A steam turbine heat rate prediction method based on smoothing filtering and time sequence mixing

Abstract

This invention belongs to the field of steam turbine performance monitoring and prediction technology in coal-fired power plants. It presents a method for predicting steam turbine heat rate based on a combination of smoothing filtering and time-series hybrid methods, comprising the following steps: (1) collecting the operating sequence data of the steam turbine in a coal-fired power plant and constructing a prediction dataset; (2) constructing an online prediction network for steam turbine heat rate based on a combination of smoothing filtering and time-series hybrid methods; (3) training the prediction network constructed in step 2 using the training set constructed in step 1; and (4) deploying the trained prediction model to the edge computing platform of a coal-fired power plant. The advantages of this invention are: the method effectively solves the problems of noise sensitivity, insufficient time-series feature modeling, poor real-time performance, and difficulty in balancing accuracy and lightweighting in existing steam turbine heat rate prediction methods. It significantly improves the accuracy, stability, and real-time performance of online heat rate prediction, providing strong technical support for intelligent operation, energy saving, and optimized scheduling of coal-fired power plant units.

Description

Technical Field

[0001] This invention relates to a method for predicting the heat rate of a steam turbine based on a hybrid smoothing filter and time series analysis, belonging to the field of steam turbine performance monitoring and prediction technology in coal-fired power plants. Background Technology

[0002] In coal-fired power plant thermal power generation systems, the steam turbine is the core equipment for energy conversion. Its heat rate is a key indicator for measuring the economic efficiency of unit operation and evaluating energy conversion efficiency. Accurate online prediction of heat rate has important engineering practical value for unit operation under different operating conditions, early warning of operation failures, and energy-saving and consumption-reducing retrofits.

[0003] During the operation of steam turbines in coal-fired power plants, their heat rate is affected by a combination of multiple parameters, including main steam pressure, temperature, unit load, and exhaust pressure. Furthermore, the operating data exhibits characteristics such as high frequency, strong noise, temporal nonlinearity, and transient fluctuations under varying operating conditions. Existing methods for predicting steam turbine heat rate face numerous technical bottlenecks, making it difficult to meet the needs of online real-time monitoring in coal-fired power plants.

[0004] (1) The traditional heat balance method relies on offline steady-state parameter calculation, which cannot adapt to dynamic operating conditions such as unit load change, start-up and shutdown. Moreover, the calculation process is cumbersome and has poor real-time performance, making it difficult to achieve online output of heat rate.

[0005] (2) Conventional machine learning prediction models are trained directly using noisy raw operating data. High-frequency noise introduced by sensor drift, operating condition fluctuations, and electromagnetic interference can easily lead to feature distortion and reduce prediction accuracy.

[0006] (3) Existing time series deep learning models focus on extracting single time domain features, and cannot simultaneously capture the long-term trend features and short-term details of heat consumption rate changes, and are not robust enough for predicting heat consumption rate under varying operating conditions.

[0007] (4) Mainstream prediction models lack efficient feature decomposition and fusion mechanisms, do not perform trend-residual separation on time series data, and do not achieve cross-scale feature interaction enhancement, which limits the model's ability to express complex time series patterns.

[0008] Extensive research has revealed that while existing lightweight time-series prediction models improve inference efficiency, they still have significant shortcomings in predicting the heat rate of steam turbines in coal-fired power plants. These models are limited by their lightweight structure, resulting in weak feature extraction capabilities and difficulty in fully capturing the multi-scale time-series features of the heat rate. When faced with data noise and transient fluctuations under varying operating conditions, they are prone to problems such as insufficient feature representation and confusion between trend and detailed features, leading to large prediction deviations and poor stability in the heat rate prediction. Consequently, they cannot provide reliable decision-making basis for optimizing the operation of coal-fired power plant units.

[0009] Therefore, facing the contradiction between the limited hardware resources at the edge of coal-fired power plants and the need for high-precision online prediction of turbine heat rate, how to construct an online prediction method for heat rate that takes into account noise suppression, multi-scale temporal feature extraction and efficient feature fusion, while maintaining lightweight characteristics, and improve prediction accuracy and robustness while ensuring real-time performance, has become a key research direction for promoting the intelligent operation of coal-fired power plants. Summary of the Invention

[0010] To address the problems of noise sensitivity, insufficient temporal feature modeling, poor online real-time performance, and difficulty in balancing accuracy and lightweight design in existing steam turbine heat rate prediction methods, this invention aims to provide a steam turbine heat rate prediction method based on smoothing filtering and temporal hybrid technology. This invention constructs a three-level network structure: an input encoding and smoothing filter decomposition module (SFED), a temporal hybrid feature extraction module (TMM), and a fusion and residual compensation output module (OUT). This achieves noise suppression, trend-residual feature decomposition, multi-scale temporal feature extraction and fusion of steam turbine operating data, and further improves prediction accuracy through linear residual compensation. Simultaneously, the network design strictly controls parameter scale and computational complexity to ensure the model maintains lightweight characteristics and adapts to the deployment requirements of coal-fired power plant edge platforms.

[0011] To achieve the aforementioned objectives and address the problems existing in the prior art, the technical solution adopted by this invention is: a method for predicting the heat rate of a steam turbine based on a combination of smoothing filtering and time-series analysis, comprising the following steps:

[0012] Step 1: Collect operating time series data of the steam turbine in the coal-fired power plant and construct a prediction dataset. Collect key operating parameters of the steam turbine in the coal-fired power plant, including 16-dimensional continuous time series variables such as main steam pressure, main steam temperature, reheat steam pressure, reheat steam temperature, main steam flow rate, exhaust steam pressure, feedwater temperature, and unit load. The sampling frequency is 1 minute, and a time series input window with a length of L=96 is constructed. The turbine heat consumption rate at the next moment in the corresponding window is the labeled data; the original data is filled with missing values ​​by linear interpolation. Outlier removal and Z-score standardization preprocessing were performed according to principles. A total of 8,000 valid time series samples were collected, covering typical operating conditions such as steady-state operation, variable load regulation, and start-up and shutdown operations of the unit. The samples were divided into a 70% training set, a 15% test set, and a 15% validation set.

[0013] Step 2: Construct an online prediction network for turbine heat rate based on smoothing filtering and temporal hybridization. This network includes an input encoding and smoothing filtering decomposition module (SFED), a temporal hybrid feature extraction module (TMM), and a fusion and residual compensation output module (OUT). Specifically, it includes the following sub-steps:

[0014] Sub-step (A) inputs timing information into the window. As the input to the input encoding and smoothing filter decomposition module SFED, this module consists of five parts: Differential Enhancement (DA), Embedded Mapping (EMB), Smoothing Filter (LSF), Channel Calibration (CAL), and Residual Decomposition (RES), which are connected in sequence. Specifically, they include:

[0015] (a) Timing input window The input is fed into the differential enhancement DA module and processed according to equation (1). Perform a first-order difference operation to extract 16-dimensional difference features, and then concatenate the difference features with the original 16-dimensional features to obtain the temporal enhancement features. ;

[0016] (1)

[0017] in, This represents the difference characteristics at time t. This represents the original 16-dimensional features at time t. This represents the original 16-dimensional features at time t-1;

[0018] (b) Temporal enhancement features The input is fed into the embedded mapping EMB, and then processed through a linear mapping layer. Mapping to model feature dimensions To obtain embedded features ;

[0019] (c) Embedded features The input is fed into a smoothing filter LSF, which smooths the embedded features. Filtering is performed by copying n boundary values ​​at both ends of the sequence using a padding method. The value of n is determined according to equation (2), ensuring that the length of the output sequence is consistent with that of the input, thus obtaining smooth features. To achieve high-frequency noise suppression,

[0020] (2)

[0021] in, This represents the kernel size of the smoothing filter, which is 25. Therefore, n is 12.

[0022] (d) Smoothing features Input into the channel calibration CAL, this channel is... Perform linear transformation calibration to obtain trend characteristics. It is used to characterize the long-term variation of the heat rate of a steam turbine;

[0023] (e) embedding features With trend characteristics The residual characteristics are input into the residual decomposition RES and obtained by calculating the element-by-element difference according to equation (3). It is used to characterize the short-term transient fluctuations and residual noise of the steam turbine heat rate.

[0024] (3)

[0025] in, This indicates the timing window length, with a value of 96. Indicates the feature dimension of the model;

[0026] Sub-step (B): Combine the input encoding with the trend characteristics output by the smoothing filter decomposition module SFED. With residual characteristics As input to the temporal blending feature extraction module (TMM), this module includes a trend blending unit (TMIX), a detail blending unit (SMIX), and a fusion unit (1), which are connected in series. Specifically, it includes:

[0027] (a) Trend characteristics This is the input to the trend fusion unit TMIX, which contains a time-frequency domain fusion feature extraction unit and a fusion unit 2, connected in series. The time-frequency domain fusion feature extraction unit contains two parallel branches; the left branch is the time-domain feature extraction unit, which uses a unidirectional GRU network to extract trend features. Perform time-domain feature extraction and output time-domain trend features. The right branch is a frequency domain feature extraction unit, which extracts trend features. Perform Fast Fourier Transform (RFFT) and learnable weighted summation sequentially. Processed with Inverse Fast Fourier Transform (IRFFT), the output frequency domain trend characteristics are obtained. The left and right branch features are input into fusion unit 2 for feature fusion. The input to fusion unit 2 is the time-domain trend feature. Frequency domain trend characteristics ,Will and By adding each element according to equation (4), the fusion trend characteristics are obtained. ;

[0028] (4)

[0029] (b) Residual characteristics As input to the Detail Mixing Unit (SMIX), this unit contains two structurally identical detail processing units connected in series, and the output is the detail features. Each detail processing unit contains one detail data processing unit and one local convolutional unit (LC), connected serially. The detail data processing unit includes one normalization processing unit (LN) and a two-branch structure, also connected serially. In the two-branch structure, the left branch is the output of the normalization processing unit LN, and the right branch is a gated mapping using Kaan-type gated linear units. The SiLU gate function is used, and the features from the left and right branches are concatenated and input into the local convolutional unit (LC).

[0030] Local convolutional units (LC) contain one one-dimensional convolutional unit (Conv1D), where the kernel size is... Set to 3, use the same padding, and set the step size to 1.

[0031] The input to fusion unit 1 is trend features. and detailed features ,Will and By adding elements one by one according to equation (5), the multi-scale fusion features are obtained. ;

[0032] (5)

[0033] (c) The fusion and residual compensation output module OUT includes one main branch and one side branch. The main and side branches perform feature concatenation, and the input of the main branch is the multi-scale fused feature. This branch contains one main prediction projection PROJ; the side branch input is the original time series window. This branch contains one linear residual compensation LRB, where:

[0034] ① The master prediction projection PROJ is used to fuse features at multiple scales according to equation (6). By performing vectorization and mapping to a scalar through a linear projection layer, the master predicted value of the heat rate is obtained. ;

[0035] (6)

[0036] in, Weighted by learnable weights, For vectorization operations, A learnable bias vector;

[0037] ② Linear residual compensation (LRB) is applied to the original time window according to equation (7). Perform vectorization operations and output linear residual compensation terms through a linear layer. This is used to compensate for the systematic bias of the master prediction model;

[0038] (7)

[0039] in, For linear residuals, learnable weights are used to weight them. The learnable bias vector for linear residuals;

[0040] ③ The master forecast value With linear residual compensation term By adding the scalars according to equation (8), the final predicted value of the turbine heat rate is obtained. ;

[0041] (8)

[0042] Step 3: Train the prediction network constructed in Step 2 using the training set built in Step 1. Use the mean squared error (MSE) as the loss function and introduce an L2 regularization term to prevent overfitting. The total loss function is... Set the training batch size to 32, the initial learning rate to 0.0005, and save the converged network parameters and the trained prediction model.

[0043] Step 4: Deploy the trained prediction model to the edge computing platform of the coal-fired power plant, collect 16-dimensional runtime sequence data of the steam turbine in real time, and construct a time sequence input window with a length of L=96. And perform the same preprocessing operation as in step 1; input the preprocessed timing sequence into the window. The input prediction model is processed sequentially through three modules: Input Encoding and Smoothing Filtering Decomposition (SFED), Temporal Hybrid Feature Extraction (TMM), and Fusion and Residual Compensation Output (OUT). The output module then calculates the predicted turbine heat rate for the current moment. .

[0044] The beneficial effects of the present invention are as follows: A method for predicting the heat rate of a steam turbine based on smoothing filtering and time series hybridization includes the following steps: (1) collecting the operating sequence data of a steam turbine in a coal-fired power plant and constructing a prediction dataset; (2) constructing an online prediction network for the heat rate of a steam turbine based on smoothing filtering and time series hybridization; (3) training the prediction network constructed in step 2 using the training set constructed in step 1; and (4) deploying the trained prediction model to the edge computing platform of a coal-fired power plant. Compared with the prior art, it has the following significant advantages: First, the input encoding and smoothing filtering decomposition module SFED effectively suppresses high-frequency noise and outliers in the steam turbine operating data through differential enhancement DA and smoothing filtering LSF. At the same time, it decomposes the time series features into trend features and residual features, realizing the effective separation of long-term trend patterns and short-term transient fluctuations, providing high-quality time series feature input for subsequent feature extraction, and solving the problem of decreased prediction accuracy caused by data noise in traditional models. Secondly, the Time Series Hybrid Feature Extraction (TMM) module, through the time-domain-frequency domain parallel structure of the Trend Hybrid Unit (TMIX), simultaneously captures the long-term time-domain trend and frequency-domain feature patterns of the heat rate. Through the gated mapping and local convolution of the Detail Hybrid Unit (SMIX), it accurately extracts short-term detail fluctuations from the residual features. This achieves full-dimensional modeling of multi-scale time series features of the heat rate, significantly improving the model's adaptability and robustness to turbine operation under varying conditions. Thirdly, the Fusion and Residual Compensation Output (OUT) module achieves efficient fusion of trend and detail features through element-wise addition. Simultaneously, it introduces a linear residual compensation mechanism, using the original time series data as input to compensate for the system bias of the main prediction model, effectively reducing the accumulation of prediction errors and further improving the accuracy of heat rate prediction, thus solving the problem of large prediction bias in a single model.

[0045] In summary, the method of this invention effectively solves the problems of noise sensitivity, insufficient modeling of time series features, poor real-time performance, and difficulty in balancing accuracy and lightweighting in existing steam turbine heat rate prediction. It significantly improves the accuracy, stability, and real-time performance of online heat rate prediction, and provides strong technical support for intelligent operation, energy saving, consumption reduction, and optimized scheduling of coal-fired power plant units. Attached Figure Description

[0046] Figure 1 This is a schematic diagram of the overall structure of the online prediction network for turbine heat rate based on a hybrid smoothing filter and time series method, as described in this invention.

[0047] Figure 2 This is a schematic diagram of the input encoding and smoothing filter decomposition module SFED of the present invention.

[0048] Figure 3 This is a schematic diagram of the temporal hybrid feature extraction module (TMM) of the present invention.

[0049] Figure 4This is a schematic diagram of the time-frequency domain parallel feature extraction structure of the trend hybrid branch TMIX of this invention.

[0050] Figure 5 This is a schematic diagram of the structure of the SMIX basic unit of the present invention.

[0051] Figure 6 This is a schematic diagram of the structure of the OUT module for fusion and residual compensation of the present invention.

[0052] Figure 7 This is a comparison chart of the turbine heat rate prediction results of the model of this invention and existing methods.

[0053] Figure 8 This is a comparison chart of the prediction results of the ablation experimental test set of the model of this invention.

[0054] Figure 9 This is a comparison chart of the error indices of the ablation experiment using the model of this invention.

[0055] Figure 10 This is a comparison chart of the fitting indices of the ablation experiment of the model of this invention.

[0056] Figure 11 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0057] The invention will be further described below with reference to the accompanying drawings, such as... Figure 11 As shown, a method for predicting the heat rate of a steam turbine based on a hybrid smoothing filter and time series analysis includes the following steps:

[0058] Step 1: Collect operating time series data of the steam turbine in the coal-fired power plant and construct a prediction dataset. Collect key operating parameters of the steam turbine in the coal-fired power plant, including 16-dimensional continuous time series variables such as main steam pressure, main steam temperature, reheat steam pressure, reheat steam temperature, main steam flow rate, exhaust steam pressure, feedwater temperature, and unit load. The sampling frequency is 1 minute, and a time series input window with a length of L=96 is constructed. The turbine heat consumption rate at the next moment in the corresponding window is the labeled data; the original data is filled with missing values ​​by linear interpolation. Outlier removal and Z-score standardization preprocessing were performed according to principles. A total of 8,000 valid time series samples were collected, covering typical operating conditions such as steady-state operation, variable load regulation, and start-up and shutdown operations of the unit. The samples were divided into a 70% training set, a 15% test set, and a 15% validation set.

[0059] Step 2: Construct an online prediction network for turbine heat rate based on a hybrid smoothing filter and time series analysis. The network structure is as follows: Figure 1As shown, the network includes an input encoding and smoothing filtering decomposition module (SFED), a temporal hybrid feature extraction module (TMM), and a fusion and residual compensation output module (OUT), specifically including the following sub-steps:

[0060] Sub-step (A) inputs timing information into the window. As the input to the input encoding and smoothing filter decomposition module SFED, the module structure is as follows: Figure 2 As shown, it includes five parts: Differential Enhancement (DA), Embedded Mapping (EMB), Smoothing Filter (LSF), Channel Calibration (CAL), and Residual Decomposition (RES), which are connected in sequence. Specifically, it includes:

[0061] (a) Timing input window The input is fed into the differential enhancement DA module and processed according to equation (1). Perform a first-order difference operation to extract 16-dimensional difference features, and then concatenate the difference features with the original 16-dimensional features to obtain the temporal enhancement features. ;

[0062] (1)

[0063] in, This represents the difference characteristics at time t. This represents the original 16-dimensional features at time t. This represents the original 16-dimensional features at time t-1;

[0064] (b) Temporal enhancement features The input is fed into the embedded mapping EMB, and then processed through a linear mapping layer. Mapping to model feature dimensions To obtain embedded features ;

[0065] (c) Embedded features The input is fed into a smoothing filter LSF, which smooths the embedded features. Filtering is performed by copying n boundary values ​​at both ends of the sequence using a padding method. The value of n is determined according to equation (2), ensuring that the length of the output sequence is consistent with that of the input, thus obtaining smooth features. To achieve high-frequency noise suppression,

[0066] (2)

[0067] in, This represents the kernel size of the smoothing filter, which is 25. Therefore, n is 12.

[0068] (d) Smoothing features Input into the channel calibration CAL, this channel is... Perform linear transformation calibration to obtain trend characteristics. It is used to characterize the long-term variation of the heat rate of a steam turbine;

[0069] (e) embedding features With trend characteristics The residual characteristics are input into the residual decomposition RES and obtained by calculating the element-by-element difference according to equation (3). It is used to characterize the short-term transient fluctuations and residual noise of the steam turbine heat rate.

[0070] (3)

[0071] in, This indicates the timing window length, with a value of 96. Indicates the feature dimension of the model;

[0072] Sub-step (B): Combine the input encoding with the trend characteristics output by the smoothing filter decomposition module SFED. With residual characteristics As input to the Temporal Hybrid Feature Extraction (TMM) module, the module structure is as follows: Figure 3 As shown, it includes a trend mixing unit TMIX, a detail mixing unit SMIX, and a fusion unit 1, which are connected in series, and specifically include:

[0073] (a) Trend characteristics As the input to the trend mixing unit TMIX, this module has the following structure: Figure 4 As shown, it includes a time-frequency domain hybrid feature extraction unit and a fusion unit 2, which are connected in series. The time-frequency domain hybrid feature extraction unit contains two parallel branches. The left branch is the time-domain feature extraction unit, which uses a unidirectional GRU network to extract trend features. Perform time-domain feature extraction and output time-domain trend features. The right branch is a frequency domain feature extraction unit, which extracts trend features. Perform Fast Fourier Transform (RFFT) and learnable weighted summation sequentially. Processed with Inverse Fast Fourier Transform (IRFFT), the output frequency domain trend characteristics are obtained. The left and right branch features are input into fusion unit 2 for feature fusion. The input to fusion unit 2 is the time-domain trend feature. Frequency domain trend characteristics ,Will and By adding each element according to equation (4), the fusion trend characteristics are obtained. ;

[0074] (4)

[0075] (b) Residual characteristics As input to the Detail Mixing Unit (SMIX), this module structure, as follows: Figure 5 As shown, it contains two identical detail processing units connected in series, and the output is detail features. Each detail processing unit contains one detail data processing unit and one local convolutional unit (LC), connected serially. The detail data processing unit includes one normalization processing unit (LN) and a two-branch structure, also connected serially. In the two-branch structure, the left branch is the output of the normalization processing unit LN, and the right branch is a gated mapping using Kaan-type gated linear units. The SiLU gate function is used, and the features from the left and right branches are concatenated and input into the local convolutional unit (LC).

[0076] Local convolutional units (LC) contain one one-dimensional convolutional unit (Conv1D), where the kernel size is... Set to 3, use the same padding, and set the step size to 1.

[0077] The input to fusion unit 1 is trend features. and detailed features ,Will and By adding elements one by one according to equation (5), the multi-scale fusion features are obtained. ;

[0078] (5)

[0079] (c) The OUT structure of the fusion and residual compensation output module, such as Figure 6 As shown, it includes one main branch and one side branch. The main and side branches perform feature concatenation, and the input of the main branch is multi-scale fused features. This branch contains one main prediction projection PROJ; the side branch input is the original time series window. This branch contains one linear residual compensation LRB, where:

[0080] ① The master prediction projection PROJ is used to fuse features at multiple scales according to equation (6). By performing vectorization and mapping to a scalar through a linear projection layer, the master predicted value of the heat rate is obtained. ;

[0081] (6)

[0082] in, Weighted by learnable weights, For vectorization operations, A learnable bias vector;

[0083] ② Linear residual compensation (LRB) is applied to the original time window according to equation (7). Perform vectorization operations and output linear residual compensation terms through a linear layer. This is used to compensate for the systematic bias of the master prediction model;

[0084] (7)

[0085] in, For linear residuals, learnable weights are used to weight them. The learnable bias vector for linear residuals;

[0086] ③ The master forecast value With linear residual compensation term By adding the scalars according to equation (8), the final predicted value of the turbine heat rate is obtained. ;

[0087] (8)

[0088] Step 3: Train the prediction network constructed in Step 2 using the training set built in Step 1. Use the mean squared error (MSE) as the loss function and introduce an L2 regularization term to prevent overfitting. The total loss function is... Set the training batch size to 32, the initial learning rate to 0.0005, and save the converged network parameters and the trained prediction model.

[0089] Step 4: Deploy the trained prediction model to the edge computing platform of the coal-fired power plant, collect 16-dimensional runtime sequence data of the steam turbine in real time, and construct a time sequence input window with a length of L=96. And perform the same preprocessing operation as in step 1; input the preprocessed timing sequence into the window. The input prediction model is processed sequentially through three modules: Input Encoding and Smoothing Filtering Decomposition (SFED), Temporal Hybrid Feature Extraction (TMM), and Fusion and Residual Compensation Output (OUT). The output module then calculates the predicted turbine heat rate for the current moment. .

[0090] To verify the effectiveness of the method of this invention, the model of this invention was compared with existing mainstream prediction methods: XGBoost, SVR, GRU and LSTM on a test set. The comparison of heat dissipation rate prediction results is shown in the figure below. Figure 7 As shown.

[0091] from Figure 7As can be seen, the model presented in this paper exhibits significant advantages in predicting the heat rate of steam turbines: it closely matches the actual value curve globally, with virtually no significant lag or deviation, and effectively reproduces the long-term time-series trend of the heat rate; in the local high-frequency fluctuation range of 80–200 sample points, it can capture the dense fluctuations of the actual value well, with high matching degree between peak and valley values, effectively preserving transient details; in the extreme operating condition fluctuation range of 800–1000 sample points, it still maintains a relatively stable fit, without significant deviation amplification; compared with the smoothing distortion problem of XGBoost and SVR, and the lag deviation phenomenon of GRU / LSTM, the model presented in this paper performs better in trend fitting, detail capture, and operating condition robustness, and can better meet the engineering requirements of high-precision online prediction of heat rate under all operating conditions of coal-fired power plants.

[0092] Because the model in this paper contains several key modules, including the smoothing filter LSF, the trend mixing unit TMIX, the detail mixing unit SMIX, and the global residual connection, the effectiveness of the model is verified by removing each module one by one and comparing the degree of performance degradation. The comparison of prediction results on the model ablation experiment test set is shown in the figure below. Figure 8 As shown. From Figure 8 As can be seen, the model in this paper is significantly better than the variant model after removing each module in terms of global trend fitting, local detail capture and robustness to extreme conditions, which intuitively verifies the necessity of each core module and the superiority of the complete model in this paper.

[0093] Figure 9 This is a comparison chart of error indices in ablation experiments, from... Figure 9 As can be seen, the complete model in this paper has the lowest values ​​among all groups in terms of the three core error indicators: mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE×100). However, after removing the smoothing filter (LSF), trend mixing unit (TMIX), detail mixing unit (SMIX), and any module of global residual connection, all errors increased to varying degrees. This directly verifies the necessity of each core module in reducing prediction error and highlights the comprehensive advantages of the complete model in this paper in terms of accuracy and stability.

[0094] Figure 10 This is a comparison chart of fitting indices from ablation experiments. Figure 10 The results show that the determination coefficient of the complete model in this paper is... Both the explained variance (EV) and the value of the model were the highest among all comparison groups, significantly outperforming the variant models with each removed module. This was achieved after removing the smoothing filter (LSF), trend mixing unit (TMIX), detail mixing unit (SMIX), and any one of the core modules of the global residual connection. Both the EV and other metrics showed varying degrees of decline, with the most significant decrease occurring after removing the Detail Mixing Unit (SMIX). Removing the Smoothing Filter (LSF) and the Trend Mixing Unit (TMIX) also led to a significant decline in the fitting performance. This fully verifies the key role of each core module in improving the model's goodness of fit and explanatory variance, further confirming the superiority of the complete model architecture and the rationality of each module's design.

[0095] In summary, the method of the present invention exhibits significant advantages in both prediction accuracy and adaptability to all operating conditions, and is suitable for online prediction of the heat rate of steam turbines in coal-fired power plants.

Claims

1. A method for predicting heat rate of a steam turbine based on a hybrid of smoothing filter and time series, characterized in that, Includes the following steps: Step 1, collect the time series data of the steam turbine of the coal-fired power plant during operation and build a prediction data set, collect the key operation parameters of the steam turbine of the coal-fired power plant, and build a time series input window with a length of L=96 The steam turbine heat consumption rate corresponding to the next moment of the window is the label data; the original data is preprocessed to obtain effective time series samples, and the training set, test set and validation set are divided in proportion; Step 2: Construct an online prediction network for turbine heat rate based on smoothing filtering and temporal hybrid technology. This network includes an input encoding and smoothing filtering decomposition module (SFED), a temporal hybrid feature extraction module (TMM), and a fusion and residual compensation output module (OUT). Specifically, it includes the following sub-steps: (A) Input timing window As the input to the input encoding and smoothing filter decomposition module SFED, this module consists of five parts: Differential Enhancement (DA), Embedded Mapping (EMB), Smoothing Filter (LSF), Channel Calibration (CAL), and Residual Decomposition (RES), which are connected in sequence. Specifically, they include: (a) Timing input window The input is fed into the differential enhancement DA module and processed according to equation (1). Perform a first-order difference operation to extract 16-dimensional difference features, and then concatenate the difference features with the original 16-dimensional features to obtain the temporal enhancement features. ; （1）； in, This represents the difference characteristics at time t. This represents the original 16-dimensional features at time t. This represents the original 16-dimensional features at time t-1; (b) Temporal enhancement features The input is fed into the embedded mapping EMB, and then processed through a linear mapping layer. Mapping to model feature dimensions To obtain embedded features ; (c) Embedded features The input is fed into a smoothing filter LSF, which smooths the embedded features. Filtering is performed by copying n boundary values ​​at both ends of the sequence using a padding method. The value of n is determined according to equation (2), ensuring that the length of the output sequence is consistent with that of the input, thus obtaining smooth features. ; （2）； in, Indicates the size of the smoothing filter kernel; (d) Smoothing features Input into the channel calibration CAL, this channel is... Perform linear transformation calibration to obtain trend characteristics. It is used to characterize the long-term variation of the heat rate of a steam turbine; (e) embedding features With trend characteristics The residual characteristics are input into the residual decomposition RES and obtained by calculating the element-by-element difference according to equation (3). It is used to characterize the short-term transient fluctuations and residual noise of the steam turbine heat rate. （3）； in, This indicates the timing window length, with a value of 96. Indicates the feature dimension of the model; (B) Trend characteristics of the input encoding and the output of the smoothing filter decomposition module SFED. With residual characteristics As input to the temporal blending feature extraction module (TMM), this module includes a trend blending unit (TMIX), a detail blending unit (SMIX), and a fusion unit (1), which are connected in series. Specifically, it includes: (a) Trend characteristics This is the input to the trend fusion unit TMIX, which contains a time-frequency domain fusion feature extraction unit and a fusion unit 2, connected in series. The time-frequency domain fusion feature extraction unit contains two parallel branches; the left branch is the time-domain feature extraction unit, which uses a unidirectional GRU network to extract trend features. Perform time-domain feature extraction and output time-domain trend features. The right branch is a frequency domain feature extraction unit, which extracts trend features. Perform Fast Fourier Transform (RFFT) and learnable weighted summation sequentially. Processed with Inverse Fast Fourier Transform (IRFFT), the output frequency domain trend characteristics are obtained. The left and right branch features are input into fusion unit 2 for feature fusion. The input to fusion unit 2 is the time-domain trend feature. Frequency domain trend characteristics ,Will and By adding each element according to equation (4), the fusion trend characteristics are obtained. ; （4）； (b) Residual characteristics As input to the Detail Mixing Unit (SMIX), this unit contains two structurally identical detail processing units connected in series, and the output is the detail features. Each detail processing unit contains one detail data processing unit and one local convolutional unit (LC), connected serially. The detail data processing unit includes one normalization processing unit (LN) and a two-branch structure, also connected serially. In the two-branch structure, the left branch is the output of the normalization processing unit LN, and the right branch is a gated mapping using Kaan-type gated linear units. The SiLU gate function is used, and the features from the left and right branches are concatenated and input into the local convolutional unit (LC). Local convolutional units (LC) contain one one-dimensional convolutional unit (Conv1D), where the kernel size is... Set to 3, use the same padding, and set the step size to 1. The input to fusion unit 1 is trend features. and detailed features ,Will and By adding elements one by one according to equation (5), the multi-scale fusion features are obtained. ; （5）； (c) The fusion and residual compensation output module OUT includes one main branch and one side branch. The main and side branches perform feature concatenation, and the input of the main branch is the multi-scale fused feature. This branch contains one main prediction projection PROJ; the side branch input is the original time series window. This branch contains one linear residual compensation LRB, where: ① The master prediction projection PROJ is used to fuse features at multiple scales according to equation (6). By performing vectorization and mapping to a scalar through a linear projection layer, the master predicted value of the heat rate is obtained. ; （6）； in, Weighted by learnable weights, For vectorization operations, A learnable bias vector; ② Linear residual compensation (LRB) is applied to the original time window according to equation (7). Perform vectorization operations and output linear residual compensation terms through a linear layer. This is used to compensate for the systematic bias of the master prediction model; （7）； in, For linear residuals, learnable weights are used to weight them. The learnable bias vector for linear residuals; ③ The master forecast value With linear residual compensation term By adding the scalars according to equation (8), the final predicted value of the turbine heat rate is obtained. ; （8）； Step 3: Train the prediction network constructed in Step 2 using the training set built in Step 1. Use the mean squared error (MSE) as the loss function and introduce an L2 regularization term to prevent overfitting. The total loss function is... Set the training batch size to 32, the initial learning rate to 0.0005, and save the converged network parameters and the trained prediction model. Step 4: Deploy the trained prediction model to the edge computing platform of the coal-fired power plant, collect 16-dimensional runtime sequence data of the steam turbine in real time, and construct a time sequence input window with a length of L=96. And perform the same preprocessing operation as in step 1; input the preprocessed timing sequence into the window. The input prediction model is processed sequentially through three modules: Input Encoding and Smoothing Filtering Decomposition (SFED), Temporal Hybrid Feature Extraction (TMM), and Fusion and Residual Compensation Output (OUT). The output module then calculates the predicted turbine heat rate for the current moment. .

2. The method for predicting turbine heat rate based on a hybrid smoothing filter and time series analysis as described in claim 1, characterized in that, The key operating parameters of the steam turbine in the coal-fired power plant include main steam pressure, main steam temperature, reheat steam pressure, reheat steam temperature, main steam flow rate, exhaust steam pressure, feedwater temperature, and unit load.

3. The method for predicting turbine heat rate based on a hybrid smoothing filter and time series analysis as described in claim 1, characterized in that, In step 1, preprocessing involves linear interpolation to fill missing values, 3 Outlier removal and Z-score standardization preprocessing are implemented.

4. The method for predicting turbine heat rate based on a hybrid smoothing filter and time series analysis as described in claim 1, characterized in that, In step 1, the sample covers the steady-state operation, variable load regulation, and start-up and shutdown operation conditions of the unit.

Images