Accurate aeration method for sewage

Abstract

The invention discloses a precise sewage aeration method, and relates to the field of water, wastewater, sewage or sludge treatment, the method comprises a feed-forward model and a feedback model, the feed-forward model collects historical data of aeration tank operation parameters, environmental parameters and oxygen demand, and a decision tree is iteratively generated to correct errors and determine a target dissolved oxygen value; the feedback model corrects the target dissolved oxygen value according to the aeration tank dissolved oxygen concentration and the mixed liquid suspended solid concentration; and controlling the power of the aeration equipment according to the target dissolved oxygen value. Through common coordination of the feedforward model and the feedback model, accurate aeration of sewage treatment is realized.

Description

Technical Field

[0001] This invention relates to the field of water, wastewater, sewage or sludge treatment, and in particular to a method for precise aeration of sewage. Background Technology

[0002] Some developed countries have developed various intelligent aeration control systems, such as sensor-based aeration control systems, which can monitor wastewater quality parameters in real time and automatically adjust aeration intensity and mode. However, their optimization effect on the synergistic removal of multiple pollutants is limited when treating complex mixed industrial wastewater. Sequencing batch aeration (SBA) processes have achieved certain results in wastewater treatment in specific industries, but their versatility and flexibility are insufficient, making it difficult to adapt to the wastewater treatment needs of different industrial sectors. Domestic industrial wastewater treatment technology has developed rapidly in recent years, with some research institutions and enterprises exploring the application of multi-mode aeration technology. Some wastewater treatment plants have adopted a simple combination of intermittent and continuous aeration modes, but switching aeration modes relies on manual experience and lacks scientific decision-making basis. While there have been some attempts in intelligent aeration control, there are significant shortcomings in process optimization and control strategies for the synergistic treatment of multiple pollutants.

[0003] Chinese Patent Publication No. CN116239233A discloses a precise aeration volume calculation system and its control method, including a data acquisition unit, a data analysis unit, and a blower control unit. From an engineering application perspective, this method expands the applicability of aeration control methods by adjusting the aeration volume in real time based on the wastewater treatment plant's influent conditions and process operating parameters. This makes the control method applicable to more application scenarios, achieving stable operation of the biological treatment section under varying influent conditions, on-demand aeration, and reduced operating costs. However, this method struggles to adapt to changes in microbial oxygen consumption characteristics, still exhibiting problems such as long dissolved oxygen stabilization periods, excessive overshoot, and low aeration volume accuracy. Summary of the Invention

[0004] This invention aims to solve the technical problem of insufficient aeration volume control precision in wastewater aeration processes. Traditional aeration methods often rely on experience to set the aeration volume, failing to accurately adjust based on dynamically changing parameters such as influent water quality and aeration tank operating status. This can easily lead to dissolved oxygen concentrations deviating from process requirements, thus affecting wastewater treatment efficiency or causing energy waste. Furthermore, this invention aims to address the problem that while existing single feedforward control models can predict oxygen demand based on historical data, they lack dynamic response capabilities to real-time operating condition fluctuations. This makes it difficult to cope with sudden changes in oxygen demand caused by the lag and interference of biological tank operating conditions, resulting in a large deviation between the target dissolved oxygen value and the actual demand. Furthermore, this invention aims to solve the problem of single... While feedback control models can adjust aeration rates based on real-time dissolved oxygen, they lack the ability to predict future changes in oxygen demand. Relying solely on current deviations for adjustment can lead to lag, and they fail to adequately incorporate key parameters such as mixed liquor suspended solids concentration for correction. This makes it difficult to adapt to changes in microbial oxygen consumption characteristics, resulting in long dissolved oxygen stabilization periods and excessive overshoot. Furthermore, this invention aims to address the shortcomings of existing models in parameter optimization and performance verification. The determination of decision tree parameters and feedback control parameters is easily influenced by subjective experience, and the lack of specific verification for abnormal operating conditions in wastewater aeration scenarios leads to weak model generalization ability and a decline in control accuracy over long-term operation.

[0005] To address the aforementioned technical problems, this invention employs the following technical solution: a precise wastewater aeration method, comprising a feedforward model and a feedback model. The feedforward model collects historical data on aeration tank operating parameters, environmental parameters, and oxygen demand, iteratively generating a decision tree to correct errors and determine the target dissolved oxygen value. The feedback model corrects the target dissolved oxygen value based on the dissolved oxygen concentration and mixed liquor suspended solids concentration in the aeration tank. The power of the aeration equipment is controlled according to the target dissolved oxygen value. By combining the feedforward and feedback models, the feedforward model determines the target dissolved oxygen value in advance based on historical data, while the feedback model dynamically corrects the target value based on real-time dissolved oxygen and mixed liquor suspended solids concentration, achieving precise control of aeration volume. This avoids the lag defect of the feedforward model and compensates for the insufficient prediction of future operating conditions by the feedback model, effectively improving wastewater aeration efficiency and reducing the energy consumption of the aeration equipment.

[0006] Preferably, before building the decision tree, the feedforward model traverses all operating parameters and environmental parameters of the aeration tank, trying different partitioning methods for each parameter. For each partitioning method tried, the model calculates the total reduction in oxygen demand prediction error between the two sets of data after the partition. By traversing parameters and trying different partitioning methods before building the decision tree, and by calculating the reduction in prediction error to screen key parameters, the feedforward model can accurately identify parameters that significantly affect oxygen demand, avoid irrelevant parameters interfering with model training, improve the accuracy of the model's oxygen demand prediction during wastewater aeration, and provide a reliable basis for subsequent determination of target dissolved oxygen values.

[0007] Preferably, the feedforward model first establishes a decision tree for the parameter that maximizes the total reduction in oxygen demand (OD) prediction error. Starting from the second decision tree, each new decision tree corrects the prediction error of the previous model. The first decision tree focuses on the parameter with the largest error reduction, and subsequent decision trees specifically correct preceding errors, making the model iteration process more directional. This gradually reduces the OD prediction deviation, enhances the model's ability to capture complex parameter relationships in wastewater aeration scenarios, further improves OD prediction accuracy, and ensures the rationality of the target dissolved oxygen value setting.

[0008] Preferably, after every 20-50 decision trees, the feedforward model inputs validation set data into the current model, calculates the error between the predicted and actual oxygen demand (OD) values, and determines whether the error continuously decreases with the increase in the number of decision trees. If not, the maximum depth of the decision trees is reduced. Verifying the error and dynamically adjusting the maximum depth of the decision trees after every 20-50 iterations allows for timely detection of model overfitting issues (such as the error ceasing to decrease). Reducing the depth avoids the model overlearning data noise, ensuring the model's stability and generalization ability under diverse wastewater aeration conditions, and extending the model's effective application period.

[0009] Preferably, after the parameters of the feedforward model are adjusted to minimize the error on the validation set, the model's generalization ability is verified using test set data. If the oxygen demand prediction error on the test set is higher than 10% of the average actual oxygen demand, the process needs to be backtracked to the data preprocessing stage for inspection, or the initialization parameters need to be readjusted and iterative training restarted until the oxygen demand prediction error on the test set is lower than 10% of the average actual oxygen demand. Using a test set error lower than 10% of the average oxygen demand as the passing standard, and backtracking or retraining when the standard is not met, the performance of the feedforward model can be strictly controlled, avoiding the use of unqualified models that lead to deviations in the target dissolved oxygen value. This ensures that the model can accurately adapt to the actual working conditions of wastewater aeration and provides accurate oxygen demand prediction support for aeration volume control.

[0010] Preferably, when the feedback model preprocesses the collected data, it uses a moving average method to take the average dissolved oxygen value over 5-10 minutes as the current valid actual value. If the dissolved oxygen data at a certain point is missing for 1-3 minutes, it is filled by interpolation using data from adjacent points during the same period. If the missing data exceeds 5 minutes, it switches to data from a backup monitoring point. Furthermore, abnormal operating condition data in the historical data are labeled and categorized. Using moving averages, tiered processing of missing values, and labeling of abnormal data for dissolved oxygen data can eliminate data noise and missing data caused by water flow disturbances and equipment failures, ensuring the data quality input to the feedback model, avoiding frequent adjustments or inaccuracies in aeration volume due to data anomalies, and improving the control stability of the feedback model.

[0011] Preferably, during the parameter tuning process of the feedback model, the dissolved oxygen control target value is first set according to the process requirements of the biological tank and dynamically adjusted according to the influent load. Then, the relevant parameters are tuned sequentially by trial and error combined with historical data. After tuning, the parameters are verified using historical abnormal data to ensure that the dissolved oxygen can return to the target range within 5-10 minutes and that the aeration rate is not over-adjusted. Dynamically setting the dissolved oxygen target value and tuning the parameters by combining trial and error with historical data, followed by verification with abnormal data, allows the parameters to adapt to different process requirements and influent load changes in the biological tank, ensuring that the dissolved oxygen quickly returns to the target range without over-aeration. This satisfies the oxygen requirements of microbial metabolism, avoids energy waste, and improves the wastewater treatment effect and economy.

[0012] Preferably, when adjusting the proportional parameters, the deviation data between real-time dissolved oxygen data and the target value is continuously collected, the rate of deviation reduction and overshoot are observed, and the proportional parameters are adjusted according to the rate of deviation reduction and overshoot. When adjusting integral-related parameters, collect dissolved oxygen static deviation data and observe the elimination of deviation under the action of integration to avoid oscillation; When adjusting the differential correlation parameters, rapid changes in dissolved oxygen are collected, and the matching degree between the rate of deviation change and the adjustment amplitude is monitored to determine the optimal differential parameters. For proportional, integral, and differential correlation parameters, the deviation reduction rate, static deviation, and rapid change data are monitored and adjusted respectively. This allows for precise optimization of each parameter's function, enabling the proportional stage to quickly control deviation, the integral stage to eliminate static errors, and the differential stage to resist interference, further improving the feedback model's control accuracy and response speed for dissolved oxygen.

[0013] Preferably, the feedback model embeds the tuned parameters into the control system. The control system receives the deviation between the calculated actual dissolved oxygen value and the target value in real time after pretreatment. The deviation signal is processed by relevant links and converted into an aeration rate adjustment signal. Embedding the tuned parameters into the control system and processing the deviation to generate the adjustment signal in real time enables efficient linkage between the feedback model and the aeration equipment, shortens the response time for deviation processing and aeration rate adjustment, avoids dissolved oxygen deviation from the target value due to signal transmission or processing delays, and ensures real-time and accurate control of the wastewater aeration process.

[0014] Preferably, the feedback model statistically analyzes the percentage of time dissolved oxygen remains within the target range and the frequency of aeration rate adjustments during operation. If dissolved oxygen fluctuates frequently or the overshoot exceeds the limit, the parameters are fine-tuned based on the latest historical data. Real-time statistical analysis of dissolved oxygen compliance rate and adjustment frequency, along with dynamic parameter fine-tuning, can promptly address the decrease in model adaptability caused by changes in wastewater aeration conditions (such as fluctuations in influent water quality and seasonal temperature changes), avoid dissolved oxygen fluctuations or over-adjustment, maintain the long-term control accuracy of the feedback model, and ensure stable wastewater treatment results.

[0015] This invention effectively solves the problem of insufficient precision in traditional wastewater aeration control by synergistically applying feedforward and feedback models, significantly improving the accuracy and adaptability of aeration volume control. The feedforward model iteratively generates a decision tree based on historical operating parameters to correct errors, enabling it to predict changes in oxygen demand and determine the initial target dissolved oxygen value in advance, avoiding aeration lag caused by a lack of prediction. The feedback model dynamically corrects the target value by combining real-time dissolved oxygen concentration and mixed liquor suspended solids concentration. Through systematic parameter tuning and data preprocessing, it accurately addresses the lag and interference of biological tank conditions, effectively suppressing dissolved oxygen overshoot and fluctuations, ensuring that microbial metabolism is always in a suitable aerobic environment. This improves wastewater treatment efficiency (such as organic matter degradation rate and nitrification efficiency) while avoiding energy waste caused by excessive aeration. Furthermore, through dual performance control of the validation and test sets, special annotation and parameter optimization of abnormal operating condition data, and dynamic monitoring and parameter updates during operation, the model has significantly enhanced its generalization ability and long-term stability, reduced control failures caused by changes in operating conditions (such as influent impact and seasonal temperature fluctuations), lowered the cost of manual intervention, and provided reliable support for the efficient, stable, and economical operation of wastewater aeration processes. Attached Figure Description

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a graph showing the oxygen demand prediction effect of the feedforward model of this invention. Detailed Implementation

[0018] like Figure 1 As shown, the present invention includes a feedforward model and a feedback model.

[0019] Example 1: This example introduces the feedforward model of the present invention. First, the Gradient Boosting Decision Tree (GBDT) algorithm is selected to construct the feedforward model. This algorithm can effectively capture the nonlinear relationship between oxygen demand and various influencing parameters, and has strong anti-interference ability against data noise, adapting to the complexity of wastewater treatment conditions.

[0020] First, data collection was conducted, focusing on collecting historical data from the past 1-2 years of operation of the aeration tanks in the wastewater treatment plant. This included real-time influent flow rate, chemical oxygen demand (COD), five-day biochemical oxygen demand (BOD5), ammonia nitrogen concentration, total nitrogen concentration, and total phosphorus concentration at the influent end; sludge concentration (MLSS), sludge age (SRT), and mixed liquor return ratio in the aeration tank operating parameters; and environmental parameters such as water temperature and pH value in the aeration tank. At the same time, actual oxygen demand data for the corresponding time period was collected. Actual oxygen demand can be calculated by combining online dissolved oxygen monitoring equipment with aeration tank volume and dissolved oxygen change rate, ensuring that each set of influencing parameter data corresponds to a clear actual oxygen demand value, forming a complete sample dataset.

[0021] Next, data preprocessing is performed. First, the integrity of the historical data for each parameter is checked one by one, and the proportion of missing values ​​for each parameter is calculated. If the proportion of missing values ​​for a parameter is less than 5%, linear interpolation of data from adjacent time points is used to fill the missing values. If the proportion of missing values ​​is between 5% and 15%, the parameter is filled by combining the historical average value of the same season and time period. If the proportion of missing values ​​exceeds 15%, the operation records of the monitoring equipment are traced back to confirm whether it is caused by equipment failure. If the original data cannot be supplemented, the samples with severe missing values ​​should be removed to avoid affecting the model training. Then, outlier detection is performed. A time series trend chart is plotted for each parameter, and the mean and three times the standard deviation of the parameter are calculated. Data exceeding the mean ± three times the standard deviation are marked as suspected outliers. Combined with the actual operating range of the wastewater treatment process (e.g., influent COD typically does not exceed 1000 mg / L, and water temperature is generally between 10-35℃), it is determined whether the suspected outlier is a genuine anomaly. If it is a monitoring error or random fluctuation, the outlier data point is directly removed. If it is a real data change caused by process anomalies (e.g., industrial wastewater impact), the data is retained and marked in the sample set to avoid losing key operating condition information. Afterwards, all parameter data are normalized, compressing the value of each parameter to the [0,1] interval. Specifically, each data point of the parameter is subtracted from its historical minimum value, and then divided by the difference between the historical maximum and minimum values ​​of the parameter, eliminating the influence of differences in units and numerical ranges between different parameters. Finally, the preprocessed dataset is randomly divided into training, validation, and test sets in a ratio of 7:2:1. When dividing the dataset, it is necessary to ensure that the temporal distribution of samples in each set is uniform, and to avoid excessive concentration of samples in a single set during a certain period of time, so as to ensure that the model can learn the patterns under different working conditions.

[0022] The model training phase then begins. First, the basic parameters of the GBDT model are initialized, setting the maximum depth of the decision trees to 5-10 layers (to avoid overfitting), the learning rate to 0.01-0.1 (balancing training speed and accuracy), and the number of decision trees to 100-500. The influencing parameters of the training set are used as input to the model, and the corresponding actual oxygen demand is used as the model's output target value, thus initiating the training process. During training, the model iteratively generates multiple decision trees. Each new decision tree corrects the error of the previous round's predictions, gradually reducing the deviation between the predicted values ​​and the actual oxygen demand. After each iteration, validation set data is input into the model for the current training phase, and the mean absolute error and mean squared error of the model on the validation set are calculated. The model hyperparameters are adjusted based on the error changes. If the validation set error continues to decrease as the number of decision trees increases, the number of decision trees is appropriately increased. If the error starts to rise after decreasing to a certain level, it indicates overfitting, and the maximum depth of the decision trees or the learning rate needs to be reduced. By repeatedly adjusting hyperparameters and comparing the validation set error, the optimal hyperparameter combination is determined. For example, when the number of decision trees is 200, the maximum depth is 8 layers, and the learning rate is 0.05, the model's mean absolute error on the validation set is minimized. At this point, training is stopped, and the optimal model parameters are saved. After training, test set data is input into the optimal model, and the prediction error on the test set is calculated. If the mean absolute error on the test set is less than 10% of the actual average oxygen demand, and the mean squared error meets the process preset requirements, the model training is considered successful. If the error exceeds the allowable range, the process needs to return to the data preprocessing stage to check for missing data or improper handling of outliers, or adjust the hyperparameters and retrain until the model's performance on the test set meets the standards.

[0023] After successful model training, the model enters the practical application and maintenance phase. A data transmission channel is established within the wastewater treatment plant's real-time operation system, enabling the feedforward model to acquire real-time data from online monitoring instruments at the influent end, including current influent flow rate, COD, and ammonia nitrogen concentration, as well as parameters from the aeration tank, such as sludge concentration, water temperature, and pH. Upon receiving this real-time data, the model automatically calls upon the saved optimal GBDT model parameters to quickly calculate the predicted oxygen demand (OD) value under the current operating conditions. This predicted value is then transmitted to the aeration system's control system. Based on the predicted OD value, the control system adjusts the blower's operating frequency or the opening of the aeration valves in advance, changing the aeration airflow to achieve feedforward control of the aeration process and avoid insufficient or excessive aeration due to lag in changes in OD. Simultaneously, a regular model update mechanism is established, collecting the latest operational data (including various influencing parameters and actual oxygen demand) every two months. The new data is processed using the previous data preprocessing methods, merged with existing historical data, and the training, validation, and test sets are redefined. The model is then retrained using the same GBDT algorithm and training process, updating model parameters to ensure it can adapt to changes in influent water quality and process conditions. Furthermore, a model operation monitoring module is set up to compare the deviation between the model's predicted oxygen demand and the actual oxygen consumption in the aeration tank (calculated from online dissolved oxygen monitoring data) in real time. When the deviation exceeds 15% for one consecutive hour, an alarm mechanism is triggered, alerting staff to investigate the cause. If the deviation is due to a decrease in model accuracy, the model update process is initiated in advance; if the deviation is due to monitoring equipment malfunction causing data anomalies, the equipment is repaired promptly to ensure the model's continuous and stable operation.

[0024] The core principle of the feedforward model training layer in this invention is to iteratively generate multiple decision trees to gradually correct prediction errors, adapting to the complex relationship between oxygen demand and multiple parameters in wastewater aeration scenarios. Before training begins, basic parameters are initialized based on the characteristics of wastewater operating conditions. The maximum depth of the decision trees is set to 5-10 layers. This is because wastewater data contains noise such as fluctuations in influent load and changes in water temperature. Decision trees that are too deep are prone to capturing irrelevant details, leading to overfitting, while those that are too shallow cannot explore the nonlinear relationship between oxygen demand and parameters such as COD and ammonia nitrogen. The learning rate is set to 0.01-0.1, which avoids both excessively high learning rates that lead to iterative instability and failure to converge to the optimal solution, and excessively low learning rates that prolong the training cycle, thus meeting the potential need for real-time response in wastewater aeration. The initial number of decision trees is set to 100-500, reserving space for subsequent adjustments through error feedback.

[0025] The training process employs an iterative approach. The first decision tree learns the most significant parameter correlations from the training set, such as the fundamental principle that oxygen demand increases synchronously with COD concentration in wastewater. At this stage, the model's predicted oxygen demand will deviate somewhat from the actual value. Starting with the second decision tree, the core task of each new decision tree is to correct the prediction errors of the previous model. For example, if the first tree did not fully consider the decrease in microbial activity and oxygen demand when water temperature is below 15℃, resulting in an overestimation of the predicted value under low-temperature conditions, the new decision tree will focus on this error range, emphasizing the correlation between water temperature and oxygen demand, and reducing the deviation under this condition by adjusting the branch logic. After each iteration, the newly generated decision tree works in conjunction with all previous decision trees to jointly output more accurate prediction results, gradually narrowing the gap with the actual oxygen demand.

[0026] During training, parameters are dynamically adjusted based on the validation set to ensure the model adapts to diverse operating conditions of wastewater aeration. Every 20-50 decision trees are iterated, the validation set data is input into the current model to calculate the error between the predicted and actual oxygen demand (OD) values. If the error continues to decrease with the increase in the number of decision trees, it indicates that the model is still effectively learning new operating conditions, such as gradually capturing the indirect impact of changes in the mixed liquor recirculation ratio on OD. In this case, the number of decision trees is appropriately increased. If the error starts to rise after decreasing to a certain level, it indicates that the model has overfitted, for example, by treating a random industrial wastewater impact as a regular pattern. In this case, the maximum depth of the decision trees needs to be reduced to simplify the model structure, or the learning rate should be lowered to slow down the iteration step size to avoid overfitting noisy data.

[0027] In the feedforward model for wastewater aeration scenarios, the core logic of the first decision tree in identifying the most significant parameter associations is achieved by evaluating the degree of influence of each parameter on the oxygen demand prediction error.

[0028] Specifically, at the start of training, the model iterates through all input parameters (such as influent COD, flow rate, water temperature, MLSS, etc.) and tries different partitioning methods for each parameter (for example, dividing COD into different intervals based on thresholds such as 200 mg / L and 300 mg / L). For each partitioning method tried, the model calculates the total reduction in oxygen demand prediction error between the two sets of data after the partitioning—the greater the reduction in error, the more significant the impact of that parameter on oxygen demand under that partitioning method.

[0029] For example, in wastewater data, when the model tests the classification of "influent COD," it finds that after classifying by 300 mg / L, the difference in oxygen demand between the high COD group and the low COD group is significant, and the overall prediction error is reduced by 30% compared to before the classification; while when testing the classification of "water temperature," the error is reduced by only 10%. At this point, "influent COD" will be identified as the most significant parameter association and become the first split node of the first decision tree, prioritizing the learning of the correlation between COD and oxygen demand.

[0030] This identification is based entirely on the patterns in the data itself, rather than pre-setting which parameter is more important, and can adaptively capture the most critical influencing factors in the current wastewater scenario.

[0031] Once the parameters are adjusted to minimize the error on the validation set, the model's generalization ability is validated using test set data. The test set includes operational data from wastewater aeration scenarios that were not included in the training, such as sudden increases in influent flow or abnormal fluctuations in MLSS. If the oxygen demand prediction error on the test set is less than 10% of the average actual oxygen demand, it indicates that the model can not only adapt to the trained routine operational conditions but also cope with unseen complex changes, meeting the requirements of wastewater aeration for model stability and adaptability. At this point, the optimal model parameters have been determined in the training layer, and the model can be used for subsequent applications. If the error does not meet the standard, it is necessary to backtrack to the data preprocessing stage to check whether key operational data has been missed, or to readjust the initialization parameters and restart iterative training until the model performance meets the requirements of the wastewater aeration scenario.

[0032] Example 2 introduces the feedback model of the present invention. Considering that the feedback model needs to dynamically adjust the aeration rate in real time according to the control target parameters, and that the operating conditions of the wastewater treatment biological tank have lag and interference, the mature and easy-to-implement PID (proportional-integral-derivative) algorithm is selected to construct the feedback model. The core is to take the dissolved oxygen (DO) concentration in the biological tank as the main control target (because DO directly reflects the balance between oxygen supply and oxygen consumption and is a key indicator for aeration rate adjustment). At the same time, the mixed liquor suspended solids (MLSS) concentration is combined to assist in the correction to ensure the accuracy of the adjustment.

[0033] First, data collection is carried out, requiring the simultaneous acquisition of real-time and historical data. For real-time data, multiple online monitoring devices within the biological tank are used to continuously collect data on DO concentration in each aerobic zone (recorded every 1-2 minutes to avoid single-point data bias; the average of 3-5 monitoring points is taken as the current actual DO value), current aeration parameters (including real-time blower airflow and aeration valve opening), MLSS concentration (collected every 5 minutes to reflect the total microbial population), and real-time COD and ammonia nitrogen concentrations at the influent end (used to determine the interference of influent load fluctuations on DO). Historical data includes DO concentration change curves over the past 6-12 months, aeration adjustment records for corresponding periods, MLSS concentration changes, influent water quality fluctuation data, and DO and aeration correlation data under abnormal process conditions (such as sludge bulking and influent shock). This data will be used for PID parameter tuning and model robustness verification.

[0034] Next, data preprocessing is performed. First, noise in the real-time DO data is addressed: because the online monitoring instrument may experience momentary jumps due to water flow disturbances and sensor contamination (such as a sudden increase in DO from 2 mg / L to 5 mg / L followed by a rapid drop), a moving average method is used to smooth the real-time DO data, taking the average DO value over 5-10 minutes as the current valid actual value to avoid frequent start-stop or large-scale adjustments to aeration volume caused by high-frequency fluctuations. Then, data missing issues are addressed: if the DO data at a certain point is missing for a short period (1-3 minutes), it is filled by interpolation using DO data from adjacent points during the same period; if the missing time exceeds 5 minutes, the data from a backup monitoring point is switched to ensure that the control signal is not interrupted. Finally, abnormal operating condition data in the historical data are labeled and classified (such as a sudden drop in DO caused by a sudden increase in influent COD, or an abnormally low MLSS causing an artificially high DO). These data need to be analyzed separately during parameter tuning to avoid interfering with parameter settings under normal operating conditions.

[0035] The next step is PID parameter tuning. First, the DO control target value is defined: based on the process requirements of the biological tank (e.g., the DO in the aerobic zone needs to be stable at 2-3 mg / L, and can be finely adjusted to 2.5-3 mg / L when nitrification is vigorous), the target DO value for different zones is set, and dynamic adjustment based on the influent load is supported (e.g., when the influent COD increases, the target DO value is temporarily increased by 0.2-0.3 mg / L). Then, the parameters are tuned using a trial-and-error method + historical data backtracking: first, the integral coefficient (I) and derivative coefficient (D) are set to 0, and only the proportional coefficient (P) is adjusted to observe the deviation between the real-time DO and the target value—if the deviation decreases slowly, P is increased appropriately; if the DO overshoot is large (e.g., the target is 2 mg / L, but the actual value rises to 3.5 mg / L), P is decreased until the DO can respond quickly and without significant overshoot. Overshoot, fix the P value; then gradually increase the I coefficient to eliminate static deviation (i.e., DO is consistently stable at 2.1-2.2 mg / L, unable to reach the 2 mg / L target). After each increase, observe whether DO gradually approaches the target value. If oscillation occurs, decrease I until the static deviation disappears; finally, add the D coefficient to suppress the rate of change of DO deviation—for example, if influent shock causes DO to drop rapidly from 2 mg / L to 1 mg / L, the D coefficient can amplify the aeration rate adjustment range in advance to reduce the DO drop. By repeatedly testing different D values ​​and combining the adjustment effects under similar disturbances in historical data, determine the optimal D value; after tuning, use historical abnormal data (such as sudden increase in influent ammonia nitrogen) to verify the parameters, ensuring that DO can return to the target range within 5-10 minutes and that the aeration rate is not over-adjusted.

[0036] First, relevant data for setting the dissolved oxygen target value were collected. The operation revolved around the process requirements of the biological tank and the influent load data: process requirement data for different reaction stages of the biological tank (such as conventional aerobic reaction and the vigorous period of nitrification reaction) were collected to determine the basic target value of dissolved oxygen (2-3 mg / L, 2.5-3 mg / L); at the same time, real-time chemical oxygen demand (COD) and ammonia nitrogen concentration data of the influent were collected, and the target value was dynamically adjusted according to its fluctuations (such as increasing the target value by 0.2-0.3 mg / L when the COD of the influent increases) to ensure that the target value is adapted to the real-time changes in oxygen consumption.

[0037] Secondly, real-time monitoring data is crucial during parameter adjustment. The operation focuses on dissolved oxygen deviation and adjustment effect data: When adjusting proportional parameters, continuously collect real-time dissolved oxygen data and deviation data from the target value, observe the rate of deviation reduction and overshoot (e.g., whether it increases from 2 mg / L to 3.5 mg / L), and increase or decrease the proportional parameter accordingly; when adjusting integral parameters, collect static dissolved oxygen deviation data (e.g., long-term stability at 2.1-2.2 mg / L), observe the elimination of deviation under integral action, and avoid oscillations; when adding derivative parameters, collect data on rapid changes in dissolved oxygen (e.g., a sudden drop from 2 mg / L to 1 mg / L due to influent shock), monitor the matching degree between the deviation change rate and the adjustment amplitude, and determine the optimal derivative parameter.

[0038] Finally, the parameter verification phase involves historical abnormal data. The operation focuses on verifying the adaptability of the operating conditions: extracting historical abnormal operating condition data from the past 6-12 months, including dissolved oxygen change curves and corresponding aeration volume adjustment records under scenarios such as sudden increases in influent ammonia nitrogen and sludge bulking. Substituting the adjusted parameters into this data, the simulation is run and the time for dissolved oxygen to return to the target range (within 5-10 minutes) and aeration volume adjustment range data are collected to verify the effectiveness of the parameters under complex operating conditions and ensure that there is no over-adjustment or under-adjustment.

[0039] In the model building and system linkage phase, the tuned PID parameters are embedded into the biological tank aeration control system: the system receives the pre-treated actual DO value in real time and calculates the deviation between the actual value and the target value; according to the PID algorithm, the deviation signal is processed by a proportional element (directly adjusting the aeration rate according to the magnitude of the deviation, with a larger adjustment amplitude for larger deviations), an integral element (eliminating static deviations based on the cumulative deviation), and a derivative element (predicting and adjusting based on the rate of change of the deviation) and then converted into an aeration rate adjustment signal; at the same time, an MLSS correction mechanism is introduced: when the real-time MLSS concentration exceeds the design range (e.g., ... For concentrations above 4000 mg / L or below 2000 mg / L, the adjustment signal is corrected based on the deviation ratio between the MLSS and the design value. For example, if the MLSS is too high (increased microbial oxygen consumption), the aeration rate adjustment range is increased by 10%-15% to prevent the DO from decreasing due to increased oxygen consumption. The adjustment signal is finally transmitted to the blower control cabinet and the aeration valve actuator to change the blower operating frequency (or number of blowers) and valve opening, thereby realizing real-time adjustment of the aeration rate. At the same time, the system collects the changes in aeration rate and DO after adjustment in real time, forming a closed-loop feedback to continuously optimize the adjustment accuracy.

[0040] In the model operation monitoring and optimization phase, during daily operation, it is necessary to statistically analyze the percentage of time DO is within the target range (it should reach more than 90%) and the frequency of aeration adjustment (avoid adjustments exceeding 5 times per hour). If DO fluctuates frequently or the overshoot exceeds the limit, the PID parameters need to be fine-tuned (e.g., in summer when the water temperature is high and microorganisms consume oxygen quickly, the proportional coefficient can be appropriately increased). Every 3-6 months, the PID parameters should be updated based on the latest historical data (including seasonal changes and influent water quality trends). For example, in winter when the water temperature is low and DO consumption is slow, the integral coefficient can be reduced to avoid excessive aeration. If new abnormal operating conditions occur (e.g., the introduction of new industrial wastewater causing abnormal DO), the data for this condition needs to be added to the historical database, and the parameters need to be readjusted to ensure that the model always adapts to the actual operating conditions.

[0041] Example 3: This example is an embodiment of the feedforward model of the present invention. First, simulation data is prepared, simulating one year of aeration tank operation data from a medium-sized wastewater treatment plant (a total of 8760 sets, one set per hour), covering key parameters affecting oxygen demand: the influent end includes the influent flow rate (range 500-1200 m³ / h). 3 The parameters for aeration tank operation include: COD (150-600 mg / L), BOD5 (80-300 mg / L), ammonia nitrogen (10-50 mg / L), total nitrogen (20-70 mg / L), and total phosphorus (2-10 mg / L); MLSS (2000-4000 mg / L), sludge age (10-20 days), and mixed liquor return ratio (50%-150%); environmental parameters include water temperature (10-35℃) and pH value (6.5-8.5); and corresponding actual oxygen demand data (30-80 kg O2 / h) are generated. Oxygen demand increases with increasing influent load and decreases with decreasing water temperature, forming a sample set that includes characteristics such as normal operating conditions, seasonal fluctuations, and influent shock.

[0042] Next, data preprocessing was performed: missing values ​​for each parameter were counted. 320 sets of total phosphorus data (3.66%) were missing, which were filled using linear interpolation of adjacent two-hour intervals (e.g., for the missing total phosphorus at 8:00 AM on January 5th, the values ​​at 7:00 AM (3.2 mg / L) and 9:00 AM (3.6 mg / L) were used to calculate a filler value of 3.4 mg / L). 1200 sets of ammonia nitrogen data (13.7%) were missing, which were filled using the average ammonia nitrogen value of 28 mg / L from the same time period over the past three years (e.g., Mondays in October each year from 9:00 AM to 10:00 AM). In outlier detection, 15 sets of COD data exceeded the mean ± 3 standard deviations (e.g., COD was 1200 mg / L at 2:00 PM on July 12th). Based on the actual process conditions (design upper limit 800 mg / L), these were determined to be monitoring errors and directly removed. Eight sets of water temperature data were below 10℃ (e.g., water temperature was 8℃ at 3:00 AM on January 20th), which is considered normal low temperature in winter and was retained and labeled "low temperature condition". Normalize all parameters, such as influent flow rate of 500 m³ / h. 3 / h corresponds to 0, 1200m3 / h corresponds to 1, and the flow rate at a certain moment is 850m. 3 / h is then converted to the result of (850-500) / (1200-500). The dataset is divided in a 7:2:1 ratio, with 6132 training sets (covering all seasons and time periods), 1752 validation sets, and 876 test sets, ensuring that each set contains samples of different working conditions.

[0043] The model then entered training: GBDT parameters were initialized, with a maximum decision tree depth of 5 layers, a learning rate of 0.05, and 100 decision trees. The 11 parameters from the training set were used as input, and the actual oxygen demand (OD) was used as output to start training. During training, the first decision tree learned the basic rule that "OD increases with COD," but the prediction error was relatively large. The 50th tree corrected the errors from the preceding trees, capturing the feature that "the decrease in OD becomes larger when the water temperature is below 15℃." The 100th tree further optimized the model, recognizing the detail that "the rate of increase in OD slows down after MLSS exceeds 3500 mg / L." Every 20 trees were evaluated using a validation set. It was found that the error started to decrease with 100 decision trees, reached its minimum with 200 trees, and rebounded with 300 trees (overfitting). Therefore, the number of decision trees was adjusted to 200. Simultaneously, the maximum depth of 6-8 layers was tested, and it was found that the validation set error was lower with 8 layers. Finally, the optimal parameters were determined: 200 decision trees, a maximum depth of 8 layers, and a learning rate of 0.05. The test set was used to verify that the average error was 8.2% of the actual average oxygen demand (below the 10% threshold), indicating that the model training was successful.

[0044] During the model application phase, the influent flow rate of 800 m³ / h was received in real time. 3 Current parameters such as h, COD 350mg / L, and water temperature 22℃ are preprocessed and input into the model. Within 15 seconds, the predicted oxygen demand (O2 / h) of 52kgO2 / h is output and transmitted to the control system, which then adjusts the blower airflow from 40m³ / h in advance. 3 / h adjusted to 45m 3 / h. New data is collected every 2 months (e.g., if the inflow rate suddenly increases to 1300m during the rainy season). 3 The model was retrained by merging the samples (per hour) with historical data and updating the model parameters, reducing the prediction error for the rainy season from 9.5% to 7.8%. During operation monitoring, when the predicted oxygen demand (O2 / h) of 55 kg O2 / h deviated from the actual O2 / h by 10% (not exceeding 15%), it was continuously observed; when the deviation exceeded 15% for 3 consecutive hours, an alarm was triggered. The investigation revealed that the COD sensor was drifting, and after calibration, it returned to normal, ensuring the stable operation of the model.

[0045] like Figure 2As shown in the chart, the feedforward model's oxygen demand (OD) prediction performance over a 30-hour test set is demonstrated. The blue line represents the actual OD, the orange line represents the model's predicted OD, and the gray shaded area represents the prediction error range. The model exhibits relatively small prediction errors under normal operating conditions (e.g., 1-10 hours, 20-30 hours), with an average error of 8.2% of the average actual OD, meeting the threshold requirement of less than 10%, demonstrating the model's accurate prediction capability under stable conditions. Under 10-15 hour influent shock conditions (COD suddenly increases to 550 mg / L) and 15-20 hour low-temperature conditions (water temperature drops to 12℃), the model still responds quickly to changes in OD, with prediction deviations controlled within 12%, verifying the model's adaptability to abnormal conditions. After parameter optimization (200 decision trees, maximum depth 8 layers, learning rate 0.05), the model's prediction curve closely matches the actual curve trend, with no significant lag or overshoot, proving the effectiveness of the model parameter tuning method. After the model was updated (e.g., by incorporating high flow data during the rainy season), the prediction accuracy was further improved, with the error decreasing from 9.5% to 7.8%, demonstrating the role of the model maintenance mechanism in ensuring long-term operational accuracy.

[0046] Example 4: This example is an embodiment of the feedback model of the present invention. First, simulation data is prepared to simulate the operation scenario of the biological tank of a small-to-medium-sized wastewater treatment plant. The target value of DO is set to 2.5 mg / L (the optimal range for nitrification in the aerobic zone), the design range of MLSS is 2500-3500 mg / L, the simulation time span is 72 hours (3 days), and the data acquisition interval is: DO every 2 minutes, MLSS every 5 minutes, and aeration rate (expressed as blower air volume) every 2 minutes. At the same time, the fluctuation of influent load is simulated (indirectly affecting DO consumption through COD and ammonia nitrogen). The generated simulation data includes three types of scenarios.

[0047] Under normal operating conditions (0-24 hours), the influent COD is stable at 300-350 mg / L, ammonia nitrogen at 25-30 mg / L, MLSS at around 3000 mg / L, and the initial DO value is 2.0 mg / L (lower than the target value), corresponding to an initial aeration rate of 30 m³ / L. 3 / h.

[0048] When the influent load suddenly increases (24-48 hours), the COD of the influent rises sharply to 500 mg / L and the ammonia nitrogen rises to 45 mg / L in 24 hours. Microbial oxygen consumption accelerates, and DO drops from 2.5 mg / L to 1.2 mg / L in 1 hour. MLSS briefly rises to 3600 mg / L (microbial proliferation). The aeration rate needs to be dynamically adjusted.

[0049] MLSS abnormality (48-72 hours): At 48 hours, due to a simulated malfunction of the sludge discharge system, MLSS gradually decreased to 2200 mg / L (below the design lower limit), while DO increased to 3.8 mg / L due to reduced microbial oxygen consumption. The aeration rate needs to be corrected by adjusting the MLSS. Next, simulation data preprocessing was performed. For "noise" in the DO data (e.g., simulating sensor fluctuations with two outliers: a sudden change to 4.0 mg / L at 6 hours and a sudden change to 0.8 mg / L at 30 hours), a 5-point moving average method was used: the outlier was taken as the center, and the average of the two preceding and following data points (a total of 5) was used for replacement. For example, the outlier at 6 hours was 4.0 mg / L, with preceding and following data points of 2.1, 2.2, 2.3, and 2.4 mg / L. After replacement, the average of these 5 data points, 2.6 mg / L, was used to eliminate high-frequency fluctuations. For MLSS data missing points: at 55 hours, three consecutive data points were missing (original data 2200, 2180, and 2150 mg / L). Linear interpolation was used to fill these missing points; that is, the first missing point at 55 hours was filled with the average of 2200 mg / L from 55 hours prior and the next valid data point of 2180 mg / L, resulting in 2190 mg / L, ensuring data continuity.

[0050] All data are aligned to the time axis. The MLSS data for every 5 minutes is interpolated to be converted into 1 data point every 2 minutes. This data is then matched with the DO and aeration volume data to form a four-column simulation dataset of "time-DO-MLSS-aeration volume".

[0051] Then, the PID parameters are tuned.

[0052] First, set the integral coefficient (I) and differential coefficient (D) to 0, and adjust only the proportional coefficient (P).

[0053] Initial P=5: Input the DO data for 0-24 hours (target 2.5 mg / L, initial 2.0 mg / L), calculate the deviation between the actual value and the target value as 0.5 mg / L, and adjust the aeration rate proportionally to the deviation according to the proportional adjustment rule. At this time, the aeration rate starts from 30 m³ / L. 3 / h increased to 32.5m 3 / h; observe subsequent DO changes: after 1 hour, DO rose to 2.3 mg / L, the deviation narrowed to 0.2 mg / L, and the aeration rate was further increased to 33.5 m³ / h. 3 / h; however, after 2 hours, the DO exceeded the target value, reaching 2.8 mg / L, and the deviation became -0.3 mg / L, causing the aeration rate to drop to 32 m³ / h. 3 The overshoot of 0.3 mg / L per hour is relatively large, so the P value needs to be reduced. Adjust P=3: Resimulate 0-24 hours, initial deviation 0.5 mg / L, adjust proportionally, aeration rate from 30 m³ / L. 3 / h increased to 31.5m 3 / h; after 1 hour, DO rose to 2.4 mg / L, a deviation of 0.1 mg / L, and the aeration rate was fine-tuned to 31.8 m³ / h. 3 / h; after 2 hours, DO stabilized at 2.5 mg / L with no overshoot, the deviation was eliminated, and P = 3 was fixed. Integral coefficient (I) is added to eliminate static bias.

[0054] Simulating 24-30 hours (after a sudden increase in influent load), DO appeared stable around 2.5 mg / L, but actually remained at 2.4 mg / L, exhibiting a static deviation of 0.1 mg / L. Assuming I = 0.2, the integral adjustment amplitude is proportional to the cumulative sum of deviations over a past period. Therefore, the sum of deviations over the past 10 minutes (5 data points, each with a deviation of 0.1 mg / L, totaling 0.5) is taken. Combined with the I value, the adjustment amplitude is calculated. The final total adjustment amplitude is the superposition of proportional and integral adjustments. The aeration rate starts from 32 m³ / L. 3 / h increased to 32.4m 3 / h; after 1 hour, DO rose to 2.5 mg / L, the static deviation was eliminated, and I = 0.2 was fixed. Adding a differential coefficient (D) suppresses the rate of change of deviation.

[0055] Simulating a sudden increase in influent load over 24 hours, the dissolved oxygen (DO) drops from 2.5 mg / L to 2.0 mg / L within 10 minutes. The rate of change of the deviation is calculated (decreased by 0.05 mg / L per minute). Assuming D = 2, the differential adjustment amplitude is proportional to the rate of change of the deviation. Therefore, the adjustment direction is to counteract the rapid change in deviation, requiring an increase in aeration rate. The total adjustment amplitude is the superposition of proportional control, integral control (currently zero due to no accumulated deviation), and differential control. The aeration rate is increased from 32 mg / L. 3 / h rapidly increased to 33.4m 3 / h; compared with the adjustment effect without D (the aeration volume adjustment range is larger without D, and DO drops to 1.8 mg / L after 30 minutes), after adding D, DO only drops to 1.9 mg / L, and the deviation change range is reduced by 30%. D=2 was determined, and the final PID parameters are P=3, I=0.2, and D=2.

[0056] A feedback model was then constructed and simulation tests were conducted.

[0057] The tuned PID parameters are embedded into the simulation control logic, and the pre-processed simulation data is input in real time.

[0058] Under normal operating conditions (0-24 hours), the model receives DO = 2.0 mg / L (actual value), with a calculation deviation of 0.5 mg / L. The adjustment range is generated through PID control logic, increasing the aeration rate from 30 m³ / L to 31.5 m³ / L. 3 / h; after 12 hours, DO stabilized at 2.5 mg / L, and the aeration rate was maintained at 31.8 m³ / h.3 / h, the DO compliance rate (2.5±0.2mg / L) reached 98%.

[0059] A sudden increase in influent load (24-48 hours) caused a sharp drop in dissolved oxygen (DO) over 24 hours. Within 10 minutes, the model adjusted the output, and the aeration rate rose to 33.4 m³ / h. 3 / h; DO recovered to 2.5 mg / L at 30 hours, with a maximum deviation of 0.7 mg / L during the period and a recovery time of 40 minutes, which was 67% shorter than that without feedback control (recovery time of 2 hours).

[0060] MLSS anomaly (48-72 hours): After 48 hours, MLSS drops to 2200 mg / L (below the design value of 300 mg / L, deviation ratio 12%). The model, based on the PID adjustment range, adds a correction amount of 12% according to the MLSS deviation ratio. For example, at a certain moment, the PID calculation requires a reduction of 0.8 m³ / s aeration rate. 3 / h, after correction the reduction was increased by 12%, and the final aeration rate was 33.4m³ / h. 3 / h decreased to 32.5m 3 / h; after 72 hours, DO decreased from 3.8 mg / L to 2.6 mg / L, returning to the target range.

[0061] Finally, the model was optimized, and the simulation results after 72 hours were statistically analyzed. The DO compliance rate was 92%, and the aeration rate was adjusted 3 times per hour (meeting the requirements). Only in the 60-65 hour period, due to the MLSS continuing to drop to 2100 mg / L, the DO briefly rose to 3.0 mg / L (overshoot 0.5 mg / L), requiring a fine adjustment of the I coefficient to 0.25. After correction, the DO overshoot was reduced to 0.3 mg / L, and the compliance rate increased to 95%. At the same time, the simulation data of this abnormal MLSS scenario was added to the database for subsequent parameter updates to ensure that the model can adapt to more operating conditions.

Claims

1. A method for precise aeration of wastewater, characterized in that, It includes a feedforward model and a feedback model. The feedforward model collects historical data on the operating parameters of the aeration tank, environmental parameters, and oxygen demand, and iteratively generates a decision tree to correct errors and determine the target dissolved oxygen value. The feedback model corrects the target dissolved oxygen value based on the dissolved oxygen concentration and the suspended solids concentration of the mixed liquor in the aeration tank. The power of the aeration equipment is controlled based on the target dissolved oxygen value.

2. The method for precise aeration of wastewater according to claim 1, characterized in that, Before establishing the decision tree, the feedforward model traverses all aeration tank operating parameters and environmental parameters, and tries different partitioning methods for each parameter. For each partitioning of a parameter, the model calculates the total reduction in the oxygen demand prediction error in the two sets of data after the partitioning.

3. The method for precise aeration of wastewater according to claim 2, characterized in that, The feedforward model first establishes a decision tree for the parameter that maximizes the total reduction in oxygen demand prediction error. Starting from the second decision tree, each new decision tree corrects the prediction error of the previous model.

4. A method for precise aeration of wastewater according to any one of claims 1 to 3, characterized in that, Every 20-50 iterations of the feedforward model, the validation set data is input into the current model to calculate the error between the predicted and actual oxygen demand values. The model then determines whether the error continues to decrease as the number of decision trees increases. If not, the maximum depth of the decision trees is reduced.

5. The method for precise aeration of wastewater according to claim 1, characterized in that, Once the parameters of the feedforward model are adjusted to minimize the error on the validation set, the model's generalization ability is verified using test set data. If the oxygen demand prediction error on the test set is higher than 10% of the average actual oxygen demand, the process needs to be backtracked to the data preprocessing stage for inspection, or the initialization parameters need to be readjusted and iterative training restarted until the oxygen demand prediction error on the test set is lower than 10% of the average actual oxygen demand.

6. The method for precise aeration of wastewater according to claim 1, characterized in that, When the feedback model preprocesses the collected data, it uses the moving average method to take the average dissolved oxygen value within 5-10 minutes as the current valid actual value. If the dissolved oxygen data at a certain point is missing for 1-3 minutes, it is filled by interpolation of the data from the same period at adjacent points. If the missing data is missing for more than 5 minutes, it switches to the data from the backup monitoring point. It also labels and classifies the abnormal operating conditions in the historical data.

7. The method for precise aeration of wastewater according to claim 6, characterized in that, During the parameter tuning process of the feedback model, the dissolved oxygen control target value is first set according to the process requirements of the biological tank and supports dynamic adjustment according to the influent load. Then, the relevant parameters are tuned sequentially by trial and error combined with historical data. After the tuning is completed, the parameters are verified by historical abnormal data to ensure that the dissolved oxygen can return to the target range within 5-10 minutes and that the aeration volume is not over-adjusted.

8. The method for precise aeration of wastewater according to claim 7, characterized in that, When adjusting the proportional parameters, continuously collect the deviation data between real-time dissolved oxygen data and the target value, observe the rate of deviation reduction and overshoot, and adjust the proportional parameters according to the rate of reduction and overshoot. When adjusting integral-related parameters, collect dissolved oxygen static deviation data and observe the elimination of deviation under the action of integration to avoid oscillation; When adjusting the differential correlation parameters, rapid changes in dissolved oxygen are collected, and the matching degree between the rate of change of deviation and the adjustment range is monitored to determine the optimal differential parameters.

9. The method for precise aeration of wastewater according to claim 1 or 6, characterized in that, The feedback model embeds the tuned parameters into the control system. The control system receives the deviation between the calculated actual dissolved oxygen value and the target value in real time after preprocessing. The deviation signal is processed by relevant links and converted into an aeration rate adjustment signal.

10. The method for precise aeration of wastewater according to claim 6, characterized in that, The feedback model calculates the percentage of time dissolved oxygen is within the target range and the frequency of aeration adjustment during operation. If dissolved oxygen fluctuates frequently or the overshoot exceeds the limit, the parameters are fine-tuned based on the latest historical data.

Images