AI-based raw water pretreatment dosing intelligent control system and method

Abstract

The invention discloses an AI-based raw water pretreatment dosing intelligent control system and method, and relates to the technical field of artificial intelligence, and the system comprises a sensor layer, an edge calculation layer, a cloud platform layer and an execution layer. The sensor layer consists of a distributed sensor network, comprises a turbidity meter, a pH electrode, an electromagnetic flowmeter, a COD (Chemical Oxygen Demand) online monitor and a distributed optical fiber temperature measuring device, covers a raw water inlet, a mixing tank and a sedimentation tank node, and transmits a signal to the edge calculation layer through an RS485 bus. According to the raw water pretreatment dosing intelligent control system and method provided by the invention, through multi-modal data fusion and a dynamic prediction model, the water quality of raw water is accurately sensed and predicted, the dosing amount is adjusted in advance, the hysteresis of a coagulation reaction is effectively compensated, the standard reaching rate of the water quality is improved, and by utilizing an NSGA-II multi-target optimization algorithm, the quality of the raw water is effectively improved. The medicament cost and the treatment effect are balanced, excessive dosing of medicaments in a traditional mode is avoided, and the operation cost is reduced.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an AI-based intelligent control system and method for chemical dosing in raw water pretreatment. Background Technology

[0002] Raw water pretreatment is a crucial step in the water treatment process. Its core objective is to remove suspended solids, colloids, organic matter, and other impurities from raw water by adding chemicals, thus providing stable water quality for subsequent advanced treatments (such as reverse osmosis and ultrafiltration). Traditional pretreatment chemical dosing systems often rely on manual experience or PID control based on a single parameter (such as turbidity or pH), using metering pumps to quantitatively add chemicals. In recent years, with the increasing demand for industrial water and stricter environmental standards, intelligent control systems based on artificial intelligence (AI) have gradually become the industry's development direction. These systems typically integrate sensor networks, data acquisition modules, and machine learning algorithms, enabling real-time monitoring of raw water quality changes (such as flow rate, water temperature, turbidity, colloid concentration, and other multi-dimensional parameters). By establishing mathematical models, they dynamically optimize chemical dosing strategies, achieving precise dosing and energy consumption control. For example, some advanced systems introduce image recognition technology to analyze the floc morphology in the flocculation tank, combining it with historical data to train neural network models, thereby predicting the optimal chemical ratio, which can reduce chemical usage by more than 20% compared to traditional methods.

[0003] However, existing raw water pretreatment dosing systems still face certain technical bottlenecks. First, traditional control methods rely on fixed parameter thresholds or manually set empirical formulas, making it difficult to adapt to dynamic fluctuations in raw water quality (such as seasonal water quality changes or sudden pollution events). This can easily lead to overdosing or underdosing of chemicals, increasing operating costs and potentially affecting the efficiency of subsequent treatment units. Second, existing systems generally lack multi-parameter collaborative analysis capabilities. For example, they adjust the dosage solely based on turbidity feedback, ignoring the combined effects of water temperature, colloidal charge characteristics, and other factors on flocculation, resulting in limited control accuracy. Third, there are stability issues in the chemical preparation process. For instance, traditional pressure dosing systems are susceptible to fluctuations in liquid level and equipment precision, leading to uneven chemical concentrations and subsequently causing scaling or water quality deterioration in subsequent treatment units. Furthermore, some systems exhibit lag in response speed; the time difference between water quality changes and chemical adjustments can cause temporary exceedances in effluent quality, especially when treating high-turbidity or highly polluted raw water. Finally, traditional systems suffer from low equipment integration, large footprint, high maintenance costs, and significant data silos, making them unsuitable for modern intelligent water management needs. To address this, we propose an AI-based intelligent control system and method for raw water pretreatment and chemical dosing. Summary of the Invention

[0004] To address the aforementioned technical issues, this paper provides an AI-based intelligent control system and method for raw water pretreatment dosing. This technical solution solves the problems of relying on fixed parameters or human experience, making it difficult to adapt to dynamic fluctuations in water quality; lacking multi-parameter collaborative analysis, resulting in limited control accuracy; reagent configuration being easily affected by equipment, leading to insufficient stability; response lag potentially causing temporary exceedances of water quality standards; low equipment integration, high maintenance costs, and the existence of data silos.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: The AI-based intelligent control system for raw water pretreatment and dosing includes: a sensor layer, an edge computing layer, a cloud platform layer, and an execution layer. The sensor layer consists of a distributed sensor network, including a turbidity meter, pH electrode, electromagnetic flow meter, COD online monitor and distributed fiber optic temperature measurement device, covering the raw water inlet, mixing tank and sedimentation tank nodes, and transmitting signals to the edge computing layer through RS485 bus; The edge computing layer includes an industrial gateway, which integrates a data cleaning module, a hybrid prediction model and an optimization control module. The hybrid prediction model combines a temporal convolutional network and a Transformer model for short-term water quality prediction and long-term trend analysis. The optimization control module includes model predictive control and the NSGA-II multi-objective optimization algorithm. The cloud platform layer is deployed based on Kubernetes containerization, integrates a digital twin simulation engine and a reinforcement learning framework, and regularly updates the model parameters of the edge computing layer. The execution layer includes a metering pump driven by a servo motor and a frequency converter, which receives instructions via the Profibus-DP protocol and dynamically adjusts the dosage. The system achieves three levels of fault tolerance through cross-validation of redundant sensors, hot backup of metering pumps, and local closed-loop edge computing, with a dosing accuracy error of ≤5% under single-point failure.

[0006] Preferably, the functional architecture of the industrial gateway includes a multi-protocol adaptation module, a real-time data processing unit, and a fault-tolerant communication interface; The multi-protocol adaptation module supports dynamic conversion of RS485, Profibus-DP and Modbus RTU protocols, enabling unified access for heterogeneous devices at the sensor layer and execution layer. The real-time data processing unit has a built-in lightweight inference engine that generates dosage optimization instructions based on the fusion prediction results of the temporal convolutional network TCN and the Transformer model. The fault-tolerant communication interface adopts a dual-channel redundancy design. The main channel is 5G wireless transmission and the backup channel is industrial Ethernet. When the signal strength of the main channel is less than a preset threshold and the packet loss rate is greater than a preset percentage, it will automatically switch to the backup channel. The industrial gateway eliminates ground potential difference interference through the P+F KFD2-UT2-EX1 signal isolator.

[0007] Preferably, the construction of the hybrid prediction model includes: Physical constraints are established based on the coagulation kinetic equation. The output of the TCN model is weighted and fused with the long-term trend prediction results of the Transformer model. The weights are dynamically adjusted according to water quality changes. The TCN model adopts an expanded causal convolutional structure with a kernel size of 3, an expansion factor of 2^layers, and 5 layers. The output layer is mapped to the predicted turbidity value through a fully connected network. The Transformer model employs a multi-head self-attention mechanism, with an input sequence length of 72 hours and an output probability distribution of the water quality trend for the next 24 hours. The model parameters are optimized using a cross-entropy loss function.

[0008] Preferably, the NSGA-II multi-objective optimization algorithm is as follows: A multi-objective function is constructed with the optimization objectives of minimizing the dosage of chemicals, maximizing the water quality compliance rate, and minimizing the residual aluminum ions. The correlation between PAC dosage and turbidity, pH value, and water temperature was analyzed using mutual information matrix analysis to generate an initial population. The optimal solution set is selected by non-dominated sorting and crowding comparison, and the offspring population is generated by simulating binary crossover and polynomial mutation. The dosing scheme is output after iterative update.

[0009] Preferably, the redundant sensor cross-validation specifically involves: When the data deviation between any two similar sensors exceeds a preset first threshold, the system automatically switches to a backup sensor, wherein the first threshold is set according to the sensor type. The metering pump hot backup includes a main metering pump and a standby metering pump. When the main metering pump fails, it is switched to the standby pump through the industrial gateway within a preset switching time window. During the switching process, the fluctuation of the dosage is controlled within a preset allowable range. The edge computing local closed loop achieves continuous control in the absence of network access through a lightweight machine learning model built into the industrial gateway. The local inference response time meets the preset real-time requirements, and it maintains data interaction with the programmable logic controller (PLC) through an industrial communication protocol.

[0010] An AI-based intelligent control method for raw water pretreatment chemical dosing, used to implement the AI-based intelligent control system for raw water pretreatment chemical dosing, includes the following steps: The raw water turbidity, pH value, flow rate, water temperature and COD data are collected in real time through a distributed sensor network, and Kalman filtering and LSTM interpolation algorithms are used for data cleaning and missing data repair. The cleaned multi-source data is spatiotemporally aligned to construct a multi-dimensional feature vector input to a hybrid prediction model. The hybrid prediction model uses a temporal convolutional network (TCN) to predict the turbidity of the sedimentation tank and combines it with a Transformer model to analyze long-term water quality trends and generate early warnings of algal blooms. Based on the prediction results, the Pareto optimal solution set is generated using the NSGA-II algorithm to balance the cost of the reagents and the treatment effect. The dosage is adjusted in advance through model predictive control (MPC) to compensate for the lag time of the coagulation reaction. The edge computing node outputs control commands to the execution layer in real time, driving the metering pump to add the agent according to the optimized dosage. At the same time, a dual-sensor deviation threshold detection and hot backup switching mechanism are adopted to ensure the dosing accuracy when a single sensor fails. The cloud platform layer regularly optimizes the parameters of the edge node model through a reinforcement learning framework and verifies the effectiveness of the control strategy based on a digital twin simulation engine.

[0011] Preferably, the data cleaning and missing data repair includes: Kalman filtering is used to eliminate electromagnetic interference noise in the sensor signal. The process noise covariance in the state equation is set to a preset first covariance value, and the observation noise covariance is set to a preset second covariance value. For missing data, the Long Short-Term Memory (LSTM) interpolation algorithm is used. The input window length is a preset time period, the number of hidden layer units is a preset number of network nodes, and the output is the sensor data at the missing time. The interpolation error does not exceed the preset error limit. Spatiotemporal alignment ensures that the time error of multi-source data is below the preset synchronization threshold and the spatial resolution meets the preset accuracy requirements by synchronizing timestamps and mapping spatial coordinates.

[0012] Preferably, the specific implementation of the model predictive control (MPC) is as follows: Using the water quality data for a predetermined future duration output from the hybrid prediction model as input, a rolling time-domain optimization problem is constructed, with the objective functions being to minimize the rate of change in chemical dosage and to minimize the sum of squared turbidity deviations. The constraints include the upper limit of metering pump flow rate, the safe threshold of reagent concentration, and the volume limit of the reaction tank; The optimal control sequence is calculated in real time by a quadratic programming solver, and the first control command is sent to the execution layer. The control period is a preset interval, and the lag compensation time error does not exceed the preset compensation threshold.

[0013] Preferably, the generation of real-time control commands for the edge computing node includes: The Pareto optimal solution set output by the NSGA-II algorithm is input into the fuzzy decision module, and the final dosage is selected according to the weight of the current working condition. The input variables of the fuzzy decision module are drug cost priority, environmental compliance priority, and treatment effect priority. The membership function adopts a preset distribution form, and the defuzzification method adopts a preset decision algorithm. The output is the metering pump frequency setting value, with a resolution that meets the preset accuracy range, and is transmitted to the frequency converter via the industrial bus protocol.

[0014] Preferably, the optimization of model parameters in the cloud platform layer includes: Periodic model retraining tasks are automatically triggered through containerized clusters, with training data consisting of historical data from edge nodes within a preset time range; A gradient descent optimizer is used, and the learning rate is adjusted according to a preset decay rule; After training, the control effect under different working conditions is simulated through a digital twin simulation engine. If the deviation between the simulation result and the measured data exceeds the preset verification threshold, a preset manual review process is triggered.

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: The intelligent control system and method for raw water pretreatment dosing proposed in this invention achieves accurate perception and prediction of raw water quality through multimodal data fusion and dynamic prediction models, allowing for advance adjustment of dosing dosage. This effectively compensates for the lag in coagulation reaction, improving the water quality compliance rate. Utilizing the NSGA-II multi-objective optimization algorithm, it balances reagent costs and treatment effects, avoiding excessive reagent addition in traditional methods, reducing operating costs and the risk of secondary pollution. The system adopts industrial-grade fault-tolerant design, such as dual-sensor cross-validation, metering pump hot backup, and edge computing local closed-loop, ensuring dosing accuracy under single-point failure and significantly improving system reliability and stability. The digital twin simulation engine and reinforcement learning framework at the cloud platform layer periodically optimize edge node model parameters, enabling continuous optimization and iteration of the control strategy and promoting the intelligentization of raw water pretreatment dosing control. Attached Figure Description

[0016] Figure 1 This is a system architecture and principle block diagram of the present invention; Figure 2 This is the control flow and flowchart of the present invention. Detailed Implementation

[0017] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0018] Reference Figure 1 As shown, the AI-based intelligent control system for raw water pretreatment and dosing includes: a sensor layer, an edge computing layer, a cloud platform layer, and an execution layer. The sensor layer consists of a distributed sensor network, including a turbidity meter, pH electrode, electromagnetic flow meter, COD online monitor and distributed fiber optic temperature measurement device, covering the raw water inlet, mixing tank and sedimentation tank nodes. It transmits signals to the edge computing layer through RS485 bus, and the deployment of multiple nodes realizes real-time monitoring of water quality parameters throughout the process. The edge computing layer includes an industrial gateway, which integrates a data cleaning module, a hybrid prediction model and an optimization control module. The hybrid prediction model combines a temporal convolutional network and a Transformer model for short-term water quality prediction and long-term trend analysis. The optimization control module includes model predictive control and the NSGA-II multi-objective optimization algorithm. Edge computing reduces cloud dependence and improves real-time response. The cloud platform layer is based on Kubernetes containerized deployment, integrates a digital twin simulation engine and reinforcement learning framework, regularly updates the edge computing layer model parameters, and the containerized architecture supports elastic expansion and seamless upgrades; The execution layer includes a servo motor-driven metering pump and a frequency converter, which receives instructions via the Profibus-DP protocol and dynamically adjusts the dosage. The high-precision execution mechanism keeps the dosage error within ±2%. The system achieves three levels of fault tolerance through redundant sensor cross-validation, metering pump hot backup, and edge computing local closed loop. Under single-point failure, the dosing accuracy error is ≤5%, and the three-level fault tolerance design ensures the continuous operation reliability of the system.

[0019] The functional architecture of the industrial gateway includes a multi-protocol adaptation module, a real-time data processing unit, and a fault-tolerant communication interface. The multi-protocol adaptation module supports dynamic conversion of RS485, Profibus-DP and Modbus RTU protocols, enabling unified access of heterogeneous devices at the sensor layer and execution layer, and reducing equipment modification costs due to protocol compatibility. The real-time data processing unit has a built-in lightweight inference engine that generates dosage optimization instructions based on the fusion prediction results of the temporal convolutional network TCN and the Transformer model. The lightweight engine makes the inference latency less than 50ms. The fault-tolerant communication interface adopts a dual-channel redundancy design. The main channel is 5G wireless transmission and the backup channel is industrial Ethernet. When the signal strength of the main channel is less than a preset threshold and the packet loss rate is greater than a preset percentage, it automatically switches to the backup channel. The dual-channel switching time is less than 200ms to ensure the continuity of instructions. The industrial gateway eliminates ground potential difference interference through the P+F KFD2-UT2-EX1 signal isolator, and the isolator has a withstand voltage rating of 2kV to prevent signal crosstalk.

[0020] The construction of the hybrid prediction model includes: Physical constraints are established based on the coagulation kinetic equation. The output of the TCN model is then weighted and fused with the long-term trend prediction results of the Transformer model. The weights are dynamically adjusted according to water quality changes to improve the interpretability of the prediction. The physical constraint formula is as follows: In the formula, For pollutant concentration over time Small changes within For a small change over time, This refers to the drug concentration. The reaction rate constant is... This represents the current turbidity concentration. The TCN model employs a dilated causal convolutional structure with a kernel size of 3, a dilation factor of 2^layers, and 5 layers. The output layer is mapped to the predicted turbidity value through a fully connected network to capture long-term temporal dependencies. Its dilated convolution formula is as follows: In the formula, For time steps The output value, The weight parameters of the convolution kernel, For the input sequence at time step eigenvalues, As the expansion factor, The kernel size; The Transformer model employs a multi-head self-attention mechanism, with an input sequence length of 72 hours and an output probability distribution of water quality trends for the next 24 hours. It optimizes model parameters using a cross-entropy loss function and enhances key feature extraction capabilities through a self-attention calculation formula, the expression of which is: In the formula, For querying the matrix, The key matrix, For value matrices, The dimension of the key vector; The NSGA-II multi-objective optimization algorithm is as follows: With the optimization objectives of minimizing reagent dosage, maximizing water quality compliance rate, and minimizing aluminum ion residue, a multi-objective function is constructed, the expression of which is: In the formula, For multi-objective functions Minimize operation, Let the objective function vector be the objective function vector of the multi-objective optimization problem. The objective function, which is directly related to the PAC dosage, needs to be minimized to reduce reagent costs. The dosage of polyaluminum chloride (PAC) is usually expressed in mg / L, and the compliance rate is the percentage of water quality parameters (such as turbidity, pH, COD, etc.) that meet the standards. The original goal of maximizing the achievement rate is transformed into minimizing its negative value to adapt to the optimization framework. To minimize aluminum ion residue and avoid environmental and health risks, This represents the concentration of residual aluminum ions after water treatment, expressed in mg / L.

[0021] The correlation between PAC dosage and turbidity, pH, and water temperature was analyzed using a mutual information matrix to generate an initial population. The correlation between variables was quantified using the mutual information formula, which is: In the formula, Let there be two random variables. for and The joint probability distribution, for and Marginal probability distribution, Mutual information value measures the nonlinear correlation between variables; The optimal solution set is selected by non-dominated sorting and crowding comparison. The offspring population is generated by simulating binary crossover and polynomial mutation. After iterative updating, the dosing scheme is output. Multi-objective optimization balances economic efficiency and environmental protection requirements.

[0022] The redundant sensor cross-validation specifically involves: When the data deviation between any two similar sensors exceeds a preset first threshold, the system automatically switches to the backup sensor. The first threshold is set according to the sensor type and is based on the ISO 15839 water quality analysis standard. The metering pump hot backup includes a main metering pump and a standby metering pump. When the main metering pump fails, it is switched to the standby pump through the industrial gateway within a preset switching time window. During the switching process, the fluctuation of the dosage is controlled within a preset allowable range, and the hot backup switching time is less than 5 seconds. The edge computing local closed loop achieves continuous control in the absence of network access through a lightweight machine learning model built into the industrial gateway. The local inference response time meets the preset real-time requirements, and it maintains data interaction with the programmable logic controller (PLC) through an industrial communication protocol. The local closed loop maintains the stability of the dosing control.

[0023] Reference Figure 2As shown, an AI-based intelligent control method for raw water pretreatment dosing, used to implement the AI-based intelligent control system for raw water pretreatment dosing, includes the following steps: Data on raw water turbidity, pH, flow rate, temperature, and COD are collected in real time using a distributed sensor network. Kalman filtering and LSTM interpolation algorithms are employed for data cleaning, missing data repair, and suppression of high-frequency noise. The Kalman filtering formula is as follows: In the formula, For time steps Prior state estimation, Here is the state transition matrix. To control the input matrix, This is the control input from the previous moment. This is an estimate of the posterior state from the previous time step; The cleaned multi-source data is spatiotemporally aligned to construct a multi-dimensional feature vector input to a hybrid prediction model. This hybrid prediction model uses a temporal convolutional network (TCN) to predict the turbidity of the sedimentation tank and combines it with a Transformer model to analyze long-term water quality trends, generating early warnings of algal blooms with a trend prediction accuracy of 90%. Based on the prediction results, the NSGA-II algorithm is used to generate a Pareto optimal solution set to balance the cost of the reagents and the treatment effect. Model predictive control (MPC) is then used to adjust the dosage in advance to compensate for the lag time of the coagulation reaction, achieving dynamic control. The MPC rolling optimization formula is as follows: In the formula, To predict the output, As the target reference value, To control the change in input (dosage), These are weighting coefficients used to balance tracking error and changes in control input. To predict the length of the time domain; The edge computing node outputs control commands to the execution layer in real time, driving the metering pump to add the agent according to the optimized dosage. At the same time, a dual-sensor deviation threshold detection and hot backup switching mechanism are adopted to ensure the dosing accuracy when a single sensor fails, and the deviation detection response time is less than 1 second. The cloud platform layer regularly optimizes the edge node model parameters through a reinforcement learning framework and verifies the effectiveness of the control strategy based on a digital twin simulation engine, with the digital twin simulation error being less than 3%.

[0024] In summary, the advantages of this invention are: significantly improved accuracy and efficiency of chemical dosing control. By deploying a multi-dimensional sensor network, comprehensive water quality sensing is achieved, data integrity is greatly improved, and water quality fluctuations can be captured in real time. Utilizing a hybrid model to fuse short-term predictions and long-term trend analysis, the dosing dosage is adjusted in advance, effectively compensating for coagulation reaction lag and ensuring stable water quality compliance. Simultaneously, the system balances reagent cost and treatment effect through a multi-objective optimization algorithm, achieving precise proportioning and reducing reagent waste and secondary pollution. Furthermore, a three-level fault-tolerant system and edge computing local closed-loop design ensure stable operation even under single-point failures, significantly reducing maintenance costs.

[0025] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. An AI-based intelligent control system for raw water pretreatment dosing, characterized in that: include: Sensor layer, edge computing layer, cloud platform layer, and execution layer; The sensor layer consists of a distributed sensor network, including a turbidity meter, pH electrode, electromagnetic flow meter, COD online monitor and distributed fiber optic temperature measurement device, covering the raw water inlet, mixing tank and sedimentation tank nodes, and transmitting signals to the edge computing layer through RS485 bus; The edge computing layer includes an industrial gateway, which integrates a data cleaning module, a hybrid prediction model and an optimization control module. The hybrid prediction model combines a temporal convolutional network and a Transformer model for short-term water quality prediction and long-term trend analysis. The optimization control module includes model predictive control and the NSGA-II multi-objective optimization algorithm. The cloud platform layer is deployed based on Kubernetes containerization, integrates a digital twin simulation engine and a reinforcement learning framework, and regularly updates the model parameters of the edge computing layer. The execution layer includes a metering pump driven by a servo motor and a frequency converter, which receives instructions via the Profibus-DP protocol and dynamically adjusts the dosage. The system achieves three levels of fault tolerance through cross-validation of redundant sensors, hot backup of metering pumps, and local closed-loop edge computing, with a dosing accuracy error of ≤5% under single-point failure.

2. The AI-based intelligent control system for raw water pretreatment dosing according to claim 1, characterized in that, The functional architecture of the industrial gateway includes a multi-protocol adaptation module, a real-time data processing unit, and a fault-tolerant communication interface. The multi-protocol adaptation module supports dynamic conversion of RS485, Profibus-DP and Modbus RTU protocols, enabling unified access for heterogeneous devices at the sensor and execution layers. The real-time data processing unit has a built-in lightweight inference engine that generates dosage optimization instructions based on the fusion prediction results of the temporal convolutional network TCN and the Transformer model. The fault-tolerant communication interface adopts a dual-channel redundancy design. The main channel is 5G wireless transmission and the backup channel is industrial Ethernet. When the signal strength of the main channel is less than a preset threshold and the packet loss rate is greater than a preset percentage, it will automatically switch to the backup channel. The industrial gateway eliminates ground potential difference interference through the P+F KFD2-UT2-EX1 signal isolator.

3. The AI-based intelligent control system for raw water pretreatment dosing as described in claim 1, characterized in that, The construction of the hybrid prediction model includes: Physical constraints are established based on the coagulation kinetic equation. The output of the TCN model is weighted and fused with the long-term trend prediction results of the Transformer model. The weights are dynamically adjusted according to water quality changes. The TCN model adopts an expanded causal convolutional structure with a kernel size of 3, an expansion factor of 2^layers, and 5 layers. The output layer is mapped to the predicted turbidity value through a fully connected network. The Transformer model employs a multi-head self-attention mechanism, with an input sequence length of 72 hours and an output probability distribution of the water quality trend for the next 24 hours. The model parameters are optimized using a cross-entropy loss function.

4. The AI-based intelligent control system for raw water pretreatment dosing according to claim 1, characterized in that, The NSGA-II multi-objective optimization algorithm is as follows: A multi-objective function is constructed with the optimization objectives of minimizing the dosage of chemicals, maximizing the water quality compliance rate, and minimizing the residual aluminum ions. The correlation between PAC dosage and turbidity, pH value, and water temperature was analyzed using mutual information matrix analysis to generate an initial population. The optimal solution set is selected by non-dominated sorting and crowding comparison, and the offspring population is generated by simulating binary crossover and polynomial mutation. The dosing scheme is output after iterative update.

5. The AI-based intelligent control system for raw water pretreatment dosing according to claim 1, characterized in that, The redundant sensor cross-validation specifically involves: When the data deviation between any two similar sensors exceeds a preset first threshold, the system automatically switches to a backup sensor, wherein the first threshold is set according to the sensor type. The metering pump hot backup includes a main metering pump and a standby metering pump. When the main metering pump fails, it is switched to the standby pump through the industrial gateway within a preset switching time window. During the switching process, the fluctuation of the dosage is controlled within a preset allowable range. The edge computing local closed loop achieves continuous control in the absence of network access through a lightweight machine learning model built into the industrial gateway. The local inference response time meets the preset real-time requirements, and it maintains data interaction with the programmable logic controller (PLC) through an industrial communication protocol.

6. An AI-based intelligent control method for chemical dosing in raw water pretreatment, characterized in that: To implement the AI-based intelligent control system for raw water pretreatment dosing as described in any one of claims 1-5, the system includes the following steps: The raw water turbidity, pH value, flow rate, water temperature and COD data are collected in real time through a distributed sensor network, and Kalman filtering and LSTM interpolation algorithms are used for data cleaning and missing data repair. The cleaned multi-source data is spatiotemporally aligned to construct a multi-dimensional feature vector input to a hybrid prediction model. The hybrid prediction model uses a temporal convolutional network (TCN) to predict the turbidity of the sedimentation tank and combines it with a Transformer model to analyze long-term water quality trends and generate early warnings of algal blooms. Based on the prediction results, the Pareto optimal solution set is generated using the NSGA-II algorithm to balance the cost of the reagents and the treatment effect. The dosage is adjusted in advance through model predictive control (MPC) to compensate for the lag time of the coagulation reaction. The edge computing node outputs control commands to the execution layer in real time, driving the metering pump to add the agent according to the optimized dosage. At the same time, a dual-sensor deviation threshold detection and hot backup switching mechanism are adopted to ensure the dosing accuracy when a single sensor fails. The cloud platform layer regularly optimizes the parameters of the edge node model through a reinforcement learning framework and verifies the effectiveness of the control strategy based on a digital twin simulation engine.

7. The AI-based intelligent control method for raw water pretreatment chemical dosing according to claim 6, characterized in that, The data cleaning and missing data repair include: Kalman filtering is used to eliminate electromagnetic interference noise in the sensor signal. The process noise covariance in the state equation is set to a preset first covariance value, and the observation noise covariance is set to a preset second covariance value. For missing data, the Long Short-Term Memory (LSTM) interpolation algorithm is used. The input window length is a preset time period, the number of hidden layer units is a preset number of network nodes, and the output is the sensor data at the missing time. The interpolation error does not exceed the preset error limit. Spatiotemporal alignment ensures that the time error of multi-source data is below the preset synchronization threshold and the spatial resolution meets the preset accuracy requirements by synchronizing timestamps and mapping spatial coordinates.

8. The AI-based intelligent control method for raw water pretreatment dosing according to claim 6, characterized in that, The specific implementation of the model predictive control (MPC) is as follows: Using the water quality data for a predetermined future duration output from the hybrid prediction model as input, a rolling time-domain optimization problem is constructed, with the objective functions being to minimize the rate of change in chemical dosage and to minimize the sum of squared turbidity deviations. The constraints include the upper limit of metering pump flow rate, the safe threshold of reagent concentration, and the volume limit of the reaction tank; The optimal control sequence is calculated in real time by a quadratic programming solver, and the first control command is sent to the execution layer. The control period is a preset interval, and the lag compensation time error does not exceed the preset compensation threshold.

9. The AI-based intelligent control method for raw water pretreatment chemical dosing according to claim 6, characterized in that, The generation of real-time control commands for the edge computing node includes: The Pareto optimal solution set output by the NSGA-II algorithm is input into the fuzzy decision module, and the final dosage is selected according to the weight of the current working condition. The input variables of the fuzzy decision module are drug cost priority, environmental compliance priority, and treatment effect priority. The membership function adopts a preset distribution form, and the defuzzification method adopts a preset decision algorithm. The output is the metering pump frequency setting value, with a resolution that meets the preset accuracy range, and is transmitted to the frequency converter via the industrial bus protocol.

10. The AI-based intelligent control method for raw water pretreatment dosing according to claim 6, characterized in that, The optimization of model parameters in the cloud platform layer includes: Periodic model retraining tasks are automatically triggered through containerized clusters, with training data consisting of historical data from edge nodes within a preset time range; A gradient descent optimizer is used, and the learning rate is adjusted according to a preset decay rule; After training, the control effect under different working conditions is simulated through a digital twin simulation engine. If the deviation between the simulation result and the measured data exceeds the preset verification threshold, a preset manual review process is triggered.

Images