Generator stator cooling water ph value control optimization method

Abstract

The present application relates to the technical field of control optimization, and discloses a method for optimizing pH value control of generator stator cooling water, which comprises the following steps: constructing a two-dimensional operating phase space of pH regulation characteristics data of the generator stator cooling water and extracting nonlinear gain characteristics; online identifying a time-varying object model of a controlled process of the generator stator cooling water pH; optimizing controller parameters based on working condition adaptive evaluation, corrosion risk constraint and feedforward search boundary shrinkage; verifying, locally correcting and issuing parameters based on continuous actual disturbance playback of optimization results; and online operating the generator stator cooling water pH regulation. In the online identification of the object model, an adaptive forgetting factor driven by the change speed of the cooling water conductivity is introduced, so that the recursive least square algorithm can accelerate model tracking when the buffer capacity is migrated and suppress parameter jitter when the working condition is stable, thereby providing a time-varying object model closer to the current actual dynamic for subsequent optimization.

Description

Technical Field

[0001] This invention relates to the field of control optimization technology, specifically a method for controlling and optimizing the pH value of generator stator cooling water. Background Technology

[0002] During the operation of large steam turbine generators or hydro turbine generators, the stator windings generate a large amount of Joule heat when carrying high currents. To ensure the safe and stable operation of the generator, this heat must be continuously removed through the stator cooling water system. The water used for cooling must not only have good thermal conductivity but also maintain extremely low electrical conductivity to ensure the electrical insulation safety of the windings.

[0003] To inhibit the corrosion of copper stators and connectors by cooling water, the industry commonly uses the addition of chemicals such as ammonia to maintain the pH of the cooling water within a slightly alkaline target range (typically between 9.0 and 9.5). However, in practical engineering applications, this chemical dosing process faces significant challenges due to its nonlinear and time-varying characteristics. The fundamental reason is that the generator stator cooling water system is not a simple pure water loop, but a complex chemical buffer system containing ammonia, dissolved carbon dioxide, and trace ions introduced through makeup water.

[0004] The characteristics of this buffer system mean that its response sensitivity to chemical dosing changes dynamically with operating conditions. Specifically, when the conductivity of the cooling water is high due to ion accumulation, the system's buffer capacity is strong, and the same change in the dosing valve position has a smaller effect on pH adjustment (low sensitivity zone). Conversely, when the system discharges wastewater and replenishes water, causing a decrease in conductivity, the buffer system is diluted, and the system's sensitivity to chemical dosing increases sharply (high sensitivity zone). If the control system fails to detect this dynamic shift in buffer capacity, pH overshoot can easily occur in the high sensitivity zone, making the cooling water highly alkaline and accelerating copper corrosion, or in the low sensitivity zone, the system may deviate from the target range for extended periods due to response lag, affecting equipment lifespan and operational safety. Therefore, a method for controlling the pH of generator stator cooling water is urgently needed. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method for controlling the pH value of generator stator cooling water. It constructs a two-dimensional operating phase space defined by pH deviation and cooling water conductivity, and uses weighted fuzzy clustering and local titration gain sorting to generate gain sensitivity level labels. This enables the control system to distinguish the nonlinear regulation sensitivity under different buffering capabilities, rather than classifying operating conditions solely based on a single pH deviation.

[0006] To achieve the above objectives, the present invention employs the following technical solution: A method for optimizing the pH value control of generator stator cooling water includes the following steps: S1. Collect historical operating data of generator stator cooling water pH control characteristics, construct a two-dimensional operating phase space based on historical operating data, extract nonlinear gain features and convert them into sensitivity level labels to obtain a structured operating sample set. S2. The pH controlled process is represented online as a first-order inertial plus pure time delay discrete object model, and the forgetting factor of the recursive least squares algorithm is adaptively adjusted by using the conductivity change rate, which reflects the dynamic changes in water quality, so that the object model can be updated with the current operating conditions. S3. Construct multiple differentiated typical working conditions for closed-loop simulation. During the particle swarm optimization process, design a fitness function that integrates gain sensitivity level labels, parameter identification confidence, alkaline overshoot corrosion risk penalty term, and dynamic search boundary contraction strategy based on feedforward jitter index. Through optimization, the obtained parameters take into account tracking accuracy, material safety, and actuator life. S4. Before outputting the final controller parameters, add a round of result verification and local correction based on continuous actual disturbance sequences, so that the parameter results undergo a complete verification close to the actual field operating conditions before they are about to be issued. S5. The optimal controller parameters, verified and corrected by S4, are sent to the execution system. Combined with real-time water quality indicators, feedforward feedback adjustment is performed to monitor the operating conditions. When deviations occur, the model and parameters are self-optimized to achieve stable pH control.

[0007] Furthermore, S1 specifically refers to: The pH deviation value is combined with the cooling water conductivity to construct a two-dimensional operating phase space, which includes the current degree of deviation of the system from the target and the current buffering capacity of the system. Weighted fuzzy C-means clustering is introduced into the two-dimensional operating phase space, and combined with the ranking of the actual dynamic response sensitivity of each region after clustering, the historical samples are divided into gain sensitivity segments with practical control significance.

[0008] Furthermore, S2 specifically refers to: Determine the controller sampling period, estimate the pure time delay based on the effective delivery volume of the dosing pipeline and the current cooling water flow rate, and determine the number of pure time delay steps based on the prior estimate of the pure time delay and the sampling period; The parameters are identified by using the rate of change of the conductivity of cooling water through an adaptive recursive least squares method.

[0009] Furthermore, S3 specifically refers to: S31. Extract a limited number of representative typical working conditions from the structured operation sample set. By comprehensively evaluating the closed-loop performance of the controller in these significantly different scenarios during the optimization process, the particle swarm algorithm is forced to find parameter combinations that are robust to all types of working conditions, rather than favoring a certain local optimal solution. S32, Calculate particle fitness by integrating section labels, identification confidence, alkaline overshoot corrosion risk, and valve action stability; S33. Search boundary shrinkage and particle update based on feedforward jitter index.

[0010] Furthermore, S4 specifically refers to: A continuous sequence of actual disturbances is extracted from recent historical operational data as a verification sequence; The controller parameter vector is fed into the control simulator, the verification sequence is fully replayed, and the overall verification performance index is calculated. The overall performance metrics will be compared with acceptable thresholds. When the overall performance index is verified to be greater than the acceptable threshold, a small-scale local correction is performed with the controller parameter vector as the center to obtain the locally corrected controller parameter vector. The locally corrected controller parameter vector is validated again, and the final controller parameter vector is output after the validation passes.

[0011] Furthermore, the pH deviation value is combined with the cooling water conductivity to construct a two-dimensional operating phase space, specifically: Read the set pH value, actual pH value, cooling water conductivity, and dosing valve position command from historical operating data according to a uniform sampling period; Validity screening of historical operation data is performed, and invalid sampling points with sensor disconnection, obvious boundary violations, duplicate sampling timestamps, and instantaneous device switching are deleted; Calculate the pH deviation value at each sampling point; Normalize the pH deviation value sequence and the cooling water conductivity sequence respectively to obtain the normalized pH deviation value and the normalized cooling water conductivity; The first normalized pH deviation value and the normalized cooling water conductivity are used to construct the second normalized pH deviation value and the third normalized cooling water conductivity. Two-dimensional operational phase space sample points.

[0012] Furthermore, weighted fuzzy C-means clustering is introduced into the two-dimensional operational phase space, and combined with the ranking of the actual dynamic response sensitivity of each region after clustering, historical samples are divided into gain sensitivity segments with practical control significance, specifically: Perform fuzzy C-means clustering on the two-dimensional operational phase space sample set; Calculate the representative local titration gain within each cluster segment; All cluster segments are reordered from largest to smallest based on representative local titration gain, and gain sensitivity level labels are generated. For each sampling point, compare its membership degree to each cluster segment, and write the sampling point into the gain sensitivity level label corresponding to the cluster segment with the highest membership degree to obtain the structured running sample set.

[0013] Furthermore, S31 specifically refers to: Multiple typical operating scenarios were extracted from the structured operational sample set according to time windows; Configure object model parameters and equivalent disturbance sequences for each typical working scenario; Define the feedback control section of the candidate controller; Define the feedforward compensation part of the candidate controller; Calculate the total dosing valve position command and perform object model simulation.

[0014] Furthermore, S32 specifically refers to: Set operating condition weights for each typical operating scenario; Construct sample-level error weights based on gain sensitivity level labels; Construct scenario-level confidence weights based on parameter estimation confidence levels; Calculate the sub-fitness for each typical working condition scenario; The overall particle fitness is obtained by weighted summation of the sub-fitnesses of all typical working scenarios.

[0015] Furthermore, S33 specifically refers to: Initialize the particle swarm search space; After each generation of fitness evaluation, the feedforward control output sequence is extracted separately and the feedforward jitter index is calculated. The average feedforward jitter index is compared with the preset jitter tolerance upper limit, and the search upper bound of the cooling water conductivity feedforward gain is dynamically narrowed. For particles that exceed the new search upper bound, redirect their velocity and synchronously correct their positions. Each particle is updated using the standard velocity-position update rule of the particle swarm optimization algorithm.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. Construct a two-dimensional operating phase space defined by pH deviation and cooling water conductivity, and use weighted fuzzy clustering and local titration gain sorting to generate gain sensitivity level labels, so that the control system can distinguish nonlinear control sensitivity under different buffering capabilities, rather than classifying the operating conditions based solely on a single pH deviation. 2. In the online identification of object models, an adaptive forgetting factor driven by the rate of change of cooling water conductivity is introduced, which enables the recursive least squares algorithm to accelerate model tracking when the buffer capacity is shifted and suppress parameter jitter when the operating conditions are stable, thereby providing a time-varying object model that is closer to the current actual dynamics for subsequent optimization. 3. During the controller parameter optimization stage, the gain sensitivity level label, parameter identification confidence, and alkaline overshoot cubic penalty term are simultaneously incorporated into the particle swarm fitness function. This ensures that the optimization process not only focuses on minimizing tracking error but also actively avoids the risk of copper component corrosion caused by high pH overshoot. Furthermore, by monitoring the feedforward jitter index online and dynamically shrinking the search boundary, the amplification effect of the cooling water conductivity feedforward channel on the high-frequency action of the valve is suppressed. 4. Before the optimization results are issued, a continuous actual disturbance sequence replay verification and local correction mechanism is introduced. The controller parameters are further corrected from "comprehensive optimal under typical working conditions" to "available continuous sequence in the field" by using complete verification under non-stationary disturbance environment. The additional disturbance risk during parameter switching is reduced by safety boundary check and integral continuity maintenance before parameter update. Attached Figure Description

[0017] Figure 1 This is a flowchart of the present invention; Figure 2 This is a diagram showing the distribution of all historical sampling points in the two-dimensional operational phase space; Figure 3 It is a weighted fuzzy C-means cluster center and segment boundary. On the basis of the two-dimensional phase space scatter plot, the cluster center of the weighted fuzzy C-means cluster and the segment boundary line determined by the maximum membership degree are superimposed. Figure 4 This is an adaptive adjustment curve of the forgetting factor; Figure 5 It is an online graph that identifies parameter change patterns; Figure 6 This is a diagram of the convergence process of fitness in particle swarm optimization. Detailed Implementation

[0018] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined in this application.

[0019] like Figure 1 As shown, this invention proposes an optimization method for controlling the pH value of generator stator cooling water, the main contents of which are as follows: S1. Two-dimensional operating phase space construction and nonlinear gain feature extraction of generator stator cooling water pH regulation characteristic data. The generator stator cooling water system maintains the cooling water within the target slightly alkaline range through chemical dosing. However, ammonia, dissolved carbon dioxide, and ions introduced by makeup water in the cooling water collectively form a buffer system, causing a significant non-linear relationship between changes in the dosing valve position and the actual pH value. The same change in the dosing valve position has varying impacts on the actual pH value under different cooling water conductivity and pH deviation conditions. Directly applying a uniform control parameter tuning method to all historical data easily yields parameter results applicable only to a single operating segment. This can lead to increased overshoot, slower recovery, or frequent valve actuation after the operating point shifts.

[0020] This invention first organizes historical operating data into a two-dimensional operating phase space that can simultaneously characterize the current control deviation and buffer capacity status. Then, it extracts nonlinear gain features based on the rate of change of actual pH value caused by the change of dosing valve position in local samples, and converts these nonlinear gain features into gain sensitivity level labels with control significance, thus obtaining a structured operating sample set. This provides a basis for operating condition labels for subsequent online model identification and particle swarm optimization. The specific steps are as follows: S101. Historical operational data regularization and two-dimensional operational phase space sample construction To address the nonlinearity between changes in the dosing valve position and pH response caused by variations in the buffer system, historical operating data needs to be processed and transformed. Traditional methods rely solely on pH deviation as the control basis, failing to reflect the differences in system buffer capacity under varying cooling water conductivity.

[0021] This step combines the pH deviation value with the cooling water conductivity to construct a two-dimensional operating phase space. The two dimensions of this space correspond to "the degree to which the system deviates from the target" and "the strength of the system's current buffering capacity," respectively. This provides a structured input basis for subsequent data-driven methods to distinguish operating segments with different control sensitivities. The specific steps are as follows: 1) Read the set pH value, actual pH value, cooling water conductivity and dosing valve position command from historical operating data according to a uniform sampling period.

[0022] The total number of historical sampling points is recorded as , This indicates the number of historical sampling points that participated in the offline analysis.

[0023] No. The set pH value for each sampling point is denoted as... , indicating the target control value; No. The actual pH value at each sampling point is recorded as follows: This indicates the actual pH value measured on-site. No. The cooling water conductivity at each sampling point is denoted as . The unit is (micro-Siemens per centimeter), used to characterize the ion concentration level and buffering capacity of cooling water; No. The dosing valve position command for each sampling point is recorded as follows: The unit is valve position percentage, used to characterize the actuator output, that is, the opening command of the actuator (dosing valve), for example... This indicates that the valve is 35% open, corresponding to the stroke position of the dosing pump or regulating valve.

[0024] 2) Validate historical operating data and delete invalid sampling points that are sensor disconnected, obviously out of bounds, have duplicate sampling timestamps, or are at the moment of device switching.

[0025] In practical implementation, data can be sorted in ascending order by sampling time. If the interval between adjacent timestamps is not equal to the preset sampling period, the segment is marked as a missing sampling segment. If the actual pH value exceeds the allowable range for engineering purposes, such as less than 7.0 or greater than 10.5, it is judged as an abnormal measurement point. If the conductivity of the cooling water is negative or remains unchanged for a long time, it is judged as an instrument failure point. For a single isolated missing point, linear interpolation between the preceding and following points can be used to fill in the gap. For segments with a continuous missing length of more than 5 sampling points, they are directly discarded.

[0026] 3) Calculate the pH deviation value for each sampling point.

[0027] No. The pH deviation value of each sampling point is denoted as , This represents the difference between the set pH value and the actual pH value at the current moment. A positive value indicates that the current actual pH value is lower than the set pH value, and a negative value indicates that the current actual pH value is higher than the set pH value. The calculation method is expressed as follows: .

[0028] 4) Normalize the pH deviation value sequence and the cooling water conductivity sequence respectively to obtain the normalized pH deviation value and the normalized cooling water conductivity. The normalized pH deviation value is denoted as... Normalized cooling water conductivity is denoted as .

[0029] To avoid the pH deviation and cooling water conductivity being too different in terms of dimensions and numerical range, which could lead to subsequent clustering results being dominated by a single dimension, a convenient approach is to use minimum-maximum normalization to map the pH deviation and cooling water conductivity to... scope.

[0030] 5) Using normalized pH deviation values and normalized cooling water conductivity Jointly construct the first Two-dimensional operational phase space sample points. These sample points are used to characterize the combined location of the "current control error state" and the "current chemical buffer state".

[0031] In one embodiment, as an example, assume the first After normalization of each sampling point: (Too low) (High conductivity, strong buffering capacity), then the two-dimensional phase space sample points are: This indicates a state of "moderate deviation but strong buffer".

[0032] Based on this, all historical operational data is converted into a two-dimensional operational phase space sample set, and subsequent segmentation can be performed directly on this sample set.

[0033] It should be noted that the purpose of constructing a two-dimensional operational phase space is to remap massive amounts of time-series data points onto a state plane. On this plane, points located close to each other indicate that they not only have similar pH control errors but also similar ion buffer environments. This allows subsequent clustering algorithms to more accurately group historical operating conditions with similar "dosing-response" characteristics into one category, rather than classifying them solely based on time or a single pH deviation.

[0034] In one embodiment, such as Figure 2 As shown, the distribution of samples and sensitivity levels in the two-dimensional operating phase space are presented, illustrating the distribution of all historical sampling points in the two-dimensional operating phase space during the pH control process of the generator stator cooling water. The horizontal axis represents the normalized pH deviation value. Dimensionless; the vertical axis represents the normalized cooling water conductivity. , dimensionless. Figure 2 Each scatter point represents the system state at a sampling time. The color indicates the sensitivity level label of the point after being divided into gain sensitivity segments: red for level 1 (most sensitive), orange / light orange for level 2, light yellow for level 3, and green for level 4 (least sensitive). By constructing a two-dimensional phase space by combining pH deviation and conductivity, it is possible to effectively distinguish operating conditions under different buffering capacities and control deviations.

[0035] S102. Gain Sensitivity Segmentation and Level Label Generation The conventional method of distinguishing operating conditions solely based on the magnitude of pH deviation is insufficient to express the changes in buffer capacity caused by the conductivity of cooling water, leading to the same pH deviation being mistakenly considered equally sensitive in different buffer zones.

[0036] This invention divides historical samples into gain sensitivity segments with practical control significance by introducing weighted fuzzy C-means clustering in a two-dimensional operating phase space and combining it with the ranking of the actual dynamic response sensitivity of each region after clustering. The specific steps are as follows: 1) Perform fuzzy C-means clustering on the two-dimensional running phase space sample set.

[0037] The total number of cluster segments is denoted as , This indicates the number of gain-sensitive segments, with a preferred value of 4, corresponding to the extremely sensitive region, sensitive region, transition region, and strong buffer zone. The fuzziness index is denoted as... , This value is used to control the degree of ambiguity in the membership of a sample to multiple cluster centers, and can be set to 2.0.

[0038] To enhance the contribution of cooling water conductivity to segmentation, a weighting coefficient is applied to the cooling water conductivity dimension when calculating the distance between sample points and cluster centers. , The distance magnification factor representing the conductivity dimension of cooling water can be taken as 1.2 to 1.8. Specifically, it is calculated based on the sample points. To the cluster center When the distance is , the standard Euclidean distance is To amplify the impact of cooling water conductivity, the distance metric was modified. Defined as ,in and The first The coordinates of each cluster center in the bias dimension and conductivity dimension. Pick arrive At this time, it will force the clustering algorithm to segment the data more precisely based on the difference in conductivity.

[0039] In the specific implementation, initialization is first performed in the two-dimensional operational phase space. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 11. 12. 13. 14. 15. 16. 17. 18. 19. 10 ... Alternatively, iteration can be stopped when the number of iterations reaches 100. Based on this, the resulting clustering is not a hard partition, but rather each sample point has a membership degree between 0 and 1 for each cluster segment, which facilitates the subsequent smooth marking of the working condition boundaries.

[0040] 2) Calculate the representative local titration gain within each cluster segment.

[0041] Representative local titration gain is denoted as , Indicates the first The average sensitivity of "actual pH change caused by dosing valve position change" within each cluster segment This represents the cluster segment index, with values ​​ranging from 1 to... .

[0042] In practical implementation, instead of directly using single-point ratios, adjacent sampling points are selected within the same cluster segment. Only sample pairs that meet the condition of "no change in the set pH value and the change in the dosing valve position reaching the minimum threshold" are retained before calculating the local titration gain. The minimum threshold is denoted as... , This represents the minimum change in the dosing valve position set to avoid an excessively small denominator; it can be taken as 0.5% of the valve position.

[0043] In one embodiment, as an example, the threshold If the change in the dosing valve position command between two adjacent sampling points... Because it is less than This means the sample does not participate in the local titration gain. The calculation is performed to prevent the calculated gain value from being artificially high due to slight valve position fluctuations or measurement noise.

[0044] In one implementation, the first The calculation method for the local titration gain corresponding to each sample pair is expressed as follows: ; in, Indicates the first The actual pH value at each sampling point; Indicates the first Local titration gain for each sample pair; This represents a very small constant, used to prevent the denominator from being zero; it can be taken as... .

[0045] In this implementation, further, all valid sample pairs within the same cluster segment are... Sort the values ​​from smallest to largest, remove the top 5% and bottom 5% extreme values, and then average the remaining values ​​to obtain the representative local titration gain of that cluster segment. This reduces the impact of measurement noise and occasional valve vibration on section sequencing.

[0046] 3) Reorder all cluster segments from largest to smallest based on representative local titration gain, and generate gain sensitivity level labels. The gain sensitivity level labels are denoted as... , Indicates the first The sensitivity level of the operating segment to which each sampling point belongs; the smaller the value, the more sensitive the operating segment.

[0047] In one implementation, when When the four level labels are reordered, they can be defined as Level 1, Level 2, Level 3, and Level 4 segments, respectively. Level 1 segment represents the segment with the largest representative local titration gain, and Level 4 segment represents the segment with the smallest representative local titration gain.

[0048] In one embodiment, for example, if the average dosing valve position change of effective sample pairs within a certain cluster segment is approximately 2%, and the average actual pH value change is approximately 0.08, then the representative local titration gain of that segment is approximately 0.04 pH units / valve position percentage; if the corresponding result for another cluster segment is approximately 0.01 pH units / valve position percentage, then the former segment should be marked with a higher sensitivity level. Based on this, the cluster number, which originally only had mathematical significance, is converted into a level label with control significance.

[0049] 4) For each sampling point, compare its membership degree to each cluster segment, and write the sampling point into the gain sensitivity level label corresponding to the cluster segment with the highest membership degree to obtain the structured running sample set.

[0050] In one embodiment, for example, suppose the sampling point After fuzzy C-means clustering, the membership vector for each segment is: The membership degree of this point to segment 2. The highest, therefore the gain sensitivity level label for this sampling point. The level corresponding to the second segment after sensitivity sorting (e.g., level 2 segment).

[0051] The structured runtime sample set is denoted as , This represents a set of historical operating samples with operating condition labels.

[0052] Based on this At least include the first The set pH value at each sampling point Actual pH value pH deviation value Cooling water conductivity Dosing valve position command and gain sensitivity level label .

[0053] It should be noted that this step does not simply use fuzzy clustering, but combines the "pH deviation value-cooling water conductivity two-dimensional operating phase space" with the "representative local titration gain ranking" so that the segment division results directly correspond to different nonlinear control sensitivities. This is achieved by incorporating the difference in cooling water buffer capacity into the control tuning process, so that subsequent controller parameter optimization can distinguish between high-sensitivity and low-sensitivity operating conditions, and avoid low-risk operating conditions from excessively dominating parameter results.

[0054] In one embodiment, such as Figure 3 As shown, the analysis of the weighted fuzzy C-means cluster center and segment boundary is based on the two-dimensional phase space scatter plot, with the cluster center of the weighted fuzzy C-means cluster (marked by the black "X" symbol) and the segment boundary line (dark blue dashed line) determined by the maximum membership degree superimposed. Figure 3 The dashed lines outline the approximate boundaries of the four gain sensitivity segments, while the cluster centers represent the typical state locations of each segment. Weights are applied to the conductivity dimension during the clustering process. Subsequently, the segmentation more precisely reflects the differences in cooling water buffer capacity, avoids the dominance of pH deviation dimension on clustering results, and embodies the key processing step of the present invention, "incorporating changes in cooling water conductivity into clustering criteria," providing an intuitive basis for the subsequent extraction of sensitivity level labels with control significance.

[0055] S2. Online identification of time-varying object model of generator stator cooling water pH controlled process. The pH change of generator stator cooling water is affected not only by the position of the dosing valve, but also by changes in flow rate, water replenishment, measuring point location, chemical delivery lag, and buffer capacity. Therefore, a fixed-object model is difficult to accurately characterize the controlled process over a long period. If a single fixed-object model is still used for subsequent parameter optimization, the obtained controller parameters are likely to become out of touch with the current field conditions.

[0056] This invention represents the pH controlled process online as a discrete object model with "first-order inertia plus pure time delay," and uses the rate of change of cooling water conductivity to adaptively adjust the forgetting factor of the recursive least squares algorithm, enabling the object model to be updated according to changes in the current operating conditions. The specific steps are as follows: S201, Determining the number of pure time delay steps 1) Determine the controller sampling period. The sampling period is denoted as... , This indicates the time interval between two consecutive calculations by the controller, for example, 2 seconds.

[0057] It should be noted that the same sampling period is used for subsequent pure time delay step calculation, discrete model identification, and particle swarm closed-loop simulation to ensure that the time base of the model and the control algorithm are consistent.

[0058] 2) Estimate the pure delay time based on the effective delivery volume of the dosing pipeline and the current cooling water flow rate.

[0059] The effective delivery volume of the dosing pipeline is denoted as , This represents the actual volume of the drug solution transported between the dosing point and the pH measuring point that has a delaying effect; the unit can be liters. The current cooling water flow rate is denoted as... , This indicates the effective flow rate of the main cooling water circuit or the relevant branch circuit at the measuring point under the current operating conditions, and the unit can be liters per second.

[0060] In practical implementation, the prior estimate of the pure time delay is denoted as... ,according to Calculated.

[0061] 3) Determine the number of pure time delay steps based on the prior estimate of the pure time delay and the sampling period. The number of pure time delay steps is denoted as... , This represents the number of discrete sampling steps from the output of the dosing valve position command to the point where an observable change in the actual pH value begins. The calculation method is expressed as follows: ; in, This indicates the rounding operation.

[0062] In one embodiment, for example, if the effective delivery volume of the dosing pipeline... The current cooling water flow rate is 120 liters. If the value is 2 liters per second, then the prior estimate of the pure time delay is... The sampling period is 60 seconds; if the sampling period is... If we take 2 seconds, then the pure lag steps are... The value is 30. This result indicates that the number of... The dosing valve position command output at each sampling time needs to be received at the next sampling time. The main impact on the actual pH value will only begin to be reflected at a specific sampling time.

[0063] It should be noted that, to avoid the pure time delay step count being fixed for a long time, causing the discrete object model to gradually deviate from the field transmission delay, the pure time delay step count is recalculated during online operation when a significant change in cooling water flow, a switch in water replenishment mode, or a switch in pH measurement points is detected. .

[0064] S202, Adaptive recursive least squares identification based on the rate of change of cooling water conductivity 1) Perform a light smoothing process on the actual pH value sequence to obtain a smoothed actual pH value sequence for identification.

[0065] Smooth the actual pH value as , This represents the actual pH value after removing high-frequency measurement jitter. In practice, a 3-point median filter can be used; when further reduction of random noise is required, a 3-point moving average can be performed after median filtering. Based on this, the active dynamics of the controlled process will not be significantly disrupted, but the interference of noise on parameter identification can be reduced.

[0066] 2) Establish a first-order discrete object model with inertia and pure time delay.

[0067] No. The identification output corresponding to each sampling point is denoted as . , Indicates the first The smoothed actual pH value of each sampling point, i.e. .

[0068] The discrete regression parameter vector is denoted as , Indicates the first The first-order object regression parameters at each sampling point are defined as follows: ,in, It represents the inertial dispersion coefficient, reflecting the speed of the process response; This represents the input action coefficient, reflecting the strength of the effect of changes in the dosing valve position on the actual pH value. This represents the bias term, used to absorb slow drift and unmodeled constant disturbances; This indicates the transpose operation.

[0069] In practical implementation, y(i) is obtained based on the discrete object model, and the calculation formula of the discrete object model is expressed as: ; in, Indicates the first The identification output corresponding to each sampling point Indicates the first Dosing valve position command for each sampling point.

[0070] 3) Construct the regression vector required for recursive least squares identification.

[0071] No. The regression vector of each sampling point is denoted as... , The input data vector used by the recursive least squares algorithm to update the parameters can be defined as follows: .

[0072] Furthermore, the recursive covariance matrix is ​​denoted as... , The covariance matrix represents the uncertainty of parameter estimation and is used to characterize the reliability of the current parameter results.

[0073] 4) The forgetting factor is adaptively set according to the rate of change of cooling water conductivity.

[0074] The rate of change of the conductivity of cooling water is denoted as , The absolute rate of change of cooling water conductivity between two adjacent sampling points can be expressed as follows: The calculation yielded that, Indicates the first Cooling water conductivity at each sampling point.

[0075] The forgetting factor is denoted as , This represents the weighting of new and old data in the recursive least squares algorithm. A larger value indicates a change in buffer capacity, making object characteristics more likely to migrate. In this case, the forgetting factor should be reduced to increase parameter tracking speed. When the value is relatively small, it indicates that the operating conditions are relatively stable. In this case, the forgetting factor should be increased to reduce parameter fluctuations.

[0076] In one implementation, two thresholds can be set. and .when season ;when season ;when When the value lies between the two, a linear interpolation is performed on the forgetting factor. This method is simple to implement and easy to put into practice.

[0077] 5) The regression parameter vector is updated point by point using a recursive least squares algorithm with an adaptive forgetting factor. and recursive covariance matrix .

[0078] In practical implementation, it is based on the recursive least squares algorithm. Using the recursive least squares algorithm, from the previous step... The updated values ​​reflect the uncertainty in the parameter estimates. Initially... Set as a larger diagonal matrix (e.g.) The recursive process is represented as follows: ; in, Represents the identity matrix; Indicates the first The recursive covariance matrix updated for each sampling point reflects the uncertainty of parameter estimation at the current time. Indicates the first The recursive covariance matrix of each sampling point; This represents the initial value of the recursive covariance matrix; express The transpose of .

[0079] In practical implementation, the initial parameter vector can be set to empirical values ​​close to the rated operating conditions, and the initial recursive covariance matrix can be set to a large diagonal matrix, for example, with 1000 diagonal elements, so that the algorithm retains strong self-adjustment capability in the initial stage. If the obtained value at a certain moment... If it is less than or equal to 0 or greater than or equal to 1, then The value is cropped to between 0.01 and 0.99 before being used for subsequent calculations, thus ensuring that the identification result corresponds to a physically achievable first-order inertial process.

[0080] 6) Convert the discrete regression parameter vector into the process gain and process time constant of the current operating condition.

[0081] No. The process gain of each sampling point is denoted as . , This indicates the steady-state amplification capability of changes in the dosing valve position to actual pH changes at the current operating point; The process time constant of each sampling point is denoted as . , It indicates how quickly the controlled process reaches a steady state at the current operating point.

[0082] In practical implementation, process gain and time constant The calculation method is expressed as follows: ; ; in, Representing the natural logarithm, due to the discrete coefficients of the first-order inertial element. Theoretically satisfied ,therefore It is a positive number, ensuring that the time constant is positive and has physical meaning.

[0083] In one implementation, to reduce the impact of identification noise on subsequent optimization, the five most recent sampling points can be used. and Perform another moving average to obtain smoother parameters for the current object model.

[0084] 7) Construct the confidence level of the parameter estimate based on the recursive covariance matrix. The confidence level of the parameter estimate is denoted as... , Indicates the first The reliability of the object model identification result for each sampling point can be set to a value between 0 and 1, with a larger value indicating a higher reliability.

[0085] In one implementation, the trace of the recursive covariance matrix can be calculated first. Then, following the principle that "the larger the trace value, the lower the confidence level", the trace value is mapped to the range of 0 to 1.

[0086] Based on this, the online object model at the current moment can be represented as: , , , .in, This indicates the steady-state gain of the dosing valve position on the pH value under the current operating conditions, expressed in pH / % %. This represents the time constant of the controlled process under the current operating conditions, in seconds.

[0087] It should be noted that, in order to make the controller parameter tuning oriented towards the "object model under the current operating conditions" rather than a fixed historical model, this invention does not simply use the recursive least squares algorithm. Instead, it introduces the rate of change of cooling water conductivity into the forgetting factor adjustment process and directly sends the identified time-varying object model and parameter estimation confidence level into the subsequent parameter optimization stage. Based on this, when water replenishment, load changes, or buffer capacity migration occur, the tuning results can still maintain good adaptability.

[0088] In one embodiment, such as Figure 4 As shown, the adaptive adjustment curve of the forgetting factor is analyzed (a: rate of change of conductivity; b: forgetting factor). It consists of two subplots, (a) and (b), sharing the same horizontal axis (sampling point number). ). Figure 4 Figure (a) shows the rate of change of the conductivity of the cooling water. (Unit: μS / cm / s), and the preset low threshold is marked. and high threshold ; Figure 4 Figure (b) shows the forgetting factor used in the recursive least squares algorithm. (Dimensionless). When the rate of change of conductivity is small, the forgetting factor remains around 0.995 to ensure parameter stability; when the rate of change increases, the forgetting factor rapidly decreases to 0.970, enhancing the ability to track new data. Experiments demonstrate the effectiveness of the present invention's strategy of "adjusting the forgetting factor online according to the rate of change of conductivity," reflecting the algorithm's ability to automatically balance identification speed and parameter smoothness between stable and abrupt changes in operating conditions.

[0089] S3. Controller parameter optimization based on adaptive evaluation of operating conditions, corrosion risk constraints, and feedforward search boundary contraction. After obtaining the structured running sample set and the online object model, the controller parameters are further optimized.

[0090] In this invention, the controller to be optimized adopts a structure of "feedback PID control + cooling water conductivity feedforward compensation". The controller parameter vector is denoted as... , The candidate controller parameter combination corresponding to a particle is defined as follows: ,in, Indicates proportional gain; Indicates the integration time constant; Represents the differential time constant; This represents the feedforward gain of the cooling water conductivity.

[0091] It should be noted that "feedback PID control + cooling water conductivity feedforward compensation" is a commonly used composite control strategy in industry. The feedback PID controller calculates the deviation between the pH setpoint and the actual pH value. This is used to calculate the baseline dosage to eliminate control errors. The cooling water conductivity feedforward compensation channel independently monitors the cooling water conductivity. Changes: When cooling water replenishment causes a decrease in conductivity (i.e., ion concentration dilution and weakened buffering capacity), this channel increases the opening of an additional dosing valve proportionally before the pH deviation occurs, in order to counteract the pH disturbance caused by the dilution effect. The two outputs are added together to form the final dosing command.

[0092] However, traditional methods often use minimizing single-condition error as the optimization objective, which can easily cause parameters to be biased towards a certain type of local operating conditions.

[0093] This invention constructs multiple differentiated typical working conditions for closed-loop simulation, and optimizes the particle swarm optimization by integrating gain sensitivity level labels, parameter identification confidence, alkaline overshoot corrosion risk penalty terms, and a dynamic search boundary contraction strategy based on feedforward jitter index into the particle swarm fitness function. This ensures that the obtained parameters balance tracking accuracy, material safety, and actuator lifespan. The specific steps are as follows: S301. Construction of Typical Operating Scenarios and Closed-Loop Simulation of Candidate Controllers Traditional optimization methods based on minimizing error under a single operating condition often result in parameters that perform well only within a specific conductivity or deviation range. When the system operates in another gain-sensitive region (e.g., switching from a strong buffer zone to a high-sensitivity region), the original controller parameters are prone to overshoot or hysteresis.

[0094] This invention does not directly simulate all historical data. Instead, it first extracts a limited number of representative typical operating scenarios from a structured operational sample set. By comprehensively evaluating the closed-loop performance of the controller under these significantly different scenarios during the optimization process, the particle swarm optimization algorithm is forced to find parameter combinations that are robust to all types of operating conditions, rather than favoring a local optimum. The specific steps are as follows: 1) From the structured running sample set Multiple typical working conditions are extracted by time window.

[0095] The total number of typical working scenarios is denoted as , This indicates the number of operating scenarios used for fitness assessment, with a preferred value of 5. The 5 typical operating scenarios can correspond to the highly sensitive steady-state operating condition, the rated operating condition, the strong buffer operating condition, the water replenishment disturbance operating condition, and the conductivity change operating condition, respectively.

[0096] Each typical operating scenario retains the set pH value trajectory, cooling water conductivity trajectory, gain sensitivity level label trajectory, and dosing valve position history trajectory within this window.

[0097] In one embodiment, for example, taking the "water replenishment disturbance condition" as an example, the trajectory within the scenario window might be: setting the pH value. Keep Unchanged; Cooling water conductivity The trajectory is From within seconds linearly decreasing to Gain sensitivity level label The corresponding change is from a level 3 section to a level 2 section; historical dosing valve position commands. To counteract the dilution effect, the valve opening is adjusted from... Gradually rise to .

[0098] 2) Configure object model parameters and equivalent disturbance sequences for each typical working condition scenario. The average process gain for a typical operating scenario is denoted as . The average process time constant is denoted as Pure delay steps are denoted as The confidence level of parameter estimation is denoted as .

[0099] In practical implementation, the online identification results can be evaluated within the time window corresponding to the specific working condition scenario. , and The average values ​​are taken separately and used as the object model parameters and confidence level for this working condition scenario.

[0100] In one embodiment, as an example, suppose the "highly sensitive steady-state condition" window includes Each sampling point was used, during which the identifier output... Group and The average process gain in this scenario That is this Arithmetic mean of the values, average process time constant The same calculation applies.

[0101] Furthermore, to make the simulation more closely resemble the actual historical process, an equivalent perturbation sequence is constructed. A typical working scenario in the first The equivalent perturbation of each simulation sampling point is denoted as . , This represents the comprehensive disturbance term in this working condition scenario that is not expressed by the simplified object model.

[0102] In one implementation, the result is obtained by back-calculation using historical measured data. This involves substituting the historical actual pH value, historical dosing valve position, and the average object model of the current working condition into the discrete object model, and saving the remaining error as an equivalent disturbance sequence for subsequent closed-loop simulation playback.

[0103] In one embodiment, as an example, assume that within a historical time window of a "water replenishment disturbance condition," the average object model parameters for that window have been obtained through online identification (e.g., steady-state gain of 0.02 PH / %, time constant of 50 seconds, and pure time delay of 30 steps). Then, two adjacent historical sampling points within that window are extracted, and their data are as follows: No. Step: The historical actual pH value was 9.10, and the historical dosing valve position was 25%. No. Step: The actual historical pH value was 9.08.

[0104] In this embodiment, the first The historical actual pH value of the first step (9.08) and the first step By substituting the historical dosing valve position commands of each step into the known average object model, a "model-predicted" first step can be calculated. The first step is to calculate the pH value, for example, to obtain 9.09. Then, the second step... The difference is obtained by subtracting the model-predicted pH value (9.09) from the actual historical pH value (9.10). That is the first equivalent perturbation of step Based on this, the model prediction error sequence in a continuous historical data is extracted as a "perturbation fingerprint" of the working condition scenario, and superimposed on the model output in subsequent simulations, so that the simulation process can more realistically reproduce the dynamic characteristics of the historical working condition.

[0105] 3) Define the feedback control part of the candidate controller.

[0106] No. A typical working scenario in the first The simulated pH deviation value of each simulated sampling point is denoted as . , This represents the difference between the set pH value and the simulated output pH value, calculated as follows: .in, Indicates the first A typical working scenario in the first The set pH value for each simulation sampling point, i.e., the control target; Indicates the first A typical working scenario in the first The simulated output PH value of each simulation sampling point is the predicted value calculated by the model.

[0107] The feedback control output is denoted as , This represents the increment of the dosing valve position generated by the feedback PID controller.

[0108] In practical implementation, a positional PID calculation method can be used. Specifically, the proportional term is directly used. The integral term is accumulated over the sampling period. The differential term takes the current value. Compared to the previous moment The difference. To prevent the integral from continuing to accumulate when the output is saturated, the integral term is frozen and no longer increases when the dosing valve position command reaches the upper or lower limit and the error direction continues to push towards saturation. Based on this, the long recovery delay caused by integral saturation can be avoided.

[0109] In one embodiment, for example, let the sampling period be... Second, , Second, Seconds. Current deviation Deviation at the previous moment If the sum of the integrals equals 2, then: Proportional term: ; Integral term: ; Differential term: ; Feedback Output .

[0110] 4) Define the feedforward compensation part of the candidate controller.

[0111] No. A typical working scenario in the first The filtered cooling water conductivity at each simulation sampling point is denoted as . , This represents the conductivity of the cooling water after a first-order low-pass filter, used to suppress conductivity measurement noise from directly entering the feedforward channel.

[0112] No. The benchmark value for cooling water conductivity under typical operating conditions is denoted as follows: , This represents the reference level of conductivity obtained at the start time or within a short window of the operating condition scenario, specifically the average conductivity at the start time or during the stable phase of the scenario.

[0113] The feedforward bias value is denoted as , The decrease in the current conductivity of the cooling water relative to the reference value can be defined as follows: When water replenishment causes a decrease in the conductivity of the cooling water, Being positive makes it easier to pass through positive values. Increase the feedforward dosage. Among them, Indicates the first The baseline value of cooling water conductivity under typical operating conditions is used as the reference starting point for feedforward compensation.

[0114] The feedforward control output is denoted as , can be represented as ,in, The feedforward gain is the cooling water conductivity, expressed as % / (μS / cm), representing the percentage increase in valve opening in the feedforward channel for every 1 μS / cm decrease in conductivity.

[0115] In practical implementation, to prevent high-frequency noise in the cooling water conductivity measurement signal from directly entering the feedforward channel and causing valve vibration, it is necessary to modify the original measurement value. Perform a first-order low-pass filter. In the simulation, a simple first-order inertial filter can be modeled, with a filtering time constant of 5-10 sampling periods. The filter output is... It can smoothly reflect the changing trend of cooling water conductivity, rather than instantaneous fluctuations.

[0116] 5) Calculate the total dosing valve position command and perform object model simulation. A typical working scenario in the first The total dosing valve position command for each simulation sampling point is denoted as... , This represents the actuator input after the feedback control output and the feedforward control output are superimposed.

[0117] The main dosing valve position command can be generated by and Add them together and trim them to the allowable range for the project, for example, 0% to 100%. Then... Substituting the discrete object model of this working condition scenario, the simulated output pH value is obtained. .

[0118] In one embodiment, as an example, suppose the feedback output Feedforward output Then the general instruction If the project's allowable range is If so, then 3.7% will be output directly.

[0119] In practical implementation, the following object simulation relationship can be used to demonstrate how the controlled object (cooling water pH value) calculates the simulated pH value at the current moment based on the pH value at the previous moment, the dosing command several steps ago, and the external disturbance at the current moment. This can be represented as: ; in, Indicates the first The first working scenario The simulated output pH value of each simulation sampling step; Indicates the first The first working scenario The simulated output pH value of each simulation sampling step; and This represents the average process gain of the operating scenario. With average process time constant The converted discrete model coefficients, i.e. , , Reflects the magnitude of system inertia. Reflects the intensity of the input effect; Indicates the first In various working scenarios, The total dosing valve position command at any given time. This represents the pure time delay steps in this scenario; Indicates the first The average process gain for a typical operating scenario, in units of (i.e., every change in the dosing valve position) The actual consequences Steady-state change).

[0120] In one embodiment, Figure 5 The online identification parameter variation curves (a: process gain; b: process time constant; c: confidence level) are displayed, consisting of three subplots (a), (b), and (c), sharing a common horizontal axis (sampling point number). (dimensionless), respectively, show the process gain obtained by the online identification of the recursive least squares algorithm. (Unit: pH / %), Process Time Constant (Unit: seconds) and parameter estimation confidence level (Dimensionless, value range 0~1). Figure 5 (a) Figure 5 Figures (b) show that the model parameters are dynamically adjusted as the operating conditions change, proving that the controlled process is time-varying; Figure 5The confidence curve (c) is high when the operating conditions are stable, but decreases when there are disturbances or sudden parameter changes, reflecting the reliability of the identification results. The experimental results demonstrate the ability of this invention to "adaptively adjust the forgetting factor by utilizing the rate of change of cooling water conductivity, enabling the object model to track changes in operating conditions in real time".

[0121] S302, particle fitness calculation for fusion section labeling, identification confidence level, alkaline overshoot corrosion risk, and valve action smoothness. 1) Set operating condition weights for each typical operating scenario. The working condition weights for each typical working scenario are denoted as follows: , This indicates the contribution of the specific operating condition to the overall fitness score. The weight of each operating condition can be obtained by normalizing the frequency of occurrence of each condition over the past 24 hours, or it can be adjusted by operators based on their risk preferences. For example, higher weights can be assigned to highly sensitive operating conditions and water replenishment disturbance conditions.

[0122] In one embodiment, for example, it is assumed that highly sensitive steady-state conditions account for a significant portion of the total operating time within the past 24 hours. Rated operating conditions account for Strong buffering conditions account for Water replenishment disturbance conditions account for Abrupt conductivity conditions account for Without manual correction, the weight vector Can be set to If operators deem the risk of water replenishment disturbance high, they can manually increase its weight, for example, by adjusting it to... .

[0123] 2) Construct sample-level error weights based on the gain sensitivity level labels. A typical working scenario in the first The gain sensitivity level label corresponding to each simulation sampling point is denoted as . The error weight is denoted as... , This indicates the error amplification factor set for the simulation sampling point due to different sensitivity to operating conditions.

[0124] In one implementation, the error weights for level 1, level 2, level 3, and level 4 segments can be set to 1.5, 1.2, 1.0, and 0.8, respectively. Based on this, the same deviation in highly sensitive segments will be given a greater cost in the fitness calculation.

[0125] 3) Construct scene-level confidence weights based on the confidence scores estimated from the parameters. The scene-level confidence weights are denoted as... , Indicates the first The credibility of the object model used for optimization evaluation in a typical working scenario.

[0126] In practical implementation, it can be directly set as follows: , or to Linear scaling is applied to keep it within the range of 0.6 to 1.0. Based on this, when the object model identification in a certain working scenario is not stable enough, the impact of that working scenario on the sub-fitness is appropriately reduced to avoid low-reliability models misleading the search direction.

[0127] 4) Calculate the sub-fitness for each typical working condition scenario. The sub-fitness of a typical working scenario is denoted as: , Represents the controller parameter vector The smaller the control performance evaluation value obtained under this working condition, the better the control quality.

[0128] In its implementation, the sub-fitness consists of three parts: a weighted pH deviation cumulative term, an alkaline overshoot corrosion risk penalty term, and a valve action change penalty term.

[0129] Furthermore, the first A typical working scenario in the first The alkaline overshoot at each simulation sampling point is denoted as... , This represents the net overshoot after the simulated output pH value exceeds the set pH value and allowable tolerance. The alkaline overshoot tolerance is denoted as... , This indicates a small overshoot that is allowed to exist but does not trigger a strong penalty; the preferred value is 0.05 PH units.

[0130] In practical implementation, the calculation method for alkaline overshoot is expressed as follows: .

[0131] Based on this, the calculation method for sub-fitness is expressed as follows: ; Weighted pH deviation cumulative term The calculation method is as follows: ; Alkaline overshoot corrosion risk penalty item The calculation method is as follows: ; Valve action change penalty item The calculation method is as follows: ; in, This represents the penalty coefficient for alkaline overshoot corrosion risk, which is preferably set to 50. This represents the penalty coefficient for changes in valve action, preferably set to 0.5; Indicates the first In a typical working scenario, at the first The total dosing valve position command for each simulation sampling step; It is used to calculate the valve position change between two adjacent control cycles to measure the smoothness of valve operation.

[0132] It should be noted that when calculating sub-fitness, The use of cubic penalty instead of linear penalty is because slight overshoot only requires moderate constraint, while significant overshoot should be rapidly amplified to proactively avoid parameter combinations that may exacerbate corrosion of copper components.

[0133] In one embodiment, for example, if the set pH value of a certain simulation sampling point is 9.10, the alkaline overshoot tolerance... Taking 0.05, when the simulated output pH value is 9.13, the alkaline overshoot is 0 because it does not exceed 9.15; when the simulated output pH value is 9.24, the alkaline overshoot is 0.09, and its cubic penalty term is... The penalty will be significantly greater than that for a small overshoot, thus pulling the optimization direction away from the danger zone.

[0134] 5) The sub-fitness of all typical working conditions is weighted and summed to obtain the particle overall fitness.

[0135] Overall fitness is denoted as , Represents the controller parameter vector The overall control performance evaluation value under all typical operating conditions is calculated as follows: .

[0136] Based on this, each particle corresponds to a comprehensive fitness level that can simultaneously reflect tracking performance, operating condition sensitivity, object model credibility, corrosion risk, and valve action smoothness.

[0137] It should be noted that, in order to expand the control objective from simply minimizing error to a comprehensive objective that takes into account tracking accuracy, material safety, and actuator lifespan, this step incorporates "gain sensitivity level label," "parameter estimation confidence level," "alkaline overshoot corrosion risk," and "valve action change" into the particle fitness calculation process simultaneously. This can significantly reduce the probability of high pH overshoot and high-frequency valve action while suppressing pH deviation.

[0138] S303, Search Boundary Shrinking and Particle Update Based on Feedforward Jitter Index 1) Initialize the particle swarm search space. The number of particles is denoted as... , This represents the number of candidate controller parameter vectors in the particle swarm, preferably 20 to 40.

[0139] right , , and Set the allowable boundaries for the project separately, among which The initial search lower bound can be 0, and the initial search upper bound is given based on historical experience. In addition to position, each particle also stores a velocity vector, its individual optimal position, and the group's optimal position.

[0140] 2) After each generation of fitness evaluation, the feedforward control output sequence is extracted separately and the feedforward jitter index is calculated. The average feedforward jitter index of the generation population is defined as follows: , Indicates the first The degree of overall valve position fluctuation caused by the particle swarm in the feedforward channel.

[0141] In practical implementation, the feedforward control output sequence can be configured for each particle and each typical working condition scenario. Calculate the sum of the absolute values ​​of the differences between adjacent sampling points, and then average them over all particles and all typical working scenarios to obtain the result. The larger this index is, the easier it is for the cooling water conductivity feedforward channel to amplify the fluctuations in cooling water conductivity into high-frequency actions.

[0142] In one embodiment, as an example, suppose a particle's feedforward output sequence in a scene is as follows: (The absolute value of the adjacent difference is) Then the sum of the sequence differences is The average value is obtained by averaging over all particles and the scene. .like This indicates an average feedforward change of 0.25% in the valve position per step.

[0143] 3) Compare the average feedforward jitter index with the preset jitter tolerance upper limit, and dynamically narrow the search upper bound of the cooling water conductivity feedforward gain. The preset jitter tolerance upper limit is denoted as... , This indicates the maximum allowable average jitter level for the feedforward channel.

[0144] In practical implementation, Based on engineering experience, a value of 0.2 can be used (i.e., the average feedforward change per step does not exceed 0.2% of the valve position). Alternatively, the value can be obtained by statistically analyzing the feedforward jitter index under the current stable operating conditions and multiplying the average value during normal operation by 1.2.

[0145] No. The upper bound for the search of the feedforward gain of the cooling water conductivity is denoted as ,when Update the upper bound of the search as follows: ; in, This indicates the operation of retrieving the maximum value; This represents a fixed search lower bound for the feedforward gain of cooling water conductivity. In engineering practice, the feedforward gain is usually a non-negative number, so this lower bound can be set to 0. This indicates the upper bound of the search contraction rate factor, which can be between 0.3 and 0.8.

[0146] Based on the update method of this search upper bound, the more severe the feedforward jitter, the lower the upper bound of the cooling water conductivity feedforward gain allowed for the next generation of search, thereby pulling the search direction from the region that is prone to valve jitter back to a more stable region.

[0147] 4) Redirect the velocity of particles that exceed the new search upper bound and synchronously correct their positions. The generation The positions of the particles in the feedforward gain dimension of cooling water conductivity are denoted as follows: The velocity component is denoted as .

[0148] when Greater than the new upper bound of the search ,and If the value is still positive, it means the particle is continuing to move towards a direction with greater feedforward gain. At this point, its velocity component is reversed and multiplied by the bounce coefficient. Rebound coefficient This represents the percentage retained after the velocity reverses, and can be taken as 0.8.

[0149] Furthermore, after velocity redirection, the feedforward gain component of the particle's cooling water conductivity is truncated to a new upper search bound. If the particle's position exceeds the bound but its velocity is already negative, only position truncation is performed, and the velocity is not reversed a second time to avoid unnecessary oscillations.

[0150] In one embodiment, as an example, suppose the first a certain particle New Upper Realm Current velocity component (Positive direction). Since the position is out of bounds and the velocity is positive, let (Reverse retain 0.8), then truncate the position to 1.0.

[0151] 5) The standard velocity-position update rule of the particle swarm optimization algorithm is used to update each particle. The particle swarm optimization algorithm is an optimization algorithm that searches for the optimal solution by simulating the foraging behavior of a flock of birds.

[0152] In practical implementation, the inertia weight is dynamically set according to the gain sensitivity level of the recent dominant operating conditions: when the recent dominant operating conditions are mainly in level 1 or level 2 segments, the inertia weight can be 0.45 to 0.60 to reduce search oscillations; when the recent dominant operating conditions are mainly in level 3 or level 4 segments, the inertia weight can be 0.65 to 0.85 to retain a strong global search capability.

[0153] After several iterations, the controller parameter vector with the minimum overall fitness is obtained, denoted as... , This represents the optimal combination of controller parameters obtained through global particle swarm optimization.

[0154] It should be noted that, in order to take advantage of the prior knowledge that the cooling water conductivity feedforward channel is most likely to induce high-frequency actions and actively narrow the dangerous search area, this step does not simply filter out the results of "excessive jitter" after optimization, but directly embeds the valve jitter risk caused by the feedforward channel into the particle swarm search boundary management process. This can significantly reduce the probability of high-frequency valve actions while retaining the ability to compensate for water replenishment disturbances in advance.

[0155] S4. Optimization result verification, local correction, and parameter distribution based on continuous actual disturbance playback. The controller parameter vector obtained by particle swarm optimization This represents the optimal result obtained under multiple typical operating conditions. However, actual field operations may encounter more complex combinations of non-stationary disturbances. Without verification using continuous real disturbance sequences, situations may still arise where "the result is optimal under typical operating conditions but recovery is slow or overshoot is large under field sequences."

[0156] Before outputting the final controller parameters, this invention adds a round of result verification and local correction based on a continuous actual disturbance sequence, so that the parameter results undergo a complete verification close to the actual field operating conditions before being issued. The specific steps are as follows: 1) Extract a continuous sequence of actual disturbances from recent historical operational data as a validation sequence. The length of the validation sequence is denoted as... , This indicates the number of consecutive sampling points used in the verification phase.

[0157] In one implementation, if the sampling period A 2-second interval can be used to select 1800 sampling points, corresponding to 3600 seconds of continuous operation data. This verification sequence prioritizes covering non-stationary events such as water replenishment, load changes, cooling water conductivity fluctuations, and actuator switching, and retains the corresponding set pH value trajectory, cooling water conductivity trajectory, and historical disturbance information.

[0158] 2) Convert the controller parameter vector The control simulator is loaded, the verification sequence is fully replayed, and the overall verification performance index is calculated. The overall verification performance index is denoted as... , Represents the controller parameter vector Overall performance evaluation value in a continuous real-world perturbation environment.

[0159] In practical implementation, the evaluation structure in S302 can be continued to verify the comprehensive performance indicators, that is, to simultaneously calculate the weighted cumulative pH deviation, the cumulative penalty for alkaline overshoot, and the change in valve action. In addition, the recovery time and the maximum overshoot can be recorded as additional safety verification indicators.

[0160] In one embodiment, as an example, suppose that after the verification sequence is replayed: the weighted pH bias is accumulated. alkaline overtuning cubic and (Corresponding to an overshoot of 0.2 cubic meters), valve action changes and , , Confidence level ,but: .

[0161] 3) Compare the overall performance metrics with the acceptable threshold. The acceptable threshold is denoted as... , This indicates the performance threshold that allows controller parameters to be directly sent to the field controller.

[0162] In one implementation, it can be... Set to 1.0 to 1.2 times the average comprehensive performance index under recent stable operating conditions. When If no actual pH value prediction result exceeding the hard protection threshold appears during the verification process, it indicates that the set of parameters meets the requirements of field application and can be directly used as the final output parameters.

[0163] In one embodiment, as an example, assume that the average verification metric calculated during recent stable operation is The acceptable threshold is set to... If the current optimization result And the maximum predicted pH value does not exceed the hard protection limit (e.g. If the verification is successful, the application can be sent directly.

[0164] 4) When At that time, with A small-scale local correction is performed around the center to obtain the locally corrected controller parameter vector. The locally corrected controller parameter vector is denoted as... , This represents the controller parameter results obtained through further optimization based on continuous actual disturbance playback.

[0165] The local search radius vector is denoted as , This indicates the allowable deviation of each parameter during the local correction phase. The maximum range. Preferably, the search radius of each dimension can be set to 5% of the corresponding global search range.

[0166] The goal of local correction is not to find the "global average optimum" again, but to make a small-scale correction that closely matches the current field disturbance environment based on the parameter results that are about to be issued. Therefore, in the specific implementation, the local correction stage will not return to all typical working conditions, but will only use the current verification sequence as the evaluation basis. The local correction algorithm can adopt a simplified version of particle swarm optimization, with the number of particles reduced to 10 and the number of iterations limited to 20 generations.

[0167] 5) Perform verification again on the locally corrected controller parameter vector. If verification passes, output the final controller parameter vector. The final controller parameter vector is denoted as... , This indicates the final combination of controller parameters sent to the field controller.

[0168] When no local correction is needed, you can directly set... When a partial correction is performed and verification passes, it can be set to... .

[0169] Before the parameters are issued, two security checks need to be performed: one is to check , , and Whether all are within the allowable boundaries of the project; another is to check whether the predicted pH value exceeds the hard protection threshold during playback. After passing the safety check, Write the data to the field controller and keep the current integral state continuous to avoid introducing additional disturbances by switching parameters.

[0170] It should be noted that, in order to avoid the disconnect between "optimal comprehensive results under typical working conditions" and "optimal results under continuous real disturbances", this invention uses a representative continuous real disturbance sequence to perform a final correction on the results without abandoning the global optimization results. This further improves the ability of the controller parameters to suppress non-stationary disturbances in real industrial environments, and achieves unified optimization in two aspects: "reducing the risk of alkaline overshoot corrosion" and "reducing the frequency of chemical dosing valve operation".

[0171] In one embodiment, Figure 6 The convergence process of the fitness in particle swarm optimization is analyzed, demonstrating the comprehensive fitness of the particle swarm optimization algorithm in the search for optimal controller parameters. The curve shows the change with the number of iterations. The horizontal axis represents the number of iterations, and the vertical axis represents the overall fitness (dimensionless; a smaller value indicates better control performance). The curve exhibits a rapid decline followed by a gradual flattening, proving that the particle swarm optimization algorithm can effectively find the parameter combination that minimizes the fitness. This fitness function integrates a weighted pH deviation, an alkaline overshoot corrosion risk penalty term, and a valve action change penalty term.

[0172] S5. Online operation process for pH control of generator stator cooling water After completing the controller parameter tuning, output the final controller parameter vector. The online operation process for deployment in the generator stator cooling water pH control system is as follows: 1) By sampling period The system reads the set pH value, actual pH value, cooling water conductivity, and dosing valve position feedback value in real time, and calculates the current value using the same method as in S1. And the normalized two-dimensional phase space position.

[0173] Furthermore, based on the cluster centers and segment ranking results obtained offline, the gain sensitivity level label of the current sampling point is determined.

[0174] In practical implementation, the distance between the current normalized two-dimensional phase space point and all cluster centers can be compared, and the level label corresponding to the cluster segment with the closest distance or the highest membership degree can be written as the current gain sensitivity level label. Based on this, the control system can know in real time whether it is in a high-sensitivity region or a strong buffer zone.

[0175] 2) Update the pure time delay step count, discrete object model parameters, process gain, process time constant, and parameter estimation confidence level according to the method in S2 to obtain the current online object model.

[0176] Furthermore, retuning trigger conditions can be set, such as: the gain sensitivity level label changing by more than one level compared to the previous period, the cooling water conductivity changing at a rate exceeding a threshold, the parameter estimation confidence continuously decreasing, or the pH deviation failing to converge to the allowable range within a preset time. Only when the trigger conditions are met will the parameter optimization and verification processes in S3 and S4 be initiated to reduce the online computational burden.

[0177] 3) When retuning is not triggered, the controller continues to use the current controller parameter vector; when retuning is triggered and S3 and S4 are completed, the latest final controller parameter vector is used. Replace the original parameters while maintaining continuous integral state. Then, calculate the new dosing valve position command using the "feedback PID control + cooling water conductivity feedforward compensation" method and output it to the dosing actuator.

[0178] In one embodiment, as an example, assuming the current system has triggered retuning and completed optimization, the final controller parameter vector obtained is: ; When running online, the status of a certain sampling period is as follows: Set the pH value to The current actual pH value is pH deviation value ; Cooling water conductivity reference value for Current conductivity after filtering for .

[0179] The calculation process is as follows: Feedback control output Calculated based on the PID algorithm, for example, obtaining ; Feedforward control output According to the formula calculate, ; Final dosing valve position command : .

[0180] In this embodiment, although the current pH value is low (the deviation is positive, requiring medication), the conductivity is also low (indicating weak buffering capacity). The system outputs through the feedback channel... In addition to the basic dosage, an extra amount was added through the feedforward channel. The dosage is adjusted to compensate for the impact of reduced buffering capacity on pH levels. The final instruction output to the valve is... .

[0181] 4) The system automatically repeats the above steps in each sampling cycle to achieve continuous online control of the pH value of the generator stator cooling water.

[0182] Based on this, the controller parameters are no longer fixed values ​​in the long term, but can be adaptively updated according to the current chemical buffer state, the dynamic characteristics of the object and the recent disturbance environment, so as to maintain good tracking performance, low risk of alkaline overshoot and relatively stable valve action characteristics in different operating sections.

[0183] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for optimizing the pH value control of generator stator cooling water, characterized in that, Includes the following steps: S1. Collect historical operating data on the pH regulation characteristics of generator stator cooling water. Construct a two-dimensional operating phase space based on the historical operating data, extract nonlinear gain features, and convert them into sensitivity level labels to obtain a structured operating sample set, specifically: The pH deviation value is combined with the cooling water conductivity to construct a two-dimensional operating phase space, which includes the degree of deviation of the system from the target and the strength of the system's current buffering capacity. In a two-dimensional operating phase space, weighted fuzzy C-means clustering is introduced, and combined with the actual dynamic response sensitivity ranking of each region after clustering, historical samples are divided into gain sensitivity segments with practical control significance. S2. The pH controlled process is represented online as a first-order discrete object model with inertia and pure time delay. The forgetting factor of the recursive least squares algorithm is adaptively adjusted by using the conductivity change rate, which reflects the dynamic changes in water quality, so that the object model can be updated with the current operating conditions. S3. Construct multiple differentiated typical working conditions for closed-loop simulation. During the particle swarm optimization process, design a fitness function that integrates gain sensitivity level labels, parameter identification confidence, alkaline overshoot corrosion risk penalty term, and dynamic search boundary contraction strategy based on feedforward jitter index. Through optimization, the obtained parameters take into account tracking accuracy, material safety, and actuator life. S4. Before outputting the final controller parameters, add a round of result verification and local correction based on continuous actual disturbance sequences, so that the parameter results undergo a complete verification close to the actual field operating conditions before they are about to be issued. S5. The optimal controller parameters, verified and corrected by S4, are sent to the execution system. Combined with real-time water quality indicators, feedforward feedback adjustment is performed to monitor the operating conditions. When deviations occur, the model and parameters are self-optimized to achieve stable pH control.

2. The method for optimizing the pH value control of generator stator cooling water according to claim 1, characterized in that, S2 specifically refers to: Determine the controller sampling period, estimate the pure time delay based on the effective delivery volume of the dosing pipeline and the current cooling water flow rate, and determine the number of pure time delay steps based on the prior estimate of the pure time delay and the sampling period; The parameters are identified by using the rate of change of the conductivity of cooling water through an adaptive recursive least squares method.

3. The method for optimizing the pH value control of generator stator cooling water according to claim 1, characterized in that, S3 specifically refers to: S31. Extract a limited number of representative typical working conditions from the structured operation sample set. By comprehensively evaluating the closed-loop performance of the controller in these significantly different scenarios during the optimization process, the particle swarm algorithm is forced to find parameter combinations that are robust to all types of working conditions, rather than favoring a certain local optimal solution. S32, Calculate particle fitness by integrating section labels, identification confidence, alkaline overshoot corrosion risk, and valve action stability; S33. Search boundary shrinkage and particle update based on feedforward jitter index.

4. The method for optimizing the pH value control of generator stator cooling water according to claim 1, characterized in that, S4 specifically refers to: A continuous sequence of actual disturbances is extracted from recent historical operational data as a verification sequence; The controller parameter vector is fed into the control simulator, the verification sequence is fully replayed, and the overall verification performance index is calculated. The overall performance metrics will be compared with acceptable thresholds. When the overall performance index is verified to be greater than the acceptable threshold, a small-scale local correction is performed with the controller parameter vector as the center to obtain the locally corrected controller parameter vector. The locally corrected controller parameter vector is validated again, and the final controller parameter vector is output after the validation passes.

5. The method for optimizing the pH value control of generator stator cooling water according to claim 1, characterized in that, By combining the pH deviation value with the cooling water conductivity, a two-dimensional operating phase space is constructed, specifically as follows: Read the set pH value, actual pH value, cooling water conductivity, and dosing valve position command from historical operating data according to a uniform sampling cycle; Validity screening of historical operation data is performed, and invalid sampling points with sensor disconnection, obvious boundary violations, duplicate sampling timestamps, and instantaneous device switching are deleted; Calculate the pH deviation value at each sampling point; Normalize the pH deviation value sequence and the cooling water conductivity sequence to obtain the normalized pH deviation value and the normalized cooling water conductivity; The i-th two-dimensional operating phase space sample point is constructed by combining the normalized pH deviation value and the normalized cooling water conductivity.

6. The method for optimizing the pH value control of generator stator cooling water according to claim 1, characterized in that, Weighted fuzzy C-means clustering is introduced into the two-dimensional operating phase space, and combined with the ranking of the actual dynamic response sensitivity of each region after clustering, historical samples are divided into gain sensitivity segments with practical control significance, specifically: Perform fuzzy C-means clustering on the two-dimensional operational phase space sample set; Calculate the representative local titration gain within each cluster segment; All cluster segments are reordered from largest to smallest based on representative local titration gain, and gain sensitivity level labels are generated. For each sampling point, compare its membership degree to each cluster segment, and write the sampling point into the gain sensitivity level label corresponding to the cluster segment with the highest membership degree to obtain the structured running sample set.

7. The method for optimizing the pH value control of generator stator cooling water according to claim 3, characterized in that, S31 specifically refers to: Multiple typical operating scenarios were extracted from the structured operational sample set according to time windows; Configure object model parameters and equivalent disturbance sequences for each typical working scenario; Define the feedback control section of the candidate controller; Define the feedforward compensation part of the candidate controller; Calculate the total dosing valve position command and perform object model simulation.

8. The method for optimizing the pH value control of generator stator cooling water according to claim 3, characterized in that, S32 specifically refers to: Set operating condition weights for each typical operating scenario; Construct sample-level error weights based on gain sensitivity level labels; Construct scenario-level confidence weights based on parameter estimation confidence levels; Calculate the sub-fitness for each typical working condition scenario; The overall particle fitness is obtained by weighted summation of the sub-fitnesses of all typical working scenarios.

9. The method for optimizing the pH value control of generator stator cooling water according to claim 3, characterized in that, S33 specifically refers to: Initialize the particle swarm search space; After each generation of fitness evaluation, the feedforward control output sequence is extracted separately and the feedforward jitter index is calculated. The average feedforward jitter index is compared with the preset jitter tolerance upper limit, and the search upper bound of the cooling water conductivity feedforward gain is dynamically narrowed. For particles that exceed the new search upper bound, redirect their velocity and synchronously correct their positions. Each particle is updated using the standard velocity-position update rule of the particle swarm optimization algorithm.

Images