Sewage treatment control system and method based on large model

Abstract

The invention belongs to the technical field of sewage treatment, and particularly relates to a sewage treatment control system and method based on a large model, and the method comprises the steps: S1, collecting multi-source heterogeneous data, carrying out the preprocessing of the multi-source heterogeneous data, dividing the data into a training set, a verification set and a test set, the multi-source heterogeneous data comprises process data, equipment operation data, environment and external data, laboratory analysis data and maintenance management data; s2, constructing a sewage treatment control optimization model based on Time-MoE, and carrying out model training; s3, inputting the preprocessed multi-source heterogeneous real-time data into the sewage treatment control optimization model, outputting a process parameter predicted value, and performing real-time regulation and control on the sewage treatment control system according to the process parameter predicted value; meanwhile, water quality data and energy consumption data after regulation and control are collected, and the sewage treatment control optimization model is updated and verified according to the water quality data and the energy consumption data. The problem of low control precision caused by one-sided model decision basis and poor adaptability is solved.

Description

Technical Field

[0001] This invention belongs to the field of wastewater treatment technology, and in particular relates to a wastewater treatment control system and method based on a large model. Background Technology

[0002] With the continuous increase in wastewater discharge, wastewater treatment has become a key link in protecting the water environment and achieving sustainable development. An efficient and stable wastewater treatment control system is of great significance for ensuring that effluent quality absolutely meets standards, for the recycling of water resources, and for reducing energy and material consumption in wastewater treatment.

[0003] Currently, wastewater treatment control technologies typically rely on mathematical models such as activated sludge models (ASM) combined with traditional PID controllers. These models simulate the process by establishing complex biochemical reaction kinetic equations, and then adjust the setpoints accordingly. However, these methods heavily depend on numerous kinetic parameters that are difficult to measure accurately. The modeling process is complex, costly, and lacks universality. Furthermore, once established, the model is relatively fixed and struggles to adapt to dynamic disturbances from internal and external factors such as influent load fluctuations, water temperature changes, and microbial community succession. This leads to decreased control accuracy and even instability in actual operation. Alternatively, traditional predictive models, such as soft measurement or predictive control based on support vector machines (SVM) or shallow neural networks, can reduce the reliance on precise mechanistic knowledge. However, its model structure is relatively simple and its representation ability is limited. It is difficult to effectively process and integrate massive amounts of heterogeneous data from multiple sources such as water quality, equipment, environment, and energy consumption. There is a problem of insufficient data utilization, resulting in one-sided decision-making basis and weak anti-interference ability of the system. Moreover, the model is a static or shallow time series model, which has a weak ability to capture complex nonlinear and long-period time series dependencies. Its prediction accuracy and foresight are limited, which cannot support high-quality predictive decision-making and has poor adaptability. It is easy to cause the control strategy to lag behind the actual operating conditions. Summary of the Invention

[0004] This invention provides a wastewater treatment control system and method based on a large model, which solves the problems of insufficient data utilization leading to biased decision-making basis and poor model adaptability leading to decreased control accuracy.

[0005] This invention provides a basic solution: a wastewater treatment control method based on a large model, specifically including the following steps: S1: Collect multi-source heterogeneous data, preprocess the multi-source heterogeneous data, and divide it into training set, validation set, and test set. The multi-source heterogeneous data includes process data, equipment operation data, environmental and external data, laboratory analysis data, and maintenance management data. S2: Construct a wastewater treatment control optimization model based on Time-MoE and train the model; S3: Input the pre-processed multi-source heterogeneous real-time data into the wastewater treatment control optimization model, output the predicted values ​​of process parameters, and adjust the wastewater treatment control system in real time according to the predicted values ​​of process parameters; at the same time, collect the adjusted water quality data and energy consumption data, and update and verify the wastewater treatment control optimization model according to the water quality data and energy consumption data.

[0006] Preferably, during preprocessing in S1, the collected data is first classified into data categories, and then appropriate preprocessing methods are applied according to the data categories. The data categories include continuous time-series data, state / event data, periodic time-series data, and high-precision discrete data.

[0007] More preferably, S1 also includes labeling the preprocessed data and dividing it into training set, validation set and test set. In the labeling process, the process parameter settings corresponding to the operating conditions where the effluent quality is stable and meets the standards and the energy consumption per unit water volume is within the set range are labeled.

[0008] Preferably, S2 includes S2-1) Construct an optimization model for wastewater treatment control; S2-2) Model training and optimization are carried out in stages; S2-2-1) Self-supervised pre-training is performed using a large amount of preprocessed unlabeled data, where the loss function is the mean squared error; S2-2-2) Supervised fine-tuning of the pre-trained model is performed using labeled data from the training set, where the loss function is smooth L1 loss; (S2-3) After the model training is completed, the final model is evaluated using a test set. The evaluation metrics include root mean square error, mean absolute error, and coefficient of determination.

[0009] More preferably, S2-2 also includes S2-2-3) Using the proximal policy optimization algorithm, the fine-tuned model is treated as an agent and placed in a wastewater treatment simulation environment built based on historical data. The model parameters are adjusted, and a new state and reward are returned. The reward function expression is as follows:

[0010] In the formula, R is, As a weighting factor for water quality compliance, As the weight of energy consumption penalty, This is an indicator function for water quality compliance. This refers to energy consumption power.

[0011] Preferably, the wastewater treatment control optimization model includes an input embedding and alignment layer, a gating network, an expert network cluster, a weighted summation and integration layer, and an output layer. The input embedding and alignment layer is used to transform the preprocessed multi-source heterogeneous data into a high-dimensional vector representation; The gated network is used to analyze the data characteristics of each input data, dynamically calculate a set of sparse weights, and determine the expert network to be activated and its corresponding contribution. The expert network cluster is used to extract in-depth temporal features of expertise from the input data, including water quality time series experts, equipment status experts, and environmental factor experts. The weighted summation and integration layer is used to perform weighted fusion of the output features of multiple activated expert networks according to the sparse weights assigned by the gated network, to generate a comprehensive feature representation. The output layer is used to map comprehensive features into predicted values ​​of process parameters.

[0012] More preferably, the weighted summation and the weighted fusion formula of the integration layer are as follows:

[0013] In the formula, As a comprehensive feature after integration, The weighting coefficients for water quality time series experts represent the importance of water quality time series characteristics for making optimal decisions under a certain operating condition. The weighting coefficients for equipment condition experts represent the importance of equipment condition characteristics under a certain operating condition. The weighting coefficients for environmental factors represent the importance of environmental factor characteristics under a certain operating condition. This is the output feature vector of the water quality time series expert network. This is the output feature vector of the device state expert network. This is the output feature vector of the environmental factors expert network.

[0014] Preferably, S3 includes S3-1) After the real-time collected data is preprocessed, it is input into the trained wastewater treatment control optimization model, and the predicted values ​​of process parameters are output. The wastewater treatment control system generates control commands based on the predicted values ​​of process parameters and sends them to the actuators to drive the actuators to act. S3-2) Continuously collect effluent water quality data and real-time energy consumption data after regulation, and update the model according to the model update strategy based on the effluent water quality data and real-time energy consumption data.

[0015] More preferably, the model update strategy includes a triggering mechanism and a model update mechanism. The trigger mechanism is to initiate a model update when the following two conditions are met: a. The residuals between the model-predicted values ​​and the actual values ​​of effluent chemical oxygen demand, effluent ammonia nitrogen, effluent total nitrogen, effluent total phosphorus, and dissolved oxygen at the end of the aerobic tank are greater than the residual setpoints; b. The variance of the data in the current sliding serial port exceeds the historical variance confidence interval; The model update mechanism uses an exponentially weighted moving average to smoothly update the model parameters, and employs priority empirical replay to assign higher sampling weights to samples with large prediction errors.

[0016] Another basic solution provided by this invention: a wastewater treatment control system based on a large model, comprising: Data acquisition and processing module: used to acquire multi-source heterogeneous data, preprocess the multi-source heterogeneous data, and divide it into training set, validation set, and test set. The multi-source heterogeneous data includes process data, equipment operation data, environmental and external data, laboratory analysis data, and maintenance and management data. Model building and training module: used to build and train a time-MoE-based wastewater treatment control optimization model; Wastewater treatment control system control module: It is used to input pretreated multi-source heterogeneous real-time data into the wastewater treatment control optimization model, output predicted values ​​of process parameters, and perform real-time control of the wastewater treatment control system based on the predicted values ​​of process parameters; at the same time, it collects water quality data and energy consumption data after control, and updates and verifies the wastewater treatment control optimization model based on the water quality data and energy consumption data.

[0017] The principles and advantages of this invention are as follows: 1. By deeply integrating multi-source heterogeneous data such as water quality, equipment, environment, and energy consumption through a large model, the shortcomings of traditional methods in decision-making are overcome, enabling control decisions to be upgraded from experience-based judgments based on single parameters to intelligent optimization based on panoramic data insights. A wastewater treatment control optimization model based on Time-MoE is constructed to predict process setpoints. The wastewater treatment control system adjusts according to the predicted values, transforming the control system from a passive response to an active intervention. This achieves real-time advanced optimization of process parameters and significantly reduces the lag of control strategies.

[0018] 2. By simultaneously using energy consumption and water quality as optimization objectives for the model, it can automatically find and execute the operating strategy with the lowest energy consumption while strictly ensuring that the effluent meets the standards, thereby achieving green, low-carbon, and efficient operation of the wastewater treatment process.

[0019] 3. By combining an expert hybrid architecture with time-series modeling, the model can capture complex nonlinear dynamic changes. Through a dual-dimensional triggering mechanism and incremental update strategy, the system has the ability to continuously learn and adapt, effectively cope with operating condition fluctuations, and maintain high control accuracy over a long period of time. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0021] The following detailed description illustrates the specific implementation method: The specific implementation process is as follows: Example 1 See Figure 1 A wastewater treatment control method based on a large model specifically includes the following steps: S1: Collect multi-source heterogeneous data, preprocess the multi-source heterogeneous data, and divide it into training set, validation set, and test set; S1-1) Collect historical data from the entire process of the wastewater treatment plant’s data acquisition and monitoring system (SCADA) and data platform (such as an industrial IoT platform or time-series database) to form multi-source heterogeneous data; The multi-source heterogeneous data includes process data, equipment operation data, environmental and external data, laboratory analysis data, and maintenance and management data.

[0022] Process data includes flow rate and water quality data for the influent unit, status data for each treatment tank in the biological treatment unit, and water quality data for the effluent unit. Water quality data includes pH, chemical oxygen demand (COD), total suspended solids (TSS), ammonia nitrogen (NH3-N), total nitrogen (TN), total phosphorus (TP), water temperature, and conductivity. Status data for each treatment tank in the biological treatment unit includes dissolved oxygen (DO), oxidation-reduction potential (ORP), mixed liquor suspended solids concentration (MLSS), and sludge age (SRT) for the aeration / aerobic tanks; ORP and nitrate (NO3-N) concentration for the anaerobic / anoxic tanks; and sludge interface meter data and effluent TSS for the secondary sedimentation tank. Data also includes the return sludge volume, excess sludge discharge volume, return sludge concentration, and excess sludge concentration for the sludge treatment unit. Equipment operation data includes the operating status, operating frequency / speed, current, and energy consumption of pumps; the operating status, air volume / pressure, current, and energy consumption of aeration equipment (blowers, surface aerators); the operating status and current of mixers; the opening degree and status of valves; and the working status and alarm status of instruments. Environmental and external data include weather data (such as rainfall, temperature, humidity, and air pressure) and temporal characteristics (hours, weekdays, and holidays); Laboratory analysis data includes manual test data performed periodically (e.g., daily, weekly) on influent, each process unit, and effluent; Maintenance management data includes equipment maintenance records, such as downtime for maintenance and records of parts replaced; as well as chemical dosing records.

[0023] S1-2) Preprocess the collected multi-source heterogeneous data, including noise filtering, data standardization, and feature extraction; During preprocessing, the collected data is first classified, and then appropriate preprocessing methods are applied according to the data category. Specifically, based on the data characteristics of wastewater treatment plants, their data categories include continuous time-series data, state / event-type data, periodic time-series data, and high-precision discrete data; For continuous time-series data, preprocessing methods include anomaly detection and noise filtering, standardization, and feature engineering. Anomaly detection and noise filtering are achieved by using an improved isolated forest algorithm to detect sensor drift faults and by using wavelet denoising technology to eliminate high-frequency interference. Standardization and feature engineering involve using Z-Score standardization to unify all continuous variables into a distribution with a mean of 0 and a standard deviation of 1; and extracting frequency domain features through Fast Fourier Transform.

[0024] Continuous time-series data is continuously generated by sensors, and the values ​​change continuously. Specific data include: influent flow rate, pH, COD, TSS, NH3-N, TP, TN, water temperature, and conductivity of the influent unit; dissolved oxygen, oxidation-reduction potential (ORP), mixed liquor suspended solids concentration (MLSS), and nitrate concentration (NO3-N) of the aeration tank of the biological treatment unit; effluent flow rate, pH, COD, TSS, NH3-N, TP, TN, and turbidity of the effluent unit; and equipment operating parameters such as blower air volume, aerator speed, return pump frequency, and energy consumption.

[0025] For status / event-type data, numerical encoding is performed, such as mapping "start / stop", "on / off", "alarm / normal" and other statuses directly to Boolean values ​​of 1 / 0; and it is converted into time series features. For device status, it is directly added to the feature matrix as a binary feature, and for maintenance events, it needs to be converted into time series features that the model can understand.

[0026] For periodic time-series data, cyclic encoding is used.

[0027] For high-precision discrete data, such as daily or weekly manual test data (COD, TN, TP, etc.), it is aligned with the online sensor data at the corresponding time point.

[0028] S1-3) Label the preprocessed data and divide it into training set, validation set, and test set; Specifically, the process parameter settings corresponding to the operating conditions where the effluent quality is stable and meets the standards and the energy consumption per unit water volume is within the set range are labeled to form a supervised learning sample {input features, output labels}. The energy consumption setting range is set according to expert experience. The process parameter settings include the air volume of the aeration system, the dosage of the carbon source addition system, the return ratio of the mixed liquor and sludge return system, and the dosage of the chemical phosphorus removal system.

[0029] The complete labeled dataset is randomly divided in a ratio of approximately 7:2:1. The training set is used to directly train the model parameters, while the validation set is used to monitor model performance, adjust hyperparameters (such as the learning rate), and perform early stopping to prevent overfitting during training. The validation set is strictly isolated and is used only after all model training is completed to finally evaluate the model's generalization performance.

[0030] S2: Construct a wastewater treatment control optimization model based on Time-MoE and train the model; S2-1) Construct an optimization model for wastewater treatment control; The wastewater treatment control optimization model includes an input embedding and alignment layer, a gated network, an expert network cluster, a weighted summation and integration layer, and an output layer. The input embedding and alignment layer is used to transform the preprocessed multi-source heterogeneous data (including standardized numerical features and categorical features) into a unified, aligned high-dimensional vector representation, providing a standardized input format for subsequent deep processing. The gated network is used to analyze the data characteristics of each input data, dynamically calculate a set of sparse weights to determine which expert networks to activate and allocate corresponding contribution levels, thereby achieving intelligent scheduling of computing resources. Specifically, the feature representations received from the input embedding layer are used to compute a weight distribution through a feedforward neural network or a lightweight Transformer, outputting a sparse weight vector.

[0031] The expert network cluster includes three expert networks for extracting deep temporal features of the area of ​​expertise from the input data; in, Expert Network A is a water quality time series expert that uses a Transformer encoder to capture the complex, long-term time-series dependencies and periodic variation patterns among water quality parameters such as COD, NH3-N, and DO. Expert Network B is a device status expert, containing convolutional layers and fully connected layers, focusing on analyzing the relationship between device operating frequency, energy consumption, state sequence and overall process efficiency. Expert network C, representing environmental factors, employs a recurrent neural network to handle the short-term impacts of external shocks such as inflow rate, temperature, and rainfall on the system.

[0032] The weighted summation and integration layer is used to perform weighted fusion of the output features of multiple activated expert networks according to the sparse weights assigned by the gated network, generating a comprehensive and balanced integrated feature representation to support the final decision. The weighted fusion formula is as follows:

[0033] In the formula, As a comprehensive feature after integration, The weighting coefficients for water quality time series experts represent the importance of water quality time series characteristics for making optimal decisions under a certain operating condition. The weighting coefficients for equipment condition experts represent the importance of equipment condition characteristics under a certain operating condition. The weighting coefficients for environmental factors represent the importance of environmental factor characteristics under a certain operating condition. This is the output feature vector of the water quality time series expert network. This is the output feature vector of the device state expert network. This is the output feature vector of the environmental factors expert network.

[0034] The output layer is used to map comprehensive features into predicted values ​​of process parameters.

[0035] S2-2) Train the model in stages and optimize the model. S2-2-1) Self-supervised pre-training is performed using a large amount of unlabeled data preprocessed in S1-2. 15% of the data points in the input sequence are randomly masked, allowing the model to predict the masked values. The loss function is mean squared error. This stage enables the model to understand the inherent temporal dynamics of parameters such as water quality and equipment. S2-2-2) Supervised fine-tuning training of the pre-trained model is performed using the labeled dataset prepared in S1-3. The predicted values ​​of process parameters output by the model are compared with the labeled optimal process parameter settings. The loss function is smooth L1 loss. At the same time, the training is monitored using a validation set to prevent overfitting. S2-2-3) Using the proximal policy optimization algorithm, the fine-tuned model is treated as an agent and placed in a wastewater treatment simulation environment built based on historical data. The model parameters are adjusted, and a new state and reward are returned. The reward function expression is as follows:

[0036] In the formula, R is, As a weighting factor for water quality compliance, As the weight of energy consumption penalty, This is an indicator function for water quality compliance. This refers to energy consumption power.

[0037] After the model training is completed (S2-3), the final model is evaluated using the test set in S1-3. The evaluation metrics include root mean square error, mean absolute error, and coefficient of determination.

[0038] S3: Input the pre-processed multi-source heterogeneous real-time data into the wastewater treatment control optimization model, output the predicted values ​​of process parameters, and adjust the wastewater treatment control system in real time according to the predicted values ​​of process parameters; at the same time, collect the adjusted water quality data and energy consumption data, and update and verify the wastewater treatment control optimization model according to the water quality data and energy consumption data.

[0039] S3-1) After the real-time collected data is preprocessed, it is input into the trained wastewater treatment control optimization model. The model outputs the predicted values ​​of process parameters. The wastewater treatment control system generates control commands based on the predicted values ​​of process parameters and sends them to the actuators to drive the actuators to act. Among them, the control commands specifically refer to the control commands for the aeration system air volume setpoint, including blower frequency commands, air branch valve opening commands, and blower start / stop commands. The control commands for the carbon source dosage setpoint include the carbon source dosing pump frequency / speed command, the carbon source dosing pump stroke / opening command, and the valve opening command; The control commands for the set value of the reflux ratio of the mixed liquor and sludge reflux system include reflux pump frequency commands and reflux valve opening commands. The control commands for the setpoint of the chemical phosphorus removal system include commands for the phosphorus removal agent dosing pump and valve opening commands. S3-2) Continuously collect effluent water quality data and real-time energy consumption data after regulation, and update the model according to the model update strategy based on the effluent water quality data and real-time energy consumption data; Model update strategies include triggering mechanisms and model update mechanisms. The triggering mechanism is as follows: Model updates are initiated when the following two conditions are met: a. The residuals between the model-predicted values ​​and the actual values ​​of effluent chemical oxygen demand, effluent ammonia nitrogen, effluent total nitrogen, effluent total phosphorus, and dissolved oxygen at the end of the aerobic tank are greater than the residual setpoints; The residual setpoints for each data point are set based on experience, and the residual setpoints are: In the formula, k is a constant. The historical standard deviation; b. The variance of the data in the current sliding serial port exceeds the historical variance confidence interval; The model update mechanism is as follows: the model parameters are updated smoothly using exponentially weighted moving average (EWMA), and priority empirical replay is used to assign higher sampling weights to samples with large prediction errors. in,

[0040] In the formula, β is the attenuation factor. For long-term model parameters, These are the parameters for the online model.

[0041] Example 2 A wastewater treatment control system based on a large model, operating the above method, includes: Data acquisition and processing module: used to acquire multi-source heterogeneous data, preprocess the multi-source heterogeneous data, and divide it into training set, validation set, and test set. The multi-source heterogeneous data includes process data, equipment operation data, environmental and external data, laboratory analysis data, and maintenance and management data. Model building and training module: used to build and train a time-MoE-based wastewater treatment control optimization model; Wastewater treatment control system control module: It is used to input pretreated multi-source heterogeneous real-time data into the wastewater treatment control optimization model, output predicted values ​​of process parameters, and perform real-time control of the wastewater treatment control system based on the predicted values ​​of process parameters; at the same time, it collects water quality data and energy consumption data after control, and updates and verifies the wastewater treatment control optimization model based on the water quality data and energy consumption data.

[0042] The above are merely embodiments of the present invention. Commonly known structures and characteristics are not described in detail here. Those skilled in the art are aware of all common technical knowledge in the field prior to the application date or priority date, are aware of all existing technologies in that field, and have the ability to apply conventional experimental methods prior to that date. Those skilled in the art can, under the guidance of this application, improve and implement this solution in combination with their own capabilities. Some typical known structures or methods should not be obstacles for those skilled in the art to implement this application. It should be noted that those skilled in the art can make several modifications and improvements without departing from the structure of the present invention. These should also be considered within the scope of protection of the present invention, and will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A large model-based sewage treatment control method, characterized by, Specifically comprising the following steps: S1: Collecting multi-source heterogeneous data, and preprocessing the multi-source heterogeneous data, and dividing the multi-source heterogeneous data into a training set, a validation set, and a test set, wherein the multi-source heterogeneous data comprises process data, equipment operation data, environmental and external data, laboratory analysis data, and maintenance management data; S2: Constructing a sewage treatment control optimization model based on Time-MoE, and performing model training; S3: Inputting the preprocessed multi-source heterogeneous real-time data into the sewage treatment control optimization model, outputting process parameter prediction values, and performing real-time regulation and control on the sewage treatment control system according to the process parameter prediction values; meanwhile, collecting water quality data and energy consumption data after the regulation and control, and updating and verifying the sewage treatment control optimization model according to the water quality data and the energy consumption data.

2. The large model-based sewage treatment control method according to claim 1, characterized by: In S1, during preprocessing, the collected data is first subjected to data category judgment, and the data is processed using a suitable preprocessing method according to the data category; the data category includes continuous time series data, state / event data, periodic time series data, and high-precision discrete data.

3. The large model-based sewage treatment control method according to claim 2, characterized by: In S1, the preprocessed data is also annotated and divided into a training set, a validation set, and a test set, wherein, during annotation, the process parameter setpoint corresponding to the working condition in which the effluent quality is stable and meets the standard and the energy consumption per unit water volume is within a set range is annotated.

4. The large model-based sewage treatment control method according to claim 1, characterized by: S2 includes S2-1) constructing a sewage treatment control optimization model; S2-2) performing model training and optimization in stages; S2-2-1) using a large amount of unannotated data after preprocessing for self-supervised pre-training, wherein the loss function is mean square error; S2-2-2) using the annotated data in the training set to perform supervised fine-tuning training on the pre-trained model, wherein the loss function is smooth L1 loss; S2-3) after model training is completed, the final model is evaluated using the test set, and the evaluation indicators include root mean square error, mean absolute error, and determination coefficient.

5. The large model-based sewage treatment control method according to claim 4, characterized by: S2-2 also includes S2-2-3) using a proximal policy optimization algorithm, taking the fine-tuned model as an agent, placing it in a sewage treatment simulation environment constructed based on historical data, adjusting the model parameters, and returning a new state and reward, wherein the reward function expression is as follows: wherein R is, is a weight for water quality compliance, is a weight for energy consumption penalty, is a water quality compliance indicator function, is an energy consumption power.

6. The large model-based sewage treatment control method according to claim 1 or 4, characterized by: The sewage treatment control optimization model comprises an input embedding and alignment layer, a gating network, an expert network cluster, a weighted summation and integration layer, and an output layer, The input embedding and alignment layer is used to convert the preprocessed multi-source heterogeneous data into a high-dimensional vector representation; The gating network is used to analyze the data characteristics of each input data, dynamically calculate a set of sparse weights, and determine the expert network to be activated and the corresponding contribution degree; The expert network cluster is used to extract deep time series features in the specialized field from the input data, including water quality time series experts, equipment state experts, and environmental factor experts; The weighted summation and integration layer is used to weight and fuse the output features of the multiple activated expert networks according to the sparse weights distributed by the gating network, to generate comprehensive feature representation; The output layer is used to map the comprehensive features to process parameter prediction values.

7. The large model-based sewage treatment control method according to claim 6, characterized by: The weighted summation and integration formula of the weighted summation and integration layer is as follows: In the formula, is the comprehensive feature after fusion, is the weight coefficient of the water quality time sequence expert, representing the importance of the water quality time sequence feature under a certain working condition to making the optimal decision, is the weight coefficient of the equipment state expert, representing the importance of the equipment state feature under a certain working condition, is the weight coefficient of the environmental factor expert, representing the importance of the environmental factor feature under a certain working condition, is the output feature vector of the water quality time sequence expert network, is the output feature vector of the equipment state expert network, is the output feature vector of the environmental factor expert network.

8. The large model-based sewage treatment control method according to claim 1, characterized by: S3 includes S3-1) After the real-time collected data is preprocessed, the data is input into the trained sewage treatment control optimization model, and process parameter prediction values are output. The sewage treatment control system generates control instructions according to the process parameter prediction values and sends the control instructions to the actuator to drive the actuator to act; S3-2) Continuously collect the regulated effluent quality data and real-time energy consumption data, and update the model according to the effluent quality data and real-time energy consumption data according to the model update strategy.

9. The large model-based sewage treatment control method according to claim 8, characterized by: The model update strategy includes a trigger mechanism and a model update mechanism, The trigger mechanism is to start model updating when the following two conditions are met: a. The residual error between the model prediction values and the actual values of the effluent chemical oxygen demand, effluent ammonia nitrogen, effluent total nitrogen, effluent total phosphorus, and dissolved oxygen at the end of the aerobic tank is greater than the residual error set value; b. The variance of the current sliding window data exceeds the historical variance confidence interval; The model update mechanism is to update the model parameters using exponential weighted moving average smoothing, and to use priority experience replay to give higher sampling weights to samples with large prediction errors.

10. A large model-based sewage treatment control system, characterized by, It includes: A data acquisition and processing module for acquiring multi-source heterogeneous data, pre-processing the multi-source heterogeneous data, and dividing the multi-source heterogeneous data into training sets, validation sets, and test sets. The multi-source heterogeneous data includes process data, equipment operation data, environmental and external data, laboratory analysis data, and maintenance management data. A model construction and training module for constructing a Time-MoE-based sewage treatment control optimization model and training the model. A sewage treatment control system regulation module for inputting preprocessed multi-source heterogeneous real-time data into the sewage treatment control optimization model, outputting process parameter prediction values, and regulating the sewage treatment control system in real time according to the process parameter prediction values. At the same time, the water quality data and energy consumption data after regulation are collected, and the sewage treatment control optimization model is updated and verified according to the water quality data and energy consumption data.

Images