Coagulant adaptive adding control method and system based on physical constraint neural network
By adopting a coagulant adaptive dosing control method based on physical constraint neural networks, the problems of weak adaptability and high chemical consumption in water treatment systems when facing complex raw water quality are solved. This method achieves stability of effluent quality and reduces operating costs, and has strong adaptability and forward-looking prediction capabilities, ensuring the safety and reliability of the system.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NANJING INST OF TECH
- Filing Date
- 2026-02-06
- Publication Date
- 2026-05-05
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing water treatment systems have weak adaptability, poor predictability, and high chemical consumption when faced with complex and ever-changing raw water quality, resulting in unstable effluent quality and increased operating costs.
An adaptive coagulant dosing control method based on physical constraint neural networks is adopted. By embedding the constraint mechanism of physical and chemical laws in the data preprocessing, neural network training and rolling optimization layers, the method realizes full-link safety control from data acquisition to control command generation. Combined with a time-series dynamic neural network model, the method predicts and optimizes the future dosage and effluent turbidity.
It achieves strong self-adaptation capability to raw water quality, improves predictability and stability of effluent quality, reduces coagulant consumption, enhances system economy and reliability, and switches to conservative control mode to ensure safety in abnormal situations.
Smart Images
Figure CN121979154A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of water treatment automatic control technology, specifically to an adaptive coagulant dosing control method and system based on a physical constraint neural network. Background Technology
[0002] The water treatment system of a thermal power plant is the lifeline for its safe, efficient, and economical operation, and raw water treatment (or makeup water treatment) is the starting point and fundamental guarantee of the entire water treatment process. Raw water is the initial water source (surface water, groundwater, or reclaimed water) for the water treatment system of a thermal power plant. Untreated raw water contains a large number and variety of impurities and can never be directly used for the plant's critical thermal equipment. The fundamental role of raw water treatment is to remove these impurities and produce high-quality, usable "makeup water" for the entire power plant. In traditional water treatment coagulation processes, coagulant dosing control mainly relies on empirical models based on fixed parameters, proportional-integral-derivative (PID) control, or feedforward-feedback composite control strategies. These methods have significant limitations when facing complex fluctuations in raw water quality (such as turbidity, temperature, pH, and pollutant composition): poor adaptability, making it difficult to adapt to sudden changes in water quality or seasonal variations; insufficient predictability, resulting in delayed control response; and low dosing accuracy, easily leading to over- or under-dosing, causing unstable effluent quality, increased reagent consumption, and higher operating costs. Furthermore, while data-driven models in emerging technologies possess some learning capabilities, they lack explicit constraints based on physicochemical mechanisms, and their outputs under extreme or unseen conditions may not conform to actual patterns. Although deep learning-based visual inspection solutions add visual inspection results as control parameters to traditional water quality parameter-based control, visual inspection itself suffers from various limitations due to application scenarios, such as the susceptibility of monitoring equipment to corrosion from raw water and the poor light transmittance of raw water. The aforementioned solutions lack sufficient engineering reliability and robustness to meet the requirements of water treatment scenarios.
[0003] In existing technologies, there is a "divide and conquer" strategy that divides working conditions based on fixed rules and calls different models, such as... Figure 1 As shown, however, this method requires a large-scale model to be trained, resulting in low efficiency.
[0004] Therefore, there is an urgent need for an intelligent dosing control method that can deeply integrate the physical mechanisms of water quality changes with real-time monitoring data and has forward-looking prediction and strong adaptive capabilities, so as to achieve precise and optimized control of the coagulation process. Summary of the Invention
[0005] This invention addresses the shortcomings of existing technologies by providing a coagulant adaptive dosing control method and system based on a physically constrained neural network. By embedding constraint mechanisms strongly coupled with the physicochemical laws of the coagulation process into the data preprocessing layer, neural network training layer, rolling optimization layer, and safety fallback layer, the entire chain from data acquisition to control command generation meets process safety requirements. This solves the problems of weak adaptability, poor predictability, and high chemical consumption in traditional control methods when facing complex and variable raw water quality, achieving stable effluent quality compliance and precise optimization of coagulant dosage.
[0006] To achieve the above objectives, the present invention adopts the following technical solution:
[0007] An adaptive dosing control method for coagulants based on a physical constraint neural network includes the following steps: Collect water quality data; Water quality data are preprocessed to obtain standardized water quality sequences; Standardized water quality sequences are input into a pre-trained time-series dynamic neural network model to predict future dosage and future effluent turbidity. At the end of each control cycle, based on the model predictive control mechanism, with the optimization objective of ensuring that the future effluent turbidity predicted by the time-series dynamic neural network model does not exceed the standard and minimizing the future dosage, rolling optimization is performed on the pre-trained time-series dynamic neural network model to solve for the optimal reagent dosing sequence. The optimal drug dosing sequence is converted into a control signal to drive the dosing pump to administer the drug.
[0008] To optimize the above technical solution, the specific measures also include: Furthermore, the water quality data includes influent flow rate, influent turbidity, influent pH value, water temperature, and sedimentation tank effluent turbidity; The water quality data preprocessing specifically involves: Z-score normalization was applied to the influent pH and temperature; Min-Max normalization was applied to the influent turbidity, influent flow rate, and sedimentation tank effluent turbidity; thus, a standardized water quality sequence for the past N control periods was obtained.
[0009] In the formula, It is the past number Standardized water quality data for each control cycle, It is the standardized water quality data for the current control period t.
[0010] Furthermore, the loss function of the temporal dynamic neural network model in the pre-training stage includes a mean squared error prediction error term and a physical constraint penalty term, expressed as follows:
[0011] In the formula, It is the total loss function. It is the mean square error. To constrain the weighting coefficients, It is a physical constraint penalty term, which is the sum of one or more of the non-negativity constraints, monotonicity constraints, and boundary constraints, with the priority being boundary constraints > monotonicity constraints > non-negativity constraints.
[0012] Furthermore, the nonnegativity constraint is used to penalize the negative dosage prediction value output by the time-series dynamic neural network model, and the expression is as follows:
[0013] In the formula, Indicates a nonnegativity constraint. This is the predicted dosage of medication.
[0014] Furthermore, the expression for the monotonicity constraint is as follows:
[0015] In the formula, This represents the monotonicity constraint, where k is the index of the water quality data other than the influent turbidity. and These represent the predicted dosage for the sample with relatively low influent turbidity and the predicted dosage for the sample with relatively high influent turbidity, respectively, in a pair of input water quality data samples.
[0016] Furthermore, the boundary constraint is used to penalize predicted dosages that exceed the process allowable range, as expressed below:
[0017] In the formula, It is a boundary constraint. This is the predicted dosage. This is the lower limit of the dosage. This is the upper limit for the amount of medicine to be administered.
[0018] Furthermore, the optimization objective of ensuring that the future effluent turbidity predicted by the time-series dynamic neural network model does not exceed the standard and minimizing the future dosage is specifically as follows: The objective function of the optimization problem is as follows:
[0019] In the formula, This represents the drug dosing sequence from the (t+1)th control period to the (t+H)th control period, where H is the prediction step size. It is the effluent turbidity predicted by the pre-trained time-series dynamic neural network model for the (t+k)th control cycle. For target turbidity, To control the weights, This represents the dosage in the (t+k)th control period; The constraints of the optimization problem are as follows: Dosage range constraints:
[0020] In the formula, This is the lower limit of the dosage. This is the upper limit for the dosage of the drug; Effluent water quality constraints:
[0021] In the formula, It is the upper limit of turbidity.
[0022] Furthermore, the specific steps of converting the optimal drug dosing sequence into a control signal to drive the dosing pump to administer the drug are as follows: The recommended dosage for each control cycle in the optimal dosing sequence is multiplied by the real-time influent flow rate to obtain the instantaneous dosing command. The formula is as follows:
[0023] In the formula, This is a dosage instruction. It is the inflow rate collected in real time. It is the dosage recommended by the model for the (t+1)th control cycle; The dosage command is converted into an analog current signal to drive the dosing pump to deliver the drug.
[0024] Furthermore, the method also includes: The security mechanism will be activated if any of the following conditions are met: The turbidity of the effluent continuously exceeds the set value for a predetermined time; The dosing equipment malfunctions and alarms, or the actual dosing amount deviates from the commanded value beyond the allowable range; The confidence level of the output of the temporal dynamic neural network model is lower than the threshold; The security mechanism includes: Immediately interrupt the control signal driving the dosing pump, automatically switch to the pre-configured PID control mode, trigger the audible and visual alarm in the control room, and record the abnormal event in the log.
[0025] This invention also proposes an adaptive coagulant dosing control system based on a physically constrained neural network, comprising: The sensing layer is used to collect water quality data; The data and cognition layer is used to preprocess water quality data to obtain a standardized water quality sequence. The standardized water quality sequence is then input into a pre-trained time-series dynamic neural network model to predict future dosage and future effluent turbidity. The optimization and decision layer is used at the end of each control cycle to perform rolling optimization on the pre-trained time-series dynamic neural network model based on the model predictive control mechanism, with the optimization objectives being that the future effluent turbidity predicted by the time-series dynamic neural network model does not exceed the standard and the future dosage is minimized, in order to solve for the optimal chemical dosing sequence. The execution and safety layer is used to convert the optimal drug dosing sequence into control signals to drive the dosing pump to administer the drug.
[0026] The beneficial effects of this invention are: (1) Strong adaptability: Through a neural network with time-series memory capability, it automatically learns and memorizes the complex nonlinear time-series relationship between water quality parameters and coagulation effect, and can effectively adapt to various fluctuations in raw water quality.
[0027] (2) Predictive and forward-looking: The model can make multi-step predictions and can be combined with the Model Predictive Control (MPC) framework to achieve proactive optimization control based on future state predictions and smooth water quality fluctuations.
[0028] (3) Integration of mechanism and data credibility: By introducing physical constraints into the loss function, chemical common sense and expert experience are explicitly encoded into the model, ensuring that the output of the artificial intelligence model conforms to basic scientific laws and improving the extrapolation robustness and engineering credibility under unknown working conditions.
[0029] (4) Economical and efficient operation: Through accurate prediction and optimized addition, the average consumption of coagulant can be significantly reduced while ensuring stable effluent quality, thus saving operating costs.
[0030] (5) The system is safe and reliable: a comprehensive safety risk protection mechanism has been designed, which can seamlessly switch to conservative control mode when the intelligent model fails or the process is abnormal, so as to ensure the continuity and safety of the water treatment process. Attached Figure Description
[0031] Figure 1 This is a schematic diagram of a typical coagulant dosing control method in the prior art.
[0032] Figure 2 This is a flowchart of an adaptive coagulant dosing control method based on a physical constraint neural network proposed in this invention.
[0033] Figure 3 This is a control principle diagram of an adaptive coagulant dosing control system based on a physical constraint neural network proposed in this invention. Detailed Implementation
[0034] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0035] Example 1 This invention proposes an adaptive coagulant dosing control method based on a physical constraint neural network. The process of this method is as follows: Figure 2 As shown, it includes the following steps: S1. Collect water quality data; water quality data includes influent flow rate, influent turbidity, influent pH value, water temperature, and sedimentation tank effluent turbidity; deploy industrial-grade sensors at key process nodes: install flow meters and turbidity meters on the raw water influent pipeline, install pH meters and thermometers near the dosing point, and install turbidity meters in the sedimentation tank effluent channel. All sensor signals are collected in real time by a programmable logic controller (PLC) and uploaded to a host industrial control computer to form the system state vector at the current moment. The vector structure is as follows:
[0036] In the formula, It is the inflow rate. It refers to the turbidity of the incoming water. It is the pH value of the influent. It's the water temperature. It refers to the turbidity of the effluent from the sedimentation tank; S2. Preprocess the water quality data to obtain a standardized water quality sequence; After receiving the raw data, the host computer first performs data cleaning, using moving average filtering to eliminate random noise, and sets reasonable ranges to remove abnormal sensor values (reasonable range: turbidity 0~4000 NTU, pH 4~10, temperature 0~40℃; values exceeding these ranges are marked as abnormal and filled with the previous period's value). Subsequently, each parameter is standardized (normalized) to eliminate the influence of dimensions, as follows: For data that approximately follows a normal distribution or has no clear boundaries (such as pH values and temperature changes), Z-score standardization can be used to make the processed data have a mean of 0 and a standard deviation of 1. The calculation method is as follows:
[0037] in, This is the original data. The mean of the samples for this feature in the training set. The standard deviation of this feature in the training set.
[0038] For data with clear boundaries or requiring confinement to a specific range (such as turbidity or flow rate), Min-Max normalization can be used to linearly map the data to the [0,1] or [-1,1] interval. The calculation method is as follows:
[0039] in, and The minimum and maximum values of this feature are for the training set.
[0040] Key design features of this embodiment: Data corresponding to non-negative constraints is normalized to [0, 1] using Min-Max to ensure consistency between the numerical space and the constraint semantics; data corresponding to monotonic constraints is standardized using Z-score to preserve relative size relationships. This preprocessing-constraint matching rule is fixed in a configuration file before model training and cannot be changed during runtime, preventing preprocessing from disrupting the constraint logic.
[0041] The statistical parameters (mean, standard deviation, minimum, and maximum) required for standardization should be calculated and fixed from the historical training dataset for subsequent real-time processing of online data. Based on the processed standardized data, a standardized state vector sequence containing the past N control cycles is constructed:
[0042] In the formula, It is the past number Standardized water quality data for each control cycle, It is the standardized water quality data for the current control period t.
[0043] S3. Input the standardized water quality sequence into the pre-trained time-series dynamic neural network model to predict the future dosage and future effluent turbidity.
[0044] The temporal dynamic neural network model employs multiple layers of recurrent neural networks with memory units. These memory units selectively retain or forget historical information through a gating mechanism, thereby effectively capturing the dynamic characteristics and hysteresis effects of water quality changes. To more clearly illustrate the working principle of these memory units, taking a Long Short-Term Memory (LSTM) network as an example, its forward propagation process can be described as follows: Forgotten Gate:
[0045] Input Gate:
[0046]
[0047] Cell status update:
[0048] Output gate:
[0049]
[0050] in Enter the current time. In hidden state, In cellular state, For the sigmoid function, This indicates element-wise multiplication. The network is followed by a fully connected layer, which maps the temporal features to the predicted values.
[0051] This invention does not limit the specific implementation structure of the temporal dynamic neural network. In one specific embodiment of this invention, a long short-term memory network structure can be used to implement the memory unit and gating mechanism. As a comparison, a gated recurrent unit (GRU) can also be used to achieve temporal modeling, but experiments in this embodiment show that LSTM has higher accuracy when the lag exceeds 3 cycles.
[0052] In the loss function during the training phase, in addition to the mean squared error (MSE) prediction error term, a physical constraint penalty term is added. The total loss function can be expressed as:
[0053] In the formula, It is the total loss function. It is the mean square error. `x` is a positive real number, representing the constraint weight coefficient, used to adjust the relative importance of the prediction error term and the physical constraint term during model training. This is achieved through offline grid search (step size 0.1). The optimal comprehensive index is then burned into the model configuration file.
[0054] It is a physical constraint penalty term, which is the sum of one or more of the non-negativity constraints, monotonicity constraints, and boundary constraints, with the priority being boundary constraints > monotonicity constraints > non-negativity constraints.
[0055] Non-negativity constraint loss: This penalty is applied to negative drug dosage predictions in the model output. The formula is as follows:
[0056] Monotonicity constraint loss: To incorporate the physicochemical principles of coagulation processes, i.e., when other input conditions (such as flow rate, pH, and temperature) are the same or similar (in this embodiment, "differences in other characteristics ≤ 5% historical standard deviation" is defined as similar), higher influent turbidity typically requires more coagulant to achieve the desired treatment effect. This constraint is achieved through the following methods: In batch training, for input features and To quantify the condition of 'similar characteristics', this embodiment defines it as follows: if for each characteristic except for influent turbidity... (Including influent flow rate, pH value, water temperature, and turbidity of sedimentation tank effluent), satisfying:
[0057] in , Features The standard deviation of the samples in the training set.
[0058] And its characteristics Then the predicted dosage value corresponding to the model must satisfy:
[0059] The corresponding penalty items are defined as follows:
[0060] in and These represent the predicted dosage for samples with low and high influent turbidity, respectively, assuming that other input features of these two samples are deemed similar. This penalty term is activated only when the predicted dosage for the low-turbidity sample is actually higher.
[0061] Boundary constraint loss: This penalty applies to predicted feed rates exceeding process limits (such as maximum equipment capacity or historical safety limits) to ensure the output remains within the feasible region. Its function is defined as follows:
[0062] This embodiment sets the priority for constraint conflicts as follows: boundary constraints (equipment capability) > monotonic constraints (process common sense) > nonnegative constraints (basic physics). When a boundary constraint penalty is triggered, the monotonic constraint penalty is automatically masked to prevent equipment overload. This priority rule is hard-coded into the training scheduler, rather than being a simple weighted sum of loss functions, ensuring absolute equipment safety under extreme operating conditions.
[0063] The temporal dynamic neural network model is trained using a deep learning framework. After training, the model can be converted to the Open Neural Network Exchange (ONNX) format or packaged as a dynamic link library (DLL) for deployment in industrial host computers.
[0064] S4. At the end of each control cycle, based on the model predictive control mechanism, with the optimization objective of ensuring that the future effluent turbidity predicted by the time-series dynamic neural network model does not exceed the standard and the future dosage is minimized, rolling optimization is performed on the pre-trained time-series dynamic neural network model to solve for the optimal dosing sequence. The objective function of the optimization problem is as follows:
[0065] In the formula, This represents the drug dosing sequence from the (t+1)th control period to the (t+H)th control period, where H is the prediction step size. It is the effluent turbidity predicted by the pre-trained time-series dynamic neural network model for the (t+k)th control cycle. For target turbidity, To control the weights, This represents the dosage in the (t+k)th control period; The constraints of the optimization problem are as follows: Dosage range constraints:
[0066] In the formula, This is the lower limit of the dosage. This is the upper limit for the dosage of the drug; Effluent water quality constraints:
[0067] In the formula, It is the upper limit of turbidity.
[0068] The optimal dosing sequence is converted into a control signal to drive the dosing pump. The recommended dosage for each control cycle in the optimal dosing sequence is multiplied by the real-time influent flow rate to obtain the instantaneous dosing command. The formula is as follows:
[0069] In the formula, This is a dosage instruction. It is the inflow rate collected in real time. This is the dosage recommended by the model for the (t+1)th control period; the dosage command generated by rolling optimization is converted into a 4-20mA analog signal (corresponding to a dosage of 0~20mg / L), denoted as the command. .
[0070] S5. Control execution and safety risk mitigation.
[0071] Execution layer receives instructions The system uses a frequency converter or regulating valve to drive the dosing pump for precise dosing. Simultaneously, the system continuously monitors: Execution feedback: Actual frequency, stroke, or flow rate of the dosing pump. Effect feedback: Real-time value of turbidity in the sedimentation tank effluent.
[0072] The system presets the following risk fallback conditions, and the safety mechanism will be activated if any one of the conditions is triggered: (1) The turbidity of the effluent exceeds the set value for a predetermined time (3 control cycles, 15 min). (2) The dosing equipment malfunctions and alarms, or the actual dosing amount deviates from the command value beyond the allowable range, with allowable deviation = command value ± 10%. (3) The confidence level of the neural network model output is lower than the threshold, or the input data is identified as a serious anomaly.
[0073] The model confidence score calculation method in this embodiment is as follows: Prediction variance is obtained through Bayesian dropout, meaning dropout remains active during the prediction phase. The next forward propagation (in this embodiment, it is taken as...) ), to obtain the predicted sample set Calculate its variance as a confidence index:
[0074] in To predict the mean. When If the confidence level is insufficient, a safety fallback mechanism is triggered. This design enables the model to have self-diagnostic capabilities, allowing it to identify uncertainties in the prediction results and improve system security and reliability.
[0075] This design differs from the traditional PID single threshold trigger, realizing AI model self-diagnosis and decoupling from process anomalies, which is the core of intelligent control safety.
[0076] The safety net measures are as follows: (1) Immediately interrupt and override the optimization control command sent by S4.
[0077] (2) Automatically switch to the pre-configured conservative PID control mode (proportional coefficient Kp=0.5).
[0078] (3) Trigger the audible and visual alarm in the control room and record the abnormal event in the log.
[0079] Once the operators intervene, the fault is resolved, and the water quality stabilizes, the system can automatically switch back to the intelligent control closed loop (after the anomaly is resolved and the water quality remains stable for 10 cycles).
[0080] Example 2 This invention proposes a coagulant adaptive dosing control system based on a physical constraint neural network, corresponding to the method in Example 1, such as... Figure 3 ,include: The sensing layer is used to collect water quality data; The data and cognition layer is used to preprocess water quality data to obtain a standardized water quality sequence. The standardized water quality sequence is then input into a pre-trained time-series dynamic neural network model to predict future dosage and future effluent turbidity. The optimization and decision layer is used at the end of each control cycle to perform rolling optimization on the pre-trained time-series dynamic neural network model based on the model predictive control mechanism, with the optimization objectives being that the future effluent turbidity predicted by the time-series dynamic neural network model does not exceed the standard and the future dosage is minimized, in order to solve for the optimal chemical dosing sequence. The execution and safety layer is used to convert the optimal drug dosing sequence into control signals to drive the dosing pump to administer the drug, and also serves as a safety risk mitigation layer.
[0081] The implementation methods of each module and its function in the system are completely consistent with the steps of the method in Implementation Example 1, so they will not be repeated here.
[0082] like Figure 1 and Figure 3 As shown, compared with the "divide and conquer" strategy of dividing working conditions based on fixed rules and calling different models in the prior art, the present invention uses a unified, physically constrained end-to-end intelligent model to process all working conditions, and achieves multi-step forward control through rolling optimization, realizing the leap from "reactive correction" to "proactive intervention".
[0083] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0084] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A method for adaptive dosing control of coagulants based on a physical constraint neural network, characterized in that, Includes the following steps: Collect water quality data; Water quality data are preprocessed to obtain standardized water quality sequences; Standardized water quality sequences are input into a pre-trained time-series dynamic neural network model to predict future dosage and future effluent turbidity. At the end of each control cycle, based on the model predictive control mechanism, with the optimization objective of ensuring that the future effluent turbidity predicted by the time-series dynamic neural network model does not exceed the standard and minimizing the future dosage, rolling optimization is performed on the pre-trained time-series dynamic neural network model to solve for the optimal reagent dosing sequence. The optimal drug dosing sequence is converted into a control signal to drive the dosing pump to administer the drug.
2. The adaptive dosing control method for coagulants based on a physical constraint neural network as described in claim 1, characterized in that, The water quality data includes influent flow rate, influent turbidity, influent pH value, water temperature, and sedimentation tank effluent turbidity; The water quality data preprocessing specifically involves: Z-score normalization was applied to the influent pH and temperature; Min-Max normalization was applied to the influent turbidity, influent flow rate, and sedimentation tank effluent turbidity; thus, a standardized water quality sequence for the past N control periods was obtained. In the formula, It is the past number Standardized water quality data for each control cycle, It is the standardized water quality data for the current control period t.
3. The adaptive dosing control method for coagulants based on a physical constraint neural network as described in claim 1, characterized in that, The loss function of the temporal dynamic neural network model in the pre-training stage includes a mean squared error prediction error term and a physical constraint penalty term, expressed as follows: In the formula, It is the total loss function. It is the mean square error. To constrain the weighting coefficients, It is a physical constraint penalty term, which is the sum of one or more of the non-negativity constraints, monotonicity constraints, and boundary constraints, with the priority being boundary constraints > monotonicity constraints > non-negativity constraints.
4. The adaptive dosing control method for coagulants based on a physical constraint neural network as described in claim 3, characterized in that, The nonnegativity constraint is used to penalize the negative dosage prediction value output by the time-series dynamic neural network model, and its expression is as follows: In the formula, Indicates a nonnegativity constraint. This is the predicted dosage of medication.
5. The adaptive dosing control method for coagulants based on a physical constraint neural network as described in claim 3, characterized in that, The expression for the monotonicity constraint is as follows: In the formula, This represents the monotonicity constraint, where k is the index of the water quality data other than the influent turbidity. and These represent the predicted dosage for the sample with relatively low influent turbidity and the predicted dosage for the sample with relatively high influent turbidity, respectively, in a pair of input water quality data samples.
6. The adaptive dosing control method for coagulants based on a physical constraint neural network as described in claim 3, characterized in that, The boundary constraint is used to penalize predicted dosages that exceed the process allowable range, and is expressed as follows: In the formula, It is a boundary constraint. This is the predicted dosage. This is the lower limit of the dosage. This is the upper limit for the amount of medicine to be administered.
7. The adaptive dosing control method for coagulants based on a physical constraint neural network as described in claim 1, characterized in that, The optimization objective, which is to ensure that the future effluent turbidity predicted by the time-series dynamic neural network model does not exceed the standard and to minimize the future chemical dosage, is as follows: The objective function of the optimization problem is as follows: In the formula, This represents the drug dosing sequence from the (t+1)th control period to the (t+H)th control period, where H is the prediction step size. It is the effluent turbidity predicted by the pre-trained time-series dynamic neural network model for the (t+k)th control cycle. For target turbidity, To control the weights, This represents the dosage in the (t+k)th control period; The constraints of the optimization problem are as follows: Dosage range constraints: In the formula, This is the lower limit of the dosage. This is the upper limit for the dosage of the drug; Effluent water quality constraints: In the formula, It is the upper limit of turbidity.
8. The adaptive dosing control method for coagulants based on a physical constraint neural network as described in claim 1, characterized in that, The specific steps of converting the optimal drug dosing sequence into a control signal to drive the dosing pump to administer the drug are as follows: The recommended dosage for each control cycle in the optimal dosing sequence is multiplied by the real-time influent flow rate to obtain the instantaneous dosing command. The formula is as follows: In the formula, This is a dosage instruction. It is the inflow rate collected in real time. It is the dosage recommended by the model for the (t+1)th control cycle; The dosage command is converted into an analog current signal to drive the dosing pump to deliver the drug.
9. The adaptive dosing control method for coagulants based on a physical constraint neural network as described in claim 1, characterized in that, The method further includes: The security mechanism will be activated if any of the following conditions are met: The turbidity of the effluent continuously exceeds the set value for a predetermined time; The dosing equipment malfunctions and alarms, or the actual dosing amount deviates from the commanded value beyond the allowable range; The confidence level of the output of the temporal dynamic neural network model is lower than the threshold; The security mechanism includes: Immediately interrupt the control signal driving the dosing pump, automatically switch to the pre-configured PID control mode, trigger the audible and visual alarm in the control room, and record the abnormal event in the log.
10. A coagulant adaptive dosing control system based on a physical constraint neural network, characterized in that, include: The sensing layer is used to collect water quality data; The data and cognition layer is used for water quality data preprocessing to obtain standardized water quality sequences; Standardized water quality sequences are input into a pre-trained time-series dynamic neural network model to predict future dosage and future effluent turbidity. The optimization and decision layer is used at the end of each control cycle to perform rolling optimization on the pre-trained time-series dynamic neural network model based on the model predictive control mechanism, with the optimization objectives being that the future effluent turbidity predicted by the time-series dynamic neural network model does not exceed the standard and the future dosage is minimized, in order to solve for the optimal chemical dosing sequence. The execution and safety layer is used to convert the optimal drug dosing sequence into control signals to drive the dosing pump to administer the drug.