Intelligent decision fusion system for multi-stage process coordination of sewage plant

Abstract

The application discloses a sewage plant multi-process section coordination intelligent decision fusion system, comprising: a perception and rule fusion layer, which is used for collecting water quality parameters, process control parameters and running state parameters of the multi-process section in real time, pre-processing data, and fusing an expert rule base; a mechanism and data driven joint modeling layer, which is used for establishing a mechanism model and a data driven model under the constraint condition of the expert rule base, and respectively predicting control variables; a collaborative optimization and fusion decision layer, which is used for executing multi-objective collaborative optimization and fusion decision across process sections, calculating dynamic fusion weights and generating final control variables under the constraint condition of the expert rule base based on the output results of the mechanism and data driven joint modeling layer, issuing the final control variables to an execution device, and realizing closed-loop optimization by combining real-time feedback and adaptive adjustment. The system improves the stability, decision accuracy and resource utilization efficiency of sewage treatment, and reduces environmental risks.

Description

Technical Field

[0001] This invention relates to the intersection of wastewater treatment and artificial intelligence, and in particular to an intelligent decision-making fusion system for the coordinated operation of multiple processes in a wastewater treatment plant. Background Technology

[0002] Wastewater treatment plants are important infrastructure for ensuring urban water environment quality and water resource security. Their operational efficiency and the stability of effluent quality are directly related to environmental protection and sustainable social development.

[0003] A typical wastewater treatment plant usually includes multiple process units such as pretreatment, biological treatment, advanced treatment and sludge treatment. There are significant material and energy coupling relationships between these units, and their operation and adjustment exhibit characteristics such as multivariability, strong coupling, nonlinearity and time-varying nature.

[0004] In actual operation, the control system must not only ensure that the effluent quality meets the standards in a long-term stable manner, but also take into account economic and environmental objectives such as energy consumption, chemical consumption, and sludge production, so as to achieve comprehensive optimization of multiple objectives.

[0005] Currently, decision-making models for different process stages in wastewater treatment plants often operate independently. However, in actual operation, the operating states of each process stage influence each other. For example, the dissolved oxygen adjustment of the aeration system not only affects the organic matter removal efficiency of the biological treatment stage but also impacts the load and energy consumption of subsequent advanced treatment stages. Without global collaborative optimization, local optimization often sacrifices overall performance.

[0006] Meanwhile, the application of real-time feedback in control optimization is also relatively limited. Existing methods usually only perform simple error correction and lack a closed-loop adaptive mechanism that combines short-term rapid adjustment with long-term strategy optimization, making it difficult to achieve continuous and stable optimal operation under different time scales and operating conditions. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent decision-making fusion system for multi-process collaboration in wastewater treatment plants, so as to achieve multi-objective optimization decision-making and dynamic fusion control across process stages.

[0008] To address this, the present invention provides an intelligent decision-making fusion system for multi-process stage collaboration in wastewater treatment plants, comprising a perception and rule fusion layer, a mechanism and data-driven joint modeling layer, and a collaborative optimization and fusion decision layer. The perception and rule fusion layer collects in-process and effluent water quality parameters, process control parameters, and operating status parameters in real time from the pretreatment, biochemical treatment, advanced treatment, and sludge treatment stages of the wastewater treatment plant, preprocesses the data, and integrates it with an expert rule base. The mechanism and data-driven joint modeling layer establishes a mechanism model and a data-driven model under the constraints of the expert rule base. The mechanism model performs process stage simulation prediction and outputs mechanism-predicted control quantities, while the data-driven model makes predictions and outputs AI-predicted control quantities. The collaborative optimization and fusion decision layer, under the constraints of the expert rule base, performs multi-objective collaborative optimization and fusion decision-making across process stages based on the output of the mechanism and data-driven joint modeling layer, calculates dynamic fusion weights, generates final control quantities, distributes the final control quantities to the execution equipment, and achieves closed-loop optimization through real-time feedback and adaptive adjustment.

[0009] This invention integrates multi-source sensing, dual-model prediction, cross-segment optimization, and closed-loop execution technologies to upgrade the "segmented management" model of wastewater treatment plants to a "full-process collaborative control" model, leading wastewater treatment plants towards digitalization and intelligence. The application of this system can reduce environmental risks and resource waste throughout the entire process, providing key technical support for wastewater treatment plants to achieve high-quality, low-carbon, and intelligent operation goals.

[0010] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the figures. Attached Figure Description

[0011] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0012] Figure 1 This is a structural block diagram of the intelligent decision-making fusion system for multi-stage process coordination in wastewater treatment plants according to the present invention;

[0013] Figure 2 This is a schematic diagram of the multi-process segment collaborative optimization process of the present invention;

[0014] Figure 3 This is a schematic diagram of the decision fusion process of the intelligent decision fusion system of the present invention. Detailed Implementation

[0015] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] Combined with reference Figures 1 to 3 The wastewater treatment plant multi-process collaborative intelligent decision-making fusion system of the present invention includes a perception and rule fusion layer, a mechanism and data-driven joint modeling layer, and a collaborative optimization and fusion decision-making layer.

[0017] The perception and rule fusion layer is used to collect influent and effluent water quality parameters, process control parameters, and operating status parameters in real time from the pretreatment, biochemical treatment, advanced treatment, and sludge treatment sections of the wastewater treatment plant. The data is preprocessed by cleaning, standardizing, temporalizing, and aligning, and then integrated with an expert rule base built based on expert experience and process specifications to provide safe boundaries and constraints for model operation.

[0018] The mechanism and data-driven joint modeling layer is used to establish mechanism models and data-driven models under the constraints of expert rule base. The mechanism model simulates the operating status of each process section based on material balance and kinetic equations and outputs mechanism predictive control quantities. The data-driven model constructs a deep learning prediction structure based on historical multivariate time series and outputs AI predictive control quantities.

[0019] The collaborative optimization and fusion decision layer is used to perform multi-objective collaborative optimization and dynamic fusion weight calculation on the mechanism predictive control quantity and the AI ​​predictive control quantity under the constraint of the expert rule base, generate the final control quantity and send it to the execution device, and at the same time combine real-time feedback to perform closed-loop adaptive adjustment, so as to achieve short-term rapid adjustment and long-term strategy optimization.

[0020] The collaborative optimization and fusion decision-making layer is the core of this system. This layer predicts control quantities through fusion mechanisms and AI-based predictions, calculates the final control quantity, and sends it to the execution devices.

[0021] During the optimization process, the system comprehensively considers multiple objectives such as water quality compliance, energy consumption, and reagent consumption, and employs a multi-objective weighted sum method for optimization. The system also features a dynamic weight fusion mechanism: when the AI ​​model has high accuracy, AI predictive control is prioritized; when the prediction error increases, it reverts to the mechanistic model control, ensuring stable system operation. This layer also utilizes closed-loop adaptive adjustment, automatically adjusting the control strategy based on real-time feedback to ensure the system's adaptability and robustness under different operating conditions, thereby improving the wastewater treatment plant's operational efficiency and water quality stability.

[0022] The working process of the perception and rule fusion layer includes the following steps S11 and S12.

[0023] S11. Collect multi-source sensing data and preprocess it; the sensing data includes influent water quality parameters, process control parameters and operating status parameters.

[0024] S12. Based on expert experience and process operation specifications, generate an expert rule base containing safety thresholds, logical constraints and triggering conditions to provide safety boundaries for subsequent model calculations.

[0025] In this invention, after collecting the sensing data with the help of an automatic sampling and detection system, the collected sensing data is preprocessed by combining the functions of relevant data processing software with practical experience in manual preprocessing.

[0026] In specific implementation, step S11, the process of preprocessing multi-source sensing data, includes the following steps S111 to S113.

[0027] S111, the data cleaning step includes noise removal, handling missing values ​​and outliers.

[0028] By performing noise filtering, missing value imputation, and outlier processing on multi-source sensing data, distortion problems caused by sensor errors, transmission interference, or sudden operating conditions during data acquisition are eliminated, ensuring the basic integrity and reliability of the sensing data and providing a high-quality data source for subsequent processing.

[0029] S112, Perceived data standardization steps, including performing zero-mean unit variance standardization.

[0030] The cleaned sensory data is processed in a planned manner. The zero-mean unit variance standardization z-score standardization method is used to unify the scale of the cleaned data. By converting sensory data of different magnitudes and dimensions into standard normal distribution data with a mean of 0 and a standard deviation of 1, the influence of scale differences between features on model training is eliminated. This avoids the dominance of model prediction results by a certain type of parameter with an excessively large value range, and ensures that each feature has an equal weight in subsequent modeling.

[0031] S113, the time-series processing of the perceived data includes constructing a multi-time-step sequence in chronological order and generating sliding window samples.

[0032] To address the temporal characteristics of the wastewater treatment process, standardized sensor data is sorted by collection timestamps to construct continuous time-series data. A sliding window technique is employed to divide the data into multiple time windows, each containing data samples from several time steps. Sampling is performed through the sliding window, generating multiple sets of time-series sample data through continuous sliding window operations.

[0033] S114, Sensing data alignment step, including interpolation and timestamp consistency processing of data with different sampling frequencies and time bases.

[0034] Interpolation and timestamp consistency processing are performed on data with different sampling frequencies and time bases. By unifying the timestamp base, the timestamps of each data source are calibrated to ensure that the sensing data collected from different data sources or at different times are strictly aligned in the time dimension.

[0035] In specific implementation, in step S12, expert experience and process specifications are transformed into a set of rule parameters, including the safety threshold range of key control variables in each process section, the logical constraint relationship between variables, and the triggering conditions of abnormal operating conditions. The set of rule parameters is stored in the expert rule base for the mechanism model and the data-driven model to call, so as to provide safety boundary constraints, operation status monitoring and boundary crossing warning functions during the model operation.

[0036] In specific implementation, the working process of the mechanism and data-driven joint modeling layer includes the following steps:

[0037] S21. Establish a mechanism model and a data-driven model under the constraints of the expert rule base;

[0038] S22. The mechanism model is used to perform process section simulation and prediction and output the mechanism prediction control quantity;

[0039] S23. The data-driven model makes predictions and outputs AI predictive control values.

[0040] The mechanism- and data-driven joint modeling layer is built upon mechanism models, data-driven models, and expert rule bases, providing dual predictive control basis for wastewater treatment processes.

[0041] The joint modeling layer of mechanism and data-driven approaches includes three parts: the construction and simulation prediction of the mechanism model, which simulates the physical, chemical and biological reactions of each process stage based on material balance and kinetic equations, and outputs the mechanism predictive control quantity; the training and prediction of the data-driven model, which uses deep learning networks to learn the patterns of historical time series data and outputs the AI ​​predictive control quantity; and the parallel collaboration of the two types of models under the constraints of the expert rule base to ensure that the predictive control quantity is within the safe operating range and to provide reliable basic data for subsequent optimization decisions.

[0042] In step S21, a mechanistic model and a data-driven model are established based on the constraints of the expert rule base. The mechanistic model is used to characterize and describe the physical, chemical and biological reaction mechanisms in the wastewater treatment process; the data-driven model is used to characterize the nonlinear mapping relationship between input features and target control quantities.

[0043] Mechanism models are used to characterize and describe the physical, chemical, and biological reaction mechanisms in wastewater treatment. They quantify processes such as flocculation and sedimentation in the pretreatment stage, microbial metabolism in the biochemical stage, and reagent reactions in the advanced treatment stage through material balance equations and reaction kinetic equations. Safety thresholds in the expert rule base are used as boundary conditions for the model, providing a theoretical explanation and a basis for stable prediction of the process essence for the system.

[0044] The data-driven model is used to characterize the nonlinear mapping relationship between input features and target control quantities. It constructs a deep learning network based on time series prediction structure, takes multivariate time series data output from the perception and rule fusion layer as input, and uses logical constraints in the expert rule base as training boundaries. By learning the hidden patterns in historical operation data, it accurately captures the complex nonlinear correlation between process variables and realizes dynamic prediction of control quantities.

[0045] In specific implementation, in step S22, the mechanism model is constructed based on the material balance and kinetic equations and the operating mechanism of each process segment, and the operating status of each process segment is simulated in real time, and the mechanism prediction control quantity is output.

[0046] In the system of this invention, the mechanistic model integrates the core operational laws of each process stage in wastewater treatment, including the kinetics of bar screen interception and grit settling in the pretreatment stage, the reaction mechanism of activated sludge in the biological treatment stage, the chemical coagulation and adsorption laws in the advanced treatment stage, and the dewatering and drying characteristics in the sludge treatment stage. By combining the mass balance equation with the reaction kinetic equation, a dynamic simulation model covering the entire process is constructed. This model can calculate in real time the changes of key parameters in each process stage under different operating conditions, such as dissolved oxygen concentration in the aeration tank, sludge interface height in the sedimentation tank, and head loss in the filter. Based on this, corresponding mechanistic predictive control quantities are output, such as aeration rate adjustment values, chemical dosage baselines, and sludge return ratios. This process-based modeling approach can maintain predictive stability under complex operating conditions, providing the system with control basis that conforms to the core laws of wastewater treatment, and supporting the scientific and reliable nature of the decision-making process.

[0047] In specific implementation, in step S23, the data-driven model is constructed based on historical multivariate time series data, and a deep learning network based on time series prediction structure is used to output AI predictive control quantities.

[0048] In the system of this invention, the data-driven model is based on historical operational data of multiple process sections processed by the perception and rule fusion layer, covering multi-dimensional time-series data such as influent and effluent water quality parameters, process control parameters, and operational status parameters. By constructing a deep learning network suitable for time series prediction, the historical data is transformed into input feature sequences and target control variable labels using the sliding window method. With safety thresholds and logical constraints in the expert rule base as training boundaries, the network parameters are iteratively optimized through the backpropagation algorithm, enabling the model to learn the hidden dynamic correlation patterns in the data.

[0049] In real-time operation, the model receives the latest multivariate time series data, and after feature extraction and pattern matching, outputs dynamic AI predictive control quantities for each process section, such as refined aeration volume adjustment values ​​based on influent fluctuations and reagent dosage correction amounts based on effluent water quality trends, thereby achieving accurate dynamic prediction of the process and making up for the limitations of mechanistic models in dealing with complex nonlinear operating conditions.

[0050] Through the above steps S21-S23, the mechanism and data-driven joint modeling layer organically integrates the mechanism model, the data-driven model, and the expert rule base. Through dual modeling, parallel prediction, and boundary constraints, it provides control outputs that are both theoretically based and data-supported for collaborative optimization decision-making.

[0051] Based on sensing data and historical operating patterns, the system can simultaneously generate mechanism-based predictive control quantities and AI-based predictive control quantities. These quantities not only conform to the core reaction mechanism of wastewater treatment but also capture the nonlinear dynamic characteristics under complex operating conditions, and all remain within the safe operating range defined by the expert rule base.

[0052] The application of the mechanism-driven and data-driven joint modeling layer not only ensures the stability and interpretability of predictive control, but also improves its adaptability and accuracy to complex operating conditions, providing a reliable dual predictive basis for subsequent collaborative optimization and fusion decision-making, and supporting the intelligent control of the wastewater treatment process.

[0053] In practice, the collaborative optimization and fusion decision-making process includes the following steps:

[0054] S31. Perform cross-process multi-objective collaborative optimization under the constraints of the expert rule base;

[0055] S32. Dynamically fusion weights of mechanistic predictive control and AI predictive control to generate fused control;

[0056] S33. The integrated control quantity is issued and executed, and closed-loop optimization is achieved by adaptive adjustment based on feedback.

[0057] The collaborative optimization and fusion decision layer is the core execution part of the intelligent decision fusion system for wastewater treatment. It is based on a multi-objective collaborative optimization and dynamic fusion weight mechanism across process stages and a closed-loop feedback adjustment strategy to achieve the optimization of the wastewater treatment process.

[0058] The work of the collaborative optimization and fusion decision layer includes three parts: multi-objective collaborative optimization, dynamic adjustment of mechanism and data-driven models, fusion weight generation of control variables, and closed-loop feedback and adaptive adjustment.

[0059] In specific implementation, step S31 involves performing cross-process multi-objective collaborative optimization under the constraints of the expert rule base. Taking into account objectives such as effluent quality compliance rate, energy consumption, and reagent dosage, a multi-objective optimization weighted sum method is used to set weights and objective functions for multi-objective optimization. The optimization objective function is as follows:

[0060]

[0061] Where Z is the overall optimization objective value; w i Let be the weight coefficient of the i-th objective, satisfying and ; Let i be the optimization objective function; Let be the i-th control vector; n is the number of optimization objectives.

[0062] Collaborative optimization employs a constrained multi-objective optimization algorithm, using Newton's iteration method to iteratively update the control input. The control input update formula is:

[0063]

[0064] in, The control vector for the dth iteration; For the joint cost function; It is the first-order gradient vector, that is, the partial derivatives of each control variable with respect to the overall target; It is the second-order Hessian matrix, i.e., the second-order partial derivative; The learning rate is adaptive and dynamically adjusted based on the iteration step size.

[0065] In specific implementation, in step S32, based on the results of multi-objective collaborative optimization control, the mechanism prediction control quantity of the collaborative optimization output is... AI predictive control quantity The formula for dynamically adjusting the fusion weights can be expressed as follows:

[0066]

[0067] in, is the weighting adjustment coefficient; E is the real-time feedback error, which is the weighted sum of the deviations between the actual and predicted values ​​of key indicators in the process section. The maximum allowable error threshold; To improve the accuracy of AI model predictions; when hour, Only use To ensure safety.

[0068] The final control quantity is obtained after dynamic weighted fusion. The final control quantity fusion formula is:

[0069]

[0070] in, For dynamic fusion weights, range ; This is the final control quantity to be issued and executed; Mechanism predictive control inputs for collaborative optimization of output; To collaboratively optimize the output of AI predictive control variables.

[0071] This dynamic fusion weighting mechanism ensures that the optimized results are prioritized when the AI ​​model has high accuracy and low error, while gradually reverting to the mechanistic model control quantity when the prediction error increases or the operation becomes unstable, so as to ensure operational safety and water quality stability.

[0072] In specific implementation, in step S33, the final fused control quantity will be... The command is sent to the corresponding device for execution, and execution feedback signals are collected in real time. and the target set value The deviation is calculated, and closed-loop optimization is achieved through adaptive adjustment based on feedback. The formula for calculating the feedback deviation is:

[0073]

[0074] in, Set a value for the target; This is the actual execution feedback value; This represents the control deviation at the current moment.

[0075] Closed-loop optimization is performed based on the deviation. The closed-loop control adopts proportional-integral-derivative (PID) regulation, and the calculation formula is as follows:

[0076]

[0077] in, For feedback adjustment signal; This is a proportionality coefficient, reflecting the intensity of the immediate response to the deviation; These are the integral coefficients used to eliminate steady-state errors; These are the differential coefficients, used to suppress rapid fluctuations in the system; This is the rate of change of deviation, used to predict deviation trends in advance.

[0078] Simultaneously, based on closed-loop optimization, an adaptive adjustment mechanism is introduced to update the dynamic fusion weights and optimization model parameters according to the feedback error trend, forming a short-term fast adjustment loop and a long-term strategy optimization loop, dynamically updating the fusion weights. AI model hyperparameters and the sensitivity coefficient of the mechanism model This ensures the stability and robustness of control performance under different operating conditions.

[0079] Based on real-time comprehensive error, with periods ranging from minutes to hours. AI accuracy The fusion weights are dynamically adjusted to achieve a short-term, rapid adjustment loop. The calculation formula is as follows:

[0080]

[0081] in, The fusion weight at the current moment, with a value range of [0,1]; To integrate the weight update step size and control the rate of weight change; For real-time comprehensive error; The maximum permissible composite error threshold; This is a metric for the accuracy of AI models within the most recent prediction window.

[0082] The AI ​​model hyperparameters are updated daily or weekly based on historical operating data and feedback evaluation results to achieve a long-term strategy optimization loop for the AI ​​model hyperparameters. The calculation formula is as follows:

[0083]

[0084] in, represents the hyperparameters of the AI ​​model; m represents the number of long-term optimization iterations. Update the learning rate for hyperparameters; For AI model hyperparameters Gradient computation operator, This is the loss function for the AI ​​model.

[0085] Parameter correction amount of computer theoretical model based on feedback data And update the sensitivity coefficient to achieve adaptive correction of the sensitivity coefficient of the mechanistic model. The calculation formula is as follows:

[0086]

[0087] in, This is a set of sensitivity coefficients for the mechanistic model, including reaction rate constants, conversion efficiency parameters, etc. Update the step size for sensitivity; This is the sensitivity correction amount calculated based on the feedback data.

[0088] This invention, at the collaborative optimization and fusion decision-making level, combines multi-objective optimization across process segments, dynamic fusion of dual models, and closed-loop adaptive adjustment to balance objectives such as water quality, energy consumption, and reagent dosage. It employs a multi-objective optimization algorithm to iteratively optimize the control quantity, calculates the fusion weight in real time to generate the final control quantity, and achieves full-process closed-loop optimization based on PID regulation and model parameter updates, ensuring efficient collaboration among various process segments.

[0089] The intelligent decision-making fusion system of this invention enables intelligent collaborative operation of the entire wastewater treatment plant process, covering multi-process optimization, cross-process operation monitoring, and global anomaly early warning. This invention has the following technical features / advantages:

[0090] 1. The intelligent decision-making fusion system of this invention can monitor key operating parameters of multiple process stages in a wastewater treatment plant in real time. The system collects data in real time through automatic sampling and intelligent sensors, and combines multi-source data preprocessing with an expert rule base to provide accurate water quality predictions and anomaly warnings, ensuring rapid response and reducing potential risks in emergency situations.

[0091] 2. The intelligent decision-making fusion system of this invention supports multi-process segment linkage intelligent control. It not only ensures the stable operation of individual process segments but also achieves overall optimization of the entire process, significantly improving the anti-interference capability and operational reliability of the processing system.

[0092] 3. The intelligent decision fusion system of this invention provides intelligent decision support throughout the entire process. It can autonomously adjust the control strategy under different operating conditions to achieve the optimal configuration of the wastewater treatment process, ensuring treatment effect and operational stability.

[0093] 4. The intelligent decision-making fusion system of this invention features end-to-end low-carbon and environmentally friendly characteristics. It prioritizes low-energy-consumption operating parameters through a multi-objective optimization algorithm, reducing ineffective energy consumption. Simultaneously, the system can monitor the pollutant removal efficiency and final emission indicators of each process stage in real time, ensuring that the treatment process meets environmental standards.

[0094] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, substitutions, or improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A smart decision-making fusion system for multi-process stage collaboration in a wastewater treatment plant, characterized in that, It includes a perception and rule fusion layer, a mechanism and data-driven joint modeling layer, and a collaborative optimization and fusion decision-making layer. The perception and rule fusion layer is used to collect influent and effluent water quality parameters, process control parameters, and operating status parameters in real time from the pretreatment section, biochemical treatment section, advanced treatment section, and sludge treatment section of the sewage treatment plant, preprocess the data, and integrate it with the expert rule base. The mechanism and data-driven joint modeling layer is used to establish a mechanism model and a data-driven model under the constraints of the expert rule base. The mechanism model performs process section simulation prediction and outputs mechanism predictive control quantity, while the data-driven model performs prediction and outputs AI predictive control quantity. The collaborative optimization and fusion decision-making layer, under the constraints of the expert rule base, performs multi-objective collaborative optimization and fusion decision-making across process sections based on the output of the mechanism and data-driven joint modeling layer. It calculates dynamic fusion weights and generates the final control quantity, which is then sent to the execution equipment. Combined with real-time feedback, adaptive adjustments are made to achieve closed-loop optimization. The working process of the collaborative optimization and fusion decision-making layer includes the following steps: S31. Perform cross-process multi-objective collaborative optimization under the constraints of the expert rule base; S32. Dynamically fusion weights of mechanistic predictive control and AI predictive control to generate fused control; S33: Integrate control inputs for execution, and perform adaptive adjustments based on feedback to achieve closed-loop optimization. In step S31, performing cross-process multi-objective collaborative optimization under the constraints of the expert rule base includes: For objectives such as effluent quality compliance rate, energy consumption, and reagent dosage, a multi-objective optimization weighted sum method is used to set weights and objective functions for multi-objective optimization. The optimization objective function is as follows: ， in, The overall optimization objective value; Let be the weight coefficient of the i-th objective, satisfying and ; Let i be the optimization objective function; Let be the i-th control vector; n is the number of optimization objectives. Collaborative optimization employs the constrained multi-objective optimization algorithm, Newton's iteration method, to iteratively update the control input. The control input update formula is: ， in, for The control vector for the next iteration; For the joint cost function; It is the first-order gradient vector, that is, the partial derivatives of each control variable with respect to the overall target; It is the second-order Hessian matrix, i.e., the second-order partial derivative; The adaptive learning rate is dynamically adjusted based on the iteration step size. In step S32, based on the multi-objective system optimization control result in S31, the mechanism prediction control quantity of the cooperative optimization output is calculated. AI predictive control quantity The formula for dynamically adjusting the fusion weights can be expressed as follows: ， in, This is the weighting adjustment coefficient; The error is fed back in real time, which is the weighted sum of the deviations between the actual and predicted values ​​of key indicators in the process section; The maximum allowable error threshold; To improve the accuracy of AI model predictions; when , Only use To ensure safety, The final control quantity is obtained after dynamic weighted fusion. The final control quantity fusion formula is: ， in, For dynamic fusion weights, range ; This is the final control quantity to be issued and executed; Mechanism predictive control inputs for collaborative optimization of output; To collaboratively optimize the output of AI predictive control variables.

2. The intelligent decision-making fusion system for multi-process stage collaboration in wastewater treatment plants according to claim 1, characterized in that, The integrated expert rule base includes: transforming expert experience and process specifications into a set of rule parameters, including the safety threshold range of key control variables for each process segment, the logical constraint relationship between variables, and the triggering conditions for abnormal operating conditions. The set of rule parameters is stored in the expert rule base for use by the mechanism model and the data-driven model, so as to provide safety boundary constraints, operation status monitoring, and out-of-bounds warning functions during model operation.

3. The intelligent decision-making fusion system for multi-process stage collaboration in wastewater treatment plants according to claim 1, characterized in that, Data preprocessing includes the following steps: S111, Data cleaning, including noise removal, handling missing values ​​and outliers; S112. Data standardization, including performing zero-mean, unit-variance standardization. S113. Data temporalization, including constructing multi-time-step sequences in chronological order and generating sliding window samples; S114. Data alignment, including interpolation and timestamp consistency processing for data with different sampling frequencies and time bases.

4. The intelligent decision-making fusion system for multi-process stage collaboration in wastewater treatment plants according to claim 1, characterized in that, The mechanistic model is used to characterize and describe the physical, chemical and biological reaction mechanisms in the wastewater treatment process. The mechanistic model is constructed based on the material balance and kinetic equations and the operating mechanism of each process section, and simulates the operating state of each process section in real time, outputting mechanistic prediction and control quantities.

5. The intelligent decision-making fusion system for multi-process stage collaboration in wastewater treatment plants according to claim 1, characterized in that, The data-driven model is used to characterize the nonlinear mapping relationship between input features and target control quantities. The data-driven model is constructed based on historical multivariate time series data and uses a deep learning network based on time series prediction structure to output AI predictive control quantities.

6. The intelligent decision-making fusion system for multi-process stage collaboration in wastewater treatment plants according to claim 1, characterized in that, In step S33, the final fusion control quantity obtained in S32 is... The command is sent to the corresponding device for execution, and execution feedback signals are collected in real time. and the target set value The deviation is calculated, and closed-loop optimization is achieved through adaptive adjustment based on feedback. The formula for calculating the feedback deviation is: ， in, Set a value for the target; This is the actual execution feedback value; The control deviation at the current moment, Closed-loop optimization is performed based on the deviation. The closed-loop control adopts proportional-integral-derivative (PID) regulation, and the calculation formula is as follows: ， in, For feedback adjustment signal; This is a proportionality coefficient, reflecting the intensity of the immediate response to the deviation; These are the integral coefficients used to eliminate steady-state errors; These are the differential coefficients, used to suppress rapid fluctuations in the system; This is the rate of change of deviation, used to predict deviation trends in advance.

7. The intelligent decision-making fusion system for multi-process stage collaboration in wastewater treatment plants according to claim 6, characterized in that, Based on closed-loop optimization, an adaptive adjustment mechanism is introduced to update the dynamic fusion weights and optimization model parameters according to the feedback error trend, forming a short-term fast adjustment loop and a long-term strategy optimization loop, and dynamically updating the fusion weights. AI model hyperparameters and the sensitivity coefficient of the mechanism model To ensure the stability and robustness of control performance under different operating conditions, Based on real-time comprehensive error, with periods ranging from minutes to hours. AI accuracy The fusion weights are dynamically adjusted to achieve a short-term, rapid adjustment loop. The calculation formula is as follows: ， in, The fusion weight at the current moment, with a value range of [0,1]; To integrate the weight update step size and control the rate of weight change; For real-time comprehensive error; The maximum allowable error threshold; This is a metric for the accuracy of AI models within the most recent prediction window. The AI ​​model hyperparameters are updated daily or weekly based on historical operating data and feedback evaluation results to achieve a long-term strategy optimization loop for the AI ​​model hyperparameters. The calculation formula is as follows: ， in, Hyperparameters of the AI ​​model; To optimize the number of iterations in the long term; Update the learning rate for hyperparameters; For the loss function of the AI ​​model, Parameter correction amount of computer theoretical model based on feedback data And update the sensitivity coefficient to achieve adaptive correction of the sensitivity coefficient of the mechanistic model. The calculation formula is as follows: ， in, This is a set of sensitivity coefficients for the mechanistic model, including reaction rate constants and conversion efficiency parameters; Update the step size for sensitivity; This is the sensitivity correction amount calculated based on the feedback data.

Images