Main steam temperature prediction method and system based on multi-scale time-frequency network
By using a multi-scale time-frequency network method, the problems of insufficient adaptive modeling of main steam temperature prediction in the frequency domain and the inadequacy of generalization ability under complex operating conditions are solved, and high-precision and robust temperature prediction is achieved, which is suitable for intelligent control systems of thermal power units.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- XIAN THERMAL POWER RES INST CO LTD
- Filing Date
- 2026-01-21
- Publication Date
- 2026-04-28
AI Technical Summary
Existing main steam temperature prediction methods are insufficient in terms of adaptive modeling of frequency domain features, collaborative extraction of multi-scale time series features, and generalization ability under complex operating conditions, making it difficult to meet the requirements of intelligent control systems for thermal power units for high-precision and robust prediction models.
A method based on multi-scale time-frequency networks is adopted to construct a main steam temperature prediction system through data preprocessing, time series decomposition, multi-scale collaborative convolution, adaptive wavelet frequency domain enhancement, and operation mode clustering, so as to achieve unified modeling of multi-scale, multi-frequency, and multi-operation mode characteristics.
It improves the robustness and accuracy of prediction results, can adapt to rapid load changes and complex operating conditions, reduces computational complexity, facilitates integration with existing thermal power unit systems, and improves safety and economy.
Smart Images
Figure CN121936299A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent operation and process parameter prediction technology of thermal power units, specifically involving a method and system for predicting main steam temperature based on a multi-scale time-frequency network. Background Technology
[0002] Main steam temperature is one of the most critical thermal parameters in the boiler and turbine systems of thermal power units. Its control accuracy directly affects the unit's thermal efficiency, equipment safety, and long-term operational reliability. In typical thermal power units, the main steam temperature needs to be stably maintained within a narrow allowable range. Excessive temperature will exacerbate oxidation and creep of the heating surface materials, shortening equipment lifespan; excessively low temperature will lead to decreased turbine efficiency, increasing coal consumption and operating costs. Therefore, accurate and stable prediction of the main steam temperature is a crucial foundation for achieving safe and economical operation and intelligent control of thermal power units.
[0003] In actual operation, the main steam temperature is affected by a combination of factors, including fuel quality, load changes, feedwater regulation, desuperheating water operation, flue gas heat exchange status, and environmental conditions. Its time series exhibits significant nonlinearity, non-stationarity, and multi-timescale characteristics. Especially under conditions of deep peak shaving, rapid load increases and decreases, and frequent start-ups and shutdowns, the main steam temperature series often simultaneously contains both slowly changing long-term trends and high-frequency disturbances. The importance of different frequency components varies significantly at different operating stages, posing a considerable challenge to the modeling capability and generalization performance of traditional prediction models.
[0004] Existing main steam temperature prediction methods mainly fall into three categories: The first category consists of statistical model-based methods, such as the Autoregressive Moving Average (ARIMA) model. These methods have simple structures but typically assume stationarity of the sequence, making them difficult to adapt to load changes and operating condition switching scenarios, and exhibiting significant prediction lag. The second category comprises prediction methods based on traditional neural networks, such as feedforward neural networks and recurrent neural networks (LSTM). While these methods can characterize time dependencies to some extent, their ability to model multi-scale features and high-frequency disturbances is limited, and they are prone to overfitting or prediction instability under complex operating conditions. The third category consists of time series prediction methods based on Transformer structures proposed in recent years. These methods have advantages in long-sequence modeling, but most focus on pure time-domain modeling or fixed frequency-domain representation, resulting in insufficient frequency-domain adaptability. Furthermore, their complex model structures and high computational costs make their deployment and application in engineering scenarios challenging.
[0005] Furthermore, existing methods generally neglect the multimodal characteristics exhibited by thermal power units under different operating stages and load ranges, lacking explicit modeling of differences in operating modes, which limits the generalization ability of models when predicting across operating conditions. In summary, current main steam temperature prediction methods still have shortcomings in areas such as adaptive modeling of frequency domain features, collaborative extraction of multi-scale time-series features, and generalization ability under complex operating conditions, making it difficult to meet the actual needs of intelligent control systems for thermal power units for high-precision, highly robust prediction models. Summary of the Invention
[0006] This invention provides a method and system for predicting main steam temperature based on multi-scale time-frequency networks. The purpose is to address the shortcomings of current main steam temperature prediction methods in terms of adaptive modeling of frequency domain features, collaborative extraction of multi-scale time-series features, and generalization ability under complex operating conditions. These shortcomings make it difficult to meet the actual needs of intelligent control systems for thermal power units for high-precision and robust prediction models.
[0007] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides a method for predicting main steam temperature based on a multi-scale time-frequency network, comprising the following steps: S1. Collect multivariate operating time series data from the distributed control system of thermal power units, and perform time alignment, outlier removal, missing data filling and normalization in sequence to obtain preprocessed multivariate time series data. S2. Perform trend-seasonal decomposition on the preprocessed multivariate time series data to obtain trend components and seasonal components. S3. Input the seasonal components into the pre-built and trained multi-scale collaborative convolutional module to extract dynamic perturbation features at different time scales and obtain multi-scale joint features. S4. Input the trend components into the pre-built and trained clustering matching module, and model and reconstruct the long-term trend pattern through the learnable cluster centers to obtain the trend prediction features. S5. Based on the pre-built and trained adaptive wavelet frequency domain enhancement module, perform adaptive wavelet decomposition and frequency domain filtering on the multi-scale joint features to obtain frequency domain enhanced features; S6. The trend prediction features and frequency domain enhancement features are fused together, and the temperature prediction results of the main steam are output through the linear prediction layer.
[0008] In some implementations, in S1, the multivariate operating time series data includes main steam temperature, unit load, fuel quantity, feedwater flow rate, and desuperheating water flow rate.
[0009] In some implementations, in S1, time alignment includes timestamp alignment and resampling of operational data from different sources and with different sampling frequencies, unifying all variables to the same sampling period; outlier removal includes identifying and removing anomalies such as sensor disconnection, sudden spikes, and long-term drift in the variable's operational time series data; missing data imputation includes imputing missing data using moving average interpolation, linear interpolation, or forward imputation methods; normalization is performed according to the following formula: ; in, These are the original observations. and These represent the mean and standard deviation of the corresponding variable in the training samples, respectively. This is the normalized value.
[0010] In some implementations, S2 specifically includes: S21. Input multivariate time series Construct a moving average operator to obtain an initial estimate of the trend component: ; in, The length of the sliding window. For time indexing, for Multivariate time series data at any given time; S22. Use copying or mirroring to complete the time series boundaries so that the moving average can be calculated at the beginning and end of the series, and calculate the trend component based on the completed series. S23. Subtract the trend component from the preprocessed multivariate time series data to obtain the seasonal component: ; S24. Ensure the following reconstruction consistency relationship is satisfied before and after decomposition: ; in, For preprocessed multivariate time series data, For seasonal portions, For trend components.
[0011] In some implementations, S3 specifically includes: S31. Parallel convolution operations are performed on the seasonal component time series using one-dimensional convolution kernels of different sizes: ; in, This is a one-dimensional convolution operation. The kernel size is the convolution kernel size. The output features of the convolution kernel of the corresponding size; S32. Concatenate the outputs of convolutions at each scale along the channel dimension to form multi-scale joint features: ; in, It is a multi-scale joint feature; S32. Map the multi-scale joint features into query vector, key vector, and value vector, and input them into a multi-head self-attention mechanism: ; in, These are the query vector, key vector, and value vector, respectively. For the attention calculation results, For matrix transpose, For vector dimensions, For activation functions; S32. Perform channel reconstruction and feature compression on the attention output to obtain the output features of the multi-scale co-convolution module.
[0012] In some implementations, S4 specifically includes: S41. Embed trend features according to the following formula: ; in, For multilayer perceptron computation, As a trend component, Embedding representation for trend features; S42, Initialization Learnable cluster centers It is used to depict trend patterns under different operating conditions; S43. Calculate the similarity using the following formula: ; in, For the first The first trend feature is embedded in the second... The similarity weight of each cluster center For the first Embedding of trend features, For the first Cluster centers, For the first Cluster centers; S44. Reconstruct trend features according to the following formula: ; in, For the first One trend prediction feature, For similarity weights, For the first Cluster centers.
[0013] In some implementations, S5 specifically includes: S51. Perform wavelet decomposition based on the second-generation wavelet lifting structure on the multi-scale joint features according to the following formula to obtain the low-frequency components and high-frequency components: ; in, For multi-scale joint features, Low-frequency components, These are high-frequency components; S52. Perform high-frequency energy calculation and normalization according to the following formula: ; in, The energy of the high-frequency components, The absolute value of the high-frequency component. Normalized energy This represents the maximum energy value. To prevent the minimum value where the denominator is zero; S53. Construct an adaptive frequency mask: ; in, For frequency mask, For fixed thresholds or learnable parameters; S54. Perform frequency domain feature reconstruction: ; in, This is an inverse wavelet transform operation. For element-wise multiplication, This is a frequency domain enhancement feature.
[0014] In some implementations, S5 specifically includes: S51. The trend prediction features and frequency domain enhancement features are fused using a feature concatenation method: ; in, The characteristics after fusion For trend prediction features, For frequency domain enhancement features; S52. Input the fused features into the linear prediction layer and output the main steam temperature prediction result. ; in, Based on the main steam temperature prediction results, These are the weight parameters for the linear prediction layer. The characteristics after fusion This is the bias parameter.
[0015] In some implementations, the multi-scale collaborative convolution module includes multi-scale convolution branches, cross-scale attention units, and feature reconstruction units; the clustering matching module includes trend feature embedding units, cluster center initialization units, similarity calculation units, and trend pattern reconstruction units; and the adaptive wavelet frequency domain enhancement module includes wavelet decomposition units, high-frequency energy calculation units, frequency mask construction units, and frequency domain reconstruction units.
[0016] This invention also provides a main steam temperature prediction system based on a multi-scale time-frequency network to implement the above-mentioned main steam temperature prediction method based on a multi-scale time-frequency network. The system includes a data preprocessing module, a time series decomposition module, a multi-scale collaborative convolution module, a clustering matching module, an adaptive wavelet frequency domain enhancement module, and a prediction output module, wherein: Data preprocessing module: used to collect multivariate operating time series data from the distributed control system of thermal power units, and sequentially perform time alignment, outlier removal, missing data completion and normalization to obtain preprocessed multivariate time series data; Time series decomposition module: used to perform trend and seasonal decomposition on preprocessed multivariate time series data to obtain trend components and seasonal components; Multi-scale co-convolution module: Used to input seasonal components into a pre-built and trained multi-scale co-convolution module to extract dynamic perturbation features at different time scales and obtain multi-scale joint features; Clustering matching module: This module takes the trend components as input to the pre-built and trained clustering matching module, and models and reconstructs the long-term trend pattern through learnable cluster centers to obtain trend prediction features. Adaptive wavelet frequency domain enhancement module: Based on the pre-built and trained adaptive wavelet frequency domain enhancement module, it performs adaptive wavelet decomposition and frequency domain filtering on multi-scale joint features to obtain frequency domain enhanced features; Prediction output module: used to fuse trend prediction features with frequency domain enhancement features, and output the temperature prediction result of the main steam through the linear prediction layer.
[0017] Compared with existing technologies, the main steam temperature prediction method and system based on multi-scale time-frequency networks of the present invention have the following advantages: This invention presents a main steam temperature prediction method based on a multi-scale time-frequency network. In the seasonal branch, it introduces an adaptive wavelet frequency domain enhancement approach. By evaluating and gating high-frequency energy, it dynamically strengthens key frequency components that significantly influence the prediction results while suppressing the influence of noise bands and irrelevant disturbances. This enables the model to adaptively cope with non-stationary disturbances caused by rapid load changes and the introduction of cooling water, improving the robustness of the prediction results. By constructing a multi-scale collaborative convolutional structure and combining it with a cross-scale attention mechanism, this invention can simultaneously extract short-period fluctuation features and long-period variation features under different time receptive fields, achieving effective cross-scale information interaction. This improves the problem of single-scale models struggling to simultaneously consider local details and global trends, enhancing the ability to express complex time-series structures. Furthermore, this invention introduces operating mode clustering matching in the trend branch. Through learnable cluster centers, it explicitly models trend features under different load intervals, peak-shaving phases, and operating modes. This allows the model to adaptively select appropriate trend prototypes for prediction based on the current operating conditions, effectively reducing the impact of operating condition switching on prediction performance and improving cross-operating condition generalization ability.
[0018] On the other hand, in the main steam temperature prediction system based on multi-scale time-frequency network of the present invention, each module can be implemented through existing deep learning frameworks and industrial computing platforms, which reduces the computational complexity and is relatively easy to integrate and deploy with the existing distributed control system and intelligent monitoring system of thermal power units. It provides data support for the early prediction and auxiliary control of main steam temperature, and improves the safety factor and economy of thermal power units through intelligent operation. Attached Figure Description
[0019] The accompanying drawings are provided to further understand the invention and constitute a part of this invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0020] Figure 1 This is a flowchart illustrating a method for predicting main steam temperature based on a multi-scale time-frequency network according to the present invention. Figure 2 This is a schematic diagram of the data preprocessing and time series decomposition module in the main steam temperature prediction method based on multi-scale time-frequency network of the present invention; Figure 3 This is a schematic diagram of the seasonal branch multi-scale collaborative convolution modeling module in the main steam temperature prediction method based on multi-scale time-frequency network of the present invention; Figure 4 This is a schematic diagram of the adaptive wavelet frequency domain enhancement module in the main steam temperature prediction method based on a multi-scale time-frequency network of the present invention; Figure 5This is a schematic diagram of the trend branch operation mode clustering and matching module in the main steam temperature prediction method based on multi-scale time-frequency network of the present invention; Figure 6 This is a schematic diagram of the trend branch and seasonal branch feature fusion and prediction output module in the main steam temperature prediction method based on multi-scale time-frequency network of the present invention; Figure 7 This is a schematic diagram illustrating the prediction performance of the MSTLNet model on the main steam temperature dataset (MST) of thermal power units in the main steam temperature prediction method based on multi-scale time-frequency networks of the present invention. Figure 8 This is a schematic diagram showing the performance comparison of different prediction models on the MST dataset in the main steam temperature prediction method based on multi-scale time-frequency network of the present invention. Figure 9 This is a schematic diagram comparing the ablation experimental performance of the MSTLNet model in the main steam temperature prediction method based on multi-scale time-frequency network of the present invention. Figure 10 This diagram illustrates a comprehensive comparison of the prediction accuracy and computational efficiency of the model on the MST dataset in the main steam temperature prediction method based on a multi-scale time-frequency network according to the present invention. Detailed Implementation
[0021] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0022] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0023] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0024] It should be noted that the apparatus and methods disclosed in the embodiments herein can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments herein. In this regard, each block in a flowchart or block diagram may represent a module, program, or part of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system to perform the specified function or action, or can be implemented using a combination of dedicated hardware and computer instructions.
[0025] In addition, the functional modules in the various embodiments of this article can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0026] How can we provide a method for predicting main steam temperature that can simultaneously characterize multi-scale temporal features, dynamically enhance key frequency domain information, and distinguish different operating modes, so as to improve prediction performance and provide reliable technical support for the intelligent operation and safe control of thermal power units?
[0027] Based on this, such as Figure 1 As shown, this invention provides a method for predicting main steam temperature based on a multi-scale time-frequency network, comprising the following steps: S1. Collect multivariate operating time series data from the distributed control system of thermal power units, and perform time alignment, outlier removal, missing data filling and normalization in sequence to obtain preprocessed multivariate time series data. S2. Perform trend-seasonal decomposition on the preprocessed multivariate time series data to obtain trend components and seasonal components. S3. Input the seasonal components into the pre-built and trained multi-scale collaborative convolutional module to extract dynamic perturbation features at different time scales and obtain multi-scale joint features. S4. Input the trend components into the pre-built and trained clustering matching module, and model and reconstruct the long-term trend pattern through the learnable cluster centers to obtain the trend prediction features. S5. Based on the pre-built and trained adaptive wavelet frequency domain enhancement module, perform adaptive wavelet decomposition and frequency domain filtering on the multi-scale joint features to obtain frequency domain enhanced features; S6. The trend prediction features and frequency domain enhancement features are fused together, and the temperature prediction results of the main steam are output through the linear prediction layer.
[0028] This invention preprocesses and decomposes the multivariate operating time series of thermal power units, dividing the original sequence into trend components and seasonal disturbance components, thereby modeling long-term evolution characteristics and short-term fluctuation characteristics respectively. In the seasonal branch, a multi-scale collaborative convolutional structure is introduced to extract dynamic disturbance information under different time sensing fields, and an adaptive wavelet frequency domain enhancement mechanism is combined to dynamically screen and strengthen key frequency components, improving the model's ability to model non-stationary disturbances and multi-frequency components. In the trend branch, an operating mode clustering and matching mechanism is constructed to explicitly model trend patterns under different load ranges and operating stages, improving the model's prediction stability and generalization ability in complex operating condition switching scenarios. Through multi-branch collaborative modeling and feature fusion, this invention achieves unified modeling of multi-scale, multi-frequency, and multi-operating mode characteristics of main steam temperature, maintaining high prediction accuracy and robustness under deep peak shaving and rapid load changes, and can be applied to multiple scenarios of thermal power units, demonstrating certain applicability.
[0029] In some operating conditions, the present invention provides a main steam temperature prediction method based on a multi-scale time-frequency network, which is carried out according to the following steps: Step A: Data Acquisition and Preprocessing; Step A1: Data Acquisition; Multivariate operating time series data are collected from the distributed control system (DCS) of thermal power units. The multivariate time series data includes at least the main steam temperature, unit load, fuel quantity, feedwater flow rate and desuperheating water flow rate, and may optionally include operating parameters closely related to the change of main steam temperature, such as flue gas damper opening, superheater outlet temperature, reheater temperature, oxygen quantity, primary air volume and secondary air volume.
[0030] Step A2: Time alignment and uniform sampling; Timestamp alignment and resampling are performed on runtime data from different sources and with different sampling frequencies to unify all variables to the same sampling period. ,in The time limit can be set to 10 seconds, 15 minutes, or 1 hour; the resampling process is implemented using mean aggregation or linear interpolation.
[0031] Step A3: Outlier removal and missing value handling; The system identifies anomalies in the collected data, such as sensor disconnection, sudden spikes, and long-term drift; it also fills in missing data using moving average interpolation, linear interpolation, or forward imputation methods to ensure the continuity of the time series.
[0032] Step A4: Normalization process; All input variables are normalized using the Z-score standardization method, and the calculation formula is as follows: ; in, These are the original observations. and These represent the mean and standard deviation of the corresponding variable in the training samples, respectively. This is the normalized value.
[0033] Step B: Time series decomposition modeling; To explicitly separate the long-term evolution trend and short-term disturbance characteristics in the main steam temperature time series, the preprocessed multivariate time series was analyzed. Perform trend-seasonal decomposition to obtain trend components. and seasonal portion It satisfies the following relationship: ; Step B1: Trend component extraction; The trend term is constructed using the moving average operator, and its expression is as follows: ; in, The length of the sliding window. For time indexing, for Multivariate time series data at any given time.
[0034] Step B2: Boundary completion processing; By using copy padding or mirror padding at the beginning and end of the time series, the moving average operator can be computed at the boundaries.
[0035] Step B3: Constructing Seasonal Components; Subtracting the trend component from the original sequence yields the seasonal (perturbation) component: ; Step B4: Decompose consistency constraints; Ensure that the reconstruction consistency relationship is satisfied before and after decomposition. To avoid decomposition bias.
[0036] in, For preprocessed multivariate time series data, For seasonal portions, For trend components.
[0037] Step C: Multi-scale collaborative convolution modeling of seasonal branches; To simultaneously characterize the perturbation features at different time scales in the main steam temperature series, the seasonal components were analyzed. Construct a multi-scale collaborative convolution module.
[0038] Step C1: Constructing multi-scale convolutional branches; Parallel convolution modeling of seasonal components is performed using multiple sets of one-dimensional convolution kernels of different sizes: ; Step C2: Multi-scale feature stitching; The outputs of convolutions at different scales are concatenated along the channel dimension to form multi-scale joint features: ; in, It is a multi-scale joint feature; Step C3: Cross-scale attention interaction; Multi-scale joint features are mapped to query vectors, key vectors, and value vectors, and then fed into a multi-head self-attention mechanism. ; in, These are the query vector, key vector, and value vector, respectively. For the attention calculation results, For matrix transpose, For vector dimensions, This is the activation function.
[0039] Step C4: Feature reconstruction output; Channel reconstruction and feature compression are performed on the attention output to obtain the output features of the multi-scale co-convolution module. .
[0040] Step D: Adaptive wavelet frequency domain enhancement; To address the issue of non-stationarity of the frequency components of the main steam temperature sequence, Adaptive wavelet frequency domain enhancement processing is performed.
[0041] Step D1: Wavelet decomposition; ; in, For multi-scale joint features, Low-frequency components, These are high-frequency components.
[0042] Step D2: High-frequency energy calculation and normalization; ; in, The energy of the high-frequency components, The absolute value of the high-frequency component. Normalized energy This represents the maximum energy value. To prevent the minimum value where the denominator is zero; Step D3: Energy gating mask construction; ; in, For frequency mask, For fixed thresholds or learnable parameters: ; Step D4: Frequency domain feature reconstruction; ; in, This is an inverse wavelet transform operation. For element-wise multiplication, For frequency domain enhancement features; Step E: Clustering and matching of the operating pattern of trend branches; To adapt to different load ranges and operating conditions, the trend component... A clustering and matching mechanism based on operating patterns is introduced.
[0043] Step E1, Trend Feature Embedding: ; in, For multilayer perceptron computation, As a trend component, This is an embedding representation of trend features.
[0044] Step E2: Cluster center initialization; initialization Learnable cluster centers It is used to characterize trend prototypes under different operating modes.
[0045] Step E3: Similarity calculation; ; in, For the first The first trend feature is embedded in the second... The similarity weight of each cluster center For the first Embedding of trend features, For the first Cluster centers, For the first Cluster centers; Step E4, Trend Pattern Reconstruction: ; in, For the first One trend prediction feature, For similarity weights, For the first Cluster centers.
[0046] Step F: Feature fusion and prediction output; Step F1, Feature Fusion: The trend prediction features and frequency domain enhancement features are fused using a feature concatenation method: ; in, The characteristics after fusion For trend prediction features, For frequency domain enhancement features; Step F2, Predict the output: The fused features are input into the linear prediction layer, and the main steam temperature prediction result is output. ; in, Based on the main steam temperature prediction results, These are the weight parameters for the linear prediction layer. The characteristics after fusion This is the bias parameter.
[0047] Furthermore, in some embodiments, the input sequence length of the present invention can preferably be 96, and the prediction step size can preferably be 12, 24, 36, or 48; the multi-scale convolutional kernel set can be expanded according to the complexity of the working conditions; frequency domain thresholding. Employing a learnable approach to enhance adaptability under varying load change rates; number of cluster centers The complexity can be set to 4–16 depending on the operating mode.
[0048] This invention also provides a main steam temperature prediction system based on a multi-scale time-frequency network, comprising a data preprocessing module, a time series decomposition module, a multi-scale collaborative convolution module, a clustering matching module, an adaptive wavelet frequency domain enhancement module, and a prediction output module, wherein: Data preprocessing module: used to collect multivariate operating time series data from the distributed control system of thermal power units, and sequentially perform time alignment, outlier removal, missing data completion and normalization to obtain preprocessed multivariate time series data; Time series decomposition module: used to perform trend and seasonal decomposition on preprocessed multivariate time series data to obtain trend components and seasonal components; Multi-scale co-convolution module: Used to input seasonal components into a pre-built and trained multi-scale co-convolution module to extract dynamic perturbation features at different time scales and obtain multi-scale joint features; Clustering matching module: This module takes the trend components as input to the pre-built and trained clustering matching module, and models and reconstructs the long-term trend pattern through learnable cluster centers to obtain trend prediction features. Adaptive wavelet frequency domain enhancement module: Based on the pre-built and trained adaptive wavelet frequency domain enhancement module, it performs adaptive wavelet decomposition and frequency domain filtering on multi-scale joint features to obtain frequency domain enhanced features; Prediction output module: used to fuse trend prediction features with frequency domain enhancement features, and output the temperature prediction result of the main steam through the linear prediction layer.
[0049] The system of the present invention achieves a main steam temperature prediction method based on a multi-scale time-frequency network by having various modules work together.
[0050] The following detailed description of the main steam temperature prediction method and system based on multi-scale time-frequency networks of the present invention will be provided through specific embodiments.
[0051] This embodiment integrates multi-scale time-domain modeling, frequency-domain adaptive enhancement, and operation mode clustering to predict the time series of main steam temperature of thermal power units. It takes multivariate operation time series data collected by the distributed control system (DCS) of thermal power units as input, and achieves high-precision prediction of main steam temperature through data preprocessing, time series decomposition, seasonal branch multi-scale modeling, frequency-domain adaptive enhancement, trend branch operation mode clustering, and feature fusion prediction.
[0052] In this embodiment, the prediction method takes a fixed-length historical time series as input and outputs the predicted value of the main steam temperature within a certain number of future time steps, which can be used for unit operation status assessment and control decision support.
[0053] like Figure 2 The diagram illustrates the process of time alignment, normalization, and trend-seasonal decomposition based on the moving average operator for multivariate operating time series data of thermal power units. This invention first collects multivariate operating time series data from the DCS system of the thermal power unit, including key variables such as main steam temperature, unit load, fuel quantity, feedwater flow rate, and desuperheating water flow rate.
[0054] The collected data underwent time alignment and unified sampling. When the sampling periods for different variables were inconsistent, a resampling method was used to unify the sampling period. Subsequently, outlier data points were removed, and missing values were filled in using moving average or linear interpolation.
[0055] After data cleaning, all variables were normalized using the Z-score standardization method to eliminate the impact of differences in the units of different variables on model training.
[0056] Based on this, a trend-seasonal decomposition is performed on the preprocessed multivariate time series. The long-term trend component is extracted using the moving average operator, and the seasonal component is obtained by subtracting the trend component from the original series, thereby achieving explicit decoupling between long-term evolutionary features and short-term disturbance features.
[0057] like Figure 3 The diagram illustrates the process of parallel feature extraction of seasonal components using one-dimensional convolutional kernels of different scales, and the achievement of multi-scale feature collaborative modeling through a cross-scale attention mechanism. This invention addresses the multi-scale perturbation features contained in seasonal components by constructing a multi-scale collaborative convolutional modeling module. This module employs multiple sets of one-dimensional convolutional kernels of different sizes to perform parallel convolution operations on the seasonal components, thereby extracting dynamic features within the receptive field at different times.
[0058] The outputs of convolutions at each scale are concatenated along the channel dimension to form a multi-scale joint feature representation. Subsequently, the multi-scale joint features are input into a cross-scale attention mechanism, which uses attention weights to achieve interaction and collaboration between features at different time scales, thereby enhancing the model's ability to model complex perturbation structures.
[0059] like Figure 4 The diagram illustrates the processing flow for obtaining frequency-domain enhanced features by performing wavelet decomposition, high-frequency energy calculation, adaptive frequency mask construction, and inverse wavelet reconstruction on multi-scale features. This invention performs adaptive wavelet frequency-domain enhancement processing on the features output by the multi-scale co-convolution module. First, wavelet decomposition is performed on the features to obtain low-frequency and high-frequency components.
[0060] Subsequently, the frequency domain energy of the high-frequency components is calculated, and an adaptive frequency mask is constructed based on the energy distribution to enhance key frequency components and suppress noise bands. Finally, the processed high-frequency and low-frequency components are reconstructed using inverse wavelet transform to obtain the enhanced feature representation in the frequency domain.
[0061] Through this frequency domain enhancement mechanism, this embodiment can effectively improve the model's sensitivity to non-stationary disturbances and key frequency components, thereby enhancing the robustness of the prediction results.
[0062] like Figure 5The diagram illustrates the process of trend feature embedding, operation pattern cluster center initialization, similarity calculation, and trend pattern reconstruction, serving to explain the modeling method of trend features under different operating conditions in this invention. To address the issue of significant differences in trend features under different load ranges and operating conditions, this invention introduces an operation pattern clustering matching mechanism into the trend branches.
[0063] First, the trend components are input into a multilayer perceptron network to obtain the trend feature embedding representation. Then, multiple learnable operating mode cluster centers are initialized to represent trend prototypes under different operating conditions.
[0064] By calculating the similarity between the trend embedding and the cluster centers, matching weights are obtained. Based on these weights, the cluster centers are then reconstructed using a weighted average to form a trend prediction representation. This process enables the model to adaptively select appropriate trend patterns according to the current operating state, improving its cross-condition prediction capabilities.
[0065] like Figure 6 The diagram illustrates the process of fusing trend prediction features with frequency-enhanced seasonal features and outputting the main steam temperature prediction result through a linear prediction layer. This invention fuses the trend prediction features output from the trend branch with the frequency-enhanced features from the seasonal branch. The fusion method employs feature concatenation, and the fused features are input into a linear prediction layer to output the predicted main steam temperature value within the future time step.
[0066] By using collaborative modeling of trend information and disturbance information, this embodiment can improve the model's adaptability to complex operating conditions while maintaining prediction accuracy.
[0067] The present invention conducted the following specific comparative experiments on predictive performance and effect analysis. To verify the performance advantages of the MSTLNet model of this invention in the main steam temperature prediction task, the entire model was implemented using the PyTorch deep learning framework and trained and tested on an NVIDIA GeForce RTX 5090 DV2 graphics processor (GPU) equipped with 24GB of video memory.
[0068] In this invention, the Adam optimizer is used for parameter updates during model training, and the initial learning rate is set to... The mean squared error (MSE) was used as the loss function for optimization. The batch size was set to 32. The input time series length was set to 96, and the prediction length was set to... This is to cover both short-term and medium-to-long-term forecasting scenarios.
[0069] To ensure the fairness of the experiment, all comparison models use the same method of dividing the training set, validation set, and test set, and the evaluation metrics are uniformly the mean squared error (MSE) and the mean absolute error (MAE).
[0070] Furthermore, this invention also compares the MSTLNet model with current mainstream time series prediction models, specifically including: iTransformer: A time series forecasting model based on an inverted Transformer structure; TimesNet: A general time series analysis framework based on two-dimensional time variation modeling; PatchTST: A Transformer forecasting model based on a time series patching mechanism; Crossformer: A multivariate time series forecasting model utilizing cross-dimensional dependencies; DLinear: A linear time series forecasting model based on trend-seasonal decomposition; FreTS: A time series forecasting model based on a frequency domain multilayer perceptron; FEDformer: A Transformer model incorporating a frequency enhancement decomposition mechanism; Autoformer: A Transformer forecasting model based on an autocorrelation decomposition mechanism.
[0071] Experimental results are as follows Figure 7 and Figure 8 As shown. Figure 7 The results show the comparison between the actual observed values of main steam temperature on the MST dataset and the predicted values of the Multi-scale Spectro-Temporal Learning Network (MSTLNet) model at different prediction step sizes, which are used to illustrate the prediction accuracy and dynamic tracking capability of the present invention under complex operating conditions. in, Figure 8 The accuracy advantage of this invention can be demonstrated by comparing its prediction performance with that of various existing mainstream time series prediction models on evaluation metrics such as mean squared error (MSE) and mean absolute error (MAE).
[0072] The results show that, on the MST main steam temperature dataset, the MSTLNet proposed in this invention performs well across four different prediction lengths. They all achieved the best or second-best prediction performance, demonstrating certain advantages.
[0073] In all prediction tasks, MSTLNet significantly outperformed the comparison models in both MSE and MAE metrics, with an average performance improvement of approximately 15%. In particular, in medium- and long-term prediction tasks with prediction step sizes of 36 and 48, MSTLNet maintained low prediction errors and high stability, demonstrating that the present invention has significant advantages in modeling complex working conditions and long-term dependencies.
[0074] like Figure 9 As shown, the present invention also verifies the function of each core module in the model through ablation experiments. Figure 9 By successively removing the Time Series Decomposition Architecture (ETDA), Clustering Matching (CM), Multi-scale Synergistic Convolution (MSC), and Adaptive Wavelet Filtering (AWF) modules, the changes in prediction error under different model structures were compared to illustrate the contribution of each core module to the overall performance improvement and their synergistic effect. The results show that Time Series Decomposition, Multi-scale Synergistic Convolution, Adaptive Wavelet Filtering, and Clustering Matching all contribute positively to prediction performance, and the synergistic effect of multiple modules can significantly improve overall performance.
[0075] like Figure 10 As shown, Figure 10 This paper comprehensively compares the present invention with existing prediction models from two dimensions: prediction accuracy and model complexity. Model complexity is measured by the number of parameters and floating-point operations, illustrating that the present invention achieves high prediction accuracy while maintaining good computational efficiency and engineering deployment applicability. The present invention provides a comprehensive comparative analysis of model prediction accuracy and computational efficiency. The results show that the present invention maintains high prediction accuracy while keeping the number of model parameters and computational complexity within a controllable range, demonstrating good engineering deployment applicability.
[0076] Furthermore, in practical engineering applications, the method of this invention can be deployed in the intelligent operation monitoring system of thermal power units. By predicting the main steam temperature in advance, it provides operators with auxiliary decision-making basis, which can be used to adjust the desuperheating water volume in advance, optimize fuel allocation strategies, and reduce equipment operation risks. The MSTLNet prediction framework of this invention is not only applicable to the prediction of main steam temperature, but can also be extended to the prediction of other key thermal parameters such as reheat steam temperature and boiler outlet flue gas temperature by adjusting the input variables and prediction targets without changing the overall structure.
[0077] In summary, this invention presents a main steam temperature prediction method and system based on a multi-scale time-frequency network. This overcomes the shortcomings of existing main steam temperature prediction methods under complex operating conditions of thermal power units, such as insufficient multi-scale time-series feature modeling capabilities, inadequate utilization of key frequency domain information, and poor adaptability to changes in operating modes. By constructing a multi-branch collaborative modeling prediction framework, it achieves effective decoupling and joint modeling of long-term trend features and short-term disturbance features in the main steam temperature time series. Furthermore, by introducing an adaptive frequency domain enhancement mechanism, it dynamically strengthens key frequency components that significantly influence the prediction results, suppressing the interference of noise and irrelevant frequency bands on model performance. Finally, through an operating mode clustering and matching structure, it explicitly models the main steam temperature evolution patterns under different load ranges and operating conditions, thereby improving the prediction accuracy and stability of the model under complex operating condition switching and non-stationary operating conditions.
[0078] Finally, it should be noted that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Anyone skilled in the art can readily implement the present invention according to the description and above. Any modifications, alterations, or equivalent variations made using the technical content disclosed above are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, or variations made to the above embodiments based on the essential technology of the present invention are still within the protection scope of the present invention.
Claims
1. A method for predicting main steam temperature based on a multi-scale time-frequency network, characterized in that, Includes the following steps: S1. Collect multivariate operating time series data from the distributed control system of thermal power units, and perform time alignment, outlier removal, missing data filling and normalization in sequence to obtain preprocessed multivariate time series data. S2. Perform trend-seasonal decomposition on the preprocessed multivariate time series data to obtain trend components and seasonal components. S3. Input the seasonal components into the pre-built and trained multi-scale collaborative convolutional module to extract dynamic perturbation features at different time scales and obtain multi-scale joint features. S4. Input the trend components into the pre-built and trained clustering matching module, and model and reconstruct the long-term trend pattern through the learnable cluster centers to obtain the trend prediction features. S5. Based on the pre-built and trained adaptive wavelet frequency domain enhancement module, perform adaptive wavelet decomposition and frequency domain filtering on the multi-scale joint features to obtain frequency domain enhanced features; S6. The trend prediction features and frequency domain enhancement features are fused together, and the temperature prediction results of the main steam are output through the linear prediction layer.
2. The main steam temperature prediction method based on multi-scale time-frequency networks according to claim 1, characterized in that, In S1, the multivariate operating time series data includes main steam temperature, unit load, fuel quantity, feedwater flow rate, and desuperheating water flow rate.
3. The main steam temperature prediction method based on multi-scale time-frequency networks according to claim 1, characterized in that, In S1, time alignment includes timestamp alignment and resampling of runtime data from different sources and with different sampling frequencies, unifying all variables to the same sampling period; outlier removal includes identifying and removing anomalies such as sensor disconnection, sudden spikes, and long-term drift in the runtime time series data of variables; missing data imputation includes imputing missing data using moving average interpolation, linear interpolation, or forward imputation methods; normalization is performed according to the following formula: ； in, These are the original observations. and These represent the mean and standard deviation of the corresponding variable in the training samples, respectively. This is the normalized value.
4. The main steam temperature prediction method based on multi-scale time-frequency networks according to claim 1, characterized in that, S2 specifically includes: S21. Input multivariate time series Construct a moving average operator to obtain an initial estimate of the trend component: ； in, The length of the sliding window. For time indexing, for Multivariate time series data at any given time; S22. Use copying or mirroring to complete the time series boundaries so that the moving average can be calculated at the beginning and end of the series, and calculate the trend component based on the completed series. S23. Subtract the trend component from the preprocessed multivariate time series data to obtain the seasonal component: ； S24. Ensure the following reconstruction consistency relationship is satisfied before and after decomposition: ； in, For preprocessed multivariate time series data, For seasonal portions, For trend components.
5. The main steam temperature prediction method based on multi-scale time-frequency networks according to claim 1, characterized in that, S3 specifically includes: S31. Parallel convolution operations are performed on the seasonal component time series using one-dimensional convolution kernels of different sizes: ； in, This is a one-dimensional convolution operation. The kernel size is the convolution kernel size. The output features of the convolution kernel of the corresponding size; S32. Concatenate the outputs of convolutions at each scale along the channel dimension to form multi-scale joint features: ； in, It is a multi-scale joint feature; S32. Map the multi-scale joint features into query vector, key vector, and value vector, and input them into a multi-head self-attention mechanism: ； in, These are the query vector, key vector, and value vector, respectively. For the attention calculation results, For matrix transpose, For vector dimensions, For activation functions; S32. Perform channel reconstruction and feature compression on the attention output to obtain the output features of the multi-scale co-convolution module.
6. The main steam temperature prediction method based on multi-scale time-frequency networks according to claim 1, characterized in that, S4 specifically includes: S41. Embed trend features according to the following formula: ； in, For multilayer perceptron computation, As a trend component, Embedding representation for trend features; S42, Initialization Learnable cluster centers It is used to depict trend patterns under different operating conditions; S43. Calculate the similarity using the following formula: ； in, For the first The first trend feature is embedded in the second... The similarity weight of each cluster center For the first Embedding of trend features, For the first Cluster centers, For the first Cluster centers; S44. Reconstruct trend features according to the following formula: ； in, For the first One trend prediction feature, For similarity weights, For the first Cluster centers.
7. The main steam temperature prediction method based on multi-scale time-frequency networks according to claim 1, characterized in that, S5 specifically includes: S51. Perform wavelet decomposition based on the second-generation wavelet lifting structure on the multi-scale joint features according to the following formula to obtain the low-frequency components and high-frequency components: ； in, For multi-scale joint features, Low-frequency components, These are high-frequency components; S52. Perform high-frequency energy calculation and normalization according to the following formula: ； in, The energy of the high-frequency components, The absolute value of the high-frequency component. Normalized energy This represents the maximum energy value. To prevent the minimum value where the denominator is zero; S53. Construct an adaptive frequency mask: ； in, For frequency mask, For fixed thresholds or learnable parameters; S54. Perform frequency domain feature reconstruction: ； in, This is an inverse wavelet transform operation. For element-wise multiplication, This is a frequency domain enhancement feature.
8. The main steam temperature prediction method based on multi-scale time-frequency networks according to claim 1, characterized in that, S5 specifically includes: S51. The trend prediction features and frequency domain enhancement features are fused using a feature concatenation method: ； in, The characteristics after fusion For trend prediction features, For frequency domain enhancement features; S52. Input the fused features into the linear prediction layer and output the main steam temperature prediction result. ； in, Based on the main steam temperature prediction results, These are the weight parameters for the linear prediction layer. The characteristics after fusion This is the bias parameter.
9. The main steam temperature prediction method based on multi-scale time-frequency networks according to claim 1, characterized in that, The multi-scale collaborative convolution module includes multi-scale convolution branches, cross-scale attention units, and feature reconstruction units; the clustering matching module includes trend feature embedding units, cluster center initialization units, similarity calculation units, and trend pattern reconstruction units. The adaptive wavelet frequency domain enhancement module includes a wavelet decomposition unit, a high-frequency energy calculation unit, a frequency mask construction unit, and a frequency domain reconstruction unit.
10. A main steam temperature prediction system based on a multi-scale time-frequency network, used to implement the main steam temperature prediction method based on a multi-scale time-frequency network as described in any one of claims 1-9, characterized in that, It includes a data preprocessing module, a time series decomposition module, a multi-scale co-convolution module, a clustering matching module, an adaptive wavelet frequency domain enhancement module, and a prediction output module, wherein: Data preprocessing module: used to collect multivariate operating time series data from the distributed control system of thermal power units, and sequentially perform time alignment, outlier removal, missing data completion and normalization to obtain preprocessed multivariate time series data; Time series decomposition module: used to perform trend and seasonal decomposition on preprocessed multivariate time series data to obtain trend components and seasonal components; Multi-scale co-convolution module: Used to input seasonal components into a pre-built and trained multi-scale co-convolution module to extract dynamic perturbation features at different time scales and obtain multi-scale joint features; Clustering matching module: This module takes the trend components as input to the pre-built and trained clustering matching module, and models and reconstructs the long-term trend pattern through learnable cluster centers to obtain trend prediction features. Adaptive wavelet frequency domain enhancement module: Based on the pre-built and trained adaptive wavelet frequency domain enhancement module, it performs adaptive wavelet decomposition and frequency domain filtering on multi-scale joint features to obtain frequency domain enhanced features; Prediction output module: used to fuse trend prediction features with frequency domain enhancement features, and output the temperature prediction result of the main steam through the linear prediction layer.