CN-121979154-A - Coagulant self-adaptive addition control method and system based on physical constraint neural network

CN121979154ACN 121979154 ACN121979154 ACN 121979154ACN-121979154-A

Abstract

The invention discloses a coagulant self-adaptive addition control method and system based on a physical constraint neural network, and relates to the technical field of automatic control of water treatment; the method comprises the steps of preprocessing water quality data to obtain a standardized water quality sequence, inputting the standardized water quality sequence into a pre-trained time sequence dynamic neural network model to predict future dosing amount and future dosing amount, when each control period is finished, based on a model prediction control mechanism, executing rolling optimization on the pre-trained time sequence dynamic neural network model to solve an optimal dosing sequence by taking the future dosing amount predicted by the time sequence dynamic neural network model as an optimization target, and converting the optimal dosing sequence into a control signal to drive a dosing pump to dose. The problems of weak self-adaptation capability, poor predictability and high medicine consumption of the traditional control method when facing complex and changeable raw water quality are solved, and stable standard reaching of the water quality and accurate optimization of coagulant addition are realized.

Inventors

GAO YANG
ZHAO ZEXU
MENG LIN
JIAO LIANGBAO
Su Yancheng
LIU BOWEN

Assignees

南京工程学院

Dates

Publication Date: 20260505
Application Date: 20260206

Claims (10)

1. The coagulant self-adaptive addition control method based on the physical constraint neural network is characterized by comprising the following steps of: Collecting water quality data; Pretreatment of water quality data to obtain a standardized water quality sequence; Inputting the standardized water quality sequence into a pre-trained time sequence dynamic neural network model, and predicting future dosage and future water turbidity; when each control period is finished, based on a model prediction control mechanism, rolling optimization is carried out on the pre-trained time sequence dynamic neural network model by taking the future water turbidity predicted by the time sequence dynamic neural network model as an optimization target, wherein the future water turbidity is not out of standard and the future dosing amount is minimum, and an optimal agent dosing sequence is solved; And converting the optimal medicament adding sequence into a control signal to drive a medicament adding pump to administrate medicaments.
2. The coagulant self-adaptive addition control method based on the physical constraint neural network according to claim 1, wherein the water quality data comprises water inflow flow, water inflow turbidity, water inflow pH value, water temperature and sedimentation tank water outflow turbidity; the pretreatment of the water quality data specifically comprises the following steps: The pH value and the water temperature of the inlet water are normalized by adopting a Z-score, the turbidity of the inlet water, the flow rate of the inlet water and the turbidity of the outlet water of a sedimentation tank are normalized by adopting a Min-Max, and a standardized water quality sequence with N control periods in the past is obtained In the formula, Is the past first Standardized water quality data for each control period, Is the standardized water quality data of the current control period t.
3. The coagulant self-adaptive addition control method based on the physical constraint neural network as set forth in claim 1, wherein the loss function of the time sequence dynamic neural network model in the pre-training stage comprises a mean square error prediction error term and a physical constraint penalty term, and the expression is as follows: In the formula, Is the function of the total loss and, Is the mean square error of the signal, In order to constrain the weight coefficients of the weights, The method is a physical constraint penalty term, wherein the physical constraint penalty term is the sum of one or more of non-negative constraint, monotonicity constraint and boundary constraint, and the priority is boundary constraint > monotonicity constraint > non-negative constraint.
4. The coagulant adaptive dosing control method based on a physical constraint neural network according to claim 3, wherein the non-negative constraint is used for applying a penalty to a predicted negative dosing value output by a time-lapse dynamic neural network model, and the expression is as follows: In the formula, Representing a non-negative constraint, and, Is the predicted dosage.
5. The coagulant adaptive addition control method based on the physical constraint neural network as set forth in claim 3, wherein the expression of the monotonicity constraint is as follows: In the formula, Represents monotonicity constraint, k is the serial number of other characteristics except the turbidity of the inlet water in the water quality data, And Representing the predicted dosage of the input water quality data samples for the samples with relatively low turbidity of the inlet water and the predicted dosage of the samples with relatively high turbidity of the inlet water respectively.
6. The coagulant adaptive addition control method based on the physical constraint neural network as set forth in claim 3, wherein the boundary constraint is used for applying a penalty to the predicted addition amount beyond the process allowable range, and the expression is as follows: In the formula, Is a boundary constraint that is set by the user, Is the predicted administration amount of the drug, Is the lower limit of the administration amount, Is the upper limit of the dosage.
7. The coagulant self-adaptive addition control method based on the physical constraint neural network according to claim 1, wherein the future turbidity of the water output predicted by the time sequence dynamic neural network model is not out of standard and the future dosage is minimum as an optimization target specifically comprises: The objective function of the optimization problem is specifically as follows: In the formula, The drug addition sequence from the (t+1) th control period to the (t+H) th control period is represented, H is the prediction step length, Is the water turbidity of the t+k control period predicted by the pre-trained time sequence dynamic neural network model, For the target turbidity to be a target value, In order to control the weight of the weight, The dosage of the t+k control period is shown; Constraint conditions for the optimization problem are as follows: And (5) constraint of the addition amount range: In the formula, Is the lower limit of the administration amount, Is the upper limit of the dosage; and (3) water quality constraint of effluent: In the formula, Is the upper turbidity limit.
8. The coagulant self-adaptive dosing control method based on the physical constraint neural network according to claim 1, wherein the step of converting the optimal drug dosing sequence into a control signal to drive a dosing pump to perform dosing is specifically as follows: Multiplying the recommended dosage of each control period in the optimal medicament dosage sequence by the water inflow acquired in real time to obtain an instantaneous dosage instruction The formula is as follows: In the formula, Is an instruction of the addition amount, Is the water inflow collected in real time, The dosage of the model in the t+1th control period is recommended; The dosing command is converted into an analog current signal to drive the dosing pump to administer the drug.
9. The coagulant adaptive addition control method based on the physical constraint neural network according to claim 1, wherein the method further comprises: triggering any one of the following conditions, then the security mechanism is started: The turbidity of the effluent continuously exceeds a set value for a preset time; The fault alarm of the dosing equipment is carried out, or the deviation between the actual dosing quantity and the instruction value exceeds the allowable range; the confidence level of the output of the time sequence dynamic neural network model is lower than a threshold value; the security mechanism includes: immediately interrupting a control signal for driving the dosing pump, automatically switching to a preconfigured PID control mode, triggering an audible and visual alarm of a control room, and recording an abnormal event in a log.
10. Coagulant self-adaptation adds control system based on physical constraint neural network, characterized by comprising: the sensing layer is used for collecting water quality data; The data and cognition layer is used for preprocessing the water quality data to obtain a standardized water quality sequence, inputting the standardized water quality sequence into a pre-trained time sequence dynamic neural network model, and predicting future dosage and future effluent turbidity; The optimization and decision layer is used for executing rolling optimization on the pre-trained time sequence dynamic neural network model based on a model prediction control mechanism and taking the future water turbidity predicted by the time sequence dynamic neural network model as an optimization target, wherein the future water turbidity is not out of standard and the future dosage is minimum, and solving the optimal medicament adding sequence; And the execution and safety layer is used for converting the optimal medicament administration sequence into a control signal to drive the medicament adding pump to administer.

Description

Coagulant self-adaptive addition control method and system based on physical constraint neural network Technical Field The invention relates to the technical field of automatic control of water treatment, in particular to a coagulant self-adaptive addition control method and system based on a physical constraint neural network. Background The water treatment system of the thermal power plant is a life line which runs safely, efficiently and economically, and raw water treatment (or makeup water treatment) is the starting point and basic guarantee of the whole water treatment process. Raw water is the primary water source (surface water, ground water or reclaimed water) of a water treatment system of a thermal power plant, and untreated raw water contains a large amount of impurities in a wide variety, so that the untreated raw water cannot be directly used for key thermal equipment of the power plant. The primary function of raw water treatment is to remove these impurities and produce high-quality and usable 'make-up water' for the whole power plant. In the traditional water treatment coagulation process, coagulant addition control mainly depends on an empirical model based on fixed parameters, proportional-integral-derivative (PID) control or a feed-forward-feedback composite control strategy. When the method faces complex fluctuation of raw water quality (such as turbidity, temperature, pH, pollutant components and the like), the method has obvious limitations of poor self-adaptive capacity, difficulty in adapting to water quality mutation or seasonal change, insufficient predictability, delayed control response, low dosing precision, easiness in excessive or insufficient dosing, unstable water quality of the yielding water, increased medicament consumption (medicament consumption) and increased running cost. In addition, in the emerging technology, although the data driving model has a certain learning capability, the data driving model lacks explicit constraint of a physical and chemical mechanism, and the output under extreme or unseen working conditions may not accord with an actual rule, while the visual detection scheme based on deep learning has the problems that the visual detection result is added as a regulating parameter on the basis of the traditional regulation and control based on the water quality parameter, the visual detection is limited by an application scene, such as that monitoring equipment is extremely easy to corrode raw water, the raw water light transmittance is poor, and the like. The project reliability and the robustness of the scheme are insufficient, and the scheme is insufficient to meet the requirements of water treatment scenes. In the prior art, a 'divide-and-conquer' strategy for dividing working conditions and calling different models based on fixed rules exists, as shown in fig. 1, but the model to be trained in the method is large in scale and low in efficiency. Therefore, an intelligent adding control method capable of deeply fusing a physical mechanism of water quality change with real-time monitoring data and having prospective prediction and strong self-adaptive capacity is needed to realize accurate optimal control of the coagulation process. Disclosure of Invention Aiming at the defects in the prior art, the invention provides a coagulant self-adaptive addition control method and system based on a physical constraint neural network. The constraint mechanisms which are strongly coupled with the physical and chemical laws of the coagulation process are respectively embedded in the data preprocessing layer, the neural network training layer, the rolling optimization layer and the safety pocket bottom layer, so that all links generated from data acquisition to control instructions meet the process safety requirements, the problems of weak self-adaptation capability, poor predictability and high drug consumption when facing complex and changeable raw water quality in the traditional control method are solved, and the stable standard reaching of the water quality and the accurate optimization of coagulant addition are realized. In order to achieve the above purpose, the present invention adopts the following technical scheme: a coagulant self-adaptive addition control method based on a physical constraint neural network comprises the following steps: Collecting water quality data; Pretreatment of water quality data to obtain a standardized water quality sequence; Inputting the standardized water quality sequence into a pre-trained time sequence dynamic neural network model, and predicting future dosage and future water turbidity; when each control period is finished, based on a model prediction control mechanism, rolling optimization is carried out on the pre-trained time sequence dynamic neural network model by taking the future water turbidity predicted by the time sequence dynamic neural network model as an optimization target, wherein the future water turbidity