Main steam temperature prediction method and system based on hybrid model

Abstract

The invention discloses a main steam temperature prediction method and system based on a hybrid model. The method comprises the following steps: S1, acquiring a historical data set of a time sequence and carrying out data preprocessing; s2, establishing a main steam temperature prediction model: S21, constructing a time sequence prediction sub-model based on a linear relation; optimizing model parameters to establish an optimal time sequence prediction sub-model by taking a composite function fusing a basic precision item and an operation risk penalty item as a target function of a global optimization algorithm; s22, outputting a first main steam temperature prediction sequence and a residual sequence through the optimal time sequence prediction sub-model; s23, training and establishing a nonlinear prediction sub-model, and outputting a residual prediction sequence; s24, superposing the first main steam temperature prediction sequence and the residual prediction sequence; and S3, acquiring a real-time value, inputting the real-time value into the main steam temperature prediction model, and outputting a main steam temperature prediction value. According to the technical scheme, the main steam temperature prediction accuracy, stability and safety can be improved.

Description

Technical Field

[0001] This invention relates to the field of thermal power generation technology, and in particular to a method and system for predicting main steam temperature based on a hybrid model. Background Technology

[0002] Main steam temperature (MST) is a core control parameter in the operation of thermal power generating units. Its stability and accuracy directly affect the unit's thermal efficiency, operational safety, and the service life of key components (such as superheaters and turbines). Under the current "dual carbon" target, thermal power units, especially large ultra-supercritical units, are required to undertake more frequent and in-depth peak-shaving tasks, leading to drastic changes in unit operating conditions. The main steam temperature system itself is a complex controlled object with large inertia, large delay, nonlinearity, and strong coupling of multiple variables, making accurate prediction of its operation under varying conditions a highly challenging technical problem.

[0003] Traditional prediction methods are mainly divided into three categories: (1) Mechanism modeling method: Based on physical laws such as thermodynamics, fluid mechanics and heat transfer, a complex mathematical model is established. The advantage of this method is that it has strong physical interpretation. However, in practical applications, due to the extreme complexity of the combustion process inside the boiler, the time-varying nature of the model parameters and the difficulty in accurately obtaining the boundary conditions, the accuracy of the model is difficult to meet the actual production needs, and the amount of calculation is huge, making it difficult to apply online. (2) Traditional statistical methods: such as the autoregressive moving average model (ARMA) or its extended form, the autoregressive integral moving average model (ARIMA). This type of model is good at capturing linear trends and periodic patterns in time series data. However, when the main steam temperature changes drastically with load, fuel composition fluctuates, or water spray desuperheating, it will exhibit strong nonlinear and non-stationary mutation characteristics. A single linear model cannot effectively fit these complex dynamics. (3) Single machine learning method: such as backpropagation (BP) neural network, support vector machine (SVM), etc. These methods have strong nonlinear mapping capabilities, but they are prone to getting trapped in local optima and have limited ability to capture the inherent temporal dependencies in time series data. They are also prone to overfitting and have insufficient model generalization ability and robustness. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method and system for predicting main steam temperature based on a hybrid model.

[0005] The technical solution adopted by this invention to solve its technical problem is: constructing a main steam temperature prediction method based on a hybrid model, the method comprising: S1. Obtain historical time series datasets and perform data preprocessing to form a model training set; the historical datasets include historical main steam temperature datasets and historical datasets of multiple feature variables related to the main steam temperature; S2. Based on the model training set, train and establish a main steam temperature prediction model, including: S21. Construct a time series prediction sub-model based on linear relationships; use the model training set as the input of the time series prediction sub-model, and use the composite function that integrates the basic accuracy term and the running risk penalty term as the objective function of the global optimization algorithm. Use the global optimization algorithm to determine the optimal model parameters of the time series prediction sub-model, and establish the optimal time series prediction sub-model based on the optimal model parameters. S22. Based on the model training set, prediction is performed using the optimal time-series prediction sub-model to output the first main steam temperature prediction sequence and the residual sequence; S23. Using the residual sequence and the corresponding operating condition characteristic variable data as input, train and establish a nonlinear prediction sub-model, and output the residual prediction sequence; S24. The first main steam temperature prediction sequence is superimposed with the residual prediction sequence as the output value of the main steam temperature prediction model; S3. Obtain the real-time main steam temperature value and the real-time values ​​of the multiple feature variables related to the main steam temperature, input them into the main steam temperature prediction model, and output the main steam temperature prediction value.

[0006] Furthermore, the multiple characteristic variables related to the main steam temperature include: unit load command, total fuel quantity, total air volume, feedwater flow rate, main steam flow rate, flow rate and valve position of each stage of water spray desuperheater and / or steam temperature and pressure at the inlet and outlet of each stage of superheater.

[0007] Further, in step S21, the operational risk penalty term includes a prediction stability penalty term, a dynamic process penalty term, and an operational safety penalty term, and the objective function is: Fitness = w1 F_accuracy+w2 F_stability+w3 F_transient+w4 F_safety Wherein, F_accuracy is the basic accuracy term, F_stability is the prediction stability penalty term, F_transient is the dynamic process penalty term, F_safety is the operational safety penalty term, and w1, w2, w3, and w4 are the weight coefficients of each sub-item.

[0008] Further, in step S21, the time series prediction sub-model based on linear relationship is an ARIMA model, and step S21 includes: Construct the ARIMA model, wherein the model parameters p, d, and q of the ARIMA model are undetermined parameters to be optimized; The model training set is used as the input of the ARIMA model, and the composite function of the fusion basic accuracy term and the running risk penalty term is used as the objective function of the global optimization algorithm. The global optimization algorithm is used to find the optimization parameters of the ARIMA model to form the optimal model parameters. The optimal time series prediction sub-model is established based on the optimal model parameters.

[0009] Further, in step S21, the global optimization algorithm is a quantum particle swarm optimization algorithm, and the step of determining the optimal model parameters of the time series prediction sub-model using the global optimization algorithm includes: Initialize the particle swarm of the quantum particle swarm optimization algorithm, wherein the position of each particle represents a set of parameter values ​​to be optimized in the ARIMA model; The training set of the model is predicted using an ARIMA model built based on each particle, and the fitness value of each particle is calculated according to the objective function. Based on the fitness value, the particle swarm is iteratively updated using the quantum particle swarm optimization algorithm; the globally optimal particle position obtained after iterative optimization is used as the optimal model parameter.

[0010] Further, in step S23, the nonlinear prediction sub-model includes a gating unit, and step S23 includes: The residual sequence and the corresponding operating condition characteristic variable data are input into the nonlinear prediction sub-model; The gating unit receives the operating condition characteristic variable data and outputs a gating signal of the same dimension as the residual sequence; The residual sequence is multiplied element-wise with the gated signal to generate an enhanced residual sequence; The enhanced residual sequence is trained based on the operating condition characteristic variable data, and the residual prediction sequence is output.

[0011] Further, in step S1, the data preprocessing step includes: non-linearly aligning the historical dataset of each feature variable related to the main steam temperature with the historical dataset of the main steam temperature, calculating the correlation with the main steam temperature, and removing feature variables with low correlation to generate a core feature variable set.

[0012] Furthermore, the nonlinear alignment step includes: The optimal nonlinear alignment path between the historical dataset of each feature variable related to the main steam temperature and the historical dataset of the main steam temperature is calculated using a dynamic time warping algorithm. For each feature variable related to the main steam temperature, the time offset at each time point is calculated according to the corresponding optimal nonlinear alignment path, and the corresponding time delay distribution is generated. The expected value of the time delay distribution is calculated as the average time delay value of the feature variable. The feature variable is then shifted and aligned according to the average time delay value.

[0013] Furthermore, in step S1, the data preprocessing step further includes: obtaining a subset of data under sparse samples and / or key operating conditions in the core feature variable set, and using a time series generative adversarial network to learn the subset of data, generating corresponding synthetic sequence samples, and expanding them into the model training set.

[0014] The present invention also constructs a main steam temperature prediction system based on a hybrid model, the system comprising: The data processing module is used to acquire historical time series datasets and perform data preprocessing to form a model training set; the historical datasets include historical main steam temperature datasets and historical datasets of multiple feature variables related to the main steam temperature; The model building module, used to train and build a main steam temperature prediction model based on the model training set, includes: The first sub-model building module is used to construct a time-series prediction sub-model based on linear relationships. The model training set is used as input to the time-series prediction sub-model, and a composite function fusing the basic accuracy term and the operational risk penalty term is used as the objective function of the global optimization algorithm. The global optimization algorithm is used to determine the optimal model parameters of the time-series prediction sub-model, and an optimal time-series prediction sub-model is established based on these optimal model parameters. Based on the model training set, prediction is performed using the optimal time-series prediction sub-model, outputting a first main steam temperature prediction sequence and a residual sequence. The second sub-model building module is used to take the residual sequence and the corresponding working condition characteristic variable data as input, train and build a nonlinear prediction sub-model, and output the residual prediction sequence. The fusion module is used to superimpose the first main steam temperature prediction sequence with the residual prediction sequence as the output value of the main steam temperature prediction model. The main steam temperature prediction module is used to obtain the real-time main steam temperature value and the real-time values ​​of the multiple feature variables related to the main steam temperature, input them into the main steam temperature prediction model, and output the main steam temperature prediction value.

[0015] Implementing this invention offers the following advantages: By using multi-source heterogeneous datasets as training sets, the model can uncover causal relationships and coupling effects between variables, thereby improving prediction accuracy. By using a composite function that integrates the basic accuracy term and the operational risk penalty term as the objective function of the global optimization algorithm, the model parameters of the time-series prediction sub-model are optimized, avoiding local optima and further improving prediction accuracy while ensuring safety and stability. The nonlinear prediction sub-model trained using dual-path inputs of residuals and operating conditions exhibits excellent adaptability to varying operating conditions, improving prediction accuracy and robustness under drastic changes. By fusing the results of the linear time-series prediction sub-model with those of the nonlinear prediction sub-model, collaborative modeling of linear trends and nonlinear abrupt changes is achieved, taking into account both the regularity and dynamic response of main steam temperature changes.

[0016] Furthermore, the optimal nonlinear alignment path between each feature variable related to the main steam temperature and the main steam temperature is calculated using a dynamic time warping algorithm, and its average time delay distribution is calculated for alignment. Then, the correlation between each feature variable and the main steam temperature is calculated for feature selection. This quantifies the nonlinear time-varying delay relationship between the feature variables and the main steam temperature, solves the information distortion problem caused by traditional fixed time delay or linear correlation analysis, provides high-quality input with clear physical meaning for the prediction model, and thus improves the prediction accuracy of the prediction model. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of a main steam temperature prediction method based on a hybrid model in one embodiment of the present invention. Detailed Implementation

[0018] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the invention are now described in detail with reference to the accompanying drawings. In the following description, specific details such as particular structures and techniques are set forth for illustrative purposes and not for limitation, so as to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will understand that the invention can also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known devices, circuits, and methods are omitted so as not to obscure the description of the invention with unnecessary detail.

[0019] like Figure 1 As shown, Figure 1 This is a flowchart of a main steam temperature prediction method based on a hybrid model in one embodiment of the present invention, which includes the following steps: S1. Obtain historical time series datasets and perform data preprocessing to form a model training set; the historical datasets include historical main steam temperature datasets and historical datasets of multiple feature variables related to main steam temperature; In this step, based on mechanistic analysis, multiple relevant characteristic variables affecting the main steam temperature are selected. Time-series data of the main steam temperature and related characteristic variables can be obtained from the power plant's distributed control system (DCS), manufacturing execution system (MES), monitoring equipment, etc., forming a historical dataset. In some optional embodiments, the multiple characteristic variables related to the main steam temperature may include: unit load command, total fuel quantity, total air volume, feedwater flow rate, main steam flow rate, flow rate and valve position of each stage of water spray desuperheater and / or steam temperature and pressure at the inlet and outlet of each stage of superheater. It should be understood that the amount of data collected can be controlled according to a set duration, such as no less than one year, or according to a set threshold for the number of sampling points, and the sampling interval can be set as needed, such as every 1 second. The collected historical dataset is preprocessed, including data cleaning, data augmentation, data time-series alignment, data normalization, etc., to ensure the validity and temporal consistency of the data. The preprocessed data forms the model training data for training the model. It should be understood that the main steam temperature data and characteristic variable data in the model training set are still time-series values. The factors affecting the main steam temperature are complex. By obtaining multiple feature variables that affect the main steam temperature as part of the model training set, the model can be used to explore the causal relationships and coupling effects in multi-source heterogeneous data, thereby improving the prediction accuracy.

[0020] S2. Train and establish a main steam temperature prediction model based on the model training set, including: S21. Construct a time series prediction sub-model based on linear relationships; use the model training set as the input of the time series prediction sub-model, and use the composite function that integrates the basic accuracy term and the running risk penalty term as the objective function of the global optimization algorithm. Use the global optimization algorithm to determine the optimal model parameters of the time series prediction sub-model, and establish the optimal time series prediction sub-model based on the optimal model parameters. In this step, the time-series prediction sub-model based on linear relationships can transform the non-stationary main steam temperature sequence into a stationary sequence, capturing the linear stationary component in the main steam temperature time series to complete the linear prediction of the main steam temperature. The performance of the time-series prediction sub-model is highly dependent on the selection of its model parameters. In this embodiment, a global optimization algorithm is used to optimize the time-series prediction sub-model to determine its model parameters.

[0021] Specifically, the global optimization algorithm aims to find the globally optimal solution of the objective function within the entire feasible domain. In this embodiment, the model parameters of the time-series prediction sub-model are used as the feasible domain for optimization, the model training set is used as the input for prediction, and a composite function integrating the basic accuracy term and the operational risk penalty term is set as the objective function of the global optimization algorithm. Iterative optimization is performed with the goal of minimizing the objective function value to find the optimal model parameters, thereby establishing the optimal time-series prediction sub-model based on these optimal model parameters. The basic accuracy term in the objective function measures the overall prediction accuracy of the time-series prediction sub-model; for example, a smaller error indicates higher accuracy. The operational risk penalty term consists of multiple operational penalty terms related to the main steam temperature. By comprehensively considering the composite function of prediction accuracy and operational risk penalty term as the objective function for model parameter optimization, not only can local optima be effectively avoided, but the safety and stability requirements of power plant operation are also incorporated, avoiding prediction behaviors that may lead to equipment damage or system oscillations. This makes the obtained optimal time-series prediction sub-model closer to the actual engineering situation, capable of coping with uncertain interference, and the prediction results more reliable and trustworthy.

[0022] S22. Based on the model training set, prediction is performed through the optimal time series prediction sub-model, and the first main steam temperature prediction sequence and residual sequence are output. In this step, the model training set is used as input, and the optimal time series prediction sub-model established in step S21 is used for prediction to obtain the linear first main steam temperature prediction sequence and the corresponding residual sequence. The residual is the difference between the actual observed value and the model prediction value. The residual sequence E contains all nonlinear and non-stationary dynamic information in the main steam temperature.

[0023] S23. Using the residual sequence and the corresponding operating condition characteristic variable data as input, train and establish a nonlinear prediction sub-model, and output the residual prediction sequence; In this step, the nonlinear prediction sub-model can be a model with nonlinear fitting capabilities, such as a BP neural network, a Long Short-Term Memory (LSTM) network, or a lightweight Transformer model. The operating condition characteristic variable data consists of the values ​​of key operating condition variables related to nonlinear dynamics selected from multiple characteristic variables related to main steam temperature in step S1, such as unit load commands, the rate of change of total fuel quantity, the rate of change of main steam flow, and the total flow rate of water spray desuperheating, forming an operating condition context vector that represents the current operating state of the system. The residual sequence obtained through the optimal time-series prediction sub-model is used as the historical dynamic input, and simultaneously, the operating condition characteristic variable data is used as the auxiliary input for the system operating state to train and establish the nonlinear prediction sub-model, outputting a residual prediction sequence. This allows the nonlinear prediction sub-model to learn the nonlinear compensation patterns corresponding to different operating conditions, improving the prediction accuracy and robustness of the model under drastic operating conditions such as deep peak shaving.

[0024] S24. The first main steam temperature prediction sequence is superimposed with the residual prediction sequence as the output value of the main steam temperature prediction model; In this step, the first main steam temperature prediction sequence output by the optimal time-series prediction model is superimposed with the residual prediction sequence output by the nonlinear prediction sub-model, and this superposition is used as the output value of the entire main steam temperature prediction model, i.e., the main steam temperature prediction value. The superposition method can be linear summation. For example, if the first main steam temperature prediction value is Y_ARIMA(t) and the residual prediction value is E_BP(t), then the main steam temperature prediction value is: Y_final(t) = Y_ARIMA(t) + E_BP(t). Through the fusion of the optimal time-series prediction sub-model and the nonlinear prediction sub-model based on the linear relationship, the collaborative modeling of linear trends and nonlinear abrupt changes is achieved, taking into account both the regularity and dynamic response of main steam temperature changes.

[0025] Understandably, after training and building the model, model validation and evaluation can be carried out. The predictive performance of the model can be comprehensively evaluated using indicators such as mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). Special attention should be paid to the prediction error of the model under variable load conditions and whether it can effectively avoid dangerous situations such as overheating / underheating.

[0026] S3. Obtain the real-time main steam temperature value and the real-time values ​​of multiple feature variables related to the main steam temperature, input them into the main steam temperature prediction model, and output the predicted main steam temperature value.

[0027] In this step, the real-time main steam temperature value and the real-time values ​​of multiple characteristic variables related to the main steam temperature are input. Using the main steam temperature prediction model established in step S2, the predicted main steam temperature value can be output. Understandably, a human-machine interactive monitoring interface can be established, allowing the system to acquire real-time data, predict using the model, and monitor the process.

[0028] The technical solution implemented in this embodiment uses multi-source heterogeneous data related to operating conditions as the model training set to construct a main steam temperature prediction model that combines a global optimization algorithm, a time-series prediction sub-model, and a nonlinear prediction sub-model. This model can adapt to complex scenarios with varying operating conditions, improve the accuracy of main steam temperature prediction, and significantly enhance the safety and stability of the model.

[0029] In some optional embodiments, in step S1, the data preprocessing step includes: nonlinearly aligning the historical dataset of each feature variable related to the main steam temperature with the historical dataset of the main steam temperature, calculating the correlation with the main steam temperature, and removing feature variables with low correlation to generate a core feature variable set. Specifically, there is a time-varying delay between the feature variables and the main steam temperature. For each historical sequence of a feature variable and the main steam temperature sequence, the nonlinear time alignment relationship is calculated and aligned using time-series alignment algorithms such as dynamic time warping or dynamic window sliding analysis based on mutual information. Subsequently, based on the aligned feature variable sequence, the correlation between each feature variable and the main steam temperature is calculated to quantify the nonlinear correlation between variables. The correlation can be calculated using mutual information (MI) or analyzed using random forest or XGBoost. A threshold is set, and feature variables with correlation below the threshold are removed, while feature variables with high correlation are retained as the core feature variable set. It should be understood that the selected core feature variable set serves as the input for subsequent data processing steps and / or model training.

[0030] By employing non-linear alignment and relevance-based feature selection, non-linear alignment preserves dynamic characteristics and improves data quality compared to linear translation. Alignment-based feature selection enhances association accuracy by removing low-relevance features and focusing on core features, allowing for model training with less and more refined data, reducing noise interference, and further improving model performance.

[0031] In some alternative embodiments, the nonlinear alignment step in the previous embodiment includes: The optimal nonlinear alignment path between the historical dataset of each feature variable related to main steam temperature and the historical dataset of main steam temperature is calculated using a dynamic time warping algorithm. For each feature variable related to the main steam temperature, the time offset at each time point is calculated according to the corresponding optimal nonlinear alignment path, and the corresponding time delay distribution is generated. The expected value of the time delay distribution is calculated as the average time delay value of the feature variable. The feature variable is then shifted and aligned according to the average time delay value.

[0032] Specifically, a Dynamic Time Warping (DTW) algorithm is used for nonlinear alignment. For each feature variable's historical dataset, i.e., its time series, DTW finds a warped path that minimizes the cumulative distance between the feature variable's time series and the main steam temperature time series—the optimal nonlinear alignment path. Subsequently, for each feature variable, based on its corresponding optimal nonlinear alignment path, the time offset of the feature variable's time series relative to the main steam temperature time series at each time point is statistically analyzed, forming a corresponding time delay distribution. The expected value of this distribution is then calculated as the average dynamic time delay value for that feature variable. Finally, based on the average dynamic time delay value, the time series of each feature variable is shifted to achieve dynamic causal time delay alignment. This embodiment, by using DTW to find the optimal nonlinear alignment path and calculating the average dynamic time delay quantization offset of each feature variable to achieve shift alignment, can characterize the dynamic change characteristics of time delay under varying operating conditions for each feature variable, conforming to industrial realities and providing high-quality input data for the model, fundamentally improving the model's prediction accuracy.

[0033] Understandably, in the data preprocessing step, before performing non-linear alignment, data cleaning can be performed first, such as removing outliers and filling in missing values, and then non-linear alignment can be performed based on the cleaned data.

[0034] In some optional embodiments, the data preprocessing step S1 further includes: obtaining a subset of data under sparse samples and / or key operating conditions in the core feature variable set, and using a time-series generative adversarial network (TimeGAN) to learn the data subset, generating corresponding synthetic sequence samples, and expanding it to the model training set. Specifically, a set of core feature variables after nonlinear alignment is obtained, and data under sparse samples or key operating conditions are formed into a data subset, wherein key operating conditions are those affecting the fluctuation of main steam temperature, such as deep peak shaving, rapid load increases and decreases, etc. Subsequently, a time-series generative adversarial network (TimeGAN) is used to learn the above data subset, synthesizing a large number of data samples with similar dynamic characteristics, and expanding it to the model training set. The technical solution of this embodiment effectively solves the model training problem under key operating conditions such as deep peak shaving and unit start-up and shutdown where historical data samples are insufficient, and improves the generalization performance and high availability of the model under various operating conditions. The main steam temperature prediction model constructed by the expanded model training set can effectively capture the main steam temperature fluctuation segment and improve the prediction accuracy of main steam temperature.

[0035] In some optional embodiments, in step S21, the operational risk penalty term includes a prediction stability penalty term, a dynamic process penalty term, and an operational safety penalty term, and the objective function is: Fitness=w1 F_accuracy+w2 F_stability+w3 F_transient+w4 F_safety Wherein, F_accuracy is the basic accuracy term, F_stability is the prediction stability penalty term, F_transient is the dynamic process penalty term, F_safety is the operational safety penalty term, and w1, w2, w3, and w4 are the weight coefficients of each sub-term.

[0036] Specifically, the basic accuracy term F_accuracy can be calculated as the root mean square error (RMSE) or mean absolute error (MAE) between the observed main steam temperature and the model's predicted main steam temperature, to measure the overall prediction accuracy of the model. The prediction stability penalty term F_stability is used to penalize high-frequency oscillations in the prediction error sequence, preventing over-adjustment of the control system. It can be measured by calculating the variance of the first difference of the prediction error sequence, such as F_stability = Var(diff(Y_true - Y_pred)), where Y_true is the observed main steam temperature and Y_pred is the model's predicted main steam temperature. The dynamic process penalty term F_transient is used to assign higher weight to the prediction error during dynamic processes where the main steam temperature changes rapidly (i.e., the difference in main steam temperature within a set time exceeds a set threshold), ensuring the model's response accuracy during varying operating conditions. The safety penalty term F_safety introduces a large step penalty function. When the predicted value exceeds the preset safe operating range (such as being higher than the over-temperature protection threshold or lower than the under-temperature limit), this term takes a maximum value, thus vetoing any parameter combination that may lead to safety risks during the optimization process. w1, w2, w3, and w4 are the weight coefficients of each sub-item, which can be preset according to the unit characteristics and safety procedures.

[0037] By setting a composite function that integrates the above-mentioned multiple risk perception terms as the objective function of the global optimization algorithm, the model parameters of the time series prediction sub-model are optimized. This is deeply integrated with the complex operating conditions of main steam temperature and the safety requirements of thermal power plants, thereby establishing a safer and more reliable optimal time series prediction sub-model. This avoids prediction behaviors that may lead to equipment damage or system oscillations, and guides the algorithm to find the optimal parameter combination under the premise of ensuring safety and stability.

[0038] In a preferred embodiment, in step S21, the time series prediction sub-model based on linear relationships is an ARIMA model, and step S21 includes: Construct an ARIMA model, where the model parameters p, d, and q are undetermined parameters to be optimized; The model training set is used as the input to the ARIMA model, and the composite function that integrates the basic accuracy term and the running risk penalty term is used as the objective function of the global optimization algorithm. The global optimization algorithm is used to find the optimal parameters of the ARIMA model and form the optimal model parameters. The optimal time series prediction sub-model is established based on the optimal model parameters.

[0039] Specifically, the Autoregressive Integrated Moving Average (ARIA) model is an improved non-stationary time series modeling method based on the ARMA model. Its mathematical model is as follows: (1) Difference processing: The original main steam temperature sequence y(t), i.e., the main steam temperature historical dataset in step S1, is subjected to d first-order differences to obtain the stationary sequence y^(d)(t). The difference formula is: When d=1, When d=2, ...... After d-th order difference, y^(d)(t) satisfies the stationarity condition; (2) ARMA Fitting: The stationary sequence y^(d)(t) satisfies the ARMA (p,q) model, and the formula is: Where y^(d)(t) is the stationary main steam temperature after d-th order differencing, p is the autoregression order, d is the differencing order (0≤d≤2), and q is the moving average order. , , ..., These are the autoregressive coefficients. , , ..., Let be the moving average coefficients, and ε(t) be the white noise sequence; (3) Inverse difference restoration: the prediction result y^(d) of the ARMA (p,q) model (t) Perform d inverse differences to obtain the linear prediction value of the original main steam temperature sequence, thus completing the linear trend modeling of non-stationary data.

[0040] Among them, the difference order d can be automatically determined by the enhanced Dickey-Fowler (ADF) test until the sequence is stationary, and the autoregression order p and the moving average order q are determined by the global optimization algorithm; or the autoregression order p, the difference order d, and the moving average order q are all determined by the global optimization algorithm.

[0041] Using multiple model parameters of the aforementioned ARIMA model as the feasible region for global optimization, and the model training set as the input for prediction, a composite function fusing the basic accuracy term and the operational risk penalty term is set as the objective function of the global optimization algorithm. Iterative optimization is performed to find the optimal p, d, and q values ​​as the optimal model parameters. Based on these optimal model parameters, an optimal time-series prediction sub-model is established. Through the ARIMA model in this embodiment, the non-stationary main steam temperature sequence is transformed into a stationary sequence. Furthermore, the autoregressive term captures the data's inherent lag dependency, and the moving average term offsets the influence of random disturbances, thus characterizing the non-stationary linear trend in the main steam temperature sequence.

[0042] In a preferred embodiment, in step S21, the global optimization algorithm is a quantum particle swarm optimization algorithm, and the step of determining the optimal model parameters of the time series prediction sub-model using the global optimization algorithm includes: Initialize the particle swarm for the quantum particle swarm optimization algorithm, where the position of each particle represents a set of parameter values ​​to be optimized in the ARIMA model; The model training set is predicted using an ARIMA model built based on each particle, and the fitness value of each particle is calculated according to the objective function. Based on the fitness value, the particle swarm is iteratively updated using the quantum particle swarm optimization algorithm; the globally optimal particle position obtained after iterative optimization is used as the optimal model parameter.

[0043] Specifically, the Quantum Particle Swarm Optimization (QPSO) algorithm is used to optimize the model parameters of the time-series prediction sub-model ARIMA. The process is as follows: Within a predefined parameter search space (e.g., p∈[1,10], q∈[1,10]), a set of parameters to be optimized in the ARIMA model, such as a set of (p,q) values, is used as a particle to initialize the QPSO particle swarm. In each iteration, an ARIMA model is constructed based on each particle. The training set is input into the ARIMA model for prediction, and the predicted values ​​are obtained. Based on the predicted values, the fitness value of each particle is calculated using an objective function, thereby iteratively updating the particle swarm until the globally optimal particle position, such as the corresponding (p,q) value, is obtained, which serves as the optimal model parameters. Compared to the standard PSO algorithm, the QPSO algorithm has better global search capabilities and faster convergence speed, effectively avoiding getting trapped in local optima. The model parameters determined in this way can improve the fitting accuracy of the ARIMA model.

[0044] In some alternative embodiments, in step S23, the nonlinear prediction sub-model includes a gating unit, and step S23 includes: Input the residual sequence and the corresponding operating condition characteristic variable data into the nonlinear prediction sub-model; The gating unit receives operating condition characteristic variable data and outputs a gating signal of the same dimension as the residual sequence; The residual sequence is multiplied element-wise with the gated signal to generate an enhanced residual sequence; The system is trained using enhanced residual sequences and operating condition characteristic variable data to output residual prediction sequences.

[0045] Specifically, a gating unit is set in the nonlinear prediction sub-model, which can be a small neural network. This gating unit receives a working condition context vector composed of working condition characteristic variable data and outputs a gating signal of the same dimension as the residual sequence output by the optimal time-series prediction sub-model, with element values ​​between (0,1). The residual sequence is then multiplied element-wise with the gating signal to generate the enhanced residual sequence Input_gated, as shown in the following formula: Input_gated = G(t) [E(t-1), ..., E(tw)] Here, G(t) is the gating signal, and E(t-1), ..., E(tw) are the residual sequences. The enhanced residual sequences Input_gated and the operating condition context vector are then used as the training set for the nonlinear prediction sub-model, outputting the residual prediction sequence. By using a gating unit that receives the operating condition context vector, larger weights can be generated for specific key operating conditions to amplify historical residual features related to that condition, focusing on important information. Simultaneously, smaller weights can suppress irrelevant noise and reduce interference. This enables subsequent models to learn complex dependencies, remain robust to noise, and adapt to dynamically changing complex operating conditions.

[0046] The main steam temperature prediction system based on a hybrid model of the present invention can be used to execute the main steam temperature prediction method based on a hybrid model described in the above embodiments. The system includes: The data processing module is used to acquire historical time series datasets and perform data preprocessing to form a model training set; the historical datasets include historical main steam temperature datasets and historical datasets of multiple feature variables related to main steam temperature. The model building module is used to train and build a main steam temperature prediction model based on the model training set, including: The first sub-model building module is used to construct a time-series prediction sub-model based on linear relationships. The model training set is used as the input of the time-series prediction sub-model, and the composite function that integrates the basic accuracy term and the running risk penalty term is used as the objective function of the global optimization algorithm. The global optimization algorithm is used to determine the optimal model parameters of the time-series prediction sub-model, and the optimal time-series prediction sub-model is built based on the optimal model parameters. Based on the model training set, prediction is performed through the optimal time-series prediction sub-model, and the first main steam temperature prediction sequence and residual sequence are output. The second sub-model building module is used to train and build a nonlinear prediction sub-model by taking the residual sequence and the corresponding working condition characteristic variable data as input, and outputting the residual prediction sequence. The fusion module is used to superimpose the first main steam temperature prediction sequence with the residual prediction sequence as the output value of the main steam temperature prediction model. The main steam temperature prediction module is used to obtain the real-time main steam temperature value and the real-time values ​​of multiple feature variables related to the main steam temperature, input them into the main steam temperature prediction model, and output the main steam temperature prediction value.

[0047] Specifically, this system is data-driven and can be deployed on ordinary servers or optimization stations without the need for high-performance GPUs. It relies on existing data acquisition systems and has extremely low implementation and maintenance costs, resulting in significant economic benefits.

[0048] It is understood that the above embodiments only illustrate preferred embodiments of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can freely combine the above technical features without departing from the concept of the present invention, and can also make several modifications and improvements, all of which fall within the protection scope of the present invention. Therefore, all equivalent transformations and modifications made with respect to the scope of the claims of the present invention should fall within the scope of the claims of the present invention.

Claims

1. A method for predicting main steam temperature based on a hybrid model, characterized in that, include: S1. Obtain historical time series datasets and perform data preprocessing to form a model training set; the historical datasets include historical main steam temperature datasets and historical datasets of multiple feature variables related to the main steam temperature; S2. Based on the model training set, train and establish a main steam temperature prediction model, including: S21. Construct a time series prediction sub-model based on linear relationships; The training set of the model is used as the input of the time series prediction sub-model, and the composite function that integrates the basic accuracy term and the running risk penalty term is used as the objective function of the global optimization algorithm. The optimal model parameters of the time series prediction sub-model are determined by the global optimization algorithm, and the optimal time series prediction sub-model is established based on the optimal model parameters. S22. Based on the model training set, prediction is performed using the optimal time-series prediction sub-model to output the first main steam temperature prediction sequence and the residual sequence; S23. Using the residual sequence and the corresponding operating condition characteristic variable data as input, train and establish a nonlinear prediction sub-model, and output the residual prediction sequence; S24. The first main steam temperature prediction sequence is superimposed with the residual prediction sequence as the output value of the main steam temperature prediction model; S3. Obtain the real-time main steam temperature value and the real-time values ​​of the multiple feature variables related to the main steam temperature, input them into the main steam temperature prediction model, and output the main steam temperature prediction value.

2. The main steam temperature prediction method based on a hybrid model according to claim 1, characterized in that, The multiple characteristic variables related to the main steam temperature include: unit load command, total fuel quantity, total air volume, feedwater flow rate, main steam flow rate, flow rate and valve position of each stage of water spray desuperheater and / or steam temperature and pressure at the inlet and outlet of each stage of superheater.

3. The main steam temperature prediction method based on a hybrid model according to claim 1, characterized in that, In step S21, the operational risk penalty term includes a prediction stability penalty term, a dynamic process penalty term, and an operational safety penalty term, and the objective function is: Fitness = w1 F_accuracy+w2 F_stability+w3 F_transient+w4 F_safety Wherein, F_accuracy is the basic accuracy term, F_stability is the prediction stability penalty term, F_transient is the dynamic process penalty term, F_safety is the operational safety penalty term, and w1, w2, w3, and w4 are the weight coefficients of each sub-item.

4. The main steam temperature prediction method based on a hybrid model according to claim 1, characterized in that, In step S21, the time series prediction sub-model based on linear relationship is an ARIMA model, and step S21 includes: Construct the ARIMA model, wherein the model parameters p, d, and q of the ARIMA model are undetermined parameters to be optimized; The model training set is used as the input of the ARIMA model, and the composite function of the fusion basic accuracy term and the running risk penalty term is used as the objective function of the global optimization algorithm. The global optimization algorithm is used to find the optimization parameters of the ARIMA model to form the optimal model parameters. The optimal time series prediction sub-model is established based on the optimal model parameters.

5. The main steam temperature prediction method based on a hybrid model according to claim 4, characterized in that, In step S21, the global optimization algorithm is a quantum particle swarm optimization algorithm, and the step of determining the optimal model parameters of the time series prediction sub-model using the global optimization algorithm includes: Initialize the particle swarm of the quantum particle swarm optimization algorithm, wherein the position of each particle represents a set of parameter values ​​to be optimized in the ARIMA model; The training set of the model is predicted using an ARIMA model built based on each particle, and the fitness value of each particle is calculated according to the objective function. Based on the fitness value, the particle swarm is iteratively updated using the quantum particle swarm optimization algorithm; the globally optimal particle position obtained after iterative optimization is used as the optimal model parameter.

6. The main steam temperature prediction method based on a hybrid model according to claim 1, characterized in that, In step S23, the nonlinear prediction sub-model includes a gating unit, and step S23 includes: The residual sequence and the corresponding operating condition characteristic variable data are input into the nonlinear prediction sub-model; The gating unit receives the operating condition characteristic variable data and outputs a gating signal of the same dimension as the residual sequence; The residual sequence is multiplied element-wise with the gated signal to generate an enhanced residual sequence; The enhanced residual sequence is trained based on the operating condition characteristic variable data, and the residual prediction sequence is output.

7. The main steam temperature prediction method based on a hybrid model according to claim 1, characterized in that, In step S1, the data preprocessing step includes: non-linearly aligning the historical dataset of each feature variable related to the main steam temperature with the historical dataset of the main steam temperature, calculating the correlation with the main steam temperature, and removing feature variables with low correlation to generate a core feature variable set.

8. The main steam temperature prediction method based on a hybrid model according to claim 7, characterized in that, The nonlinear alignment step includes: The optimal nonlinear alignment path between the historical dataset of each feature variable related to the main steam temperature and the historical dataset of the main steam temperature is calculated using a dynamic time warping algorithm. For each feature variable related to the main steam temperature, the time offset at each time point is calculated according to the corresponding optimal nonlinear alignment path, and the corresponding time delay distribution is generated. The expected value of the time delay distribution is calculated as the average time delay value of the feature variable. The feature variable is then shifted and aligned according to the average time delay value.

9. The main steam temperature prediction method based on a hybrid model according to claim 7, characterized in that, In step S1, the data preprocessing step further includes: obtaining a data subset under sparse samples and / or key operating conditions in the core feature variable set, and using a time series generative adversarial network to learn the data subset, generating corresponding synthetic sequence samples, and expanding them into the model training set.

10. A main steam temperature prediction system based on a hybrid model, characterized in that, include: The data processing module is used to acquire historical time series datasets and perform data preprocessing to form a model training set; the historical datasets include historical main steam temperature datasets and historical datasets of multiple feature variables related to the main steam temperature; The model building module, used to train and build a main steam temperature prediction model based on the model training set, includes: The first sub-model building module is used to construct a time-series prediction sub-model based on linear relationships. The model training set is used as input to the time-series prediction sub-model, and a composite function fusing the basic accuracy term and the operational risk penalty term is used as the objective function of the global optimization algorithm. The global optimization algorithm is used to determine the optimal model parameters of the time-series prediction sub-model, and an optimal time-series prediction sub-model is established based on these optimal model parameters. Based on the model training set, prediction is performed using the optimal time-series prediction sub-model, outputting a first main steam temperature prediction sequence and a residual sequence. The second sub-model building module is used to take the residual sequence and the corresponding working condition characteristic variable data as input, train and build a nonlinear prediction sub-model, and output the residual prediction sequence. The fusion module is used to superimpose the first main steam temperature prediction sequence with the residual prediction sequence as the output value of the main steam temperature prediction model. The main steam temperature prediction module is used to obtain the real-time main steam temperature value and the real-time values ​​of the multiple feature variables related to the main steam temperature, input them into the main steam temperature prediction model, and output the main steam temperature prediction value.

Images