Method and device for predicting carbon emission data of gas turbine of thermal power plant

Abstract

The invention discloses a method and device for predicting carbon emission data of a gas turbine of a thermal power plant, and the method comprises the steps: carrying out the standardization processing of a trend term through reversible instance normalization, and enabling m variable channels to be clustered into n channel clusters through employing a K-Means algorithm for the trend term after the standardization processing; through a cluster-driven prediction module, applying an exclusive independent prediction model for each channel cluster to carry out parallel prediction, and carrying out inverse standardization on a prediction result to obtain a final trend prediction value, and carrying out structured multi-scale wavelet decomposition on a seasonal item to obtain a final approximate component and a plurality of high-frequency detail components; performing collaborative coding on the final approximate component to obtain depth context representation; performing dynamic fusion on the depth context representation and the plurality of high-frequency detail components by adopting an attention cross-scale fusion module to obtain fusion representation; and reconstructing a seasonal component according to the fusion representation and the plurality of high-frequency detail components through inverse wavelet transform to obtain a seasonal predicted value and a trend predicted value to form a final long-term time sequence prediction result, and outputting the result.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of carbon emissions, and particularly relates to a method and device for predicting carbon emission data of a gas turbine of a thermal power plant, and is particularly applied to the prediction of carbon emissions and related working condition parameters of the gas turbine of the thermal power plant. BACKGROUND

[0002] As a key power unit, the running state of the gas turbine is not only related to the power generation efficiency, but also directly affects the emission level of pollutants such as carbon monoxide (CO) and nitrogen oxides (NOx). Therefore, accurately predicting the multi-dimensional system state of the carbon emission and related working condition parameters (such as temperature, pressure, vibration, etc.) of the gas turbine in the future period of time has great theoretical and application value for realizing predictive maintenance, optimizing combustion control, and ensuring safe and environmentally friendly production.

[0003] However, the time series data collected from the gas turbine sensors has typical complexity, which brings great challenges to the prediction task. First, the data presents significant multi-scale dynamics, including both long-period trends caused by grid load scheduling and short-time high-frequency fluctuations caused by combustion process disturbances. Second, the system state is tightly coupled by multiple variables, and there is a complex nonlinear causal relationship between the data of each sensor. Third, the data shows strong non-stationarity, and its statistical properties will dynamically drift with factors such as device start-stop, working condition switching, and environmental changes.

[0004] The existing prediction techniques have obvious deficiencies in dealing with the above challenges. Traditional statistical models (such as ARIMA) are difficult to depict the nonlinear relationship between multiple variables. The model based on recurrent neural network (RNN) is prone to gradient problems when processing long sequences. Although the standard Transformer model can capture global dependencies, its inherent permutation invariance makes it less sensitive to the local structure and order of time series. Some improved methods try to introduce decomposition ideas, but mostly use fixed and non-adaptive decomposition strategies (such as Fourier transform or fixed wavelet basis), which cannot flexibly match the dynamic time-frequency characteristics of the gas turbine under different operating conditions, resulting in inaccurate information decomposition. In addition, when fusing multi-scale information, existing models usually use simple information superposition or fixed hierarchical structure, lacking a dynamic mechanism to distinguish and weight the importance of different scale information, resulting in key fault precursors or high-frequency disturbance information being smoothed out or disturbed by noise information.

[0005] Therefore, there is an urgent need for a unified prediction framework that can adaptively analyze the multi-scale characteristics of industrial data, cooperatively model global and local information, and dynamically fuse multi-source features, to meet the demand for high-precision, multi-dimensional prediction in modern industrial intelligent operation and maintenance. SUMMARY

[0006] This invention provides a method and apparatus for predicting carbon emissions from gas turbines in thermal power plants. The invention aims to address the technical problems of existing prediction methods, which struggle to adaptively analyze the multi-scale dynamic characteristics of industrial data, collaboratively model global and local information, and dynamically fuse multi-source features, resulting in insufficient prediction accuracy and generalization ability. Details are described below:

[0007] A first aspect: a method for predicting carbon emission data from gas turbines in thermal power plants, the method comprising:

[0008] Obtain and decompose the time series data of the gas turbine to obtain the trend term. and seasonal items

[0009] For trend items The trend term is obtained by applying reversible instance normalization. The standardized trend terms are clustered into n channel clusters using the K-Means algorithm, with m variable channels grouped together. A cluster-driven prediction module then applies a dedicated, independent prediction model to each channel cluster for parallel prediction, and the prediction results are inversely standardized to obtain the final trend prediction value.

[0010] For seasonal items Structured multi-scale wavelet decomposition is performed to obtain the final approximate component and multiple high-frequency detail components; the final approximate component is co-coded to obtain a deep context representation.

[0011] An attention-based cross-scale fusion module is used to dynamically fuse the deep context representation with the multiple high-frequency detail components to obtain a fused representation;

[0012] By using inverse wavelet transform, the seasonal components are reconstructed based on the fused representation and the multiple high-frequency detail components to obtain the seasonal prediction value.

[0013] Seasonal forecast and trend forecast values The final long-term time series prediction results are then output.

[0014] The process of clustering m variable channels into n channel clusters is as follows:

[0015]

[0016] Among them, id m It is the index of the cluster to which channel m belongs. Indicates based on index id m The selected multilayer perceptron prediction model specific to this cluster; This represents the normalized predicted trend term output by the model.

[0017] Among them, the seasonal item Structured multi-scale wavelet decomposition is performed as follows:

[0018] Feature tensor The process will proceed sequentially through a GELU activation function and a layer normalization operation, as shown below:

[0019]

[0020] in, For input tensor I k After an expansion rate of δ i The depth can be used to separate the output features obtained after convolution. One using expansion ratio δ i Depth-separable convolution operations, For a set of different preset expansion rates, Given the input tensor of a K-level multi-scale convolutional block, FusionConv is a fusion convolution operation, typically a 1x1 convolution. The final output of the entire multi-scale convolutional block is the learnable predictive operator P. k Or update operator U k The calculation results are as follows: LayerNorm is the layer normalization operation, and GELU is the GELU activation function.

[0021] The co-coding of the final approximate components is as follows:

[0022] Current layer input A multi-head self-attention mechanism is used to capture the global long-range dependencies of a sequence. The output is residually connected to the input and then normalized through layers to obtain a preliminary global context-aware representation, denoted as...

[0023] Subsequently The core component of CITE, the convolutional module, explicitly enhances the perception and modeling of local temporal patterns. This structure includes depthwise separable convolutions for efficient extraction of local features, pointwise convolutions for feature interaction and dimension adjustment, and gated linear units for feature selection and the introduction of nonlinearity, and integrates a residual learning mechanism.

[0024] First The channel dimension is expanded from D to 2D through a 1×1 pointwise convolution, providing a richer representation space for subsequent feature extraction. A depthwise convolution then independently captures local temporal patterns in each channel. This is followed by a gated linear unit as a non-linear gate to adaptively select and pass information, generating the gated representation. Features after gating The image is projected back to the original dimension D through another pointwise convolution, thus refining the local features and obtaining...

[0025] The integrated information is then fed into a standard position-wise feedforward network for final feature refinement, producing the output of the current CITE layer. The low-frequency characterization output after processing by all CITE layers is denoted as It contains both long-term dependencies of the sequence and rich local dynamic details.

[0026] A second aspect is an apparatus for predicting carbon emission data from gas turbines in thermal power plants, the apparatus comprising: a processor and a memory, the memory storing program instructions, the processor calling the program instructions stored in the memory to cause the apparatus to perform the method described in any of the first aspects.

[0027] Thirdly, an apparatus for predicting carbon emission data from gas turbines in thermal power plants, the apparatus comprising:

[0028] Data Acquisition and Preprocessing Module: Responsible for establishing connections with external data sources, acquiring raw, multivariable gas turbine time series data, and performing preprocessing operations;

[0029] Sequence decomposition module: Performs preliminary pattern decoupling on the standard data stream after preprocessing, using moving average decomposition to decompose the complex input sequence into two subsequences with different physical meanings and frequency characteristics, which are then sent to dedicated prediction branches for processing.

[0030] Trend prediction module: Models and predicts trend components. Channel clustering unit analyzes the time series characteristics of each channel of multivariate trend components and applies K-Means clustering algorithm to automatically divide variable channels with similar dynamic patterns into the same cluster. Then, the parallel prediction unit configures an independent, lightweight prediction model for each cluster and performs parallel long-range prediction of all channel data in the cluster.

[0031] The seasonal prediction module: The adaptive decomposition unit parses the input seasonal components into a low-frequency approximate component and multiple high-frequency detail components at different scales; the co-coding unit receives the low-frequency approximate component and performs co-feature extraction of global dependencies and local structures through its internally integrated self-attention and convolution mechanism to generate a deep context representation; the cross-scale fusion unit uses the context representation as a query to dynamically retrieve and fuse the most relevant information from multiple levels of high-frequency details to achieve information enhancement; the symmetric reconstruction unit starts with the fused representation and reconstructs the final multivariate prediction sequence of the seasonal components through an inverse transformation symmetrical to the decomposition process;

[0032] Prediction Combination Module: Receives the trend prediction value output from the trend prediction module and the seasonal prediction value output from the seasonality prediction module; adds these two prediction results with different time scales point by point to combine them into a complete final multivariate prediction result that includes both long-term evolution patterns and short-term fluctuation details.

[0033] Fourth 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.

[0034] The beneficial effects of the technical solution provided by this invention are:

[0035] 1. Adaptive time-frequency analysis capability: Through the SMW (Structured Multi-Scale Wavelet Decomposition) module, this invention can adaptively learn the optimal time-frequency decomposition method based on the actual operating data of the gas turbine, accurately capture the dynamic patterns under different operating conditions, and solve the problems of existing technologies relying on fixed priors and poor adaptability.

[0036] 2. Global-Local Collaborative Modeling: The CITE (Convolutional Fusion Transformer Encoder) encoder organically combines the global receptive field of the Transformer with the local perception capability of convolution, enabling it to simultaneously capture the long-term operating trends (global) that affect the state of the gas turbine and the instantaneous fluctuations (local) during the combustion process, resulting in more comprehensive and in-depth information extraction.

[0037] 3. Dynamic and efficient cross-scale fusion: The ACF (Attention Cross-Scale Fusion) module breaks through the limitations of traditional fixed fusion. It dynamically judges the importance of high-frequency details at each scale to the core trend in a data-driven manner and performs weighted fusion, which effectively enhances key signals, suppresses noise, and solves the problem of cross-scale information barriers.

[0038] 4. Refined Multivariate Prediction: By clustering and differentially modeling the multi-sensor channels in the trend components, the complex correlations between different physical quantities can be handled more precisely. This results in a comprehensive and high-precision prediction of the entire system state (all variables), enabling coordinated forecasting of the entire system state, including pollutants and key operating parameters. This precise grasp of the dynamics of multivariate correlations provides a data foundation for proactive combustion process optimization and control, thereby more effectively suppressing pollutant generation at its source and improving overall energy efficiency. Attached Figure Description

[0039] Figure 1This is a schematic diagram of the overall process of a method for predicting carbon emissions from gas turbines in thermal power plants according to the present invention;

[0040] Figure 2 This is a flowchart of the overall algorithm of the present invention;

[0041] Figure 3 This is a schematic diagram of the structure of the multi-scale convolutional block of the present invention;

[0042] Figure 4 This is a schematic diagram of the structure of the collaborative encoder module of the present invention;

[0043] Figure 5 This is a schematic diagram of the attention cross-scale fusion module of the present invention;

[0044] Figure 6 This is a schematic diagram of the structure of the carbon emission prediction system for gas turbines in thermal power plants according to the present invention;

[0045] Figure 7 This is a schematic diagram of another embodiment of the carbon emission data prediction system for gas turbines in thermal power plants according to the present invention. Detailed Implementation

[0046] 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.

[0047] Example 1

[0048] This invention proposes a method for predicting carbon emission data from gas turbines in thermal power plants. (See also...) Figure 1 and Figure 2 The method includes the following steps:

[0049] S101: Acquire and decompose the time series data of the gas turbine to obtain the trend term. and seasonal items

[0050] Specifically, this step involves inputting a historical multivariate time series dataset. The dataset is derived from gas turbine emissions datasets, where L is the length of the historical sequence and M is the number of sensor channels. The input sequence is in the real number field. A moving average-based decomposition module is used to decompose the input sequence into a trend term. and seasonal items

[0051] The decomposition module based on moving average is well known to those skilled in the art, and will not be described in detail in this embodiment of the invention.

[0052] S102: First, analyze the trend item. The result is obtained by applying Reversible Instance Normalization (RevIN) for standardization. Subsequently, the trend term is fed into ClusteredMLP for prediction, which uses K-Means clustering to divide the m variable channels into n clusters;

[0053]

[0054] Among them, id m It is the index of the cluster to which channel m belongs. To indicate based on index id m The selected multilayer perceptron (MLP) prediction model specific to this cluster; This represents the normalized predicted trend term output by the model.

[0055] Each MLP consists of two fully connected layers and a ReLU activation function. The predicted normalized trend is then scaled back to its original scale using the inverse RevIN transform to obtain the final prediction of the trend term. B represents the batch size, M represents the number of variable channels, corresponding to the number of sensors on the gas turbine, and T represents the length of the prediction sequence, i.e., the number of future time steps to be predicted.

[0056] 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.

[0057] S103: For seasonal items Perform structured multi-scale wavelet decomposition;

[0058] Among them, the seasonal items As the initial approximate component Input a decomposition module (SWM) consisting of multiple levels of learnable wavelet transforms. Each level of the transform is based on a lifting scheme and includes segmentation, prediction, and update steps. The K-th level decomposition process is as follows:

[0059]

[0060] in, To express [the opinion / towards] The even-position subsequence obtained after segmentation; To represent the approximate component of the (k-1)th level, it is the input signal of the kth level decomposition; To express [the opinion / towards] The odd-position subsequence obtained after segmentation; To represent the detail components obtained from the k-th level decomposition; P k To represent the learnable predictive operator of the k-th level; To represent the new approximate components obtained from the k-th level decomposition; Uk Let represent the learnable update operator for the k-th level.

[0061] The core innovation of this module lies in its prediction (P k ) and update (U k The operator consists of a multi-scale convolutional block (MSCB) (e.g.) Figure 3 (As shown in the diagram) This design enables the wavelet filter to adaptively learn from the data and possesses strong local multi-scale sensing capabilities. Instead of using a standard single-scale convolutional kernel, MSCB innovatively integrates a set of kernels with different preset dilation rates in parallel. The depth can be separable one-dimensional convolution. Where N d It is the quantity of the expansion rate. For input I k Right now or The multi-scale convolution module operation is first computed through parallel depthwise separable convolutions, that is: for each preset dilation rate δ i A depthwise convolution Defined as intra-channel convolution, these depthwise convolution outputs from different dilation rates are concatenated along the channel dimension to form a convolution with M·N channels. d Feature tensor Subsequently, this concatenated tensor undergoes cross-channel information integration and feature dimension adjustment via a fusion convolution, reducing its channel count from M·N. d Restore to M. The output of MSCB is denoted as... Ultimately, in order to obtain operator P k or U k The output. The process involves sequentially applying a GELU activation function and a layer normalization (LayerNorm) operation. The calculation process is shown below:

[0062]

[0063] in, For input tensor I k After an expansion rate of δ i The depth can be used to separate the output features obtained after convolution. One using expansion ratio δ i Depth-separable convolution operations, For a set of different preset expansion rates, Given the input tensor of a K-level multi-scale convolutional block, FusionConv is a fusion convolution operation, typically a 1x1 convolution, used to integrate the concatenated multi-scale features and adjust the channel dimensions. The final output of the entire multi-scale convolutional block is the learnable predictive operator P. kOr update operator U k The calculation results are as follows: LayerNorm is the layer normalization operation, and GELU is the GELU activation function.

[0064] By iteratively performing the decomposition, this module is able to extract the original seasonal items. Decomposed into a series of detail components at different frequency scales And a final, lowest-frequency approximation component It is considered to carry the most important long-term evolutionary patterns in the seasonal components and is directly used as input for subsequent CITE modules; the original detail components at each level The original detail components are stored for subsequent inverse wavelet transform to accurately recover high-frequency information. They are also processed by their respective dedicated linear projection layers to generate projected detail components. This is not only a key source for constructing high-frequency context in subsequent attention cross-scale modules (ACF), but is also stored to aid in the final signal reconstruction. The approximate components generated by intermediate-level decomposition, excluding the lowest frequency, are... They will also be processed through their respective linear projection layers to obtain... It is stored as an approximate residual skip connection in the decoding stage.

[0065] S104: Perform a co-encoder on the final approximate components;

[0066] The lowest frequency approximation component Input into the co-encoder (CITE) (e.g.) Figure 4 (As shown). This module is an innovative enhancement to the standard Transformer architecture. It consists of N stacked encoder layers. The processing flow of each layer is as follows:

[0067] First, the input of the current layer (Initial input is) This approach captures global long-range dependencies in a sequence using a standard multi-head self-attention (MHSA) mechanism. Its output is residually concatenated with the input and then subjected to layer normalization (LayerNorm) to obtain a preliminary global context-aware representation, denoted as...

[0068] Subsequently The core component input into CITE—the ConvModule—explicitly enhances the perception and modeling of local temporal patterns. This architecture includes depthwise separable convolutions for efficient local feature extraction, pointwise convolutions for feature interaction and dimensionality adjustment, and gated linear units (GLUs) for feature selection and introducing non-linearity, along with an integrated residual learning mechanism. First, [the following is a convolutional mechanism]... The channel dimension is expanded from D to 2D through a 1×1 pointwise convolution, providing a richer representation space for subsequent feature extraction. A depthwise convolution then independently captures local temporal patterns in each channel. This is followed by a gated linear unit (GLU) to adaptively select and pass information as a non-linear gate, generating the gated representation. The calculation process is as follows:

[0069]

[0070] Features after gating The image is projected back to the original dimension D through another pointwise convolution, thus refining the local features. The output of the entire operation sequence and Residual connections are performed to form the final convolutional augmented representation. The calculation is as follows:

[0071]

[0072] This integrated information is then fed into a standard position-wise feedforward network (FFN) for final feature refinement, producing the output of the current CITE layer. The calculation process is as follows:

[0073]

[0074] Dropout is a regularization technique used to prevent overfitting.

[0075] Each layer of CITE collaboratively performs dual modeling of global and local information. Finally, the low-frequency representation output after processing by all CITE layers is denoted as... It contains both long-term dependencies of the sequence and rich local dynamic details.

[0076] S105: Employs an attention cross-scale fusion (ACF) module (such as...) Figure 5 (As shown). The ACF module first projects the detail components of each level. By using a linear projection layer to uniformly map the feature dimensions to the model's hidden dimension Dmodel, we obtain the adjusted multi-channel detail features. Subsequently, in order to extract detailed features from multiple channels The context is aggregated into a compact and representative single-vector representation c. kIt employs a dynamic channel attention mechanism based on feature mean aggregation. This mechanism posits that the importance of a channel at a specific scale can be represented by the average strength of its current feature representation. Specifically, for The feature vector of Dmodel dimension for each channel m in the model. First, calculate its mean across the feature dimensions as the raw attention score for that channel, denoted as . Original fractions Attention weights for each channel are generated by normalizing across all channels using the Softmax function. Ultimately, the detailed context c of level K k By analyzing the feature vectors of all channels We obtain the following by weighting and summing the corresponding attention weights:

[0077]

[0078] ACF further performs cross-scale information aggregation and deep fusion.

[0079] First, all detailed context vectors are stacked hierarchically to form a unified cross-scale key-value pair matrix that integrates high-frequency fluctuation information from different frequency scales. `Stack` represents the tensor stacking operation. This represents the single-vector high-frequency context representation corresponding to the k-th decomposition level, where dim represents the dimension along which the stacking operation proceeds, and K... wl D represents the total number of levels in the wavelet decomposition. model Indicates the model dimension.

[0080] The core of the ACF module is a deep fusion structure composed of N identical CrossScaleFusionLayer layers stacked together, where each fusion layer j receives data from the layer above. As a query (j>0), it represents the low-frequency global context representation from the CITE module. As the initial query, i.e. Reuse As a shared key and value, the query dynamically retrieves data from... The most relevant multi-scale high-frequency information is retrieved and integrated to obtain the attention output A. cross,j This is achieved using standard scaled dot product attention, and the calculation process is as follows:

[0081]

[0082] in, To integrate high-frequency information from all levels into a unified cross-scale key-value pair matrix, it serves as both Key and Value in this calculation; softmax is the Softmax function used to normalize the attention score into a probability distribution; d k The dimension of the key vector is used as a scaling factor to stabilize the gradient; The output of the (j-1)th fusion layer is used as the query for the j-th fusion layer.

[0083] To more effectively fuse the outputs of the cross-attention and feedforward networks and stabilize the information flow, a dual-gated residual connection (GRC) mechanism is employed. In fusion layer j, the cross-attention output A... cross,j Its corresponding input query Q j The fusion is performed using the first GRC. This module concatenates the two features along the feature dimension, then calculates a dynamic gating value using a dedicated linear transformation layer and a sigmoid activation function. This gating value is ultimately applied to A. cross,j This adaptively modulates the intensity of its information flow. Finally, the residuals are added to and normalized with the original input of this module to obtain the intermediate... The calculation process is as follows:

[0084]

[0085] in, The intermediate fused feature after the first gated residual connection (GRC); Q j Query is the input to the j-th layer fusion; σ is the Sigmoid activation function, whose output is used as the dynamic gating value; Linear is a linear transformation layer.

[0086] Then F′ j It is fed into the FFN module for nonlinear feature transformation to obtain A fnn,j Then A fnn,j With F′ j The final output of the current fusion layer j is produced by fusing the second GRC in exactly the same way.

[0087]

[0088] Among them, A fnn,j intermediate feature F′ j The output after nonlinear transformation by a feedforward neural network (FFN).

[0089] This dual GRC structure provides the model with the ability to refine and context-awarely control the information flow at each fusion layer, ensuring the enhancement of effective information and the suppression of noise.

[0090] The result obtained after N layers of such fusion operations is then connected to the initial input of the ACF module through a residual connection. The summation and normalization result in a fused representation denoted as . The final fused representation, which includes deep low-frequency context and multi-scale high-frequency details, will serve as the input basis for subsequent decoding stages.

[0091] S106: Reconstruct seasonal components through inverse wavelet transform and combine them for final prediction;

[0092] First, the fused multi-scale low-frequency representations are obtained through linear projection. Transformed into the lowest frequency approximation component Using the same dimension as the initial approximate component of the inverse wavelet transform Subsequently, from k=K w Step by step upwards to k=1, execute K w The inverse lifting scheme is used in each reconstruction level k. The approximate output for inverse lifting is derived from the current approximate component. Approximate projection components of the corresponding level stored in the encoding stage It is added together; while detailed input The construction is more refined, specifically consisting of the detail components of the projection. Compared to the original detail components after linear transformation specific to the current decoding level k They are added together. These two dynamically combined inputs work together on the current inverse transform level.

[0093] The core steps of the inverse boosting scheme include inverse update, inverse prediction, and interleaving, and it uses a learnable operator U. k and P k Also based on the MSCB construction, it forms a symmetry with the decomposition process. The specific calculation process is as follows:

[0094]

[0095] in, This refers to the even-position subsequence recovered through the reverse update step; For the approximate components that serve as input in the k-th level reconstruction; For the detail components that serve as input in the k-th level reconstruction; P is the odd-position subsequence recovered through the inverse prediction step; k U is a learnable inverse predictor of the kth level; k For the k-th level learnable inverse update operator; For even-numbered positions in the reconstructed higher-level approximation components; The recovered even-position subsequence; The recovered odd-position subsequence; These are the odd positions in the reconstructed higher-level approximate components.

[0096] After K w After the inverse transformation, a fully reconstructed sequence of seasonal terms is obtained. The sequence is then passed through a final linear projection layer to map its length from L to the desired prediction length T, yielding the prediction result for the seasonal term.

[0097] S107: Final Long-Term Time Series Forecast Output By trend item and seasonal items Adding them together, we get:

[0098]

[0099] The final long-term time series prediction output It provides a high-dimensional, system-level prediction of the future operating state of gas turbines. It does not predict a single emission indicator in isolation, but rather reveals the future synergistic evolution relationships among all variables, including key operating parameters and pollutants. This holistic understanding of multivariate dynamics enables advanced control systems to achieve optimized regulation based on the global state. For example, by anticipating changes in operating conditions, it can adjust combustion strategies in advance to suppress pollution at its source, rather than simply responding to emission results, thereby achieving more efficient pollution control, energy efficiency improvement, and equipment reliability assurance.

[0100] In summary, this invention provides an end-to-end, high-precision multivariate time series forecasting solution by performing refined, multi-scale modeling of data from the specific industrial scenario of gas turbines.

[0101] Example 2

[0102] To objectively and comprehensively verify the effectiveness and advancement of the method proposed in this invention, this section will detail the experimental setup, comparison model, and evaluation indicators, and analyze the experimental results on publicly available real industrial datasets.

[0103] 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.

[0104] The dataset contains 36,733 data points arranged in chronological order, with a sampling frequency of hourly. Each data point includes 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).

[0105] 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).

[0106] To ensure the fairness and reproducibility of the experiment, this embodiment follows the standard time series partitioning method, dividing the dataset into a training set (70%), a validation set (10%), and a test set (20%) in chronological order.

[0107] 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.

[0108] 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).

[0109] Table 1 Comparison of results of the present invention and different models on the Gas Turbine dataset.

[0110]

[0111] The experimental results in Table 1 show that:

[0112] Overall performance is optimal. Across all prediction lengths, the method proposed in this invention achieves the lowest error values ​​on both the MSE and MAE evaluation metrics, significantly outperforming all baseline models compared. This fully demonstrates that the technical solution proposed in this invention possesses the strongest overall predictive capability and the highest accuracy.

[0113] In challenging long-term time-series forecasting tasks where the forecast length increases to 720 hours, the performance advantage of the method in this invention is particularly evident. Compared to other models whose performance deteriorates significantly due to error accumulation, the method in this invention maintains a low forecast error, demonstrating its excellent long-term forecast stability and robustness.

[0114] The superior performance of this invention lies in its unique collaborative design. Adaptive multi-scale wavelet decomposition can capture the time-frequency characteristics of gas turbines under complex operating conditions more accurately than fixed decomposition methods; the collaborative encoder effectively combines global and local information, overcoming the limitations of a single model; and the attention cross-scale fusion mechanism dynamically enhances key information and suppresses noise, which is crucial for achieving high-precision long-term prediction.

[0115] In summary, through comprehensive comparative experiments on real industrial datasets, the method proposed in this invention has been fully demonstrated to have significant superiority and practical value in handling complex multivariate industrial time series forecasting tasks.

[0116] Example 3

[0117] like Figure 6 As shown, this embodiment of the invention provides an industrial carbon emission prediction device, which is a modular, end-to-end prediction framework. Logically, the device includes the following functional modules that work together to achieve high-precision prediction of multivariate time series such as those from gas turbines:

[0118] Data Acquisition and Preprocessing Module: As the input portal of the device, this module is responsible for establishing a connection with external data sources (such as real-time databases, historical data warehouses, or SCADA systems) and acquiring raw, multivariable gas turbine time series data. After acquisition, it will perform a series of preprocessing operations to eliminate the dimensional differences between different sensor variables, thereby providing high-quality, uniformly formatted data input for subsequent prediction modules.

[0119] Sequence Decomposition Module: This module receives the standard data stream processed by the data acquisition and preprocessing modules and performs initial pattern decoupling on it. It employs a moving average decomposition algorithm to break down the complex input sequence into two subsequences with different physical meanings and frequency characteristics: a long-term, slowly changing trend component and a short-term, periodic seasonal component. These two components are then separately fed into dedicated prediction branches for further processing.

[0120] Trend Prediction Module: This module is specifically responsible for modeling and predicting trend components. It comprises two core units. First, the channel clustering unit analyzes the time-series characteristics of each channel (i.e., different sensors) of the multivariate trend component and applies the K-Means clustering algorithm to automatically group variable channels with similar dynamic patterns into the same cluster. Then, the parallel prediction unit configures an independent, lightweight prediction model (such as MLP) for each cluster and performs parallel long-range predictions on all channel data within that cluster. Through this "divide and conquer" strategy, the module can differentiate the modeling of variables with different physical meanings, ultimately outputting high-precision multivariate trend prediction values.

[0121] Seasonal Prediction Module: As the core prediction engine of the system, this module is responsible for deep and structured modeling of the seasonal components, which contain complex short-term fluctuations. Its internal processing pipeline first uses an adaptive decomposition unit (SMW) to adaptively and hierarchically resolve the input seasonal components into a low-frequency approximation component and multiple high-frequency detail components at different scales. Subsequently, a co-coding unit (CITE) receives the low-frequency approximation component and uses its integrated self-attention and convolution mechanisms to extract collaborative features of global dependencies and local structures, generating a deep context representation. Next, a cross-scale fusion unit (ACF) uses this context representation as a query to actively retrieve and fuse the most relevant information from multiple levels of high-frequency details to achieve information enhancement. Finally, a symmetric reconstruction unit (inverse SMW) uses the fused representation as a starting point and, through an inverse transformation symmetrical to the decomposition process, accurately reconstructs the final multivariate prediction sequence of the seasonal components.

[0122] The forecasting combination module serves as the final decision fusion center. This module receives trend forecasts from the trend forecasting module and seasonal forecasts from the seasonality forecasting module. It adds these two forecasts, which have different time scales, point by point, to form a complete final multivariate forecast that includes both long-term evolution patterns and short-term fluctuation details.

[0123] Results storage and output module: As the output portal, this module is responsible for persistently storing the final prediction results generated by the prediction combination module. At the same time, it also provides a standardized data interface that can output the prediction results to upper-layer applications in real time, such as early warning systems for equipment health management, optimization control systems for energy saving and consumption reduction, or visual dashboards for operation and maintenance personnel monitoring.

[0124] In summary, through comprehensive comparative experiments on real industrial datasets, the method proposed in this invention has been fully demonstrated to have significant superiority and practical value in handling complex multivariate industrial time series forecasting tasks.

[0125] Example 4

[0126] like Figure 7 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.

[0127] The system specifically includes:

[0128] Processor: Serves as the computing and control core of the system. The processor connects to other components via the system bus. It is responsible for executing the "industrial carbon emission prediction program" stored in memory to implement all the method steps described in the embodiments of this invention, including data decomposition, trend prediction, seasonal prediction, and the combination of final results, specifically:

[0129] Obtain and decompose the time series data of the gas turbine to obtain the trend term. and seasonal items

[0130] For trend items The trend term is obtained by applying reversible instance normalization. The standardized trend terms are clustered into n channel clusters using the K-Means algorithm, with m variable channels grouped together. A cluster-driven prediction module then applies a dedicated, independent prediction model to each channel cluster for parallel prediction, and the prediction results are inversely standardized to obtain the final trend prediction value.

[0131] For seasonal items Structured multi-scale wavelet decomposition is performed to obtain the final approximate component and multiple high-frequency detail components; the final approximate component is co-coded to obtain a deep context representation.

[0132] An attention-based cross-scale fusion module is used to dynamically fuse the deep context representation with the multiple high-frequency detail components to obtain a fused representation;

[0133] By using inverse wavelet transform, the seasonal components are reconstructed based on the fused representation and the multiple high-frequency detail components to obtain the seasonal prediction value.

[0134] Seasonal forecast and trend forecast values The final long-term time series prediction results are then output.

[0135] Clustering m variable channels into n channel clusters is as follows:

[0136]

[0137] Among them, id m It is the index of the cluster to which channel m belongs. Indicates based on index id m The selected multilayer perceptron prediction model specific to this cluster; This represents the normalized predicted trend term output by the model.

[0138] Among them, the seasonal items Structured multi-scale wavelet decomposition is performed as follows:

[0139] Feature tensor The process will proceed sequentially through a GELU activation function and a layer normalization operation, as shown below:

[0140]

[0141] in, For input tensor I k After an expansion rate of δ i The depth can be used to separate the output features obtained after convolution. One using expansion ratio δ i Depth-separable convolution operations, For a set of different preset expansion rates, Given the input tensor of a K-level multi-scale convolutional block, FusionConv is a fusion convolution operation, typically a 1x1 convolution. The final output of the entire multi-scale convolutional block is the learnable predictive operator P. k Or update operator U k The calculation results are as follows: LayerNorm is the layer normalization operation, and GELU is the GELU activation function.

[0142] The co-coding of the final approximate components is as follows:

[0143] Current layer input A multi-head self-attention mechanism is used to capture the global long-range dependencies of a sequence. The output is residually connected to the input and then normalized through layers to obtain a preliminary global context-aware representation, denoted as...

[0144] Subsequently The core component of CITE, the convolutional module, explicitly enhances the perception and modeling of local temporal patterns. This structure includes depthwise separable convolutions for efficient extraction of local features, pointwise convolutions for feature interaction and dimension adjustment, and gated linear units for feature selection and the introduction of nonlinearity, and integrates a residual learning mechanism.

[0145] First The channel dimension is expanded from D to 2D through a 1×1 pointwise convolution, providing a richer representation space for subsequent feature extraction. A depthwise convolution then independently captures local temporal patterns in each channel. This is followed by a gated linear unit as a non-linear gate to adaptively select and pass information, generating the gated representation. Features after gating The image is projected back to the original dimension D through another pointwise convolution, thus refining the local features and obtaining...

[0146] The integrated information is then fed into a standard position-wise feedforward network for final feature refinement, producing the output of the current CITE layer. The low-frequency characterization output after processing by all CITE layers is denoted as It contains both long-term dependencies of the sequence and rich local dynamic details.

[0147] 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.

[0148] 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.

[0149] 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.

[0150] Memory: This is the system's data and program storage center. Memory can store various software modules and data, including:

[0151] Operating system: Manages and coordinates the hardware and software resources of a system.

[0152] Network communication module: The underlying protocol stack that provides support for data transmission and reception via network interfaces.

[0153] User Interface Module: Drives the user interface and processes input from the user.

[0154] 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.

[0155] 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.

[0156] 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.

[0157] 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.

[0158] 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 method for predicting carbon emissions from gas turbines in thermal power plants, characterized in that, The method includes: Obtain and decompose the time series data of the gas turbine to obtain the trend term. and seasonal items For trend items The trend term is obtained by applying reversible instance normalization. The standardized trend terms are clustered into n channel clusters using the K-Means algorithm, with m variable channels grouped together. A cluster-driven prediction module then applies a dedicated, independent prediction model to each channel cluster for parallel prediction, and the prediction results are inversely standardized to obtain the final trend prediction value. For seasonal items Structured multi-scale wavelet decomposition is performed to obtain the final approximate component and multiple high-frequency detail components; the final approximate component is co-coded to obtain a deep context representation. An attention-based cross-scale fusion module is used to dynamically fuse the deep context representation with the multiple high-frequency detail components to obtain a fused representation; By using inverse wavelet transform, the seasonal components are reconstructed based on the fused representation and the multiple high-frequency detail components to obtain the seasonal prediction value. Seasonal forecast and trend forecast values The final long-term time series prediction results are then output.

2. The method for predicting carbon emissions from gas turbines in thermal power plants according to claim 1, characterized in that, The process of clustering m variable channels into n channel clusters is as follows: Among them, id m It is the index of the cluster to which channel m belongs. Indicates based on index id m The selected multilayer perceptron prediction model specific to this cluster; This represents the normalized predicted trend term output by the model.

3. The method for predicting carbon emissions from gas turbines in thermal power plants according to claim 1, characterized in that, The seasonal items Structured multi-scale wavelet decomposition is performed as follows: Feature tensor The process will proceed sequentially through a GELU activation function and a layer normalization operation, as shown below: in, For input tensor I k After an expansion rate of δ i The depth can be used to separate the output features obtained after convolution. One using expansion ratio δ i Depth-separable convolution operations, For a set of different preset expansion rates, Given the input tensor of a K-level multi-scale convolutional block, FusionConv is a fusion convolution operation, typically a 1x1 convolution. The final output of the entire multi-scale convolutional block is the learnable predictive operator P. k Or update operator U k The calculation results are as follows: LayerNorm is the layer normalization operation, and GELU is the GELU activation function.

4. The method for predicting carbon emissions from gas turbines in thermal power plants according to claim 1, characterized in that, The collaborative encoding of the final approximate components is as follows: Current layer input A multi-head self-attention mechanism is used to capture the global long-range dependencies of a sequence. The output is residually connected to the input and then normalized through layers to obtain a preliminary global context-aware representation, denoted as... Subsequently The core component of CITE, the convolutional module, explicitly enhances the perception and modeling of local temporal patterns. This structure includes depthwise separable convolutions for efficient extraction of local features, pointwise convolutions for feature interaction and dimension adjustment, and gated linear units for feature selection and the introduction of nonlinearity, and integrates a residual learning mechanism. First The channel dimension is expanded from D to 2D through a 1×1 pointwise convolution, providing a richer representation space for subsequent feature extraction. A depthwise convolution then independently captures local temporal patterns in each channel. This is followed by a gated linear unit as a non-linear gate to adaptively select and pass information, generating the gated representation. Features after gating The image is projected back to the original dimension D through another pointwise convolution, thus refining the local features and obtaining... The integrated information is then fed into a standard position-wise feedforward network for final feature refinement, producing the output of the current CITE layer. The low-frequency characterization output after processing by all CITE layers is denoted as It contains both long-term dependencies of the sequence and rich local dynamic details.

5. A device for predicting carbon emission data from gas turbines in thermal power plants, 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 1-4.

6. A device for predicting carbon emission data from gas turbines in thermal power plants, characterized in that, The device includes: Data Acquisition and Preprocessing Module: Responsible for establishing connections with external data sources, acquiring raw, multivariable gas turbine time series data, and performing preprocessing operations; Sequence decomposition module: Performs preliminary pattern decoupling on the standard data stream after preprocessing, using moving average decomposition to decompose the complex input sequence into two subsequences with different physical meanings and frequency characteristics, which are then sent to dedicated prediction branches for processing. Trend prediction module: Models and predicts trend components. Channel clustering unit analyzes the time series characteristics of each channel of multivariate trend components and applies K-Means clustering algorithm to automatically divide variable channels with similar dynamic patterns into the same cluster. Then, the parallel prediction unit configures an independent, lightweight prediction model for each cluster and performs parallel long-range prediction of all channel data in the cluster. The seasonal prediction module: The adaptive decomposition unit parses the input seasonal components into a low-frequency approximate component and multiple high-frequency detail components at different scales; the co-coding unit receives the low-frequency approximate component and performs co-feature extraction of global dependencies and local structures through its internally integrated self-attention and convolution mechanism to generate a deep context representation; the cross-scale fusion unit uses the context representation as a query to dynamically retrieve and fuse the most relevant information from multiple levels of high-frequency details to achieve information enhancement; the symmetric reconstruction unit starts with the fused representation and reconstructs the final multivariate prediction sequence of the seasonal components through an inverse transformation symmetrical to the decomposition process; Prediction Combination Module: Receives the trend prediction value output from the trend prediction module and the seasonal prediction value output from the seasonality prediction module; adds these two prediction results with different time scales point by point to combine them into a complete final multivariate prediction result that includes both long-term evolution patterns and short-term fluctuation details.

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 1-4.

Images