Intelligent filtering process control method and system for chemical wastewater treatment

Abstract

The invention relates to the technical field of chemical wastewater treatment, and discloses an intelligent filtering process control method and system for chemical wastewater treatment, and the method comprises the steps: collecting the water quality parameters of chemical wastewater and the operation parameters of filtering equipment in real time through a sensor, and carrying out the preprocessing of the collected data, so as to obtain the processed data; the method comprises the following steps: establishing a prediction sub-model of a filtering process by adopting TCN-Attention, establishing a control sub-model by adopting MPC-DRL and embedding the prediction sub-model to obtain a control strategy evaluation model, training the established control strategy evaluation model by utilizing processed data, and performing hyper-parameter optimization by adopting CMA-ES-BNN; inputting the data collected in real time into the trained control strategy evaluation model, and outputting the index parameters of the filtering process and the optimal control strategy through the control strategy evaluation model; according to the invention, real-time monitoring and intelligent control of the chemical wastewater treatment and filtration process are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical wastewater treatment, and in particular to a method and system for controlling an intelligent filtration process of chemical wastewater treatment. Background Art

[0002] Chemical wastewater has complex components and contains a large amount of organic matter, inorganic matter, heavy metal ions and other pollutants. If directly discharged, it will cause serious harm to the environment. Filtration is an important part of the chemical wastewater treatment process. Through filtration, impurities such as suspended matter and colloids in the wastewater can be removed, and the water quality of the wastewater can be improved. However, the traditional chemical wastewater treatment filtration process control method has the following problems: 1. The setting of control parameters mainly relies on manual experience, which lacks scientificity and accuracy, making it difficult to achieve optimal control of the filtration process; 2. The response to changes in wastewater quality is not timely. When the wastewater quality fluctuates, the control parameters of the filtration process cannot be adjusted in time, resulting in unstable filtration effects; 3. The operating efficiency of the filtration equipment is low and the energy consumption is high, which increases the cost of wastewater treatment. Summary of the Invention

[0003] The purpose of the present invention is to solve the above problems and to design a method and system for controlling the intelligent filtration process of chemical wastewater treatment.

[0004] A first aspect of the present invention provides a method for controlling an intelligent filtration process for treating chemical wastewater, the method comprising the following steps: Use sensors to collect water quality parameters of chemical wastewater and operating parameters of filtration equipment in real time, pre-process the collected data, and obtain processed data; First, TCN-Attention is used to establish a prediction sub-model for the filtering process. Then, MPC-DRL is used to establish a control sub-model and embed it into the prediction sub-model to obtain a control strategy evaluation model. The established control strategy evaluation model is trained using the processed data, and CMA-ES-BNN is used for hyperparameter optimization. The real-time collected data is input into the trained control strategy evaluation model, and the control strategy evaluation model outputs the index parameters of the filtering process and the optimal control strategy.

[0005] Optionally, in a first implementation of the first aspect of the present invention, the step of using sensors to collect water quality parameters of chemical wastewater and operating parameters of filtration equipment in real time, and preprocessing the collected data to obtain processed data includes: Using a combination of isolation forest and LSTM autoencoder architecture, we perform hierarchical anomaly detection on the collected data and remove abnormal data; Build a multi-source data fusion architecture based on graph neural networks, and model the data after exception processing as a weighted directed graph; The spatiotemporal graph convolutional network is used as the core operator. The spatial dependencies of the weighted directed graph are captured through the graph convolution operation in the spatial dimension. The temporal dynamic features of the parameter sequence are extracted by combining the convolution operation in the temporal dimension to form a high-order feature representation that integrates the spatiotemporal information to obtain the processed data.

[0006] Optionally, in a second implementation of the first aspect of the present invention, the use of a combined architecture of an isolation forest and an LSTM autoencoder to perform hierarchical anomaly detection on the collected data and eliminate abnormal data includes: The isolation forest algorithm is used to calculate the path length from the data point to the isolated tree in the collected data, locate the discrete outliers that deviate from the normal distribution, and mark them; The encoder of the LSTM autoencoder receives the time series data from the collected data, extracts the temporal features of the data through a multi-layer LSTM network, and maps the high-dimensional input into a low-dimensional latent vector; The decoder restores the low-dimensional latent vector generated by the encoder through the reverse LSTM network to obtain the reconstructed data; The difference between the collected data and the reconstructed data is calculated to obtain the reconstruction error. When the reconstruction error exceeds the preset threshold, it is determined that the sensor is drifting or the device is operating abnormally. Abnormal data is identified and removed through dual processing of isolation forest and LSTM autoencoder.

[0007] Optionally, in a third implementation of the first aspect of the present invention, the step of first using TCN-Attention to establish a prediction sub-model for the filtering process includes: Five convolutional layers are set in the TCN layer, the convolution kernel size of each layer is 3, the expansion factor increases exponentially, and each layer is followed by BatchNorm and ReLU activation functions. A multi-head self-attention mechanism is adopted with four heads. The feature sequence output by the TCN layer is mapped to different subspaces for attention calculation, capturing the complex dependencies within the sequence. The attention output is mapped to the prediction target dimension through the fully connected layer to establish a prediction sub-model.

[0008] Optionally, in a fourth implementation of the first aspect of the present invention, the control sub-model is established using MPC-DRL and embedded in the prediction sub-model to obtain a control strategy evaluation model, including: The prediction domain and control domain of the MPC controller are set, and the Actor-Critic architecture is adopted. The Actor network outputs the control action, and the Critic network evaluates the action value. The MPC controller calls the prediction sub-model in each control cycle to obtain the future state prediction. The DRL intelligent agent learns the optimization strategy of the MPC controller, uses the optimization result of the MPC controller as the supervision signal, trains the DRL network, and establishes the control sub-model.

[0009] Optionally, in a fifth implementation of the first aspect of the present invention, the processing data is used to train the established control strategy evaluation model, and CMA-ES-BNN is used to perform hyperparameter optimization, including: Determine the range of hyperparameters to be optimized and use the covariance matrix adaptive evolution strategy to initialize the search process; Each candidate parameter combination is evaluated and ranked according to their fitness. The core parameters of CMA-ES are updated, the mean of the elite parameters is calculated, the mean vector is updated, the search is guided towards more optimal areas, the covariance matrix is ​​adjusted, and an evolutionary path memory mechanism is introduced to feed back the parameter coupling relationships discovered during the evaluation process to the BNN network. Iterate continuously until the maximum number of iterations is reached, and finally select the parameter combination with the highest fitness as the optimal hyperparameter configuration of the control strategy evaluation model.

[0010] Optionally, in a sixth implementation of the first aspect of the present invention, inputting the real-time collected data into a trained control strategy evaluation model, and outputting the index parameters of the filtering process and the optimal control strategy through the control strategy evaluation model, includes: The real-time collected data is processed and input into the trained control strategy evaluation model. Through the causal convolution and dilated convolution operations of the TCN layer, features at different time scales are extracted. The weight distribution of features at each time step is calculated through the self-attention mechanism, focusing on the temporal pattern of future state prediction to obtain the prediction results. Based on the prediction results, the DRL agent conducts a strategy search in the preset action space and generates candidate control sequences for multiple time steps in the future through Monte Carlo tree search. Each candidate sequence evaluates the multi-objective reward function value in a simulation environment, and the control sequence with the largest cumulative reward is selected as the optimal strategy. Finally, the indicator parameters of the filtering process and the optimal control strategy are output.

[0011] A second aspect of the present invention provides an intelligent filtration process control system for chemical wastewater treatment, the system comprising: The data processing module is used to use sensors to collect water quality parameters of chemical wastewater and operating parameters of filtration equipment in real time, and pre-process the collected data to obtain processed data; The model building module is used to first use TCN-Attention to establish a prediction sub-model for the filtering process, then use MPC-DRL to establish a control sub-model and embed it into the prediction sub-model to obtain a control strategy evaluation model. The established control strategy evaluation model is trained using the processed data, and CMA-ES-BNN is used for hyperparameter optimization. The control evaluation module is used to input the real-time collected data into the trained control strategy evaluation model, and output the index parameters of the filtering process and the optimal control strategy through the control strategy evaluation model.

[0012] The third aspect of the present invention provides an intelligent filtration process control device for chemical wastewater treatment, which includes a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory to enable the intelligent filtration process control device for chemical wastewater treatment to execute the various steps of the intelligent filtration process control method for chemical wastewater treatment as described in any of the above items.

[0013] A fourth aspect of the present invention provides a computer-readable storage medium having instructions stored thereon, which, when executed by a processor, implement the various steps of the intelligent filtration process control method for chemical wastewater treatment as described in any of the above items.

[0014] In the technical solution provided by the present invention, sensors are used to collect water quality parameters of chemical wastewater and operating parameters of filtration equipment in real time, and the collected data are preprocessed to obtain processed data; TCN-Attention is first used to establish a prediction sub-model of the filtration process, and then MPC-DRL is used to establish a control sub-model and embed the prediction sub-model to obtain a control strategy evaluation model, the established control strategy evaluation model is trained using the processed data, and CMA-ES-BNN is used to perform hyperparameter optimization; the real-time collected data is input into the trained control strategy evaluation model, and the control strategy evaluation model outputs the index parameters of the filtration process and the optimal control strategy; the present invention realizes real-time monitoring and intelligent control of the chemical wastewater treatment filtration process, improves the setting accuracy of the control parameters, can respond to changes in wastewater quality in a timely manner, dynamically adjusts the control parameters of the filtration process, and can adapt to the chemical wastewater treatment needs of different water qualities and working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Various other advantages and benefits will become apparent to those skilled in the art by reading the following detailed description of the preferred embodiment.The accompanying drawings are only for the purpose of illustrating the preferred embodiment and are not to be considered as limiting the present invention.

[0016] Figure 1Flowchart of the intelligent filtration process control method for chemical wastewater treatment provided by an embodiment of the present invention; Figure 2 A schematic diagram of the structure of an intelligent filtration process control system for chemical wastewater treatment provided by an embodiment of the present invention; Figure 3 This is a structural diagram of the intelligent filtration process control equipment for chemical wastewater treatment provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The terms "first," "second," "third," "fourth," and so forth (if any) in the description and claims of the present invention and in the accompanying drawings are used to distinguish similar items and are not necessarily used to describe a particular order or precedence. It should be understood that the terms used in this manner are interchangeable where appropriate, so that the embodiments described herein can be implemented in an order other than that shown or described herein. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, apparatus, product, or device that includes a series of steps or elements is not necessarily limited to those steps or elements expressly listed, but may include other steps or elements not expressly listed or inherent to such process, method, product, or device.

[0018] For ease of understanding, the specific process of the embodiment of the present invention is described below. Figure 1 The flow chart of the intelligent filtration process control method for chemical wastewater treatment provided by an embodiment of the present invention specifically includes the following steps: Step 101: Using sensors to collect water quality parameters of chemical wastewater and operating parameters of filtration equipment in real time, and preprocessing the collected data to obtain processed data; In this embodiment, multiple types of sensors are deployed, such as water quality sensors: pH, turbidity, pollutant concentration, dissolved oxygen, etc.; equipment sensors: filtration pressure, flow, motor speed, backwash frequency, etc.; water quality parameters and filtration equipment operating parameters in the chemical wastewater treatment process are collected in real time at preset time intervals, and the data's timestamp, sensor number, parameter name and corresponding value are recorded.

[0019] In this embodiment, a glass electrode pH sensor is used for pH measurement. Its working principle is to use the potential difference between the glass electrode and the reference electrode to change with the pH value of the solution. In the complex composition environment of chemical wastewater, this sensor can accurately and quickly respond to changes in pH value and is suitable for real-time monitoring of the acidity and alkalinity of wastewater. The glass electrode is highly selective for hydrogen ions, and the reference electrode provides a stable reference potential. The combination of the two can accurately measure the pH value of the solution, meeting the strict requirements for real-time pH monitoring in chemical wastewater treatment. The suspended solids concentration is determined using an optical sensor based on the principle of light scattering. When light passes through wastewater containing suspended solids, the suspended solids scatter the light. The suspended solids concentration can be determined by detecting the intensity of the scattered light. This sensor has high sensitivity and accuracy and can effectively detect the content of suspended solids in wastewater. Its working process is as follows: the light source emits light of a specific wavelength, which is collimated by a lens and then irradiated into the wastewater sample. The scattered light generated by the suspended solids is received by the photodetector, and the signal processing circuit converts the optical signal into an electrical signal, which is then used to calculate the suspended solids concentration. This sensor is suitable for wastewater with various turbidity ranges and performs particularly well in scenarios where the suspended solids concentration in chemical wastewater varies greatly. The filtration pressure uses a pressure transmitter, which converts the pressure signal into an electrical signal. By measuring the pressure in the pipeline, it provides an important basis for the control of the filtration process. Its sturdy structure can adapt to the vibration and pressure fluctuation environment when the filtration equipment is running. The pressure transmitter is usually composed of a pressure sensitive element, a signal amplification circuit and an output interface. The pressure sensitive element converts the pressure it feels into a tiny electrical signal change. After amplification by the amplification circuit, it outputs a standard analog signal or digital signal, which is easy to receive and process by the data acquisition system, ensuring accurate monitoring of pressure changes during the filtration process. The flow rate is measured using an electromagnetic flowmeter. According to the law of electromagnetic induction, when a conductive liquid flows in a magnetic field, an induced electromotive force is generated. The flow value can be obtained by measuring the induced electromotive force. This sensor has a wide measurement range and high accuracy, and is suitable for measuring the flow rate of conductive liquids in chemical wastewater. Its main components include a measuring tube, an excitation system, electrodes, and a signal processing unit. The excitation system generates a uniform magnetic field. The conductive liquid flows in the measuring tube and cuts the magnetic lines of force, generating an induced electromotive force on the electrodes. The signal processing unit amplifies, filters, and converts the induced electromotive force to obtain accurate flow data. It is not affected by factors such as fluid density, viscosity, temperature, and pressure, and is suitable for accurate measurement of chemical wastewater flow. The temperature is measured using a thermocouple sensor. Utilizing the principle that thermoelectric potential will be generated when the temperatures at both ends of a closed loop composed of two conductors made of different materials are different, the sensor can stably measure temperature in a high-temperature environment, meeting the temperature monitoring needs of the filtration equipment during operation. The thermocouple sensor consists of a thermocouple, insulating material, a protective sleeve, and a junction box. One end of the thermocouple made of two different materials is welded to form the measuring end, and the other end is the reference end. When the temperature of the measuring end changes, the thermoelectric potential changes accordingly. By measuring the thermoelectric potential and combining it with reference end temperature compensation, the measured temperature value can be accurately obtained. The sensor has the characteristics of a wide measurement range, fast response speed, and high reliability, and is suitable for real-time temperature monitoring in a high-temperature environment of chemical filtration equipment.

[0020] In this embodiment, a combined architecture of isolation forest and LSTM autoencoder is used to perform hierarchical anomaly detection on the collected data and eliminate abnormal data; a multi-source data fusion architecture is constructed based on a graph neural network, and the data after anomaly processing is modeled as a weighted directed graph; a spatiotemporal graph convolutional network is used as the core operator, and the spatial dependency of the weighted directed graph is captured through the graph convolution operation in the spatial dimension, and the temporal dynamic features of the parameter sequence are extracted in combination with the convolution operation in the time dimension to form a high-order feature representation that fused the spatiotemporal information to obtain the processed data.

[0021] In this example, a multi-source data fusion architecture is constructed based on a graph neural network, modeled as a weighted directed graph. Each sensor is represented as a graph node, with node features including real-time collected water quality parameters and equipment operating parameters. Edge weights are determined by both the physical connection between sensors and data correlation, creating a graph structure that reflects the coupling relationship between the spatial layout of sensors and their parameters. Data correlation uses the Pearson correlation coefficient to calculate the dynamic correlation between parameters.

[0022] In this embodiment, the isolation forest algorithm is used to calculate the path length from the data point to the isolation tree in the collected data, locate and mark discrete outliers that deviate from the normal distribution; the encoder of the LSTM autoencoder receives the time series data in the collected data, extracts the temporal features of the data through a multi-layer LSTM network, and maps the high-dimensional input into a low-dimensional latent vector; the decoder restores the low-dimensional latent vector generated by the encoder through a reverse LSTM network to obtain reconstructed data; the difference between the collected data and the reconstructed data is calculated to obtain a reconstruction error. When the reconstruction error exceeds a preset threshold, it is determined to be sensor drift or abnormal equipment operation status; through dual processing of the isolation forest and LSTM autoencoder, abnormal data is determined and eliminated.

[0023] In this embodiment, the encoder of the LSTM autoencoder receives preprocessed time series data, such as a time-varying sequence of filtration pressure and time series data of suspended solids concentration. A multi-layer LSTM network is used to extract the time series features of the data, mapping the high-dimensional input into a low-dimensional latent vector. The latent vector contains the core dynamic pattern of the data under normal operating conditions, such as the pattern of pressure gradually increasing with the use time of the filter material during the filtration process. Based on the latent vector generated by the encoder, the decoder attempts to restore the original input data through a reverse LSTM network or a fully connected layer to obtain reconstructed data of the same dimension as the input data, such as the pressure and concentration values ​​at each moment predicted based on the latent vector. The original input data is compared with the decoded reconstructed data time-step and parameter-by-parameter to calculate the difference between the two. The single-point errors of all time steps and all parameters are aggregated in a preset manner to obtain an overall reconstruction error. When the filtration system operates normally, the encoder and decoder can accurately restore the data after training, and the reconstruction error is small. If sensor drift or progressive equipment anomalies occur, the time series pattern of the input data will deviate from the normal pattern during training, making it difficult for the decoder to accurately restore it, resulting in a significantly increased reconstruction error. By setting a threshold to compare the reconstruction error, such time series pattern anomalies can be identified.

[0024] Step 102: First, TCN-Attention is used to establish a prediction sub-model for the filtering process. Then, MPC-DRL is used to establish a control sub-model and embed it into the prediction sub-model to obtain a control strategy evaluation model. The established control strategy evaluation model is trained using the processed data, and CMA-ES-BNN is used to optimize hyperparameters. In this embodiment, five convolutional layers are set in the TCN layer, the convolution kernel size of each layer is 3, the expansion factor increases exponentially, each layer is followed by BatchNorm and ReLU activation functions, and a multi-head self-attention mechanism is adopted with four heads. The feature sequence output by the TCN layer is mapped to different subspaces for attention calculation, the complex dependencies within the sequence are captured, and the attention output is mapped to the prediction target dimension through the fully connected layer to establish a prediction sub-model.

[0025] In this embodiment, the prediction domain and control domain of the MPC controller are set, and the Actor-Critic architecture is adopted. The Actor network outputs the control action, and the Critic network evaluates the action value. The MPC controller calls the prediction sub-model in each control cycle to obtain the future state prediction. The DRL intelligent agent learns the optimization strategy of the MPC controller, uses the optimization result of the MPC controller as a supervision signal, trains the DRL network, and establishes the control sub-model.

[0026] In this embodiment, the range of hyperparameters to be optimized is determined, and the covariance matrix adaptive evolution strategy is used to initialize the search process; each candidate parameter combination is evaluated, ranked according to the fitness of the candidate parameters, the core parameters of CMA-ES are updated, the mean of the elite parameters is calculated, the mean vector is updated, the search is guided to shift to a more optimal area, the covariance matrix is ​​adjusted, and an evolutionary path memory mechanism is introduced to feed back the parameter coupling relationship discovered during the evaluation process to the BNN network; continuous iteration is performed until the maximum number of iterations is reached, and finally the parameter combination with the highest fitness is selected as the optimal hyperparameter configuration of the control strategy evaluation model.

[0027] In this embodiment, a temporal convolutional network combined with an attention mechanism is used to construct a prediction sub-model, focusing on the long-range dependency modeling of the dynamic characteristics of the filtration process. First, pre-processed multi-dimensional time series data, such as water quality parameters, equipment operating parameters, and historical control parameters, are input into the TCN. Causal convolution and dilated convolution are used to extract multi-scale temporal features layer by layer, such as short-term flow fluctuations, medium-term filter material resistance changes, and long-term water quality trends. Residual connections are used to maintain gradient conduction efficiency. A self-attention module is connected to the TCN output layer to calculate the weight coefficients of features at different time steps, thereby enhancing the influence of key time series patterns on the prediction results, such as the abnormal flow fluctuation signal before the sudden increase in filtration pressure. Finally, the model outputs the predicted values ​​of indicator parameters for multiple time steps in the future, including key process variables such as filtration resistance, suspended solids removal rate, and energy consumption, providing forward-looking input for control strategy generation. The model predictive control (MPC) framework is combined with deep reinforcement learning (DRL). A control sub-model is established and embedded in the prediction sub-model to form a closed-loop optimization system. The state space is first defined as the measured parameters of the filtration process at the current moment, the future state output by the prediction sub-model, and the historical control action sequence. The action space is a set of adjustable equipment operating parameters, such as the filter pressure adjustment range, backwash frequency, and filter media replacement cycle adjustment. Each action corresponds to a discrete or continuous adjustment strategy. A multi-objective reward function is designed, comprehensively considering factors such as water quality treatment effect, equipment operating efficiency, and control smoothness. Water quality effect (such as the suspended solids removal rate reaching the standard), equipment operating efficiency (such as the energy saving rate), and control smoothness (such as the fluctuation of the parameter adjustment range). At each decision step, the control sub-model first obtains a predicted process state trajectory for a period of time in the future through the prediction sub-model. Then, based on the DRL algorithm, it searches in the action space for a control strategy that maximizes the reward function, such as PPO or DDPG. This forms a rolling optimization sequence containing multiple future adjustment plans. Only the first action is executed and the state is updated in the next iteration, achieving real-time optimal control in a dynamic environment. The covariance matrix adaptive evolutionary strategy (CMA-ES) combined with Bayesian neural network (BNN) is used to jointly train and optimize the hyperparameters of the control strategy evaluation model. First, the TCN network structure parameters of the prediction sub-model, the hyperparameters of the DRL agent, and the rolling optimization time domain of the MPC are used as optimization variables to construct a high-dimensional parameter search space. Among them, the TCN network structure parameters such as the number of convolution kernels and the expansion factor sequence, and the hyperparameters of the DRL agent such as the learning rate and the experience replay buffer size are used. BNN is used to quantify the model prediction uncertainty, and random noise is introduced into the training data to simulate sensor measurement errors and working condition fluctuations. Through Monte Carlo d The ropout algorithm estimates the confidence interval of the prediction results and provides CMA-ES with a fitness function that includes uncertainty, such as the control strategy robustness indicator under a 95% confidence level. During the training process, CMA-ES dynamically explores the optimal parameter combination in the search space by iteratively updating the covariance matrix of the parameter distribution. At the same time, the BNN continuously learns the data distribution characteristics, optimizes the generalization ability of the prediction sub-model and the decision robustness of the control sub-model, until the control accuracy and strategy robustness of the model on the validation set reach the preset threshold, where the control accuracy, such as the indicator parameter prediction error, is less than 5%, and the strategy robustness, such as the adjustment delay in the event of a sudden change in water quality, is less than 2 time steps.

[0028] In this embodiment, the TCN layer is configured with 5 convolutional layers, each with a convolution kernel size of 3, and an exponential growth factor to ensure that the receptive field covers a sufficiently long time series. Each layer is followed by a BatchNorm and ReLU activation function; the Attention layer uses a multi-head self-attention mechanism with 4 heads to map the feature sequence output by the TCN to different subspaces for attention calculation, capturing the complex dependencies within the sequence; the attention output is mapped to the prediction target dimension through a fully connected layer, such as predicting the filtrate concentration, pressure and other parameters for the next 5 time steps; weight initialization: the Xavier initialization method is used to ensure that the variance of the input and output of each layer of the network remains consistent to avoid gradient vanishing or exploding; bias initialization: all biases are initialized to 0 to ensure stability in the initial stage of model training; The MPC controller's prediction horizon Np = 10 and its control horizon Nc = 5, predicting the system state for the next 10 time steps and optimizing only the control input for the first five. A weight coefficient, or objective function, was set, taking into account both maximizing filtration efficiency and minimizing energy consumption. Physical constraints included a pressure cap of 1.0 MPa, a flow range of 20-80 m³ / h, and a valve opening range of 0-100%. The DRL agent's state space encompassed physical parameters such as current pressure, flow, and concentration, as well as state predictions for the next three time steps output by the prediction submodel. The action space was defined as a continuous space, encompassing control variables such as feed pump speed and valve opening. The policy network employed an actor-critic architecture, with the actor network outputting control actions and the critic network evaluating their value. The TD3 algorithm was used to improve training stability. The prediction submodel was coupled with the MPC, with the MPC calling the prediction submodel at each control cycle to obtain future state predictions for use in optimizing the control sequence. MPC and DRL were integrated: the DRL agent learned the MPC optimization strategy and used the MPC optimization results as supervisory signals to guide DRL network training.

[0029] Step 103: input the real-time collected data into the trained control strategy evaluation model, and output the index parameters of the filtering process and the optimal control strategy through the control strategy evaluation model.

[0030] In this embodiment, the real-time collected data is processed and input into the trained control strategy evaluation model. The causal convolution and dilated convolution operations of the TCN layer are used to extract features of different time scales. The weight distribution of the features of each time step is calculated through the self-attention mechanism, focusing on the temporal pattern of future state prediction to obtain the prediction results. Based on the prediction results, the DRL agent performs strategy search in the preset action space, and generates candidate control sequences for multiple future time steps through Monte Carlo tree search. Each candidate sequence evaluates the multi-objective reward function value in a simulated environment, and the control sequence with the largest cumulative reward is selected as the optimal strategy. Finally, the indicator parameters of the filtering process and the optimal control strategy are output.

[0031] In this embodiment, the pressure, flow rate, concentration and other parameters of the filtration process are collected in real time by sensors with a sampling frequency of 1 Hz. The normalization parameters saved in the training phase are used to normalize the real-time data. The normalized data of the first 10 time steps at the current moment are input into the prediction sub-model to obtain the state prediction for the next 5 time steps. The current state and the predicted state are input into the MPC-DRL control sub-model to generate the optimal control sequence. The control decision is converted into an actual control signal and sent to the actuator. The key indicator parameters such as filtration rate, filter cake moisture content, energy consumption, etc. are calculated and output in real time. The current state, prediction results and control decisions are displayed in the form of charts through the human-machine interface. When the prediction results or control parameters exceed the safety range, the sound and light alarm system is triggered to remind the operator to intervene.

[0032] See also Figure 2 , a schematic diagram of the structure of an intelligent filtration process control system for chemical wastewater treatment provided by an embodiment of the present invention, the system includes: The data processing module is used to use sensors to collect water quality parameters of chemical wastewater and operating parameters of filtration equipment in real time, and pre-process the collected data to obtain processed data; The model building module is used to first use TCN-Attention to establish a prediction sub-model for the filtering process, then use MPC-DRL to establish a control sub-model and embed it into the prediction sub-model to obtain a control strategy evaluation model. The established control strategy evaluation model is trained using the processed data, and CMA-ES-BNN is used for hyperparameter optimization. The control evaluation module is used to input the real-time collected data into the trained control strategy evaluation model, and output the index parameters of the filtering process and the optimal control strategy through the control strategy evaluation model.

[0033] Figure 3This is a schematic diagram of the structure of an intelligent filtration process control device for chemical wastewater treatment provided by an embodiment of the present invention. The intelligent filtration process control device 600 for chemical wastewater treatment may vary significantly depending on configuration or performance. It may include one or more central processing units (CPUs) 610 (e.g., one or more processors), a memory 620, and one or more storage media 630 (e.g., one or more mass storage devices) storing application programs 633 or data 632. The memory 620 and storage medium 630 may be either transient or persistent storage. The program stored in the storage medium 630 may include one or more modules (not shown), each of which may include a series of instructions and operations within the intelligent filtration process control device 600 for chemical wastewater treatment. Furthermore, the processor 610 may be configured to communicate with the storage medium 630, executing the series of instructions and operations stored in the storage medium 630 on the intelligent filtration process control device 600 to implement the method provided in the above embodiment.

[0034] The chemical wastewater treatment intelligent filtration process control device 600 may further include one or more power supplies 640, one or more wired or wireless network interfaces 650, one or more input and output interfaces 660, and / or one or more operating devices 631, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be understood by those skilled in the art that Figure 3 The structure of the intelligent filtration process control equipment for chemical wastewater treatment shown does not constitute a limitation on the computer equipment provided by the present invention, and may include more or fewer components than shown in the figure, or a combination of certain components, or a different arrangement of components.

[0035] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions. When the instructions are run on a computer, the computer executes the various steps of the intelligent filtration process control method for chemical wastewater treatment provided in the above embodiments.

[0036] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described equipment, devices, and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0037] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the portion that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0038] The above shows and describes the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely preferred examples of the present invention and are not intended to limit the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. The intelligent filtration process control method for chemical wastewater treatment is characterized by: The method comprises the following steps: Use sensors to collect water quality parameters of chemical wastewater and operating parameters of filtration equipment in real time, pre-process the collected data, and obtain processed data; First, TCN-Attention is used to establish a prediction sub-model for the filtering process. Then, MPC-DRL is used to establish a control sub-model and embed it into the prediction sub-model to obtain a control strategy evaluation model. The established control strategy evaluation model is trained using the processed data, and CMA-ES-BNN is used for hyperparameter optimization. The real-time collected data is input into the trained control strategy evaluation model, and the control strategy evaluation model outputs the index parameters of the filtering process and the optimal control strategy.

2. The intelligent filtration process control method for chemical wastewater treatment according to claim 1, characterized in that: The method of using sensors to collect water quality parameters of chemical wastewater and operating parameters of filtration equipment in real time and preprocessing the collected data to obtain processed data includes: Using a combination of isolation forest and LSTM autoencoder architecture, we perform hierarchical anomaly detection on the collected data and remove abnormal data; Build a multi-source data fusion architecture based on graph neural networks, and model the data after exception processing as a weighted directed graph; The spatiotemporal graph convolutional network is used as the core operator. The spatial dependencies of the weighted directed graph are captured through the graph convolution operation in the spatial dimension. The temporal dynamic features of the parameter sequence are extracted by combining the convolution operation in the temporal dimension to form a high-order feature representation that integrates the spatiotemporal information to obtain the processed data.

3. The intelligent filtration process control method for chemical wastewater treatment according to claim 2, characterized in that: The isolation forest and LSTM autoencoder architecture is used to perform hierarchical anomaly detection on the collected data and remove abnormal data, including: The isolation forest algorithm is used to calculate the path length from the data point to the isolated tree in the collected data, locate the discrete outliers that deviate from the normal distribution, and mark them; The encoder of the LSTM autoencoder receives the time series data from the collected data, extracts the temporal features of the data through a multi-layer LSTM network, and maps the high-dimensional input into a low-dimensional latent vector; The decoder restores the low-dimensional latent vector generated by the encoder through the reverse LSTM network to obtain the reconstructed data; The difference between the collected data and the reconstructed data is calculated to obtain the reconstruction error. When the reconstruction error exceeds the preset threshold, it is determined that the sensor is drifting or the device is operating abnormally. Abnormal data is identified and removed through dual processing of isolation forest and LSTM autoencoder.

4. The intelligent filtration process control method for chemical wastewater treatment according to claim 1, characterized in that: The prediction sub-model of the filtering process is first established using TCN-Attention, including: Five convolutional layers are set in the TCN layer, the convolution kernel size of each layer is 3, the expansion factor increases exponentially, and each layer is followed by BatchNorm and ReLU activation functions. A multi-head self-attention mechanism is adopted with four heads. The feature sequence output by the TCN layer is mapped to different subspaces for attention calculation, capturing the complex dependencies within the sequence. The attention output is mapped to the prediction target dimension through the fully connected layer to establish a prediction sub-model.

5. The intelligent filtration process control method for chemical wastewater treatment according to claim 1, characterized in that: The control sub-model is then established using MPC-DRL and embedded into the prediction sub-model to obtain a control strategy evaluation model, including: The prediction domain and control domain of the MPC controller are set, and the Actor-Critic architecture is adopted. The Actor network outputs the control action, and the Critic network evaluates the action value. The MPC controller calls the prediction sub-model in each control cycle to obtain the future state prediction. The DRL intelligent agent learns the optimization strategy of the MPC controller, uses the optimization result of the MPC controller as the supervision signal, trains the DRL network, and establishes the control sub-model.

6. The intelligent filtration process control method for chemical wastewater treatment according to claim 1, characterized in that: The processed data is used to train the established control strategy evaluation model, and CMA-ES-BNN is used to optimize hyperparameters, including: Determine the range of hyperparameters to be optimized and use the covariance matrix adaptive evolution strategy to initialize the search process; Each candidate parameter combination is evaluated and ranked according to their fitness. The core parameters of CMA-ES are updated, the mean of the elite parameters is calculated, the mean vector is updated, the search is guided towards more optimal areas, the covariance matrix is ​​adjusted, and an evolutionary path memory mechanism is introduced to feed back the parameter coupling relationships discovered during the evaluation process to the BNN network. Iterate continuously until the maximum number of iterations is reached, and finally select the parameter combination with the highest fitness as the optimal hyperparameter configuration of the control strategy evaluation model.

7. The intelligent filtration process control method for chemical wastewater treatment according to claim 1, characterized in that: The real-time collected data is input into the trained control strategy evaluation model, and the control strategy evaluation model outputs the index parameters of the filtering process and the optimal control strategy, including: The real-time collected data is processed and input into the trained control strategy evaluation model. Through the causal convolution and dilated convolution operations of the TCN layer, features at different time scales are extracted. The weight distribution of features at each time step is calculated through the self-attention mechanism, focusing on the temporal pattern of future state prediction to obtain the prediction results. Based on the prediction results, the DRL agent conducts a strategy search in the preset action space and generates candidate control sequences for multiple time steps in the future through Monte Carlo tree search. Each candidate sequence evaluates the multi-objective reward function value in a simulation environment, and the control sequence with the largest cumulative reward is selected as the optimal strategy. Finally, the indicator parameters of the filtering process and the optimal control strategy are output.

8. Intelligent filtration process control system for chemical wastewater treatment, characterized by: The system includes: The data processing module is used to use sensors to collect water quality parameters of chemical wastewater and operating parameters of filtration equipment in real time, and pre-process the collected data to obtain processed data; The model building module is used to first use TCN-Attention to establish a prediction sub-model for the filtering process, then use MPC-DRL to establish a control sub-model and embed it into the prediction sub-model to obtain a control strategy evaluation model. The established control strategy evaluation model is trained using the processed data, and CMA-ES-BNN is used for hyperparameter optimization. The control evaluation module is used to input the real-time collected data into the trained control strategy evaluation model, and output the index parameters of the filtering process and the optimal control strategy through the control strategy evaluation model.

9. An intelligent filtration process control device for chemical wastewater treatment, characterized in that: The intelligent filtration process control equipment for chemical wastewater treatment includes a memory and at least one processor, wherein instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the intelligent filtration process control equipment for chemical wastewater treatment executes each step of the intelligent filtration process control method for chemical wastewater treatment as described in any one of claims 1-7.

10. A computer-readable storage medium having instructions stored thereon, characterized in that: When the instructions are executed by the processor, the various steps of the intelligent filtration process control method for chemical wastewater treatment as described in any one of claims 1-7 are implemented.

Images