Exponential smoothing and hybrid expert-based carbon emission prediction method and device
By employing parallel tensor exponential moving average decomposition and multi-scenario techniques, the modeling challenges of long-term trends and high-frequency fluctuations in gas turbine carbon emission prediction were solved, resulting in more accurate predictions.
Patent Information
- Application Number
- CN202610528025.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-04-21
- Publication Date
- 2026-06-09
AI Technical Summary
Existing prediction methods struggle to accurately model both the long-term trends and high-frequency fluctuations of gas turbine carbon emissions under complex operating conditions, thus limiting their predictive performance.
A carbon emission prediction method based on exponential smoothing and hybrid expert is adopted. The trend component and seasonal component are decoupled by a parallel tensor exponential moving average decomposition module, a multi-scale trend pyramid is constructed for prediction, and a time-frequency sensing residual hybrid expert module is used for adaptive routing to reconstruct the final prediction result.
It achieves accurate capture of carbon emission trends and adaptive decoupling of high-frequency fluctuations, improving the stability and accuracy of predictions and maintaining high-efficiency prediction performance under complex operating conditions.
Smart Images
Figure CN122175096A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of carbon emission and industrial time series prediction, and in particular to a carbon emission prediction method and apparatus based on exponential smoothing and hybrid expert methods, specifically applied to long-term high-precision prediction of carbon emissions and related operating parameters. Background Technology
[0002] As a key power unit in thermal power plants, the operating status of gas turbines directly affects the emission levels of pollutants such as carbon monoxide (CO) and nitrogen oxides (NOx). Therefore, accurately predicting the multi-dimensional system status of gas turbines, including carbon emissions, over a future period is of great value for predictive maintenance and optimized combustion control.
[0003] However, multivariate time series data of carbon emissions are significantly complex: First, the series contains smoothed trends such as equipment aging and long-term load scheduling, as well as high-frequency residuals caused by combustion chamber disturbances, and the two are highly coupled; Second, there are heterogeneous dynamic patterns in the high-frequency residuals (e.g., periodic oscillations, sudden nonlinear fluctuations, and pure sensor noise), which are often difficult to model uniformly using existing prediction techniques.
[0004] Existing decomposition methods based on fixed-window moving averages are not sensitive enough to recent changes, while conventional prediction models lack a mechanism to adaptively select processing logic based on input state, resulting in limited prediction performance under complex conditions. Summary of the Invention
[0005] This invention provides a carbon emission prediction method and apparatus based on exponential smoothing and hybrid expert methods. The invention aims to address the technical problem that existing prediction methods, when processing gas turbine carbon emission data, struggle to accurately model the long-term carbon emission trend caused by load changes, and adaptively analyze high-frequency carbon emission fluctuations (residuals) caused by local combustion disturbances. Furthermore, they struggle to uniformly model the heterogeneous dynamic patterns in the carbon emission generation process, thus limiting the performance of carbon emission prediction under complex operating conditions. Details are described below:
[0006] A first aspect is a carbon emission prediction method based on exponential smoothing and hybrid expert methods, the method comprising:
[0007] The preprocessed gas turbine multivariate time series data sequence is decoupled into trend components and seasonal components using a parallel tensor-based exponential moving average decomposition module.
[0008] A multi-scale trend pyramid is constructed to predict trend components and obtain trend prediction results; a dual-branch representation structure is used to perform deep feature encoding on seasonal components to obtain cross-scale context representation based on local fine-grained dynamic features.
[0009] Adaptive route prediction is performed using a time-frequency sensing residual hybrid expert module based on local fine-grained dynamic features and cross-scale context representation. The gating probabilities of each expert are obtained, and the seasonal component prediction results are obtained based on the gating probabilities and seasonal component prediction.
[0010] The prediction results are reconstructed based on the seasonal component prediction results and the trend prediction results to obtain the final multivariate prediction results. The final multivariate prediction results are then output to the carbon emission control device, and feedforward emission reduction control of gas turbine carbon emissions is achieved by adjusting the combustion parameters in advance.
[0011] The parallel tensor-based exponential moving average decomposition module is as follows:
[0012] Construct the decay tensor: ,in, , It is a smoothing factor;
[0013] Define the weight tensor:
[0014]
[0015] Based on this, the trend component is calculated as follows:
[0016]
[0017] in This represents element-wise multiplication. This indicates a prefix accumulation operation. Let be the weight tensor and D be the decay tensor.
[0018] The multi-scale trend pyramid is as follows:
[0019] For trend components Includes low-frequency macroscopic change information, set Different scale levels, corresponding to scale indexes , of which The downsampling factor for each scale is :
[0020] For the Each scale, the trend component is represented as:
[0021]
[0022] in, For the first Trend components at each scale are represented. This is a one-dimensional average pooling operation; The pooling kernel size, The pooling step size;
[0023] The final trend prediction result is expressed as follows:
[0024]
[0025] in, For linear interpolation operations, For the first Prediction results at various scales Where S is the interpolation target length, and S is the total number of decomposition scale levels. For globally learnable weight vectors, It is a normalized exponential function.
[0026] The method includes processing using a multi-scale context encoder and a local temporal residual encoder;
[0027] Multi-scale context encoder: for seasonal components After adjusting the dimensions, flattening, and performing batch normalization, the original shape is restored to obtain a standardized input representation. ; Representation of standardized input Feature representations at different scales are extracted through multiple one-dimensional average pooling operations of different scales. Then, a step-by-step fusion approach is used to pass contextual information, for the first... Each scale, its features are first obtained through a feedforward network mapping. After length alignment of the coarser-scale features through linear mapping, they are added to adjacent finer-scale features for residual fusion. When the above recursive fusion process reaches... At that time, the final fine-scale feature representation obtained ,Will With standardized input representation Residual fusion is performed to obtain a hybrid multi-scale feature representation. Finally The rearranged and linearly mapped representation is projected onto a unified latent space dimension to obtain a cross-scale context representation. ;
[0028] Local Temporal Residual Encoder: The standardized seasonal components are divided into multiple continuous feature blocks along the time dimension to obtain the local residual representation at the current moment; the above compensation modeling is repeated for all blocks along the time axis to obtain the local residual feature flow. The local residual characteristic flow Rearranged and mapped to a unified latent space dimension, local fine-grained dynamic features are obtained. .
[0029] The time-frequency sensing residual hybrid expert module includes:
[0030] Context-Aware Dynamic Gating: Transforming Cross-Scale Context Representation Representation of local features The features are concatenated along the feature dimension, and routing scores for each expert are generated through linear mapping and nonlinear activation. ; Perform route scoring Normalization yields the gating probabilities for each expert. ;
[0031]
[0032] in, This indicates that the introduced random disturbance term follows a normal distribution with a mean of 0 and a variance of 0.01. , For linear mapping operations, This indicates a feature concatenation operation;
[0033]
[0034] in, Indicates the total number of experts; They represent the first The and the first Routing scores of individual experts; multi-frequency heterogeneous expert computation: including Fourier experts, identity suppression experts, and multilayer perceptron experts; Fourier experts map features to the frequency domain through Fast Fourier Transform, then perform frequency domain transformation through a linear layer and inversely transform back to the time domain; Fourier experts, identity suppression experts, and multilayer perceptron experts are used as candidate expert pools, and the highest-scoring experts are selected based on gating probabilities. Several experts participated in the actual calculations, and the outputs of these experts were weighted and summed with their corresponding gating probabilities to obtain the seasonal component prediction results. :
[0035]
[0036] in, For the selected number The forward computation output of an expert For the first The gating probability of an expert The front with the highest gating probability A set of indexes for experts, where j is the expert index.
[0037] A second aspect is a carbon emission prediction device based on exponential smoothing and hybrid expert methods, the device comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the device to perform the method described in any of the first aspects.
[0038] Third aspect, a computer-readable storage medium storing a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of the first aspects.
[0039] The beneficial effects of the technical solution provided by this invention are:
[0040] 1. More accurate capture of carbon emission trends: The tensor-quantized parallel exponential moving average (EMA) decomposition operator is used to give higher weight to recent changes in the operating conditions of the gas turbine, which can more sensitively capture the long-term evolution trend of carbon emissions caused by equipment aging or load adjustment, with accurate decomposition and avoidance of serial overhead.
[0041] 2. Long-term stable carbon emission forecasting: Multi-scale trend pyramid (MSTP) effectively overcomes data fluctuations caused by cross-cycle load scheduling by integrating low-frequency information at different time granularities, thus improving the stability of macro-level carbon emission trend forecasting.
[0042] 3. Adaptive decoupling of complex carbon emission fluctuations: The time-frequency sensing residual hybrid expert module (ResidualMoE) realizes expert routing through dynamic gating, Fourier experts fit periodic emission characteristics, MLP experts capture instantaneous carbon emission anomalies caused by sudden changes in combustion conditions, and identity experts filter sensor noise, thereby achieving accurate prediction of high-frequency carbon emission fluctuations and reducing environmental pollution. Attached Figure Description
[0043] Figure 1 A schematic diagram of the overall process of a carbon emission prediction method based on exponential smoothing and hybrid expert methods provided by the present invention;
[0044] Figure 2 This is a flowchart of the overall algorithm of the present invention;
[0045] Figure 3 This is a schematic diagram of the structure of the time-frequency sensing residual hybrid expert module;
[0046] Figure 4 This is a schematic diagram of the carbon emission data prediction device of the present invention;
[0047] Figure 5 This is a schematic diagram of another embodiment of the carbon emission data prediction device of the present invention. Detailed Implementation
[0048] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below.
[0049] Example 1
[0050] This invention proposes a carbon emission prediction method based on exponential smoothing and hybrid expert methods. (See also...) Figure 1 and Figure 2 The method includes the following steps:
[0051] S101: Acquire multivariate time series data of gas turbine and perform stabilization preprocessing;
[0052] Specifically, this step involves: acquiring a historical multivariate time series data set. ,in Indicates the batch size. The length of the historical sequence. The number of channels is a variable. The set is a real number set. To mitigate the distribution drift caused by operating condition switching or environmental changes in real gas turbine time series data, the input sequence is first subjected to Reversible Instance Normalization (RevIN) for stabilization preprocessing to obtain the normalized sequence.
[0053] The inverse transform of RevIN is well known to those skilled in the art, and will not be described in detail in this embodiment of the invention.
[0054] S102: The preprocessed gas turbine multivariate time series data is decoupled into trend components and seasonal components using a parallel tensor-based exponential moving average decomposition module (EMA).
[0055] The Exponential Moving Average (EMA) decomposition module assigns higher weight to more recent operating condition changes, making it more suitable for characterizing the gradual trend information of the gas turbine. To avoid the serial overhead caused by time-step recursive calculations, tensor quantization parallelism is used to complete the calculations.
[0056] Construct the decay tensor:
[0057] in, , This is a smoothing factor.
[0058] Define the weight tensor:
[0059]
[0060] Based on this, the trend component is calculated as follows:
[0061]
[0062] in This represents element-wise multiplication. This indicates a prefix accumulation operation. Let be the weight tensor and D be the decay tensor.
[0063] The corresponding seasonal component is defined as follows:
[0064]
[0065] S103: Construct a multi-scale trend pyramid to predict trend components and obtain trend prediction results;
[0066] For trend components Includes low-frequency macroscopic change information, set Different scale levels, corresponding to scale indexes , of which The downsampling factor for each scale is :
[0067] For the Each scale, the trend component is represented as:
[0068]
[0069] in, For the first Trend components at each scale are represented. This is a one-dimensional average pooling operation; The pooling kernel size, This is the pooling step size.
[0070] At each scale, an independent linear predictor is used to map the trend representation to obtain the prediction results at the corresponding scale. Subsequently, linear interpolation was used to uniformly map the prediction results at each scale to the target length. Introducing globally learnable weight vectors and through Normalization yields the fusion weights for each scale. The final trend prediction result is expressed as:
[0071]
[0072] in, For linear interpolation operations, For the first Prediction results at various scales Where S is the interpolation target length, and S is the total number of decomposition scale levels. For globally learnable weight vectors, It is a normalized exponential function.
[0073] S104: A dual-branch representation structure is used to encode the seasonal components with deep features, obtaining local fine-grained dynamic features. Cross-scale context representation ;
[0074] Among them, seasonal components The model incorporates complex high-frequency combustion fluctuations (high-frequency carbon emission fluctuations caused by transient changes in local operating conditions), and is processed using a multi-scale context encoder and a local temporal residual encoder.
[0075] Multi-scale context encoder: for seasonal components After adjusting the dimensions, flattening, and performing batch normalization, the original shape is restored to obtain a standardized input representation. ;
[0076]
[0077] in, This is a one-dimensional batch normalization operation. For flattening operation.
[0078] For standardized input representation Feature representations at different scales are extracted through multiple one-dimensional average pooling operations of different scales. :
[0079] in , This represents the total number of scales.
[0080] Then, a step-by-step fusion approach is used to pass contextual information. For the first... Each scale, its features are first obtained through a feedforward network mapping. :
[0081]
[0082] in, , This is the weight matrix. , For bias terms, is the activation function for the Gaussian error linear unit.
[0083] After length alignment of the coarser-scale features through linear mapping, they are added to adjacent finer-scale features for residual fusion.
[0084]
[0085] in, ; For the first Feature representation at each scale.
[0086] When the above recursive fusion is executed to At that time, the final fine-scale feature representation obtained :
[0087]
[0088] Will With standardized input representation Residual fusion is performed to obtain a hybrid multi-scale feature representation. :
[0089]
[0090] Finally The rearranged and linearly mapped representation is projected onto a unified latent space dimension to obtain a cross-scale context representation. :
[0091]
[0092] in, and These are the projection weights and the bias, respectively. This is a tensor rearrangement operation.
[0093] Local temporal residual encoder: Divides the standardized seasonal components into multiple consecutive feature blocks along the time dimension. Let the current target block be... The previous block was Using adjacent historical blocks as local reference information, compensation modeling is performed on the current block to form the local residual representation at the current moment. :
[0094]
[0095] in, Show the normalized features of the previous block. and The weights and biases of the aggregation layer; This is a random deactivation operation. Let be the activation function of the Gaussian error linear unit.
[0096] The above compensation modeling is repeated for all blocks along the time axis to obtain the local residual feature flow. The local residual characteristic flow Rearranged and mapped to a unified latent space dimension, local fine-grained dynamic features are obtained. :
[0097]
[0098] in and These are the weights and biases of the feature map, respectively.
[0099] S105: Based on local fine-grained dynamic features Cross-scale context representation Adaptive route prediction is performed using a time-frequency sensing residual hybrid expert module (ResidualMoE) to obtain the gating probability of each expert. Based on gating probability Combined with seasonal component prediction, obtain the seasonal component prediction results. ;
[0100] Context-Aware Dynamic Gating: Transforming Cross-Scale Context Representation Representation of local features The features are concatenated along the feature dimension, and routing scores for each expert are generated through linear mapping and nonlinear activation. :
[0101]
[0102] in, This indicates that the introduced random disturbance term follows a normal distribution with a mean of 0 and a variance of 0.01. ; Perform route scoring Normalization yields the gating probabilities for each expert. , For linear mapping operations, This indicates a feature splicing operation.
[0103]
[0104] in, Indicates the total number of experts; They represent the first The and the first The routing score of each expert.
[0105] Multi-frequency heterogeneous expert computation includes Fourier experts, identity experts, and multilayer perceptron experts; Fourier experts map features to the frequency domain through fast Fourier transform, then perform frequency domain transformation through a linear layer and inverse transform back to the time domain.
[0106] (1) Fourier experts are used to model frequency domain structure information:
[0107]
[0108] in, This is the output of a Fourier expert. For inverse Fast Fourier Transform, For Fast Fourier Transform.
[0109] (2) The identity suppression expert is used as a noise filter, where the weight matrix is... and bias terms All are forcibly initialized to zero matrices to filter out unpredictable random walks or pure noise. The calculation process is as follows:
[0110]
[0111] in, This represents the output of the identity suppression expert.
[0112] (3) Multilayer perceptron experts directly perform nonlinear mapping in the time-domain feature space, for the first... A multilayer perceptron expert, whose computational process is used to model complex nonlinear dynamic modes:
[0113]
[0114] in, and These represent the hidden layer and output layer weight matrices, respectively. and For the corresponding bias term; For the first The output of a multilayer perceptron expert.
[0115] The model uses the Fourier experts, identity inhibition experts, and multilayer perceptron experts from (1) to (3) above as a pool of candidate experts, and selects the one with the highest score based on the gating probability. Several experts participated in the actual calculations, and the outputs of these experts were weighted and summed with their corresponding gating probabilities to obtain the seasonal component prediction results. :
[0116]
[0117] in, For the selected number The forward computation output of an expert For the first The gating probability of an expert The front with the highest gating probability A set of indexes for experts, where j is the expert index.
[0118] S106: Based on seasonal component prediction results and trend forecast results Reconstruct the prediction results to obtain the final multivariate prediction results. ;
[0119] The final normalized forecast output is obtained by adding the trend term and the seasonal term:
[0120]
[0121] in, This is the final normalized prediction output.
[0122] Then, the inverse RevIN transformation is performed to restore the original scale, yielding the final multivariate prediction results. :
[0123]
[0124] The inverse transform of RevIN is well known to those skilled in the art, and will not be described in detail in this embodiment of the invention.
[0125] S107: Present the final multivariate prediction results The output is sent to the carbon emission control device, and by adjusting the combustion parameters in advance, feedforward emission reduction control of gas turbine carbon emissions is achieved.
[0126] Example 2
[0127] To objectively and comprehensively verify the effectiveness and advancement of the method proposed in the embodiments of the present invention, this section will describe in detail the experimental setup, comparison model, and evaluation indicators, and analyze the experimental results on publicly available real industrial datasets.
[0128] This invention employs a publicly available gas turbine emissions dataset (Gas Turbine CO and NOx Emission Data Set) to validate the predictive performance of the proposed method. This dataset originates from a real-world gas turbine system in a thermal power plant and is collected in real-time by multiple sensors, serving as a typical benchmark for validating industrial time series forecasting algorithms.
[0129] The dataset contains 36,733 data points arranged in chronological order, with a sampling frequency of hourly. Each data point contains 11 multivariable sensor readings, such as ambient temperature (AT), ambient pressure (AP), ambient humidity (AH), turbine inlet temperature (TIT), and the concentrations of target emissions carbon monoxide (CO) and nitrogen oxides (NOx).
[0130] The task of this experiment is to predict multivariate variables, that is, to use data of all 11 variables over a historical period (e.g., the past 96 hours) to predict the values of all 11 variables over a longer future period (e.g., the next 96, 192, 336, or 720 hours).
[0131] To ensure the fairness and reproducibility of the experiment, this embodiment follows the standard time series partitioning method, dividing the dataset into training set (70%), validation set (10%) and test set (20%) in chronological order.
[0132] To quantify the prediction accuracy of the method of this invention, two standard evaluation metrics widely used in the field of time series forecasting are adopted, namely MSE and MAE.
[0133] All comparative models were trained and tested using their publicly available optimal parameter configurations and under the same data partitioning and experimental conditions to ensure fairness. In the embodiments of this invention, the proposed method was comprehensively compared with the aforementioned baseline models on a gas turbine dataset, and the experimental results are shown in Table 1. The table shows the MSE and MAE indices of each model at different prediction lengths (96, 192, 336, and 720 hours).
[0134] Table 1. Comparison of results of the present invention and different models on the Gas Turbine dataset.
[0135]
[0136] The experimental results in Table 1 show that:
[0137] The Exponential Smoothing and Time-Frequency Hybrid Expert Model (ExpMoE) proposed in this invention was comprehensively compared with several mainstream advanced time series forecasting models on publicly available gas turbine datasets. Experiments show that, across all forecast lengths from short to long periods, the method of this invention achieves the best prediction accuracy across all evaluation metrics, fully validating its overall superiority in complex, multivariate industrial forecasting scenarios.
[0138] In short- to medium-term forecasting, the model is able to keenly capture recent local dynamic fluctuations of gas turbines; while in the most challenging ultra-long-term forecasting tasks, when most comparative models suffer significant performance degradation due to error accumulation, the method of this invention still maintains extremely low prediction error, demonstrating excellent long-term prediction stability and robustness.
[0139] The superior performance of this invention lies in its unique "trend-seasonal" dual-branch collaboration and dynamic routing design. On the one hand, the tensor exponential moving average and multi-scale trend pyramid provide a solid macro-trend anchor for long-term forecasting; on the other hand, the time-frequency sensing residual hybrid expert module successfully decouples the periodic patterns and local abrupt changes in high-frequency components by adaptively assigning different experts, and effectively filters out the interference of random noise, thus surpassing the comparative model in overall performance.
[0140] Furthermore, in practical applications, carbon emission prediction devices adjust the fuel-to-air ratio in advance based on the prediction results, achieving feedforward intervention. This method reduces the actual generation of carbon monoxide and nitrogen oxides at the source, effectively reducing the risk of power plants exceeding emission standards and improving the atmospheric environment.
[0141] Example 3
[0142] like Figure 4 As shown, this embodiment of the invention provides a device for predicting carbon emission data from gas turbines in thermal power plants. This device is a modular end-to-end prediction framework, logically comprising the following functional modules:
[0143] Data acquisition and preprocessing module: responsible for establishing connections with external data sources, acquiring raw gas turbine time series data, and performing reversible instance normalization operations to eliminate dimensional differences between sensor variables.
[0144] Sequence decomposition module: Employs a tensor-quantized parallel exponential moving average (EMA) algorithm to decouple complex input sequences into a trend component with long-term smoothing characteristics and a seasonal component containing high-frequency fluctuations.
[0145] Trend prediction module: Built-in multi-scale trend pyramid (MSTP) performs feature extraction and independent linear prediction of trend components at different downsampling scales, and then dynamically fuses the results of each scale through interpolation and softmax weights.
[0146] Seasonal prediction module: Integrates a dual-branch representation structure and a hybrid expert network. The Multi-Scale Context Encoder (MSCE) and the Local Temporal Residual Encoder (LTRE) are responsible for extracting cross-scale global context and local fine-grained dynamic features, respectively. The Time-Frequency Aware Residual Hybrid Expert Module (ResidualMoE) utilizes a dynamic gating mechanism to adaptively route features to Fourier experts, identity experts, or MLP experts, and outputs weighted high-frequency prediction values.
[0147] The prediction combination module and result output module add the trend prediction value and the seasonal prediction value point by point, perform reverse normalization, persist the data, and output the prediction result to the upper layer application in real time.
[0148] Example 4
[0149] like Figure 5 The diagram shown is a hardware structure schematic of an industrial carbon emission prediction system provided in an embodiment of the present invention. This system can be a general-purpose computing device, such as a server, workstation, or embedded industrial computer.
[0150] The system specifically includes:
[0151] Processor: As the computing and control core of the system, it is responsible for executing the "carbon emission prediction program" stored in the memory to implement the complete prediction method steps described in Example 1 of this specification (including EMA decomposition, MSTP trend prediction, dual-branch coding, ResidualMoE expert routing, etc.).
[0152] System bus: As a communication link connecting various hardware components within the system, it is responsible for transmitting data and control signals between the processor, memory, user interface, network interface, and display screen.
[0153] User interface: Used to enable interaction between operators and the system. It can include input devices such as a keyboard, mouse, and touchscreen, allowing users to configure prediction tasks, set parameters, and start or stop programs. As shown in the figure, the user interface can also be embedded in or connected to a display screen for real-time feedback of operational information.
[0154] Network interface: Used to enable communication between the system and external networks. Through this interface, the system can obtain sensor data from the gas turbine from the factory's real-time database or cloud platform, and send the prediction results to other related industrial application systems.
[0155] Memory: This is the system's data and program storage center. Memory can store various software modules and data, including:
[0156] Operating system: Manages and coordinates the hardware and software resources of a system.
[0157] Network communication module: The underlying protocol stack that provides support for data transmission and reception via network interfaces.
[0158] User Interface Module: Drives the user interface and processes input from the user.
[0159] Carbon emission prediction program: This is the core software implementation of this invention. When this program is executed by the processor, it implements the complete prediction method flow described in this specification.
[0160] Display screen: As an independent output device, it is directly connected to the system bus to intuitively display the final prediction results, performance index reports and system operating status information to users.
[0161] Unless otherwise specified, the model numbers of the various devices in this embodiment of the invention are not limited, and any device that can perform the above functions is acceptable.
[0162] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.
[0163] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. A carbon emission prediction method based on exponential smoothing and hybrid expert methods, characterized in that, The method includes: The preprocessed gas turbine multivariate time series data sequence is decoupled into trend components and seasonal components using a parallel tensor-based exponential moving average decomposition module. A multi-scale trend pyramid is constructed to predict trend components and obtain trend prediction results; a dual-branch representation structure is used to perform deep feature encoding on seasonal components to obtain cross-scale context representation based on local fine-grained dynamic features. Adaptive route prediction is performed using a time-frequency sensing residual hybrid expert module based on local fine-grained dynamic features and cross-scale context representation. The gating probabilities of each expert are obtained, and the seasonal component prediction results are obtained based on the gating probabilities and seasonal component prediction. The prediction results are reconstructed based on the seasonal component prediction results and the trend prediction results to obtain the final multivariate prediction results. The final multivariate prediction results are then output to the carbon emission control device, and feedforward emission reduction control of gas turbine carbon emissions is achieved by adjusting the combustion parameters in advance.
2. The carbon emission prediction method based on exponential smoothing and hybrid expert methods according to claim 1, characterized in that, The parallel tensor-based exponential moving average decomposition module is as follows: Construct the decay tensor: ,in, , It is a smoothing factor; Define the weight tensor: ； Based on this, the trend component is calculated as follows: ； in This represents element-wise multiplication. This indicates a prefix accumulation operation. Let be the weight tensor and D be the decay tensor.
3. The carbon emission prediction method based on exponential smoothing and hybrid expert methods according to claim 1, characterized in that, The multi-scale trend pyramid is as follows: For trend components Includes low-frequency macroscopic change information, set Different scale levels, corresponding to scale indexes , of which The downsampling factor for each scale is : For the Each scale, the trend component is represented as: ； in, For the first Trend components at each scale are represented. This is a one-dimensional average pooling operation; The pooling kernel size, The pooling step size; The final trend prediction result is expressed as follows: ； in, For linear interpolation operations, For the first Prediction results at various scales Where S is the interpolation target length, and S is the total number of decomposition scale levels. For globally learnable weight vectors, It is a normalized exponential function.
4. The carbon emission prediction method based on exponential smoothing and hybrid expert methods according to claim 1, characterized in that, The method includes processing using a multi-scale context encoder and a local temporal residual encoder; Multi-scale context encoder: for seasonal components After adjusting the dimensions, flattening, and performing batch normalization, the original shape is restored to obtain a standardized input representation. ; Representation of standardized input Feature representations at different scales are extracted through multiple one-dimensional average pooling operations of different scales. Then, a step-by-step fusion approach is used to pass contextual information, for the first... Each scale, its features are first obtained through a feedforward network mapping. After length alignment of the coarser-scale features through linear mapping, they are added to adjacent finer-scale features for residual fusion. When the above recursive fusion process reaches... At that time, the final fine-scale feature representation obtained ,Will With standardized input representation Residual fusion is performed to obtain a hybrid multi-scale feature representation. Finally The rearranged and linearly mapped representation is projected onto a unified latent space dimension to obtain a cross-scale context representation. ; Local Temporal Residual Encoder: The standardized seasonal components are divided into multiple continuous feature blocks along the time dimension to obtain the local residual representation at the current moment; the above compensation modeling is repeated for all blocks along the time axis to obtain the local residual feature flow. The local residual characteristic flow Rearranged and mapped to a unified latent space dimension, local fine-grained dynamic features are obtained. .
5. The carbon emission prediction method based on exponential smoothing and hybrid expert methods according to claim 1, characterized in that, The time-frequency sensing residual hybrid expert module includes: Context-Aware Dynamic Gating: Transforming Cross-Scale Context Representation Representation of local features The features are concatenated along the feature dimension, and routing scores for each expert are generated through linear mapping and nonlinear activation. ; Perform route scoring Normalization yields the gating probabilities for each expert. ; ； in, This indicates that the introduced random disturbance term follows a normal distribution with a mean of 0 and a variance of 0.
01. , For linear mapping operations, This indicates a feature concatenation operation; ； in, Indicates the total number of experts; They represent the first The and the first Routing scores of individual experts; multi-frequency heterogeneous expert computation: including Fourier experts, identity suppression experts, and multilayer perceptron experts; Fourier experts map features to the frequency domain through Fast Fourier Transform, then perform frequency domain transformation through a linear layer and inversely transform back to the time domain; Fourier experts, identity suppression experts, and multilayer perceptron experts are used as candidate expert pools, and the highest-scoring experts are selected based on gating probabilities. Several experts participated in the actual calculations, and the outputs of these experts were weighted and summed with their corresponding gating probabilities to obtain the seasonal component prediction results. : ； in, For the selected number The forward computation output of an expert For the first The gating probability of an expert The front with the highest gating probability A set of indexes for experts, where j is the expert index.
6. A carbon emission prediction device based on exponential smoothing and hybrid expert methods, characterized in that, The device includes a processor and a memory, the memory storing program instructions, the processor invoking the program instructions stored in the memory to cause the device to perform the method according to any one of claims 1-5.
7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method described in any one of claims 1-5.