Intelligent control method and system for water treatment agent

Abstract

The invention relates to the technical field of water treatment, and discloses an intelligent control method and system for a water treatment agent, and the method comprises the steps: collecting water quality parameters and flow data in real time, dynamically constructing a time sequence characteristic space, and integrating a current parameter value, a historical lagging parameter value and interaction characteristics; preprocessing the data, extracting key interaction features, and constructing a multi-dimensional feature space; the processed data are input into a learning prediction model integrated by three basic learners including a multilayer perceptron, a random forest and support vector regression, the learning prediction model is fused through a dynamic weighting strategy, the weight is adaptively adjusted according to prediction consistency, and a prediction value is output; and adding is controlled according to the predicted value, dynamic correction is carried out in combination with historical deviation, and meanwhile, adding amount range limitation and change rate constraint per minute are implemented. The problems of slow response, high error rate and weak compatibility in the prior art are solved, and the purposes of high control precision, low cost, strong anti-interference capability, high automation level and high response speed are achieved.

Description

Technical Field

[0001] This invention relates to the field of water treatment technology, and in particular to an intelligent control method and system for water treatment agents. Background Technology

[0002] Mine water is a major type of industrial wastewater generated during coal mining, characterized by its complex and variable quality, high suspended solids content, and large fluctuations in turbidity. Coagulation is a key process in mine water treatment. By adding coagulants, colloids and fine suspended solids in the water are aggregated into larger flocs, facilitating subsequent sedimentation and separation. Accurate control of the coagulant dosage directly affects the treatment effect, operating costs, and effluent quality.

[0003] Currently, the following methods are mainly used for controlling the addition of coagulants in mine water: 1. Manual experience-based addition method: Operators manually adjust the amount of coagulant added based on experience and visual observation of water quality. This method relies entirely on the operator's experience and subjective judgment, resulting in low dosage accuracy, difficulty in adapting to rapid changes in water quality, and high labor costs. 2. Single-parameter automatic control method: Based on a single water quality parameter (such as turbidity or pH value), a simple linear or piecewise function relationship is established to automatically adjust the amount of coagulant added. 3. Traditional PID control method: Using a classic proportional-integral-derivative (PID) controller, a certain water quality parameter is used as the control variable to achieve automatic adjustment of coagulant addition. This method has a mature control algorithm, but it is difficult to handle multivariate coupling and nonlinear relationships. 4. Simple neural network control method: Some studies use a single artificial neural network to establish a mapping relationship between water quality parameters and dosage.

[0004] For example, Chinese patent CN120387085A discloses a water treatment chemical dosing verification and prediction system and method, providing the following technical solutions: training a prediction model by extracting characteristic parameters and target parameters of the water plant's chemical dosing process environment; cleaning the target data collected from various collection points during chemical dosing in the water plant using a host computer program; segmenting data by time dimension for different seasons and scaling different input parameters proportionally; inputting parameter changes from different groups into the trained prediction model for real-time prediction of dosing amount; and verifying the system through a simulated inclined tube device to ensure high-precision control during actual chemical dosing. This invention, through the integration of big data, the internet, and AI technologies, achieves accurate prediction and automatic adjustment of chemical dosing, not only improving water quality stability but also reducing alum dosage, achieving the goal of economical and efficient water treatment. However, the aforementioned water treatment chemical dosing verification and prediction system and method are insufficient in terms of processing efficiency, response speed, and understanding of complex queries. Its poor compatibility and limited functionality restrict the overall user experience and practical application scope. Summary of the Invention

[0005] This invention solves the problems of slow response, shallow understanding, high error rate and weak compatibility in the prior art, and proposes an intelligent control method and system for water treatment agents, which achieves the goals of high control accuracy, low cost, strong anti-interference ability, high level of automation and fast response speed.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent control of water treatment agents, comprising: Real-time collection of water quality parameters and flow data, dynamic construction of time-series feature space, and integration of current parameter values, historical lagged parameter values, and interactive features; The data is preprocessed to extract key interaction features and construct a multi-dimensional feature space. The processed data is input into a learning prediction model that integrates three basic learners: multilayer perceptron, random forest, and support vector regression. The model is fused using a dynamic weighting strategy, with the weights adaptively adjusted based on prediction consistency. The model outputs predicted values. The addition of data is controlled based on the predicted values, and dynamic corrections are made in conjunction with historical biases. At the same time, the addition range is limited and the rate of change per minute is constrained.

[0007] It significantly improves control accuracy and system adaptability. By integrating current and historical parameters in a multi-dimensional feature space, it captures dynamic changes and interactive effects in water quality, reducing prediction errors. Combined with ensemble learning and adaptive correction, it ensures that the effluent water quality consistently meets standards. At the same time, dosage limits and rate of change constraints prevent fluctuations, enhance system safety and reagent utilization, and reduce the frequency of human intervention.

[0008] An intelligent control system for a water treatment agent includes: The data acquisition and preprocessing unit is connected to the sensor group and flow meter installed on the inlet pool and pipeline; The integrated learning prediction unit connects to the data acquisition and preprocessing unit, receives the processed data, and outputs the predicted dosage through the model. The intelligent dosing control unit is connected to an integrated learning and prediction unit, which receives the predicted dosing amount and makes dynamic corrections to generate the final control command. The safety interlock module is connected to the intelligent dosing control unit to impose dosing range restrictions and change rate constraints on control commands; the actuator includes a solenoid valve group and a metering pump, the metering pump is connected to the coagulant storage tank, and the safety interlock module is connected to the solenoid valve group and the metering pump.

[0009] It achieves closed-loop intelligent control, and its modular design supports efficient data flow and decision-making. The integrated learning unit combined with safety interlocks ensures a short response time for the complete process from prediction to execution, thereby improving the level of automation. The actuator, such as the metering pump, works with safety constraints to reduce operating costs. At the same time, it is easy to deploy, compatible with existing equipment, and shortens the online cycle.

[0010] Preferably, the dynamic weighted strategy fusion is as follows: first, the standard deviation of the predicted values ​​of the three basic learners, namely, multilayer perceptron, random forest, and support vector regression, is calculated. If the prediction consistency is less than a set value, a fixed weight allocation is adopted, wherein the weight of multilayer perceptron > the weight of random forest > the weight of support vector regression. If the prediction consistency is greater than or equal to the set value, the weight of the model closest to the median is increased.

[0011] It effectively improves prediction stability and anti-interference ability. When the model divergence is small, it prioritizes the nonlinear fitting of the neural network. When the divergence is large, it strengthens the weight of the median model to suppress the influence of noise. The dynamic adjustment based on the standard deviation ensures that the integrated model maintains stable prediction under sensor failure or water quality change, and reduces the overall error growth rate.

[0012] Preferably, the data preprocessing specifically includes: outlier handling, using an improved quartile method to define the effective data range, calculating the interquartile range based on the set quartiles of the data, defining the effective data range, using the standard deviation as a substitute when the interquartile range is zero, and using the moving average method to substitute data exceeding the effective data range; establishing an autoregressive time series model for turbidity parameters, where the current turbidity value is calculated by multiplying the previous turbidity value by a coefficient and adding a random perturbation; establishing a dynamic model for pH values, where the current value is calculated by multiplying the values ​​from the previous two time points by two different coefficients and adding a random perturbation; and establishing a slowly varying model for temperature, where the current value is calculated by multiplying the values ​​from the previous two time points by two different coefficients and adding a random perturbation.

[0013] The preprocessing steps significantly improve data quality and model input reliability. Outlier handling reduces noise interference and maintains data continuity by improving the quartile method and moving average method. Autoregressive time series models are established for different parameters to accurately capture the variation patterns of turbidity, pH value and temperature. An 18-dimensional extended feature space is extracted to enhance the information dimension, provide richer time series features for ensemble learning and enhance prediction accuracy.

[0014] Preferably, the historical deviation is dynamically corrected by a correction coefficient. An initial value is set, and the coefficient is reduced when the actual processing effect is better than expected for several consecutive cycles, and increased when the effect is worse than expected. The correction coefficient is limited to a set range. The final dosage is the product of the predicted value and the correction coefficient.

[0015] The dynamic correction mechanism effectively compensates for systematic biases in the model, adjusts the correction coefficient based on historical biases to ensure long-term control accuracy, and limits the coefficient range to 0.8 to 1.2 to avoid dosing fluctuations caused by over-adjustment. Combined with predicted value output, it reduces reagent waste, improves the stability of effluent water quality, and increases the compliance rate.

[0016] Preferably, the water quality parameters include turbidity, pH value, temperature, and suspended solids concentration; the historical lag parameters include the first-order lag characteristics of water quality parameters and flow data at the current time and the two previous times, as well as the second-order lag characteristics of turbidity and pH value; the key interaction features include the interaction term between turbidity and pH value, the interaction term between suspended solids and temperature, the ratio of turbidity to suspended solids, the square term of pH value deviation, the square root term of turbidity, and discrete features based on flow rate values.

[0017] The model input is comprehensively optimized, and water quality parameters cover key influencing factors, historical lags, and interactive features such as turbidity-pH interaction terms, thereby strengthening the synergistic effect between parameters. An 18-dimensional extended space is constructed to improve information utilization and significantly enhance prediction accuracy, while also adapting to the complex and variable nature of mine water quality.

[0018] Preferably, the prediction consistency is the standard deviation, and the output prediction value needs to constrain the dosage within a set range. If the prediction value is greater than or less than the set range, the corresponding maximum or minimum value of the set range is output.

[0019] The constraint mechanism ensures the safety and efficiency of dosing, quantifies consistency through standard deviation, and automatically limits the range when it exceeds the limit; it prevents waste of reagents due to overdosing or underdosing from affecting the treatment effect, and maintains stable system operation and reduces the risk of secondary pollution by combining the change rate limit.

[0020] Preferably, the processed data includes a basic dosage, which is specifically a weighted sum of water quality parameters and flow data. The specific weighting coefficient is determined through field tests based on the characteristics of mine water quality.

[0021] The basic dosage calculation enhances the model's interpretability and adaptability, and the weighting coefficients are customized based on field tests; it provides an initial prediction benchmark, compensates for ensemble learning bias, simplifies the deployment process, shortens the debugging cycle, and ensures consistent control accuracy in different mine scenarios.

[0022] Preferably, the dynamic correction based on historical deviations also includes performance evaluation and model updates. Specifically, the ensemble learning model is retrained using the most recent date's running data after a set time interval. The training metric is the mean absolute percentage error, which is 1 divided by the number of samples and then multiplied by the absolute value of the difference between the actual value and the predicted value of each sample, divided by the sum of the actual values. When the error exceeds a set value, the model parameters are updated.

[0023] A regular update mechanism ensures long-term high performance of the system, and assessments and monthly retraining adapt to seasonal changes in water quality. When the error exceeds the threshold, the model is automatically optimized to reduce manual maintenance, maintain the mean absolute percentage error, and ensure continuous improvement in prediction accuracy.

[0024] Preferably, the control and dosing also includes automatically switching to a preset value control mode when any sensor malfunctions, maintaining dosing control based on historical average values, and issuing an alarm signal to remind for timely handling.

[0025] The fault handling function greatly improves the system's reliability and continuity. When a sensor fails, it switches to preset value control to avoid processing interruption. Combined with alarm signals, it can quickly respond to maintenance, reduce downtime risks, and ensure that the effluent water quality still meets the standards under abnormal operating conditions.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows.

[0027] 1. This invention employs a multi-dimensional feature space and a heterogeneous ensemble learning strategy, significantly improving the control accuracy of coagulant dosing. By integrating the time-series dynamic changes of multiple key parameters such as turbidity and pH, the model predictions highly match actual needs, avoiding the error accumulation and bias problems of traditional methods. This innovation greatly enhances the stability of effluent quality, significantly reduces turbidity fluctuations, and consistently meets treatment standards, effectively addressing rapid changes in mine water quality. The system demonstrates strong adaptability in practical applications, ensuring efficient and reliable treatment processes and reducing reliance on manual intervention. Embedded experimental data graphs visually demonstrate the close correlation between predicted and actual values, verifying the model's accuracy and unbiasedness.

[0028] 2. The intelligent dosing mechanism of this invention directly reduces the amount of coagulant used. By dynamically adjusting the dosing strategy as needed, it effectively avoids the resource waste caused by traditional overdosing. This optimization not only significantly reduces reagent procurement costs but also lowers overall operating costs. Simultaneously, environmental benefits are also apparent, such as reducing carbon emissions and sludge production, alleviating solid waste disposal pressure, and providing an economical and efficient path for sustainable water treatment. The system design emphasizes cost control and ecological balance, adapting to different mine conditions and ensuring long-term economic and environmental sustainability. This comprehensive benefit makes the technology easy to promote in practical deployment, achieving resource optimization without complex maintenance.

[0029] 3. The heterogeneous integrated learning architecture of this invention significantly improves the system's anti-interference capability and stability. Even under sensor noise or sudden changes in water quality, it can still maintain stable control and reduce the risk of operational interruption. The automation mechanism significantly reduces the frequency of manual maintenance through adaptive correction of historical deviations and periodic model updates. The response speed is as fast as real-time, and it can promptly handle abnormal operating conditions such as sudden increases in turbidity caused by heavy rain. Multiple safety protection strategies further ensure the reliability of continuous operation. Attached Figure Description

[0030] Figure 1 This is an overall flowchart of an intelligent control method for a water treatment agent according to the present invention.

[0031] Figure 2 This is a block diagram of an intelligent control system for a water treatment agent according to the present invention.

[0032] Figure 3 This is a comparison and analysis chart of actual and predicted values ​​in the experimental data of this invention.

[0033] Figure 4 This is a residual analysis graph in the experimental data graph of this invention.

[0034] Figure 5 This is a histogram of the residual distribution in the experimental data graph of this invention.

[0035] Figure 6 This is a statistical chart showing the percentage distribution of errors in the experimental data of this invention. Detailed Implementation

[0036] See Figure 1-6 As shown, a smart control method for water treatment agents includes: Real-time collection of water quality parameters and flow data, dynamic construction of time-series feature space, and integration of current parameter values, historical lagged parameter values, and interactive features; The data is preprocessed to extract key interaction features and construct a multi-dimensional feature space. The processed data is input into a learning prediction model that integrates three basic learners: multilayer perceptron, random forest, and support vector regression. The model is fused using a dynamic weighting strategy, with the weights adaptively adjusted based on prediction consistency. The model outputs predicted values. The addition of data is controlled based on the predicted values, and dynamic corrections are made in conjunction with historical biases. At the same time, the addition range is limited and the rate of change per minute is constrained.

[0037] An intelligent control system for a water treatment agent includes: The data acquisition and preprocessing unit is connected to the sensor group and flow meter installed on the inlet pool and pipeline; The integrated learning prediction unit connects to the data acquisition and preprocessing unit, receives the processed data, and outputs the predicted dosage through the model. The intelligent dosing control unit is connected to an integrated learning and prediction unit, which receives the predicted dosing amount and makes dynamic corrections to generate the final control command. The safety interlock module is connected to the intelligent dosing control unit to impose dosing range restrictions and change rate constraints on control commands; the actuator includes a solenoid valve group and a metering pump, the metering pump is connected to the coagulant storage tank, and the safety interlock module is connected to the solenoid valve group and the metering pump.

[0038] like Figure 1 In one embodiment shown, Figure 1This is an overall flowchart of an intelligent control method for water treatment agents according to the present invention. First, the present invention collects water quality parameters and flow data in real time, dynamically constructs a temporal feature space, and integrates current parameter values, historical lag parameter values, and interaction features. Water quality parameters include turbidity, pH value, temperature, and suspended solids concentration. Historical lag parameters include the first-order lag features of water quality parameters and flow data at the current time and the two time points prior, as well as the second-order lag features of turbidity and pH value. Key interaction features include the interaction term between turbidity and pH value, the interaction term between suspended solids and temperature, the ratio of turbidity to suspended solids, the square term of pH value deviation, the square root term of turbidity, and discrete features based on flow rate values.

[0039] Next, the data is preprocessed. Preprocessing includes outlier handling, employing an improved quartile method to define the valid data range. Specifically, the interquartile range is calculated by determining the quartiles set for the data. When the interquartile range is zero, the standard deviation is used as a substitute; data exceeding the valid data range are replaced using a moving average. Simultaneously, an autoregressive time-series model is established for the turbidity parameter, where the current turbidity value is calculated by multiplying the previous turbidity value by a coefficient and adding a random perturbation. A dynamic model is established for pH, where the current value is calculated by multiplying the values ​​from the previous two time points by two different coefficients and adding a random perturbation. A slowly varying temperature model is established, where the current value is calculated by multiplying the values ​​from the previous two time points by two different coefficients and adding a random perturbation.

[0040] Next, the processed data is input into a learning prediction model integrating three basic learners: Multilayer Perceptron, Random Forest, and Support Vector Regression. This model is fused using a dynamic weighting strategy, with weights adaptively adjusted based on prediction consistency. Specifically, the standard deviation of the predicted values ​​from the three basic learners is first calculated. If the prediction consistency is less than a set value, a fixed weight allocation is used, with the weight of the Multilayer Perceptron being greater than that of the Random Forest, which is greater than that of the Support Vector Regression. If the prediction consistency is greater than or equal to the set value, the weight of the model closest to the median is increased. The output predicted value must constrain the injection amount within a set range. If the predicted value is greater than or less than the set range, the corresponding maximum or minimum value within the set range is output. The processed data includes the basic injection amount, which is specifically a weighted sum of water quality parameters and flow rate data. The specific weighting coefficients are determined through field experiments based on the characteristics of the mine water quality.

[0041] Then, the dosage is controlled based on the predicted value, and dynamically corrected by incorporating historical deviations. Dynamic correction of historical deviations is achieved through a correction coefficient. An initial value is set, and the coefficient decreases when the actual treatment effect is better than expected for several consecutive cycles, and increases when the effect is worse than expected. The correction coefficient is limited to a set range; the final dosage is the product of the predicted value and the correction coefficient. Simultaneously, dosage range limits and minute-by-minute change rate constraints are implemented. Dosing control also includes automatically switching to a preset value control mode when any sensor malfunctions, maintaining dosage control based on historical averages, and issuing an alarm signal to prompt timely action.

[0042] Finally, dynamic correction based on historical bias also includes performance evaluation and model updates. Specifically, the ensemble learning model is retrained using the most recent running data every set time interval. The training metric is the mean absolute percentage error, which is calculated as 1 divided by the number of samples, multiplied by the absolute value of the difference between the actual and predicted values ​​of each sample, and divided by the sum of the actual values. When the error exceeds a set value, the model parameters are updated.

[0043] In another embodiment, the technical problem to be solved by the present invention, in view of the shortcomings of the prior art, is as follows: 1. Inconsistent treatment results due to low precision in coagulant dosing. The shortcomings of existing technologies: Existing single-parameter control methods (such as those based solely on turbidity control) or simple multi-parameter linear combination methods cannot accurately reflect the complex relationships between water quality parameters, leading to large deviations in coagulant dosage and low control accuracy, typically with errors of 15-25%, resulting in unstable effluent quality and low compliance rates. The technical problem solved by this invention: How to improve the accuracy of coagulant dosage, reduce the control error to within 5%, and ensure stable compliance of effluent quality.

[0044] 2. The problem of high operating costs due to serious waste of reagents. Disadvantages of existing technologies: Due to insufficient precision in dosing control, existing technologies often employ a conservative strategy of overdosing to ensure treatment effectiveness, resulting in coagulant dosages that are 20-30% higher than the theoretical optimal value. This leads to significant waste of reagents, high operating costs, and the potential for secondary pollution from overdosing. Technical problem solved by this invention: How to achieve precise coagulant dosing, reduce reagent waste, and lower operating costs by 15-25%.

[0045] 3. Poor system anti-interference capability leads to instability in the control system. The shortcomings of existing technologies: Single algorithm models are easily affected by factors such as sensor noise, sudden changes in water quality, and equipment failures, resulting in poor system robustness. When abnormal operating conditions occur, control failures or significant fluctuations can easily occur, affecting the stable operation of the entire treatment system. The technical problem solved by this invention: How to improve the anti-interference capability and stability of the control system, maintaining effective control even under abnormal operating conditions.

[0046] 4. Poor adaptability leads to the need for frequent manual adjustments. The shortcomings of existing technologies: Existing control methods lack learning capabilities and cannot automatically adapt to dynamic characteristics such as seasonal changes in mine water quality and changes in operating conditions. They require frequent parameter adjustments and system maintenance by operators, resulting in high labor costs, and the lag in adjustments affects the treatment effect. The technical problem solved by this invention: How to achieve adaptive adjustment of the control system, reduce manual intervention, and improve the system's automation level and long-term stability.

[0047] Therefore, the present invention employs the following technical means: The online multi-parameter water quality monitoring unit is the sensing foundation of the entire system, responsible for real-time acquisition of key water quality parameters of the mine water. This unit is equipped with five high-precision online sensors for monitoring turbidity, pH, temperature, suspended solids concentration, and flow rate. The turbidity sensor uses a light scattering measurement principle with a range of 0-1000 NTU. The system monitors influent turbidity in the range of 10-100 NTU, with a control target of effluent turbidity ≤10 NTU and a measurement accuracy of ±2%. It is installed on the straight section of the influent pipeline, at least 10 times the pipe diameter away from bends to ensure flow stability. To prevent probe surface contamination from affecting measurement accuracy, the system is equipped with an automatic cleaning device that performs a cleaning procedure every 4 hours. The pH sensor uses a composite electrode design, covering the full range of 0-14 with an accuracy of ±0.01%. It adopts a flow cell installation method and integrates temperature compensation to ensure measurement accuracy under different temperature conditions. The system also features an automatic calibration function, automatically performing two-point calibration every 7 days to maintain long-term measurement stability.

[0048] The temperature sensor uses an industrial-grade PT100 platinum resistance thermometer, with a measurement range of -20 to 100℃ and an accuracy of ±0.1℃. It is integrated with the pH sensor in the same flow cell, achieving temperature compensation while reducing installation space. The suspended solids concentration sensor is based on the optical scattering principle, with a range of 0-1000 mg / L. The system monitors the influent suspended solids concentration range of 50-250 mg / L, with a control target of ≤30 mg / L in the effluent, an accuracy of ±3%, and uses a vertical installation method to effectively avoid bubble interference. The flow meter uses an electromagnetic measurement principle, with a range of 0-1000 m³ / h. 3 / h, accuracy ±0.5%, installed on a straight pipe section, ensuring a straight pipe section requirement of 10 times the pipe diameter before and 5 times the pipe diameter after, to guarantee the accuracy of flow measurement.

[0049] This invention constructs a five-dimensional feature space, which increases the information dimension by more than 150% compared to traditional single-parameter or dual-parameter control methods, providing a rich data foundation for subsequent intelligent prediction. Random forest feature importance analysis based on experimental data shows that the contribution of each water quality parameter to the prediction of coagulant dosage varies significantly. Turbidity has the highest importance (0.385), serving as the primary factor determining dosage; suspended solids concentration is second (0.298), synergistically characterizing particulate matter load in water with turbidity; pH ​​(0.156) significantly affects flocculation effect; flow rate (0.098) affects residence time and mixing intensity; and temperature (0.063) affects reactivity.

[0050] Pearson correlation coefficient analysis revealed the intrinsic relationships between parameters. This analysis employed the statistical method of calculating the Pearson product-moment correlation coefficient: first, the deviations of the two variable values ​​from their respective means for each sample point were calculated; then, the sum of the products of these deviations was calculated. The denominator was calculated as the square root (i.e., standard deviation) of the sum of the squares of the deviations of the two variables. Finally, the correlation coefficient r was equal to the sum of the deviation products divided by the product of the two standard deviations. This coefficient ranges from -1 to 1 and is used to quantify the positive or negative correlation between parameters. The calculation results showed a strong positive correlation between turbidity and suspended solids concentration (r = 0.847), a moderate negative correlation between pH and temperature (r = -0.312), and a weak positive correlation between flow rate and turbidity (r = 0.289). Significant interaction effects existed among the parameters, such as the turbidity-pH synergistic effect and the suspended solids-temperature coupling effect. When all parameters were in an optimal synergistic state, the dosage control accuracy could reach ±1.5 mg / L. This multidimensional parameter relationship analysis provides important prior knowledge for the ensemble learning model, significantly improving prediction accuracy.

[0051] The data acquisition and preprocessing unit is responsible for acquiring raw data from the sensors and performing necessary processing to ensure that the data quality meets the model input requirements. Data acquisition adopts a fixed cycle of 1 minute, reading data from each sensor via Modbus RTU / TCP protocol to form structured data records containing timestamps, turbidity, pH value, temperature, suspended solids concentration, and flow rate.

[0052] Data quality checking is the first step in preprocessing. The system first checks data integrity to ensure all fields have valid values. Then, a range check is performed: influent turbidity is typically between 10-100 NTU, effluent turbidity is controlled between 1-10 NTU, pH is between 6.0-9.0, temperature is between 5-35℃, influent suspended solids concentration is between 50-250 mg / L, effluent suspended solids concentration is controlled between 10-30 mg / L, and flow rate is between 100-600 m³ / L. 3 Data outside the / h range will be marked as outliers.

[0053] To address the time-varying characteristics of mine water quality, this invention features a specially designed full-parameter time-series feature extraction mechanism. The system establishes complete time-series models for all key water quality parameters: 1. Extraction of turbidity time series features Establish an autoregressive time series model: the turbidity value at the current time (t) is equal to 0.8 times the turbidity value at the previous time (t-1), plus a random disturbance term ε. T (t) represents a random perturbation within the range [-5,5]NTU.

[0054] 2. Extraction of pH value time-series features A dynamic pH value change model is established: the current pH value is composed of 0.75 times the previous pH value (t-1) and 0.2 times the pH value of the two previous times (t-2), plus a random perturbation ε. pH (t) represents a random disturbance in the range [-0.3, 0.3].

[0055] 3. Temperature time series feature extraction Establish a model for the slowly changing temperature characteristics: The current temperature Temp(t) is calculated as the sum of 0.85 times the temperature at the previous time step (t-1) and 0.1 times the temperature at the two time steps before that (t-2), plus the perturbation term ε. Temp (t) represents a random perturbation within the range of [-2, 2]℃.

[0056] 4. Extraction of Temporal Features of Suspended Solids Concentration A suspended solids concentration fluctuation model is established: the current suspended solids concentration SS(t) is equal to the sum of 0.7 times the value at the previous time step (t-1) and 0.25 times the value at the two previous time steps (t-2), plus the disturbance ε. SS (t) represents a random perturbation within the range of [-15, 15] mg / L.

[0057] 5. Traffic flow time series feature extraction Establish a flow rate trend model: The current flow rate F(t) is composed of 0.6 times the flow rate at the previous time step (t-1) and 0.3 times the flow rate at the two time steps before that (t-2), plus the disturbance ε. F(t), where [-30, 30]m 3 Random disturbances within the range of / h.

[0058] Through full-parameter temporal modeling, the system constructs an 18-dimensional extended feature space, specifically including: Five original features at the current time step: T(t), pH(t), Temp(t), SS(t), F(t) Five first-order hysteresis characteristics: T(t-1), pH(t-1), Temp(t-1), SS(t-1), F(t-1) Two second-order hysteresis characteristics of turbidity: T(t-2) and pH(t-2) (turbidity and pH are the most sensitive to changes). Six interactive and non-linear features: Turbidity-pH interaction term: T(t) multiplied by |pH(t) - 7.0|; Suspended matter-temperature interaction term: SS(t) multiplied by Temp(t) / 1000; Turbidity-suspended solids ratio: T(t) divided by (SS(t) + 1e-6); The square term of pH deviation: [pH(t) - 7.2] 2 ; Turbidity square root term: √T(t); Flow rate classification characteristics: Based on F(t), it is divided into low (0-200), medium (200-400), and high (400-600) m. 3 / h has three levels.

[0059] This multidimensional time-series modeling method can effectively capture the dynamic changes in water quality and the complex relationships between parameters, providing richer time-dimensional information for the prediction model. Compared with the traditional five-dimensional feature space, the information dimension is increased by 260%.

[0060] Outlier handling employs a modified quartile method. First, the 5th and 95th quartiles of the data are calculated as Q1 and Q3, respectively. Then, the interquartile range (IQR) is calculated as Q3 - Q1. When the IQR is zero, it indicates that the data distribution is too concentrated; in this case, the standard deviation is used instead of the IQR. The effective data range is defined as [Q1 - 1.5 × IQR, Q3 + 1.5 × IQR]. Data outside this range are replaced using a moving average method, ensuring data continuity while avoiding interference from outliers on the model.

[0061] Data standardization is the final step in preprocessing. The Z-score standardization method is used to standardize the original feature values ​​X to X0. scaledThe calculation method is as follows: subtract the mean μ of the feature, and then divide by its standard deviation σ. For example, the turbidity value is first subtracted from the historical average turbidity value, and then divided by the turbidity standard deviation. Standardization eliminates the influence of different feature dimensions, improving the training efficiency and prediction accuracy of subsequent machine learning models.

[0062] The ensemble learning prediction unit is the intelligent core of the entire system, responsible for predicting the optimal coagulant dosage based on the pretreated water quality data. This unit adopts a heterogeneous ensemble learning architecture, which includes three complementary basic learners: Multilayer Perceptron (MLP), Random Forest (RF), and Support Vector Regression (SVR).

[0063] The reasons for choosing these three models are as follows: MLP is good at handling complex nonlinear relationships and can learn high-order interaction features between water quality parameters; RF has good noise resistance and feature selection capabilities, can automatically identify important features, and has relatively low requirements for data quality; SVR has strong generalization ability, performs stably in small sample cases, and can effectively handle regression problems in high-dimensional feature spaces.

[0064] The multilayer perceptron submodule employs a feedforward neural network structure. The input layer contains five neurons, corresponding to five standardized water quality features. The hidden layer is designed as a single layer with 100 neurons, and the activation function is the hyperbolic tangent (tanh), which has better nonlinear expression capabilities compared to the ReLU function. The output layer is a single neuron, directly outputting the predicted coagulant dosage. The network training uses the Adam optimization algorithm, combining the advantages of momentum and adaptive learning rates. The L2 regularization parameter α is set to 0.00842, and the initial learning rate is 0.00282; these parameters are obtained through random search optimization. An early stopping mechanism is enabled during training; training automatically stops when the validation set performance shows no improvement after 10 consecutive iterations, effectively preventing overfitting.

[0065] The Random Forest submodule constructs an ensemble model containing 100 decision trees, with a maximum depth of 10 layers per tree, a minimum number of samples for node splits of 5, and a minimum number of samples for leaf nodes of 2. These parameter settings balance model complexity and generalization ability. At each node split, the square root of the total number of features is randomly selected for optimal split point search, increasing the model's randomness and diversity. A bootstrap sampling method with replacement is used to generate training subsets for each tree, further improving the ensemble's performance.

[0066] The Support Vector Regression submodule uses a Radial Basis Function (RBF) as its kernel function, which maps the original feature space to a high-dimensional space, enabling nonlinear regression. The penalty parameter C is set to 1.0, balancing model complexity and training error. The ε-insensitive loss parameter is set to 0.1, ignoring samples with prediction errors less than 0.1, thus improving the model's robustness. The kernel function parameter γ employs an automatic scaling strategy, adaptively adjusting based on the number of features and variance.

[0067] The voting integrator is responsible for fusing the prediction results of the three basic learners, employing a dynamic weighting strategy: the final predicted coagulant dosage D is a weighted sum of the predicted values ​​Di from the three basic models (MLP, RF, SVR), with weights wi dynamically adjusted based on prediction consistency. The weight calculation is based on dynamic adjustment according to prediction consistency. The standard deviation of the three model predictions is calculated as a measure of uncertainty. When the prediction consistency is high (standard deviation < 2.0), the basic weights [0.4, 0.35, 0.25] are used, corresponding to MLP, RF, and SVR respectively. When the prediction discrepancy is large, the weight of the model closest to the median is increased to physically constrain the prediction results and ensure that the dosage is within the range of [5.0, 50.0] mg / L.

[0068] The core innovation of this invention lies in employing a heterogeneous ensemble learning strategy and a dynamic weighting mechanism, fully leveraging the nonlinear fitting ability of neural networks, the feature selection ability and anti-overfitting properties of random forests, and the generalization ability of support vector machines. Experimental results based on 3000 samples show that the mean absolute percentage error (MAPE) of the ensemble model is reduced to 2.89%, and the coefficient of determination R0 is [missing value]. 2 The performance reached 0.9303, significantly outperforming any single model. The performance comparison of each basic learner is as follows: MLP's MAPE is 2.71%, R... 2 It is 0.9372; the MAPE of RF is 3.17%, R 2 The value was 0.9171; the MAPE of SVR was 3.49%, and R... 2 The value is 0.9058; the MAPE of the ensemble model is 2.89%, and R0 is... 2 It is 0.9303.

[0069] The intelligent dosing control unit is responsible for translating the predicted dosage into actual control actions. The dosage calculation module first calculates the baseline dosage based on the mechanistic model: the baseline dosage is a linear combination of five water quality parameters: turbidity (T) multiplied by a coefficient a1, pH deviation squared ((pH - 7.0)). 2Multiply a1 by a2, temperature by a3, suspended solids concentration (SS) by a4, and flow rate (F) by a5. The coefficients a1, a2, a3, a4, and a5 need to be determined through field tests based on the specific mine water quality characteristics. Typical reference values ​​are 0.1, 0.3, 0.05, 0.15, and 0.01, respectively. In practical applications, these coefficients should be dynamically adjusted and optimized based on operational data.

[0070] In practical applications, the system uses the product of the predicted value and the historical correction coefficient as the final dosage. The correction coefficient is dynamically adjusted based on historical control deviations, with an initial value of 1.0. The adjustment rules are as follows: when the actual treatment effect is better than expected for five consecutive cycles, the correction coefficient decreases by 0.02; when the actual treatment effect is worse than expected for five consecutive cycles, the correction coefficient increases by 0.02. The correction coefficient range is limited to [0.8, 1.2] to avoid over-adjustment. This adaptive correction mechanism can compensate for the systematic deviation of the model prediction and improve long-term control accuracy.

[0071] The dosing pump control module is responsible for converting the calculated dosage into control signals for the actuators. The system uses a diaphragm metering pump as the actuator for coagulant dosing, with a flow rate range of 0-100 L / h. Precise flow rate regulation is achieved through a frequency converter. The control signal uses an industrial standard 4-20mA current signal, and the conversion formula is: 4 plus 16 multiplied by the ratio of the dosing flow rate to the maximum flow rate Q. Simultaneously, an electromagnetic flowmeter monitors the actual dosing flow rate in real time, forming a closed-loop control system to ensure dosing accuracy.

[0072] The safety interlock module incorporates multiple safety strategies. The upper limit for dosage is set at 50 mg / L to prevent overdosing, waste, and secondary contamination; the lower limit is set at 5 mg / L to ensure basic treatment effectiveness; and the rate of change is limited to no more than 5 mg / L per minute to prevent drastic fluctuations in dosage from impacting the treatment system. If any sensor malfunctions, the system automatically switches to a preset value control mode, maintaining basic dosage control based on historical averages, while simultaneously issuing an alarm signal to alert maintenance personnel for timely intervention.

[0073] The implementation of the method of this invention includes three stages: system initialization, real-time operation, and optimization and maintenance. In the system initialization stage, sensor calibration is performed first, using standard solutions to calibrate the zero point and range of each sensor to ensure measurement accuracy. Then, communication link testing is conducted to verify normal data communication between the sensors, controller, and host computer. Next, pre-trained model parameters are loaded, including the weight parameters of the three basic learners and data normalization parameters. Finally, safety interlock testing is performed to simulate various fault conditions and verify the normal operation of the safety protection functions.

[0074] During real-time operation, the system continuously executes data acquisition, preprocessing, prediction, and control processes according to a set cycle. A set of water quality data is collected every minute from five sensors, and after quality checks, it is stored in a real-time database. The preprocessing module performs outlier detection, time-series feature extraction, and standardization on the data. The processed data is input into an ensemble learning model, where three basic learners make predictions in parallel, and a voting ensemble calculates the average to obtain the final prediction result. The control module calculates the actual dosage based on the predicted value and correction coefficients, and after a safety check, outputs a control signal to the metering pump. The entire process is executed within 10 seconds to ensure real-time control.

[0075] The optimization and maintenance phase mainly includes effect evaluation and model updates. The system uses multiple indicators to evaluate model performance, the most important of which is the Mean Absolute Percentage Error (MAPE): MAPE calculates the average absolute percentage deviation between the actual dosage and the predicted value. The specific steps are: for each sample, calculate the absolute value of the difference between the actual and predicted values, divide it by the actual value, and then multiply by 100%; finally, average the results for all samples. For example, the MAPE for n samples is 100% divided by n multiplied by the sum of the absolute deviations of each sample. Based on these evaluation indicators, the system regularly updates the correction coefficients and performs monthly model retraining, updating model parameters using the most recent month's operating data to adapt to long-term changes in water quality characteristics.

[0076] The technological innovations of this invention are mainly reflected in four aspects. First, the construction of a multi-dimensional feature space: for the first time, five key parameters—turbidity, pH, temperature, suspended solids concentration, and flow rate—are simultaneously incorporated into the control model, improving information utilization by more than 150% compared to traditional methods, thus laying a data foundation for accurate prediction. Second, the fusion of temporal features: innovatively introducing a turbidity time series model, capturing the dynamic changes in water quality through autoregressive modeling, and improving the model's adaptability to water quality fluctuations. Third, the heterogeneous integrated learning architecture: combining the advantages of three heterogeneous models—multilayer perceptron, random forest, and support vector regression—and achieving complementary advantages through voting ensemble, improving prediction accuracy to 97.11%. Fourth, the adaptive optimization mechanism: through real-time correction of historical biases and periodic model retraining, ensuring the system maintains high accuracy and stability during long-term operation.

[0077] Experimental results demonstrate that this invention achieves significant technological advancements. In terms of prediction accuracy, the average absolute percentage error is reduced to 2.89%, a nearly 90% improvement compared to the 15-25% error rate of traditional methods. Regarding reagent savings, precise dosing avoids overdosing, reducing reagent usage by 15-20% and significantly lowering operating costs. As for system stability, the robustness of ensemble learning and the adaptive mechanism ensure stable control under various operating conditions.

[0078] To adapt to different application scenarios and technical conditions, this invention also provides three alternative technical solutions. The deep learning solution uses Long Short-Term Memory (LSTM) networks or Gated Recurrent Units (GRUs) instead of Multilayer Perceptrons, which can better model time-series features and is particularly suitable for application scenarios with drastic water quality changes. However, this solution requires more training data and computing resources, resulting in relatively high implementation costs. The online learning solution uses an online gradient descent algorithm, which can update model parameters in real time and quickly adapt to water quality changes, making it suitable for scenarios with frequent changes in water quality characteristics. However, online learning has relatively poor stability and requires stricter parameter constraints and security protection. The fuzzy control solution builds a fuzzy rule base based on expert experience and achieves dosing control through fuzzy inference. It has the advantages of good interpretability and ease of understanding, making it suitable for scenarios with high requirements for system transparency. However, the accuracy of fuzzy control is generally lower than that of machine learning methods, with an error typically in the range of 12-15%.

[0079] In system implementation, five sensors—turbidity, pH, temperature, suspended solids concentration, and flow rate—are essential; the absence of any one will significantly impact prediction accuracy. Standardization steps in data preprocessing are also mandatory; otherwise, features with different dimensions will affect model training. Ensemble learning requires at least two base learners to demonstrate its ensemble effect, with three learners being the recommended configuration. A dosing control actuator is fundamental to system operation and must be included. Temporal feature extraction, a third base learner, an online optimization module, and a graphical user interface are optional configurations that can further improve system performance but are not essential.

[0080] This invention, through innovative system architecture design and advanced algorithm implementation, successfully solves technical problems such as low control accuracy of mine water coagulant addition, serious waste of reagents, and poor system stability. It achieves significant results with a control accuracy improvement of more than 50% and a reagent saving of 15-20%, demonstrating important technological advancement and broad application prospects.

[0081] The intelligent dosing control method and system for mine water coagulants based on ensemble learning proposed in this invention has the following technical advantages and beneficial effects compared with the prior art: 1. Control precision is significantly improved, and the stability of processing results is enhanced. This invention constructs an 18-dimensional extended feature space including turbidity, pH value, temperature, suspended solids concentration, and flow rate, and employs a dynamic weighted ensemble learning strategy using three heterogeneous models: multilayer perceptron, random forest, and support vector regression, to achieve ultra-accurate prediction of coagulant dosage. Based on experimental data from 3000 samples, the system's mean absolute percentage error (MAPE) is 2.89%, and the coefficient of determination R0 is [missing value]. 2Reaching 0.9303, the prediction accuracy is improved by about 90% compared to existing single-parameter control methods (with errors typically between 15-25%) or simple PID control (with errors around 12-18%).

[0082] The significant improvement in control precision directly enhanced the stability of the effluent quality. Experimental results showed that the residual standard deviation was only 1.64 mg / L, and the residual mean was close to zero (0.0137), indicating no systematic bias in the prediction. Because the dosage was highly matched to actual needs, it avoided the problems of poor treatment effect due to insufficient dosage and reagent residue caused by excessive dosage. Actual operating data showed that the standard deviation of effluent turbidity decreased from ±3.2 NTU to ±1.5 NTU, the suspended solids removal rate remained stable at over 95%, and the effluent compliance rate increased from 85% to 98%.

[0083] 2. Reagent consumption is significantly reduced, and operating costs are effectively controlled. Traditional control methods, due to precision limitations, typically employ a conservative strategy of appropriately overdosing to ensure treatment effectiveness. This invention achieves on-demand dosing through precise prediction, dynamically adjusting the dosage based on real-time changes in water quality parameters. Actual measurement data shows that, while maintaining the same treatment effect, the average coagulant dosage is reduced by 18-22%.

[0084] 3. The system's anti-interference capability is enhanced, and its operational reliability is improved. The dynamic weighted ensemble learning architecture employed in this invention exhibits exceptional robustness. Three fundamental learners learn data features from different perspectives: MLP focuses on nonlinear fitting, RF on feature selection, and SVR on generalization ability. When one model is affected by noise, the others can still maintain normal predictions. The dynamic weight allocation mechanism effectively suppresses prediction bias in individual models.

[0085] Noise immunity experiments show that when individual sensor data contains ±10% random noise, the overall prediction error of the integrated system only increases by 2-3 percentage points (from 2.89% to 5.1%), while the error of a single neural network model increases by 8-12 percentage points (from 2.71% to 10.8%). Furthermore, the system's multiple safety protection mechanisms (dosage limit of 5-50 mg / L, rate of change limit of 5 mg / L / min, automatic fault switching, etc.) further improve operational reliability, enhancing continuous operation stability by over 65%.

[0086] 4. Increased automation reduces manual maintenance workload. This invention introduces an adaptive correction mechanism based on historical deviations and a periodic model update function, enabling the system to automatically adapt to slow changes in water quality characteristics. Traditional control systems require operators to frequently adjust control parameters based on seasonal changes, water quality fluctuations, and other factors, averaging 2-3 manual interventions per week. In contrast, under normal operating conditions, this system only requires a routine check and model performance evaluation once a month, reducing manual maintenance workload by approximately 50%.

[0087] 5. The response speed meets the requirements of real-time control. The algorithm of this invention is optimized so that a complete control cycle (including data acquisition, preprocessing, model prediction, and control output) can be completed within 8-10 seconds. Compared with some control methods based on complex optimization algorithms (response time 30-60 seconds), the response speed is improved by 3-6 times. The faster response speed enables the system to respond promptly to rapid changes in water quality, such as the sudden increase in turbidity caused by mine water inflow during heavy rain, effectively preventing the treatment system from being impacted.

[0088] 6. Implementation and deployment are relatively simple, and it has good applicability. This invention employs a modular design and standardized interfaces, making it compatible with existing water quality monitoring equipment and control systems. During system deployment, historical operational data is used for model pre-training. After the new system goes live, it typically only requires 10-15 days of parameter fine-tuning to reach a stable operating state. Compared to complex control systems that require long-term debugging (usually 1-2 months), the deployment cycle is shortened by more than 60%.

[0089] The system also provides three alternative technical solutions (deep learning, online learning, and fuzzy control), allowing users to choose the appropriate solution based on the technical conditions and personnel skill levels of different mines, thus enhancing the value of the technology's promotion and application.

[0090] 7. Environmental benefits are gradually becoming apparent. The environmental benefits of precise dosing control are mainly reflected in the following aspects: First, it reduces the amount of coagulant used by 18-22%, correspondingly reducing energy consumption and carbon emissions during chemical production and transportation; second, it improves the stability of effluent quality (the compliance rate increases from 85% to 98%), significantly reducing the pollution load on receiving water bodies; third, it reduces sludge production by 12-15%, alleviating the pressure on solid waste disposal; and fourth, it avoids the risk of secondary pollution caused by excessive dosing. Preliminary calculations show that treating 10,000 tons of mine water can reduce CO2 equivalent emissions by approximately 1.8-2.2 tons.

[0091] It should be noted that the above-mentioned technical effects are based on laboratory test data, and the actual application effects may vary due to factors such as mine water quality characteristics, equipment conditions, and operation and management levels. The core value of this invention lies in providing a more intelligent and refined control method, offering a feasible solution for the technological upgrading of the mine water treatment industry.

[0092] like Figure 2 In one embodiment shown, Figure 2 This is a block diagram of an intelligent control system for a water treatment agent according to the present invention. The intelligent control system for a water treatment agent includes: The data acquisition and preprocessing unit is connected to the sensor group and flow meter installed on the inlet pool and pipeline; The integrated learning prediction unit connects to the data acquisition and preprocessing unit, receives the processed data, and outputs the predicted dosage through the model. The intelligent dosing control unit is connected to an integrated learning and prediction unit, which receives the predicted dosing amount and makes dynamic corrections to generate the final control command. The safety interlock module is connected to the intelligent dosing control unit to impose dosing range restrictions and change rate constraints on control commands; the actuator includes a solenoid valve group and a metering pump, the metering pump is connected to the coagulant storage tank, and the safety interlock module is connected to the solenoid valve group and the metering pump.

[0093] In another embodiment, the present invention provides an intelligent dosing control system for mine water coagulants based on ensemble learning. This system adopts a modular design and mainly includes five functional modules: an online monitoring unit for multiple water quality parameters, a data acquisition and preprocessing unit, an ensemble learning prediction unit, an intelligent dosing control unit, and a human-machine interaction and management unit. The units communicate in real-time via industrial Ethernet and fieldbus, forming a complete closed-loop control system. The overall system architecture adopts a layered design: the bottom layer is the field equipment layer, including various sensors and actuators; the middle layer is the control layer, realizing data processing and intelligent decision-making; and the top layer is the management layer, providing human-machine interaction and system management functions.

[0094] This system is an integrated water treatment monitoring and automatic dosing control system. It starts from the inlet tank, and the water flows through the coagulation reaction tank and sedimentation tank in sequence, and finally reaches the outlet tank.

[0095] Throughout the process, various sensors and flow meters are responsible for real-time monitoring: turbidity sensors, pH sensors, temperature sensors, and suspended solids sensors collect key water quality parameters from each treatment unit and the effluent tank, while flow meters monitor water flow data. All this real-time data is fed into the data acquisition and preprocessing unit. After processing the raw data, it is sent to the integrated learning prediction unit for model calculation, and the data is stored in the historical database for later retrieval. The integrated learning prediction unit combines real-time and historical data to output prediction commands to the intelligent dosing control unit.

[0096] The intelligent dosing control unit generates a control signal based on this signal. After being verified by the safety interlock module, the signal ultimately drives the actuators—namely, the metering pump controlling the coagulant storage tank and the solenoid valve group controlling the coagulation reaction tank—thereby achieving precise and automatic adjustment of the agent dosing.

[0097] In addition, data from the intelligent dosing control unit, safety interlock module, and various sensors are all uploaded to the operator workstation, providing personnel with a comprehensive monitoring interface. This workstation also connects to the reporting system and remote monitoring terminal, enabling data recording, report generation, and remote management functions.

[0098] like Figure 3 In one embodiment shown, Figure 3 This is a comparative analysis chart of actual and predicted values ​​in the experimental data of this invention. The experiment was conducted on a desktop computer with an Intel U9 processor and 32GB of memory, based on a test of 3000 samples. A scatter plot comparing the actual coagulant dosage with the predicted values ​​of the ensemble learning model is shown for 600 test samples. The data points are mainly distributed within the dosage range of 28-50 mg / L, closely distributed along the diagonal, indicating a high correlation between the predicted and actual values. The coefficient of determination R for linear fitting is... 2 The accuracy reached 0.9303, verifying the model's extremely high predictive accuracy.

[0099] like Figure 4 In one embodiment shown, Figure 4 This is a residual analysis graph from the experimental data of this invention. It shows the distribution of the predicted residuals with respect to the predicted values. The residuals are randomly distributed near the zero line, with no obvious systematic bias or heteroscedasticity. Most of the residuals are controlled within ±2 mg / L, indicating that the model has excellent stability and unbiasedness. The residual distribution is most concentrated in the predicted value range of 35-45 mg / L, indicating the highest prediction accuracy.

[0100] like Figure 5 In one embodiment shown, Figure 5This is a histogram of the residual distribution in the experimental data of this invention. It shows the statistical distribution characteristics of the prediction error. The residuals exhibit a standard normal distribution, with the center located around 0.0137 and a standard deviation of 1.64 mg / L. Approximately 78% of the prediction errors are within ±1.5 mg / L, and 95% of the errors are within ±3.2 mg / L, which conforms to the statistical law of normal distribution, proving the high reliability of the model's predictions.

[0101] like Figure 6 In one embodiment shown, Figure 6 This is a statistical graph showing the percentage distribution of errors in the experimental data of this invention. It quantifies the distribution of relative prediction errors. Approximately 45% of the prediction samples have errors within 2%, 75% have errors within 5%, and 92% have errors within 8%, with an average absolute percentage error of 2.89%, far exceeding the 15-25% error level of traditional methods.

[0102] This invention is not limited to the above-described embodiments. Any changes made to its shape or material composition, or any structural design using the methods provided by this invention, are considered variations of this invention and should be considered within the scope of protection of this invention.

Claims

1. A method for intelligent control of water treatment agents, characterized in that, include: Real-time collection of water quality parameters and flow data, dynamic construction of time-series feature space, and integration of current parameter values, historical lagged parameter values, and interactive features; The data is preprocessed to extract key interaction features and construct a multi-dimensional feature space. The processed data is input into a learning prediction model that integrates three basic learners: multilayer perceptron, random forest, and support vector regression. The model is fused using a dynamic weighting strategy, with the weights adaptively adjusted based on prediction consistency, and the predicted value is output. Dosing is controlled based on predicted values ​​and dynamically corrected by incorporating historical deviations. At the same time, limits on the range of dosage and constraints on the rate of change per minute are implemented.

2. The intelligent control method for a water treatment agent according to claim 1, characterized in that, The dynamic weighted fusion strategy is as follows: First, calculate the standard deviation of the predicted values ​​of the three basic learners: Multilayer Perceptron, Random Forest, and Support Vector Regression. If the prediction consistency is less than a set value, a fixed weight allocation is adopted, where the weight of Multilayer Perceptron > the weight of Random Forest > the weight of Support Vector Regression. If the prediction consistency is greater than or equal to the set value, increase the weight of the model closest to the median.

3. The intelligent control method for a water treatment agent according to claim 1, characterized in that, The data preprocessing specifically includes: Outlier handling employs an improved quartile method to define the valid data range. The interquartile range is calculated by determining the quartiles set for the data. When the interquartile range is zero, the standard deviation is used as a substitute. Data exceeding the valid data range is replaced by the moving average method. An autoregressive time series model is established for the turbidity parameter. The turbidity value at the current time is calculated by multiplying the turbidity value at the previous time by a coefficient and adding random perturbation. A dynamic model is established for pH value, where the current value is calculated by multiplying the values ​​of the previous two time points by two different coefficients and adding a random perturbation; a slowly varying model is established for temperature value, where the current value is calculated by multiplying the values ​​of the previous two time points by two different coefficients and adding a random perturbation.

4. The intelligent control method for a water treatment agent according to claim 1, characterized in that, The historical deviation is dynamically corrected through a correction coefficient. An initial value is set, and the coefficient decreases when the actual processing effect is better than expected for several consecutive cycles, and increases when the effect is worse than expected. The correction coefficient is limited to a set range. The final dosage is the product of the predicted value and the correction coefficient.

5. The intelligent control method for a water treatment agent according to claim 1, characterized in that, The water quality parameters include turbidity, pH value, temperature, and suspended solids concentration. The historical lag parameters include the first-order lag characteristics of the water quality parameters and flow data at the current time and the two previous times, as well as the second-order lag characteristics of turbidity and pH value. The key interactive features include the interaction terms of turbidity and pH, the interaction terms of suspended solids and temperature, the ratio of turbidity to suspended solids, the square term of pH deviation, the square root term of turbidity, and discrete features based on flow rate.

6. The intelligent control method for a water treatment agent according to claim 1 or 2, characterized in that, The prediction consistency is the standard deviation. The output prediction value needs to be constrained to be within a set range. If the prediction value is greater than or less than the set range, the corresponding maximum or minimum value of the set range will be output.

7. The intelligent control method for a water treatment agent according to claim 1 or 5, characterized in that, The processed data includes a basic dosage, which is specifically a weighted sum of water quality parameters and flow data. The specific weighting coefficient is determined through field tests based on the characteristics of mine water quality.

8. The intelligent control method for a water treatment agent according to claim 1, characterized in that, The dynamic correction based on historical deviations also includes performance evaluation and model updates. Specifically, the ensemble learning model is retrained using the most recent date's running data after a set time interval. The training metric is the mean absolute percentage error, which is 1 divided by the number of samples and then multiplied by the absolute value of the difference between the actual value and the predicted value of each sample, divided by the sum of the actual values. When the error exceeds a set value, the model parameters are updated.

9. The intelligent control method for a water treatment agent according to claim 1, characterized in that, The control and dosing also includes automatically switching to a preset value control mode when any sensor malfunctions, maintaining dosing control based on historical average values, and issuing an alarm signal to remind users to handle the issue promptly.

10. An intelligent control system for a water treatment agent, employing the intelligent control method for a water treatment agent as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing unit is connected to the sensor group and flow meter installed on the inlet pool and pipeline; The integrated learning prediction unit connects to the data acquisition and preprocessing unit, receives the processed data, and outputs the predicted dosage through the model. The intelligent dosing control unit is connected to an integrated learning and prediction unit, which receives the predicted dosing amount and makes dynamic corrections to generate the final control command. The safety interlock module connects to the intelligent dosing control unit and imposes dosing range limits and change rate constraints on control commands; The actuator includes a solenoid valve assembly and a metering pump. The metering pump is connected to the coagulant storage tank, and the safety interlock module is connected to the solenoid valve assembly and the metering pump.

Images