamic feature vector. The carbon emission prediction unit is used to obtain carbon emission prediction values ​​based on the static feature vector, the fused dynamic feature vector, and the prediction output network including the baseline prediction branch 01 and the fluctuation prediction branch 02. Specifically: a baseline carbon emission scalar is obtained based on the static feature vector and the baseline prediction branch 01; a carbon emission fluctuation scalar is obtained based on the fused dynamic feature vector and the fluctuation prediction branch 02; and a carbon emission prediction value is obtained based on the baseline carbon emission scalar and the carbon emission fluctuation scalar. The result output module is used to sort the dynamic features in the dynamic feature matrix by contribution based on the expert gating weight distribution, obtain the target dynamic features that meet the preset screening conditions, and form a carbon emission prediction result based on the target dynamic features and the carbon emission prediction value.

[0078] It should be noted that the device embodiments described above are merely illustrative, and some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Furthermore, in the accompanying drawings of the device embodiments provided by this invention, the connection relationships between modules indicate that they have communication connections, which can specifically be implemented as one or more communication buses or signal lines. Those skilled in the art can understand and implement this without any creative effort.

[0079] Example 4: Based on the above embodiments of the power plant carbon emission prediction method, another embodiment of the present invention provides a terminal device, which includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the power plant carbon emission prediction method of any embodiment of the present invention.

[0080] For example, in this embodiment, the computer program can be divided into one or more modules, which are stored in the memory and executed by the processor to complete the present invention. The one or more modules may be a series of computer program instruction segments capable of performing a specific function, which describe the execution process of the computer program in the terminal device.

[0081] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.

[0082] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the terminal device, connecting all parts of the terminal device via various interfaces and lines.

[0083] Based on the above-described method embodiments, another embodiment of the present invention provides a computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to execute the power plant carbon emission prediction method described in any of the above-described method embodiments of the present invention.

[0084] The modules / units integrated in the device / terminal equipment, if implemented as software functional units and sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.

[0085] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting carbon emissions from power plants, characterized in that, The method, applied to a pre-built carbon emission prediction model, which includes a pre-built hybrid expert network, a gating network, and a prediction output network, comprises: Static feature vectors and dynamic feature matrices within the current sliding time window are obtained based on real-time collected power plant operation data. The hybrid expert output vector is obtained based on the dynamic feature matrix and the hybrid expert network. The expert gating weight distribution is obtained based on the static feature vector, dynamic feature matrix, and gating network. The hybrid expert output vector is weighted and summed based on the expert gating weight distribution to obtain a fused dynamic feature vector. Based on the static feature vector, the fused dynamic feature vector, and the prediction output network including the baseline prediction branch and the fluctuation prediction branch, the carbon emission prediction value is obtained. Specifically: a baseline carbon emission scalar is obtained based on the static feature vector and the baseline prediction branch; a carbon emission fluctuation scalar is obtained based on the fused dynamic feature vector and the fluctuation prediction branch; and the carbon emission prediction value is obtained based on the baseline carbon emission scalar and the carbon emission fluctuation scalar. Based on the expert gating weight distribution, the dynamic features in the dynamic feature matrix are sorted by contribution to obtain target dynamic features that meet the preset screening conditions. Carbon emission prediction results are generated based on the target's dynamic characteristics and the predicted carbon emission values.

2. The method for predicting carbon emissions from power plants as described in claim 1, characterized in that, The process of obtaining static feature vectors and dynamic feature matrices within the current sliding time window based on real-time collected power plant operation data includes: Based on real-time collected power plant operation data, a set of static attribute features and an initial set of dynamic time-series features are obtained; Based on the static attribute feature set, an encoding process is performed to obtain a static feature vector; First-order difference calculation is performed based on the initial dynamic time series feature set to obtain the load differential feature sequence; Based on the load differential feature sequence, the initial dynamic time series feature set is expanded in dimension to obtain a dynamic time series feature set; The dynamic time-series feature set is subjected to sliding window truncation and standardization processing to obtain the dynamic feature matrix within the current sliding time window.

3. The method for predicting carbon emissions from power plants as described in claim 1, characterized in that, Also includes: Based on the dynamic feature matrix, obtain the absolute value of the load and the differential characteristic of the load at the end of the current sliding time window; The current operating condition label is obtained based on the absolute value of the load, the differential characteristics of the load, and the preset operating condition judgment threshold, and the current operating condition label is used as an auxiliary result for carbon emission prediction.

4. The method for predicting carbon emissions from power plants as described in claim 1, characterized in that, The hybrid expert network includes a steady-state expert network and a transient expert network. Obtaining the hybrid expert output vector based on the dynamic feature matrix and the hybrid expert network includes: The steady-state output result is obtained based on the dynamic feature matrix and the steady-state expert network in the hybrid expert network. The transient output result is obtained based on the dynamic feature matrix and the transient expert network in the hybrid expert network. The hybrid expert output vector is obtained based on the steady-state output result and the transient output result.

5. The method for predicting carbon emissions from power plants as described in claim 4, characterized in that, The step of obtaining steady-state output results based on the dynamic feature matrix and the steady-state expert network in the hybrid expert network includes: The steady-state expert network includes a low-load steady-state expert channel and a high-load steady-state expert channel; The dynamic feature matrix is ​​input into the low-load steady-state expert channel and the high-load steady-state expert channel in the steady-state expert network for parallel computation to obtain the low-load steady-state output result and the high-load steady-state output result, and then a steady-state output result is formed based on the low-load steady-state output result and the high-load steady-state output result.

6. The method for predicting carbon emissions from power plants as described in claim 4, characterized in that, The process of obtaining transient output results based on the dynamic feature matrix and the transient expert network in the hybrid expert network includes: The hidden state of the dynamic feature matrix is ​​extracted based on the gated recurrent unit in the transient expert network to obtain the hidden state of each time step. For each time step: a linear transformation is performed based on the hidden layer state and preset transformation parameters to obtain linear hidden layer data, and a preset activation function is used to numerically map the linear hidden layer data to obtain an energy score. Then, an inner product is calculated based on the energy score and a preset context query vector to obtain the matching degree. The matching degree of all time steps is subjected to exponential transformation and normalization to obtain the attention weight distribution of time steps; The hidden layer states are weighted and summed based on the time-step attention weight distribution to obtain the transient output result.

7. The method for predicting carbon emissions from power plants as described in claim 6, characterized in that, The process of obtaining transient output results based on the dynamic feature matrix and the transient expert network in the hybrid expert network includes: The transient expert network includes a load increase transient expert channel and a load decrease transient expert channel; The dynamic feature matrix is ​​input into the load increase transient expert channel and the load decrease transient expert channel in the transient expert network for parallel calculation to obtain the load increase transient output result and the load decrease transient output result, and then a transient output result is formed based on the load increase transient output result and the load decrease transient output result.

8. The method for predicting carbon emissions from power plants as described in claim 7, characterized in that, The step of ranking the dynamic features in the dynamic feature matrix by contribution based on the expert gating weight distribution to obtain target dynamic features that meet preset screening conditions includes: Based on the expert gating weight distribution, the time-step attention weight distribution of the load-increasing transient expert channel and the time-step attention weight distribution of the load-reducing transient expert channel are weighted and fused to obtain global attention; For each time step within the current sliding time window, a contribution sequence is obtained based on the dynamic feature matrix and global attention, and the target dynamic features in the dynamic feature matrix that meet the preset filtering conditions are output based on the contribution sequence.

9. The method for predicting carbon emissions from power plants as described in claim 1, characterized in that, The pre-construction process of the carbon emission prediction model includes: An initial carbon emission prediction model is constructed based on a pre-built hybrid expert network, gating network, and prediction output network. For any training sample in the preset power plant operation data training set, obtain the corresponding sample static feature vector and sample dynamic feature matrix; obtain the sample hybrid expert output vector based on the sample dynamic feature matrix and the hybrid expert network in the initial carbon emission prediction model; obtain the sample gating weight distribution based on the sample static feature vector, sample dynamic feature matrix and the gating network in the initial carbon emission prediction model; obtain the sample fusion dynamic feature vector based on the sample gating weight distribution and the sample hybrid expert output vector; and obtain the sample carbon emission prediction value based on the sample static feature vector, sample fusion dynamic feature vector and the prediction output network in the initial carbon emission prediction model. The mean square error is calculated based on the standardized real labels of the preset power plant operation data training set and the sample carbon emission prediction values ​​to obtain the single sample principal loss, and the principal loss is obtained based on the single sample principal loss of all the training samples. The single-sample gating entropy is calculated based on the sample gating weight distribution, and the single-sample gating entropy of all the training samples is summed to obtain the total entropy value. The total entropy value is inverted and normalized to obtain the entropy regularization loss; The joint loss function is obtained by weighted summation of the main loss and the entropy regularization loss. The initial carbon emission prediction model is updated by backpropagation based on the joint loss function until the preset parameter update stopping condition is met, thus obtaining the carbon emission prediction model.

10. A power plant carbon emission prediction system, characterized in that, The system includes a feature extraction module, a carbon emission model prediction module, a feature fusion module, and a result output module. The carbon emission model prediction module includes a hybrid expert output unit, a weight distribution calculation unit, and a carbon emission prediction unit, wherein: The feature extraction module is used to obtain static feature vectors and dynamic feature matrices within the current sliding time window based on real-time collected power plant operation data; The hybrid expert output unit is used to obtain the hybrid expert output vector based on the dynamic feature matrix and the hybrid expert network; The weight distribution calculation unit is used to obtain the expert gate weight distribution based on the static feature vector, dynamic feature matrix, and gated network; The feature fusion module is used to perform a weighted summation of the hybrid expert output vector based on the expert gating weight distribution to obtain a fused dynamic feature vector. The carbon emission prediction unit is used to obtain carbon emission prediction values ​​based on the static feature vector, the fused dynamic feature vector, and the prediction output network including a baseline prediction branch and a fluctuation prediction branch. Specifically: a baseline carbon emission scalar is obtained based on the static feature vector and the baseline prediction branch; a carbon emission fluctuation scalar is obtained based on the fused dynamic feature vector and the fluctuation prediction branch; and a carbon emission prediction value is obtained based on the baseline carbon emission scalar and the carbon emission fluctuation scalar. The result output module is used to sort the dynamic features in the dynamic feature matrix by contribution based on the expert gating weight distribution, obtain the target dynamic features that meet the preset screening conditions, and form a carbon emission prediction result based on the target dynamic features and the carbon emission prediction value.

Images