A method and system for predicting main steam temperature based on query-enhanced relationship large model

Abstract

The application discloses a kind of based on query enhancement relation big model's main steam temperature prediction method and system, belong to thermal power plant thermal control and intelligent prediction field.The method utilizes multi-scale feature refining module to construct multi-resolution feature pyramid by frequency domain cycle identification and hierarchical pooling, realize the noise suppression and feature enhancement of noisy industrial time series data;Phase perception double query modeling module adopts time query and relationship query double parallel structure, respectively extracts global cycle priori and variable dynamic physical coupling dependence, and realizes feature adaptive cooperation by phase guide gate fusion;Adaptive prediction module completes main steam temperature multi-step prediction and uncertainty quantification.Through the modeling paradigm of feature refining and relationship discovery decoupling, the problem that traditional method is sensitive to non-stationary disturbance under AGC flexible peak shaving scene, it is difficult to consider prediction accuracy and robustness, provides reliable technical support for safe and economic operation of coal-fired unit.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring and advanced control technology for thermal power generation processes, specifically involving a method and system for predicting main steam temperature based on a query-enhanced relational large model. Background Technology

[0002] Main steam temperature is a core control parameter that determines the safety boundary and thermal efficiency of thermal power generating units. In the context of high-proportion grid connection of new energy sources and deep, flexible peak shaving by the units, accurate and forward-looking prediction of main steam temperature is a crucial prerequisite for coordinated unit control, proactive temperature regulation, and risk warning. Deviations in main steam temperature control directly affect the safety and economy of unit operation: overheating accelerates the aging of superheaters and high-temperature components of the turbine, shortening equipment lifespan; underheating reduces the unit's cycle thermal efficiency and increases coal consumption for power generation. Therefore, achieving high-precision and robust prediction of main steam temperature is of critical significance for the safe and economical operation of thermal power units and the improvement of their deep peak-shaving capabilities.

[0003] Currently, methods for predicting main steam temperature are mainly divided into two categories: traditional mechanistic modeling and data-driven methods. Traditional mechanistic modeling methods are based on thermodynamics and fluid mechanics principles, describing the dynamic behavior of main steam temperature by establishing physical equations for the boiler superheating system. Typical methods include lumped parameter methods, distributed parameter methods, and simplified models based on heat balance. Although these methods possess clear physical interpretability and can reflect the basic laws governing changes in main steam temperature, they rely on precise thermodynamic boundary conditions, such as flue gas temperature, flue gas flow rate, and surface fouling coefficient. In real industrial environments, these boundary conditions are difficult to obtain accurately, which limits the accuracy of the model. The main steam temperature system has dynamic characteristics of large inertia, large hysteresis, strong nonlinearity, and strong coupling, making accurate mechanism description extremely difficult. Especially under the non-stationary operating conditions of high-frequency load changes in Automatic Generation Control (AGC), the prediction accuracy of traditional mechanism models will drop sharply, making it difficult to meet the high-precision prediction requirements of flexible peak-shaving modes. The parameter calibration of mechanism models usually relies on field tests, which is time-consuming and labor-intensive, and the model has limited generalization ability, making it difficult to adapt to the characteristic changes of different units or different operating cycles of the same unit.

[0004] In recent years, with the rapid development of industrial big data technology and artificial intelligence algorithms, data-driven methods have received widespread attention in the field of main steam temperature prediction. Existing data-driven methods mainly include recurrent neural networks (RNN), long short-term memory networks (LSTM), gated recurrent units (GRU), temporal convolutional networks (TCN), and various Transformer variants. These methods achieve future predictions of main steam temperature by learning the temporal dependencies in historical operating data. However, existing data-driven methods still have systemic defects in the main steam temperature prediction scenario: 1) Insufficient modeling of coupling relationships between variables, and the inherent limitations of mainstream multivariate temporal modeling strategies. Channel-independent strategies (such as DLinear and PatchTST) process each sensor channel independently, with no information interaction between channels during prediction. Although this strategy has a certain robustness to industrial noise, it completely ignores the physical coupling relationship between variables in the boiler-turbine system, losing the core correlation information that determines the change of main steam temperature; channel-mixed strategies (such as traditional Transformer variants) achieve fully connected interaction between all variables through attention mechanisms. However, industrial time-series data often contains a large number of variables that are unrelated to or weakly correlated with the prediction target. Fully connected interactions can lead to the model being affected by irrelevant variables and industrial noise, causing overfitting or oversmoothing problems. Although static graph / clustering strategies seek a balance between the two, their pre-set static relationship graphs are difficult to adapt to the dynamic changes in variable coupling relationships caused by operating condition switching, and cannot capture transient correlation changes during peak shaving. 2) Insufficient capture of multi-scale time-series characteristics. Thermal power operation data contains multiple time-scale patterns: load changes show daily / weekly periodic patterns, while fuel fluctuations, valve regulation, etc., show second-level transient responses. Most existing methods directly perform time-series dependency mining on noisy raw data, mixing feature purification and relationship discovery, and lack an effective suppression mechanism for high-frequency industrial noise in the raw signal. This makes the model susceptible to noise interference and generates false time-series correlations, making it difficult to simultaneously capture the strong daily / weekly periodic patterns and second-level transient fluctuations in thermal power data. Moreover, Transformer-type architectures often face deployment bottlenecks in industrial fields due to excessive secondary computation complexity and parameter count. 3) Insufficient adaptability to non-stationary industrial scenarios. Under flexible peak-shaving operation mode, thermal power units frequently experience extreme operating conditions such as load changes, start-ups and shutdowns, and coal quality fluctuations. The operating data exhibits strong non-stationarity and significant differences in distribution across operating modes. Most existing methods have failed to systematically address the inherent challenges of strong non-stationarity, high noise, and differences in distribution across operating modes in industrial data under AGC flexible peak-shaving. Under extreme operating conditions such as frequent load changes, start-ups and shutdowns, and coal quality fluctuations, the predictive performance will significantly degrade, making it difficult to meet the predictive requirements of thermal power units operating under all conditions.In summary, existing technologies still lack a novel modeling framework that can simultaneously achieve feature purification of noisy data, accurately capture multi-scale temporal dynamics, adaptively model the dynamic coupling relationship between variables, and overcome the challenges of non-stationary industrial scenarios.

[0005] In summary, existing main steam temperature prediction technologies have significant shortcomings in the following aspects: First, they lack an effective feature extraction mechanism for noisy industrial data, resulting in insufficient model robustness; second, they cannot accurately capture multi-scale temporal dynamic characteristics, making it difficult to simultaneously model long-term periodic patterns and short-term transient responses; third, they lack the ability to adaptively model the dynamic coupling relationships between variables, making them unable to adapt to abrupt changes in relationships under non-stationary operating conditions; fourth, the computational complexity of the models is too high, making it difficult to meet the requirements of real-time industrial deployment; and fifth, they lack uncertainty estimation capabilities, making it impossible to quantify the reliability of prediction results. Therefore, there is an urgent need for a novel main steam temperature prediction modeling framework that can simultaneously achieve feature extraction of noisy data, accurately capture multi-scale temporal dynamics, adaptively model the dynamic coupling relationships between variables, and overcome the challenges of non-stationary industrial scenarios, in order to support the safe and economical operation of thermal power units under deep peak shaving conditions. Summary of the Invention

[0006] To address the technical challenge of existing main steam temperature prediction methods in simultaneously achieving feature refinement, multi-scale time-series dynamic capture, and adaptive modeling of dynamic coupling relationships between variables in noisy industrial data, this invention aims to provide a main steam temperature prediction method and system based on Query-Enhanced Relation Network (QuERNet). This method constructs a deep network architecture including a multi-scale feature refinement module and a dual-query modeling mechanism: First, high-frequency transient noise is filtered out through multi-scale decomposition to generate refined features; then, time query (TQ) and relation query (RQ) are used to extract pure periodic priors and decoupled physical dependencies based on the refined features, respectively; finally, a phase-guided mechanism adaptively fuses cross-domain features and aggregates and outputs the mean and variance predictions of the main steam temperature from top to bottom. This invention effectively improves the prediction accuracy and model robustness under complex and variable operating conditions by explicitly separating the feature refinement and relation discovery stages.

[0007] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, the present invention provides a method for predicting the main steam temperature based on a query-enhanced relational large model, comprising the following steps: Step 1: Obtain multivariate time series data during the operation of the coal-fired power unit, preprocess the data, and obtain preprocessed multivariate time series window data; Step 2: Based on the preprocessed multivariate time-series window data, perform frequency domain period recognition and multi-resolution pooling to obtain multi-scale time-series enhancement features; Step 3: Based on the multi-scale temporal enhancement features obtained in Step 2, execute the time query branch and the relation query branch in parallel, extract the periodic prior features and the dynamic coupling features between variables respectively, and fuse the two through the phase-guided gating fusion mechanism to obtain the feature representation of fused multi-scale spatiotemporal semantics. Step 4: Based on the feature representation of fused multi-scale spatiotemporal semantics obtained in Step 3, decode to obtain the multi-step prediction result of the main steam temperature and the corresponding prediction uncertainty quantification value.

[0008] Preferably, in step 1, the preprocessing includes: missing value imputation, outlier detection and correction, normalization, and constructing the preprocessed multivariate time-series window data using a sliding window.

[0009] Preferably, the multivariate time series data includes at least three of the following: main steam temperature, fuel flow rate, desuperheater valve opening, feedwater flow rate, superheater damper opening, burnout air valve position, and unit load.

[0010] In step 2, the specific processing steps of the multi-scale feature refinement module include: By aggregating the spectral energies of all sensor channels using Fast Fourier Transform, the dominant periodic components can be identified. Based on the identified period length, one-dimensional average pooling is performed on the original sequence to obtain pooled sequence representations with different time resolutions; Pooling sequence representations of different scales are concatenated along the time dimension to construct a multi-scale comprehensive feature map; By using pointwise linear transformation and activation functions, the multi-scale feature map is projected onto a high-dimensional feature space, and the multi-scale temporal enhancement feature is output.

[0011] In step 3, the specific processing procedure for the time query branch includes: For each scale of temporal enhancement features, identify time anchor points within the period and generate node embeddings containing absolute position and periodic phase information; Using phase as a discrete index, query vectors are extracted from a globally shared learnable dictionary. Query-guided temporal attention aggregation is then performed on each variable channel to obtain time-query-enhanced temporal features.

[0012] Step 3, the specific processing steps for relation query branching and feature fusion include: Generate time-series aggregated feature embeddings for each variable, calculate the original correlation between variables, construct positive correlation collaboration graphs and negative correlation competition graphs respectively, and obtain variable relationship features for relationship query enhancement through graph propagation; Dynamic gating vectors are generated based on phase-aware anchor embedding. Adaptive convex combination fusion is performed on the temporal features enhanced by time query and the variable relationship features enhanced by relationship query to obtain single-scale interactive enhancement features. A top-down multi-scale aggregation strategy is adopted to upsample coarse-scale features and iteratively fuse them with fine-scale features to output the feature representation of the fused multi-scale spatiotemporal semantics.

[0013] Preferably, when constructing the positive correlation collaboration graph and the negative correlation competition graph, the top K variables in correlation ranking are extracted to form the positive correlation collaboration set, and the bottom K variables in correlation ranking are extracted to form the negative correlation competition set. Softmax normalization is then applied to the edge weights of both sets, where K is a preset sparse connection hyperparameter. By extracting the top K variables in correlation ranking to form the positive correlation collaboration set and the bottom K variables in correlation ranking to form the negative correlation competition set, the graph structure is made sparse, effectively reducing computational complexity and avoiding information redundancy and overfitting risks caused by fully connected graphs. This solves the problem of excessive computational complexity in existing technology models, which makes it difficult to meet the requirements of real-time industrial deployment. Applying Softmax normalization to the edge weights of both sets completely preserves the complete gradient information of the collaboration and competition relationships between variables, enabling the model to dynamically learn the strength changes of the coupling relationships between variables during end-to-end training. Compared to static graph strategies, phase gating can accurately capture transient coupling changes caused by rapid switching of operating conditions.

[0014] In step 4, the multi-step prediction results of the main steam temperature and the corresponding quantified value of the prediction uncertainty are obtained. The specific process includes: A time series prediction network is shared across all variable channels. The time series prediction network consists of one-dimensional convolutional layers and multilayer perceptrons. The historical temporal dimension of the features is aligned to the target prediction duration through a one-dimensional convolutional layer, and the features are decoded by a multi-layer perceptron to generate a multivariate prediction sequence. By using a separate linear head, the predicted mean and variance sequences of the main steam temperature target channel are extracted from the multivariate prediction sequence. Efficient feature decoding is achieved by sharing a time-series prediction network (one-dimensional convolutional layer + multilayer perceptron) among the variable channels, reducing the number of parameters, mitigating overfitting risk, and solving the problem of excessive model computational complexity. The one-dimensional convolutional layer aligns the historical time-series dimension to the target prediction duration, enabling flexible adjustment of the prediction horizon. Simultaneously extracting the predicted mean and variance through the separate linear head provides not only point predictions of the main steam temperature but also additional calibration of the predicted variance, offering highly valuable confidence intervals and operational safety boundary assessment references for power plant operators and intelligent control systems.

[0015] End-to-end optimization training is performed using a composite loss function, which includes: the mean square error term between the predicted mean and the true value of the main steam temperature, the negative log-likelihood term of the Gaussian distribution, the L1 norm penalty term of the normalized edge weights of the positive and negative correlation graphs, and the constraint term of minimizing the cosine similarity between different phase vectors in the global dictionary of the time query.

[0016] Secondly, the present invention provides a system for implementing the aforementioned main steam temperature prediction method based on a query-enhanced relational large model, comprising: The data acquisition module is used to acquire multivariate time series data during the operation of coal-fired power units; The multi-scale feature refinement module is used to refine the preprocessed multivariate time-series window data at multiple scales and output the refined feature representation. The phase-aware dual-query modeling module is used to receive the refined feature representation, perform phase-aware dual-query modeling, and output a feature representation that integrates multi-scale spatiotemporal semantics. An adaptive prediction module is used to receive the feature representation fused with multi-scale spatiotemporal semantics, and output the multi-step prediction result of the main steam temperature and the quantified value of the prediction uncertainty.

[0017] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the main steam temperature prediction method based on the query-enhanced relational large model.

[0018] Fourthly, the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the main steam temperature prediction method based on a query-enhanced relational large model.

[0019] Compared with the prior art, the present invention has the following beneficial effects: This invention provides a main steam temperature prediction method based on Query Enhancement Relationship Network (QuERNet). By constructing a multi-scale feature refinement module through "frequency domain periodicity identification and multi-resolution pooling," it achieves noise suppression and feature enhancement of noisy industrial data. This fundamentally isolates transient spikes, step changes, and frequent start-stop interference during power plant load shifting, solving the problems of insufficient effective feature purification mechanisms and inadequate model robustness in existing technologies. Through parallel execution of time and relation queries, the time query branch extracts globally noise-free periodic priors from a discrete dictionary, while the relation query branch decouples the dynamic physical coupling dependencies between variables, achieving accurate decoupling and mapping of dual complex dynamics. This solves the problem of existing technologies being unable to simultaneously capture multi-scale temporal dynamics and adaptively model dynamic coupling relationships between variables. Through a "phase-guided gating fusion mechanism," it adaptively fuses cross-domain features, avoiding the over-smoothing problem caused by channel mixing strategies. Simultaneously, phase gating accurately captures transient coupling changes caused by rapid switching of operating conditions. Finally, it outputs the predicted value of the main steam temperature. This method explicitly separates the feature refinement and relationship discovery stages, effectively improving the prediction accuracy and model robustness under complex and variable operating conditions, and providing reliable technical support for the safe and economical operation of thermal power units under the background of deep peak shaving.

[0020] Furthermore, by aggregating the dominant periodic components of multi-channel spectral energy using Fast Fourier Transform (FFT), the system can adaptively capture strong daily / weekly periodic patterns in coal-fired power unit operation data without requiring manual setting of the decomposition scale. One-dimensional average pooling is performed based on the identified period length to extract sequence trends within a non-overlapping window, effectively filtering out industrial noise such as sensor spikes and high-frequency disturbances while preserving the underlying temporal semantics. Pooled sequences at different scales are concatenated and projected into a high-dimensional space via linear transformation and GeLU activation, constructing a multi-resolution feature pyramid. This feature refinement mechanism ensures that the subsequent spatiotemporal graph neural network architecture is built upon highly reliable and pure semantic features, significantly improving the model's predictive stability and data quality tolerance, and solving the robustness problem caused by directly modeling on raw, noisy data in traditional methods.

[0021] Furthermore, by generating node embeddings containing absolute position and periodic phase information, the model can perceive periodic positional information in time-series data. The time query branch extracts query vectors from a globally shared learnable dictionary using phase as a discrete index, achieving pure noise-independent periodic prior extraction and effectively avoiding the oversmoothing problem caused by channel mixing strategies. The relationship query branch, by constructing positive correlation collaboration graphs and negative correlation competition graphs, completely and losslessly preserves the collaborative and competitive characteristics between industrial variables caused by physical and thermodynamic coupling, and achieves adaptive modeling of dynamic coupling relationships between variables through decoupling graph propagation. Phase-guided gating fusion generates dynamic gating vectors based on anchor point embedding, achieving adaptive convex combination fusion of two types of heterogeneous features, and can accurately capture transient coupling changes caused by rapid switching of operating conditions through phase gating. The top-down multi-scale aggregation strategy preserves the original temporal resolution and outputs rich feature representations that fuse multi-scale spatiotemporal semantics. This claim achieves accurate decoupling and mapping of dual complex dynamics, solving the technical problem that existing technologies cannot simultaneously capture multi-scale temporal dynamics and adaptively model dynamic coupling relationships between variables.

[0022] The system provided by this invention, through modular design, breaks through the inherent limitations of traditional methods that directly model on raw, high-noise data. By employing a modeling paradigm of explicit decoupling between feature refinement and relation discovery, it ensures that the spatiotemporal graph neural network architecture is built upon highly reliable and pure semantic features. Through a dual-channel mechanism of time query and relation query, it achieves precise decoupling and mapping of dual complex dynamics. Furthermore, by outputting a refined average prediction value of the main steam temperature from the prediction module, it additionally provides calibration prediction variance, achieving comprehensive safety assurance that balances precise control with uncertainty quantification. This system can be directly deployed in the control system or intelligent monitoring platform of coal-fired power units, providing a complete hardware and software solution for the safe and economical operation of thermal power units under deep peak shaving conditions. Attached Figure Description

[0023] Figure 1 This is a flowchart of a main steam temperature prediction method based on Query Enhanced Relationship Network (QuERNet) as described in this invention.

[0024] Figure 2 This is a flowchart illustrating the workflow of the multi-scale feature refinement module in this invention.

[0025] Figure 3 This is a flowchart of the workflow of the dual-query modeling and interaction module in this invention.

[0026] Figure 4 This is a schematic diagram of the physical system and prediction motivation for predicting the main steam temperature of a thermal power plant in an embodiment of the present invention.

[0027] Figure 5This is a schematic diagram of the overall architecture of the Query Enhancement Relationship Network (QuERNet) described in this invention; Figure 6 The chart shows a comparison of the prediction performance of normalized MSP based on different models. (a) is the prediction trajectory of the normalized MSP (prediction duration H=96), and (b) is the absolute prediction error. Curve showing the change over time. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0029] Query-Enhanced Relation Network (QuERNet): This invention proposes a deep network architecture whose core idea is to enhance feature representation through a "query" mechanism. It includes two parallel branches: time query and relation query, which are used to extract time-series periodic priors and physical coupling dependencies between variables, respectively, hence the name "Query-Enhanced Relation Network".

[0030] Multi-scale feature refinement refers to the process of performing multi-resolution transformation on the original noisy time-series data. Through frequency domain periodicity identification and hierarchical pooling operations, a multi-resolution feature pyramid is constructed to achieve noise suppression and enhancement of key time-series features.

[0031] Time Query (TQ): A mechanism that retrieves query vectors from a globally shared learnable dictionary using the periodic phase as a discrete index. This branch processes each sensor channel independently, extracting noise-independent purified periodic prior features through query-guided temporal attention aggregation.

[0032] Relation Query (RQ): A mechanism for modeling dynamic physical coupling dependencies among multiple variables. This branch calculates the original correlation between variables, constructs a positive correlation collaboration graph and a negative correlation competition graph, updates the query vector through a decoupling graph propagation network, and obtains a feature representation of the variable relationships.

[0033] Phase-aware refers to the model's ability to identify and utilize periodic phase information in time-series data. In this invention, the time anchor embedding includes absolute position and periodic phase information, and a dynamic gating vector is generated accordingly to guide the adaptive fusion of time query features and relational query features.

[0034] Phase-guided Gated Fusion: A feature fusion strategy. Based on phase-aware anchor embedding, it generates dynamic gated vectors and performs adaptive convex combination fusion on time-based query enhancement features and relation-based query enhancement features, achieving dynamic synergy between the two under different operating conditions.

[0035] Positive Correlation Graph & Negative Competition Graph: These are graph structures used in this invention to decouple and model the relationships between variables. Based on the correlation ranking between variables: the top K variables are extracted to form a positive correlation correlation graph, representing the physical coupling relationship of cooperative changes between variables; the bottom K variables are extracted to form a negative correlation competition graph, representing the physical coupling relationship of competitive suppression between variables. Both types of graphs are subjected to Softmax normalization to fully preserve the gradient information of cooperative and competitive relationships between variables.

[0036] Multi-scale aggregation refers to the iterative fusion of interactive enhancement features at different scales using a top-down strategy. Starting from the coarsest scale, coarse-scale features are upsampled and then concatenated with fine-scale features to ultimately output a feature representation that retains the original temporal resolution and fuses multi-scale spatiotemporal semantics.

[0037] The Adaptive Prediction Module is a time-series prediction network composed of one-dimensional convolutional layers and a multi-layer perceptron (MLP). This module shares parameters across all variable channels, aligns historical time-series features to the target prediction duration, decodes and generates a multivariate prediction sequence, and outputs the predicted mean and variance of the main steam temperature target channel.

[0038] The present invention will now be described in further detail with reference to the accompanying drawings: Example 1 This embodiment takes a supercritical 600MW coal-fired unit as an example, collects its operating data under the AGC flexible peak shaving mode, and describes in detail the main steam temperature prediction method based on the query-enhanced relational model (QuERNet) of the present invention.

[0039] (a) Application Scenarios like Figure 4 As shown, under modern AGC-R operation mode, the daily load of thermal power generating units changes extremely frequently, leading to high-frequency transient disturbances and non-stationary fluctuations in key thermodynamic parameters (such as main steam temperature). Main steam temperature is a core control parameter determining the safety boundary and thermal efficiency of thermal power generating units, and its accurate forward prediction is a crucial prerequisite for coordinated unit control, proactive temperature regulation, and risk warning. This invention addresses this industrial scenario by proposing a prediction method that can simultaneously achieve noisy data feature purification, accurately capture multi-scale temporal dynamics, adaptively model the dynamic coupling relationship between variables, and overcome the challenges of non-stationary industrial scenarios.

[0040] like Figure 1 As shown, the overall calculation process of this invention includes the following steps: Step 1: Obtain multivariate time series data during the operation of the coal-fired power unit, preprocess the data, and obtain preprocessed multivariate time series window data; Step 2: Based on the preprocessed multivariate time-series window data, perform frequency domain period recognition and multi-resolution pooling to obtain multi-scale time-series enhancement features; Step 3: Based on the multi-scale temporal enhancement features obtained in Step 2, execute the time query branch and the relation query branch in parallel, extract the periodic prior features and the dynamic coupling features between variables respectively, and fuse the two through the phase-guided gating fusion mechanism to obtain the feature representation of fused multi-scale spatiotemporal semantics. Step 4: Based on the feature representation of fused multi-scale spatiotemporal semantics obtained in Step 3, decode to obtain the multi-step prediction result of the main steam temperature and the corresponding prediction uncertainty quantification value.

[0041] Step 1: Data Acquisition and Preprocessing This embodiment continuously collects 90 days of operating data from the unit's distributed control system (DCS), with a sampling interval of 1 minute. Nine variables strongly correlated with the main steam temperature are selected: main steam temperature (target variable), fuel flow rate, desuperheater valve opening, feedwater flow rate, superheater damper opening, burnout air valve position, unit load, flue gas oxygen content, and economizer outlet water temperature.

[0042] Data preprocessing is performed as follows: Step 11: Outlier handling: For each sensor channel, outliers are identified using the 3σ principle, the mean and standard deviation are calculated, and values ​​outside the range of [μ-3σ, μ+3σ] are replaced by linear interpolation.

[0043] Step 12: Normalization: Use the Z-Score normalization method to make the mean of each variable channel 0 and the standard deviation 1;

[0044] in μ x σ is the mean of this channel. x The standard deviation is denoted as .

[0045] Step 13: Sequence window partitioning: Set the historical window length L=96 (corresponding to 96 minutes) and the prediction duration H=12 (corresponding to 12 minutes); use the sliding window method to generate samples with a sliding step size of 5 minutes, and obtain approximately 25,000 samples in total, which are divided into training set, validation set and test set in a ratio of 7:2:1.

[0046] Step 2: Based on the preprocessed multivariate time-series window data, perform frequency domain period recognition and multi-resolution pooling to obtain multi-scale time-series enhancement features.

[0047] like Figure 2 As shown in the workflow diagram, this step aims to filter out high-frequency noise and extract reliable temporal semantics. The specific processing flow is as follows: (1) Frequency Domain Period Identification: Collect multivariate time series data with a historical time window length of L, and first identify the dominant periodic component inherent in the operational data. Perform Fast Fourier Transform (FFT) on the time series data of each sensor channel to calculate the amplitude spectrum. Assuming the sequence length of a certain channel is L=96, pad with zeros to 256 points before performing FFT:

[0048] in, Indicates the first d The amplitude spectrum function of each sensor channel represents different frequencies. f The corresponding amplitude intensity; d This refers to the index of the sensor channel (variable), with a value range. d =1,2,…,D, where D is the total number of variables; f Frequency, measured in Hz, represents how quickly a signal changes. F represents the Fourier transform operator, used to convert a time-domain signal to the frequency domain. x 1:L,d Indicates the first d The time series data of each channel within the time window 1 to L. L This represents the length of the history window, which is the number of time steps in the input sequence (in this embodiment, L=96, corresponding to 96 minutes). This indicates that the amplitude spectrum is a real vector of length L (the Fourier transform result is symmetric, so only the first half is valid).

[0049] Aggregating spectral energy across all channels to capture system-level periodic modes:

[0050] in, The average amplitude spectrum across all channels represents the system-level spectral energy distribution. D This represents the total number of sensor channels (or variables in a multivariate time series). The entire formula's function is to include all... Amplitude spectrum of each channel Summate the results and then divide by the total number. This allows for the calculation of the average amplitude spectrum across all channels. This helps to capture cyclical patterns at the system level of the entire thermal power generating unit, rather than being limited to local patterns from a single sensor.

[0051] Select the one with the most significant energy K After ignoring the DC component (f=0), select the three frequency components with the largest amplitude. , , Calculate the corresponding period length. :

[0052] in, : indicates the identified first The length of each period determines the window size for average pooling operations in the subsequent multi-scale feature refinement module. : Indicates the length of the historical time window, i.e., the total number of time steps of the multivariate time series input into the model. In the embodiment, The typical value is 96 (corresponding to 96 minutes). : Represents the dominant frequency component identified through Fast Fourier Transform (FFT). In the formula... This means that the system selected the top 3 frequencies with the most significant amplitude spectrum energy. This is the ceiling notation in mathematics. Its purpose is to ensure that the calculated period length is accurate. The integer value is used to perform pooling operations on non-overlapping windows. : Represents the index of the frequency component. This algorithm selects different... The values ​​are used to construct a multi-resolution feature pyramid, thereby simultaneously capturing temporal dynamic characteristics at different scales. The core significance of the formula lies in transforming the abstract frequency features identified in the frequency domain into specific, discrete "time period lengths" in the time domain, thus providing a data-driven physical basis for subsequent denoising and feature extraction in the model.

[0053] (2) Hierarchical average pooling: For each identified period One-dimensional average pooling is applied to the non-overlapping window of the original sequence to generate scaled features, thereby suppressing high-frequency noise and extracting underlying trends.

[0054] in, : Represents the new eigenvalue after pooling. Specifically refers to the... Each sensor channel (variable) is used in a window with a size of [missing information]. After pooling, the first element in the newly obtained sequence The value of each time step. : Indicates the pooling window size (i.e., period length). This is the number of periods calculated in the previous step (Fast Fourier Transform, FFT). The length of each dominant cycle. (The formula begins with...) Used to calculate the average value of the data within a window. : Represents the index of the variable (sensor channel). For example, in thermal power generation data, This likely represents the main steam temperature. This could represent fuel flow, etc. : Represents the time index of the new sequence after pooling (i.e., which pooling window). If the original sequence length is... Then the length of the new sequence is usually . : Represents the time step index of the original time series. It is the traversal variable for the summation symbol. : Represents a specific observation in the original time series. That is, the _th __ The variable in the original first... Data at each time step. and These are the lower and upper bounds of the summation symbol, precisely pinpointing the first... The pooling window extracts the "start position" and "end position" from the original time series. Since the stride is equal to the window size... This ensures that there is no overlap between the windows.

[0055] when p k At 24, the 96-bit sequence is compressed into a 4-bit pooled sequence, with each pooling value representing the trend mean over 24 minutes. This effectively suppresses high-frequency sensor noise and extracts underlying trend information. (3) Multi-scale feature splicing: Pooling sequences of different scales are spliced ​​along the time dimension to construct a multi-scale comprehensive feature map; (4) Feature embedding and enhancement: By performing pointwise linear mapping and applying the Gaussian error linear unit (GeLU) activation function, the multi-scale map is projected onto a high-dimensional feature space, outputting a multi-resolution feature pyramid:

[0056] in, : Represents a multi-scale composite feature map (the concatenated features). It is composed of different scales (i.e., different pooling windows) from step 22. The pooled sequences generated are spliced ​​together along the time dimension to form the basic data. and : These represent the learnable weight matrix and bias vector, respectively. They form the core of the point-wise linear transformation, used to project features into a high-dimensional space. : Represents the activation function of the Gaussian Error Linear Unit. An activation function that combines nonlinear mapping and regularization properties. : Represents the final output multi-scale temporal enhancement features (multi-resolution feature pyramid). It is the pure semantic input that this module ultimately passes to the downstream "dual query modeling module". : Indicates the output feature The high-dimensional real tensor space to which it belongs, in This represents the total time step of the pooling after concatenation. For example, if the pooling sequence lengths for the three cycles are 4, 2, and 1 respectively, then the concatenated time step... . : Indicates the total number of variables (number of sensor channels). For example, if there are a total of 9 variables such as main steam temperature and fuel flow rate, then... . : Represents the enhanced high-dimensional hidden feature dimension. For example, expanding the original scalar time-series numerical value to A dimensional vector representation.

[0057] In this embodiment, the concatenation shape of the pooled sequence is (number of variables D, total pooling length ∑kL / p) k )=(9,4+2+1)=(9,7), after linear transformation and GeLU activation, the output multi-scale temporal enhancement feature X is (9,7,64), which serves as the basic input for subsequent modules.

[0058] Step 3: Based on the multi-scale temporal enhancement features obtained in Step 2, execute the time query branch and the relation query branch in parallel, extract the periodic prior features and the dynamic coupling features between variables respectively, and fuse the two through the phase-guided gating fusion mechanism to obtain the feature representation of fused multi-scale spatiotemporal semantics.

[0059] like Figure 5 As shown in the overall architecture diagram, this module uses a dual-parallel structure of Time Queries (TQ) and Relational Queries (RQ) to extract purified global periodic priors and dynamic physical coupling dependencies between variables, respectively. Figure 3As shown, this module includes TQ branches and RQ branches, and achieves feature fusion through a cross-domain interaction mechanism.

[0060] (1) Embedding of time anchor points and phase sensing nodes For each scale of temporal enhancement features, identify time anchors within the period, and generate a node embedding for each anchor that includes absolute position and periodic phase information: (2) Construct the time query TQ branch Using phase as a discrete index, purified noise-independent query vectors are extracted from a globally shared learnable dictionary. Query-guided temporal attention aggregation is performed independently on each variable channel to obtain a TQ-enhanced temporal feature representation. Time anchors are set at regular intervals, and phase-aware node embeddings containing absolute time and relative phase are calculated. This phase is used as a discrete index from a globally shared learnable dictionary. Retrieving noiseless time query vectors :

[0061] in, : indicates scale Next, the The time query vector corresponding to each time anchor point represents the "ideal" characteristic representation of that specific time point within the global cyclical pattern. : Indicates scale A globally shared learnable dictionary. It is essentially a parameter matrix (lookup table) that stores features of all possible phases at that scale. : indicates the first Anchor points in the cycle The discrete phase index within the cycle. It determines the current position within the cycle (e.g., which hour of the day). This is the symbol for matrix slicing or indexing. Its meaning is to extract data from a dictionary matrix... Extract the first The entire row vector corresponding to the row.

[0062] Calculate the scaling dot product attention score between the global query vector and the local refined features. :

[0063] in, : Represents the attention score or weight. It specifically refers to the weights used in processing the first... When the number of sensor channels (variables) is 1, the 1st A global query of the i-th time anchor point within the local neighborhood... The degree of attention paid to the features of each time step. : Represents the noiseless temporal query vector retrieved from the global dictionary. It acts as Query(Q) in the attention mechanism here. : Indicates multi-scale Next, the The variables at local time step The refined feature vector. It serves as both the Key (K) in the attention mechanism and the Value (V) for subsequent aggregation. : Represents the transpose of a matrix or vector, used to calculate the dot product between the query vector and the feature vector to measure their similarity. : Represents feature dimension The square root of is used as a scaling factor. : indicates surrounding the first A local temporal neighborhood is defined by a time anchor point. : The standard Softmax normalization function.

[0064] The temporal context of each variable is independently aggregated to form a time representation. :

[0065] in, This refers to the enhanced temporal representation of time queries (TQ). Specifically, it refers to multi-scale temporal feature representation. Next, the The variable (sensor channel) in the first... The final feature vector after purification by the attention mechanism at each time anchor point. : indicates the first The local time neighborhood where each time anchor point is located All time steps within Perform a summation operation. : This represents the attention score calculated by the formula in the previous step. It is a scalar weight that sums to 1, representing the attention score within the local neighborhood. The importance of each time step to the current anchor point. : indicates scale Next, the The variables at local time step The feature vector. In standard attention mechanisms, it acts as a weighted summation of the Value (V) vector (which can also be represented as...). (in the form of).

[0066] (3) Constructing Relational Query (RQ) Branches For each variable, a time-series aggregated feature embedding is generated, the original correlation between variables is calculated, and positive correlation synergy graphs and negative correlation competition graphs are constructed respectively. RQ-enhanced variable relationship feature representations are obtained through decoupled graph propagation. (4) Based on phase-aware anchor embedding, dynamic gated vectors are generated. Through phase-guided gated fusion mechanism, adaptive convex combination fusion is performed on TQ enhanced temporal features and RQ enhanced relational features to obtain single-scale interactive enhanced features; like Figure 5 As shown, the original temporal correlation scores between physical variables are calculated, and the structures of the positive collaboration graph and the negative competition graph are extracted. Edge weights are normalized by independently applying Softmax.

[0067] in, : Indicates multi-scale Below, variables With variables The final normalized edge weights in the positively correlated collaboration graph. This range is... between. : indicates scale Below, variables With variables The original temporal relevance score between them. This is typically obtained by extracting the temporal embeddings of the variables and calculating the inner product or similarity. : Represents a variable The positively correlated local neighborhood. It is determined by the original correlation score. Sort in descending order and extract the top-ranked items. The set of the variables with the highest scores (i.e., top-) (Node set). Traditional graph networks typically use ReLU to truncate (reset to zero) negative correlations. However, in thermal power plant control, competing relationships (such as the negative physical feedback of a decrease in main steam temperature due to an increase in desuperheating water flow) are crucial for transient regulation. This invention fully preserves the bipolar gradient by exponentially taking the original positive and negative Logit values ​​separately, thus avoiding the loss of coupling information between key control variables.

[0068] (5) Multiscale polymerization A top-down multi-scale aggregation strategy is adopted, starting from the coarsest scale. After upsampling the coarsest scale features, they are iteratively fused with the fine-scale features, and finally output a feature representation that retains the original temporal resolution and fuses multi-scale spatiotemporal semantics. Relational query vectors are propagated through a decoupled graph to extract spatial representations. Subsequently, anchor points were used for embedding. Generate phase-guided dynamic gating vector :

[0069] in, : Indicates multi-scale Below is the first A dynamically gated vector generated at each time anchor point. It is a vector with a range of... The weight vector between them is used to control the proportion of subsequent cross-domain feature fusion. : indicates the Sigmoid activation function. : indicates the first The phase-aware node embedding for each time anchor point contains the absolute time position of that anchor point in the sequence and its relative phase information within the current period. and : These represent the learnable parameter matrix and bias vector used in generating the gating, respectively. They are optimized through end-to-end training of the model on a large amount of historical running data.

[0070] like Figure 3 and Figure 5 As shown, the fusion features are obtained by executing the cross-domain interaction mechanism. :

[0071] in, : Indicates multi-scale Next, the The final single-scale interactive enhancement feature obtained from each time anchor point is a comprehensive feature representation that integrates temporal regularity and spatial physical coupling. : Indicates time-enhanced time-series features for time queries. It represents pure periodic prior patterns (such as daily load trends) extracted from the global dictionary. : Indicates the variable relational feature of relational queries. It represents the real-time, dynamic coupling dependency between physical variables (such as the transient cooperation and competition between desuperheating water flow rate and main steam temperature). : indicates element-wise multiplication. : indicates and Complementary weights ensure that the fusion of the two features is a smooth process.

[0072] Step 4: Based on the feature representation of fused multi-scale spatiotemporal semantics obtained in Step 3, decode to obtain the multi-step prediction result of the main steam temperature and the corresponding prediction uncertainty quantification value.

[0073] (1) Shared prediction network configuration like Figure 1 End steps and Figure 5As shown in the aggregation module, starting from the coarsest scale, a top-down strategy is used to iteratively upsample and fuse interactive enhancement features from each scale: Assuming there are K scales in total, starting from the coarsest scale k=K, the features are fused layer by layer upwards to the finest scale k=1. The fusion formula is as follows:

[0074] in : indicates scale At a finer scale, the fused comprehensive feature tensor is a final feature representation that incorporates cross-scale spatiotemporal semantics. : Indicates scale The single-scale interactive enhancement features are obtained by fusing time-based and relation-based queries at this scale. : Indicates the value derived from the scale Upsampling is performed on the fused features at a coarser scale. The purpose is to align the low temporal resolution at the coarser scale to the high resolution at the finer scale through interpolation. : indicates a concatenation operation, which means concatenating two feature vectors end to end along the channel dimension.

[0075] After top-down iterative fusion, the final feature representation H is obtained, which retains the original temporal resolution and fuses multi-scale spatiotemporal semantics. agg ∈R L×C×Dmodel (In this embodiment, L=96, C=8, D) model =128) (2) Timing alignment and decoding Input the fused features into the adaptive prediction network The network consists of one-dimensional convolutional layers and multi-layer perceptrons: through the prediction network composed of one-dimensional convolutions and multi-layer perceptrons, predictions for the future horizon are generated. :

[0076] in, This represents the final multivariate prediction sequence. It is a sequence that includes all future target time steps (prediction duration). The tensor matrix of predicted values ​​for all sensor channels within the range. : Represents an adaptive temporal prediction network. In this patent architecture, it is specifically composed of a one-dimensional convolutional layer and a multilayer perceptron (MLP). : Represents the final feature representation output after top-down multi-scale aggregation.

[0077] Training the above model (1) Training environment configuration

[0078] (2) Hyperparameter settings

[0079] (3) Composite loss function The overall training objective consists of prediction accuracy loss, uncertainty calibration loss, graph sparsity loss, and query diversity loss, which are jointly optimized using a composite loss function:

[0080] in : Represents the total composite loss. It is the final objective function for model backpropagation and parameter updates. : Represents the mean squared error loss. Used to measure the predicted value. Compared with the actual observed value The deviation between them ensures the accuracy of the prediction. : Represents graph sparsity loss. By applying regularization constraints to the graph structure in the relation query branches, the model is forced to retain only the most critical physical coupling edges, eliminating redundant noise and preventing overfitting. : Indicates query diversity loss. Used in constraint-time query dictionaries. The vectors in the dictionary ensure that different phase indices can learn distinct and recognizable temporal patterns, thus avoiding dictionary collapse. : Represents the loss weight hyperparameter. Used to balance the contribution ratio of various losses in the overall optimization process.

[0081] In this embodiment, the loss weights are set to: λ1=0.05, λ2=0.01.

[0082] (4) Optimization strategy Mixed Precision Training (AMP): Enable torch.cuda.amp. The noise in industrial data naturally provides a certain degree of numerical stability. FP16 calculations under AMP have almost no impact on MSE, but can save 30%-40% of GPU memory and speed up the process.

[0083] Validation and early stopping: Monitor the MSE of prediction step 96 (H=96) separately. Due to the cumulative effect of error, if the MSE_{H=96} increases for 10 consecutive epochs on the validation set instead of decreasing, training should be stopped immediately.

[0084] Step 25: Prediction Results and Performance Evaluation See Figure 6In terms of predicting trajectories using the Normalized SSP (H=96), different models showed varying performance. The QuERNet (Ours) model (blue solid line) was able to track the changing trend of the true value (black solid line) relatively well, while the PatchTST (orange dashed line) and Transformer (red dotted line) models exhibited lag and oversmoothing in certain periods, especially near the AGC transient region, where the deviation between the predicted and true values ​​was quite significant.

[0085] Regarding the absolute prediction error curves, the QuERNet model exhibits a relatively small overall absolute prediction error, with errors within most time steps falling below the tight error boundary (<0.022), indicating high prediction accuracy. In contrast, the PatchTST and Transformer models show relatively large absolute prediction errors, exceeding the baseline error peak (>0.04) at certain time points, suggesting significant uncertainty in the prediction process and lower reliability of the prediction results.

[0086] Overall, the QuERNet model demonstrates superior performance in predicting normalized MSP values, more accurately tracking changes in the true value with smaller prediction errors, showing a clear advantage over PatchTST and Transformer models. This is attributed to its well-designed adaptive prediction module, optimized training process, and joint optimization of the composite loss function, among other factors.

[0087] (1) Comparative experiment Data preparation: The USC-1000 dataset was used, which contains 28 key thermal parameters and has a sampling resolution of 1 second to ensure that the data has sufficient sample size and diversity to evaluate the performance of different models. The model of this invention was constructed and optimized in accordance with the above method.

[0088] Model selection and configuration: Select multiple comparison models, including LSTM, TCN, Informer, Autoformer, PatchTST, DLinear, CrossGNN, and the QuERNet model of this invention.

[0089] Model Training: Each model is trained using the training set. During training, the same Adam optimizer and ReduceLROnPlateau learning rate decay strategy are used to ensure consistency in training conditions. Each model undergoes multiple rounds of iterative training on the training set. After each round of training, the model is evaluated using the validation set, and the training progress is monitored based on the evaluation results (loss function values ​​on the validation set).

[0090] Performance Evaluation: Evaluate the performance of each trained model on the test set. Calculate the mean squared error (MSE) and mean absolute error (MAE) for different prediction lengths (H = 24, H = 48, H = 96).

[0091] Record the MSE and MA values ​​of each model at different prediction lengths and calculate the average value (Avg.) to comprehensively evaluate the performance of the model at different prediction durations, as shown in Table 1.

[0092] Table 1: Performance comparison on the USC-1000 dataset

[0093] As shown in Table 1, with the prediction duration H increasing from 24 to 96, the performance of traditional time series models such as LSTM and TCN degrades most drastically (MSE growth rate exceeding 50%), demonstrating their inadequacy in handling ultra-long sequence dependencies and error accumulation. While Transformer variants (Informer and Autoformer) possess the potential for long sequence modeling, their sparse attention or autocorrelation mechanisms still capture spurious time dependencies in high-noise industrial environments, causing the MSE to rise to 0.495 and 0.456 at H=96, respectively. In contrast, the QuERNet of this invention achieves an MSE of only 0.381 at H=96, a reduction of approximately 8.9% compared to the optimal baseline PatchTST.

[0094] (2) Ablation test Data preparation: Power plant datasets (USC-1000, ZN-1000, FX-DCS) were used. Data preprocessing was performed, including normalization, and the datasets were divided into training, validation, and test sets, with the division ratios similar to those in the comparative experiments.

[0095] Variant construction: w / o MS (removing the multi-scale module): The FFT periodicity identification and hierarchical pooling structures are removed. The original time-series data is directly fed into the subsequent dual-query module after linear embedding. This variant is used to verify the necessity of multi-scale refining for noise suppression.

[0096] w / o PTQ (Removing Phase-Aware Temporal Query): Replaces the temporal query branch with a standard multi-head self-attention mechanism. Instead of retrieving phase priors from a discrete dictionary, the temporal anchors directly apply attention weights to locally refined features. This variant is used to evaluate the contribution of globally purified cyclic priors to prediction robustness.

[0097] w / o DRQ (Decoupling Query): This variant cancels the decoupling construction of the positive / negative correlation graph. After calculating the correlation between variables, a ReLU activation function is applied to truncate all negative values, retaining only positively correlated edges for graph propagation. This variant is used to examine the crucial role of preserving competing negative gradient relationships in thermal process modeling.

[0098] Training and Evaluation: Each variant model and the full QuERNet model were trained independently under the same hardware and software environment and hyperparameter configuration. To ensure statistical significance, each experiment was run three times, and the average MSE and MAE were used as the final reported values. The evaluation horizon was uniformly taken as the average of H={24, 48, 96}.

[0099] Model Training: The full model and its variants are trained using the training set. A similar training strategy to the comparative experiments is employed, including the same optimizer, initial learning rate, and number of training epochs. During training, the validation set is used to monitor and adjust the model.

[0100] Performance Evaluation: The full model and each variant model were evaluated on the test set, and the MSE and MAE values ​​of each model were calculated on the three datasets. By comparing the performance metrics of the full model and the variant models, the impact of components such as multi-scale modules, temporal queries, and relational queries on model performance was analyzed, and the effectiveness of their combined effect was verified. The results are shown in Table 2.

[0101] Table 2: Ablation Study Results of Power Plant Datasets (Average) )

[0102] As shown in Table 2, the MSE and MAE of the method of this invention are lower than those of the variant model with each component removed on the three datasets. This verifies the effectiveness of the combined effect of the multi-scale module, time query, and relational query components. Their cooperation enables the algorithm to better handle time series prediction tasks and achieves good performance on the power plant dataset.

[0103] Example 2 This embodiment is basically the same as Embodiment 1, except that: (1) Different data sources: This embodiment uses the operating data of a subcritical 300MW coal-fired unit, with a sampling interval of 30 seconds, a historical window length of L=120 (corresponding to 60 minutes), and a prediction duration of H=20 (corresponding to 10 minutes).

[0104] (2) Different variable selection: Seven variables were selected: main steam temperature, fuel flow rate, desuperheater valve opening, feedwater flow rate, unit load, flue gas oxygen content, and superheater damper opening.

[0105] (3) Different period identification parameters: Select the two frequency components with the strongest energy, and the corresponding period lengths p∈{60,120}.

[0106] (4) Sparse connection parameter K=2.

[0107] Experimental results show that this embodiment also achieved a significantly better prediction effect than the baseline model, verifying the generalization ability of the method of the present invention under different units and different parameter settings.

[0108] Example 3 This embodiment is basically the same as Embodiment 1, except that the value of the balance coefficient in the composite loss function is different: λ1=0.05, λ2=0.005, λ3=0.03 Experimental results show that this set of coefficients can also achieve good prediction performance, verifying the robustness of the method of the present invention to hyperparameter settings.

[0109] Example 4 This embodiment provides a main steam temperature prediction system for coal-fired power units that implements the above method, including: Data acquisition module: Used to collect multivariate time-series data from the DCS system in real time, and perform normalization processing and sequence window division.

[0110] Multi-scale feature refinement module: used to perform FFT periodic identification, hierarchical average pooling and feature embedding on preprocessed data, and output multi-scale temporal enhanced features.

[0111] Phase-aware query enhancement module: Includes time query branch and relation query branch, used to extract periodic prior features and dynamic coupling features between variables in parallel, and fuse the two through phase-guided gating fusion mechanism to obtain feature representation of fused multi-scale spatiotemporal semantics.

[0112] Adaptive prediction module: used to decode the feature representation of the fused multi-scale spatiotemporal semantics, and generate multi-step prediction results of the main steam temperature and the corresponding prediction uncertainty quantification value.

[0113] Output module: Used to output prediction results, 95% confidence intervals, and visualization.

[0114] The system can be deployed on the generator set's DCS historical station or an independent intelligent prediction server, with a single inference time of approximately 15ms (GPU) / 45ms (CPU), meeting the real-time requirements of industry.

[0115] The above content is only for illustrating the technical concept of the present invention and should not be construed as limiting the scope of protection of the present invention. Any modifications made to the technical solution based on the technical concept proposed in this invention shall fall within the scope of protection of the claims of this invention.

Claims

1. A method for predicting the main steam temperature based on a query-enhanced relational large model, characterized in that, Includes the following steps: Step 1: Obtain multivariate time series data during the operation of the coal-fired power unit, preprocess the data, and obtain preprocessed multivariate time series window data; Step 2: Based on the preprocessed multivariate time-series window data, perform frequency domain period recognition and multi-resolution pooling to obtain multi-scale time-series enhancement features; Step 3: Based on the multi-scale temporal enhancement features obtained in Step 2, execute the time query branch and the relation query branch in parallel, extract the periodic prior features and the dynamic coupling features between variables respectively, and fuse the two through the phase-guided gating fusion mechanism to obtain the feature representation of fused multi-scale spatiotemporal semantics. Step 4: Based on the feature representation of fused multi-scale spatiotemporal semantics obtained in Step 3, decode to obtain the multi-step prediction result of the main steam temperature and the corresponding prediction uncertainty quantification value.

2. The method for predicting the main steam temperature based on a query-enhanced relational large model according to claim 1, characterized in that, In step 2, the specific processing steps of the multi-scale feature refinement module include: By aggregating the spectral energies of all sensor channels using Fast Fourier Transform, the dominant periodic components can be identified. Based on the identified period length, one-dimensional average pooling is performed on the original sequence to obtain pooled sequence representations with different time resolutions; Pooling sequence representations of different scales are concatenated along the time dimension to construct a multi-scale comprehensive feature map; By using pointwise linear transformation and activation functions, the multi-scale feature map is projected onto a high-dimensional feature space, and the multi-scale temporal enhancement feature is output.

3. The method for predicting the main steam temperature based on a query-enhanced relational large model according to claim 1, characterized in that, In step 3, the specific processing procedure for the time query branch includes: For each scale of temporal enhancement features, identify time anchor points within the period and generate node embeddings containing absolute position and periodic phase information; Using phase as a discrete index, query vectors are extracted from a globally shared learnable dictionary. Query-guided temporal attention aggregation is then performed on each variable channel to obtain time-query-enhanced temporal features.

4. The method for predicting the main steam temperature based on a query-enhanced relational large model according to claim 1, characterized in that, Step 3, the specific processing steps for relation query branching and feature fusion include: Generate time-series aggregated feature embeddings for each variable, calculate the original correlation between variables, construct positive correlation collaboration graphs and negative correlation competition graphs respectively, and obtain variable relationship features for relationship query enhancement through graph propagation; Dynamic gating vectors are generated based on phase-aware anchor embedding. Adaptive convex combination fusion is performed on the temporal features enhanced by time query and the variable relationship features enhanced by relationship query to obtain single-scale interactive enhancement features. A top-down multi-scale aggregation strategy is adopted to upsample coarse-scale features and iteratively fuse them with fine-scale features to output the feature representation of the fused multi-scale spatiotemporal semantics.

5. The method for predicting the main steam temperature based on a query-enhanced relational large model according to claim 4, characterized in that, When constructing the positive correlation collaboration graph and the negative correlation competition graph, the top K variables in the correlation ranking are extracted to form the positive correlation collaboration set, and the bottom K variables in the correlation ranking are extracted to form the negative correlation competition set. Softmax normalization is performed on the edge weights of the two sets respectively, where K is a preset sparse connection hyperparameter.

6. The method for predicting the main steam temperature based on a query-enhanced relational large model according to claim 1, characterized in that, In step 4, the multi-step prediction results of the main steam temperature and the corresponding quantified value of the prediction uncertainty are obtained. The specific process includes: A time series prediction network is shared across all variable channels. The time series prediction network consists of one-dimensional convolutional layers and multilayer perceptrons. The historical temporal dimension of the features is aligned to the target prediction duration through a one-dimensional convolutional layer, and the features are decoded by a multi-layer perceptron to generate a multivariate prediction sequence. By separating the linear head, the predicted mean and predicted variance sequences of the main steam temperature target channel are extracted from the multivariate prediction sequence.

7. The method for predicting the main steam temperature based on a query-enhanced relational large model according to claim 1, characterized in that, End-to-end optimization training is performed using a composite loss function, which includes: the mean square error term between the predicted mean and the true value of the main steam temperature, the negative log-likelihood term of the Gaussian distribution, the L1 norm penalty term of the normalized edge weights of the positive and negative correlation graphs, and the constraint term of minimizing the cosine similarity between different phase vectors in the global dictionary of the time query.

8. A system for implementing the main steam temperature prediction method based on a query-enhanced relational large model as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire multivariate time series data during the operation of coal-fired power units and perform preprocessing to obtain preprocessed multivariate time series window data. The multi-scale feature refinement module is used to refine the preprocessed multivariate time-series window data at multiple scales and output the refined feature representation. The phase-aware dual-query modeling module is used to receive the refined feature representation, perform phase-aware dual-query modeling, and output a feature representation that integrates multi-scale spatiotemporal semantics. An adaptive prediction module is used to receive the feature representation fused with multi-scale spatiotemporal semantics, and output the multi-step prediction result of the main steam temperature and the quantified value of the prediction uncertainty.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the main steam temperature prediction method based on the query-enhanced relational large model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the main steam temperature prediction method based on the query-enhanced relational large model as described in any one of claims 1 to 7.

Images