Data-driven super large air blowing alkali method bromine extraction intelligent energy-saving control system

Abstract

The application relates to the technical field of automation control, and discloses a data-driven super-large air blowing alkali bromine extraction intelligent energy-saving control system, which comprises a multi-parameter detection array, a data fusion pretreatment module, a bromine extraction prediction module, a collaborative optimization control module and a model incremental updating module.The multi-parameter detection array collects asynchronous process data and actual bromine concentration in real time; the data fusion pretreatment module performs drift correction and time sequence alignment based on a reference benchmark to generate a characteristic sequence; the bromine extraction prediction module outputs continuous prediction of bromine concentration through a transformer encoder network; the collaborative optimization control module fuses dynamic electricity prices and material costs to construct a cost function, and solves optimal air flow and acid addition control instructions in a rolling manner; and the model incremental updating module triggers gradient descent based on error monitoring to realize online parameter replacement.The application eliminates data drift error, breaks the high-energy-consumption bottleneck caused by static setting, realizes global dynamic optimization and energy saving and consumption reduction in the bromine extraction process.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, specifically to a data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction. Background Technology

[0002] The air-blowing alkaline process is currently the main industrial process for extracting bromine from seawater. This process primarily includes stages such as an acidification oxidation tower, an air-blowing tower, a primary alkaline absorption tower, a secondary alkaline absorption tower, a mist eliminator, and a closed-loop fan system, supplemented by auxiliary pumps such as alkaline circulation pumps, raw material pumps, and acid pumps. The acidification oxidation tower and the air-blowing tower are the core components of the bromine extraction production line, where complex fluid dynamics, gas-liquid two-phase mass transfer processes, and chemical reaction kinetics occur. In ultra-large-scale bromine extraction production lines (taking the air-blowing tower as an example, the traditional tower diameter is typically around 6.5m, with a processing capacity of approximately 1200m³),... 3 / h; while the diameter of ultra-large towers can reach 16m, and the single tower capacity can reach up to 9000m³. 3 The seawater flow is enormous and the background water quality (such as temperature, salinity, and initial bromide ion concentration) fluctuates continuously due to natural seasonal, climatic, and tidal changes, which places strict requirements on the control of process parameters in the bromine extraction process.

[0003] Because the bromine extraction process has large time lags and strong nonlinear characteristics, and at ultra-large scales, the waste of raw materials and energy costs caused by the lag fluctuations of traditional static control are magnified many times over. Conventional control algorithms rely solely on the deviation between the setpoint and the current measured value for passive adjustment, making it difficult to compensate for process disturbances caused by fluctuations in seawater conditions in advance, which easily leads to fluctuations in the terminal blow-out rate. At the data acquisition level, the bromine extraction process site environment is highly corrosive, and solid-state sensors that are in direct contact with the reaction fluid exhibit potential drift after long-term operation, causing the measured values ​​to gradually deviate from the true values. At the same time, different types of detection instruments on site (such as spectrometers and electrochemical probes) differ in hardware signal response time and sampling period, resulting in asynchronous misalignment of the output data sequence on the time axis. Existing systems typically lack cleaning, alignment, and drift correction mechanisms for multi-source asynchronous sensor data. Directly substituting noisy and misaligned data into the control loop can cause malfunctions in the actuators.

[0004] In terms of system execution and cost control, existing operating methods generally rely on fixed threshold settings. To ensure that the final bromine concentration meets the standard, the control system typically drives the raw material pump and acid pump chlorination valve continuously according to a manually set conservative upper limit, maintaining a high airflow and excessive reagent dosing. This static control mode does not take into account the operating boundaries of the equipment, dynamically changing real-time material costs, and peak and off-peak electricity prices, causing the production line to maintain high-power operation during peak grid load periods or when raw water conditions are good, resulting in redundant consumption of electrical energy and chemical raw materials.

[0005] Some bromine extraction processes have attempted to introduce data-driven predictive models trained on historical datasets. However, the bromine extraction reaction mechanism and equipment operating status undergo long-term changes due to pipeline scaling, catalyst decay, and seasonal fluctuations. When static predictive models trained and solidified offline are continuously deployed in the field, the error between their output predictions and the actual process conditions gradually accumulates over time. Existing control systems lack online closed-loop monitoring and incremental fine-tuning mechanisms for the model's internal parameters, and cannot adaptively update weights based on new data collected in real-time. This leads to irreversible performance degradation in the predictive optimization capabilities of the control system during long-term operation. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction. This system solves the problems of control distortion caused by data drift from multi-source asynchronous sensors in existing bromine extraction process control systems, energy consumption and reagent redundancy waste caused by conventional passive feedback and static conservative settings, and control performance degradation caused by the lack of adaptive update mechanism in static prediction models during long-term operation.

[0007] To achieve the above objectives, the present invention provides the following technical solution: This invention provides a data-driven, ultra-large-scale intelligent energy-saving control system for air-blowing alkaline bromine extraction, comprising: The multi-parameter detection array module includes a nuclear magnetic resonance analyzer, a transistor pH sensor, and a temperature and salinity probe, which are used to collect data on the bromine extraction process status and the actual bromine concentration. The data fusion preprocessing module is used to receive the bromine extraction process status data, perform outlier removal, time series alignment and standardization normalization operations, and output a feature vector sequence. The bromine extraction prediction module is used to input the feature vector sequence into the internal transformer encoder network model for processing, and output a continuous bromine extraction rate prediction sequence and a predicted bromine concentration. The collaborative optimization control module is used to receive the continuous bromine extraction rate prediction value sequence, perform a cost function minimization calculation under constraints, and output gas flow rate command, acid addition rate command and chlorination rate command. The model incremental update module is used to receive the actual bromine concentration and the predicted bromine concentration, calculate the average absolute error between the actual bromine concentration and the predicted bromine concentration, and extract historical data to perform gradient descent training on the transformer encoder network model and update the parameters when the average absolute error is greater than a set threshold.

[0008] Preferably, the multi-parameter detection array module includes: The nuclear magnetic resonance concentration detection unit includes the nuclear magnetic resonance analyzer, which is used to collect the actual bromine concentration and chlorine residue in the bromine extraction process status data; A transistor pH detection unit, including the transistor pH sensor, is used to collect the pH value within the bromine extraction process status data; The temperature and salinity detection unit includes the temperature and salinity probe, which is used to collect temperature and salinity information within the bromine extraction process status data.

[0009] Preferably, the data fusion preprocessing module includes: The noise localization and isolation unit uses the isolated forest algorithm to locate and isolate outliers within the bromine extraction process status data; The sensor drift correction unit uses an extended Kalman filter to construct the system state equation and observation equation, and outputs a corrected pH value. The sampling time alignment unit uses a network time protocol and an asynchronous interpolation algorithm to convert the bromine extraction process status data with different sampling frequencies, along with the corrected pH value, into the same time series. The standardization and normalization unit performs numerical normalization calculations on the data within the same time series and outputs the feature vector sequence.

[0010] Preferably, the sensor drift correction unit uses the actual bromine concentration as an unbiased reference benchmark, establishes a state vector including the actual pH value, offset voltage, and the actual bromine concentration, substitutes the state vector into the system state equation and the observation equation to perform state estimation calculation to obtain the estimated offset voltage, and performs addition and subtraction operations on the pH value inside the bromine extraction process state data and the estimated offset voltage to obtain the corrected pH value.

[0011] Preferably, the bromine extraction prediction module includes: Principal component feature dimensionality reduction unit performs orthogonal decomposition dimensionality reduction on the feature vector sequence using principal component analysis, and outputs a dimensionality-reduced feature sequence; The multi-head self-attention inference unit inputs the dimensionality-reduced feature sequence into the multi-head self-attention layer inside the transformer encoder network model, calculates the correlation weight parameters between features at different times, and uses the correlation weight parameters to perform weighted summation and aggregation on the value vector inside the dimensionality-reduced feature sequence to obtain the hidden layer representation sequence. The multilayer perceptron mapping unit maps the hidden layer representation sequence to the continuous bromine extraction rate prediction sequence using a multilayer perceptron.

[0012] Preferably, the multilayer perceptron mapping unit includes a fully connected neural network layer. The multilayer perceptron mapping unit inputs the hidden layer representation sequence into the fully connected neural network layer for weight matrix multiplication and nonlinear activation function operation to generate a numerical sequence in the future control time domain as the predicted value sequence of the continuous bromine extraction rate, and simultaneously outputs the corresponding predicted bromine concentration.

[0013] Preferably, the collaborative optimization control module includes: A multi-objective cost function construction unit establishes the cost function composed of the sum of a first penalty sub-item, a second penalty sub-item, a third penalty sub-item, and a fourth penalty sub-item; The operation constraint unit establishes a set of constraint conditions, including system dynamics equality constraints, equipment operation boundary inequality constraints, and terminal target concentration relaxation constraints, as the constraint conditions. The sequential quadratic programming solution unit uses the sequential quadratic programming algorithm to perform rolling numerical solutions on the cost function and the set of constraints within a set control period, and outputs the airflow rate command, the acid addition rate command, and the chlorination rate command.

[0014] Preferably, the first penalty term established by the multi-objective cost function construction unit corresponds to the product of the square term of the airflow rate command and the first weighting coefficient; the second penalty term corresponds to the product of the square term of the difference between the set target bromine concentration and the predicted bromine concentration and the second weighting coefficient; the third penalty term corresponds to the product of the square term of the acid addition rate command and the third weighting coefficient; and the fourth penalty term corresponds to the product of the square term of the chlorination rate command and the fourth weighting coefficient. The first weighting coefficient, the second weighting coefficient, the third weighting coefficient, and the fourth weighting coefficient are dynamically assigned values ​​based on real-time material costs and grid prices.

[0015] Preferably, the model incremental update module includes: The absolute error monitoring unit receives the actual bromine concentration and the predicted bromine concentration, and calculates the average absolute error between the actual bromine concentration and the predicted bromine concentration within a set time window. A historical sample cache unit stores historical experience data packages containing multi-source aligned feature vectors and the actual bromine concentration labeling results; The gradient descent parameter update unit extracts the historical experience data package when the mean absolute error value is greater than a set threshold, performs gradient descent training on the transformer encoder network model, and replaces and updates the network weight parameters inside the transformer encoder network model.

[0016] Preferably, the gradient descent parameter update unit inputs the extracted historical experience data package into the transformer encoder network model for backpropagation calculation, modifies the network weight parameters and feature mapping matrix parameters inside the transformer encoder network model according to the calculated gradient values, and after the modification is completed, directly replaces the network weight parameters and feature mapping matrix parameters in the bromine extraction prediction module to perform online inference calculation.

[0017] This invention provides a data-driven, ultra-large-scale intelligent and energy-saving control system for air-blowing alkaline bromine extraction. It offers the following advantages: 1. This invention utilizes an extended Kalman filter within the data fusion preprocessing module to use the bromine concentration collected by a nuclear magnetic resonance analyzer as an unbiased reference. This allows for state estimation and numerical compensation of the voltage drift generated by the transistor pH sensor over time. This solves the potential drift problem of the sensor in complex seawater environments. Furthermore, by combining an interpolation algorithm, the sampling timing of multi-source asynchronous data is aligned, providing a stable process feature vector for the control system's operation that is free from time phase deviations.

[0018] 2. This invention employs principal component analysis combined with a transformer encoder network to construct a prediction model. Orthogonal decomposition is used to process high-dimensional state inputs, reducing the computational burden caused by multicollinearity among temperature, salinity, and the flow rates of various reagents. Subsequently, a multi-head self-attention layer is used to calculate the correlation weights between nodes in the time series. This enables the objective extraction of nonlinear reaction kinetics characteristics under the coupling of multiple process variables, outputting a continuous bromine extraction rate prediction sequence.

[0019] 3. This invention constructs a closed feedback loop between a collaborative optimization control module and a model incremental update module. Under the constraints of equipment execution boundaries, it synchronously outputs multiple control commands by solving a multi-objective cost function that includes wind turbine power consumption and reagent consumption. Simultaneously, by monitoring the absolute error value, it uses historical cached data to fine-tune the gradient descent parameters of the model when the error exceeds the limit. This changes the isolated adjustment method of a single loop, realizes synchronous linkage of multiple devices, and enables the predictive control model to have online parameter correction capabilities to adapt to the time-varying characteristics of seawater quality.

[0020] 4. This invention effectively reduces bromine loss in blow-out wastewater and improves the economic benchmark for bromine extraction in the industry. Through multi-dimensional parameter monitoring and predictive adjustment, when faced with raw seawater fluctuating at around 60 ppm, the residual bromine concentration in blow-out wastewater can be stabilized and reduced from the traditional 18-20 ppm to the range of 12-15 ppm, thereby increasing the overall bromine extraction rate from the traditional 65%-70% to 75%-80%. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 Figure (a) shows the comparison curve of the actual bromine concentration tracking and control effect, and Figure (b) shows the comparison curve of the cumulative comprehensive cost of system operation (including power consumption and reagent consumption). Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Reference Figure 1 This invention provides a data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction process. This system is applied to an ultra-large-scale air-blowing alkaline bromine extraction production line, which includes an acidification oxidation tower, an air blowing tower, a primary alkaline absorption tower, a secondary alkaline absorption tower, a mist eliminator, and a closed-loop fan system. It also includes the control loops for external alkaline circulation pumps, raw material pumps, acid pumps, and other pumps, as well as the alkalinization reactor, air blowing tower, compressor, acid pump, and chlorination valve.

[0024] The control system includes a multi-parameter detection array module, a data fusion preprocessing module, a bromine extraction prediction module, a collaborative optimization control module, and a model incremental update module.

[0025] The multi-parameter detection array module is installed at the inlet and outlet pipelines of the alkalization reactor and along the flow path of the air blowing tower in the bromine extraction production line. The multi-parameter detection array module includes a nuclear magnetic resonance analyzer, a transistor pH sensor, and temperature and salinity probes, used to acquire process status data and actual bromine concentration online during the operation of the bromine extraction process.

[0026] The signal output of the multi-parameter detection array module is connected to the signal input of the data fusion preprocessing module via an industrial Ethernet bus. The data fusion preprocessing module receives the process status data transmitted by the multi-parameter detection array module, performs outlier removal, time series alignment, and normalization operations through its built-in data cleaning and alignment logic, and generates a continuous feature vector sequence at the output.

[0027] The signal output of the data fusion preprocessing module is communicatively connected to the signal input of the bromine extraction prediction module. The bromine extraction prediction module contains a transformer encoder network model. It receives a sequence of feature vectors as the input matrix to the network model, performs forward propagation calculations, and outputs a continuous sequence of predicted bromine extraction rates and the corresponding predicted bromine concentrations.

[0028] The signal output of the bromine extraction prediction module is connected to the signal input of the collaborative optimization control module. The collaborative optimization control module receives a continuous sequence of predicted bromine extraction rates. Internally, the collaborative optimization control module is equipped with a mathematical programming solver, which performs a minimization operation on the multi-objective cost function under the set system dynamics and equipment operation boundary constraints. The control command output of the collaborative optimization control module is connected to the electrical drive terminals of the compressor, acid pump, raw material pump, and alkali circulation pump on the production line via a field-programmable logic controller (FPGA), used to issue airflow rate commands, acid addition rate commands, raw material pump flow rate commands, and alkali circulation rate commands to the actuators. The rich array of actuators, combined with the oxidation potential and pressure drop data obtained by the multi-dimensional physicochemical detection unit, enables the control system to comprehensively cover the adjustment needs of fluid dynamics and reaction kinetics.

[0029] The concentration signal output of the multi-parameter detection array module and the prediction data output of the bromine extraction prediction module are both connected to the data input of the model incremental update module. The model incremental update module synchronously receives the actual bromine concentration and the predicted bromine concentration, and calculates the average absolute error between the actual bromine concentration and the predicted bromine concentration within a set time window.

[0030] The parameter update control terminal of the model incremental update module is connected to the network parameter configuration interface of the bromine extraction prediction module. When the model incremental update module determines that the mean absolute error value is greater than the set trigger threshold, it extracts historical process state features and corresponding actual bromine concentrations from the internal memory as a training sample set, performs gradient descent backpropagation calculation on the transformer encoder network model, generates an updated network weight matrix, and replaces the original network weight parameters in the bromine extraction prediction module through the parameter configuration interface.

[0031] Reference Figure 1 The multi-parameter detection array module includes a nuclear magnetic resonance concentration detection unit, a transistor pH detection unit, and a temperature and salinity detection unit. The various sensor components of the multi-parameter detection array module are installed in the inlet and outlet pipes of the bromine extraction production line and at the outlet node of the air blow-out tower via flange connections or bypass pipelines, for online acquisition of bromine extraction process status data and actual bromine concentration.

[0032] The nuclear magnetic resonance (NMR) concentration detection unit includes an NMR analyzer. The NMR analyzer is connected to the main drain pipe at the bottom of the air blow-out tower via a bypass sampling pump. The NMR analyzer internally contains a radio frequency transmitter and a magnetic field coil. It continuously acquires the NMR signals of seawater samples flowing through the test chamber according to a set first sampling cycle. After Fourier transform processing by the internal processor, it outputs the actual bromine concentration and chlorine residue levels within the bromine extraction process status data.

[0033] The discrete sequence of actual bromine concentration output by the nuclear magnetic resonance analyzer is represented as follows: The output discrete sequence of chlorine residue is represented as follows: .in, The discrete sampling time step sequence number is in positive integer form. This represents the first sampling period time interval value of the nuclear magnetic resonance analyzer.

[0034] The transistor pH detection unit includes a transistor pH sensor. The probe housing of the transistor pH sensor is screwed directly into the fluid in the mixing reaction zone of the bromine extraction production line via a threaded interface. The transistor pH sensor consists of a measuring electrode containing an ion-sensitive metal-oxide-semiconductor field-effect transistor and a reference electrode, used to collect the pH value within the bromine extraction process status data. The transistor pH sensor outputs a raw voltage signal according to a set second sampling period, and the potential response process at its internal measuring electrode interface follows the Nernst mathematical mechanism. The formula for the raw voltage signal output by the transistor pH sensor is: ; In the formula, For continuous time variables; For transistor pH sensors in continuous time variables The original output voltage value; This is the fixed reference potential value for the reference electrode; It is the ideal gas constant; The temperature variable is included in the bromine extraction process status data. It is Faraday's constant; This represents the actual pH value of the fluid. This refers to the unmeasurable offset voltage parameter generated by the transistor pH sensor under long-term immersion in seawater. The built-in analog-to-digital converter circuit of the transistor pH sensor... Discretized sampling is performed, and the output sampling interval is the second sampling period. The discrete sequence of pH values.

[0035] The temperature and salinity detection unit includes a temperature and salinity probe. The probe and the transistor pH sensor are installed side-by-side at the same pipe cross-section. The temperature and salinity probe internally encapsulates a thermistor temperature sensing element and a dual-electrode conductivity measuring element, operating according to a set third sampling cycle. The temperature and salinity data within the bromine extraction process are collected. The temperature and salinity probes output a discrete temperature sequence. The voltage signal is synchronously transmitted to the transistor pH detection unit and substituted into the original voltage signal formula to perform hardware-level temperature reference value compensation. This is the sampling sequence number.

[0036] The multi-parameter detection array module continuously transmits discrete-time series data generated by the NMR analyzer, transistor pH sensor, and temperature-salinity probe to the bus via an industrial Ethernet interface. Due to the hardware-level difference between the analysis time of the NMR analyzer and the electrical signal response time of the solid-state sensor, the first sampling period... Second sampling period and the third sampling period The unequal values ​​cause the bromine extraction process status data output by the multi-parameter detection array module and the actual bromine concentration to exhibit an asynchronous discrete time sequence with non-overlapping sampling nodes.

[0037] Reference Figure 1 The data fusion preprocessing module is internally connected to a noise localization and isolation unit, a sensor drift correction unit, a sampling timing alignment unit, and a standardization and normalization unit.

[0038] The signal input of the noise localization and isolation unit is connected to the multi-parameter detection array module. The noise localization and isolation unit receives bromine extraction process status data and actual bromine concentration, and constructs an isolation forest model containing multiple random binary trees in its internal memory. Within the isolation forest model, the noise localization and isolation unit randomly selects a feature dimension within the bromine extraction process status data and randomly chooses a split value between the maximum and minimum values ​​to perform binary partitioning of the data points, calculating the average path length of each data point in the forest. The noise localization and isolation unit compares the average path length with an internally set isolation depth threshold, marking data points with average path lengths less than the isolation depth threshold as outlier noise. It then performs location isolation and data removal operations on the marked outlier noise in the data sequence.

[0039] The input of the sensor drift correction unit is connected to the output of the noise localization and isolation unit. The sensor drift correction unit uses the actual bromine concentration output from the NMR analyzer (after outlier removal) as an unbiased reference standard to calculate the voltage drift of the transistor pH sensor online. An extended Kalman filter is deployed within the sensor drift correction unit. The state equation and observation equation of the nonlinear system established by the extended Kalman filter are as follows.

[0040] Define the discrete time step as The state vector of a control system is defined as follows: ,in This represents the actual pH value of the system. This represents the offset voltage of the transistor pH sensor. This represents the actual bromine concentration. The state equation of the control system is expressed as: ; In the formula, A nonlinear transfer function representing the evolution of the system state over time; Represents the system process noise vector, with a covariance matrix. ; Representative at the The state vector of the control system at the previous discrete time step; This represents the state vector of the control system.

[0041] The control system observation vector is defined as ,in This represents the raw pH measurement value acquired by the transistor pH sensor. This represents the actual bromine concentration measured by the nuclear magnetic resonance analyzer. The observation equation is expressed as: ; In the formula, The observation function represents the mapping relationship between real state variables and sensor measurements; Represents the measurement noise vector, with a covariance matrix. ; Represents the state vector of the control system; Indicates the first The observation vector of the control system at discrete time steps. Due to Including the combined effects of the actual pH value and the offset voltage, the extended Kalman filter performs a first-order linearization expansion of the nonlinear observation function by calculating the Jacobian matrix during operation, and then calculates the Kalman gain matrix.

[0042] The extended Kalman filter utilizes the calculated Kalman gain matrix, combined with the current-time observation vector containing NMR measurement results. Feedback correction is performed on the prior state vector to output the posterior state estimate vector. The sensor drift correction unit estimates the vector from the posterior state. Extract the estimated offset voltage component The sensor drift correction unit uses the estimated offset voltage component to perform a subtraction operation on the original pH measurement value to obtain the corrected pH value. The correction formula is expressed as: ; In the formula, This represents the corrected pH value that eliminates sensor potential drift error; Representative at the Each discrete time step represents the raw pH measurement value acquired by the transistor pH sensor to be processed. This indicates the estimated offset voltage component.

[0043] The sampling timing alignment unit receives each process parameter and corrected pH value after noise isolation. Due to the different sampling periods of the sensor hardware, data points are misaligned on the time axis. The sampling timing alignment unit deploys a network time protocol within the local area network to add a uniformly standardized timestamp to all received data packets. The sampling timing alignment unit sets a uniform control step size time interval. Using linear interpolation mathematical algorithms, at a unified time point Numerical calculations are performed on the asynchronously acquired data nodes, where To align the time series numbers, after interpolation calculations, the multi-source process state parameters are mapped and concatenated into the same time series dimension, forming a discrete multidimensional vector group.

[0044] The normalization unit receives a group of multidimensional vectors. The normalization unit internally contains... A sliding data window at each time point. At any current time point... For each characteristic variable within the bromine extraction process status data, the standardized normalization unit calculates the historical mean and historical standard deviation within the sliding data window, and then performs Z-score normalization algebraic operations using the mean and standard deviation. The calculation formula is as follows: ; In the formula, Represents the current time node The unstandardized merged feature vectors; This represents the mean vector calculated from the corresponding feature variables within the sliding data window. This represents the standard deviation vector calculated for the corresponding feature variable within the sliding data window. This represents the sequence of feature vectors output by the computation. The normalization unit sends the feature vector sequence, containing dimensionally uniform data, to the system bus, completing the feature preprocessing process.

[0045] Reference Figure 1 The bromine extraction prediction module contains a principal component feature dimensionality reduction unit, a multi-head self-attention inference unit, and a multilayer perceptron mapping unit. These three units are connected in series according to the data signal flow direction.

[0046] The principal component feature reduction unit receives the feature vector sequence sent by the data fusion preprocessing module at its signal input end. The unit calculates the feature covariance matrix of the feature vector sequence, performs eigenvalue decomposition on the covariance matrix, and obtains the set of eigenvalues ​​and the corresponding set of feature vectors. The unit sorts the eigenvalues ​​from largest to smallest and selects the leading feature vectors whose cumulative variance contribution rate reaches a set threshold to form the projection matrix. The unit uses the projection matrix to perform linear projection calculation on the feature vector sequence; the dimensionality reduction formula is expressed as: ; In the formula, The principal component dimension reduction vector represents the current time step; It represents a single feature vector within the feature vector sequence at the current time step; This represents the projection matrix formed by the extracted preceding feature vectors. Multiple principal component dimensionality-reduced vectors arranged in chronological order are concatenated and combined to output the dimensionality-reduced feature sequence.

[0047] The input of the multi-head self-attention inference unit is connected to the output of the principal component feature dimensionality reduction unit. The multi-head self-attention inference unit uses the dimensionality-reduced feature sequence as input to the transformer encoder network model. The transformer encoder network model contains a multi-head self-attention layer. For the h-th attention head within the multi-head self-attention layer, the multi-head self-attention inference unit performs matrix multiplication operations on the dimensionality-reduced feature sequence with three independent linear mapping weight matrices, generating a query vector matrix, a key vector matrix, and a value vector matrix. The generation process can be expressed mathematically as follows: ; In the formula, Representing the A query vector matrix with attention heads; Representing the The key vector matrix of each attention head; Representing the A matrix of value vectors for each attention head; The input feature matrix represents the sequence of reduced-dimensional features. , and These represent the parameters of three linear mapping weight matrices obtained through backpropagation training within the multi-head self-attention layer.

[0048] The multi-head self-attention inference unit performs matrix dot product multiplication on the transposes of the query vector matrix and the key vector matrix. The result of the dot product multiplication is then divided by a scaling factor, and a normalized exponential function is used to calculate the association weight parameters between features at different time points. The multi-head self-attention inference unit then uses these association weight parameters to perform a weighted summation aggregation operation on the value vector matrix. The mathematical formula for this aggregation operation is: ; In the formula, Representing the The weighted output matrix of each attention head; Representing the A query vector matrix with attention heads; Representing the A matrix of value vectors for each attention head; The transpose of the key vector matrix; The numerical dimension parameter represents the vectors inside the key vector matrix. The result of the square root operation on the numerical dimension parameter is used as the scaling factor mentioned above. This represents the normalized exponential function. The multi-head self-attention inference unit concatenates the weighted output matrices of all attention heads along the feature dimension, multiplies the concatenated matrix with the output transformation matrix, and then outputs the hidden layer representation sequence.

[0049] The input of the multilayer perceptron mapping unit is connected to the output of the multi-head self-attention inference unit. The multilayer perceptron mapping unit contains a feedforward fully connected neural network layer. The multilayer perceptron mapping unit receives the hidden layer representation sequence and passes it to the internal feedforward fully connected neural network layer to perform weight matrix multiplication and nonlinear activation function operations. The calculation formula is expressed as follows: ; In the formula, The numerical result matrix represents the output of the multilayer perceptron mapping unit. The numerical result matrix contains a sequence of predicted bromine extraction rates in the future control time domain and the corresponding predicted bromine concentrations. The hidden layer representation sequence representing the output of the multi-head self-attention inference unit; The matrix representing the network weight parameters of a feedforward fully connected neural network layer; Represents the network bias vector; This represents a nonlinear activation function. The multilayer perceptron mapping unit establishes a mathematical mapping relationship between the hidden layer representation sequence and the reaction kinetic prediction parameters through nonlinear operations, and outputs a continuous sequence of predicted bromine extraction rates and predicted bromine concentrations to the bus.

[0050] Reference Figure 1The collaborative optimization control module is internally connected to a multi-objective cost function construction unit, an operation constraint unit, and a sequential quadratic programming solution unit. The input of the collaborative optimization control module receives the continuous bromine extraction rate prediction sequence and the predicted bromine concentration output by the bromine extraction prediction module.

[0051] The multi-objective cost function construction unit internally establishes a cost function composed of the sum of the first, second, third, and fourth penalty sub-terms. Within the defined prediction time domain, the multi-objective cost function construction unit performs summation calculations on the algebraic terms at the discrete control nodes. The mathematical formula for the cost function established by the multi-objective cost function construction unit is expressed as follows: ; In the formula, This represents the total calculated value of the cost function; This represents the total number of discrete control steps in the prediction time domain; This represents the current discrete control step sequence number; Represents airflow rate command. That is, the squared term of the airflow rate command and the first weighting coefficient. The product of these terms is used as the first penalty term; This represents setting the target bromine concentration. This represents the predicted bromine concentration. This is the squared term of the difference between the target bromine concentration and the predicted bromine concentration, plus the second weighting coefficient. The product of these terms is used as the second penalty term; This represents the acid addition rate instruction. That is, the squared term of the acid addition rate command and the third weighting coefficient. The product of these is used as the third penalty term; This represents the chlorine flow rate instruction. That is, the squared term of the chlorination rate command and the fourth weighting coefficient. The product of these factors is used as the fourth penalty term. The multi-objective cost function construction unit has an external data receiving interface for receiving real-time material cost and grid price data. The multi-objective cost function construction unit utilizes a linear mapping proportional function to adjust the first weighting coefficient based on real-time material cost and grid price. Second weighting coefficient Third weighting coefficient and the fourth weighting coefficient Dynamic numerical assignment is performed, increasing the first weighting coefficient during peak electricity price periods. The assigned value.

[0052] The operational constraint unit establishes a set of constraints, including system dynamics equality constraints, equipment operational boundary inequality constraints, and terminal target concentration relaxation constraints. Equipment operational boundary inequality constraints limit the absolute upper and lower limits of airflow rate commands, acid addition rate commands, and chlorination rate commands, as well as the range of rate changes per unit time for the executing components, through numerical interval expressions. Terminal target concentration relaxation constraints introduce non-negative relaxation variables at the end nodes of the prediction time domain, using mathematical inequalities... To impose restrictions, among which To relax variables and prevent the mathematical solver from encountering infeasible solutions under the constraints of equipment operation boundaries, system dynamics equality constraints require that the control sequence must satisfy the input-output algebraic mapping relationship established by the forward propagation operation of the transformer encoder network model. The operation constraint unit transmits the established set of constraints to the sequence quadratic programming solver unit.

[0053] The sequential quadratic programming solver unit receives both the cost function and the set of constraints. Within a set control period, the unit uses a sequential quadratic programming algorithm to find a set of control variable time series that minimizes the sum of the cost function under the current system state. In each iteration, the algorithm approximates the nonlinear cost function as a quadratic objective function at the current iteration point using Taylor expansion, and approximates the nonlinear constraints in the constraint set as linear constraints using first-order Taylor expansion, thus constructing a sub-quadratic programming mathematical model. The unit obtains the iterative search direction by solving the sub-quadratic programming mathematical model and calculates the iteration step size using a straight-line search procedure, repeatedly updating the iteration point position until the norm of the objective function gradient is less than the set convergence tolerance.

[0054] After completing the rolling numerical solution of the sequential quadratic programming algorithm, the sequential quadratic programming solution unit extracts the numerical components corresponding to the first control cycle from the obtained optimal control sequence and outputs the airflow rate command, acid addition rate command, and chlorination rate command at the output terminal. The collaborative optimization control module sends the above commands to the electrical execution control loop in the field, and repeats the above rolling solution and command issuance operation after acquiring new process status data in the next control cycle.

[0055] Reference Figure 1 The model incremental update module contains an absolute error monitoring unit, a historical sample caching unit, and a gradient descent parameter update unit.

[0056] The input of the model incremental update module is communicatively connected to the multi-parameter detection array module, the data fusion preprocessing module, and the bromine extraction prediction module. The historical sample cache unit internally uses a circular queue data structure with a fixed storage capacity. During each control cycle of the system's continuous operation, the historical sample cache unit receives the dimensionality-reduced feature vector sequence output by the data fusion preprocessing module and uses the actual bromine concentration with the same time-series label as the true result marker; the two are combined to form a historical experience data package. The historical sample cache unit stores the historical experience data packages in chronological order into the circular queue data structure; when the number of stored data packages reaches the capacity limit, the historical sample cache unit overwrites and discards the earliest stored historical experience data package according to a first-in-first-out (FIFO) logic, maintaining the temporal proximity characteristic of the stored samples.

[0057] The absolute error monitoring unit has a fixed-length sliding time window internally configured, which includes... A series of continuous discrete sampling nodes. The absolute error monitoring unit synchronously receives the actual bromine concentration and the predicted bromine concentration at the same timestamp, and calculates the average absolute error between the two using algebraic absolute value calculation. The formula for calculating the average absolute error is expressed as: ; In the formula, This represents the average absolute error calculated within the sliding time window. This represents the discrete sampling nodes contained within the sliding time window; Represents the sequence number of discrete sampling nodes within the sliding time window; Represents the sampling node The actual bromine concentration is detected and time-aligned by the multi-parameter detection array module. This represents the corresponding sampling node. The predicted bromine concentration is output by the bromine extraction prediction module. The absolute error monitoring unit calculates the average absolute error value for the current sliding time window and compares it with a pre-configured threshold value stored in the internal read-only memory.

[0058] The hardware trigger terminal of the gradient descent parameter update unit is connected to the output terminal of the absolute error monitoring unit. When the absolute error monitoring unit determines that the mean absolute error value is greater than a set threshold, it sends a level trigger signal to the gradient descent parameter update unit. After receiving the trigger signal, the gradient descent parameter update unit randomly extracts a specified number of historical experience data packets from the historical sample cache unit to form a training batch matrix. The gradient descent parameter update unit uses the training batch matrix as input, performs forward propagation in the mirror structure of the transformer encoder network model, and uses the mean squared error loss function to calculate the loss value between the output value of the mirror structure and the actual bromine concentration labeling result.

[0059] The gradient descent parameter update unit, based on the calculated loss value, performs backpropagation calculations using the chain rule to obtain the gradient matrix of the internal parameter matrix of the transformer encoder network model with respect to the loss value. The gradient descent parameter update unit employs an adaptive moment estimation mathematical optimization algorithm to estimate the first and second moments of the gradient matrix, thereby determining the iteration step size for the current backpropagation step. The specific iterative update formula for the network parameters is expressed as follows: ; In the formula, This represents the updated set of network parameters generated by the calculation. The parameter set contains the updated network weight parameters and feature mapping matrix parameters. This represents the old parameter set of the transformer encoder network before the update; This represents the set learning rate hyperparameter value; This represents using the training batch matrix The loss value calculated under the old parameter set; This represents the gradient operator that takes the partial derivative with respect to the old set of parameters.

[0060] The gradient descent parameter update unit iteratively executes the extraction, calculation, and differentiation steps described above within a set training cycle until the calculated loss value converges or the maximum number of iterations is reached. After training is complete, the gradient descent parameter update unit uses a direct memory access controller to write the finally generated updated network weight parameters and feature mapping matrix parameters into the parameter storage sector of the bromine extraction prediction module, completing the replacement of the network weight parameters within the transformer encoder network model. After the parameter replacement is complete, the system internally sets a reset flag, and the bromine extraction prediction module calls the replaced network weight parameters and feature mapping matrix parameters to execute the online inference calculation process for the next cycle.

[0061] Specific application examples: Figure 2 Figure (a) is a comparison curve of the system operation effect according to a specific application embodiment of the present invention, and Figure (b) is a comparison curve of the actual bromine concentration tracking and control effect.

[0062] This embodiment deploys the control system of the present invention in a large-scale seawater bromine extraction plant with a daily output of hundreds of tons along the coast. The main technological challenges faced by this plant are: the seawater temperature varies between day and night, and the tides cause drastic fluctuations in the salinity and initial bromide ion concentration of the influent. When using traditional single-loop PID control, in order to ensure the final blow-out rate, it is often necessary to rely on manual experience to set an excessively high fan speed and an excessive amount of sulfuric acid added.

[0063] After the control system of this invention intervenes, the operation process is as follows: The multi-parameter detection array module installed on the pipeline collects the state of seawater fluid in real time. Due to the corrosiveness and deposits in seawater, the transistor pH sensor begins to drift after one week of operation. At this time, the extended Kalman filter inside the data fusion preprocessing module estimates the offset voltage of the pH sensor by calling the unbiased actual bromine concentration given by the nuclear magnetic resonance analyzer and performs digital subtraction, so that the pH data obtained by the control system always remains true and reliable.

[0064] The cleaned and aligned feature vectors are then fed into the bromine extraction prediction module. The transformer encoder network model captures the nonlinear characteristic that a decrease in temperature leads to a slowdown in the reaction rate, predicting a downward trend in the bromine extraction rate over the next few cycles.

[0065] After receiving the predicted signal, the collaborative optimization control module retrieves the current grid price data. Assuming it's currently during off-peak electricity pricing at night, the multi-objective cost function construction unit automatically reduces the penalty weighting coefficient for the airflow rate command (allowing a moderate increase in fan power consumption) while maintaining a strict penalty on the sulfuric acid dosage. The sequential quadratic programming solver, under the boundary constraints of the equipment's maximum rotational speed, outputs a set of control commands to moderately increase the fan speed and maintain a stable acid pump speed, which are then sent to the field. Through this spatiotemporal collaborative adjustment, the system cleverly utilizes off-peak electricity to reduce overall operating costs while maintaining the target terminal concentration.

[0066] After three months of continuous operation, the seawater entered summer conditions, and the background water quality changed. The sliding window of the model incremental update module detected that the average absolute error between the actual bromine concentration and the predicted bromine concentration exceeded the set threshold. It then automatically extracted historical cache samples from the past two weeks, performed a backpropagation fine-tuning of the weight matrix of the transformer network in the background, and hot-replaced the new parameters in the working area, so that the system could restore its high-precision prediction capability.

[0067] Reference Figure 2 To verify the technical effectiveness of this invention, a comparative experiment was conducted in the aforementioned bromine extraction plant for 30 days (a total of 720 control cycles, once per hour). The first 15 days used the existing traditional PID cascade control system, and the latter 15 days seamlessly switched to the intelligent energy-saving control system of this invention. The statistical data are as follows: Under traditional control methods, the large hysteresis characteristic often results in a residual bromine concentration in the blown waste liquid as high as 18-20 ppm, causing the overall bromine extraction rate to hover only between 65% and 70%. The system of this invention, by virtue of the feedforward compensation of the predictive model and the collaborative linkage of multiple devices (blowers, raw material pumps, acid and alkali pumps, etc.), successfully controls the residual bromine concentration in the blown waste liquid at a low level of 12-15 ppm, thereby increasing the overall bromine extraction rate to 75%-80%.

[0068] This system utilizes a multi-objective cost function and an electricity price linkage mechanism to mitigate excessive acid addition while optimizing the operating power of the blowers. During the statistical period, the comprehensive electricity consumption per ton of bromine decreased by 8.5%, and the consumption of sulfuric acid per ton of bromine decreased by 5.2%.

[0069] Thanks to the digital drift correction of the data fusion preprocessing module, the calibration and cleaning cycle of the on-site transistor pH sensor has been extended from the original 7 days to more than 30 days, reducing the frequency of manual maintenance.

Claims

1. A data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction, characterized in that: include: The multi-parameter detection array module includes a nuclear magnetic resonance analyzer, a transistor pH sensor, a temperature and salinity probe, a redox potential probe, a pressure and flow transmitter, and a solution colorimeter, which are used to collect data on the bromine extraction process status and the actual residual bromine concentration in the waste liquid. The data fusion preprocessing module is used to receive the bromine extraction process status data, perform outlier removal, time series alignment and standardization normalization operations, and output a feature vector sequence. The bromine extraction prediction module is used to input the feature vector sequence into the internal transformer encoder network model for processing, and output a continuous bromine extraction rate prediction sequence and a predicted bromine concentration. The collaborative optimization control module is used to receive the continuous bromine extraction rate prediction value sequence, perform a cost function minimization calculation under constraints, and output gas flow rate command, acid addition rate command and chlorination rate command. The model incremental update module is used to receive the actual bromine concentration and the predicted bromine concentration, calculate the average absolute error between the actual bromine concentration and the predicted bromine concentration, and extract historical data to perform gradient descent training on the transformer encoder network model and update the parameters when the average absolute error is greater than a set threshold.

2. The data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction according to claim 1, characterized in that, The multi-parameter detection array module includes: The nuclear magnetic resonance concentration detection unit includes the nuclear magnetic resonance analyzer, which is used to collect the bromine extraction process status data, the actual bromine concentration, and the residual chlorine amount; A transistor pH detection unit, including the transistor pH sensor, is used to collect the pH value within the bromine extraction process status data; A temperature and salinity detection unit, including the temperature and salinity probe, is used to collect temperature and salinity information within the bromine extraction process status data. The multidimensional physicochemical detection unit includes the redox potential probe, the pressure and flow transmitter, and the solution colorimetric spectrometer, which are used to collect fluid oxidation potential, pipeline flow and pressure drop, and solution colorimetric verification information within the bromine extraction process status data.

3. The data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction according to claim 1, characterized in that, The data fusion preprocessing module includes: The noise localization and isolation unit uses the isolated forest algorithm to locate and isolate outliers within the bromine extraction process status data; The sensor drift correction unit uses an extended Kalman filter to construct the system state equation and observation equation, and outputs a corrected pH value. The sampling time alignment unit uses a network time protocol and an asynchronous interpolation algorithm to convert the bromine extraction process status data with different sampling frequencies, along with the corrected pH value, into the same time series. The standardization and normalization unit performs numerical normalization calculations on the data within the same time series and outputs the feature vector sequence.

4. The data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction according to claim 3, characterized in that, The sensor drift correction unit uses the actual bromine concentration as an unbiased reference benchmark to establish a state vector containing the true pH value, offset voltage, and the actual bromine concentration. It substitutes the state vector into the system state equation and the observation equation to perform state estimation calculations to obtain the estimated offset voltage. It then performs addition and subtraction operations on the pH value inside the bromine extraction process state data and the estimated offset voltage to obtain the corrected pH value.

5. The data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction according to claim 1, characterized in that, The bromine extraction prediction module includes: Principal component feature dimensionality reduction unit performs orthogonal decomposition dimensionality reduction on the feature vector sequence using principal component analysis, and outputs a dimensionality-reduced feature sequence; The multi-head self-attention inference unit inputs the dimensionality-reduced feature sequence into the multi-head self-attention layer inside the transformer encoder network model, calculates the correlation weight parameters between features at different times, and uses the correlation weight parameters to perform weighted summation and aggregation on the value vector inside the dimensionality-reduced feature sequence to obtain the hidden layer representation sequence. The multilayer perceptron mapping unit maps the hidden layer representation sequence to the continuous bromine extraction rate prediction sequence using a multilayer perceptron.

6. The data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction according to claim 5, characterized in that, The multilayer perceptron mapping unit includes a fully connected neural network layer. The multilayer perceptron mapping unit inputs the hidden layer representation sequence into the fully connected neural network layer to perform weight matrix multiplication and nonlinear activation function operations to generate a numerical sequence in the future control time domain as the continuous bromine extraction rate prediction value sequence, and simultaneously outputs the corresponding predicted bromine concentration. The predicted bromine concentration is a predicted value of the residual bromine concentration in the waste liquid after the bromine extraction process.

7. The data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction according to claim 1, characterized in that, The collaborative optimization control module includes: A multi-objective cost function construction unit establishes the cost function composed of the sum of a first penalty sub-item, a second penalty sub-item, a third penalty sub-item, and a fourth penalty sub-item; The operation constraint unit establishes a set of constraint conditions, including system dynamics equality constraints, equipment operation boundary inequality constraints, and terminal target concentration relaxation constraints, as the constraint conditions. The sequential quadratic programming solution unit uses the sequential quadratic programming algorithm to perform rolling numerical solutions on the cost function and the set of constraints within a set control period, and outputs the airflow rate command, the acid addition rate command, and the chlorination rate command.

8. The data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction according to claim 7, characterized in that, The first penalty term established by the multi-objective cost function construction unit corresponds to the product of the square term of the airflow rate command and the first weighting coefficient; the second penalty term corresponds to the product of the square term of the difference between the set target residual bromine concentration in the waste liquid and the predicted residual bromine concentration in the waste liquid and the second weighting coefficient; the third penalty term corresponds to the product of the square term of the acid addition rate command and the third weighting coefficient; and the fourth penalty term corresponds to the product of the square term of the chlorination rate command and the fourth weighting coefficient. The first weighting coefficient, the second weighting coefficient, the third weighting coefficient, and the fourth weighting coefficient are dynamically assigned values ​​based on real-time material costs and grid prices.

9. The data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction according to claim 1, characterized in that, The model incremental update module includes: The absolute error monitoring unit receives the actual bromine concentration and the predicted bromine concentration, and calculates the average absolute error between the actual bromine concentration and the predicted bromine concentration within a set time window. A historical sample cache unit stores historical experience data packages containing multi-source aligned feature vectors and the actual bromine concentration labeling results; The gradient descent parameter update unit extracts the historical experience data package when the mean absolute error value is greater than a set threshold, performs gradient descent training on the transformer encoder network model, and replaces and updates the network weight parameters inside the transformer encoder network model.

10. The data-driven intelligent energy-saving control system for ultra-large-scale air-blowing alkaline bromine extraction according to claim 9, characterized in that, The gradient descent parameter update unit inputs the extracted historical experience data package into the transformer encoder network model for backpropagation calculation. Based on the calculated gradient value, it modifies the network weight parameters and feature mapping matrix parameters inside the transformer encoder network model. After the modification is completed, the network weight parameters and feature mapping matrix parameters are directly replaced in the bromine extraction prediction module to perform online inference calculation.

Images