A circulating water treatment closed-loop optimization method and system based on water quality prediction

Abstract

The application relates to the technical field of circulating water treatment, and discloses a circulating water treatment closed-loop optimization method and system based on water quality prediction, wherein the circulating water treatment closed-loop optimization method based on water quality prediction comprises the following steps: constructing a time sequence knowledge graph supporting a time dimension, which is used for representing time sequence evolution characteristics of water quality parameters and relationships thereof; constructing a multi-time scale prediction engine, which is used for predicting future water quality states; developing a prediction type constraint generator, which is used for converting the future water quality prediction into current decision constraints; introducing a time and space sensitive multi-objective optimization algorithm, which is used for optimizing a reagent proportioning strategy; realizing a multi-time scale collaborative optimization system, which is used for further balancing short-term response and long-term stability; and constructing an environment adaptive active evolution system, which is used for realizing automatic adjustment of the system with environment changes; the application realizes fundamental transformation of a circulating water treatment control mode, and provides a brand-new closed-loop optimization paradigm for a circulating water treatment system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of circulating water treatment, more particularly, it relates to a circulating water treatment closed-loop optimization method and system based on water quality prediction. BACKGROUND

[0002] The circulating water treatment system is widely used in industrial cooling water, central air conditioning cooling tower, chemical production process and other fields, and its core goal is to ensure the efficient and stable operation of the system by monitoring and regulating water quality parameters. The existing circulating water treatment system usually includes water quality monitoring equipment, reagent adding equipment, filtering equipment and circulating pump and other infrastructure, and cooperates with certain automatic control means to realize water quality management.

[0003] However, the existing technology still has many deficiencies in the automation and intelligentization of circulating water treatment, mainly in the following aspects:

[0004] The current mainstream circulating water treatment control system mostly adopts a feedback-based reactive control strategy, that is, adjustment is only performed after detecting that the water quality parameter exceeds the set threshold. This control method has obvious hysteresis, and the effect of control measures such as reagent addition often needs a certain time to appear, which makes it difficult for the system to respond to rapid changes in water quality in a timely manner, especially in the case of sudden water quality fluctuations, the system response is not timely enough, affecting the water quality stability; the water quality prediction model and the control system in the existing technology are mostly independent of each other, and the prediction results are difficult to directly used for control decision-making, due to the lack of efficient data interface and collaborative mechanism between the prediction system and the control system, the prediction information cannot fully guide the control operation such as reagent addition, limiting the application of the prediction-driven feedforward control strategy; the existing circulating water treatment system has limitations in the control demand of multiple time scales, some systems pay too much attention to short-term water quality optimization, ignoring the stability of long-term operation; while other systems emphasize long-term stability, sacrificing the rapid response capability to short-term water quality fluctuations, how to balance the water quality control effect in different time scales such as hours, days and weeks is a problem that needs to be solved by the current technology; there are complex dynamic correlations between water quality parameters (such as pH value, conductivity, turbidity, residual chlorine, etc.) in the circulating water system, these correlations change with time and environmental conditions, the existing technology mostly uses static models or simple time series analysis methods, which is difficult to fully capture and model the time evolution rule of water quality parameters, resulting in limited understanding and prediction ability of the dynamic behavior of the system; the existing circulating water treatment system generally lacks self-adaptation capability to external environmental changes (such as seasonal change, raw water quality fluctuation, etc.), the system parameters usually need to be adjusted regularly by artificial, which is difficult to realize the self-adaptation and continuous optimization of environmental disturbance, affecting the intelligent level and operation efficiency of the system.

[0005] In summary, the existing circulating water treatment technology has obvious deficiencies in response timeliness, prediction and control cooperation, multi-time scale balance, complex time sequence relationship modeling and environmental self-adaptive ability, and a new technical solution is needed to improve the intelligentization and optimization level of the circulating water treatment system. SUMMARY

[0006] The application provides a circulating water treatment closed-loop optimization method and system based on water quality prediction, which solves the technical problems of response lag, prediction and control separation, multi-time scale balance difficulty and poor environmental adaptability in related technologies.

[0007] The application provides a circulating water treatment closed-loop optimization method based on water quality prediction, which comprises:

[0008] A time sequence knowledge graph supporting time dimension is constructed to represent the time sequence evolution characteristics of water quality parameters and their relationships;

[0009] Based on the time sequence knowledge graph, a multi-time scale prediction engine is constructed to predict future water quality state;

[0010] A prediction-based constraint generator based on the prediction results of future water quality state is developed to convert future water quality prediction into current decision constraints;

[0011] A spatiotemporal sensitive multi-objective optimization algorithm is introduced to optimize the reagent ratio strategy according to the prediction results and current decision constraints;

[0012] Based on the optimized reagent ratio strategy, a multi-time scale collaborative optimization system is realized to further balance short-term response and long-term stability;

[0013] Based on the optimization results of the multi-time scale collaborative optimization system, an environmental adaptive active evolution system is constructed to realize automatic adjustment of the system with environmental changes.

[0014] Further, the step of constructing the time sequence knowledge graph supporting the time dimension comprises:

[0015] The water quality parameters are defined as entity nodes in the graph, and each node contains parameter name, value and timestamp attributes;

[0016] A relationship discovery algorithm based on Granger causality test is applied to identify the potential causal relationship between parameters, and the potential causal relationship is represented as a directed edge in the graph;

[0017] A graph neural network-based time sequence pattern learning method is used to extract the evolution law of water quality parameters over time from historical data.

[0018] Further, the multi-time scale prediction engine is constructed based on a hybrid architecture of graph convolution network and gated recurrent unit, comprising:

[0019] a spatial feature extraction module for capturing spatial dependency among nodes;

[0020] a temporal feature extraction module for capturing time series characteristics;

[0021] an attention fusion layer for fusing spatial features and temporal features;

[0022] a multi-time scale prediction head including three parallel branches of short-term prediction, medium-term prediction and long-term prediction.

[0023] Further, the multi-time scale prediction engine also introduces Monte Carlo dropout technology to provide a confidence interval for each prediction result, obtains a prediction distribution through random dropout forward propagation, calculates the mean as the final prediction value and the variance as the uncertainty measure.

[0024] Further, the prediction formula constraint generator includes:

[0025] a key threshold and target interval definition module for defining a target range and an alarm threshold for each water quality parameter;

[0026] a prediction trajectory risk assessment module for risk analysis of the predicted water quality parameter trajectory, considering parameter importance, deviation degree and out-of-bound probability comprehensively;

[0027] a dynamic constraint generation module for generating constraint conditions for control decisions based on the risk assessment results.

[0028] Further, the spatio-temporal sensitive multi-objective optimization algorithm includes:

[0029] a drug ratio model is constructed to describe the response model of the relationship between the amount of drug added and the water quality parameters;

[0030] a target function is constructed for three time scales of short-term, medium-term and long-term, respectively, focusing on response speed, stability and cost-effectiveness;

[0031] An improved particle swarm optimization algorithm is used, which uses continuous-discrete hybrid coding, adaptive penalty function and multi-population cooperative evolution strategy to solve multi-objective optimization problems.

[0032] Further, the multi-time scale collaborative optimization system includes:

[0033] a three-layer control architecture of emergency response layer, daily control layer and strategy optimization layer is established;

[0034] a layer information transmission and decision coordination system is constructed to ensure that control decisions of different time scales are compatible and consistent as a whole;

[0035] According to the system state and the prediction result, the activation priority of different time scale control layers is dynamically adjusted.

[0036] Further, the environment adaptive active evolution system comprises:

[0037] A model performance monitoring and evaluation module continuously monitors the performance of the prediction model, calculates the prediction error index by comparing the predicted value with the actual observation value, and evaluates the performance of the prediction model.

[0038] An incremental learning and model updating module updates the time series knowledge graph and the prediction model by using the incremental learning method.

[0039] A seasonal pattern adaptive module identifies and learns the seasonal variation pattern of the water quality parameter.

[0040] A control effect feedback and optimization module records the control operation and its effect, and optimizes the control strategy by using the reinforcement learning method.

[0041] Further, the incremental learning and model updating module comprises three key components, namely, a selective memory module, an elastic weight adjustment algorithm and a knowledge distillation framework, which are used to evaluate the value of new data, protect key knowledge and balance new and old knowledge.

[0042] The application provides a circulating water treatment closed-loop optimization system based on water quality prediction, which is used for executing the above-mentioned circulating water treatment closed-loop optimization method based on water quality prediction.

[0043] A time series knowledge graph module is used to represent the time series evolution characteristics of the water quality parameters and their relationships.

[0044] A multi-time scale prediction engine is used to predict the future water quality state based on the time series knowledge graph.

[0045] A prediction constraint generator is used to convert the future water quality prediction into current decision constraints.

[0046] A multi-objective optimization module is used to optimize the reagent proportioning strategy according to the prediction result and the constraint condition.

[0047] A multi-time scale collaborative control module is used to balance the relationship between short-term response and long-term stability.

[0048] An environment adaptive active evolution module is used to realize automatic adjustment of the system with the change of the environment.

[0049] The application has the beneficial effects that the time series knowledge graph is used to capture the dynamic correlation and time series evolution law between the water quality parameters, high-precision prediction of the water quality change is realized, the prediction accuracy is improved compared with the traditional method, and the performance is excellent in predicting water quality mutation events.

[0050] Early intervention based on multi-time scale prediction results, transforming traditional passive response control into active prevention control, enabling the system to deal with potential water quality problems in advance and reducing water quality fluctuation risk;

[0051] Through multi-objective optimization and predictive constraint generation, the drug ratio strategy is optimized under the premise of ensuring water quality stability, reducing drug usage, and reducing energy consumption;

[0052] The environmental adaptive active evolution system enables the application to autonomously adapt to external factors such as seasonal changes and raw water quality fluctuations, without the need for manual intervention to maintain optimal performance;

[0053] The multi-time scale collaborative optimization system enables the application to quickly respond to short-term disturbances and maintain long-term stability, improving water quality stability compared to traditional systems;

[0054] Through a unified time sequence driving framework, the traditional separated prediction system and control system are integrated into a single architecture, reducing interface conversion complexity, reducing system parameter quantity, and improving running efficiency;

[0055] The application realizes the fundamental transformation of the circulating water treatment control mode, upgrading from the traditional "measurement, judgment, reaction" mode to the "prediction, prevention, optimization" mode, providing a new closed-loop optimization paradigm for circulating water treatment systems. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is a flowchart of a circulating water treatment closed-loop optimization method based on water quality prediction in the application;

[0057] Figure 2 is a broken line graph of the prediction trend of different water quality parameters within 24 hours in the application;

[0058] Figure 3 is a scatter plot of the Pareto optimal solution set obtained by the multi-objective optimization algorithm in the application;

[0059] Figure 4 is a radar chart of the performance difference between the application method and the traditional control method in five dimensions;

[0060] Figure 5 is an area chart of the accuracy rate change trend of the incremental learning model and the traditional fixed model during long-term use;

[0061] Figure 6 is a two-axis graph of the relationship between water temperature and microbial activity index in a circulating water system with seasonal changes;

[0062] Figure 7 is a column chart of the performance difference between the traditional water quality control method and the application method in three key indicators. DETAILED DESCRIPTION

[0063] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is merely meant to provide a better understanding of the subject matter described herein and can include changes, modifications, additions or omissions of the functions and arrangements of the elements discussed without departing from the scope of the present disclosure. Various examples can omit, substitute, or add various processes or components in addition to those described or in lieu thereof. Also, some examples can be described and / or claimed as including or excluding portions of features in addition to those described or in lieu thereof.

[0064] In at least one embodiment of the present disclosure, a closed-loop optimization method for circulating water treatment based on water quality prediction is disclosed, as shown in Figure 1 The method comprises the following steps:

[0065] Step 1, construct a time dimension supported time series knowledge graph, which is used to represent the time series evolution characteristics of water quality parameters and their relationships;

[0066] This step first collects historical water quality parameter data of the circulating water system, including pH value, conductivity, turbidity, residual chlorine, hardness, microbial count and other parameters, as well as environmental data such as temperature, humidity, seasonal information, etc. The collected data is preprocessed, including outlier detection, missing value filling and data standardization. Then the time series knowledge graph is constructed through the following sub-steps. The time series knowledge graph constructs pH value, conductivity, turbidity, residual chlorine, hardness, microbial count and other key water quality parameters as graph nodes, discovers the dynamic correlation between parameters (such as the change of microbial activity caused by the change of pH value, the relationship between conductivity and fouling risk, etc.) through Granger causality test, and records the evolution characteristics of these relationships over time:

[0067] Water quality parameters (such as pH value, conductivity, turbidity, residual chlorine, etc.) have different dimensions and numerical ranges, and need to be standardized to map them to the same numerical interval (such as 0-1 or -1 to 1) for fair comparison in subsequent graph construction and model training. For pH value, which has a fixed range, minimum-maximum normalization can be used. For conductivity, turbidity and other parameters that theoretically have no upper limit, Z-score standardization or log transformation and then normalization should be used. Seasonal information and other classification data need to be converted into numerical features through one-hot encoding.

[0068] Step 1.1, entity definition and attribute extraction;

[0069] Water quality parameters are defined as entity nodes in the graph, each node containing parameter name, value, timestamp and other attributes. A sliding time window method is used to calculate statistical features (such as mean, variance, rate of change, etc.) for each parameter as the time series feature attributes of the node.

[0070] Step 1.2, relationship discovery and graph construction;

[0071] The relationship discovery algorithm based on Granger causality test is applied to identify the potential causal relationships between parameters and represent them as directed edges in the graph. The relationship strength is quantified by the time-lagged correlation coefficient, and the relationship type is determined according to the direction of correlation (positive or negative).

[0072] Before calculating the time-lagged correlation coefficient between different water quality parameters, the parameters need to be standardized to eliminate the influence of dimension and numerical range differences. For example, pH values are usually in the range of 1-14, while conductivity may be several hundred to several thousand microsiemens per centimeter. Direct calculation of correlation will result in the dominance of parameters with large values. Z-score standardization or min-max normalization should be performed on all parameter sequences to ensure the fairness of correlation calculation.

[0073] The mathematical expression of the time series relationship is as follows:

[0074] ;

[0075] where, represents the parameter at time , the influence relationship strength of parameter at time , is the time-lagged correlation function, and represent the time series data of parameter and parameter , respectively.

[0076] The time-lagged correlation function is specifically implemented by calculating the cross-correlation coefficient of two time series at a given time lag, and combining the p-value of Granger causality test to form a comprehensive score. When the score exceeds the preset threshold, it is confirmed that there is a causal relationship and a directed edge connection is established.

[0077] Step 1.3, time series evolution pattern extraction;

[0078] The time series pattern learning method based on graph neural network is adopted to extract the evolution rule of water quality parameters over time from historical data. The graph state transition matrix is constructed to capture the evolution patterns of the system under different conditions. The dynamic evolution process of the time series graph can be expressed as: for any time The state of the knowledge graph is described by its rate of change over time, which captures the dynamic behavior of the system. This rate of change depends on three key factors: the current graph state (reflecting the current values of water quality parameters and their relationship structure), the control vector (including the addition of various reagents and adjustment operations), and the environmental factor vector (such as temperature, humidity, and other external conditions). These three factors are integrated through a state transition function to obtain the rate of change of the graph state, i.e., the instantaneous change trend of each node attribute and relationship strength. The state transition function is implemented through a hybrid model combining graph neural networks and differential equations: first, the graph neural network processes the current graph state to calculate the basic change trend of each node under the current relationship structure; then, the change trends are adjusted according to the control vector to reflect the impact of control operations on the system state; finally, the modulation effect of environmental factors on the change rate is considered to generate the final state derivative value. Through this mechanism, the system can simulate and predict the continuous evolution of water quality parameters over time, providing a theoretical basis for prediction and control.

[0079] The time-series knowledge graph in this application has the following specific structure:

[0080] Multi-level node system: The graph nodes are divided into three categories: parameter nodes, aggregation nodes, and environmental nodes. Parameter nodes represent single water quality parameters; aggregation nodes represent parameter combinations, capturing the synergistic effects of multiple parameters; environmental nodes represent external environmental factors.

[0081] Time-sensitive edges: The edges in the graph contain time delay attributes, representing the time required for the influence to propagate from the source node to the target node. At the same time, the edge weights can change over time, representing the dynamic characteristics of relationship strength.

[0082] Context sub-graphs: Special context sub-graphs are constructed for different operating conditions (such as high load, low temperature, etc.), each capturing the parameter relationships under specific conditions.

[0083] Memory structure: The graph contains both long-term steady-state relationships and short-term dynamic relationships, with long-term relationships representing stable causal relationships and short-term relationships capturing temporary strong correlations.

[0084] The time-series knowledge graph maintains the latest state through a dynamic update module: when new observation data is received, the node attribute values are first updated, then the edge relationship strength is calculated based on the new data, and if necessary, new edges are added or weakly related edges are removed, and finally the graph state transition matrix is updated.

[0085] Step 2, based on the time-series knowledge graph, build a multi-time scale prediction engine to predict future water quality states;

[0086] This step uses the time series knowledge graph constructed in step 1 to construct a multi-time scale prediction engine that can provide short-term, medium-term and long-term predictions simultaneously. The prediction engine provides differentiated predictions for different water quality parameter characteristics, such as hourly-level accurate predictions for rapidly changing parameters like pH, day-level trend predictions for slowly changing parameters like hardness, and week-level change trends for seasonal parameters like microbial indicators. It also considers the interaction between parameters, such as the accelerating effect of temperature rise on residual chlorine decay rate, to accurately predict the combined effect. Specifically, it includes:

[0087] Step 2.1, constructing a graph neural network prediction model;

[0088] According to the method provided in the present application, a time series graph prediction model is constructed based on a hybrid architecture of graph convolution network (GCN, Graph Convolutional Network) and gated recurrent unit (GRU, Gated Recurrent Unit). The model takes graph node features and topology as input, extracts spatial dependencies through multi-layer graph convolution, and captures time series characteristics through recurrent network.

[0089] Specifically, the graph neural network prediction model of the present application includes the following key components:

[0090] Spatial feature extraction module: composed of three layers of graph convolution network, each layer containing 32, 64, and 128 convolution kernels, used to capture the spatial dependency between nodes. The graph convolution operation is implemented as follows: first, obtain the adjacency matrix of the graph and add self-loop connection to form the augmented adjacency matrix; then calculate the degree matrix of the nodes, i.e. the number of connections of each node; then normalize the augmented adjacency matrix to convert it to a symmetric normalized form; then multiply the node feature matrix of the current layer with the normalized adjacency matrix, and then multiply it with the learnable weight matrix; finally, apply the ReLU activation function to the obtained feature representation to generate the node feature representation of the next layer. This process realizes the aggregation and transformation of node features, enabling each node to integrate the information of its neighbor nodes.

[0091] Time series feature extraction module: composed of bidirectional GRU with hidden layer dimension of 256, used to capture time series characteristics. The historical feature sequence of each node is processed by GRU to form a time series feature representation.

[0092] Attention fusion layer: adopts a multi-head self-attention structure (8 attention heads) to fuse spatial features and time series features and calculate the importance weights between different features.

[0093] Multi-time scale prediction head: contains three parallel branches, corresponding to short-term, medium-term and long-term prediction respectively. Each branch contains two fully connected layers (dimensions are 128 and 64 respectively) and an output layer with a dimension equal to the number of predicted parameters.

[0094] It should be noted that the model training adopts a self-supervised learning method, constructs training samples through historical data, uses mean square error as the loss function, and adds an L2 regularization term to prevent overfitting. The model is trained using the Adam optimizer with a learning rate of 0.001 and a batch size of 32.

[0095] Step 2.2, multi-time scale prediction framework construction;

[0096] A hierarchical prediction strategy is adopted, and a special prediction head is constructed for different prediction time spans. Short-term prediction (1-6 hours) focuses on detail accuracy, medium-term prediction (1-3 days) balances accuracy and trend, and long-term prediction (1-2 weeks) mainly captures macro trends. Each prediction head pays attention to different time scale features through an attention structure.

[0097] Step 2.3, prediction uncertainty quantification;

[0098] The Monte Carlo dropout technique is introduced to provide a confidence interval for each prediction result. By performing multiple random dropout forward propagations, a prediction distribution is obtained, and the mean is taken as the final prediction value and the variance is taken as the uncertainty measure. Prediction confidence is used for risk assessment in subsequent control decisions.

[0099] Optionally, in some embodiments, an ensemble learning method can be used, in which multiple models with different structures or parameters are trained, and their prediction results are considered comprehensively to further improve prediction accuracy and robustness.

[0100] As shown in Figure 2 , the prediction trends of different water quality parameters (pH, conductivity, turbidity) within 24 hours are shown. The solid line in the figure represents the predicted value, and the dashed line area represents the confidence interval of the prediction, intuitively presenting the short-term prediction capability of the multi-time scale prediction engine. As can be seen from the figure, the system can accurately capture the trend of water quality parameters and quantify the uncertainty of the prediction through the confidence interval, providing a reliable basis for subsequent control decisions.

[0101] Step 3, develop a prediction-based constraint generator based on the prediction results of future water quality states, convert future water quality predictions into current decision constraints;

[0102] This step, based on the prediction results of step 2, generates constraints to guide current control decisions. The constraint generator prioritizes constraints according to the importance of water quality parameters, achieving prediction-based feedforward control. This prediction-driven constraint generation mechanism transforms circulating water treatment from a passive response to an active prevention approach. The system no longer waits until water quality parameters actually exceed standards before taking action, but intervenes precisely before problems occur. Specifically, this includes:

[0103] Step 3.1, defining key thresholds and target intervals;

[0104] Target ranges and warning thresholds are defined for each water quality parameter, including the ideal operating range, acceptable fluctuation range, and absolute limit. Weighting coefficients are assigned to different parameters based on their importance to construct a parameter priority matrix.

[0105] Step 3.2, Risk assessment of predicted trajectory;

[0106] Risk analysis is performed on the predicted water quality parameter trajectories to identify potential boundary violations and unstable trends. The degree of risk is quantified by the deviation between the predicted parameter values ​​and the target range, as well as the probability of boundary violations. The risk assessment calculation method is as follows: first, the importance weight of each water quality parameter is determined. The system first assesses the parameter's impact on system safety and performance. Then, it calculates the normalized Euclidean distance between the predicted parameter value and the target value; a larger distance indicates a greater deviation from the target. Next, it calculates the probability that the parameter value will exceed the allowable range. This probability is based on the uncertainty distribution of the prediction and is obtained by calculating the probability density of the predicted distribution falling outside the allowable range. Finally, it multiplies the parameter weight, distance metric, and probability of exceeding the limit to obtain a comprehensive risk assessment value. This risk assessment value comprehensively considers the parameter's importance, degree of deviation, and likelihood of exceeding the limit, providing a basis for subsequent constraint generation. For high-risk parameters, the system generates stricter control constraints; for low-risk parameters, it provides greater control flexibility.

[0107] The target values ​​and allowable fluctuation ranges for different water quality parameters vary greatly. Directly calculating the Euclidean distance will lead to parameters with larger fluctuation ranges dominating the risk assessment results. The deviation of each parameter should be divided by its allowable fluctuation range to obtain the relative deviation, and then the normalized Euclidean distance should be calculated. For calculating the probability of exceeding the range, the uncertainty distribution needs to be standardized to a unified probability space to ensure that the risks of exceeding the limits for different parameters are comparable.

[0108] Step 3.3, Dynamic Constraint Generation;

[0109] Based on the risk assessment results, constraints for the current control decision are generated. For high-risk predictions, mandatory constraints are generated; for medium-risk predictions, soft constraints are generated; and for low-risk predictions, optimization objectives are generated. These constraints are expressed as upper and lower limits or target intervals for control variables (such as the amount of reagent added).

[0110] It should be understood that in some embodiments, the constraint generator can also consider factors such as operating cost and equipment life of the system to generate multi-dimensional constraint conditions, achieving more comprehensive optimization control.

[0111] Step 4, introduce a spatiotemporal sensitive multi-objective optimization algorithm to optimize the dosage ratio strategy according to the prediction results and current decision constraints;

[0112] This step is based on the constraints generated in step 3, and uses a multi-objective optimization algorithm to optimize the efficiency of reagent use and energy consumption while meeting water quality parameter targets. Specifically, it includes:

[0113] Step 4.1, reagent dosage ratio model construction;

[0114] A response model of reagent dosage and water quality parameters is constructed, considering the interaction and hysteresis effect between reagents. This model is trained based on historical control data and can predict the impact of different reagent dosage schemes on water quality parameters. The method for determining the optimal control strategy is as follows: first, define the control variable set , including the combination of reagent dosage and addition time; then define the cost function, which considers multiple target indicators under the current graph state and the predicted future graph state; then calculate the cost value of each possible control strategy; finally, select the control strategy with the minimum cost value as the optimal scheme. The response model of reagent dosage and water quality parameters uses a hybrid modeling method combining multi-level recurrent neural networks and physical and chemical constraint rules; the model inputs include the current water quality state vector (pH value, conductivity, turbidity, etc.), 24-hour historical water quality trajectory, reagent operation vector (reagent dosage and time point), environmental condition vector (temperature, humidity, system load), and system state identifier; the output is the future 1-24 hour water quality parameter evolution trajectory, reagent effect index, parameter fluctuation risk assessment, and reagent residual concentration prediction. The interaction between reagents is modeled through a special interaction feature layer, which captures the synergistic or antagonistic relationship between reagents, such as the neutralization effect of oxidizing bactericides and reducing cleaning agents, the influence of scale inhibitors on pH adjuster, etc., and automatically extracts multi-order interaction features through a nonlinear mapping function. The hysteresis effect uses a time delay embedding method to distinguish between fast response relationships (minute level), medium response relationships (hour level), and slow response relationships (day level), and dynamically calculates the residual influence of historical operations through memory units. The online learning mechanism is based on the incremental gradient descent method, including sliding window sample management, differentiated learning rate scheduling, abnormal sample processing, model structure adaptation, and parameter importance evaluation.

[0115] The specific implementation of the cost function is as follows: first, normalize each evaluation index (including the cost of reagents, energy consumption, water quality standard degree, and system stability) to make the dimensions consistent and the numerical range uniform; then assign weight coefficients to each index, which are dynamically adjusted through system state monitoring and historical data analysis; then calculate the weighted sum of each index, where the reagent cost part considers the unit price and usage of each type of reagent, the energy consumption part integrates pump power, heating / cooling energy consumption, etc., the water quality standard degree part evaluates the degree of agreement between each water quality parameter and the target value, and the system stability part calculates the fluctuation amplitude of the key parameters; finally, apply a smoothing function to handle boundary conditions to ensure a reasonable optimization direction under extreme conditions.

[0116] The reagent ratio model in this application is specifically implemented as a three-layer structure:

[0117] Reagent response sub-model: a separate response model is established for each reagent to describe its influence on each water quality parameter. These sub-models are based on a combination of regression models trained based on physical and chemical principles and historical data, which can handle nonlinear response relationships.

[0118] Reagent interaction layer: models the synergistic and antagonistic effects between different reagents, using second-order interaction terms to capture the mutual influence between reagents.

[0119] Time response layer: simulates the time-varying curve of reagent effect through an exponential decay function, accurately reflecting the dynamic response process of water quality parameters after reagent addition.

[0120] The reagent ratio model training uses a two-stage method: first, initialize the basic response relationship using historical data, and then continuously optimize the model parameters through online learning to adapt to changes in system characteristics.

[0121] Step 4.2, multi-time scale objective function construction;

[0122] Construct the objective function for three time scales: short-term (response speed), medium-term (stability), and long-term (cost-effectiveness), focusing on different optimization priorities. Use the weighted sum method to integrate the three objectives into a comprehensive evaluation function, and the weight coefficients can be dynamically adjusted according to the system state.

[0123] The short-term objective function focuses on quickly correcting deviations, and its calculation method is as follows: first, calculate the absolute deviation between the predicted value and the target value of each water quality parameter at the short-term prediction time point; then multiply the deviation of each parameter by the corresponding parameter importance coefficient , which represents the weight of different parameters on system performance; finally, sum all the weighted deviations to get the overall short-term objective function value. The smaller the value, the higher the short-term water quality control accuracy.

[0124] The medium-term objective function focuses on parameter stability, and its calculation method is as follows: first, calculate the variance of each water quality parameter in the medium-term prediction period, and the variance value represents the fluctuation degree of the parameter in the period; then multiply the variance of each parameter by the corresponding stability weight coefficient , reflecting the difference in the stability requirements of different parameters; finally, sum all the weighted variances to obtain the overall medium-term objective function value. The smaller the value, the higher the stability of the medium-term water quality parameters.

[0125] The long-term objective function optimizes resource use efficiency, and its calculation method is as follows: first, calculate the costs of reagents, energy consumption, and equipment maintenance, respectively; then multiply these three types of costs by the corresponding cost weight coefficients 、 and , reflecting the importance of different cost types; finally, sum the weighted costs of each type to obtain the overall long-term objective function value. The smaller the value, the higher the resource use efficiency.

[0126] The near-term objective function focuses on parameter deviation, the medium-term objective function focuses on parameter variance, and the long-term objective function focuses on cost efficiency. These indicators have different dimensions and numerical ranges. Before weighting and calculation, the target function values of the three time scales need to be normalized. Relative to historical typical values, proportional normalization or Z-score standardization can be used to ensure the comparability of targets at different time scales in comprehensive evaluation.

[0127] The specific implementation of the parameter importance coefficient is as follows: through data analysis of historical water quality problems, the correlation of each parameter with the overall performance and safety of the system is calculated, and the basic weight is set according to expert knowledge, and is dynamically adjusted according to the current system operation state. Key safety indicators (such as pH value, residual chlorine, etc.) are given higher weights, and parameters with less impact are given lower weights.

[0128] The specific implementation of the stability weight coefficient is as follows: based on the allowed fluctuation range of each parameter and its impact on system stability. Parameters sensitive to fluctuations (such as microbial indicators) are given higher weights, and parameters that allow a certain range of fluctuations (such as conductivity) are given lower weights. The weight value is calculated by analyzing the difference characteristics of historical stable operation data and problem data, and is fine-tuned in combination with operation experience.

[0129] The specific implementation of the cost weight coefficients 、 and The specific implementation is: according to the main target of system optimization and user demand to determine the relative importance of three cost types (reagent cost, energy consumption and equipment maintenance cost). In different system states and operation stages, these weights are dynamically adjusted, for example, increase the energy weight in the high energy consumption season, increase the maintenance cost weight in the equipment aging stage, and increase the reagent cost weight in the period of rising reagent cost. The weight adjustment is based on system state monitoring, cost analysis and economic benefit calculation of prediction.

[0130] The comprehensive evaluation function integrates the above three time scale objective functions by weighted sum method, and the calculation method is: the short-term objective function value is multiplied by the short-term weight coefficient , the medium-term objective function value is multiplied by the medium-term weight coefficient , and the long-term objective function value is multiplied by the long-term weight coefficient Then add them to get the comprehensive evaluation value. The weight coefficient is dynamically adjusted according to the system state, and the long-term weight is increased when the system is stable, and the short-term weight is increased when facing potential risks, to balance the relationship between immediate response and long-term optimization.

[0131] Step 4.3, the optimization algorithm is executed;

[0132] According to the method provided in the application, an improved particle swarm optimization algorithm is used to solve the multi-objective optimization problem. The algorithm is improved according to the characteristics of the reagent proportioning problem, including processing discrete variables, constraint processing and local search strategy enhancement. The algorithm outputs the optimal reagent proportioning scheme, including the addition amount and addition time of various reagents.

[0133] The improved particle swarm optimization algorithm in the application has the following characteristics:

[0134] Mixed coding strategy: using continuous-discrete mixed coding representation scheme, continuous variables represent reagent addition amount, and discrete variables represent addition time.

[0135] Constraint processing module: using adaptive penalty function method to process hard constraints and soft constraints to ensure the feasibility of the solution.

[0136] Multi-population cooperative evolution: maintain multiple sub-populations for parallel search, each sub-population focuses on different optimization directions, and share high-quality solutions through regular information exchange.

[0137] Local search enhancement: apply local optimization techniques such as pattern search to potential solutions to speed up the convergence process.

[0138] Adaptive parameter adjustment: dynamically adjust algorithm parameters (inertia weight, learning factor, etc.) according to search progress, balance global exploration and local development.

[0139] The algorithm execution process is as follows:

[0140] Initialize multiple sub-populations, each of which is for a different optimization objective;

[0141] Evaluate the objective function value and constraint violation degree for each particle; update the individual optimal solution and the global optimal solution; adjust the particle position and velocity according to the update rule; apply local search to promising solutions; exchange information between sub-populations; dynamically adjust algorithm parameters; repeat until the termination condition is met; output the Pareto optimal solution set, and select the final scheme according to the decision maker's preference.

[0142] As Figure 3 shown, the Pareto optimal solution set obtained by the multi-objective optimization algorithm in the present application is shown. Each point in the figure represents a feasible drug ratio scheme, the horizontal axis represents the system resource consumption (including drug cost and energy consumption), and the vertical axis represents the system stability (the inverse of the degree of water quality parameter fluctuation). The curve in the figure is the Pareto frontier, which represents the trade-off relationship between system resource consumption and system stability, verifying the effectiveness of the multi-objective optimization algorithm in the present application. Decision makers can select a suitable balance point on the frontier as the final control scheme according to actual needs.

[0143] Step 5, based on the optimized drug ratio strategy, a multi-time scale collaborative optimization system is realized to further balance short-term response and long-term stability;

[0144] This step constructs a multi-level control architecture to coordinate control strategies at different time scales to achieve the goal of both fast response to short-term disturbances and long-term stability. Specifically, it includes:

[0145] Step 5.1, hierarchical control framework construction;

[0146] A three-layer control architecture including emergency response layer, daily control layer and strategy optimization layer is established. The emergency response layer handles sudden abnormalities with a response time of minutes; the daily control layer performs prediction-based feedforward control with a time scale of hours; the strategy optimization layer is responsible for long-term parameter optimization with a time scale of days or weeks.

[0147] Step 5.2, interlayer coordination system;

[0148] An interlayer information transmission and decision coordination system is constructed to ensure that control decisions at different time scales are compatible and consistent as a whole. The upper layer control provides constraints and guidance for the lower layer, and the lower layer control provides execution feedback and state updates for the upper layer.

[0149] Step 5.3, dynamic adjustment module;

[0150] According to the system state and prediction results, the activation priority of different time scale control layers is dynamically adjusted. In normal operation, the long-term optimization strategy plays a leading role; in the face of potential risks, the medium-term control strengthens intervention; in the event of a sudden event, the short-term response layer takes over the system.

[0151] It should be noted that in some embodiments, the method provided by the present application can also set a transition buffer module between each control layer to ensure smooth transition of control strategy switching between layers and avoid control shock.

[0152] As Figure 4 shown, the performance differences between the method of the present application and the traditional control method are comprehensively evaluated from five dimensions: prediction accuracy, response speed, resource efficiency, environmental adaptability, and system stability. The blue polygon in the figure represents the performance of the multi-time scale collaborative optimization system of the present application, and the red polygon represents the performance of the traditional control method. From the figure, it can be clearly seen that the method of the present application is superior to the traditional method in all five dimensions, especially in prediction accuracy, environmental adaptability, and system stability, fully demonstrating the comprehensive performance advantage of the method and the effect of multi-time scale collaborative optimization.

[0153] Step 6, based on the optimization results of the multi-time scale collaborative optimization system, an environment adaptive active evolution system is constructed to realize automatic adjustment of the system with environmental changes;

[0154] This step enables the system to adapt to environmental changes independently, continuously optimize the model and control strategy, and form a true closed-loop optimization system. Specifically, it includes:

[0155] Step 6.1, model performance monitoring and evaluation;

[0156] The performance of the prediction model is continuously monitored, and the prediction error index is calculated by comparing the predicted value with the actual observed value. When the error exceeds the preset threshold or shows a continuous growth trend, the model update process is triggered.

[0157] Specifically, the system uses a multi-index comprehensive evaluation method, including:

[0158] Root mean square error: evaluates the average deviation of the predicted value from the actual value;

[0159] Mean absolute percentage error: evaluates the relative prediction error;

[0160] R² determination coefficient: evaluates the degree of explanation of the model to the data variation;

[0161] Prediction bias accumulation: detects the systematic bias of the prediction;

[0162] Critical event prediction rate: evaluates the prediction ability of the model for important events (such as parameter mutation).

[0163] The indicators of RMSE, MAPE, R², PBCS and critical event prediction rate have different numerical ranges and meanings. When setting the trigger threshold and making comprehensive evaluation, it is necessary to standardize these indicators to a unified scoring interval (such as 0-1). The percentile conversion based on historical performance distribution or fuzzy logic scoring method can be used to make comprehensive comparison of different types of performance indicators.

[0164] The specific implementation of PBCS is to calculate the signed cumulative value of all prediction deviations in a time window, rather than simply accumulating the absolute error. This method can detect the systematic deviation trend of the prediction model, such as continuously overestimating or underestimating a certain parameter. In specific calculation, first obtain the difference sequence of prediction values and actual values at consecutive time points, and keep the positive and negative signs. Then calculate the cumulative sum of these differences. Finally, divide the cumulative sum by the number of time points to get the standardized PBCS value. When the absolute value of PBCS exceeds the preset threshold and shows a monotonic increasing trend, it indicates that the model has systematic deviation and needs to trigger the model adjustment mechanism.

[0165] The system sets two levels of trigger thresholds: warning threshold and update threshold. When the performance indicator exceeds the warning threshold, the system increases the monitoring frequency. When the update threshold is exceeded or the warning threshold is maintained for consecutive periods, the model update process is triggered.

[0166] Step 6.2, incremental learning and model updating;

[0167] When updating the time series knowledge graph and prediction model using incremental learning method, there is no need to completely retrain. New data is first used to update the graph structure, and then used to fine-tune the prediction model parameters, maintaining the model's sensitivity to the latest data.

[0168] The incremental learning system of the present application contains three key components:

[0169] Selective memory module: evaluate the information value of new data, select high-value samples to join the training set, and remove redundant or outdated samples to maintain the representativeness and timeliness of training data.

[0170] Elastic weight adjustment algorithm: calculate the importance of parameters according to the Fisher information matrix, impose stronger regularization constraints on important parameters, protect the learned key knowledge, and allow the model to adapt to new patterns; first, divide the water quality parameters into fast-changing parameter groups (pH value, residual chlorine) and slow-changing parameter groups (conductivity, hardness), and calculate the second derivative matrix of the log-likelihood function with respect to the model parameters for each group, and then synthesize the overall Fisher matrix. When calculating, special consideration is given to the sensitivity of parameters under different operating conditions (high temperature, low temperature, high load, etc.), and higher weights are given to key points according to historical water quality fluctuation records. When the system detects seasonal changes or sudden changes in raw water quality, the Fisher matrix guides the model to retain key knowledge (such as the long-term correlation between pH value and scaling risk), while allowing adaptive updates (such as the seasonal relationship between microbial activity and temperature). The knowledge distillation framework ensures that the model can quickly adapt to the latest water quality changes (such as the response characteristics of the system after the addition of a new corrosion inhibitor), while retaining stable relationship patterns between original water quality parameters (such as the mapping relationship between conductivity and corrosion rate), ensuring the continuous effectiveness of the circulating water system prediction model.

[0171] Knowledge distillation framework: maintain a teacher model to save historical knowledge, and minimize the deviation from the prediction of the teacher model while fitting new data, to balance new and old knowledge.

[0172] The update process adopts a phased strategy: first, update the graph structure, including adding / removing edges and adjusting edge weights; then fix the graph structure and update the node representation and prediction model parameters; finally, perform global fine-tuning to integrate all components.

[0173] As shown in Figure 5 , the accuracy trends of the incremental learning model and the traditional fixed model during long-term use are compared. The blue area in the figure represents the accuracy of the incremental learning model of the present application, and the red area represents the accuracy of the traditional fixed model. The horizontal axis represents the system running time (months). As can be clearly seen from the figure, the accuracy of the traditional fixed model shows a clear downward trend over time, while the incremental learning model of the present application can maintain a high level of accuracy and quickly recover its performance after changes in system characteristics (such as seasonal transitions). The figure demonstrates the effectiveness of the incremental learning and model update mechanism in the present application, and visually demonstrates the self-evolution ability of the system, providing a guarantee for long-term stable operation.

[0174] Step 6.3, seasonal pattern adaptation;

[0175] Recognize and learn seasonal variation patterns of water quality parameters, and build specialized sub-graphs for different seasons. The system can automatically switch and mix corresponding models and strategies according to the current season and environmental conditions, achieving stable performance throughout the year.

[0176] The seasonal pattern adaptive system includes the following core functions:

[0177] Seasonal pattern recognition: Fourier analysis and wavelet transform are used to detect periodic patterns in time series, automatically identify seasonal cycles and characteristics;

[0178] Seasonal feature extraction: build feature representation for each identified seasonal cycle, capture the typical characteristics and relationship patterns of water quality parameters in that season; Seasonal feature extraction in the circulating water treatment system uses multi-level feature analysis method, first decomposes the water quality time series data into different frequency components through wavelet transform, and identifies seasonal cycles; Then apply dimensionality reduction techniques to extract seasonal discriminant feature vectors from high-dimensional water quality parameter space, including temperature-pH correlation features, conductivity-microbial activity joint features, etc. A seasonal feature extraction template library is established for the characteristics of the circulating water system: the summer template focuses on extracting the feature combination of accelerated microbial reproduction, intensified pH fluctuation, and increased residual chlorine consumption rate under the condition of rising water temperature; The winter template focuses on the characteristics of the change of scaling tendency and the decrease of microbial activity in low temperature environment; The spring and autumn transition season extracts the turbidity fluctuation characteristics caused by rainfall changes. The system determines the current operating seasonal pattern by calculating the similarity between the current water quality parameter sequence and each seasonal template, and automatically calls the corresponding prediction model and control strategy, such as actively increasing the frequency of germicide addition in summer mode, adjusting the pH target range to 7.2-7.8, and adjusting the blowdown cycle according to the water temperature-microbial relationship; The winter mode optimizes the scale inhibitor formula and adjusts the proportion of corrosion inhibitor and dispersant to 2:1 to respond to the change of scaling risk under low temperature conditions, realizing intelligent seasonal regulation and control of the circulating water treatment system;

[0179] Seasonal sub-model library: build specialized prediction models and control strategies for different seasonal conditions to form a model library;

[0180] Smooth transition system: build a gradual switching algorithm during the seasonal transition period to achieve smooth transition of the model and avoid sudden changes in control strategy;

[0181] Seasonal precursor detection: based on environmental parameters and early indicators, predict seasonal changes in advance, preset corresponding models, and achieve seamless adaptation.

[0182] In addition, in some embodiments, the method provided by the present application can also combine historical annual data to build a long-term memory library of seasonal patterns, further improving the accuracy of seasonal prediction.

[0183] For example, Figure 6As shown, the relationship between water temperature and microbial activity index in the circulating water system is shown with seasonal variation. The column chart represents the monthly average water temperature change, and the line chart represents the corresponding microbial activity index change. As can be seen from the figure, the microbial activity index and water temperature show a clear seasonal correlation pattern, but the relationship is not simply linear, reflecting the complex seasonal factors in the system. The figure verifies the application scenario of the system seasonal pattern adaptive function, intuitively demonstrates the necessity of the environment adaptation-active evolution system, and proves that the invention can effectively cope with the changes in system characteristics caused by seasonal changes.

[0184] Step 6.4, control effect feedback and optimization;

[0185] Record each control operation and its effect, and build a control-response database. Regularly analyze the control effect, evaluate the effectiveness of the control strategy, and optimize the control strategy using reinforcement learning method to improve the overall performance of the system.

[0186] The control effect feedback and optimization module uses a model-based reinforcement learning framework, which specifically includes:

[0187] Control-response database: records each control operation, system state and response effect, and builds a structured database;

[0188] Effect evaluation engine: comprehensive evaluation of control effect through multi-dimensional indicators (water quality stability, resource efficiency, energy consumption, etc.);

[0189] Strategy optimizer: based on the double-delay deep deterministic policy gradient algorithm, learn the optimal control strategy from historical control data;

[0190] Simulation verification environment: based on the time sequence knowledge graph to build a simulation environment for verifying the effectiveness and safety of the new strategy;

[0191] Gradual deployment system: adopt conservative strategy deployment scheme, test new strategy in limited range first, and gradually expand application range after verification.

[0192] Through this closed-loop feedback and optimization system, the system can continuously learn from actual operation experience, continuously optimize the control strategy, and form a truly environment adaptation-active evolution intelligent control system.

[0193] As Figure 7As shown, the performance differences of the traditional water quality control method and the method of the present application in the three key indicators of prediction accuracy, early intervention time, and reagent use efficiency are intuitively compared. The blue column in the figure represents the traditional method, and the green column represents the method of the present application. From the figure, it can be clearly seen that the prediction accuracy of the method of the present application is improved by about 45%, the early intervention time is improved from 0 hours (passive response) of the traditional method to an average of 12 hours (active prevention), and the reagent use efficiency is improved by about 30%. These data intuitively show the technical advantages of the present application, verify the actual performance of the above technical effects, and fully prove the innovation value and application potential of the method of the present application in the field of circulating water treatment.

[0194] A circulating water treatment closed-loop optimization system based on water quality prediction is used to perform the above-mentioned circulating water treatment closed-loop optimization method based on water quality prediction, comprising:

[0195] A time sequence knowledge graph module is used to represent the time sequence evolution characteristics of water quality parameters and their relationships.

[0196] A multi-time scale prediction engine is used to predict future water quality states based on the time sequence knowledge graph.

[0197] A predictive constraint generator is used to convert future water quality predictions into current decision constraints.

[0198] A multi-objective optimization module is used to optimize reagent proportioning strategies according to prediction results and constraint conditions.

[0199] A multi-time scale collaborative control module is used to balance the relationship between short-term response and long-term stability.

[0200] An environmental adaptive active evolution module is used to automatically adjust the system as the environment changes.

[0201] The embodiments of the present application are described above, but the embodiments are not limited to the specific implementation described above. The specific implementation described above is only illustrative, not limiting, and those skilled in the art can make more forms of equivalent embodiments under the inspiration of the embodiments, which are all within the protection of the embodiments.

Claims

1. A circulating water treatment closed loop optimization method based on water quality prediction, characterized in that, The method comprises the following steps: Constructing a time series knowledge graph supporting the time dimension to represent the time series evolution characteristics of water quality parameters and their relationships; Based on the time series knowledge graph, a multi-time scale prediction engine is constructed to predict the future water quality state; Developing a predictive constraint generator based on the prediction results of the future water quality state to convert the future water quality prediction into current decision constraints; Introducing a spatio-temporal sensitive multi-objective optimization algorithm to optimize the dosage ratio strategy according to the prediction results and current decision constraints, and based on the generated constraints, using a multi-objective optimization algorithm to optimize the efficiency of the dosage and energy consumption while meeting the water quality parameter targets, specifically including: Dosage ratio model construction, constructing a response model of the relationship between dosage addition and water quality parameters, considering the interaction and hysteresis effect between dosages; Multi-time scale objective function construction, constructing objective functions of three time scales of short-term response speed, medium-term stability and long-term cost benefit, focusing on different optimization priorities; Optimization algorithm execution, using an improved particle swarm optimization algorithm, using continuous-discrete hybrid coding, adaptive penalty function and multi-population co-evolution strategy to solve the multi-objective optimization problem; Based on the optimized dosage ratio strategy, a multi-time scale collaborative optimization system is realized to construct a multi-level control architecture to coordinate the control strategy at different time scales, specifically including: Hierarchical control framework construction, establishing a three-layer control architecture including an emergency response layer, a daily control layer and a strategy optimization layer; Interlayer coordination system, constructing an interlayer information transmission and decision coordination system to ensure that the control decisions of different time scales are compatible and consistent as a whole; Dynamic adjustment module, dynamically adjusting the activation priority of different time scale control layers according to the system state and prediction results; Based on the optimization results of the multi-time scale collaborative optimization system, an environmental adaptive active evolution system is constructed to realize automatic adjustment of the system with environmental changes.

2. The method of claim 1, wherein, The step of constructing a time series knowledge graph supporting the time dimension comprises: Defining water quality parameters as entity nodes in the graph, each node containing parameter name, value and timestamp attributes; Applying a relationship discovery algorithm based on Granger causality test to identify potential causal relationships between parameters and represent the potential causal relationships as directed edges in the graph; Using a graph neural network-based time series pattern learning method to extract the evolution rules of water quality parameters over time from historical data.

3. The method of claim 1, wherein, The multi-time scale prediction engine is constructed based on a hybrid architecture of graph convolution network and gated recurrent unit, including: Spatial feature extraction module for capturing spatial dependency between nodes; Time series feature extraction module for capturing time series characteristics; Attention fusion layer for fusing spatial features and time series features; Multi-time scale prediction head including three parallel branches of short-term prediction, medium-term prediction and long-term prediction.

4. The method of claim 3, wherein the method further comprises: The multi-time scale prediction engine also introduces Monte Carlo dropout technology to provide confidence intervals for each prediction result, obtains prediction distribution through random dropout forward propagation, calculates the mean as the final prediction value and the variance as the uncertainty measure.

5. The method of claim 1, wherein, The predictive constraint generator comprises: The key threshold and target interval definition module defines the target range and warning threshold for each water quality parameter. The predicted trajectory risk assessment module performs risk analysis on the predicted water quality parameter trajectory, taking into account the importance of the parameter, the degree of deviation, and the probability of crossing the boundary. The dynamic constraint generation module generates constraint conditions for control decisions based on the risk assessment results.

6. The method of claim 1, wherein, The environment-adaptive active evolution system includes: The model performance monitoring and evaluation module continuously monitors the performance of the prediction model by comparing the predicted values with the actual observed values to calculate the prediction error indicators. The incremental learning and model updating module updates the time series knowledge graph and prediction model using incremental learning methods. The seasonal pattern adaptive module identifies and learns the seasonal variation patterns of water quality parameters. The control effect feedback and optimization module records control operations and their effects and optimizes control strategies using reinforcement learning methods.

7. The method of claim 6, wherein the method further comprises: The incremental learning and model updating module includes three key components: selective memory module, elastic weight adjustment algorithm, and knowledge distillation framework, which are used to evaluate the value of new data, protect key knowledge, and balance new and old knowledge.

8. A circulating water treatment closed loop optimization system based on water quality prediction, characterized in that, A closed-loop optimization method for circulating water treatment based on water quality prediction is provided, which includes: The time series knowledge graph module is used to represent the time series evolution characteristics of water quality parameters and their relationships. The multi-time scale prediction engine is used to predict future water quality based on the time series knowledge graph. The predictive constraint generator is used to convert future water quality prediction into current decision constraints. The multi-objective optimization module is used to optimize the reagent proportioning strategy according to the prediction results and constraint conditions. The multi-time scale collaborative control module is used to balance the relationship between short-term response and long-term stability. The environment-adaptive active evolution module is used to automatically adjust the system to changes in the environment.

Images