Fluid purification method and system based on lightweight transformer and guide plate collaborative regulation

Abstract

The application discloses a fluid purification method and system based on light-weighted Transform and guide plate collaborative regulation, the method comprising: dividing multiple independent monitoring and control areas on the cross section of the fluid medium channel inlet, and collecting the multi-physical field parameters of each area in real time; preprocessing the multi-physical field parameters, and constructing a space-time feature tensor according to the area dimension and the time dimension; inputting the space-time feature tensor into a light-weighted Transform model, and outputting a three-element collaborative control instruction; adjusting the purification execution mechanism and the dynamic guide plate execution mechanism corresponding to each area according to the three-element collaborative control instruction; collecting the total pollutant concentration of the fluid medium channel end outlet in real time, and when the total pollutant concentration of the end outlet continuously deviates from the set target value, applying a global correction amount to the pollutant weight coefficient to form a closed-loop collaborative control. The problems of uneven flow field distribution, excessive dosing, denitration efficiency fluctuation and ammonia escape in the prior art are solved.

Description

Technical Field

[0001] This invention relates to the field of industrial fluid flow and pollutant synergistic control technology, specifically a fluid purification method and system based on the synergistic regulation of a lightweight Transformer and a baffle plate. Background Technology

[0002] With increasingly stringent environmental protection requirements and the transformation of the energy structure, my country's coal-fired power units are facing the challenge of normalized "deep peak shaving" operations. Against this backdrop, unit loads need to be frequently and rapidly adjusted between as low as 30% of rated load and even lower than full load. This drastic change in operating conditions leads to significant spatiotemporal unsteady fluctuations in key parameters such as flue gas flow, temperature, and pollutant concentration, severely testing the traditional independent control mode of the "three islands" (desulfurization, denitrification, and dust removal systems).

[0003] Specifically, existing technologies typically design and control desulfurization, denitrification, and dust removal systems separately, with each subsystem independently adjusted based on local parameters (e.g., the denitrification system only relies on inlet NOx concentration and flow rate). This "three-island separation" control strategy reveals inherent flaws under wide load and rapid changing operating conditions, such as deep peak shaving.

[0004] The flow field organization is disconnected from the reaction process: static guide plates or fixed flow field designs cannot adapt to changes in flow velocity and direction caused by large load changes, resulting in uneven distribution of flue gas flow field, which in turn affects the ammonia mixing effect on the surface of the denitrification catalyst and the dust collection efficiency of the dust collector.

[0005] Lagging control response and lack of coordination: Conventional zoned ammonia injection control is based only on the NOx concentration feedback at the current point and fails to anticipate the changing trends of temperature field and flow velocity field (for example, temperature fluctuations will significantly affect the denitrification reaction rate and ammonia slip), resulting in excessive or insufficient ammonia injection, and failing to achieve the optimal balance between denitrification efficiency and ammonia slip.

[0006] The overall energy efficiency and stability of the system are reduced: the lack of collaborative optimization based on multi-physics coupling among the three islands makes it easy to generate control oscillations when dealing with rapid load changes. This not only increases the ineffective waste of reducing agents (such as ammonia) and energy consumption, but may also exacerbate the risk of equipment wear or blockage due to flow field distortion.

[0007] In the biological treatment tank in a municipal wastewater treatment plant's influent distribution and carbon source addition system, the liquid purification process involves municipal wastewater undergoing pretreatment via coarse and fine screens and a grit chamber before entering the anaerobic-anoxic-aerobic (AAO) biological treatment tank. At the end of the main influent channel, multiple parallel biological treatment tank corridors distribute the influent and add carbon sources (such as sodium acetate). Currently, the following problems exist: Uneven flow distribution: Due to differences in pipeline layout and resistance, the inflow rate of each corridor can vary by ±20%, resulting in inconsistent hydraulic residence time.

[0008] The water quality fluctuates dramatically: the concentrations of chemical oxygen demand (COD), ammonia nitrogen (NH3-N), and total nitrogen (TN) in the influent vary greatly with time of day (morning and evening peaks) and season.

[0009] Inefficient carbon source addition: Traditional methods rely on online instrument signals from the main inlet pipe to allocate carbon sources in a fixed ratio. This fails to detect the actual hydraulic and water quality differences in each corridor, resulting in insufficient carbon sources and low denitrification efficiency in some corridors, while excessive carbon sources in others lead to waste and potential secondary pollution. Summary of the Invention

[0010] The technical problem to be solved by this invention is to solve the problems of uneven flow, excessive dosing, and low reaction efficiency caused by flow field disturbance in gas and liquid flow systems.

[0011] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A fluid purification method based on the coordinated control of a lightweight Transformer and a baffle plate, adaptable to gas and liquid flow systems, includes: Multiple independent monitoring and control zones are divided on the cross-section of the fluid medium channel inlet, and multi-physics parameters of each zone are collected in real time. After preprocessing the multiphysics parameters, a spatiotemporal feature tensor is constructed according to the region dimension and the time dimension. The spatiotemporal feature tensor is input into a lightweight Transformer model, and the output is a ternary cooperative control command; According to the three-element coordinated control command, the corresponding purification actuators and dynamic guide vane actuators in each area are adjusted; The total pollutant concentration at the end outlet of the fluid medium channel is collected in real time. When the total pollutant concentration at the end outlet continues to deviate from the set target value, a global correction is applied to the pollutant weighting coefficient to form a closed-loop collaborative control.

[0012] Furthermore, the lightweight Transformer model architecture includes: a feature embedding layer, a position encoding layer, a dual-head self-attention layer, a feedforward neural network layer, and a parameter generation layer; In the gas flow system, the feature embedding and position encoding layer maps the learnable embedding matrix of the spatiotemporal feature tensor into a high-dimensional feature vector and adds spatiotemporal position encoding to obtain a region-time embedding representation. A spatial coupling head and a temporal evolution head are set in the dual-head attention layer. Based on high-dimensional feature vectors, the spatial coupling head calculates the coupling relationship between flow velocity and temperature between regions and outputs spatial coupling feature vectors. The temporal evolution head analyzes the changing trend of pollutant concentration over time and outputs temporal evolution feature vectors. The outputs of the dual-head attention layer are then concatenated. The concatenated feature vector is input into the feedforward neural network layer for nonlinear transformation, and then mapped to the final three-element collaborative control command through the parameter generation layer, including: temperature weight coefficient, pollutant weight coefficient and dynamic guide vane opening angle.

[0013] Furthermore, the opening degree of the ammonia injection regulating valve in the purification actuator is determined by the following formula: ; In the formula, This is the final ammonia injection rate. This is the baseline ammonia injection rate calculated based on the inlet NOx concentration and flue gas flow rate. , To set the weighting coefficients for the project, Temperature weighting coefficient, This represents the pollutant weighting coefficient.

[0014] Furthermore, according to the three-element coordinated control command, the corresponding purification actuators and dynamic guide vane actuators in each area are adjusted to perform execution control and safety constraints: when the pollutant concentration in any area is lower than the preset safety threshold, the pollutant weight coefficient of that area is forcibly set to zero, and the opening angle of the corresponding dynamic guide vane is adjusted to the fully open position to avoid over-spraying and maintain flow field stability.

[0015] Furthermore, the dynamic deflector opening angle is determined in the three-dimensional collaborative control command of the deflector opening and is subject to dual strategy constraints, including: airflow uniformity main control logic, concentration protection forced coverage logic, and execution priority logic; The main control logic for flow uniformity is as follows: ; In the formula, For the first The dynamic deflector opening angle of the area This is expressed as the average cross-sectional velocity at the inlet of the fluid medium channel. For the first Real-time flow rate of the region; The logic for forced coverage of concentration protection is as follows: In the formula, For logical statements, represented as if So ; For the first ammonia concentration in the area Indicates as forced full opening Deflectors in the area, This indicates that the forced shutdown is the first one. Ammonia spraying in the area For the first Pollutant weighting coefficients for the region; The execution priority logic is as follows: In the first priority, the condition is that the lightweight Transformer model inference is successful, and the action is that the deflector uses the output dynamic deflector opening angle; in the second priority, the condition is... The action is forced. In the third priority, if the sensor fails, the action is to maintain the previous valid value; in the fourth priority, if the system stops suddenly, the action is to lock the current position.

[0016] Furthermore, when the total pollutant concentration at the terminal outlet continuously deviates from the set target value, a global correction is applied to the pollutant weighting coefficients to form a closed-loop coordinated control. The correction logic is as follows: if... If this continues for a period of time, a global correction will be applied to the pollutant weighting coefficients. ;like Then the global correction applied to the pollutant weight coefficients is a fixed value; where, the global correction... It can be obtained through the following formula: ; In the formula, The average NOx concentration. This is a function that takes the minimum value.

[0017] Furthermore, in liquid flow systems, the lightweight Transformer model differs from that in gas flow systems, except for the different multiphysics parameters collected, in the dual-head attention layer. The spatial coupling head learns the coupling relationship between flow velocity, suspended solids concentration and baffle opening, while the time evolution head learns the changing trends of ammonia nitrogen and COD concentrations. The ternary collaborative control commands output by the lightweight Transformer model are the temperature-viscosity weighting coefficient, the ammonia nitrogen / COD pollutant weighting coefficient, and the dynamic baffle opening angle.

[0018] Furthermore, purify the implementing agency Carbon source addition; wherein, the final carbon source addition amount for each zone is determined by the following formula: ; In the formula, This refers to the final carbon source addition amount. Basic carbon source addition amount , These are fixed weights in the liquid system, determined based on actual process adjustments. This is the temperature-viscosity weighting coefficient. This represents the ammonia nitrogen / COD pollutant weighting coefficient.

[0019] Furthermore, the control logic of the guide vane in the liquid flow system includes: main control logic: adjusting the opening angle of the guide vane based on the dynamic output of the lightweight Transformer model; The mandatory protection logic includes: Anti-precipitation logic: If the first Regional flow velocity If the speed remains below 0.3 m / s, then the area will be forced to... ←85° and trigger an alarm; Low load protection: If the first Regional ammonia nitrogen concentration If the concentration remains consistently below 2 mg / L, the nitrogen removal load of the corridor is considered extremely low, and forced nitrogen removal is required. ←0, stop adding carbon source and set ←90° to maintain basic flow rate and prevent siltation.

[0020] This invention provides a system for applying the fluid purification method based on the synergistic control of a lightweight Transformer and a baffle plate described above, comprising: The physical parameter module is used to divide the cross-section of the fluid medium channel inlet into multiple independent monitoring and control areas, and to collect multi-physical field parameters of each area in real time. The feature module is used to preprocess the multiphysics parameters and construct the spatiotemporal feature tensor according to the region dimension and the time dimension. The inference module is used to input spatiotemporal feature tensors into a lightweight Transformer model and output ternary cooperative control commands. The adjustment module is used to adjust the purification actuators and dynamic guide vane actuators corresponding to each area according to the three-element coordinated control command; The feedback correction module is used to collect the total pollutant concentration at the end outlet of the fluid medium channel in real time. When the total pollutant concentration at the end outlet continues to deviate from the set target value, a global correction amount is applied to the pollutant weight coefficient to form a closed-loop collaborative control.

[0021] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves joint regulation of ammonia injection quantity and airflow distribution by performing spatiotemporal coupling modeling of multi-physics field parameters within the flue, and outputting coordinated control commands such as temperature weighting coefficient, pollutant weighting coefficient, and dynamic guide vane opening angle. This solves problems in existing technologies such as uneven flow field distribution, excessive chemical dosing, fluctuating denitrification efficiency, and high ammonia slip. This invention successfully solves the core problems that have long existed in liquid flow and reaction systems, such as uneven distribution, crude addition, and delayed response. It realizes the technological transfer and expansion from "gas flue gas purification" to "liquid fluid control", demonstrating its strong universality and engineering application value.

[0022] This invention is applicable to the intelligent control of flow organization and pollutant purification in gas (such as flue gas from coal-fired power plants, steel sintering, cement kilns, and waste incineration) and liquid (such as chemical processes, wastewater treatment, and cooling circulating water) flow systems. Unlike traditional methods, this invention achieves feedforward-feedback coordinated control of multiple physics fields (flow velocity, temperature, pollutant concentration, static pressure, and particulate matter concentration) through a lightweight Transformer model. This solves the flow field distortion problem caused by wet flue gas or high-viscosity fluids, and the model parameters are ≤64 KB, making it suitable for local deployment with industrial PLCs. Attached Figure Description

[0023] Figure 1 This is a flowchart of a fluid purification method based on the coordinated control of a lightweight Transformer and a baffle plate, according to an embodiment of the present invention.

[0024] Figure 2 This is a process flow diagram of the gas flow system in an embodiment of the present invention.

[0025] Figure 3 This is a schematic diagram showing the arrangement of the sensor and the flow guide plate in the area according to an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the lightweight Transformer model structure according to an embodiment of the present invention.

[0027] Figure 5 This is a flowchart illustrating the control logic of the gas flow system according to an embodiment of the present invention.

[0028] Figure 6 This is a schematic diagram of the guide vane angle control according to an embodiment of the present invention. Detailed Implementation

[0029] To facilitate understanding of the technical solution of the present invention by those skilled in the art, the technical solution of the present invention will now be further described in conjunction with the accompanying drawings.

[0030] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0031] Example 1, Application in gas flow systems - Intelligent control method for flue gas purification; Please see Figure 1 As shown, the present invention provides a fluid purification method based on the coordinated control of a lightweight Transformer and a baffle plate.

[0032] In this embodiment, hardware and system deployment are performed before executing the fluid purification method.

[0033] Please see Figure 2 and Figure 3 As shown, in this embodiment, the fluid purification system of the present invention relies on a highly reliable and highly synchronous multi-physics sensing network. This network consists of four parts: a regionalized sensor array, an edge signal conditioning unit, an industrial communication bus, and an anti-interference installation structure, ensuring stable operation in the high-temperature, high-dust, and high-vibration environment of coal-fired flue gas. The gas flow system, arranged according to the flue gas flow direction, includes an electrostatic precipitator 100, a desulfurization tower 200, an SCR reactor 300, and a PLC control unit 400. Furthermore, the SCR reactor is equipped with a multi-physics sensor 101 and a dynamic guide plate 102.

[0034] 1. Flue area division and measuring point layout Regional division: On the horizontal flue cross-section 1-2 meters in front of the SCR reactor 300 inlet, it is evenly divided into 6 independent monitoring and control zones (referred to as Zone 1 to Zone 6) along the width direction. Each zone is approximately 0.8-1.2 meters wide (depending on the total width of the flue), covering the entire flow cross section.

[0035] Measurement point location: An integrated measuring point unit is set up in each region, located 5 cm downstream of the geometric center of the region, to avoid the influence of the wall boundary layer and to truly reflect the state of the mainstream area.

[0036] 2. Sensor Selection and Technical Parameters Each measuring point unit integrates the following five types of sensors, all led out via high-temperature resistant armored cables, and have a built-in self-cleaning / anti-clogging structure, as shown in Table 1: Table 1 Sensor Selection and Technical Parameters

[0037] 3. Conditioning and Synchronous Data Collection Edge processing unit: Each area is equipped with an IP66-rated signal conditioning box, containing: 3.1) 4–20mA / 0–10V signal isolation module (suppressing common-mode interference); 3.2) 16-bit high-precision ADC (sampling rate 1kHz, actual usage 1Hz); 3.3) Temperature compensation circuit (for the cold junction of thermocouples); 3.4) Backflush control solenoid valve drive circuit.

[0038] Time synchronization mechanism: All six acquisition cards are synchronized via the IEEE 1588 PTP precision clock protocol to ensure that the timestamp error of data in each area is less than 1 ms, which meets the requirements of spatiotemporal modeling.

[0039] In this embodiment, data preprocessing (completed in the edge PLC) includes: 1. Moving average filtering (window = 5 s) → suppresses high-frequency noise; 2. Outlier removal (3σ principle) → Prevents data corruption caused by momentary sensor failure; 3. Feature normalization → mapping to the [0,1] interval to adapt to AI model input.

[0040] 4. On-site and system integration: Fieldbus: PROFINET industrial Ethernet is used to connect the conditioning boxes of the 6 zones to the main control PLC (such as Siemens S7-1515F), with a communication cycle of ≤10 ms.

[0041] Data format: Upload one structure per cycle; Redundancy design: Critical signals (such as NOx and flow rate) simultaneously output 4–20mA analog signals to the DCS system as backup inputs in case of AI failure. In the event of communication interruption, the PLC retains the last valid value and triggers an alarm.

[0042] 5. Designed to withstand harsh environments, see Table 2 for details.

[0043] Table 2 Design for Harsh Environments

[0044] 6. Calibration and Maintenance Online calibration: NOx analyzer: automatic zero / range calibration every 24 hours (with N2 and standard gas); Flow sensor: zero drift correction using the "zero wind speed" state during shutdown.

[0045] Maintenance interface: All sensors can be pulled out from the maintenance platform outside the flue without stopping the machine or entering the flue; the conditioning box panel has LED status indicator lights (power / communication / fault).

[0046] 7. Data integration with AI systems: The preprocessed data is organized and fed into the lightweight Transformer model as follows: Time dimension: Cache the most recent 10 periods (10 seconds of data); Spatial dimensions: arranged in order from region 1 to 6; Final input: A floating-point array of 6 (regions) x 10 (time steps) x 5 (features) is formed and passed to the AI ​​inference engine via shared memory.

[0047] S10 divides the cross-section of the fluid medium channel inlet into multiple independent monitoring and control zones, and collects multi-physics parameters of each zone in real time.

[0048] In this embodiment, the area division and data acquisition are as follows: multiple independent monitoring and control areas are divided on the cross-section of the flue gas purification reactor inlet flue, and multi-physical field parameters of each area are collected in real time. The multi-physical field parameters include at least flue gas velocity, temperature, pollutant concentration, static pressure and particulate matter concentration.

[0049] In this embodiment, the system is deployed as follows: Process flow: Electrostatic precipitator → Limestone-gypsum wet desulfurization → SCR denitrification; SCR inlet flue gas temperature: 50℃, humidity >10%, containing trace droplets; Flue cross-section: 6m (width) × 4m (height); Divide the area into 6 regions (equal along the width direction); Deployment per region: Micro differential pressure Pitot tube (flow rate v); PT10 thermal resistor (temperature T); in-situ laser NOx analyzer (C, response <1s); static pressure sensor (P); light scattering dust meter (PM); all signals are connected to Siemens S7-1515FPLC, sampling period 1s.

[0050] S20 preprocesses the multiphysics parameters and constructs a spatiotemporal feature tensor according to the region dimension and the time dimension.

[0051] In this embodiment, data preprocessing and feature construction are performed: the collected multiphysics parameters are filtered, outlier removed and normalized, and spatiotemporal feature tensors are constructed according to the regional and time dimensions as input data for the lightweight Transformer model.

[0052] S30 inputs the spatiotemporal feature tensor into the lightweight Transformer model and outputs a three-dimensional collaborative control command, including: temperature weight coefficient, pollutant weight coefficient, and dynamic deflector opening angle.

[0053] In this embodiment, the lightweight Transformer model state representation is generated by inputting a spatiotemporal feature tensor into a lightweight Transformer model deployed in an industrial programmable logic controller (PLC), and generating a region-time state representation through feature embedding and spatiotemporal location encoding.

[0054] The lightweight Transformer model includes at least two self-attention heads, where: The primary focus is on modeling the spatial coupling relationships between different regions, with an emphasis on extracting the influence characteristics of uneven flow velocity and temperature distribution on the flow field organization.

[0055] The second attention focus is used to model the evolution characteristics of pollutant concentration over time and extract information on the trend of pollutant concentration changes.

[0056] Through the attention mechanism described above, the context state representation vector corresponding to each region is obtained.

[0057] In this embodiment, the cooperative control parameters are generated as follows: the context state representation vector is input to the parameter generation layer, and the parameter generation layer outputs a ternary cooperative control command, including a temperature weighting coefficient. Pollutant weighting coefficient And the opening angle α of the dynamic guide vane.

[0058] Among them, the temperature weighting coefficient Pollutant weighting coefficient The intermediate modulation parameter reflects the impact of temperature distribution and pollutant concentration changes on dosing requirements under current operating conditions, and is not a direct measure of equipment operation. The dynamic deflector opening angle α is generated by weighted fusion of the output characteristics of the spatially coupled attention head and the temporally evolved attention head, and is used to comprehensively reflect the control requirements of flow velocity, temperature, and pollutant distribution on airflow organization.

[0059] In this embodiment, 1. Input data Constructing the spatiotemporal feature tensor ,in, For the spatiotemporal feature tensor based on multiphysics, For real numbers, Represented as a spatial dimension, it refers to the number of independent monitoring areas (area 1 to area 6) divided by the cross-section of the flue. It is represented by the time dimension, the length of the sliding time window, which is the historical data of the most recent 10 seconds (sampling frequency 1Hz). It is represented as a feature dimension, consisting of 5 types of physical quantities collected in each region at each time step.

[0060] Historical data for 6 regions × 5 physical quantities (v, T, C, P, PM) × 10 seconds; data are filtered by moving average and normalized by Min-Max.

[0061] 2. Model Structure Input the spatiotemporal feature tensor into the encoder of the lightweight Transformer model and execute: 2.1 Feature Embedding and Spatiotemporal Location Encoding: In the formula, For the first Each region at time The embedding vector is typically 32-dimensional. The learnable feature embedding matrix maps the 5-dimensional original physical quantities to a high-dimensional semantic space. For the first Each region at time The flue gas velocity (m / s). For the first Each region at time The temperature (°C). For the first Each region at time NOx concentration (ppm) For the first Each region at time The static pressure (Pa). For the first Each region at time The particulate matter concentration (mg / m³). It is a spatiotemporal location encoding vector containing a region index. and time step This information enables the lightweight Transformer model to distinguish states such as "Region 1 current" and "Region 2 past". (Bold text) This is the matrix transpose.

[0062] 2.2, Two-headed self-attention calculation ; In the formula, This is a comprehensive feature representation extracted and fused through dual attention mechanisms in both spatial and temporal dimensions. For the concatenation operation, the outputs of the two attention heads are merged. The primary focus, also known as the spatial coupling focus, concentrates on the spatial coupling relationship between regions (such as the influence of high-speed regions on low-speed regions). The second attention head, also known as the time evolution head, focuses on the dynamics of time evolution (such as the upward trend of NOx concentration). To output the projection matrix, the concatenated vector is reduced back to its original dimension. The specific calculation of the two-headed self-attention mechanism is shown in Table 3.

[0063] Table 3 Examples of Two-Head Self-Attention Calculation

[0064] The input of the lightweight Transformer model is The output consists of three control parameters. Among them, the temperature weighting coefficient Dominated by the spatial coupling head, it reflects the sensitivity of temperature to the amount of ammonia injected. For example, when the temperature in region 1 is high (T=55℃) and the flow rate is low (v=5m / s), (Increase ammonia injection rate). Pollutant weighting coefficient. The time evolution head is dominant, reflecting the trend of NOx concentration changes. For example, when the NOx concentration in region 2 increases (dC / dt = +2 ppm / min), (Increase ammonia injection rate). Deflector opening angle. It is determined by both the spatial coupling head and the temporal evolution head. For example: high-speed region (v>15m / s) → =30°; Low-speed zone (v<8m / s) → =90°.

[0065] In this embodiment, more specifically, the lightweight Transformer model architecture mainly includes a feature embedding layer, a position encoding layer, a dual-head self-attention layer, a feedforward neural network layer, and a parameter generation layer. The specific process is as follows: Step 1: Feature Embedding and Location Encoding Input tensor Through learnable embedding matrices Mapped to high-dimensional feature vectors And add spatiotemporal location coding The region-time embedding representation is obtained: ; Step 2, dual-head self-attention computation, the embedding representation is processed by two parallel self-attention heads: (Spatial Coupling Head): Calculates the coupling relationship between flow velocity and temperature between regions and outputs a spatial coupling feature vector. ; (Time Evolution Head): Analyzes the trend of pollutant concentration changes over time and outputs a time evolution feature vector. The outputs of the two heads are concatenated. ; This is the concatenated vector of the outputs of the two-headed self-attention method.

[0066] Step 3: Feedforward Neural Network Layer and Parameter Generation Layer concatenated feature vectors The input is fed into a feedforward neural network layer for nonlinear transformation, and then passed through a parameter generation layer (a fully connected layer, mapped to three control parameters): ; In the formula, For the first The temperature weighting coefficient for the region is used to adjust the sensitivity of the ammonia injection rate to temperature. For the first The regional pollutant weighting coefficient is used to adjust the response of ammonia injection rate to NOx concentration. For the first The dynamic deflector opening angle of the region is used to actively balance the airflow. The output layer weight matrix maps the hidden states to three control parameters. The first output of the Transformer encoder Region context representation vector, This is the output layer bias vector.

[0067] Among them, constraints are imposed on the control parameters. , , .

[0068] In this embodiment, the ternary instruction generation mechanism is explained as follows: Although this model only has two attention heads, through the linear transformation of the parameter generation layer, the fused features of the dual-head output are mapped to three independent control parameters: temperature weighting coefficient. The flow velocity-temperature coupling characteristics extracted by the spatial coupling head are the main factors, with pollutant weighting coefficients as the primary factor. The concentration trend characteristics extracted by the time evolution head are the main factors, and the baffle opening angle is the primary factor. The design is jointly determined by the output characteristics of two attention heads, comprehensively reflecting the synergistic control requirements of velocity distribution, temperature field, and pollutant concentration field on airflow organization. This design maintains the model's lightweight nature (≤64KB of parameters) while achieving multi-physics collaborative decision-making, possessing clear interpretability and engineering adjustability.

[0069] In this embodiment, training and deployment: Training data: 6 months of historical data (1.5 million samples); Tags: Generated by expert rules (High-speed zone → α=30°, NOx<15ppm →) =0, α=90°); Loss function: Weighted MSE (λ²=0.5, emphasis) (accuracy) Deployment: PyTorch → ONNX → C code → S7-1515FPLC; Inference latency: <45ms.

[0070] S40 adjusts the purification actuators and dynamic guide vane actuators corresponding to each area according to the three-element coordinated control command.

[0071] In this embodiment, control and safety constraints are implemented: based on the three-dimensional coordinated control command, the purification actuators and dynamic guide vane actuators corresponding to each region are adjusted. When the pollutant concentration in any region is lower than a preset safety threshold, the pollutant weight coefficient of that region is forcibly set to zero, and the opening angle of the corresponding dynamic guide vane is adjusted to the fully open position to avoid over-spraying and maintain flow field stability.

[0072] In this embodiment, the purification actuator is specifically an ammonia injection actuator, wherein the opening degree of the ammonia injection regulating valve in the ammonia injection actuator is determined by the following formula: ; In the formula, This is the final ammonia injection rate. This is the baseline ammonia injection rate calculated based on the inlet NOx concentration and flue gas flow rate. , By setting weighting coefficients for the project and using data from the PLC configuration table as fixed constants, this design retains the stability of the original control system while introducing the dynamic optimization capability driven by the lightweight Transformer model.

[0073] In this embodiment, the intelligent control of the dynamic guide vane includes: The air deflector mechanical structure includes an independently adjustable air deflector assembly installed in each area, comprising: Louver blades: High-temperature resistant alloy steel (1Cr18Ni9Ti), dimensions 300mm×50mm×3mm; Rotating shaft: Supported by heat-resistant bearings, allowing continuous rotation from 0° to 90°; Servo drive unit: IP67 protection rated DC servo motor (MAXONEC45), supports Modbus TCP; Safety mechanisms: hard limit switch (0° / 95°), electromagnetic brake (power failure self-locking), jamming detection (current surge alarm).

[0074] Deflector opening Outputted by a lightweight Transformer model and subject to dual policy constraints, specifically including airflow uniformity control logic, concentration protection mandatory coverage logic, and execution priority logic. The airflow uniformity control logic includes: ; in, Expressed as the cross-sectional average velocity, and, It is a high-speed zone. This is the low-speed region. At the same time, it can be seen that simple calculations based solely on flow velocity cannot meet the requirements of multi-factor coupled control, for the following reasons: Limitations of basic physical calculations: They only consider flow velocity and neglect key factors such as temperature, NOx concentration, and particulate matter concentration. For example, when the flow velocity in region 1 is high (v=18m / s) but the temperature is low (T=45℃), the baffle opening needs to be reduced; when the temperature is high (T=55℃), the opening needs to be increased. It can be seen that the lightweight Transformer model is necessary to control the baffle opening angle adjustment, as summarized in Table 4.

[0075] Table 4 Comparison of the methods for deflector opening between the present invention and traditional methods

[0076] Verification using a real-world case: In a certain region (Region 3), the flow velocity was high (v=17m / s) but the NOx concentration was low (C=12ppm). Traditional method: α=30° (only based on high flow velocity), In this invention, α=25° (because the NOx concentration is low, the opening of the guide vane needs to be reduced to avoid over-spraying). Result: The ammonia escape rate in this region decreased from 1.2ppm to 0.7ppm.

[0077] Concentration protection forced overlay logic: ; where, in the formula, For logical statements, represented as if So , Represented as the first ammonia concentration in the area Indicates as forced full opening Deflectors in the area, This indicates that the forced shutdown is the first one. Even with uniform airflow, the ammonia injection system forces the deflectors to fully open to compensate for the reduced disturbance caused by closing the ammonia injection, thus maintaining the residence time of particulate matter.

[0078] Execution priority is shown in Table 5: Table 5. Priority Logic of the Air Deflector

[0079] In this embodiment, hybrid control and safety reversal: Actuator, ammonia injection valve ← Variable frequency fan ← Servo guide plate ← .

[0080] In this embodiment, it should be noted that, , Its positioning is not as a "final execution instruction," but rather as an "intermediate variable in intelligent decision-making." Temperature weighting coefficient. and pollutant weighting coefficient These are not directly equivalent to "ammonia injection valve opening" or "fan speed"; they are control strategy parameters output by the lightweight Transformer model, used to dynamically modulate the subsequent execution instruction generation logic. Using "weighting coefficients" instead of directly outputting the execution quantity has the following advantages: 1. Compatibility with existing control systems: Power plants already have basic ammonia injection control loops (based on denitrification efficiency setpoints); this invention, as an "upper-level intelligent optimization layer," does not overturn the original control architecture, only providing correction signals; engineering deployment costs are low, and it is easy to modify; 2. Decoupling modeling complexity: Directly outputting "valve opening" requires the model to learn the nonlinear characteristics of the entire actuator (such as the valve flow characteristic curve); while outputting "weighting coefficients" only requires learning the relationship between the operating conditions and control requirements, making the model lighter (≤64 KB); 3. Compatibility with safety fallback mechanisms and mandatory coverage strategies (such as... (Time-stop spray) can be directly applied to the final executed instruction without interfering with the output of the lightweight Transformer model; if , If it is directly equal to the valve opening, then safety logic is difficult to insert.

[0081] In this embodiment, a safety mechanism is set up: when the lightweight Transformer model fails, it automatically switches to the rule mode (normal mode); when the flow guide plate is stuck, it triggers the collaborative compensation of the adjacent area; all operation logs are uploaded to the cloud to support remote diagnosis.

[0082] S50 collects the total pollutant concentration at the end outlet of the fluid medium channel in real time. When the total pollutant concentration at the end outlet continues to deviate from the set target value, a global correction is applied to the pollutant weighting coefficient to form a closed-loop collaborative control.

[0083] In this embodiment, the outlet feedback correction involves collecting the total pollutant concentration at the outlet of the flue gas purification reactor. When the outlet concentration continuously deviates from the set target value, the pollutant weighting coefficient is adjusted. Apply a global correction to form a closed-loop collaborative control.

[0084] In this embodiment, the NOx feedback correction mechanism at the outlet is as follows: NOx at the CEMS outlet is collected (updated every 30 seconds); the 10-minute moving average NOx concentration is calculated. Correction logic: If Lasting 5 minutes → ,in, Pollutant weighting coefficients Apply a global correction. This is a function that takes the minimum value. If → .

[0085] Case study: Retrofitting the denitrification system of Unit #2 in a 300MW coal-fired power plant Hardware Deployment: Zone Division: The flue is 6 meters wide and divided into 6 zones, each 1 meter wide; Sensor Configuration (per zone): Micro differential pressure Pitot tube (Rosemount3051CD) → flow rate, K-type armored thermocouple → temperature, TDLAS laser analyzer (SickGM700) → NOx concentration, high temperature pressure transmitter → static pressure, light scattering dust meter (DURAGD-R290) → particulate matter concentration.

[0086] Actuators: Zoned ammonia injection regulating valve (6 units), servo-driven dynamic guide plate (6 sets, MAXON EC45 motor, 0°~90° continuously adjustable).

[0087] Control process: Collect data from 6 regions × 5 dimensions × 10 seconds, input the lightweight Transformer model to output ternary commands; determine the local NOx concentration threshold, execute ammonia injection + baffle adjustment, and perform closed-loop correction of outlet NOx concentration feedback.

[0088] Lightweight Transformer model parameters: input dimension 6×10×5; hidden layer 32-dimensional, number of main attention heads 2; total parameters 47.6KB, inference latency <50 ms (Siemens S7-1515F PLC).

[0089] The actual test results (June 2025, for 30 consecutive days) are shown in Table 6.

[0090] Table 6. Measured Results of Denitrification System Retrofit for Unit #2 of a 300MW Coal-fired Power Plant

[0091] Example 2, Application in Liquid Flow Systems - Intelligent Control System for Influent Distribution and Carbon Source Addition in Biological Tanks of Municipal Wastewater Treatment Plants, to solve the problems of uneven distribution, carbon source waste, and unstable denitrification efficiency caused by fluctuations in influent flow rate and water quality.

[0092] 1. System Deployment and Process Background Process flow: Municipal wastewater is pretreated by coarse and fine screens and a grit chamber before entering the anaerobic-anoxic-aerobic (AAO) biological treatment tank. This system is deployed at the end of the main inlet channel of the biological treatment tank and is responsible for distributing the inlet water and adding carbon sources (such as sodium acetate) to the six parallel biological treatment tank corridors.

[0093] Key challenges include: uneven flow distribution, with influent flow deviations of up to ±20% in each corridor due to differences in pipeline layout and resistance, resulting in inconsistent hydraulic retention times; drastic water quality fluctuations, with influent COD, ammonia nitrogen (NH3-N), and total nitrogen (TN) concentrations varying significantly with time of day (morning and evening peaks) and seasons; and inefficient carbon source addition, where traditional methods rely on online instrument signals from the main influent pipe to allocate carbon sources at fixed ratios, failing to detect actual hydraulic and water quality differences in each corridor, leading to insufficient carbon sources in some corridors (low denitrification efficiency) and excessive carbon sources in others (wasteful and potentially causing secondary pollution).

[0094] Suitable locations for modification: Install dynamic flow distribution plates and carbon source dosing points at the end of the main water intake channel and at the entrances of the branch channels that divide the water into each corridor.

[0095] 2. Hardware Deployment Zone division: The cross-section of the main channel is evenly divided into 6 independent monitoring and control zones along the width direction, corresponding to 6 downstream corridors.

[0096] Sensor configuration (per area): Flow velocity v: An ultrasonic Doppler flow meter (such as the Japanese Fuji FSC series) is used, with a range of 0.1-5m / s and an accuracy of ±1%FS. It is a non-contact measurement with no risk of blockage.

[0097] Temperature T: A PT100 platinum resistance thermometer with a 316L stainless steel sheath is used, which is directly inserted into the fluid.

[0098] Key water quality parameters C1 and C2: C1 is the concentration of ammonia nitrogen (NH3-N). An online ammonia nitrogen analyzer using ultraviolet spectrophotometry (such as the German WTB Amtax Compact) is employed, with a range of 0-50 mg / L and a response time of <5 min. C2 (a comprehensive indicator) is either chemical oxygen demand (COD) or total organic carbon (TOC). An online UV-VIS spectrometer (such as the German s:canspectro::lyser) is used to monitor COD trends in real time and to assess the carbon-to-nitrogen ratio.

[0099] Static pressure P: A diffused silicon pressure transmitter (such as E+HCerabarS) is used, with a tantalum diaphragm material that is corrosion resistant.

[0100] Suspended solids concentration (SS): A laser scattering suspended solids concentration meter (such as the Japanese Aqualight TSS-100) is used, with a range of 0-5000 mg / L.

[0101] In a multiphysics sensing network for liquid flow fields, adjustments are made to monitoring parameters, guide vane actuators, control models and logic, and system integration and safety.

[0102] In this embodiment, the adjustment of monitoring parameters (sensor array) includes: flow rate measurement, pollutant concentration measurement, particulate matter measurement, pressure and temperature measurement.

[0103] In flow rate measurement, a micro-differential Pitot tube is used for gas, while an electromagnetic flow meter, ultrasonic flow meter, or Doppler velocity meter is used for liquid. These devices have different requirements for the conductivity and cleanliness of the liquid, and the installation method also needs to be adjusted accordingly (such as full pipe installation, avoiding air bubbles).

[0104] In pollutant concentration measurement, gaseous samples are measured for specific gas concentrations such as NOx and SO2 (e.g., using TDLAS). Liquid samples require measurement of chemical oxygen demand (COD), biochemical oxygen demand (BOD), total organic carbon (TOC), pH, conductivity, turbidity, specific ion concentrations (e.g., ammonia nitrogen, phosphate), or oil concentration, depending on the target pollutant definition. Corresponding sensors include online water quality analyzers, UV-VIS spectrometers, and ion-selective electrodes, with different response times and installation methods compared to gas sensors.

[0105] In particulate matter measurement, gaseous dust concentration is measured using the light scattering method. Liquids are adapted for measuring suspended solids (SS) concentration or turbidity, using a turbidimeter or laser scattering particulate matter analyzer. Liquid transparency, color interference, and anti-fouling design must be considered.

[0106] Pressure and temperature measurement principles are similar, but the diaphragm material (such as Hastelloy or tantalum diaphragm) and sealing structure of liquid pressure sensors must be resistant to liquid corrosion and permeation. Temperature sensors, on the other hand, must consider the heat capacity and heat transfer characteristics of the liquid.

[0107] In this embodiment, the adjustment of the baffle actuator (actuator layer) must be based on the chemical properties of the liquid (acidity, alkalinity, chloride ion content, oxidizing properties) using corrosion-resistant materials such as 316L stainless steel, duplex steel, titanium alloy, PVC, PP, or plastic / rubber-lined structures. Regarding drive and sealing, the rotating shaft must be equipped with a reliable mechanical seal or packing seal to prevent liquid leakage. The drive motor needs a higher waterproof rating (e.g., IP68), and in flammable and explosive environments, explosion-proof certification (Ex d / IIC T6) is also required. In terms of hydrodynamic design, liquids have high density and inertia, resulting in hydrodynamic loads on the baffle that are much greater than those on gases. Mechanical calculations and structural design must be re-performed to ensure sufficient shaft strength and drive torque, and to consider impact loads during water hammer or sudden flow changes.

[0108] In this embodiment, regarding the adjustment of the control model and logic (decision layer), in terms of changes in the physical field coupling relationship, in liquids, temperature has a significant impact on viscosity, directly affecting flow resistance and velocity distribution. Therefore, in the "spatial coupling head" of the lightweight Transformer, the coupling weights of temperature and velocity need to be retrained and redefined. The mixing of liquid contaminants relies more on convection than diffusion, and the velocity distribution has a more critical impact on the uniformity of the concentration field, which needs to be strengthened in the attention mechanism. Regarding timescale adjustments, liquid systems typically respond more slowly than gases (due to greater inertia and larger pipe capacity). The data sampling period, the model input time window length (originally 10 seconds), and the output frequency of control commands may need to be extended to accommodate the slower fluid dynamics.

[0109] In this embodiment, regarding system integration and safety adjustments (system level), in terms of installation and maintenance, the liquid system needs to consider interfaces for discharge, cleaning, and antifreeze (for outdoor or cold regions). Sensor probes need to be easily disassembled and cleaned online or equipped with automatic cleaning devices (such as ultrasonic cleaning). In terms of communication and signaling, long-distance liquid pipelines may require fieldbus or wireless transmission solutions more adapted to humid environments. In the safety fallback strategy, the spare values ​​in the rule control table need to be reset according to typical operating conditions of the liquid process (such as different influent volumes and seasonal variations in water quality).

[0110] Execution agency configuration: Dynamic flow distribution board: Blade material: Made of 316L stainless steel, with a polished surface to reduce adhesion.

[0111] Drive unit: It adopts an IP68 waterproof and explosion-proof (Ex d) DC servo motor and is equipped with a double-end mechanical seal rotating shaft to ensure long-term underwater operation without leakage.

[0112] Opening range: 0° (fully closed, used for flow throttling) to 90° (fully open). The throttling position is commonly used in high flow areas and is set at 20°-40°; the fully open position is used in low flow areas or for forced flushing and is set at 80°-90°.

[0113] Carbon source dosing valve: It adopts a corrosion-resistant diaphragm regulating valve (such as GEMÜ from Germany), which is linked with the metering pump and accepts a 4-20mA regulating signal from the PLC.

[0114] Control unit: All area sensor signals are connected to a central industrial PLC (such as Siemens S7-1500), and the sampling period is set to 5 seconds (slower than the gas system, which is in line with the dynamic response speed of liquids).

[0115] 3. Liquid adaptation and training of lightweight Transformer models Input data tensor: The dimension has been adjusted to: 6: Six areas (spaces).

[0116] 12: The time window length is 60 seconds (because the sampling period is 5 seconds, there are 12 time steps).

[0117] 6. Characteristic dimensions: flow rate (v), temperature (T), ammonia nitrogen concentration (C1), COD concentration (C2), static pressure (P), and suspended solids concentration (SS).

[0118] Model structure adjustment: The core structure remains unchanged; it is still a dual-head attention encoder.

[0119] Spatial Coupled Head (Head1): Focus on learning the coupling relationship between flow velocity, suspended solids concentration and baffle opening, as well as the effect of temperature on fluid viscosity (indirectly affecting flow velocity distribution).

[0120] Time Evolution Head (Head2): Focuses on studying the changing trends of ammonia nitrogen and COD concentrations to predict carbon source demand.

[0121] Output: Ternary instruction .

[0122] Temperature-viscosity weighting coefficient. At low temperatures, liquid viscosity increases, leading to increased flow resistance. Increase the size to indicate that the baffle opening needs to be increased or the carbon source mixing strategy needs to be adjusted.

[0123] Ammonia nitrogen / COD pollutant weighting coefficient. Directly used to adjust the carbon source dosage.

[0124] The opening angle of the dynamic guide vane is used to balance the water inflow of each channel.

[0125] Training data and loss function: Training data: Collect one year of historical operating data (including different operating conditions such as rainy season, dry season, and holidays), approximately 800,000 sets of samples.

[0126] Tag generation: Based on the optimization objective of achieving TN standards in the effluent of the corridor while minimizing total carbon source consumption, the ideal conditions at each time point are derived by using historical best operating time data. , , The value serves as a supervisory label.

[0127] Loss function: Weighted mean squared error (MSE) is used, where given (Carbon source control) higher weight ( =0.7), to emphasize the goal of energy conservation and emission reduction.

[0128] 4. Dedicated control logic for liquid systems Calculation of basic carbon source addition: The basic carbon source dosage is calculated based on the total ammonia nitrogen load in the influent main and the set carbon-nitrogen ratio. The final carbon source dosage for each zone is determined by the following formula: ; in, , These are fixed weights in the liquid system, determined based on actual process adjustments. This refers to the final carbon source addition amount. This refers to the amount of basic carbon source added.

[0129] Intelligent control logic of the deflector: Main control logic: Model output The aim is to increase the flow rate in each region. Approaching the average flow velocity It also takes into account the balance of suspended solids concentration (to prevent sedimentation).

[0130] Forced protection logic: Anti-sedimentation logic: If the flow rate in a certain area... If the flow rate remains below 0.3 m / s (the critical velocity for preventing sedimentation), then the area will be forced to... ←85° (fully open) and trigger an alarm.

[0131] Low load protection: If the ammonia nitrogen concentration in a certain area... If the concentration remains consistently below 2 mg / L, the nitrogen removal load of the corridor is considered extremely low, and forced nitrogen removal is required. ←0 (Stop adding carbon source), and set ←90° to maintain basic flow velocity and prevent sedimentation. Among these, This represents the region index for dividing the total channel cross-section.

[0132] Export feedback correction (closed loop): Collect the TN value of the effluent at the end of the corridor of each biological treatment tank (update cycle 15 minutes).

[0133] If the total nitrogen (TN) of the effluent from a certain channel exceeds a set value (e.g., 12 mg / L) for three consecutive cycles, then the channel's... Apply a positive, gradual correction amount ( ).

[0134] If the total nitrogen (TN) of all effluent from all corridors continues to meet the standards and the total carbon source addition is higher than the historical benchmark, then global optimization will be triggered for fine-tuning. value.

[0135] 5. Actual test results: A continuous 90-day test was conducted at a wastewater treatment plant with a daily processing capacity of 100,000 tons (covering spring and summer). See Table 7 for details.

[0136] Table 7. Performance Verification in Liquid Purification Scenarios

[0137] 6. Conclusion This embodiment fully verifies the effectiveness and superiority of the present invention in liquid purification scenarios. By adapting sensors, actuators, model features, and control logic to the characteristics of the liquid medium, the present invention successfully solves the core problems that have long existed in liquid flow and reaction systems, such as uneven distribution, coarse dosing, and lag response. It realizes the technological migration and expansion from "gas flue gas purification" to "liquid fluid control," demonstrating its strong universality and engineering application value.

[0138] Example 3 This invention provides a system for applying the fluid purification method based on the synergistic control of a lightweight Transformer and a baffle plate described above, comprising: The physical parameter module is used to divide the cross-section of the fluid medium channel inlet into multiple independent monitoring and control areas, and to collect multi-physical field parameters of each area in real time.

[0139] The feature module is used to preprocess the multiphysics parameters and construct the spatiotemporal feature tensor according to the region dimension and the time dimension.

[0140] The inference module is used to input spatiotemporal feature tensors into a lightweight Transformer model and output ternary cooperative control commands.

[0141] The adjustment module is used to adjust the purification actuators and dynamic guide vane actuators corresponding to each area according to the three-element coordinated control command.

[0142] The feedback correction module is used to collect the total pollutant concentration at the end outlet of the fluid medium channel in real time. When the total pollutant concentration at the end outlet continues to deviate from the set target value, a global correction amount is applied to the pollutant weight coefficient to form a closed-loop collaborative control.

[0143] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention, and no reference numerals in the claims should be construed as limiting the scope of the claims.

[0144] The above embodiments are merely examples of implementation methods of the invention. The scope of protection of the present invention is not limited to the above embodiments. For those skilled in the art, several modifications and improvements can be made without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention.

Claims

1. A fluid purification method based on the synergistic control of a lightweight Transformer and a baffle plate, characterized in that, Adaptable to gas and liquid flow systems, including: Multiple independent monitoring and control zones are divided on the cross-section of the fluid medium channel inlet, and multi-physics parameters of each zone are collected in real time. After preprocessing the multiphysics parameters, a spatiotemporal feature tensor is constructed according to the region dimension and the time dimension. The spatiotemporal feature tensor is input into a lightweight Transformer model, and the output is a ternary cooperative control command; According to the three-element coordinated control command, the corresponding purification actuators and dynamic guide vane actuators in each area are adjusted; The total pollutant concentration at the end outlet of the fluid medium channel is collected in real time. When the total pollutant concentration at the end outlet continues to deviate from the set target value, a global correction is applied to the pollutant weighting coefficient to form a closed-loop collaborative control.

2. The fluid purification method based on the coordinated control of a lightweight Transformer and a baffle plate according to claim 1, characterized in that, The lightweight Transformer model architecture includes: a feature embedding layer, a position encoding layer, a dual-head self-attention layer, a feedforward neural network layer, and a parameter generation layer; In the gas flow system, the feature embedding and position encoding layer maps the learnable embedding matrix of the spatiotemporal feature tensor into a high-dimensional feature vector and adds spatiotemporal position encoding to obtain a region-time embedding representation. A spatial coupling head and a temporal evolution head are set in the dual-head attention layer. Based on high-dimensional feature vectors, the spatial coupling head calculates the coupling relationship between flow velocity and temperature between regions and outputs spatial coupling feature vectors. The temporal evolution head analyzes the changing trend of pollutant concentration over time and outputs temporal evolution feature vectors. The outputs of the dual-head attention layer are then concatenated. The concatenated feature vector is input into the feedforward neural network layer for nonlinear transformation, and then mapped to the final three-element collaborative control command through the parameter generation layer, including: temperature weight coefficient, pollutant weight coefficient and dynamic guide vane opening angle.

3. The fluid purification method based on the synergistic control of a lightweight Transformer and a baffle plate according to claim 2, characterized in that, The opening degree of the ammonia injection regulating valve in the purification actuator is determined by the following formula: ； In the formula, This is the final ammonia injection rate. This is the baseline ammonia injection rate calculated based on the inlet NOx concentration and flue gas flow rate. , To set the weighting coefficients for the project, Temperature weighting coefficient, This represents the pollutant weighting coefficient.

4. The fluid purification method based on the coordinated control of a lightweight Transformer and a baffle plate according to claim 2, characterized in that, According to the three-element coordinated control command, the corresponding purification actuators and dynamic guide vane actuators in each area are adjusted to perform execution control and safety constraints: when the pollutant concentration in any area is lower than the preset safety threshold, the pollutant weight coefficient of that area is forcibly set to zero, and the opening angle of the corresponding dynamic guide vane is adjusted to the fully open position to avoid over-spraying and maintain flow field stability.

5. The fluid purification method based on the synergistic control of a lightweight Transformer and a baffle plate according to claim 2, characterized in that, The deflector opening angle is dynamically determined in the three-dimensional collaborative control command and is subject to dual strategy constraints, including: airflow uniformity main control logic, concentration protection forced coverage logic, and execution priority logic. The main control logic for flow uniformity is as follows: ； In the formula, For the first The dynamic deflector opening angle of the area This is expressed as the average cross-sectional velocity at the inlet of the fluid medium channel. For the first Real-time flow rate of the region; The logic for forced coverage of concentration protection is as follows: In the formula, For logical statements, represented as if So ; For the first ammonia concentration in the area Indicates as forced full opening Deflectors in the area, This indicates that the forced shutdown is the first one. Ammonia spraying in the area For the first Pollutant weighting coefficients for the region; The execution priority logic is as follows: In the first priority, the condition is that the lightweight Transformer model inference is successful, and the action is that the deflector uses the output dynamic deflector opening angle; in the second priority, the condition is... The action is forced. In the third priority, if the sensor fails, the action is to maintain the previous valid value; in the fourth priority, if the system stops suddenly, the action is to lock the current position.

6. The fluid purification method based on the coordinated control of a lightweight Transformer and a baffle plate according to claim 2, characterized in that, When the total pollutant concentration at the terminal outlet continuously deviates from the set target value, a global correction is applied to the pollutant weighting coefficients to form a closed-loop collaborative control. The correction logic is as follows: If... If this continues for a period of time, a global correction will be applied to the pollutant weighting coefficients. ;like Then the global correction applied to the pollutant weight coefficients is a fixed value; where, the global correction... It can be obtained through the following formula: ； In the formula, The average NOx concentration. This is a function that takes the minimum value.

7. The fluid purification method based on the coordinated control of a lightweight Transformer and a baffle plate according to claim 2, characterized in that, In liquid flow systems, the lightweight Transformer model differs from that in gas flow systems, except for the different multiphysics parameters collected, in the dual-head attention layer. The spatial coupling head learns the coupling relationship between flow velocity, suspended solids concentration and baffle opening, while the time evolution head learns the changing trends of ammonia nitrogen and COD concentrations. The ternary collaborative control commands output by the lightweight Transformer model are the temperature-viscosity weighting coefficient, the ammonia nitrogen / COD pollutant weighting coefficient, and the dynamic baffle opening angle.

8. The fluid purification method based on the synergistic control of a lightweight Transformer and a baffle plate according to claim 7, characterized in that, Purification implementation agency Carbon source addition; wherein, the final carbon source addition amount for each zone is determined by the following formula: ； In the formula, This refers to the final carbon source addition amount. Basic carbon source addition amount , These are fixed weights in the liquid system, determined based on actual process adjustments. This is the temperature-viscosity weighting coefficient. This represents the ammonia nitrogen / COD pollutant weighting coefficient.

9. The fluid purification method based on the coordinated control of a lightweight Transformer and a baffle plate according to claim 7, characterized in that, The control logic of the guide vane in a liquid flow system includes: Main control logic: Adjust the opening angle of the dynamic guide vane based on the output of the lightweight Transformer model; The mandatory protection logic includes: Anti-precipitation logic: If the first Regional flow velocity If the speed remains below 0.3 m / s, then the area will be forced to... ←85° and trigger an alarm; Low load protection: If the first Regional ammonia nitrogen concentration If the concentration remains consistently below 2 mg / L, the nitrogen removal load of the corridor is considered extremely low, and forced nitrogen removal is required. ←0, stop adding carbon source and set ←90° to maintain basic flow rate and prevent siltation.

10. A system applying the fluid purification method based on the coordinated control of a lightweight Transformer and a baffle as described in any one of claims 1-9, characterized in that, include: The physical parameter module is used to divide the cross-section of the fluid medium channel inlet into multiple independent monitoring and control areas, and to collect multi-physical field parameters of each area in real time. The feature module is used to preprocess the multiphysics parameters and construct the spatiotemporal feature tensor according to the region dimension and the time dimension. The inference module is used to input spatiotemporal feature tensors into a lightweight Transformer model and output ternary cooperative control commands. The adjustment module is used to adjust the purification actuators and dynamic guide vane actuators corresponding to each area according to the three-element coordinated control command; The feedback correction module is used to collect the total pollutant concentration at the end outlet of the fluid medium channel in real time. When the total pollutant concentration at the end outlet continues to deviate from the set target value, a global correction amount is applied to the pollutant weight coefficient to form a closed-loop collaborative control.

Images