Intelligent water quality prediction and regulation method in descaling process of thermal power plant

Abstract

The invention provides an intelligent water quality prediction, regulation and control method in a thermal power plant descaling process, and relates to the technical field of water quality prediction.The method comprises the specific steps that data recorded in the chemical cleaning and descaling process of a thermal power plant boiler is collected and preprocessed, and a thermal power plant descaling water quality data set is constructed; introducing a local fluctuation suppression weight, a water temperature deviation adjustment weight, a medicament mutation suppression factor and a time attenuation weight to generate an enhanced delay vector; constructing a global state vector through covariable weighting and direction perception expression; and forming prediction input by combining a global state vector, a dosing variable quantity and temperature deviation, optimizing by adopting heterovariance robust task loss based on Huber loss and an uncertainty factor, and finally realizing multi-step prediction of the pH value, the Langier saturation index, the calcium ion concentration, the silicate concentration, the conductivity and the turbidity in the future 15 minutes.

Description

Technical Field

[0001] This invention relates to the field of water quality prediction technology, and in particular to an intelligent water quality prediction and control method in the descaling process of thermal power plants. Background Technology

[0002] During the long-term operation of thermal power plants, boilers and pipelines are prone to scale formation due to the deposition of salts, particles, and metal ions in the water. This leads to decreased heat exchange efficiency, increased energy consumption, and risks of overheating or corrosion. To ensure safe operation, regular descaling is necessary. However, relying solely on manual sampling is inefficient and infrequent, which can result in incomplete descaling or excessive use of chemicals, increasing economic and environmental burdens. Therefore, there is an urgent need to develop an intelligent water quality prediction and control method that integrates real-time monitoring, predictive modeling, and automatic regulation to achieve advance prediction and precise dosing, thereby improving both economic efficiency and safety.

[0003] Publication No. CN104318325A proposes a nonlinear autoregressive exogenous input model optimized by a genetic algorithm. In real-time monitoring, the system detects whether the data exceeds the standard. If it is a sudden pollution event, the improved dynamic time warping algorithm is used to match historical pollution patterns and update the model. Finally, the system predicts future water quality changes based on real-time data and the model. Publication No. CN110308705A proposes a model trained using a random forest classification algorithm. By inputting test samples into the trained model, the predicted water quality value can be obtained. Based on the prediction results, the relevant equipment can be automatically controlled to realize water quality prediction based on big data and artificial intelligence.

[0004] However, existing technologies have the following shortcomings in predicting water quality during descaling in thermal power plants: First, data acquisition dimensions are limited, lacking multi-node and multi-parameter system fusion, and preprocessing methods are insufficient, making it difficult to ensure the integrity and stability of input data; second, time-series modeling fails to effectively distinguish key influencing factors such as local fluctuations, temperature deviations, and chemical mutations, and the physicochemical coupling relationships between variables are not fully utilized; third, there is a lack of identification and suppression mechanisms for abnormal situations such as sudden chemical addition and disturbances in cleaning conditions, resulting in insufficient model stability under complex operating conditions; finally, existing prediction models generally adopt simple linear modeling and a single error function, without considering heteroscedasticity and adaptive adjustment of uncertainty, resulting in insufficient prediction accuracy and robustness, making it difficult to meet the actual needs of dynamic water quality prediction and control in the descaling process of thermal power plants. Summary of the Invention

[0005] This invention proposes a method for predicting and controlling the descaling water quality in thermal power plants, aiming to solve the problems of single data acquisition, insufficient utilization of variable correlation, poor adaptability to abnormal operating conditions, and low prediction accuracy in existing technologies. First, the collected multi-source water quality monitoring data are standardized, missing values ​​are imputed, and anomalies are removed to construct a complete dataset. Then, local fluctuation suppression weights, water temperature deviation adjustment weights, reagent mutation suppression factors, and time decay weights are introduced to correct historical data point by point, resulting in an enhanced delay vector. In the covariate modeling stage, coupling weights are constructed based on variable correlations, and combined with local rate of change factors and cleaning condition adjustment factors to form a covariate weighted matrix, highlighting the dynamic coupling relationship between variables. Furthermore, the matrix is ​​divided into local sub-blocks, and a direction-aware expression is obtained through joint weighting of direction-aware weights and anomaly-sensitive factors, thereby enhancing the model's ability to distinguish different time and variable dimensions. In the prediction stage, local mean, slope, and fluctuation variance are calculated, and a global state vector is generated by combining anomaly weights. This vector is then concatenated with exogenous process quantities such as dosage changes and temperature deviations to form the prediction input, enabling multi-step prediction of pH, Langelinie saturation index, calcium ion concentration, silicate ion concentration, conductivity, and turbidity for the next 15 minutes. Finally, during training, a heteroscedastic robust task loss based on Huber loss and uncertainty factors is introduced, combined with hyperparameter setting and iterative optimization to achieve stable model convergence. This invention can effectively improve the accuracy and robustness of descaling water quality prediction in thermal power plants, providing a reliable basis for dynamic control.

[0006] A method for intelligent water quality prediction and control during the descaling process in thermal power plants, the specific method being as follows: S1. Obtain the data recorded during the chemical cleaning and descaling process of the boiler in the thermal power plant, and preprocess the original thermal power plant descaling water quality dataset to construct the thermal power plant descaling water quality dataset. S2. Based on the scale removal water quality dataset of thermal power plants, the local fluctuation suppression weight is obtained by calculating the local variance, the water temperature deviation adjustment weight is obtained by the water temperature difference, the chemical mutation suppression factor is obtained by the dosage at adjacent time points, and the time decay weight is obtained by calculating the decay coefficient. The factors are fused point by point to form an enhanced delay vector. S3. The enhanced delay vector correlation calculation yields the covariate coupling weights, which are then used to weight the water quality variables and construct a covariate fusion vector. The local rate of change factor for water quality changes at adjacent time points is calculated, and the operating condition adjustment factor is obtained by combining the cleaning conditions to form a covariate weighting matrix. S4. Divide the covariate weighting matrix into local sub-blocks by row and width, obtain the time direction weight based on the difference of the mean of adjacent time slices, obtain the variable perception weight based on the difference of the variable mean, obtain the anomaly sensitivity factor based on the distribution difference, and jointly weight it with the local sub-blocks to obtain the direction perception expression. S5. Calculate the mean, trend slope and fluctuation variance based on the direction-aware expression, construct a global state vector with the abnormal weights obtained from the abnormality sensitivity factor, and concatenate it with the drug dosage change and temperature deviation to form an autoregressive input. The predicted output result is obtained through a linear vector. S6. Construct a prediction model for descaling water quality in thermal power plants. Input the descaling water quality dataset of thermal power plants, use joint loss as the loss function, and sequentially go through steps S2 to S5. Then, iterate and train until convergence to complete the training of the prediction model for descaling water quality in thermal power plants. S7. Adjust the type and dosage of added chemicals based on the prediction output of the thermal power plant descaling water quality prediction model.

[0007] Preferably, for the construction of the thermal power plant boiler and chemical cleaning descaling system descaling water quality dataset, online monitoring points are set up at key nodes such as the makeup water outlet, after condensate polishing, boiler feed water inlet, boiler blowdown branch, before and after the circulating water heat exchanger, and the inlet and outlet of the chemical cleaning loop, and the data is verified by manual analysis. Multi-source water quality parameters include pH, redox potential, conductivity, turbidity, calcium ion concentration, magnesium ion concentration, dissolved oxygen concentration, silicate concentration, total hardness, chloride ion concentration, sulfate concentration, temperature, flow rate, pressure, corrosion resistance, and electrochemical noise signal. The original water quality data underwent unit conversion, outlier removal, missing value imputation, and feature scaling and standardization to construct a thermal power plant descaling water quality dataset. After unifying the units of the collected data, outliers were removed using boundary rules and Hampel filtering, short-term missing values ​​were imputed using interpolation, and all features were standardized. Finally, samples were constructed based on a 60-minute sliding window to generate predictive labels for the main water quality indicators for the next 15 minutes, resulting in a thermal power plant descaling water quality dataset for intelligent water quality prediction and control models.

[0008] First, the local variance of each time point is calculated based on the water quality change amplitude at each time point and the two consecutive time points before and after it. Then, the local fluctuation suppression weight is constructed by normalizing the local variance by taking its reciprocal. By combining the difference between the current circulating water temperature and the optimal descaling temperature of the power plant pipeline at the sampling moment, the water temperature deviation adjustment weight is calculated. The water temperature deviation adjustment weight is larger when the temperature is close to the optimal value. By comparing the change in drug dosage between the current time and the previous time, a drug mutation inhibition factor is obtained, which reduces the impact of the current time slice when a sudden change occurs in the drug dosage. Calculate the time decay weight based on the time interval between each historical sample and the current moment; The enhanced historical sample at each time point is obtained by multiplying the local fluctuation suppression weight, water temperature deviation adjustment weight, drug mutation suppression factor, time decay weight, and the historical water quality value at the corresponding time. All enhanced historical samples are then arranged in chronological order to construct an enhanced delay vector.

[0009] Preferably, in step S2, the local variance of each time point is calculated based on the water quality change amplitude at each time point and the two consecutive time points before and after it. The reciprocal of the sum of the local variance and 1 is used as the local fluctuation suppression weight, making samples with smaller historical fluctuations more reliable in subsequent processing, thus helping to remove local abnormal peaks caused by load disturbances. Subsequently, the water temperature deviation adjustment weight is calculated by combining the difference between the circulating water temperature at the current sampling time and the optimal descaling temperature of the thermal power plant pipeline. The weight corresponding to the temperature being close to the optimal value is greater, thereby increasing the contribution of high-temperature sensitive sections in historical samples and helping to highlight the data characteristics of scale-sensitive areas. By comparing the reagent dosage between the current time and the previous time, The magnitude of dosage changes is used to obtain the drug mutation inhibition factor. Sudden changes in drug dosage automatically reduce the effectiveness of the current time slice to avoid interference with prediction results and enhance the model's ability to learn from normal water quality evolution processes. Finally, based on the time interval between each historical sample and the current moment, a time decay weight is calculated, and the proportion of each sample is estimated based on the time decay weight, thus ensuring that subsequent models pay more attention to recent historical information that has a direct impact on the current operating state. By fusing the local fluctuation suppression weight, water temperature deviation adjustment weight, drug mutation inhibition factor, and time decay weight, an enhanced historical sample is obtained for each time point, and all enhanced historical samples are arranged in chronological order to form an enhanced delay vector.

[0010] Preferably, in step S3, the correlation between the enhanced delay vector and the enhanced delay vector of the candidate variable is calculated to obtain the degree of coupling between the variables. The correlation value is then normalized in the candidate set to form a covariate coupling weight representing the coupling strength between different variables. At each time slice, the three covariates with the highest correlation to the target variable are selected. The enhanced delayed samples of the variables and the enhanced delayed samples of the target variable are weighted together and combined. The resulting coupling weights are used as superposition coefficients to form a covariate fusion vector from the delayed samples of each variable. The local rate of change factor is obtained by measuring the magnitude of water quality changes at adjacent times. The local rate of change factor reflects the change factor of fluctuation characteristics in a short period of time, and strengthens the characteristic influence in the range of violent fluctuations. The operating conditions recorded during the operation of the thermal power plant pipeline are binarized and labeled to distinguish between normal operation and cleaning stages. The operating condition adjustment factor is obtained through normalization calculation. Abnormal fluctuations in reagent dosage and circulation flow rate are detected. The operating condition adjustment factor reduces the weight of the current sample in subsequent processing. The covariate fusion vector, local rate of change factor, and operating condition adjustment factor are calculated by dot product and arranged sequentially to form a covariate weighted matrix, thereby achieving joint modeling of multivariate time series features.

[0011] Furthermore, the covariate coupling weights are first obtained by calculating the correlation between different water quality indicators, making each feature data dependent and thus truly reflecting the multidimensional coupling relationship of circulating water quality in thermal power plants. Then, based on the covariate coupling weights, the multivariate features are weighted and combined to obtain a covariate fusion vector, which is used to integrate multi-source indicator information and avoid prediction bias caused by fluctuations in a single indicator. A local rate of change factor is obtained by comparing the magnitude of variable changes at adjacent times, used to emphasize the short-term abrupt changes in pH, conductivity, or calcium ion concentration after reagent addition, enabling the model to keenly capture abnormal water quality changes. Simultaneously, an operating condition adjustment factor is introduced, combining the abrupt changes in reagent dosage and circulating water flow rate, to weaken samples at abnormal times and reduce interference from extreme operating conditions. Finally, the covariate fusion vector, rate of change factor, and operating condition adjustment factor are expanded within a time sliding window to form a covariate weighting matrix, thus taking into account both time evolution characteristics and operating condition constraints, providing a comprehensive and stable input basis for subsequent water quality prediction and control.

[0012] Preferably, in step S4, the starting row and starting column of the window are used as indices, and the covariate weighting matrix is ​​divided into multiple local sub-blocks by setting the row height and column width of the sliding window. Each sub-block corresponds to a fixed time window and variable combination, resulting in multiple local sub-blocks. The differences between water quality monitoring values ​​of adjacent time slices in local sub-blocks are calculated separately. The differences at the same time slice location are summed and averaged to obtain the overall change corresponding to the current time slice. The overall change index of continuous time slices is differentially calculated to obtain the time change amplitude that can reflect the degree of difference between adjacent time slices. The obtained time variation amplitude is combined with the preset scaling factor, and then the intermediate quantity corresponding to each time slice is converted into a standardized weight coefficient through normalization. The normalized result is then embedded into the time dimension in the order of row index to complete the construction of the time direction weight. In the local sub-block, determine the column index where the target variable is located. For each other column variable, calculate the difference between the value of the column variable in each time row and the value of the target variable in the same row. Sum the difference results of all time rows and take the average to obtain the difference degree of the current column variable relative to the target variable. Use the average difference degree as the variable difference degree of the current column variable. The product of the difference result of each column variable in the local sub-block and the set variable direction scaling factor is normalized. The difference result is transformed by the exponential mapping function to obtain the corresponding mapping value. The mapping values ​​of all column variables are summed and used as the denominator of the normalization. The mapping value of each column variable is divided by the sum of the mapping values ​​of all column variables to obtain the variable perception weight of the current column variable. By calculating the KL divergence between the water quality distribution within each local sub-block and the water quality distribution during the reference stable period, and multiplying it with the anomaly amplification factor, the anomaly sensitivity factor is obtained, which amplifies the water quality changes under abnormal operating conditions. A direction-aware representation is constructed by performing a dot product operation on the local sub-block data and the calculated time direction weights, variable direction weights, and anomaly sensitivity factors.

[0013] Furthermore, firstly, by setting a fixed sliding window, the covariate weighted matrix is ​​divided into multiple local sub-blocks according to the time and variable dimensions. Each sub-block corresponds to a fixed time window and variable combination. This operation decomposes the original matrix into multiple local time-variable combinations, enabling the processing of fine-grained features of thermal power plant water quality data, capturing changes in specific water quality characteristics at specific time periods, and avoiding over-smoothing and information loss. Within each local sub-block, the direction-aware weights of time and variable directions are derived by calculating the differences between time and variables. The time direction weights reflect the impact of different time segments on the target water quality changes, while the variable direction weights emphasize the degree of coupling between different water quality variables. The calculation process enhances the time-awareness of the variables. The relative importance of variables is considered to ensure that the model pays sufficient attention to the key drivers of water quality changes. The difference between the water quality distribution of a local sub-block and the water quality distribution during the reference stable period is calculated, and the difference is used to generate anomaly sensitivity factors. Specifically, if the difference between the current water quality pattern and the normal water quality distribution is large, the anomaly sensitivity factor will be amplified, increasing the weight of the local sub-block in the model, and the model can provide sensitivity under abnormal conditions. The data of each local sub-block is multiplied by the calculated direction-aware weights and anomaly sensitivity factors to construct a direction-aware representation. Through this weighting method, the model can comprehensively consider time evolution, interactions between variables, and abnormal changes, generating a richer representation of water quality change characteristics.

[0014] Preferably, in step S5, the global state vector is obtained by calculating the time slice mean, trend slope and fluctuation variance of the direction-aware expression, and then performing linear aggregation on the abnormal weights obtained by normalizing and weighting the abnormality sensitivity factor. The dosage change is obtained by calculating the difference between the dosage at the previous moment and the dosage at the current moment. The temperature deviation is obtained by calculating the difference between the current temperature and the optimal descaling temperature. The global state vector obtained in the previous step is concatenated with the dosage change and temperature deviation to form an autoregressive input. Linear calculation is then performed to obtain the multi-index prediction results for the next 15 minutes for the descaling scenario.

[0015] Furthermore, firstly, statistical calculations are performed on the direction-aware expression to obtain the mean, trend slope, and variance of each time slice. These three indicators reflect the stability level, response trend, and disturbance intensity of water quality, respectively. Subsequently, the statistical values ​​of the mean, trend slope, and variance of the time slices are normalized and weighted by combining anomaly sensitivity factors, highlighting the weight values ​​of periods with high influence. The global state vector at the current moment is obtained by weighted summation. The global state vector is used for subsequent water quality prediction, comprehensively reflecting the evolution of the target water quality. The obtained global state vector is concatenated with the dosage change and temperature deviation to form an autoregressive input vector. The autoregressive input vector is used for multi-step prediction through a linear regression model to output the predicted water quality indicators for the next 15 minutes. The predicted indicators include: pH value, Langelinie saturation index, calcium ion concentration, silicate ion concentration, conductivity, and turbidity. This step can accurately predict water quality changes and provide a decision-making basis for the control during the descaling process.

[0016] Preferably, in step S6, based on the robust loss of the Huber function, the Huber function calculates the difference between the predicted value and the true value. If the absolute value of the difference is less than the threshold, the square of the difference between the predicted value and the true value is used as the loss value; if the absolute value of the difference is greater than or equal to the threshold, the absolute value of the difference between the predicted value and the true value minus the threshold is multiplied by the threshold as the loss value. A dynamic uncertainty factor is introduced for each prediction target, and its value is the standard deviation between the true value and the predicted value. The uncertainty factor is combined with the robust loss of the Huber function to form the heteroscedasticity robust task loss. By setting hyperparameters and training the model using a dataset of descaling water quality from thermal power plants, the training process for the descaling water quality prediction model is finally completed.

[0017] Furthermore, a heteroscedasticity robust task loss function is combined with Huber loss and uncertainty factor adjustment to address complex prediction errors and optimize model performance. First, the number of prediction steps H and the target water quality index set K are set and used as the core structure of the model. When calculating the loss, a weighted adjustment is first introduced based on the time decay factor. Then, the prediction error of each target index is processed using Huber loss. This loss function uses squared error when the error is small, and switches to linear loss when the error is large, ensuring that the model can effectively handle outlier data and avoid oversensitivity to large errors. To adaptively adjust the model's weights for each prediction target, an uncertainty factor is introduced. The uncertainty factor is dynamically adjusted based on the prediction error of each target index. By introducing the uncertainty factor, the model can automatically adjust the learning process when facing predictions with high uncertainty, thereby improving overall prediction accuracy and robustness. Finally, the model is trained by setting hyperparameters and a joint loss function, and optimized using the Adam optimizer, thus completing the training process of the descaling water quality prediction model.

[0018] The technical effects and advantages provided by the present invention in the above technical solution are as follows:

[0019] This application introduces a multi-factor correction mechanism, including local fluctuation suppression, water temperature deviation adjustment, reagent mutation suppression, and time decay, to perform point-by-point correction on historical water quality data and construct a stable enhanced delay vector. This effectively reduces the interference of sudden reagent addition, water temperature fluctuation, and sensor anomalies on data modeling, achieving a smoother and more reliable expression of time-series features. This design breaks through the limitations of existing methods that directly rely on the original sampling sequence, not only improving the stability and robustness of the input data, but also providing a more realistic physical evolution basis for subsequent feature modeling.

[0020] This application establishes a covariate weighted matrix by combining the effects of covariate coupling weights, local rate of change factors, and cleaning condition adjustment factors. Furthermore, it combines direction-aware weights and anomaly sensitivity factors to form a direction-aware expression, which can explicitly express the physicochemical coupling relationship between multiple variables and highlight the key features under time-series abrupt changes and abnormal conditions. This not only enhances the global adaptability of the prediction model under complex conditions but also improves the sensitivity and accuracy to water quality changes across variables and time periods, effectively making up for the shortcomings of existing methods in variable correlation and anomaly detection.

[0021] In the prediction output and model training stages, this application combines the global state vector with exogenous process quantities such as dosage changes and temperature deviations to achieve direct prediction of multiple indicators for the next 15 minutes. It also uses a heteroscedastic robust task loss based on Huber loss and dynamic uncertainty factor for optimization. This method not only ensures the interpretability and causal consistency of the prediction results, but also dynamically adjusts the training weights of each target indicator, reducing the impact of outliers and high uncertainty data on the model. This improves the training convergence speed and prediction accuracy across operating conditions, ultimately providing a reliable basis for reagent dosing and process control in the descaling process of thermal power plants. Attached Figure Description

[0022] Figure 1 This is a flowchart of an intelligent water quality prediction and control method for the descaling process in a thermal power plant, provided by the present invention.

[0023] Figure 2 This is a structural diagram of constructing an enhanced delay vector provided by the present invention.

[0024] Figure 3 This is a structural diagram of constructing a covariate weighted matrix provided by the present invention.

[0025] Figure 4 This is a structural diagram of the construction direction-aware expression provided by the present invention.

[0026] Figure 5 This is a structural diagram of the predicted output results of descaling water quality in thermal power plants provided by the present invention.

[0027] Figure 6 This is a training loss diagram of the descaling water quality prediction model for thermal power plants provided by this invention.

[0028] Figure 7 This is a diagram showing the difference between the predicted and actual values ​​after dimensionality reduction provided by this invention.

[0029] Figure 8 This is a pH value and conductivity prediction graph provided by the present invention.

[0030] Figure 9 This is a calcium ion concentration and turbidity prediction chart provided by the present invention.

[0031] Figure 10 This is a graph showing the prediction of silicate concentration and Langelinie saturation index provided by the present invention. Detailed Implementation

[0032] This invention proposes an intelligent water quality prediction and control method for the descaling process in thermal power plants, aiming to solve the problems of single data acquisition, insufficient utilization of variable correlation, poor adaptability to abnormal operating conditions, and low prediction accuracy in existing technologies. The method first standardizes, imputes missing values, and removes anomalies from the collected multi-source water quality monitoring data to construct a thermal power plant descaling water quality dataset. Then, it introduces local fluctuation suppression weights, water temperature deviation adjustment weights, reagent mutation suppression factors, and time decay weights to correct historical data point by point, obtaining an enhanced delay vector. In the covariate modeling stage, coupling weights are constructed based on variable correlations and combined with local change rate factors and cleaning condition adjustment factors to form a covariate weighted matrix to highlight the dynamic coupling relationship between variables. Furthermore, the matrix is ​​divided into local sub-blocks... By jointly weighting the direction-aware weights and the anomaly sensitivity factor, a direction-aware expression is obtained, thereby enhancing the model's ability to distinguish different time and variable dimensions. In the prediction stage, the local mean, slope, and fluctuation variance are calculated, and a global state vector is generated by combining the anomaly weights. This vector is then concatenated with the dosage change and temperature deviation to form the prediction input, enabling multi-step prediction of pH, Langelinie saturation index, calcium ion concentration, silicate ion concentration, conductivity, and turbidity for the next 15 minutes. Finally, during the training process, a heteroscedasticity robust task loss based on Huber loss and uncertainty factor is introduced. Combined with hyperparameter setting and iterative optimization, the model achieves stable convergence. This invention can effectively improve the accuracy and robustness of scale removal water quality prediction in thermal power plants, providing a reliable basis for dynamic control.

[0033] Please see Figure 1 As shown in the embodiment of this application, a method for intelligent water quality prediction and control in the descaling process of a thermal power plant includes the following specific steps.

[0034] S1. Obtain the data recorded during the chemical cleaning and descaling process of the boiler in the thermal power plant, and preprocess the original thermal power plant descaling water quality dataset to construct the power plant descaling water quality dataset.

[0035] In the boiler and chemical cleaning and descaling system of thermal power plants, multiple online monitoring points are deployed at the makeup water outlet, after condensate polishing, boiler feedwater inlet, boiler blowdown branch, front end of circulating water heat exchanger, rear end of circulating water heat exchanger, chemical cleaning loop inlet, and chemical cleaning loop outlet. These monitoring points are equipped with pH sensors, oxidation-reduction potential sensors, conductivity sensors, turbidity sensors, calcium ion sensors, magnesium ion sensors, dissolved oxygen sensors, silicate analyzers, total hardness analyzers, chloride ion analyzers, sulfate analyzers, temperature sensors, flow sensors, pressure sensors, corrosion resistance probes, and electrical... A chemical noise monitoring device is used to achieve real-time acquisition of multi-source water quality parameters and operating parameters. The sampling frequency is set to once per minute during normal operation and once every 10 seconds during the chemical dosing and cleaning transition phases. Manual timed sampling is also performed. The online monitoring data is verified using ion chromatography, inductively coupled plasma optical emission spectrometry, titration analysis, and spectrophotometry. The acquired data undergoes unit conversion, unifying hardness, calcium ion, magnesium ion, and silicate concentrations to calcium carbonate equivalents in milligrams per liter. Conductivity is converted to 25 degrees Celsius using a temperature coefficient of 0.019 per degree Celsius. Baseline values; Dissolved oxygen values ​​are converted to a 25°C baseline value; In the outlier removal process, invalid data is first removed based on physical boundary conditions, then Hampel filtering with a window length of 21 points and a threshold of 3 times the standard deviation is used to remove pulse-type discrete values, and the first-order difference between adjacent sampling points is calculated. Outlier marking and local smoothing are performed based on the set rate change upper limit; In the missing value handling process, cubic spline interpolation is used to fill in continuous gaps of no more than 3 minutes, and exponentially weighted moving average is used to fill in continuous gaps of more than 3 minutes but no more than 30 minutes. The half-life is set to 10 minutes, and continuous gaps exceeding 3 minutes are considered missing. Sequences with a duration of 0 minutes are directly discarded. In the feature scaling stage, robust scaling is first performed using the median and interquartile range, then standardized using the mean and standard deviation. Finally, all continuous features are linearly mapped to a numerical range of -1 to 1. Ultimately, multi-point time series samples are constructed using a fixed sliding time window with a historical window length of 60 minutes and a step size of 5 minutes. These samples are aligned with the time of reagent addition and generate prediction labels for pH, conductivity, calcium ion concentration, turbidity, silicate concentration, and Langelinie saturation index for the next 15 minutes, forming a thermal power plant descaling water quality dataset for intelligent water quality prediction and control models.

[0036] In this embodiment, multi-point time series samples are constructed according to a fixed sliding time window, and the historical window length is... The sampling data is divided into time segments with a time interval of 60 minutes and a step size of 5 minutes. The mathematical model is as follows: ; in, For the first Water quality observations at various times, totaling 16 variables. For A historical window that ends at a specific moment; Generate predictive labels for pH, conductivity, calcium ion concentration, turbidity, silicate concentration, and Langelier saturation index for the next 15 minutes. For each sample, using the drug administration time as the anchor point, roll forward 15 minutes to extract target variables and construct supervisory labels. The mathematical model is as follows: All target variables are minute-by-minute curves predicting the next 15 minutes.

[0037] S2. Based on the scale removal water quality dataset of thermal power plants, the local fluctuation suppression weight is obtained by calculating the local variance, the water temperature deviation adjustment weight is obtained by the water temperature difference, the chemical mutation suppression factor is obtained by the dosage at adjacent time points, the time decay weight is obtained by calculating the decay coefficient, and the factors are fused point by point to form an enhanced delay vector.

[0038] Furthermore, in step S2, the enhanced delay vector is constructed, and the process is as follows: Figure 2 As shown, the specific steps for constructing the enhanced delay vector are as follows.

[0039] S21. First, the local variance of each time point is calculated based on the water quality change amplitude of each time point and the two consecutive time points before and after it. Then, the local fluctuation suppression weight is constructed by normalizing the local variance by taking its reciprocal. In this embodiment, to reduce the impact of instantaneous fluctuations in circulating water during high-load change phases on historical delayed inputs, the local variance is first calculated based on the water quality values ​​at the current time point and two consecutive time points before and after it. The reciprocal of the sum of the local variance and 1 is used as the local fluctuation suppression weight, giving a larger weight to time periods with smaller fluctuations to improve the reliability of stable samples. The mathematical model for the local fluctuation suppression weight is as follows: ; in, This is the relative position index of the variable within the historical time window. For the first Variables in The value at time, The sliding window length for the local variance is initially set to 2, and is used to control the range of data involved in the calculation of local fluctuations. Let Variance function represent the degree of local fluctuation. For the first in the sample The index in the index The weighting of local fluctuations in the data is adjusted accordingly; the smaller the fluctuation, the lower the weighting. The larger the value, the better it helps stabilize data. S22. Calculate the water temperature deviation adjustment weight by combining the difference between the current circulating water temperature and the optimal descaling temperature of the thermal power plant pipeline. The water temperature deviation adjustment weight is larger when the temperature is close to the optimal value. In this embodiment, scaling is most likely to occur in the circulating water of thermal power plants at around 55°C. The present invention further calculates the water temperature deviation weight based on the difference between the real-time water temperature and this optimal temperature. The mathematical model is as follows: ; in, For circulating water in The real-time temperature is obtained from on-site sensors. The optimal descaling temperature is 55℃. For the first in the sample The index in the index The water temperature deviation of each variable has a higher adjustment weight; the closer the water temperature is to 55℃, the higher the weight. S23. By comparing the change in drug dosage between the current time and the previous time, a drug mutation inhibition factor is obtained, which reduces the influence of the current time slice when a sudden change occurs in the drug dosage. In this embodiment, to reduce the unstable bias caused by instantaneous drug administration to historical input, a drug mutation inhibition factor is calculated based on the magnitude of the change in drug dosage between the current and previous times. The mathematical model is as follows: ; in, for Time of the first The dosage of the drug corresponding to each variable. This is the drug mutation penalty coefficient, initially set to 0.45. For the first in the sample The index in the index The drug mutation inhibitory factor for each variable has a rapidly decreasing weight when the dosage changes drastically, in order to avoid interference with the spread. S24. Calculate the time decay weight based on the time interval between each historical sample and the current time. In this embodiment, to highlight the weight of water quality changes at similar times in the prediction, a weight that gradually decreases over time is constructed based on the time interval between the current historical sample and the time of drug administration. The closer the time is to the current time, the greater the weight of the sample, thereby increasing its influence in the model input. The specific method for the time decay weight is as follows: the decay coefficient that controls the decay rate of the weight over time is multiplied by the time index of the current historical sample, and then an exponential operation is performed to obtain the time decay weight. The initial value of the attenuation coefficient is set to 0.09; S25. The enhanced historical sample at each time point is obtained by multiplying the local fluctuation suppression weight, water temperature deviation adjustment weight, drug mutation suppression factor, time decay weight and the historical water quality value at the corresponding time. All enhanced historical samples are arranged in chronological order to construct the enhanced delay vector. In this embodiment, the local fluctuation suppression weight, water temperature deviation adjustment weight, drug mutation inhibition factor, and time decay weight are multiplied point by point, and the product result is then multiplied by the historical water quality value at the corresponding time point to obtain the enhanced historical sample for each time point. The mathematical model is as follows: ; in, for Time of the first Enhanced historical samples corresponding to each variable; Finally, the enhanced historical samples arranged in chronological order are concatenated to form the enhanced delay vector, the mathematical model of which is: ; Among them, A series of enhanced historical samples were spliced ​​together. for Time of the first The enhanced delay vector corresponding to each variable is used as the subsequent input.

[0040] S3. Based on the correlation of the enhanced delay vector, the covariate coupling weights are calculated, and the water quality variables are weighted to construct the covariate fusion vector; the local change rate factor of water quality changes at adjacent time points is calculated, and the working condition adjustment factor is obtained by combining the cleaning conditions to form the covariate weighting matrix.

[0041] Furthermore, in step S3, a covariate weighting matrix is ​​constructed, the process of which is as follows: Figure 3 As shown, the specific steps for constructing the covariate weighting matrix are as follows.

[0042] S31. Calculate the correlation between the enhanced delay vector and the enhanced delay vector of the candidate variable to obtain the degree of coupling between the variables. Normalize the correlation value in the candidate set to form the covariate coupling weight representing the coupling strength between different variables. In this embodiment, historical correlation statistics are performed using the output enhanced delay vector. Combined with multi-index water quality samples aligned with the drug dosing time, cross-variable coupling strength is identified and screened to obtain covariate coupling weights, which represent the correlation between various water quality indicators. The top three indicator sets are selected to highlight the role of strongly correlated indicators in the prediction target. First, at each time position... Above, the enhanced delay vector Stable Enhanced Delay Vector with Candidate Covariates The correlation of the stable delay vector is calculated to obtain the degree of coupling between multiple variables, and the covariate coupling weights are constructed based on this. The mathematical model is as follows: ; in, For variables For variables Covariate coupling weights, for Time of the first A stable augmented delay vector of candidate covariates, For variables The set of the three candidate variables with the highest absolute correlation values, excluding those containing , Let be the enhanced delay vector of the p-th candidate variable at time t, and let Corr be the Pearson correlation coefficient. The correlation between the enhanced delay vectors of the k-th variable and the j-th variable at time t. For the k-th variable and the candidate set Sum of the absolute values ​​of the correlations among all variables; S32. At each time slice, select the three covariates that rank highest in relevance to the target variable, and combine the enhanced delayed samples of the variables with the enhanced delayed samples of the target variable in a weighted combination. Use the resulting coupling weights as superposition coefficients to form a covariate fusion vector from the delayed samples of each variable. In this embodiment, by utilizing covariate coupling weights, the stable enhanced delayed samples of the target water quality index and its three preceding related indicators at the same time position are weighted and superimposed to perform cross-variable information fusion and consistency amplification, resulting in a covariate fusion vector. Its mathematical model is: ; in, for Enhanced historical samples corresponding to the j-th variable at time j; by superimposing coupled variable information, the target index can reflect the dynamic relationship between multiple variables, enhancing the sensitivity of prediction to the evolution of water quality; S33. By measuring the amplitude of water quality changes at adjacent times, the local rate of change factor is obtained. The local rate of change factor reflects the change factor of fluctuation characteristics in a short period of time, and strengthens the characteristic influence in the range of violent fluctuations. In this embodiment, to highlight time slices with significant changing trends in the historical time series, a local rate of change factor is constructed. By comparing the magnitude of water quality variable changes at adjacent times, a local rate of change factor that reflects short-term fluctuation characteristics is obtained. Its mathematical model is: ; When the conductivity changes abruptly shortly after the agent is added, the rate of change factor will rapidly amplify the conductivity information, improve the sensitivity to short-term anomalies and drastic fluctuations, and avoid prediction lag of the system under operating conditions. S34. The operating conditions recorded during the operation of the thermal power plant pipeline are binarized and marked to distinguish between normal operation and cleaning stages. The operating condition adjustment factor is obtained through normalization calculation. Abnormal fluctuations in the dosage of reagents and circulation flow rate are detected. The operating condition adjustment factor reduces the weight of the current sample in subsequent processing. In this embodiment, the water quality variables are in an unstable state during the descaling and cleaning stage. To suppress the interference during the descaling and cleaning stage, an operating condition adjustment factor is introduced, the mathematical model of which is as follows: ; in, The initial value under normal operating conditions is 0, and it is 1 during cleaning. Operating condition adjustment factor. The value is 1 during normal operation and 0.5 during cleaning. By reading the cleaning and normal operating condition flags of the thermal power plant pipeline, it is possible to identify whether the current state is unsteady. When the current state is identified as cleaning, the effectiveness of the current time slice is reduced. The adjustment factor can automatically weaken the weight of water quality signals in the cleaning and abnormal stages in subsequent modeling, thereby avoiding unstructured noise interference caused by manual operation and sudden operating conditions, and improving the stability and reliability of the prediction results. S35. The covariate fusion vector, local rate of change factor and operating condition adjustment factor are embedded into the time sliding window and arranged in sequence to form a covariate weighting matrix to achieve joint modeling of multivariate time series features. In this embodiment, the covariate fusion vector is expanded over time according to a fixed historical window and sliding step size, and local rate of change factors and cleaning condition adjustment factors are introduced at corresponding positions in the matrix for joint weighting, thus completing the structured organization of time and variables, and obtaining the covariate weighting matrix; the mathematical model of the covariate weighting matrix is: ; in, For the k-th variable at position The covariate weighting elements expand the one-dimensional time trajectory into a two-dimensional structure, containing three layers of information: local rate of change, variable coupling, and operating condition adjustment. The covariate weighting matrix not only retains the temporal sequence characteristics of historical water quality evolution, but also integrates the coupling relationship between variables and operating condition correction information, providing high-quality input for subsequent prediction and control models, thereby improving the ability to characterize and control the water quality change trend during the descaling process of thermal power plants.

[0043] S4. Divide the covariate weighting matrix into local sub-blocks by row and width, obtain the time direction weight based on the difference in the mean of adjacent time slices, obtain the variable perception weight based on the difference in the mean of variables, obtain the anomaly sensitivity factor based on the distribution difference, and jointly weight it with the local sub-blocks to obtain the direction perception expression.

[0044] Furthermore, in step S4, a direction-aware representation is constructed, the process of which is as follows: Figure 4 As shown, the specific steps for constructing a direction-aware representation are as follows.

[0045] S41. Using the starting row and starting column of the window as indexes, the covariate weighting matrix is ​​divided into multiple local sub-blocks by setting the row height and column width of the sliding window. Each sub-block corresponds to a fixed time window and variable combination, resulting in multiple local sub-blocks. In this embodiment, the row height of the sliding window is first set according to the structure of the covariate weighting matrix. With column width The covariate weighting matrix is ​​divided into multiple local sub-blocks, each corresponding to a fixed time window, including multiple combinations of water quality variables. The mathematical model for dividing into local sub-blocks is as follows: ; in, From A local sub-block extracted from the middle. The row height of the sub-block. The column width of the sub-block, , The starting row and starting column indices for the sub-block; In this embodiment, This represents the local water quality characteristics under the current time window t and the combination of variables. This indicates the number of time slices covered, with an initial value of 4 or 5 minutes for sampling, corresponding to 20 minutes, controlling the time span of each local sub-block. This indicates the number of variables included, with an initial value of 3, controlling the coupling scale of variables within each local sub-block; S42. Calculate the differences between water quality monitoring values ​​of adjacent time slices in the local sub-blocks respectively. Sum the difference values ​​at the same time slice location and take the average to obtain the overall change corresponding to the current time slice. Perform differential operation on the overall change index of continuous time slices to obtain the time change amplitude that can reflect the degree of difference between adjacent time slices. Perform dot product with the obtained time change amplitude and the preset scaling factor. Then, through normalization processing, convert the intermediate quantities corresponding to each time slice into standardized weight coefficients. Embed the normalized results into the time dimension in the order of row index to complete the construction of time direction weights. In this embodiment, after the sub-block division is completed, the coupling relationship between the temporal evolution characteristics and variables within each local sub-block is analyzed, and the weights in the time direction and the variable direction are calculated respectively. In the time direction, the water quality change amplitude between adjacent time slices is compared, and the weight value of different time slices on water quality evolution is obtained using an exponential weighting method. The mathematical model of time-aware weighting is as follows: ; in, For the sub-block The temporal variation of a row is calculated as the difference between the mean values ​​of adjacent time slices. For the sub-block The range of time variation of the line The scaling factor is the time scaling factor. The weight is for the time direction, ranging from [0,1]. In this embodiment, The initial value is 0.1, which controls the sensitivity of time differences to the weights. The mathematical model for the calculation is as follows: .

[0046] S43. In the local sub-block, determine the column index of the target variable. For each other column variable, calculate the difference between the value of the column variable in each time row and the value of the target variable in the same row. Sum the difference results of all time rows and take the average to obtain the difference degree of the current column variable relative to the target variable. Take the average difference degree as the variable difference degree of the current column variable. Multiply the difference degree result of each column variable in the local sub-block with the set variable direction scaling factor and normalize it. Transform the difference degree result through the exponential mapping function to obtain the corresponding mapping value. Sum the mapping values ​​of all column variables and use it as the denominator of the normalization. Divide the mapping value of each column variable by the sum of the mapping values ​​of all column variables to obtain the variable perception weight of the current column variable. In this embodiment, weights are calculated by analyzing the coupling differences between different water quality variables in the variable direction. When constructing variable direction weights, an exponential normalization method is used to transform correlations into different weight values, making indicators highly coupled with the target variable have a stronger influence in the modeling. The variable-aware weight mathematical model is as follows: ; in, For the sub-block The variance of a column is calculated as the mean difference between the variable and the target variable. For the sub-block Variable variability in the column The scaling factor for the direction of the variable. The directional weight of the variable, ranging from [0,1]. In this embodiment, The initial value is 0.1, controlling the sensitivity of the differences in control variables to the weights. The mathematical model for the calculation is as follows: ; in, The column index of the target variable in the sub-block; S44. By calculating the KL divergence between the water quality distribution in each local sub-block and the water quality distribution during the reference stable period, and multiplying it with the anomaly amplification coefficient, the anomaly sensitivity factor is obtained, which amplifies the water quality changes under abnormal operating conditions. In this embodiment, this step calculates the KL divergence by comparing the water quality distribution of each local sub-block with the water quality distribution during the reference stable period. This measures the degree of deviation between the current local state and the normal state. Combined with the anomaly amplification factor, the deviation is transformed into an anomaly sensitivity factor. The mathematical model for the anomaly sensitivity factor is as follows: ; in, For the current local sub-block's abnormal sensitivity factor, This is the amplification factor, initially set to 0.5, which controls the degree of amplification of the anomalous factor. To reference the water quality distribution during the stable period, the initial values ​​were obtained from the statistical distribution of the mean values ​​of samples under normal operating conditions over the past 7 days. This represents the current water quality distribution in a local sub-block. The KL divergence between the local sub-block and the reference stable period distribution measures the difference between the current local sub-block and the normal water quality state. The local sub-blocks correspond to the drastic fluctuations in conductivity and turbidity during the cleaning period. Their difference from the reference distribution will increase significantly, and the anomaly sensitivity factor will also be amplified. The model will automatically increase its attention to the anomaly sub-blocks in subsequent processing. S45. Perform a dot product operation on the local sub-block data and the calculated time direction weight, variable direction weight and anomaly sensitivity factor to construct a direction-aware representation. In this embodiment, the mathematical model for direction-aware representation is: ; in, The row index of the sub-block, with an initial value range of... , The column index for the sub-block, with an initial value range of... , The sub-block elements, after being expressed with orientation awareness, contain multi-dimensional weighted information. The element value is a local sub-block.

[0047] S5. Construct a global state vector based on the mean, trend slope, fluctuation variance, and anomaly weights calculated from the direction-aware expression. Concatenate the global state vector with the dosage change and temperature deviation to form an assembly vector. Obtain the predicted output of the descaling water quality of the thermal power plant through the linear vector.

[0048] Furthermore, in step S5, the predicted output results of descaling water quality in thermal power plants are constructed, and the process is as follows: Figure 5 As shown, the specific steps for constructing the predicted output results of descaling water quality in thermal power plants are as follows.

[0049] S51. By calculating the time slice mean, trend slope and fluctuation variance of the direction-aware expression, and combining it with the anomaly sensitivity factor for normalization weighting, linear aggregation is performed to obtain the global state vector. In this embodiment, in each direction-aware expression sub-block element, rows represent time slices and columns represent variable combinations. The column mean of each row is calculated to compress the multivariate information within the same time slice into a single scalar. The mathematical model for the column mean of each row is as follows: ; in, The column mean is used to summarize the multivariate coupling strength of the elements in the current direction-aware expression sub-block. Subsequently, the time slice mean, trend slope, and variance are extracted along the time dimension. The mathematical model for the time slice mean is as follows: ; in, The mean of the current time slice in the direction-aware expression sub-block element is used to characterize the stability level of the direction-aware expression sub-block element; The mathematical model for the trend slope is: ; in, The trend slope of the current time slice in the direction-aware expression sub-block element is represented by the average first-order difference of the row mean, which is used to characterize the rising and falling trends of the direction-aware expression sub-block element and reflect the short-term response after drug administration. The mathematical model for volatility variance is: ; in, The variance of the current time slice in the direction-aware expression sub-block element is used to characterize the local fluctuations caused by washing and backwashing; Since the influence values ​​of perceptual expression sub-block elements differ across different directions, an anomaly sensitivity factor is introduced as the weighting criterion and normalized to construct an anomaly weight. This ensures that perceptual expression sub-block elements with higher influence values ​​receive higher weights in the global aggregation. The mathematical model for the anomaly weight is as follows: ; in, These are abnormal weight values. For the initial behavior The starting column is The abnormal sensitivity factor of the local sub-block; The mean, trend slope, and variance of each directional perception sub-block are mapped to a unified global state space. The global state vector at the current moment is obtained by weighting and summing the values ​​according to the anomaly weights; the mathematical model of the global state space is: ; The mathematical model for the global state vector is: ; in, For a fixed mapping matrix, The initial value is set to 32. For a moment The global state vector provides a unique input representation for subsequent operations.

[0050] S52. By concatenating the global state vector obtained in the previous step with the dosage change and temperature deviation into an autoregressive input, linear calculation is performed to obtain the multi-index prediction results for the next 15 minutes for the descaling scenario. In this embodiment, water quality changes are affected by variations in dosage and temperature deviations. The method for calculating the variation in dosage is the difference between the dosage at the previous moment and the dosage at the current moment, and its mathematical model is as follows: ; in, For the current dosage, This refers to the dosage administered at the previous moment. The change in drug dosage at time t is used to capture instantaneous impacts; The temperature deviation is calculated as follows: the current temperature value and the optimal descaling temperature. The mathematical model for the difference is: ; in, This represents the temperature value at time t. The autoregressive input is obtained by concatenating the changes in drug dosage, temperature deviation, and global state vector. Its mathematical model is as follows: ; The concatenation provides an explicit input vector; The autoregressive input is then linearly transformed to finally output the multi-indicator prediction results for the next 15 minutes. The mathematical model is as follows: ; in, Here, is the intercept term, initially set to 0; and is a linear coefficient vector, with initial values ​​following a normal distribution. This is a linear coefficient vector, initially composed of random numbers following a normal distribution, whose values ​​are adaptively adjusted during training. The model predicts the results of multiple indicators for the next 15 minutes at the current time t, including pH value, Langelier saturation index, calcium ion concentration, silicate ion concentration, conductivity, and turbidity.

[0051] S6. Construct a prediction model for descaling water quality in thermal power plants. Input the descaling water quality dataset of thermal power plants, use joint loss as the loss function, and sequentially go through steps S2 to S5. Then, iterate and train until convergence to complete the training of the prediction model for descaling water quality in thermal power plants.

[0052] Furthermore, in step S6, the specific steps for training the descaling water quality prediction model for thermal power plants are as follows.

[0053] S61. Robust loss based on the Huber function: The Huber function calculates the difference between the predicted value and the true value. If the absolute value of the difference is less than the threshold, the square of the difference between the predicted value and the true value is used as the loss value. If the absolute value of the difference is greater than or equal to the threshold, the absolute value of the difference between the predicted value and the true value minus the threshold is multiplied by the threshold to use as the loss value. A dynamic uncertainty factor is introduced for each prediction target. Its value is the standard deviation between the true value and the predicted value. The uncertainty factor is combined with the robust loss of the Huber function to form the heteroscedasticity robust task loss. In this embodiment, a heteroscedasticity robust task loss is introduced to optimize the water quality prediction model for descaling in thermal power plants, while considering adaptive uncertainty adjustment. The heteroscedasticity robust task loss function uses the Huber loss function to robustly process the prediction error, reducing the interference of outliers and noise, thereby improving the model's robustness to outliers. Simultaneously, a dynamic uncertainty factor is introduced, adaptively adjusting according to the uncertainty of each prediction target to automatically reduce the impact of noise, allowing the model to assign high weights to high-confidence predictions and suppress low-confidence predictions. The mathematical model for the heteroscedasticity robust task loss is as follows: ; in, The prediction steps are set to 15. The target water quality index set includes: pH value, Langelier saturation index, calcium ion concentration, silicate ion concentration, conductivity, and turbidity. For the true value, This is the inflection threshold for the Huber loss function, initially set to 0.5. To automatically adjust the uncertainty factor, representing the uncertainty of each predicted target, the confidence weights of each output term are automatically adjusted, with an initial value set to 0.1. This represents the loss value for heteroscedasticity robust tasks. In this embodiment, the Huber loss is an improvement on the standard mean squared error loss, exhibiting robustness and effectively handling outliers. When the error is small, the Huber loss is non-linear, maintaining smoothness; however, when the error is large, the Huber loss becomes linear, reducing the impact of outliers on the training process. The mathematical model of the Huber loss function is as follows: ; Where m is equivalent to ; In this embodiment, the method for calculating the automatic adjustment uncertainty factor is as follows: The standard deviation.

[0054] S62. By setting hyperparameters and training the descaling water quality dataset from thermal power plants, the training process of the descaling water quality prediction model is finally completed. In this embodiment, a model for predicting and controlling descaling water quality in thermal power plants based on multidimensional water quality sensor data is constructed. The model is input into a thermal power plant descaling water quality dataset and sequentially undergoes multiple stages, including enhanced delay vector, covariate weighted matrix, direction-aware representation, and assembly vector. Further enhancements are made by introducing local fluctuation suppression weights, water temperature deviation adjustment weights, reagent mutation suppression factors, time decay weights, covariate coupling weights, local change rate factors and operating condition adjustment factors, direction-aware weights, anomaly sensitivity factors, local sub-blocks, and a global state vector. This strengthens the model's accurate response capability in cross-time period water quality prediction and descaling reagent dosage control. During training, a heteroscedasticity robust task loss optimization strategy is introduced. By constraining the model's stability, prediction accuracy, and response sensitivity to convergence, the model is guided to learn the relationship between water quality evolution patterns and control strategies, improving its adaptability to abnormal states such as water quality mutations and cleaning disturbances. Ultimately, a thermal power plant descaling water quality prediction and control model with high prediction accuracy, excellent response stability, and wide operating condition adaptability is formed, effectively improving the dynamic water quality control capability and operating condition change response capability during the thermal power plant descaling process.

[0055] Furthermore, the descaling water quality prediction and control model for thermal power plants proposed in this invention is implemented using the Python programming language and developed using the PyTorch deep learning framework. During training, the Adam optimizer is selected, with an initial learning rate set to 0.001. An exponential decay strategy is used to dynamically adjust the learning rate during training. The batch size is set to 64, and the number of training epochs is 150. The loss function is optimized using the heteroscedasticity robust task loss function, and an additional dynamic uncertainty factor is introduced to adaptively adjust according to the uncertainty of the prediction target.

[0056] Furthermore, the collected data on descaling water quality from thermal power plants is input into the constructed prediction dataset for descaling water quality from thermal power plants for training and validation. The joint loss convergence curve is shown in the figure below. Figure 6 As shown in the figure, as the number of training rounds increases, the task loss value gradually decreases from a relatively high level in the initial stage and tends to stabilize in the later stage. This indicates that the water quality prediction model for descaling in thermal power plants converges and stabilizes during the iteration process, and can effectively complete the extraction of water quality time series features and multivariate coupled modeling. Figure 7 To analyze the distribution of points between the actual and predicted values ​​after dimensionality reduction, the graph shows that the predicted values ​​of the thermal power plant descaling water quality prediction model are close to the actual values, indicating that the model can accurately predict the water quality conditions for the next 15 minutes. The prediction effect is as follows: Figure 8 , 9 As shown in Figure 10, the horizontal axis represents time in minutes, and the vertical axis represents the corresponding water quality index values, including pH, conductivity, calcium ion concentration, turbidity, silicate concentration, and Langelier saturation index. The figure shows that the actual sampled values ​​are presented as dots, reflecting the dynamic changes of the circulating water in the thermal power plant under the influence of chemical dosing and temperature disturbances. At the 75th minute, the predicted value of the thermal power plant descaling water quality prediction model is marked with a triangle. The results show that the predicted point is highly close to the actual value, indicating that the thermal power plant descaling water quality prediction model can accurately capture the water quality evolution pattern. Compared with traditional methods, this thermal power plant descaling water quality prediction model exhibits higher fitting accuracy and stronger anomaly sensitivity in multi-index joint prediction. It can not only reduce the interference of sudden chemical addition and non-stationary factors such as water temperature fluctuations on the prediction results, but also maintain the physical consistency and trend continuity among multiple indicators.

[0057] S7. Based on the prediction output of the thermal power plant descaling water quality prediction model, adjust the type and metering of the added reagents. The specific steps are as follows.

[0058] In this embodiment, the proposed descaling water quality prediction model for thermal power plants can output multiple indicators for the next 15 minutes, including pH, conductivity, calcium ion concentration, turbidity, silicate concentration, and Langelier saturation index. Based on the prediction results, the model dynamically adjusts the type and dosage of reagents added. When the predicted pH is below the optimal reaction range, the system increases the dosage of alkaline reagents to improve solution stability. If the pH is above the upper limit, the dosage of alkaline reagents is reduced or an acidic buffer is added appropriately to maintain the reaction environment. In cases of abnormally high conductivity prediction, the system reduces the addition of electrolyte reagents to avoid excessive water hardness. Conversely, when conductivity is too low, auxiliary reagents are added appropriately to enhance ion activity and ensure the descaling reaction is successful. The effectiveness of the prediction is assessed; if the predicted calcium ion concentration increases and may exceed the allowable upper limit, the dosage of chelating agent is increased to reduce the calcium ion level in the water, thereby avoiding the risk of scaling; for the predicted results of turbidity and silicate concentration, when an upward trend is observed, the dosage ratio of dispersant and flocculant is appropriately increased to suppress particulate deposition and secondary pollution; finally, if the prediction results of the Langelinie saturation index indicate that the system has a tendency to scale, the feedforward suppression of scaling risk is achieved by preferentially adjusting the total amount and type of reagents; through the above prediction-driven dosage control mechanism, this application can achieve intelligent adjustment of reagent type and dosage, ensuring the water quality stability and system operation safety of the circulating cooling water in the descaling process of thermal power plants.

[0059] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements are all within the scope of protection of the present invention.

Claims

1. An intelligent water quality prediction and regulation method in a power plant descaling process, characterized in that, The method comprises the following steps: Collecting multi-source water quality data of a thermal power plant to construct a water quality data set for descaling of the thermal power plant; According to the water quality data set for descaling of the thermal power plant, a local fluctuation suppression weight is calculated by local variance, a water temperature deviation adjustment weight is calculated by a water temperature difference, a reagent mutation suppression factor is calculated by a reagent dosage at adjacent time points, a time decay weight is calculated according to a decay coefficient, and the factors are fused point by point to form an enhanced delay vector; A covariant coupling weight is calculated by correlation of the enhanced delay vector, water quality variables are weighted, a covariant fusion vector is constructed, a local change rate factor is calculated according to water quality changes at adjacent time points, a working condition adjustment factor is obtained in combination with a cleaning working condition, and a covariant weighted matrix is formed; The covariant weighted matrix is divided into local subblocks according to rows and columns, a time direction weight is obtained based on a mean value difference between adjacent time slices, a variable perception weight is obtained by a variable mean value difference, an abnormal sensitivity factor is obtained based on a distribution difference, and the local subblocks are jointly weighted to obtain a direction perception expression; A mean value, a trend slope and a fluctuation variance are calculated according to the direction perception expression, a global state vector is constructed by combining an abnormal weight obtained from the abnormal sensitivity factor, and a self-recurrence input is spliced by combining a reagent change amount and a temperature deviation, and a prediction output result is obtained through a linear vector; A water quality prediction model for descaling of a thermal power plant is constructed, the water quality data set for descaling of the thermal power plant is input, a joint heteroscedastic robust task loss function is constructed, and iterative training is performed until convergence, thereby completing training of the water quality prediction model for descaling of the thermal power plant.

2. The method according to claim 1, wherein, Raw water quality monitoring data of a thermal power plant boiler at different operation stages in a chemical cleaning and descaling process are acquired, including multi-source water quality parameters collected at nodes of a make-up water outlet, condensate water fine treatment, a boiler feed water inlet, a boiler blowdown branch, a front end of a circulating water heat exchanger, a rear end of the circulating water heat exchanger, a chemical cleaning loop inlet and a chemical cleaning loop outlet; The multi-source water quality parameters include pH value, oxidation-reduction potential, conductivity, turbidity, calcium ion concentration, magnesium ion concentration, dissolved oxygen concentration, silicate concentration, total hardness, chloride ion concentration, sulfate concentration, temperature, flow rate, pressure, corrosion resistance value and electrochemical noise signal; The raw water quality data are processed by unit conversion, abnormal value elimination, missing value interpolation and feature scaling standardization to construct a water quality data set for descaling of a thermal power plant.

3. The method of claim 1, wherein the method further comprises: Firstly, a local variance of each time point is calculated based on water quality change amplitudes at the time point and two continuous time points before and after the time point, and a local fluctuation suppression weight is constructed by taking the inverse of the local variance and normalizing. A water temperature deviation adjustment weight is calculated by combining a difference between a circulating water temperature at a current time point and an optimal descaling temperature of a pipeline of the thermal power plant, and the water temperature deviation adjustment weight has a large value when the temperature is close to the optimal value. A reagent mutation suppression factor is obtained by comparing a reagent dosage change amplitude between a current time point and a previous time point, and the influence degree of the current time slice is reduced when the reagent dosage has a mutation. A time decay weight is calculated according to a time interval between each historical sample and the current time point. The enhanced historical sample at each time point is obtained by multiplying the local fluctuation suppression weight, the water temperature deviation adjustment weight, the reagent mutation suppression factor, the time decay weight and the historical water quality value at the corresponding time, and all the enhanced historical samples are arranged in time sequence to construct an enhanced delay vector.

4. The method of claim 1, wherein the method further comprises: The enhanced delay vector and the enhanced delay vector of the candidate variable are calculated to obtain the coupling degree between the variables, and the correlation value is normalized in the candidate set to form a covariant coupling weight representing the coupling strength between different variables. In each time slice, the top three covariates with high correlation with the target variable are selected, and the enhanced delay samples of the variables and the enhanced delay samples of the target variable are combined together, and the coupling weight is used as the superposition coefficient to form a covariant fusion vector. The local change rate factor is obtained by measuring the water quality change amplitude of adjacent time points, which reflects the change factor of fluctuation characteristics in a short time, and strengthens the influence of the characteristics in the fluctuation intense interval. The working condition state recorded during the operation of the pipeline of the thermal power plant is binary labeled to distinguish between normal operation and cleaning stage, and a working condition adjustment factor is obtained by normalization calculation to detect abnormal fluctuations in reagent dosage and circulating flow rate, and the working condition adjustment factor reduces the weight of the current time sample in subsequent processing. The point product calculation is performed on the covariant fusion vector, the local change rate factor and the working condition adjustment factor, and the covariant weighted matrix is formed in sequence to realize the joint modeling of the multi-variable time sequence characteristics.

5. The method of claim 1, wherein the method further comprises: The starting row and column of the window are used as the index, the covariant weighted matrix is divided into multiple local sub-blocks by setting the row height and column width of the sliding window, each sub-block corresponds to a fixed time window and variable combination, and multiple local sub-blocks are obtained. The difference between the water quality monitoring values of adjacent time slices in the local sub-block is calculated, and the difference values at the same time slice position are summed and averaged to obtain the overall change corresponding to the current time slice. The overall change of the continuous time slices is differentiated to obtain the time change amplitude. The time change amplitude is combined with the preset scaling factor and normalized to form a standardized weight coefficient, and the time dimension is embedded in the row index order to construct the time direction weight.

6. The method of claim 1, wherein the method further comprises: The column index of the target variable is determined in the local sub-block, the difference between the values of the remaining variables in each time row and the values of the target variable in the same row is calculated, and the difference results are summed and averaged row by row to obtain the variable difference degree of each column variable relative to the target variable. The variable difference degree is multiplied by the variable direction scaling factor and normalized, and the mapping value is obtained through the nonlinear mapping function transformation. The mapping values of all column variables are added as the normalization denominator, and the variable perception weight is determined by the ratio of the mapping value of each column variable to the normalization denominator. The abnormal sensitivity factor is obtained by calculating the KL divergence between the water quality distribution in each local sub-block and the reference stable period water quality distribution and multiplying it by the abnormal amplification coefficient, which amplifies the water quality change under abnormal working conditions. The direction perception expression is constructed by dot product operation of the local sub-block data, the time direction weight, the variable direction weight and the abnormal sensitive factor.

7. The method of claim 1, wherein the method further comprises: The global state vector is obtained by linear aggregation of the abnormal weight obtained by calculation of the time slice mean, the trend slope and the fluctuation variance of the direction perception expression, and normalization weighting of the abnormal sensitive factor; The drug change amount is obtained by calculating the difference between the drug amount of the last moment and the current moment, and the temperature deviation is obtained by calculating the difference between the temperature value of the current moment and the optimal descaling temperature. The global state vector obtained in the last step, the drug change amount and the temperature deviation are spliced into an autoregressive input, and linear calculation is performed to obtain a 15-minute multi-index prediction result for the descaling scene. 8.The method of claim 1, wherein the method further comprises: determining a water quality of the water in the water tank based on the water quality data; and controlling the water quality of the water in the water tank based on the water quality data. Based on the robust loss of Huber function, the Huber function calculates the difference between the predicted value and the true value. If the absolute value of the difference is less than the turning threshold, the square of the difference between the predicted value and the true value is taken as the loss value. If the absolute value of the difference is greater than or equal to the turning threshold, the absolute value of the difference between the predicted value and the true value minus the threshold is multiplied by the threshold to form the loss value. A dynamic uncertainty factor is introduced on each prediction target, and the value of the uncertainty factor is the standard deviation of the true value and the predicted value. The uncertainty factor is combined with the robust loss of Huber function to form a heteroscedastic robust task loss. The training process of the descaling water quality prediction model is finally completed by setting hyperparameters and training with the descaling water quality data set of the thermal power plant.

Images