Intelligent optimization method for heat tracing system of power plant based on environment temperature fluctuation

Abstract

The invention provides a power plant heat tracing system intelligent optimization method based on environment temperature fluctuation, and relates to the technical field of temperature control of a power plant heat tracing system, and the method comprises the following steps: S1, data collection and preprocessing: deploying a sensor array, collecting an operation data signal # imgabs0 # of a power plant heat tracing pipeline, carrying out the preprocessing of the operation data signal # imgabs1 #, and storing the preprocessed operation data signal # imgabs0 #; outputting a standardized feature vector; s2, model establishment and parameter optimization: constructing a physical-data dual-drive heat transfer model, and inputting standardized feature vector optimization initial parameters; compared with the prior art, the method has the following beneficial effects: firstly, control parameters can be automatically adjusted according to factors such as thermal load change, external environment temperature fluctuation and equipment aging; and secondly, the temperature trend in the future short time can be predicted by constructing a heat transfer dynamic model, the control strategy is adjusted in advance, the temperature is prevented from being out of control, heat dissipation can be increased in advance or heat source input can be adjusted in advance, and safe and efficient operation of the system is effectively guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of temperature control of the tracing heating system in a power plant, and particularly to an intelligent optimization method for the tracing heating system in a power plant based on ambient temperature fluctuations. Background Art

[0002] In the prior art, the tracing heating system in a power plant is mainly used to prevent media such as water, steam or other liquids in pipelines from freezing or solidifying in a low-temperature environment, ensuring the normal operation of the system, and the temperature control of electric tracing heating usually adopts a control algorithm with fixed parameters. This method can meet the basic temperature control requirements to a certain extent, but in the face of complex and changeable operating conditions, it often shows the problem of insufficient adaptability.

[0003] When the heat load changes, the ambient temperature fluctuates or the heat transfer efficiency changes due to equipment aging, the control algorithm with fixed parameters cannot adjust the control parameters in time, which may lead to unstable temperature control. Throughout the year, the ambient temperature changes greatly, and the traditional tracing heating system still controls according to fixed parameters. In summer, the temperature may be too high, and in winter, it may be too low, unable to meet the requirements of precise and stable temperature for the system, thereby affecting the normal operation of the system. At the same time, the traditional method reacts only when the temperature is about to get out of control, which is not convenient for predicting and adjusting the control strategy in advance, posing a potential threat to the safe and efficient operation of the system. Therefore, the existing temperature control technology for the tracing heating system in a power plant needs to be improved to adapt to complex working conditions and improve the adaptability, stability and safety of the system. Summary of the Invention

[0004] Aiming at the deficiencies of the above-mentioned prior art, the technical problem to be solved by the present invention is to provide an intelligent optimization method for the tracing heating system in a power plant based on ambient temperature fluctuations, which can automatically adjust the control parameters based on real-time working conditions and predict the temperature trend in advance, reducing temperature fluctuations.

[0005] To solve the above technical problem, the technical solution adopted by the present invention is: The present invention provides an intelligent optimization method for the tracing heating system in a power plant based on ambient temperature fluctuations, including the following steps: S1. Data acquisition and preprocessing: Deploy a sensor array and collect the operation data signals of the tracing heating pipelines in the power plant , and perform preprocessing on the operation data signals to output a standardized feature vector; S2. Model establishment and parameter optimization: Construct a physics-data dual-driven heat transfer model and input the standardized feature vector to optimize the initial parameters; S3. Seasonal intelligent control: Construct a differential parameter library according to seasonal characteristics and perform intelligent switching in the transition season according to the optimized heat transfer model; S4, Fault Self-Healing and Model Update: Establish a fault classification response mechanism and construct an online evolution framework for the heat transfer model.

[0006] In the preferred solution, in step S1, for the operating data signal The specific steps of preprocessing are as follows: Use a 4-layer Coiflet wavelet decomposition and reconstruction algorithm to denoise the operating data signal After denoising, the signal Is used as the input data matrix , Separate the low-rank normal data and sparse abnormal data through RobustPCA, and use the spatio-temporal K-nearest neighbor interpolation method for data repair. The repaired data is input to PCA, calculate and screen the principal component Z, and then through the weighted fusion of PCA and mutual information method, construct a fusion feature vector ; Among them, the sensor array includes temperature sensors, current sensors, voltage sensors, and flow sensors; The operating data signal Includes pipeline temperature , Heating tape power , Ambient temperature , Medium flow rate .

[0007] In the preferred solution, step S1 also includes the following steps: Based on a 10-second sliding window, fit a cubic spline curve to calculate the temperature change rate r; Let , , Be three change rate thresholds, and ( ), Dynamically switch the sampling frequency According to the relationship between the temperature change rate r and the threshold. The formula is as follows: ; Among them, Is the acquisition frequency.

[0008] In the preferred solution, the specific steps in step S2 are as follows: S21, Physical-Data Dual-Driven Heat Transfer Model Construction: Based on the Fourier equation, construct a steady-state heat transfer model, introduce the medium flow rate to nonlinearly correct the thermal conductivity, calculate the steady-state temperature distribution, and obtain the pipeline basic temperature distribution , According to LSTM to capture the temporal dynamic characteristics, input the normalized historical temperature , Heating tape power , Ambient temperature , Medium flow rate , Aging factor , and update the hidden state at the current moment through the internal control mechanism , output the hidden state mapped to the temperature fluctuation component through the fully connected layer to generate a dynamic correction term , to correct the steady-state temperature of the steady-state heat transfer model; The Fourier pipeline basic temperature distribution and the LSTM dynamic correction term are superimposed to perform temperature prediction; S22. Bayesian hyperparameter optimization: Use Bayesian optimization to adjust the number of nodes in the LSTM hidden layer , learning rate , regularization coefficient , according to the historical evaluation results, construct a hyperparameter distribution, use the Monte Carlo sampling method to sample the hyperparameter space, calculate the expected improvement value of each sampling point, and select the hyperparameter combination with the largest expected improvement value as the hyperparameters for the next iteration through the expected improvement criterion; Among them, the optimal hyperparameter combination is: , , ; Add Gaussian noise to the LSTM weights Generate 100 groups of prediction results, calculate the mean and standard deviation based on the prediction results to obtain a 95% confidence interval, and evaluate the model uncertainty; S23. Dynamically update the aging compensation factor: Fit the aging historical data through linear regression and incrementally update the aging factor online , and embed the aging factor into the temperature prediction formula to compensate for the decrease in thermal efficiency caused by equipment aging.

[0009] In the preferred solution, in step S21, the hidden state is mapped through the fully connected layer to the temperature fluctuation component, and the calculation formula of the dynamic correction term is as follows: ; The Fourier pipeline basic temperature distribution and the LSTM dynamic correction term The temperature prediction output after superposition is expressed by the formula as follows: ; Among them, is the weight matrix of the fully connected layer, is the bias vector, is the hidden state output by the LSTM; In step S22, the calculation formula for screening the hyperparameter combination by maximizing the expected improvement is as follows: ; According to the mean and standard deviation of the prediction results, the formula for outputting the confidence interval is as follows: ; Among them, the mean represents the average level of the prediction results, and the formula is , and the standard deviation measures the degree of dispersion of the prediction results, and the formula is ; In step S23, the formula for adjusting the attenuation rate in real time according to the new data is as follows: ; Among them, is the learning rate, which is used to control the update step size, is the currently observed aging factor, is the aging factor predicted according to the previous model; Embed the aging factor into the prediction formula as follows: .

[0010] In the preferred solution, the specific steps in step S3 are as follows: S31. Construction of the seasonal characteristic parameter library: Collect historical environmental temperature data , and according to the historical environmental temperature data mean and standard deviation , define the reference temperature , and quantify the seasonal characteristic index through the formula , and establish a control parameter library corresponding to different seasons according to the quantization results; S32. Intelligent switching in the transition season: Use the trained LSTM model to input the current and recent environmental temperature data, and predict the temperature sequence for the next 72 time steps , calculate the mean of the time series and the predicted temperature sequence, calculate the temperature trend slope through a specific formula , set a threshold, when 72 consecutive time steps meet specific conditions, trigger the seasonal mode switch according to the positive and negative of the temperature trend slope ; S33. Dynamic weight allocation: Dynamically allocate the winter weight and the summer weight according to the seasonal characteristic index and the real-time temperature deviation , fuse the winter and summer control parameters to obtain the final adjustment step size and the response time ; S34. Predictive reactive control: Define the input variable as the real-time temperature deviation and the temperature change rate are respectively processed by Gaussian and triangular membership functions for fuzzification, and the corresponding duty cycle adjustment amount is mapped according to the fuzzified input variables The duty cycle adjustment amount is obtained by defuzzifying the output through the centroid method Combined with the season sensitivity coefficient , aging factor and time factor, a heat balance equation is established to dynamically adjust the power of the heating tape. According to the duty cycle adjustment amount of the fuzzy output , season weight and response time dynamically optimize the PWM duty cycle; The aging factor is coupled with the season parameters to calculate the dynamic attenuation factor , and the power is finally corrected by integrating the heat balance power , dynamic attenuation factor and the real-time temperature deviation .

[0011] In the preferred solution, in step S31, the formula for quantifying the season characteristic index is as follows: ; wherein, represents the relationship between the dispersion degree and the average level of temperature data, represents non-linear adjustment according to the ratio of the average temperature to the reference temperature; The formula for storing the parameter settings from different seasons to the parameter library is: ; wherein, when , it is determined as the winter mode, and the adjustment step size is set, and the response time seconds; when , it is determined as the summer mode, and the adjustment step size is set, and the response time seconds; In step S32, the calculation formula for the temperature trend slope is: ; wherein, the numerator part represents the degree of co-variation of time and temperature deviating from their respective means, and the denominator part is used for normalization processing; The formula for determining the switched season mode according to the positive and negative of the temperature trend slope is: ; wherein, the specific condition is when , indicating that the temperature is on a downward trend, switch to winter mode; When , indicating that the temperature is on an upward trend, switch to summer mode; In step S33, first calculate the winter weight: , then calculate the summer weight: ; Among them, the seasonal characteristic index rises, then the winter weight shows an upward trend, represents the absolute value of the temperature deviation, and the winter weight increases with the increase; Substitute the parameter settings in the parameter library to adjust the step size and the response time , then the final adjustment step size and the response time are expressed by the following formula: ; ; In step S34, according to the duty cycle adjustment amount output by fuzzy control, the seasonal weight, and the response time , the formula for dynamically optimizing the PWM duty cycle is expressed as follows: ; Among them, is the current duty cycle, is the winter weight, is the final response time; The power after comprehensive thermal balance adjustment, the dynamic attenuation factor , and the real-time temperature deviation , the power is finally corrected through the following formula: .

[0012] In the preferred solution, the specific steps in step S4 are as follows: S41. Sensor fault detection and compensation: Obtain the historical temperature data of the sensor , and calculate the historical temperature standard deviation . Trigger the Bayesian anomaly probability calculation through the redundant consistency score, locate the faulty sensor, and repair the data of the faulty sensor according to the spatio-temporal Kriging interpolation method; S42. Detection and switching of the heating tape damage: Determine the abnormal harmonic energy ratio to drive the power adjustment of the standby system; Among them, the abnormal proportion of harmonic energy includes the abnormal proportion of the energy in the characteristic frequency band of the current signal detected by frequency-domain analysis and the abnormal correlation between the detected power input and the temperature response by covariance analysis; S43. Model update incremental training: allocate data weights according to the prediction error, and adjust the LSTM weights using the online Bayesian update formula; S44. Q-learning reinforcement learning optimization: construct the state-action space and the reward function, and update the policy network parameters; S45. Aging factor dynamic calibration: establish that the aging factor follows an exponential decay model , set the state variables, obtain the state transition equation, and construct the observation equation. According to the observation equation, obtain the linearized observation equation, and use the extended Kalman filter to estimate the aging rate in real time , and embed the estimated aging factor into the heat transfer model to correct the thermal conductivity .

[0013] In the preferred solution, in step S41, the redundant consistency score formula in the sensor array is: ; Among them, is the historical temperature standard deviation, represents the maximum value of the temperature difference between any two sensors in the sensor array, represents a reference value for normalization processing; If the consistency score < 0.8, it is determined that there is an abnormal sensor; When there is an abnormal sensor, the Bayesian abnormal probability calculation is triggered and is represented by the following formula: ; Among them, is the weight of the th Gaussian distribution, is the mean value, is the variance; If , then mark the sensor as faulty; The formula for repairing the data of the faulty sensor using the spatio-temporal Kriging interpolation method is as follows: ; Among them, is the position of the current faulty sensor and the position of the neighboring sensor is the Euclidean distance, is the current time and the measurement time of the neighboring sensor The difference is the weight calculated according to the spatio-temporal distance between the adjacent sensor and the faulty sensor and are the spatial and temporal scale parameters; In step S42, the specific calculation steps for the frequency-domain analysis to detect the abnormal energy proportion in the characteristic frequency band of the current signal are as follows: For the current signal of the heating cable perform a fast Fourier transform to obtain the frequency-domain signal , and the formula for extracting the energy of the 1 kHz frequency band is: ; calculate the total energy as: ; where indicates that the frequency band range of the energy signal is 1 kHz ± 50 Hz, represents the total energy of the entire frequency domain; If , then it is determined that the energy proportion is abnormal as an open circuit; The specific calculation steps for covariance analysis to detect the abnormal correlation between power input and temperature response are as follows: ; where is the temperature change, is the power change, is the average value of the temperature change , is the average value of the power change ; If , then it is determined that the energy proportion is abnormal as the failure of the heating cable; When it is determined that the energy proportion is abnormal, the formula for adjusting the power of the standby heating cable according to the power compensation formula is: ; where is the nominal power of the standby heating cable, is the compensation gain, is the reference temperature.

[0014] In the preferred solution, in step S43, obtain the model predicted temperature and the actual measured temperature , calculate the prediction error according to the model predicted temperature and the actual measured temperature , and calculate the variance estimate value of the error according to the prediction error , and through the prediction error and the variance estimate value Define data weights , the formula is as follows: ; Assume that the model parameters follow a Gaussian distribution and are updated by verifying the Bayesian update formula: ; Among them, is the observation noise variance, and the numerator part represents the correction of the parameter mean by the new data, and the denominator is used to balance the influence of the new data and the prior information; In step S44, the state space defines the state as: ; Among them, is the temperature deviation, is the temperature change rate, is the aging factor, is the seasonal characteristic index; The action space defines the state as: ; Among them, is the power adjustment amount of the tracing cable, is the PWM duty cycle adjustment amount; Calculate the total power, and the reward function is: ; Among them, , , are the weight coefficients, measures the magnitude of the temperature deviation, represents the total power consumption of the system, is used to control the energy consumption, measures the change amplitude of the action, is used to avoid frequent adjustment of the system; Use the Actor-Critic framework to update the policy network , and the policy network outputs the probability distribution of the action according to the current state and updates the parameters of the policy network according to the following policy gradient formula : ; Among them, is the action value function, is the state value function; In step S45, establish the formula for the aging factor following an exponential decay model as: ; Wherein, is the initial value of the aging factor, is the aging rate, is the noise term; The extended Kalman filter is used to estimate the aging rate in real time The formula is: ; Wherein, is the Kalman gain, which is dynamically adjusted according to the current state and observed data; The real-time estimated aging factor is embedded in the heat transfer model to correct the thermal conductivity The formula is: .

[0015] The present invention provides an intelligent optimization method for the tracing heating system of a power plant based on ambient temperature fluctuations. Compared with the prior art, it has the following beneficial effects: First, the present invention can automatically adjust control parameters according to factors such as heat load changes, ambient temperature fluctuations, and equipment aging. Compared with the fixed-parameter control algorithms in the prior art, it has higher adaptability. In different seasons, it can adjust parameters such as the flow rate of the cooling medium and the heating power in a timely manner according to the change of the ambient temperature, ensuring that the temperature is accurately and stably within the set range; Second, the present invention can predict the temperature trend in the short term in the future by constructing a heat transfer dynamic model, and adjust the control strategy in advance to avoid temperature runaway. While the prior art usually reacts when the temperature is about to get out of control and cannot prevent problems in advance. During peak electricity consumption periods, it can increase heat dissipation or adjust the heat source input in advance, effectively ensuring the safe and efficient operation of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] The present invention will be further described below in conjunction with the drawings and embodiments: Figure 1 is the main flow chart of the present invention; Figure 2 is the present invention Figure 1 in the data preprocessing flow chart; Figure 3 is the present invention Figure 2 in the dynamic sampling adjustment flow chart; Figure 4 is the present invention Figure 1 in the heat transfer model construction flow chart; Figure 5 is the present invention Figure 1 in the seasonal adaptive control flow chart; Figure 6 is the present inventionFigure 1 Fault self-healing and update flow chart in . DETAILED DESCRIPTION

[0017] In order to better understand the purpose, structure and function of the present invention, the embodiments and features in the embodiments of the present application can be combined with each other without conflict. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0018] Example 1 like Figures 1 to 6 As shown, this embodiment provides an intelligent optimization method for a power plant heating system based on ambient temperature fluctuations, and the specific steps are as follows: S1. The specific steps of data collection and preprocessing are as follows: S11. Sensor deployment and data collection: Deploy redundant sensor arrays in key areas of the pipeline to collect and output measured operating data signals in real time. , including pipe temperature , Heating cable power , Ambient temperature , medium flow ; Among them, the sensor array consists of 5 sensors in each group, and the sensors include temperature sensors, current and voltage sensors, and flow sensors; For the data collected by each group of sensor arrays, temperature data is taken as an example, and other types of data are processed in the same way.

[0019] The temperature sensor covers the following temperature ranges:

[0020] First, calculate multiple sets of temperature sensor measurements The median M, , the median can avoid the impact of individual extreme values ​​on the overall data, and then calculate each measurement value Absolute deviation from the median M: , which determines how far each data point deviates from the average level, and then sets a deviation threshold ,like , then determine It is an outlier and is removed; For non-outliers, the standard deviation-based weighted average method is used for fusion, and the standard deviation of non-outliers is calculated, and then the weight Calculate the fused temperature value based on the sensor's historical stability data ; S12. Data preprocessing: Use Coiflet wavelet to process the collected operation data signals Decompose the running data signal into 4 layers Decompose at different scales and positions to obtain coefficients of different frequency components, achieving the purpose of extracting low-frequency (approximate) and high-frequency (detail) components.

[0021] Determine the threshold by estimating the noise variance, process the detail coefficients, effectively suppress high-frequency noise, while retaining the effective signal, and reconstruct the detail coefficients after threshold processing to generate the denoised signal. Realize the restoration and denoising of the signal.

[0022] S122. Outlier rejection: Use the Robust PCA algorithm to process the data. According to the denoised signal Arrange it in a data matrix according to the time series. For the data matrix Decompose it into a low-rank matrix L and a sparse matrix S. By solving Under the constraint condition Separate normal data from abnormal data; Use the Alternating Direction Method of Multipliers (ADMM) to iteratively solve and Through the way of alternating optimization, gradually approach the optimal and matrices, improve the accuracy of outlier separation. For the non-zero elements (i.e., outlier positions) in the sparse matrix Use Nearest neighbor interpolation ( ) for repair. Calculate the repaired value of the outlier by using the weighted interpolation method to obtain the repaired outlier.

[0023] Calculate the covariance matrix of the denoised and repaired data Select the eigenvectors corresponding to the first 3 largest eigenvalues to form a projection matrix Through the principal component projection formula Reduce the data to a 3D principal component space Realize data dimensionality reduction, reduce data redundancy, and calculate the principal component () and the mutual information with the target temperature Use the mutual information calculation method based on kernel density estimation. First, estimate the joint probability density function and the marginal probability density functions and 、 .

[0024] The correlation between the principal components and the target temperature is measured by calculating the mutual information, and the principal components that contribute the most to temperature prediction are selected. The selected principal components are fused with the dynamic change data of the medium flow rate, and the medium flow rate data is introduced as , then the final eigenvector is: ; A feature vector containing various key information is formed by the above fusion formula with the selected principal components and the flow rate data, providing more comprehensive and representative input data for subsequent model training.

[0025] S13. Dynamic sampling frequency setting based on the sensor array The sensor array calculates the temperature change rate based on a sliding window; Among them, the window size = 10 seconds; Let the temperature data within the window be: ; Among them, is the number of data points within the window; The change of temperature over time is approximated by fitting a cubic spline curve , and let the expression of the cubic spline curve on each sub-interval be: ; The coefficients are determined by satisfying certain boundary conditions and continuity conditions, and then the temperature change rate at the midpoint time of the window is calculated: , that is: ; Among them, is the starting point of the sub-interval containing ; Calculating the temperature change rate by fitting a cubic spline curve can more accurately capture the trend of temperature change, especially in the case of complex temperature changes.

[0026] Set three change rate thresholds , , ([[]] ); If , then the sampling frequency ; if , then the sampling frequency ; if , then the sampling frequency ; if , then the sampling frequency .

[0027] Among them, , , , , , , .

[0028] S2. Selection and Optimization of Heat Transfer Dynamic Model S21. Construction of Physical Data Double-Driven Model S211. Physical Driving Layer (Fourier Heat Conduction Equation) Ignoring the time term, the basic temperature distribution of the pipeline is calculated based on the Fourier equation , and the specific formula is as follows: ; where is the Hamiltonian operator. In the Cartesian coordinate system , this equation represents the heat conduction in the medium under steady state. The divergence of the heat flux density is 0, that is, there is no heat accumulation in space.

[0029] In actual solution, for complex pipeline geometries, the finite element method is adopted.

[0030] Suppose the pipeline area is discretized into elements. For each element , according to the Galerkin method, multiply the Fourier equation: by the test function and integrate over the element to obtain the following: ; Assemble the equations of all elements to form a linear equation system: ; where is the stiffness matrix, is the nodal temperature vector, is the load vector.

[0031] Solve the above linear equation system to obtain the basic temperature distribution of the pipeline .

[0032] S212. Data-Driven Layer (LSTM Dynamic Correction) The input sequence includes historical temperature , tracing tape power , ambient temperature , medium flow rate , aging factor . To eliminate the dimension difference and improve the model convergence speed, the input features are normalized, and the formula is: ; where represents different input features (such as , , etc.), is the mean of the feature , and is the standard deviation of the feature .

[0033] Next, the formula for calculating the mean of the feature is as follows: ; where is the number of samples; Calculate the standard deviation of the feature : ; The LSTM hidden state update formula is expressed as follows: ; where is the normalized input sequence, is the hidden state at the previous moment, is the parameter set of the LSTM, including the weight matrix and the bias vector.

[0034] Inside the LSTM cell, the flow of information is controlled by the input gate , the forget gate , the output gate and the memory cell . By calculating the input gate , the forget gate , the output gate and the memory cell , the final hidden state is: ; where is the sigmoid function, is the element-wise multiplication.

[0035] Through the above formula, the LSTM can capture the temporal dynamic characteristics and output the hidden state at the current moment, effectively handling the long-term dependencies in the time series.

[0036] Map the hidden state to the temperature fluctuation component through the fully connected layer. The formula is as follows: ; where is the weight matrix of the fully connected layer, is the bias vector.

[0037] The above formula maps the hidden state output by the LSTM to the temperature fluctuation space, generating the dynamic correction term , which is used to correct the steady-state temperature of the physical model to reflect the dynamic changes in the actual heat transfer process.

[0038] S22. Model Training and Optimization S221. Hybrid model fusion: Combining the steady-state solution of the physical model with the LSTM dynamic correction term for temperature prediction, the formula is as follows: ; where the steady-state temperature calculated by the physical model provides a basic temperature distribution, and the LSTM dynamic correction term considers the dynamic factors in the heat transfer process. The sum of the two forms a complete temperature prediction , improving the generalization of the model through this physical data dual-driven method, enabling it to better adapt to different working conditions.

[0039] S222. Bayesian hyperparameter optimization (TreestructuredParzenEstimator, TPE) To improve efficiency, select a part of the hyperparameter space as the search space; where the search space includes the number of nodes in the LSTM hidden layer , learning rate , regularization coefficient .

[0040] Among them, the number of nodes in the hidden layer takes values in the range of [64, 256], and the learning rate takes values in the range of ; the regularization coefficient takes values in the range of , and its role is to control the model complexity and prevent overfitting; TPE probability modeling: Based on historical evaluation results, construct a hyperparameter distribution, and the formula is: ; where represents the hyperparameter set, represents the historical evaluation data, is the likelihood function, indicating the probability of observing the data under the hyperparameter , is the prior distribution, representing the initial understanding of the hyperparameter .

[0041] Through Bayesian inference, use historical data to update the prior distribution to obtain the posterior distribution , thus selecting high-potential hyperparameter combinations. In the above calculations, a Gaussian process is used to estimate the likelihood function: ; Let the historical evaluation data be , ; where is a hyperparameter combination, is the corresponding model evaluation metric.

[0042] The Gaussian process assumes that the data follows a Gaussian distribution and measures the similarity between hyperparameters through a kernel function. The radial basis function kernel is used: ; where, is the signal variance, is the length scale parameter. By calculating and optimizing the Gaussian process, the estimated value of the likelihood function is obtained, and then the posterior distribution is calculated. In this process, the values of the hyperparameters are all within , , range.

[0043] Expected Improvement (EI) maximization: Select the hyperparameters that maximize the expected improvement. The formula is: ; where, is the model evaluation metric under the hyperparameters , is the current optimal model evaluation metric. This formula balances exploration and exploitation, searches for hyperparameter combinations in the hyperparameter space that may bring greater improvement, and accelerates model convergence.

[0044] In actual calculations, the Monte Carlo sampling method is used to approximately calculate the expected improvement. Then, the hyperparameter space is sampled, the expected improvement value of each sampling point is calculated, and the hyperparameters with the largest expected improvement value are selected as the hyperparameters for the next iteration.

[0045] S223. Uncertainty Evaluation (Monte Carlo Sampling) Weight Perturbation Sampling: Add Gaussian noise to the LSTM weights , and generate 100 groups of prediction results. The formula is: ; where, is the noise vector for the th sampling, represents the output of the LSTM for the input under the weights after adding noise.

[0046] The uncertainty of the model parameters is simulated through the above formula. Considering that in actual training, there is a certain randomness in the model weights, different weights may lead to different prediction results.

[0047] The mean and standard deviation of the prediction results are statistically calculated. The formula is: ; ; Then, the 95% confidence interval is output: ; Among them, the mean value represents the average level of the prediction result, and the standard deviation measures the degree of dispersion of the prediction result. Similarly, the optimized values of the hyperparameters will indirectly affect the degree of dispersion and the mean value of the prediction result, and thus change the range of the confidence interval, making the uncertainty evaluation result more consistent with the actual performance of the model under specific hyperparameter settings.

[0048] S224. Dynamic update of the aging compensation factor Linear regression fitting of the aging historical data: Let the aging factor decay linearly with time, and the formula is: ; Fit (annual decay rate) through the least squares method. Collect the historical data of the aging factor changing with time , , and define the objective function: ; For with respect to take the derivative and set the derivative to 0, that is: ; Solving the above equation can obtain the estimated value of , so as to determine the decay relationship of the aging factor with time. The overall performance of the model is affected by the optimization result of the hyperparameters, and the calculation of the aging compensation factor depends on the learning and fitting ability of the model to the data. Therefore, the change of the hyperparameter value within the specified range will indirectly affect the fitting result, making the calculation of the aging factor more consistent with the actual performance of the model.

[0049] Online incremental update: Adjust the decay rate daily according to the new data, and the formula is: ; Among them, is the learning rate, used to control the update step size, is the currently observed aging factor, is the aging factor predicted by the previous model.

[0050] Temperature prediction correction: Embed the aging factor into the prediction formula, and the formula is: .

[0051] In practical applications, as the equipment ages, the thermal efficiency will decrease. The temperature prediction value is corrected through this formula to improve the prediction accuracy. Since the calculation of the aging factor is indirectly affected by the hyperparameter optimization, the final temperature prediction correction result will also change accordingly, making the predicted value more in line with the actual situation.

[0052] S3. Seasonal intelligent control strategy S31. Construction of Differentiated Parameter Library S311. Quantification of Seasonal Characteristics First, collect historical ambient temperature data , and calculate the mean of the historical ambient temperature data . Then, calculate the standard deviation of the historical ambient temperature as: ; The above formula is used to measure the dispersion degree of historical ambient temperature data

[0053] Then, define the reference temperature , and quantify the seasonal characteristics through the formula:

[0054] Among them, reflects the relationship between the dispersion degree and the average level of temperature data , and then makes a non-linear adjustment according to the ratio of the average temperature to the reference temperature

[0055] In winter, the ambient temperature is low and fluctuates greatly is small is relatively large, making ; In summer, the ambient temperature is high and fluctuates little is large is relatively small, resulting in , thus realizing the quantitative distinction of seasonal characteristics

[0056] S312. Parameter Pre-storage Rule According to the quantified seasonal characteristic index , pre-store the control parameters of different seasons to construct a parameter library

[0057] When , it is determined as the winter mode, and set the adjustment step size , and the response time seconds. In winter, the temperature is low. A larger adjustment step size can respond more quickly to temperature changes, and shortening the response time can compensate for heat dissipation losses in time to ensure the stable operation of the system

[0058] When , it is determined as the summer mode , seconds. In summer, the temperature is high. A smaller adjustment step size can avoid over-regulation, and a longer response time helps to prevent the system from adjusting frequently and improve the system stability

[0059] The formula for storing the parameter settings of different seasons is: ; ​During the control process, corresponding parameters can be quickly called according to seasonal characteristics.

[0060] S32. Intelligent switching in the transitional season S321. Temperature trend prediction Using the trained LSTM model, input the current and recent ambient temperature data, and output the temperature sequence for the next 3 days (with each hour as a time step, a total of 72 time steps). .

[0061] Calculate the mean of the time series and the mean of the temperature sequence . These two formulas are used to obtain the average levels of the time series and the predicted temperature sequence respectively.

[0062] Temperature trend slope The calculation formula is: ; Among them, the numerator part reflects the degree of co-variation of time and temperature deviating from their respective means, and the denominator part is used for normalization. By calculating this slope, the future temperature change trend can be accurately judged.

[0063] S322. Switching conditions Set a threshold. When it satisfies for 3 consecutive days (i.e., 72 time steps), trigger the seasonal mode switch, and determine the switched seasonal mode according to the positive or negative of the temperature trend slope. The formula is: ; When , it indicates that the temperature decline trend is obvious, and switch to the winter mode; when , it shows that the temperature rise trend is significant, and switch to the summer mode to achieve intelligent switching in the transitional season.

[0064] S33. Dynamic weight allocation S331. Weight calculation: According to the seasonal characteristic index and the real-time temperature deviation calculate the winter weight and the summer weight . First, calculate the winter weight: ; Among them, The larger it is, the closer the season is to winter, and the greater its impact on the weight; represents the absolute value of the temperature deviation. The greater the temperature deviation, the greater the winter weight will be accordingly.

[0065] Then, through the calculation formula of the summer weight: .

[0066] S332. Control parameter fusion: Use the calculated weights to fuse the control parameters in winter and summer to obtain the final adjustment step size and response time. The formula for the final adjustment step size is: ; where 2 is the adjustment step size in winter and 1 is the adjustment step size in summer.

[0067] According to the weight distribution, when it is closer to winter, is larger, and the final adjustment step size is closer to the adjustment step size in winter to achieve fast response; when it is close to summer, is larger, and the adjustment step size is closer to the adjustment step size in summer to avoid over-adjustment. The formula for the final response time is: ; Achieve a balance between fast response in winter and stability in summer for the response time through weight distribution.

[0068] S34. Predictive reactive control S341. Fuzzy logic controller design: Definition of input variables, real-time temperature deviation : ; where is the set target temperature, is the actually measured temperature. This formula is used to calculate the difference between the actual temperature and the target temperature, providing a basis for subsequent control.

[0069] Temperature change rate is: ; where is the temperature at the current moment, is the temperature at the previous moment, is the time interval. This formula is used to measure how fast the temperature changes over time.

[0070] For the actual temperature deviation , use a Gaussian membership function: , ; where represents the value of the actual temperature deviation , is the center value of the membership function, is the standard deviation. Different values correspond to different fuzzy sets, such as "negative large" (NB), "negative small" (NS), "zero" (ZO), "positive small" (PS), "positive large" (PB).

[0071] By adjusting and The value can define the coverage range and membership degree of each fuzzy set for the temperature deviation according to the actual situation, realizing the fuzzy processing of the temperature deviation. For the temperature change rate , a triangular membership function is adopted: , ; Among them, represents the value of the temperature change rate , is the vertex value of the membership function, is the base width. Different values correspond to the three fuzzy sets of "negative" (Neg), "zero" (Zero), and "positive" (Pos). By adjusting and , the division and membership degree of each fuzzy set for the temperature change rate can be determined according to actual needs.

[0072] Formulate a fuzzy rule base based on actual operation experience and control objectives; If and , then the duty cycle adjustment amount , indicating that when the temperature deviation is large positive and the temperature change rate is positive, the duty cycle needs to be increased significantly; If and , then the duty cycle adjustment amount , that is, when the temperature deviation is small negative and the temperature change rate is negative, the duty cycle is appropriately reduced.

[0073] Defuzzification (centroid method) converts the fuzzy output into an exact duty cycle adjustment amount through a formula, and the formula is as follows: ; Among them, is the membership degree of the th rule, is the duty cycle adjustment amount corresponding to the th rule, is the total number of rules. Based on the centroid method principle, this formula comprehensively considers the influence of all rules and calculates an exact duty cycle adjustment amount for subsequent adjustment of the PWM duty cycle.

[0074] S342, Energy balance formula and power compensation: According to the law of conservation of energy, considering the season sensitivity coefficient , aging factor and time factor, dynamically adjust the power of the heating tape through a formula, and the formula is as follows: ; Among them, is the base power, and the season sensitivity coefficient It is 0.15 in winter and 0.05 in summer, reflecting the sensitivity to temperature changes in different seasons; is the aging influence coefficient, is the current time, is the initial time. It is used to describe the cumulative effect of equipment aging over time. As time goes by, the influence of aging on power gradually increases.

[0075] Combined with the duty cycle adjustment amount of the fuzzy output and the season weight, optimize the PWM duty cycle through the formula as follows: ; Among them, is the current duty cycle, is the winter weight, is the final response time. In winter, is larger, is smaller, the value of is larger, making the duty cycle adjustment amplitude larger to achieve fast response; in summer, is smaller, is larger, and the duty cycle adjustment amplitude is relatively smaller to ensure system stability.

[0076] S343. Aging and season coupling compensation Dynamic decay factor: Coupling the aging factor with the season parameter, calculate the dynamic decay factor through the following formula: ; Among them, in winter, , the influence of the aging factor is amplified; in summer, , , the influence of the aging factor is weakened, realizing the dynamic adjustment of the aging influence with seasons. Final power correction: Considering the power after thermal balance adjustment, the dynamic decay factor and the real-time temperature deviation , correct the power finally through the following formula: ; S4. Fault self-healing S41. Sensor fault detection and compensation: First, collect the historical temperature data of the sensor within a period of time (such as the past 24 hours) , and calculate its historical temperature standard deviation: ; Among them, is the number of historical temperature data, is the average value of the historical temperature data. This formula is used to measure the dispersion degree of the historical temperature data.

[0077] For the temperature data collected by the sensor array at the current moment ( representing different sensor numbers), calculate the consistency index within the sensor array: ; wherein, represents the maximum value of the temperature difference between any two sensors within the sensor array, is a reference value for normalization. This formula evaluates the consistency of the sensor array by comparing the temperature differences between sensors with the historical temperature fluctuations.

[0078] If the consistency score < 0.8, it indicates that the temperature differences between sensors are too large, exceeding the normal fluctuation range, and an abnormal sensor is determined to exist.

[0079] Assume that the sensor data follows a Gaussian mixture model (GMM), that is, the formula is: ; wherein, is the number of Gaussian distributions in the Gaussian mixture model, is the weight of the th Gaussian distribution, satisfying , is the Gaussian probability density function with a mean of and a variance of .

[0080] Estimate the model parameters , and through the expectation-maximization (EM) algorithm.

[0081] In the E-step of the EM algorithm, calculate the probability that each data point belongs to the th Gaussian distribution as: ; In the M-step, update the parameters as: ; ; ; Repeat the E-step and M-step until the parameters converge.

[0082] Calculate the anomaly probability as: ; wherein, this formula represents the maximum probability that a sensor measurement value does not belong to any Gaussian distribution. If , it indicates that the sensor measurement value has a large difference from the normal data distribution, and mark the sensor as faulty.

[0083] Faulty sensor data is repaired based on SpatioTemporal Kriging interpolation.

[0084] First, determine the set of neighboring sensors , and the neighboring sensors are selected according to spatial distance and time sequence; Let the sensor position be represented in a two-dimensional space . For the spatial distance, calculate the Euclidean distance between the current faulty sensor position and the neighboring sensor position as: ; For the time distance, calculate the difference between the current time and the measurement time of the neighboring sensor as .

[0085] Use the following formula for repair: ; where is the weight calculated according to the spatio-temporal distance between the neighboring sensor and the faulty sensor, and are the spatial and temporal scale parameters used to adjust the influence range of spatio-temporal correlation.

[0086] The weight can be calculated by the inverse distance weighting method, that is, the formula is: ; where this formula uses the spatio-temporal correlation of neighboring sensors to repair the data, making the repaired data more in line with the actual situation.

[0087] S42. Detection and switching of tracing heating cable damage: Analysis of harmonic energy ratio: Perform a fast Fourier transform (FFT) on the tracing heating cable current signal to obtain the frequency domain signal .

[0088] According to the Fourier transform formula: ; It is implemented on a computer through the discrete Fourier transform (DFT), that is: , ; where is the number of sampling points.

[0089] Extract the energy in the 1kHz frequency band: ; Calculate the total energy as: ; where indicates that the frequency band range of the energy signal is 1kHz ± 50Hz, Represents the total energy in the entire frequency domain.

[0090] If , it indicates that the energy in the 1kHz frequency band accounts for a relatively large proportion, exceeding the normal range, and it is determined to be an open circuit.

[0091] Power-temperature dynamic correlation test: The formula for calculating the temperature change is: ; The formula for the power change is: ; Among them, is the time interval (such as 1 minute). Calculate the average value of the temperature change and the average value of the power change .

[0092] Calculate the dynamic correlation coefficient through the following formula: ; If , it indicates a strong negative correlation between power and temperature, that is, the power is normal but the temperature drops abnormally, then it is determined that the tracing tape fails. After determining that the tracing tape fails, start the standby tracing tape.

[0093] Calculate the temperature loss as: ; Among them, is the set target temperature, is the currently actual measured temperature.

[0094] Adjust the power of the standby tracing tape according to the following power compensation formula: ; Among them, is the nominal power of the standby tracing tape, is the compensation gain, generally set to 1.2, is the reference temperature, including the average temperature during normal operation. This formula compensates the power of the standby tracing tape according to the temperature loss to ensure that the system can maintain normal operation.

[0095] S5. Closed-loop model update and optimization S51. Incremental training (online learning): For the predicted temperature of the model and the actually measured temperature , calculate the prediction error as: ; Calculate the variance estimate of the error as: ; Among them, is the number of recent prediction error data, is the average value of the prediction error.

[0096] Define the data weight: ; Among them, the prediction error For large samples, their weights are smaller. In this way, when the model is updated, the influence of noise samples on the model is suppressed, ensuring that the model update is more accurate.

[0097] Suppose the model parameters follow a Gaussian distribution. Initially, the prior distribution of the parameters is set as: ; Among them, is the prior mean, is the prior covariance.

[0098] For the new sample data ( is the input feature matrix, is the corresponding output vector), update according to the following online Bayesian update formula: ; Among them, is the observation noise variance. In this formula, the numerator part represents the correction of the parameter mean by the new data, and the denominator is used to balance the influence of the new data and the prior information. By continuously updating with new data, the LSTM weights are dynamically adjusted to enable the model to adapt to the drift of the data distribution.

[0099] S52, Q-learning reinforcement learning optimization The state space defines the state as: ; Among them, is the temperature deviation, is the temperature change rate, is the aging factor, is the seasonal feature index. These state variables comprehensively reflect the current operating state of the system.

[0100] The action space defines the state as: ; Among them, is the adjustment amount of the tracing cable power, is the adjustment amount of the PWM duty cycle. The action space represents the control operations that the system can take.

[0101] Calculate the total power as: ; Among them, is the current tracing cable power.

[0102] Then the reward function is: ; Among them, , , is a weight coefficient used to balance different optimization objectives. Measures the magnitude of the temperature deviation, The larger it is, the higher the degree of emphasis on temperature stability; Represents the total power consumption of the system, Used to control energy consumption; Measures the change amplitude of the action, Used to avoid frequent system adjustments.

[0103] Update the policy network using the Actor-Critic framework , first, the policy network Outputs the probability distribution of the action according to the current state through a neural network, with its parameters being .

[0104] Calculate the action value function as: ; the state value function is: ; Represents the expected cumulative reward after taking action in state , which can be iteratively calculated through the following Bellman equation: ; where is the discount factor, usually taking a value of 0.9, used to balance the current reward and future rewards.

[0105] Represents the value in state , which can be obtained by taking the weighted average of for all possible actions, that is: ; Update the policy network parameters according to the following policy gradient formula: ; where this formula calculates the gradient of the policy network parameters, enabling the policy network to update in the direction of maximizing the cumulative reward and avoiding local optima.

[0106] S53. Dynamic calibration of the aging factor Assume that the aging factor follows an exponential decay model as: ; where is the initial value of the aging factor, is the aging rate, is the noise term, following a Gaussian distribution with a mean of 0 and a variance of .

[0107] Use the Extended Kalman Filter (EKF) to estimate the aging rate in real time , the specific steps are as follows: First, linearize the aging factor that follows an exponential decay model. Let the state variable , then the state transition equation is: ; the observation equation is: ; Perform a Taylor expansion of the observation equation at the current estimated value to obtain the linearized observation equation as follows: ; where is the estimated value of .

[0108] Calculate the Kalman gain: ; where is the state estimation error covariance, is the Jacobian matrix of the linearized observation equation, is the observation noise covariance, then: ; ; Update the aging rate through the following formula : ; where is dynamically adjusted according to the current state and observation data to achieve real-time tracking of the aging rate.

[0109] Embed the real-time estimated aging factor into the heat transfer model to correct the thermal conductivity.

[0110] Assume the original thermal conductivity is , and obtain the corrected effective thermal conductivity through the following formula: ; Example 2 Combined with Example 1 for further illustration, the embodiments of the present application provide an electronic device. The electronic device may include: A memory, a processor, and a computer program stored in the memory and executable on the processor.

[0111] When the processor executes the program, it implements the intelligent optimization method for the power plant tracing system based on environmental temperature fluctuations provided in the above embodiments.

[0112] Furthermore, the electronic device further includes: A communication interface for communication between the memory and the processor.

[0113] A memory for storing a computer program executable on the processor.

[0114] The memory may include high-speed RAM memory and may also include non-volatile memory, such as at least one disk memory.

[0115] If the memory, processor, and communication interface are implemented independently, the communication interface, memory, and processor can be interconnected via a bus to complete communication with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 4 only a thick line is used to represent it in the figure, but it does not mean that there is only one bus or one type of bus.

[0116] The processor may include one or more processing units. For example, the processor may include an application processor (AP), an Application Specific Integrated Circuit (ASIC), a modem processor, a Central Processing Unit (CPU), an image signal processor (ISP), a controller, a memory, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural-network processing unit (NPU), etc. Among them, different processing units can be independent devices or integrated in one or more processors. Among them, the controller can be the nerve center and command center. The controller can generate operation control signals according to the instruction operation code and timing signals to complete the control of fetching and executing instructions. A memory can also be provided in the processor to store instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can save the instructions or data that the processor has just used or recycled. If the processor needs to use the instruction or data again, it can be directly called from the said memory. This avoids repeated accesses, reduces the waiting time of the processor, and thus improves the efficiency of the system.

[0117] An embodiment of the present application further provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the intelligent optimization method of the power plant tracing system based on ambient temperature fluctuation as described above is implemented.

[0118] An embodiment of the present application further provides a computer program product, the computer program can run computer instructions, and when the computer instructions are executed by a processor, the intelligent optimization method of the power plant tracing system based on ambient temperature fluctuation as described above is implemented.

[0119] The storage medium mentioned above may be a read-only memory, a magnetic disk or an optical disc, etc. Although the embodiments of the present application have been shown and described above, it can be understood that the above embodiments are exemplary and should not be construed as limiting the present application. Those of ordinary skill in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present application.

Claims

1. An intelligent optimization method for a power plant heating system based on ambient temperature fluctuations, characterized in that: The following steps are involved: S1. Data acquisition and preprocessing: Deploy sensor arrays and collect operating data signals of the power plant heating pipelines , for the running data signal Perform preprocessing and output standardized feature vector; S2, model building and parameter optimization: construct a physical-data dual-driven heat transfer model, input standardized feature vectors to optimize initial parameters; S3, Seasonal intelligent control: Build a differentiated parameter library based on seasonal characteristics, and perform intelligent switching of transition seasons based on the optimized heat transfer model; S4. Fault self-healing and model updating: Establish a fault classification response mechanism and build an online evolution framework for the heat transfer model.

2. The intelligent optimization method for the power plant heating system based on ambient temperature fluctuation according to claim 1 is characterized in that: In step S1, the operating data signal The specific steps of preprocessing are: using the 4-layer Coiflet wavelet decomposition and reconstruction algorithm to reconstruct the running data signal De-noising is performed, and the signal after denoising is As input data matrix , RobustPCA is used to separate low-rank normal data and sparse abnormal data, and the spatiotemporal K-nearest neighbor interpolation method is used to repair the data. The repaired data is input into PCA, the principal component Z is calculated and screened, and then weighted fusion is performed through PCA and mutual information method to construct a fusion feature vector ; Wherein, the sensor array includes a temperature sensor, a current sensor, a voltage sensor and a flow sensor; Operation data signal Including pipe temperature , Heating cable power , Ambient temperature , medium flow .

3. The intelligent optimization method for the power plant heating system based on ambient temperature fluctuation according to claim 2 is characterized in that: Step S1 also includes the following steps: The temperature change rate r was calculated based on fitting a cubic spline curve with a 10-second sliding window; set up , , are three change rate thresholds respectively, and ( ), dynamically switch the sampling frequency according to the relationship between the temperature change rate r and the threshold The formula is as follows: ； in, is the collection frequency.

4. The intelligent optimization method for the power plant heating system based on ambient temperature fluctuation according to claim 1 is characterized in that: The specific steps in step S2 are as follows: S21. Physical-data dual-driven heat transfer model construction: A steady-state heat transfer model is constructed based on the Fourier equation, and the nonlinear correction of the medium flow rate is introduced to calculate the steady-state temperature distribution and obtain the basic temperature distribution of the pipeline. , according to LSTM to capture the dynamic characteristics of time series, input the normalized historical temperature , Heating cable power , Ambient temperature , medium flow , aging factors , and updates the hidden state at the current moment through the internal gate control mechanism , the output is mapped to the temperature fluctuation component through the fully connected layer to generate a dynamic correction term , to correct the steady-state temperature of the steady-state heat transfer model; Fourier pipeline foundation temperature distribution Dynamic correction term with LSTM Superposition can be used for temperature prediction; S22. Bayesian hyperparameter optimization: Use Bayesian optimization to adjust the number of LSTM hidden layer nodes , learning rate , regularization coefficient , based on the historical evaluation results, construct the hyperparameter distribution, use the Monte Carlo sampling method to sample the hyperparameter space, calculate the expected improvement value of each sampling point, and use the expected improvement criterion to select the hyperparameter combination with the largest expected improvement value as the hyperparameter for the next iteration; Among them, the optimal hyperparameter combination is: , , ; Adding Gaussian noise to LSTM weights Generate 100 sets of prediction results, calculate the mean and standard deviation based on the prediction results to obtain a 95% confidence interval and evaluate the model uncertainty; S23, Dynamic update of aging compensation factor: Fit aging history data through linear regression and update aging factor incrementally online , the aging factor Embedded temperature prediction formulas compensate for thermal efficiency loss due to equipment aging.

5. The intelligent optimization method for the power plant heating system based on ambient temperature fluctuation according to claim 4 is characterized in that: In step S21, the hidden state is mapped through the fully connected layer To the temperature fluctuation component, dynamic correction term The calculation formula is as follows: ； Fourier Pipe Foundation Temperature Distribution Dynamic correction term with LSTM Temperature prediction of the output after superposition The formula is as follows: ； in, is the weight matrix of the fully connected layer, is the bias vector, is the hidden state of LSTM output; In step S22, the calculation formula for selecting the hyperparameter combination by maximizing the expected value is as follows: ； Based on the mean and standard deviation of the predicted results, the formula for the output confidence interval is as follows: ； Among them, the mean Represents the average level of the prediction results, and the formula is , standard deviation To measure the discreteness of the prediction results, the formula is: ; In step S23, the formula for adjusting the attenuation rate in real time according to the new data is as follows: ； in, is the learning rate, which is used to control the step size of the update, is the currently observed aging factor, is the aging factor predicted by the previous model; The aging factor is embedded in the prediction formula as follows: 。 6. The intelligent optimization method for the power plant heating system based on ambient temperature fluctuation according to claim 1 is characterized in that: In step S3, the specific steps are as follows: S31. Construction of seasonal characteristic parameter library: Collection of historical ambient temperature data , based on historical ambient temperature data The mean and standard deviation , define the reference temperature , the seasonal characteristic index is quantified by the formula , establish the control parameter library corresponding to different seasons according to the quantification results; S32, Transition Season Intelligent Switching: Use the trained LSTM model to input current and recent ambient temperature data and predict the temperature series for the next 72 time steps , calculate the mean of the time series and the predicted temperature series, and calculate the temperature trend slope through a specific formula , set the threshold, when 72 consecutive time steps meet the specific conditions, according to the temperature trend slope Positive and negative triggers seasonal mode switching; S33. Dynamic weight allocation: based on seasonal characteristic index and real-time temperature deviation Dynamically assign winter weights and summer weights , integrating winter and summer control parameters, and obtaining the final adjustment step length and response time ; S34, Predictive Reactive Control: Input variables define real-time temperature deviation and temperature change rate , respectively, using Gaussian and triangular membership functions for fuzzy processing, and mapping the fuzzy input variables to the corresponding duty cycle adjustment , the duty cycle adjustment is obtained by defuzzifying the output using the centroid method , combined with the seasonal sensitivity coefficient , aging factors and time factors, establish a heat balance equation to dynamically adjust the heating cable power, and adjust the duty cycle according to the fuzzy output , seasonal weights and response times Dynamically optimize PWM duty cycle; Aging Factor Calculation of dynamic attenuation factor coupled with seasonal parameters , comprehensive thermal balance power , dynamic attenuation factor and real-time temperature deviation Make a final correction to the power.

7. The intelligent optimization method for the power plant heating system based on ambient temperature fluctuation according to claim 6 is characterized in that: In step S31, the seasonal characteristic index is quantified The formula is as follows: ； in, Indicates the relationship between the dispersion of temperature data and the average level, It indicates that nonlinear adjustment is performed according to the ratio of the average temperature to the reference temperature; The parameter setting formula for storing different seasons in the parameter library is: ； Among them, when When the winter mode is set, set the adjustment step length , response time Second; when When the summer mode is set, set the adjustment step length , response time Second; In step S32, the temperature trend slope The calculation formula is: ； The numerator represents the degree of coordinated change of time and temperature from their respective means, and the denominator is used for normalization. According to the temperature trend slope The positive or negative value of determines the switching seasonal mode formula: ； Among them, the specific conditions are , indicating that the temperature is decreasing and switching to winter mode; when , indicating that the temperature is on the rise, switching to summer mode; In step S33, the winter weight is first calculated: , then calculate the summer weights: ; Among them, the seasonal characteristic index Increase, then the winter weight The trend is upward. Indicates the absolute value of temperature deviation, winter weight Follow increase and grow; Substitute the parameters in the parameter library to set the adjustment step size and response time , then the final adjustment step size and response time The formula is as follows: ； ； In step S34, the duty cycle adjustment value is adjusted according to the fuzzy output , seasonal weights and response times The formula for dynamically optimizing the PWM duty cycle is as follows: ； in, is the current duty cycle, is the winter weight, is the final response time; Power after comprehensive heat balance adjustment , Dynamic attenuation factor and real-time temperature deviation , the power is finally corrected by the following formula: 。 8. The intelligent optimization method for the power plant heating system based on ambient temperature fluctuation according to claim 1 is characterized in that: In step S4, the specific steps are as follows: S41, sensor fault detection and compensation: obtain sensor historical temperature data , and calculate the historical temperature standard deviation ,The Bayesian anomaly probability calculation is triggered by the redundant consistency score to locate the faulty sensor, and the faulty sensor data is repaired according to the spatiotemporal Kriging interpolation method; S42, Heating cable damage detection and switching: drive the standby system power adjustment by determining the abnormal proportion of harmonic energy; Among them, the abnormal harmonic energy ratio includes the abnormal energy ratio of the characteristic frequency band of the current signal detected by frequency domain analysis and the abnormal correlation between power input and temperature response detected by covariance analysis; S43, model update incremental training: assign data weights according to prediction errors, and use the online Bayesian update formula to adjust LSTM weights; S44, Q-learning reinforcement learning optimization: constructing state-action space and reward function, and updating policy network using Actor-Critic framework parameter; S45. Dynamic calibration of aging factors: Establishing an exponential decay model for aging factors , set the state variables, derive the state transfer equation, and construct the observation equation. According to the observation equation, the linearized observation equation is obtained, and the aging rate is estimated in real time using the extended Kalman filter. , the estimated aging factor Corrected thermal conductivity by embedding heat transfer model .

9. The intelligent optimization method for the power plant heating system based on ambient temperature fluctuation according to claim 8 is characterized in that: In step S41, the redundancy consistency score formula within the sensor array is: ； in, is the historical temperature standard deviation, Indicates the maximum value of the temperature difference between any two sensors in the sensor array. It is represented as a reference value and used for normalization; If the consistency score is <0.8, it is determined that there is an abnormal sensor; When there is an abnormal sensor, the Bayesian abnormal probability calculation is triggered and expressed as the following formula: ； in, It is The weights of a Gaussian distribution, is the mean, is the variance; like , then mark the sensor For failure; The formula for repairing faulty sensor data using the spatiotemporal kriging interpolation method is as follows: ； in, Current fault sensor location With proximity sensor location The Euclidean distance of For current time Measuring time with proximity sensors The difference, is the weight calculated based on the temporal and spatial distance between the neighboring sensors and the faulty sensor, and are the spatial and temporal scale parameters; In step S42, the specific calculation steps of the frequency domain analysis to detect the abnormal energy proportion of the characteristic frequency band of the current signal are as follows: Current signal for heating cable Perform fast Fourier transform to obtain frequency domain signal , the formula for extracting 1kHz frequency band energy is: ; Calculate the total energy as: ; in, The frequency range of the energy signal is 1kHz±50Hz. Expressed as the total energy in the entire frequency domain; like , then the energy ratio abnormality is determined to be a circuit breaker; The specific calculation steps for covariance analysis to detect abnormal correlation between power input and temperature response are as follows: ； in, is the temperature change, is the power change, is the temperature change The average value of Power variation The average value of like , then the abnormal energy ratio is determined to be the failure of the heating belt; When it is determined that the energy ratio is abnormal, the formula for adjusting the standby heating cable power is as follows according to the power compensation formula: ； in, is the nominal power of the standby heating cable, is the compensation gain, is the reference temperature.

10. The intelligent optimization method for a power plant heating system based on ambient temperature fluctuation according to claim 8, characterized in that: In step S43, the model predicted temperature is obtained and the actual measured temperature , the temperature is predicted according to the model and the actual measured temperature Calculating prediction error , and according to the prediction error Compute an estimate of the variance of the error , through the prediction error and variance estimates Defining data weights , the formula is as follows: ； Set model parameters Obey the Gaussian distribution and then verify the Bayesian update formula for update: ； in, is the observation noise variance, the numerator Indicates the correction of the parameter mean by the new data, the denominator Used to balance the impact of new data and prior information; In step S44, the state space defines the state as: ; in, is the temperature deviation, is the rate of temperature change, It is the aging factor. is the seasonal characteristic index; The action space defines the state as: ; in, is the heating cable power adjustment, is the PWM duty cycle adjustment; Calculating the total power, the reward function is: ； in, , , is the weight coefficient, Measure the magnitude of temperature deviation, represents the total power consumption of the system, To control energy consumption, Measuring the range of motion, Used to avoid frequent system adjustments; Updating the policy network using the Actor-Critic framework , policy network According to the current status Output Action The probability distribution of , and update the policy network according to the following policy gradient formula parameter: ； in, is the action value function, is the state value function; In step S45, an exponential decay model of the aging factor is established. The formula is: ； in, is the initial value of the aging factor, is the aging rate, is the noise term; Extended Kalman Filter for Real-time Estimation of Aging Rate The formula is: ； in, is the Kalman gain, which is dynamically adjusted according to the current state and observation data; The aging factor estimated in real time Corrected thermal conductivity by embedding heat transfer model The formula is: 。

Images