Power plant concentrated water intelligent processing system

Abstract

The present application relates to the technical field of water treatment control, and discloses an intelligent concentrated water treatment system for power plants, which comprises a pretreatment dosing module, a membrane concentration module, a concentrated water buffer module, an MVR crystallization module and a central control module; the central control module acquires the running state data of each physical module to construct a state vector, inputs a long short-term memory network model to output a membrane pollution deterioration rate factor and an inorganic salt fouling probability value; then, a model prediction control objective function is constructed based on the deterioration rate factor, a control instruction sequence is generated by rolling solution and is executed; the system is internally provided with an online dilution control logic based on the apparent viscosity of the crystal slurry and a scale prevention feedforward intervention mechanism based on double thresholds. The present application combines data prediction with a prediction control model, converts conventional feedback control into feedforward adaptive adjustment, prolongs the running cycle of the membrane assembly, reduces the risk of crystallization pipeline blockage and the comprehensive operation cost of the system.

Description

Technical Field

[0001] This invention relates to the field of water treatment control technology, specifically to an intelligent concentrated water treatment system for power plants. Background Technology

[0002] In existing power plant concentrate treatment processes, a physical flow-sheet architecture combining pretreatment chemical dosing, reverse osmosis membrane concentration, and evaporation crystallization is adopted. In actual operation, the underlying actuators of each process section mostly adopt independent single-loop feedback control modes. This control logic mainly relies on the comparison of a single sensor value with a set threshold, which is a post-event response mechanism. Since fluctuations in raw water quality and quantity will have a cumulative effect within the system, this segmented and lagging control method cannot predict the deterioration trend of the reverse osmosis membrane module in advance. When the system detects an increase in transmembrane pressure or a decrease in permeate flux and triggers an alarm, irreversible inorganic salt scaling or organic colloidal fouling has already formed on the membrane surface.

[0003] Meanwhile, in the final evaporation and crystallization stage, the apparent viscosity of the crystal slurry changes dynamically as the concentration process progresses. Existing treatment systems lack a linkage mechanism between the rheological state at the crystallization end and the concentration load of the front-end membrane. When the viscosity exceeds the limit inside the crystallizer, the system cannot spontaneously call upon the internal low-concentration fluid for intervention, increasing the risk of physical blockage in the crystallization pipeline and pump valves. In addition, existing anti-scaling chemical dosing and cleaning procedures are mostly executed according to fixed time cycles or flow ratios, failing to establish a mathematical correlation with the real-time fouling probability and hydraulic residence time of the equipment. This results in the dosage of chemicals frequently deviating from actual needs, increasing the overall operating energy consumption and chemical costs of the treatment system. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides an intelligent power plant concentrate treatment system, which solves the problems of control lag, membrane scaling, and crystallization pipeline blockage caused by the independent feedback of each process section and the lack of a prediction mechanism in existing power plant concentrate treatment systems.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a smart power plant concentrate treatment system, comprising:

[0006] The pretreatment dosing module is used to receive raw water from the power plant and adjust the hardness and alkalinity of the water by adding chemical agents; The membrane concentration module has its liquid input end connected to the liquid output end of the pretreatment dosing module, and is used to concentrate and separate water through reverse osmosis to produce fresh water and concentrated water. The concentrate buffer module has its liquid input end connected to the concentrate output end of the membrane concentration module. It is used to store the concentrate discharged from the membrane concentration module and to regulate the influent flow rate and water quality fluctuations. The MVR crystallization module has a liquid input end connected to the liquid output end of the concentrate buffer module, which is used to perform mechanical vapor recompression evaporation crystallization treatment on the concentrate and complete solid-liquid separation. The central control module establishes communication connections with the pretreatment dosing module, the membrane concentration module, the concentrate buffer module, and the MVR crystallization module, respectively, to synchronously acquire the operating status data of each physical module at discrete time steps and construct a global system state vector corresponding to the current moment based on the operating status data. The global system state vector is input into a pre-trained long short-term memory network model, which outputs a membrane fouling degradation rate factor and an inorganic salt fouling probability value. Based on the membrane fouling degradation rate factor, a model predicts the control objective function. The model predicts the minimum value of the control objective function in the control time domain to generate a control command sequence, and the control command sequence is sent to the actuators corresponding to each physical module. The system calculates membrane degradation trend indicators using a long short-term memory network model and uses these indicators as calculation parameters for the model's predictive control objective function. The system control logic changes from feedback regulation after parameters exceed limits to feedforward adaptive regulation.

[0007] Furthermore, the global system state vector is generated by one-dimensional transpose and stacking of the preprocessing side state sub-vector, the membrane system side state sub-vector, and the crystallization side state sub-vector; After acquiring the operating status data, the central control module uses a moving average filtering algorithm to remove high-frequency burst noise. Then, based on the historical upper and lower limits of various physical parameters, it maps the operating status data to a preset dimensionless value range using a maximum and minimum value normalization method. The central control module extracts the membrane system side state vectors corresponding to multiple consecutive discrete time steps within the time sliding window and combines them to form a two-dimensional temporal feature matrix, and uses the two-dimensional temporal feature matrix as the basic input data stream of the pre-trained long short-term memory network model; The above steps eliminate the interference of parameter dimension differences between different process sections on network weights through filtering and data normalization. The temporal combination method meets the input format requirements of the long short-term memory network model.

[0008] Furthermore, the pre-trained long short-term memory network model adopts a structural topology that includes an input layer, multiple long short-term memory hidden layers, and a two-branch fully connected output layer. The two-dimensional temporal feature matrix is ​​passed through the input layer to the long short-term memory hidden layer to extract the associated feature weights, and the hidden state feature vector is output at the end of the long short-term memory hidden layer. The hidden state feature vector is synchronously distributed to the numerical prediction branch structure and the class probability branch structure. The numerical prediction branch structure is used to map and output the predicted value of the membrane fouling index at the corresponding future prediction time; the category probability branch structure is used to directly output a probability distribution vector with a dimension of 2, and the two numerical components of the probability distribution vector indicate the probability of inorganic salt fouling and the probability of organic colloidal fouling, respectively. The Long Short-Term Memory Network model adopts a dual-branch output structure and shares the hidden layer state vector. In a single forward propagation calculation, it simultaneously outputs continuous membrane fouling index predictions and discrete fault classification probabilities, reducing the computational load of the model and maintaining the correlation of the results features.

[0009] Furthermore, the membrane fouling degradation rate factor is calculated as follows: The central control module extracts the permeate flux and transmembrane pressure difference parameters corresponding to the current time step, and calculates the actual membrane fouling index at the current time according to the reverse osmosis membrane permeability calibration rules. Calculate the difference between the predicted membrane fouling index at the future prediction time output by the numerical prediction branch structure and the actual membrane fouling index. The membrane fouling degradation rate factor is calculated by dividing the obtained difference by the preset future prediction step time span parameter. By calculating the difference between the predicted and measured values ​​and dividing it by the time span parameter, the rate of change of the membrane module state is quantified and used as a constraint condition for the operation of the control algorithm.

[0010] Furthermore, the model predicts the control objective function by algebraically summing and integrating the cumulative cost over the prediction time span with the control increment smoothing constraint term to construct a complete discrete global dynamic objective function. The accumulated cost is based on the integral sum of the single-step operating cost within the system's predicted time domain over time steps; the single-step operating cost is composed of the system's dosing cost evaluation function, operating power consumption evaluation function, and dynamic state penalty term. The objective function integrates power consumption, drug consumption, and penalty terms to solve for the control command sequence that balances the requirements for stable system operation and economic cost.

[0011] Furthermore, the control command sequence includes the frequency of alkaline agent addition, the frequency of acidic agent addition, the frequency of scale inhibitor addition, the operating frequency of the high-pressure variable frequency pump, and the speed of the steam compressor. The central control module extracts the alkaline agent dosing frequency, acidic agent dosing frequency, and scale inhibitor dosing frequency to calculate the system dosing cost evaluation function, and extracts the high-pressure variable frequency pump operating frequency and steam compressor speed to calculate the operating power consumption evaluation function; The central control module substitutes the membrane fouling degradation rate factor into a preset exponential function to calculate a dynamic penalty coefficient, and multiplies the dynamic penalty coefficient with the operating frequency parameter of the high-pressure variable frequency pump to construct the dynamic state penalty term; the central control module calls the sequential quadratic programming solver algorithm to solve for the minimum value of the global dynamic objective function; The above steps convert the membrane fouling degradation rate factor into a penalty term using an exponential function; when membrane fouling is predicted, the penalty term increases, and the system reduces the concentration load and slows down membrane fouling by limiting the operating frequency of the high-pressure variable frequency pump.

[0012] Furthermore, the central control module reads the apparent viscosity value of the slurry contained in the global system state vector in real time at continuous time steps; When the apparent viscosity of the slurry is greater than or equal to the safety rheological critical constant pre-cured in the controller memory, the central control module outputs a high-level trigger control word to activate the reverse safety constraint and online dilution control logic within the system, and applies physical operation boundary restrictions to the optimization solution space of the model predictive control algorithm. Set the apparent viscosity judgment condition for the crystal slurry to prevent the output of instructions that cause the viscosity inside the crystallizer to exceed the limit during model optimization, and avoid blockage of pipes and equipment caused by excessive viscosity.

[0013] Furthermore, the specific execution steps of the reverse security constraint and online dilution control logic include: Within the control time domain, the upper limit of the steam compressor speed of the MVR crystallization module is limited, and a forced derating command to reduce the frequency of the high-pressure variable frequency pump is simultaneously issued to the membrane concentration module. The membrane concentration module continuously generates low-concentration dilute and concentrated water under low-frequency operation and discharges it into the concentrated water buffer module. The concentrated water buffer module pumps the low-concentration dilute and concentrated water after physical displacement mixing into the thermal circulation pipeline of the MVR crystallization module for rheological gradation dilution until the apparent viscosity of the crystal slurry is lower than the lower limit reset value for multiple consecutive time steps, at which point the reverse safety constraint command is revoked. The low-concentration concentrate generated by the membrane concentration module under frequency reduction operation is introduced into the crystallization module for dilution, thereby reducing the viscosity of the crystal slurry inside the system and completing the system's own water regulation.

[0014] Furthermore, when the membrane fouling degradation rate factor is greater than or equal to the set degradation rate warning threshold and the inorganic salt scaling probability value is greater than or equal to the set inorganic salt scaling confidence threshold, the central control module activates the anti-scaling self-healing intervention dual feedforward trigger flag in the system's underlying program. The central control module truncates the optimization iterative output path of the model predictive control objective function and sends a pulse dosing command sequence to the pretreatment dosing module, specifically stopping the dosing of alkaline agents and adding acidic agents and scale inhibitors at the maximum rated dosage. The anti-scaling process is triggered by a combination of rate and probability thresholds, pausing model prediction and solving, increasing the dosage of acidic agents and scale inhibitors in the system, and cleaning the deposits on the membrane surface.

[0015] Furthermore, the time during which the pretreatment dosing module executes the pulse dosing command sequence is the intervention time window. The threshold length of the intervention time window is dynamically calculated and calibrated based on the hydraulic retention time obtained by dividing the total internal volume of the membrane concentration module by the current influent flow rate and multiplying it by a specific dimensionless constant multiple. When the cumulative intervention time reaches the calculated threshold of the intervention time window length, the central control module clears the physical activation state of the anti-scaling and self-healing intervention dual feedforward trigger flag and resumes issuing the optimal control command sequence generated based on the model predictive control objective function. The intervention execution time is calculated based on the system influent flow rate and membrane volume, the dosage of cleaning agent is controlled, and the original predictive control process of the system is restored after the intervention is completed.

[0016] This invention provides an intelligent concentrated wastewater treatment system for power plants. It has the following beneficial effects: 1. This invention combines a long short-term memory network model with a model predictive control algorithm to introduce the membrane fouling degradation rate factor into the global dynamic objective function. This design changes the traditional feedback control method that relies on a single physical quantity threshold. The system can dynamically adjust the operating frequency of the high-pressure variable frequency pump and the dosage of chemical reagents according to the predicted membrane fouling trend. This mechanism completes feedforward adaptive adjustment before the membrane module becomes severely fouled, effectively extending the continuous operation cycle of the reverse osmosis membrane and reducing the system's operating power consumption.

[0017] 2. This invention introduces a reverse safety constraint based on the apparent viscosity of the crystal slurry and an online dilution control logic. When the system determines that the internal viscosity of the crystallizer reaches a preset critical constant, the central control module forcibly reduces the frequency of the high-pressure variable frequency pump of the front-end membrane concentration module to generate low-concentration dilute water, which is then pumped to the end MVR crystallization module for rheological gradation dilution. The operation utilizes the material characteristic differences before and after the process section within the system to solve the risk of pipeline blockage caused by high viscosity at the end, and avoids disrupting the original water balance of the system due to the introduction of external fresh water sources.

[0018] 3. This invention establishes a dual feedforward intervention mechanism based on the membrane fouling degradation rate factor and the probability value of inorganic salt scaling, and issues pulse dosing commands in conjunction with a dynamically calibrated intervention time window. After confirming the scaling trend, the system truncates the model prediction optimization calculation and instead executes a short-term high-concentration acid agent and scale inhibitor dosing, and the dosing duration is directly related to the hydraulic residence time under the current influent flow rate. The mechanism can complete the dissolution and stripping of the attached substances in the early stage of crystallization, which not only avoids the waste of agents caused by conventional cleaning with a fixed duration, but also ensures a smooth handover of the control process after the abnormal state is handled. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the system architecture of the present invention; Figure 2 This is a schematic diagram of the method flow of the present invention; Figure 3 This is a schematic diagram of the system state perception and vectorized data flow of the present invention; Figure 4 This is a schematic diagram of the membrane state temporal prediction and classification network structure of the present invention; Figure 5 This is a schematic diagram of the global dynamic optimization control logic flow of the present invention; Figure 6 Figure 1 shows the hysteresis constraint and de-capacity elution time-series response curves based on rheological state of the present invention. Figure 2 shows the rheological state monitoring and hysteresis interval of the MVR crystallization module, Figure 3 shows the operating frequency of the high-pressure variable frequency pump, and Figure 4 shows the speed of the steam compressor. Figure 7 The following is a diagram of the water quality driven dual feedforward triggering and scale prevention self-healing pulse intervention action of the present invention. In the diagram, (a) is the monitoring of membrane fouling degradation rate factor, (b) is the monitoring of inorganic salt scaling probability component, (c) is the Boolean control word of bottom layer intervention flag bit, and (d) is the execution action diagram of metering pump. Figure 8 This is a time series comparison diagram of the transmembrane pressure difference of the present invention; Figure 9 This is a scatter plot showing the power consumption of the compressor and the apparent viscosity of the slurry in this invention.

[0020] Among them, 10 is the pretreatment dosing module; 20 is the membrane concentration module; 30 is the concentrate buffer module; 40 is the MVR crystallization module; and 50 is the central control module. Detailed Implementation

[0021] 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.

[0022] See attached document Figure 1 This invention provides an intelligent power plant concentrate treatment system, the system comprising: The pretreatment dosing module 10 is used to receive raw water from the power plant and adjust the hardness and alkalinity of the water by adding chemical agents. The liquid output end of the pretreatment dosing module 10 is connected to the liquid input end of the membrane concentration module 20. The membrane concentration module 20 is used to concentrate and separate water through reverse osmosis to produce fresh water and concentrated water, and the concentrated water output end of the membrane concentration module 20 is connected to the liquid input end of the concentrated water buffer module 30. The concentrate buffer module 30 is used to store the concentrate discharged from the membrane concentration module 20 and to regulate the influent flow rate and water quality fluctuations. The liquid output end of the concentrate buffer module 30 is connected to the liquid input end of the MVR crystallization module 40. MVR crystallization module 40 is used to perform mechanical vapor recompression evaporation crystallization treatment on the concentrated water supplied by concentrated water buffer module 30 and to complete solid-liquid separation. The central control module 50 establishes communication connections with the pretreatment dosing module 10, the membrane concentration module 20, the concentrate buffer module 30, and the MVR crystallization module 40, respectively, to synchronously collect the operating status data of each physical module, and calculate and output control commands based on the preset algorithm model to send to the corresponding actuators of each physical module.

[0023] See attached document Figure 2 This invention provides a control method for a power plant concentrate intelligent treatment system, the method comprising the following steps: S1. The central control module 50 synchronously acquires the operating status data of the pretreatment dosing module 10, the membrane concentration module 20 and the MVR crystallization module 40 at discrete time steps, and constructs the global system state vector corresponding to the current moment based on the operating status data. S2. The central control module 50 inputs the state subsequence of the membrane concentration module 20 within the continuous historical time window into the pre-trained long short-term memory network model. The long short-term memory network model forward propagates and outputs the membrane fouling degradation rate factor at the future predicted time, as well as the probability distribution of the current membrane fouling type being inorganic salt fouling. S3. The central control module 50 constructs a model predictive control objective function with the goal of minimizing the operating cost of each module. The membrane fouling degradation rate factor extracted in step S2 is introduced into the model predictive control objective function as a dynamic penalty term parameter. The central control module 50 solves the minimum value of the model predictive control objective function in the control time domain to generate a sequence of control commands, which are then sent to the pretreatment dosing module 10, the membrane concentration module 20, and the MVR crystallization module 40. S4. The central control module 50 monitors the apparent viscosity of the crystal slurry in the MVR crystallization module 40 in real time. When the apparent viscosity of the crystal slurry exceeds the preset safe rheological critical constant, the central control module 50 applies constraints to the model predictive control solution domain to limit the upper limit of the compressor speed of the MVR crystallization module 40. At the same time, it sends a forced reduction command to the membrane concentration module 20 to reduce the frequency of the high-pressure variable frequency pump. The low-concentration concentrate generated by the membrane concentration module 20 enters the MVR crystallization module 40 through the concentrate buffer module 30 for online dilution and elution until the apparent viscosity of the crystal slurry is lower than the preset threshold. S5. When the membrane fouling degradation rate factor output in step S2 is greater than the set warning threshold and the inorganic salt scaling probability is greater than the set confidence threshold, the central control module 50 blocks the regular dosing command issued to the pretreatment dosing module 10 in step S3. Within the set intervention time window, the pretreatment dosing module 10 executes a pulse dosing command sequence, specifically stopping the dosing of alkaline agents and adding acidic agents and scale inhibitors at the maximum rated dosing amount. The resulting slightly acidic influent enters the membrane concentration module 20 to dissolve the inorganic salt crystals on the membrane surface. After the intervention time window ends, the central control module 50 resumes issuing the optimization control command of step S3.

[0024] See attached document Figure 3 Step S1, in its specific implementation, includes the following steps: S101, the central control module 50 establishes a communication link with each distributed control unit on site through the industrial communication protocol. The central control module 50 sets a unified discrete time step and sends synchronous sampling instructions to the sensors and equipment frequency converters distributed in each physical process pipeline. The central control module 50 performs time stamp synchronization management for data acquisition time alignment of multi-source heterogeneous devices based on the timing calibration method of the edge computing gateway; After acquiring the raw physical data, the central control module 50 uses a moving average filtering algorithm to remove high-frequency burst noise. Then, based on the historical upper and lower limits of each physical parameter, it maps the raw physical data to a preset dimensionless value range using a maximum and minimum value normalization method. The preset dimensionless value range is from 0 to 1.

[0025] S102. After completing the spatiotemporal alignment and filtering normalization processing, the central control module 50 acquires the real-time operating status data of the pretreatment dosing module 10. The specific data acquisition process includes: The influent flow rate is obtained by an electromagnetic flow meter deployed on the main influent pipe, the raw water hardness and alkalinity are obtained by an online titration water quality analyzer deployed on the influent pipeline, and the dosing frequency of alkaline agents, acidic agents and scale inhibitors are obtained by reading the feedback signals of the frequency converters that control each dosing pump. After obtaining the above physical quantities, the central control module 50 will sequentially arrange the influent flow rate, raw water hardness, raw water alkalinity, alkaline agent dosing frequency, acidic agent dosing frequency and scale inhibitor dosing frequency at the corresponding time step to construct a pretreatment side state sub-vector. The pretreatment side state sub-vector sequence serves as an ordered data set to characterize the basic material input attributes of the pretreatment dosing module 10.

[0026] S103, the central control module 50 synchronously acquires real-time operating status data of the membrane concentration module 20 and the concentrate buffer module 30 based on the characteristics of the influent. The specific data acquisition process includes: The high-pressure variable frequency pump operating frequency is obtained by reading the communication register of the high-pressure pump driver; the transmembrane pressure difference is obtained by subtracting the values ​​of the pressure transmitters distributed at the inlet and concentrate ends of the reverse osmosis membrane module; the permeate flux is obtained by the flow meter in the permeate pipeline; the salinity on the concentrate side is obtained by converting the online conductivity meter at the concentrate end; and the buffer level is obtained by the level transmitter installed inside the concentrate buffer module 30. The central control module 50 constructs a membrane system-side state vector based on the acquired high-pressure variable frequency pump operating frequency, transmembrane pressure difference, permeate flux, concentrate-side salinity, and buffer level. The membrane system-side state vector maps the fluid dynamics response characteristics of the membrane concentration module 20 at a specific moment to the material separation equilibrium relationship.

[0027] S104. The central control module 50 acquires real-time operating status data of the MVR crystal module 40. The specific data acquisition process includes: The steam compressor speed is obtained by reading data from the steam compressor control cabinet, and the crystal slurry circulation concentration is obtained by an online density meter installed on the crystallizer circulation pipeline; The central control module 50 collects the operating current of the crystallization circulation pump motor and calculates the apparent viscosity of the crystal slurry based on the mapping model between motor torque and fluid apparent viscosity. As the viscosity of the crystal slurry increases, the fluid resistance of the pump impeller rotation increases, and the output torque of the drive motor increases synchronously with the operating current. The central control module 50 executes the anti-scaling failure logic by using the above-mentioned electrical parameters to calculate the rheological state. The central control module 50 integrates the steam compressor speed, crystal slurry circulation concentration, and crystal slurry apparent viscosity to construct a crystallization-side state vector.

[0028] S105. After completing the state extraction, the central control module 50 performs a matrix merging and concatenation operation on the sub-vectors of the three spatial dimensions based on the same time step identifier. The specific operation includes: The pretreatment-side state vector, the membrane system-side state vector, and the crystallization-side state vector are transposed and stacked in one dimension to generate a global system state vector. The global system state vector reflects the comprehensive characteristics of the flow field morphology and thermodynamic cross-section of the entire physical process at a specific moment. The central control module 50 stores the global system state vectors generated at different discrete time steps into the database in a time sequence to form a continuous time series data matrix; The formula for concatenating the global system state vector is shown below: ; In the formula: For time step The global system state vector; For the discrete time step variables of the system; For time step The preprocessed side state sub-vectors; For time step The membrane system side state vector; For time step The state vector of the crystallization side; A mathematical symbol used to represent the transpose of a vector or matrix.

[0029] See attached document Figure 4 Step S2, in its specific implementation, includes the following steps: S201, the central control module 50 extracts a fixed-length time sliding window from the continuous time series data matrix based on the set historical backtracking step size parameter. The set historical backtracking step size parameter ranges from 12 to 24 discrete time steps. The central control module 50 extracts the membrane system side state vectors corresponding to multiple consecutive discrete time steps within the time sliding window and combines them to form a two-dimensional time series feature matrix. The number of rows in the two-dimensional time series feature matrix corresponds to the number of historical backtracking time steps, and the number of columns in the two-dimensional time series feature matrix corresponds to the physical parameter dimension of the membrane system side state vectors. The central control module 50 uses a two-dimensional time-series feature matrix composed of the high-pressure variable frequency pump operating frequency, transmembrane pressure difference, and water flux as the basic input data stream for the pre-constructed long short-term memory network model.

[0030] S202, The long short-term memory network model deployed in the central control module 50 adopts a structural topology that includes an input layer, multiple long short-term memory hidden layers, and a double-branch fully connected output layer; The two-dimensional temporal feature matrix is ​​passed to the long short-term memory hidden layer through the input layer. The forget gate, input gate and output gate inside the long short-term memory hidden layer extract the correlation feature weights of the membrane operation parameters over the time span according to the temporal gating calculation rules. The Long Short-Term Memory (LSTM) network model outputs a hidden state feature vector at the end of the LSM hidden layer through forward propagation. The hidden state feature vector represents the implicit numerical information of the historical water permeability degradation trend of the reverse osmosis membrane. The central control module 50 synchronously distributes the hidden state feature vectors to the numerical prediction branch structure and the class probability branch structure at the back end of the long short-term memory network model.

[0031] S203. In the numerical prediction branch structure, the hidden state feature vector is mapped through a fully connected layer containing a linear activation function to output the predicted value of the membrane fouling index at the corresponding future prediction time. The central control module 50 extracts the permeate flux and transmembrane pressure difference parameters corresponding to the current time step. The central control module 50 calculates the actual membrane fouling index at the current time according to the reverse osmosis membrane permeability calibration rules. The central control module 50 calculates the difference between the predicted membrane fouling index output by the long short-term memory network model and the actual membrane fouling index. The central control module 50 divides the obtained difference by the preset future prediction step time span parameter to calculate the membrane fouling degradation rate factor. The preset future prediction step time span parameter ranges from 3 to 6 discrete time steps. The membrane fouling degradation rate factor characterizes the deterioration acceleration of the micropore channels inside the membrane concentration module 20 due to physical blockage by entrapment. The formula for calculating the membrane fouling degradation rate factor is as follows: ; In the formula: For time step The membrane fouling degradation rate factor obtained from the evaluation; For the discrete time step variables of the system; The corresponding future prediction time output of the forward propagation of the Long Short-Term Memory network model The predicted value of the membrane fouling index; To utilize the corresponding time step The current actual membrane fouling index is calculated from the measured operating data. The preset time span parameter for future prediction steps.

[0032] S204. In the category probability branch structure, the hidden state feature vector is processed by another set of parallel fully connected layers to extract classification mapping features. The central control module 50 extracts the preprocessing side state sub-vector containing the raw water hardness characteristics from the global system state vector. The central control module 50 then uses the preprocessing side state sub-vector as a physical feedforward constraint and splices it to the input endpoint of the fully connected layer. The spliced ​​multidimensional fusion feature vector is mapped to a normalized probability space with a sum of 1 by the Softmax classification activation function. The long short-term memory network model directly outputs a probability distribution vector with a dimension of 2 at the end of the category probability branch structure. The two numerical components of the probability distribution vector indicate the probability of inorganic salt fouling and organic colloidal fouling on the surface of the current membrane concentration module 20, respectively.

[0033] S205. During the training of the Long Short-Term Memory Network model, a training sample set is generated by randomly selecting slices of continuously running data from the historical database stored in the central control module 50. Based on the historical physical cleaning records of the membrane concentration module 20 and the laboratory water quality reports, add continuous numerical labels of membrane fouling index at corresponding time nodes and discrete monothermic coding labels that characterize the actual scaling or fouling type to the training sample set. The central control module 50 uses the mean square error function to calculate the regression loss between the predicted value of the branch output and the continuous value label, and the central control module 50 uses the cross entropy function to calculate the classification probability of the branch output and the classification loss between the one-hot encoded label. The central control module 50 constructs a multi-task joint loss function by weighting and summing the regression loss and classification loss according to a preset ratio. The preset ratio ranges from 60% to 80% for regression loss and 20% to 40% for classification loss. The central control module 50 uses the gradient descent optimization algorithm to update the connection weights of neurons in each layer of the long short-term memory network model in reverse until the multi-task joint loss function converges numerically.

[0034] See attached document Figure 5 Step S3, in its specific implementation, includes the following steps: S301, The central control module 50 constructs a basic economic operating cost model of the physical process module in the discrete time dimension based on the physical properties of the control variables; The central control module 50 extracts the alkaline agent dosing frequency, acidic agent dosing frequency, and scale inhibitor dosing frequency from the control command sequence. The central control module 50 multiplies the three dosing frequencies by the corresponding preset chemical agent unit price coefficients to calculate the system dosing cost evaluation function. The chemical agent unit price coefficients are calculated based on the on-site purchase contract amount and the dosing volume of the metering pump in a single stroke. The central control module 50 extracts the high-pressure variable frequency pump operating frequency and the steam compressor speed from the control command sequence. Based on the fluid machinery similarity law, the central control module 50 multiplies the cubic value of the high-pressure variable frequency pump operating frequency and the cubic value of the steam compressor speed by the corresponding power consumption conversion coefficients to calculate the operating power consumption evaluation function of the membrane concentration module 20 and the MVR crystallization module 40. The power consumption conversion coefficient is calibrated by fitting the constant of the characteristic curve of the pump valve pipeline resistance and the rated power on the motor nameplate.

[0035] S302, the central control module 50 extracts the membrane fouling degradation rate factor output by the long short-term memory network model. The central control module 50 uses the membrane fouling degradation rate factor as an independent variable and substitutes it into the preset exponential function to calculate and generate a dynamic penalty coefficient. The base of the preset exponential function is set as a natural constant or a positive real number within a specific range based on the experience curve of irreversible fouling development of reverse osmosis membranes. The value range of the positive real number within the specific range is 1.2 to 1.5. The central control module 50 multiplies the dynamic penalty coefficient with the high-pressure variable frequency pump operating frequency parameter to construct a dynamic state penalty term. The value of the dynamic state penalty term increases exponentially with the rate of increase of membrane surface fouling. The central control module 50 uses the sum of a polynomial composed of the dosing cost evaluation function, the operating power consumption evaluation function, and the dynamic state penalty term as the single-step operating cost in the system prediction time domain. The dynamic state penalty term intervenes and changes the equipment operation optimization trend generated by power consumption accounting, so that the target evaluation cost of high pollution risk equipment in the system model increases.

[0036] S303, the central control module 50 integrates all single-step running costs within the set model predictive control prediction time domain span. The set model predictive control prediction time domain span ranges from 10 to 15 control steps. The central control module 50 introduces a control increment penalty diagonal matrix and accumulates the product of the control increment penalty diagonal matrix and the actuator action change between adjacent control steps to generate a control increment smoothing constraint term; the internal elements of the control increment penalty diagonal matrix are assigned values ​​based on the mechanical fatigue life limit of the field equipment, and the internal elements of the control increment penalty diagonal matrix restrict the frequency jitter action of the corresponding physical actuator. The central control module 50 integrates the cumulative cost over the predicted time span with the smoothing constraint term of the control increment through algebraic addition to construct a complete discrete global dynamic objective function. The mathematical expression of the global dynamic objective function is constructed based on the standard form of discrete-time extrapolation and rolling optimization, and consists of two parts: the total cost and the total control increment. The mathematical expression for the global dynamic objective function is shown below: ; In the formula: To predict the global economic objective function in the time domain; The index variable is accumulated for the prediction step size in the prediction time domain; For predicting the time domain length parameter; To step at the current time The next predicted future Step-by-step cost evaluation function for drug administration; For the current discrete time step variable of the system; To step at the current time The next predicted future A step-by-step evaluation function for the operating power consumption of the membrane system; To step at the current time The next predicted future The power consumption evaluation function for the crystallization system in step; The dynamic penalty coefficient is calculated and generated based on the membrane fouling degradation rate factor. For time steps The membrane fouling degradation rate factor obtained from the evaluation; To step at the current time The next predicted future The operating frequency of the high-voltage variable frequency pump; To accumulate index variables for the control step size within the control time domain; To control the time domain length parameter; The control increment penalty diagonal matrix characterizing control smoothness; For time steps The control increment of the control instruction sequence.

[0037] S304, the central control module 50 calls the sequence quadratic programming solver algorithm to find the optimal control command sequence that minimizes the global dynamic objective function within the physical operation constraint boundary of the current system state; the physical operation constraint boundary covers the minimum flow frequency threshold and maximum rated frequency threshold of the variable frequency pump, as well as the upper and lower limits of the compressor speed; When the membrane fouling degradation rate factor is too large, causing the dynamic state penalty term of the objective function to increase, the sequential quadratic programming solver algorithm generates the optimal control solution by reducing the frequency of the variable frequency pump and increasing the speed of the compressor based on the gradient descent logic. The sequential quadratic programming solver algorithm drives the fluid process system to complete the cross-phase transfer of physical water treatment load from membrane separation to thermal evaporation. The central control module 50 extracts the first item of the optimal control instruction sequence as the current actual control instruction to be issued to the underlying mechanism. In the next time step, the central control module 50 repeatedly executes the rolling optimization calculation process based on the latest global system state vector.

[0038] See attached document Figure 6 Step S4, in its specific implementation, includes the following steps: S401, the central control module 50 reads the apparent viscosity value of the crystal slurry contained in the global system state vector in real time under continuous time steps, and the central control module 50 performs a numerical comparison calculation with the extracted apparent viscosity value of the crystal slurry and the safety rheological critical constant pre-cured in the controller memory. The critical rheological constant for safety is calibrated offline based on the cross-sectional area of ​​the heat exchange tubes inside the MVR crystallization module 40 and the mechanical parameters of the circulating pump head limit, and is a fixed constant within a specific Pascal-second range, which ranges from 1.5 Pascal-seconds to 2.5 Pascal-seconds. The central control module 50 has a built-in logic judgment unit. When the actual apparent viscosity of the crystal slurry is greater than or equal to the safety rheological critical constant, it outputs a high-level trigger control word. The high-level trigger control word activates the reverse safety constraint and online dilution control logic inside the system.

[0039] The specific mathematical expression of the logical decision-triggered operation is shown below: ; In the formula: For the current time step The safety constraint trigger Boolean control word output by the lower central control module; For the discrete time step variables of the system; For time steps The real-time apparent viscosity of the slurry contained in the global system state vector; It is the safety rheological critical constant pre-programmed into the controller memory.

[0040] S402. After receiving the high-level trigger status signal, the central control module 50 applies physical operation boundary constraints to the optimization solution space of the model predictive control algorithm. The central control module 50 overwrites the prediction instructions output by the sequential quadratic programming solver algorithm to the MVR crystallization module 40, and limits the upper limit of the steam compressor speed in the control time domain; the central control module 50 simultaneously issues a forced derating instruction to the membrane concentration module 20 to reduce the frequency of the high-pressure variable frequency pump. The forced derating command bypasses the iterative optimization process of the economic objective function. The underlying logic lowers the operating frequency of the high-voltage variable frequency pump to the lower limit protection range of 30% to 50% of the rated frequency of the equipment. The linkage action restricts the quality transfer process of continuously delivering high-load materials from the front-end process to the back-end crystallization process.

[0041] After receiving the forced reduction command, the high-pressure variable frequency pump inside the S403 membrane concentration module 20 reduces the total flow rate of the feed water and the cross-flow velocity on the surface of the reverse osmosis membrane; the decrease in the transmembrane driving pressure inside the reverse osmosis membrane module causes the physical retention separation effect to weaken and the permeate flux to drop, and the decrease in the water recovery rate on the concentrate side causes the total dissolved solids concentration inside the discharge fluid to leave the high concentration enrichment state. The membrane concentration module 20 continuously generates low-concentration dilute and concentrated water under low-frequency operation and discharges it into the concentrate buffer module 30. The high-concentration fluid inside the concentrate buffer module 30 undergoes physical mixing and replacement with the continuously flowing low-concentration dilute and concentrated water.

[0042] S404, the concentrated water buffer module 30 pumps the low-concentration dilute concentrated water after physical displacement mixing into the thermal circulation pipeline of the MVR crystallization module 40. The high-viscosity crystal slurry inside the MVR crystallization module 40 is continuously diluted by the low-concentration dilute concentrated water through rheological gradation. After low-concentration dilute and concentrated water enters the evaporation cycle system, it increases the solvent ratio of the liquid phase and widens the thermodynamic gradient space between the actual supersaturated concentration and the reference saturated concentration. The system performs high-temperature hot rinsing and elution of the crystal slurry agglomerates while maintaining the high-temperature thermal field inside the MVR crystallization module 40. The central control module 50 introduces hysteresis comparison logic including a lower limit reset value to monitor the apparent viscosity of the slurry. When the apparent viscosity of the slurry is lower than the lower limit reset value for multiple consecutive time steps, the central control module 50 cancels the reverse safety constraint command and resumes the system's rolling optimization control process. The value range of the lower limit reset value is 60% to 70% of the value of the safety rheological critical constant. When the reverse safety constraint logic and the scale prevention and self-healing intervention logic are triggered simultaneously at the same time step, the central control module 50 prioritizes the control action of the reverse safety constraint logic according to the principle that physical fluid anti-clogging is superior to chemical water quality anti-scaling.

[0043] See attached document Figure 7 Step S5, in its specific implementation, includes the following steps: S501, the central control module 50 reads the membrane fouling degradation rate factor and the inorganic salt scaling probability component in the probability distribution vector output by the long short-term memory network model during the discrete control cycle. The central control module 50 extracts the degradation rate warning threshold from the memory and compares it with the inorganic salt scaling confidence threshold. The degradation rate warning threshold is calculated and calibrated offline based on the statistical distribution mean of degradation factors during the historical normal operation cycle of the reverse osmosis membrane, plus three standard deviations. The value range of the degradation rate warning threshold is 0.05 per hour to 0.15 per hour, and the inorganic salt scaling confidence threshold is set within the probability range of 80% to 95%. The Boolean logic judgment unit built into the central control module 50 outputs a logic high level when the membrane fouling degradation rate factor is greater than or equal to the degradation rate warning threshold and the inorganic salt scaling probability component is greater than or equal to the inorganic salt scaling confidence threshold. The logic high level activates the anti-scaling self-healing intervention dual feedforward trigger flag in the system's underlying program. The specific mathematical expression of the dual feedforward triggering logic for anti-scaling and self-healing intervention is shown below: ; In the formula: For time steps The anti-scaling and self-healing intervention trigger control word; For the discrete time step variables of the system; The logic high-level state constant of the dual feedforward trigger flag for activating the anti-scaling and self-healing intervention; This is a logic low-level state constant corresponding to when the trigger condition is not met; For time steps The membrane fouling degradation rate factor obtained from the evaluation; The degradation rate warning threshold is set based on historical statistical distribution; For time steps The inorganic salt scaling probability component in the probability distribution vector output by the lower neural network; The preset confidence threshold for inorganic salt scaling; This corresponds to all other normal operating states of the system besides the aforementioned dual triggering conditions.

[0044] S502, after the central control module 50 detects that the anti-scaling and self-healing intervention dual feedforward trigger flag is activated, it cuts off the optimization iteration output path of the global dynamic objective function at the software logic layer; the central control module 50 rewrites the underlying drive instructions sent to the pretreatment dosing module 10 according to the preset strategy. The central control module 50 forces the alkaline agent dosing frequency in the control command sequence to a zero value, and simultaneously forces the acid agent dosing frequency and scale inhibitor dosing frequency to the maximum rated operating frequency boundary of the metering pump mechanically permissible. The pulse-driven forced rewrite mechanism bypasses the conventional economic optimization control network and applies the highest priority chemical dosing intervention to the physical actuators. The control switching action blocks the evolutionary trajectory of the fluid process system, which is continuously shifting towards the high-pollution risk range based on the basic operating cost model.

[0045] S503, after receiving the rewrite instruction, the pretreatment dosing module 10 injects acidic and scale inhibitors into the raw water main at high frequency and cuts off the dosing of alkaline agents. The feed water is transformed into a slightly acidic chemically unbalanced fluid under the physical mixing action of the pulsed reagent and pumped into the membrane concentration module 20. During the cross-flow scouring process on the surface of the reverse osmosis membrane, the slightly acidic fluid undergoes a chemical displacement and dissolution reaction between hydrogen ions and the calcium and magnesium inorganic salt crystal nuclei initially attached to the membrane surface. The system maintains the chemical intervention state until the cumulative pulse duration within the intervention time window reaches the preset length threshold. The length threshold of the intervention time window is dynamically calculated and calibrated based on the hydraulic residence time obtained by dividing the total internal volume of the membrane concentration module 20 by the current influent flow rate and multiplying it by a specific dimensionless constant multiple. The specific dimensionless constant multiple ranges from 1.5 times to 2.0 times. The physical calculation logic establishes the duration span required for slightly acidic water to replace the physical flow channel of the entire membrane module.

[0046] S504, the high-precision hardware timer inside the central control module 50 generates a hardware interrupt signal after the accumulated intervention time reaches the calculated intervention time window length threshold. The central control module 50 responds to the interrupt signal and executes the self-healing intervention reset program to clear the physical activation state of the anti-scaling self-healing intervention dual feedforward trigger flag. The central control module 50 cancels the pulse-type forced rewrite constraint on the underlying drive command of the pretreatment dosing module 10 and releases the control output channel. The central control module 50 resumes sending the optimal optimization control sequence, which includes the global dynamic objective function of dosing cost and equipment power consumption, to the field physical execution link. The system regression is based on the finite-time-domain rolling optimization control process of global system state vector and operating cost accounting. The central control module 50 synchronously resets the state calculation variables of the membrane fouling degradation rate factor and enters the cycle state of the next monitoring cycle.

[0047] To aid in understanding the technical solution of this invention, the following provides an application example and test example based on a zero-discharge retrofit project for circulating cooling water wastewater in a 600MW thermal power plant.

[0048] In the actual operating conditions of this thermal power plant, the concentration of calcium and magnesium ions in the circulating cooling wastewater received by the pretreatment dosing module fluctuates between 400 mg / L and 1200 mg / L. The membrane concentration module uses an anti-fouling reverse osmosis membrane module, and the MVR crystallization module uses a forced circulation evaporator crystallizer.

[0049] During the actual operation of the project, after the system starts up, the central control module synchronously acquires 105.3m data based on distributed sensors. 3 The central control module combines the physical quantities of the influent flow rate ( / h), raw water hardness (845.2 mg / L), high-pressure variable frequency pump frequency (38.4 Hz), transmembrane pressure difference (1.24 MPa), and apparent viscosity of crystal slurry (1.62 Pa·s) to form a global system state vector.

[0050] The Long Short-Term Memory (LSTM) network model continuously receives the state sequence of the membrane system. When changes in raw water quality increase the amount of residue trapped in the micropores of the membrane module, the membrane fouling degradation rate factor output by the model during forward propagation increases from 0.02 / h to 0.08 / h, and the probability of inorganic salt scaling reaches 88.5%. At this point, the central control module triggers water quality-driven anti-scaling and self-healing intervention pulse control. The system interrupts the global dynamic objective function optimization path and forces the pretreatment dosing module to stop adding sodium hydroxide solution. The operating frequency of the metering pumps for hydrochloric acid and polycarboxylic acid scale inhibitors is increased to 100% of the rated limit. A slightly acidic unbalanced fluid with a pH of 4.2 enters the membrane concentration module and undergoes a chemical displacement dissolution reaction after 45 minutes of continuous operation to remove the inorganic salt crystal nuclei attached to the membrane surface. After the intervention, the transmembrane pressure difference drops to 0.98 MPa, and the system switches back to the global dynamic optimization control process.

[0051] When the apparent viscosity of the crystal slurry in the MVR crystallization module rises to 2.65 Pa·s during a specific operating cycle, a rheological state reverse safety constraint is triggered. The central control module limits the compressor speed and forces the high-pressure variable frequency pump of the membrane concentration module to reduce its frequency to 25 Hz. The reverse osmosis permeate flux decreases, and the low-concentration dilute water with a total dissolved solids concentration of 15000 mg / L is pumped into the MVR crystallization module through the buffer module. After 1.5 hours of high-temperature hot flushing and elution, the apparent viscosity of the crystal slurry is diluted to below the lower limit reset value of 1.45 Pa·s. The system then releases the safety constraint and resumes normal operation.

[0052] To verify the actual technical effect of the present invention, a 60-day comparative experiment was conducted in a thermal power plant project. For the first 30 days, the existing PID control system on site was used as the control group, and for the last 30 days, the intelligent power plant concentrate treatment system control method provided by the present invention was used as the experimental group.

[0053] The specific experimental verification data and effect comparison records are shown in Table 1 and Table 2.

[0054] Table 1: Comparison of macroscopic operating indicators of the two control systems throughout their entire lifecycle.

[0055]

[0056] Table 2: Discrete Sampling and Tracking Table of Core State Parameters for Local Time Periods

[0057] From Table 1 and Table 2 and appendices Figure 8 and attached Figure 9It can be seen that after introducing dynamic state penalty terms and rolling optimization algorithms, the daily power consumption of the membrane system in the experimental group was reduced by 364.35 kWh / d compared with the control group, and the daily power consumption of the MVR crystallization module was reduced by 1432.32 kWh / d compared with the control group. The experimental group relied on anti-scaling self-healing intervention pulse control to control the number of offline chemical cleanings of the membrane module within the cycle to 1 time and reduce the cost of chemical agents per ton of water to 3.28 RMB / t. When encountering water quality changes on the 13th to 14th day, the apparent viscosity of the crystal slurry in the control group increased to 2.68 Pa·s and the extreme value of the membrane degradation rate assessment reached 0.134 / h. The experimental group activated three anti-scaling self-healing intervention actions in the corresponding period and constrained the peak value of the apparent viscosity of the crystal slurry to 1.84 Pa·s and suppressed the extreme value of the membrane degradation rate to within 0.091 / h. The experimental group system reduced the downtime of MVR pipeline clumping to 0 hours based on the rheological state reverse safety constraint logic.

[0058] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart concentrated wastewater treatment system for power plants, characterized in that, include: The pretreatment dosing module is used to receive raw water from the power plant and adjust the hardness and alkalinity of the water by adding chemical agents; The membrane concentration module has its liquid input end connected to the liquid output end of the pretreatment dosing module, and is used to concentrate and separate water through reverse osmosis to produce fresh water and concentrated water. The concentrate buffer module has its liquid input end connected to the concentrate output end of the membrane concentration module. It is used to store the concentrate discharged from the membrane concentration module and to regulate the influent flow rate and water quality fluctuations. The MVR crystallization module has a liquid input end connected to the liquid output end of the concentrate buffer module, which is used to perform mechanical vapor recompression evaporation crystallization treatment on the concentrate and complete solid-liquid separation. The central control module establishes communication connections with the pretreatment dosing module, the membrane concentration module, the concentrate buffer module, and the MVR crystallization module, respectively, to synchronously acquire the operating status data of each physical module at discrete time steps and construct a global system state vector corresponding to the current moment based on the operating status data. The global system state vector is input into a pre-trained long short-term memory network model, which outputs a membrane fouling degradation rate factor and an inorganic salt scaling probability value. Based on the membrane fouling degradation rate factor, a model is constructed to predict the control objective function. The minimum value of the model-predicted control objective function is solved in the control time domain to generate a control command sequence, and the control command sequence is sent to the actuators corresponding to each physical module.

2. The intelligent power plant concentrate treatment system according to claim 1, characterized in that: The global system state vector is generated by one-dimensional transpose and stacking of the preprocessing side state vector, the membrane system side state vector, and the crystallization side state vector. After acquiring the operating status data, the central control module uses a moving average filtering algorithm to remove high-frequency burst noise. Then, based on the historical upper and lower limits of various physical parameters, it maps the operating status data to a preset dimensionless value range using a maximum and minimum value normalization method. The central control module extracts the membrane system side state vectors corresponding to multiple consecutive discrete time steps within the time sliding window and combines them to form a two-dimensional temporal feature matrix. The two-dimensional temporal feature matrix is ​​then used as the basic input data stream of the pre-trained long short-term memory network model.

3. The intelligent power plant concentrate treatment system according to claim 2, characterized in that: The pre-trained long short-term memory network model adopts a structural topology that includes an input layer, multiple long short-term memory hidden layers, and a two-branch fully connected output layer. The two-dimensional temporal feature matrix is ​​passed through the input layer to the long short-term memory hidden layer to extract the associated feature weights, and the hidden state feature vector is output at the end of the long short-term memory hidden layer. The hidden state feature vector is synchronously distributed to the numerical prediction branch structure and the class probability branch structure. The numerical prediction branch structure is used to map and output the predicted value of the membrane fouling index at the corresponding future prediction time; the category probability branch structure is used to directly output a probability distribution vector with a dimension of 2, and the two numerical components of the probability distribution vector indicate the probability of inorganic salt fouling and the probability of organic colloidal fouling, respectively.

4. The intelligent power plant concentrate treatment system according to claim 3, characterized in that: The membrane fouling degradation rate factor is calculated as follows: The central control module extracts the permeate flux and transmembrane pressure difference parameters corresponding to the current time step, and calculates the actual membrane fouling index at the current time according to the reverse osmosis membrane permeability calibration rules. Calculate the difference between the predicted membrane fouling index at the future prediction time output by the numerical prediction branch structure and the actual membrane fouling index. The membrane fouling degradation rate factor is calculated by dividing the obtained difference by the preset future prediction step time span parameter.

5. The intelligent power plant concentrate treatment system according to claim 1, characterized in that: The model predicts the control objective function, which is constructed by algebraically summing and integrating the cumulative cost over the prediction time span with the control increment smoothing constraint term to obtain a complete discrete global dynamic objective function. The accumulated cost is based on the integral sum of the single-step operating cost over time steps within the system's predicted time domain; the single-step operating cost is composed of the system's dosing cost evaluation function, operating power consumption evaluation function, and dynamic state penalty term.

6. The intelligent power plant concentrate treatment system according to claim 5, characterized in that: The control command sequence includes the frequency of alkaline agent addition, the frequency of acidic agent addition, the frequency of scale inhibitor addition, the operating frequency of the high-pressure variable frequency pump, and the speed of the steam compressor. The central control module extracts the alkaline agent dosing frequency, acidic agent dosing frequency, and scale inhibitor dosing frequency to calculate the system dosing cost evaluation function, and extracts the high-pressure variable frequency pump operating frequency and steam compressor speed to calculate the operating power consumption evaluation function; The central control module substitutes the membrane fouling degradation rate factor into a preset exponential function to calculate a dynamic penalty coefficient, and multiplies the dynamic penalty coefficient with the operating frequency parameter of the high-pressure variable frequency pump to construct the dynamic state penalty term. The central control module invokes a sequential quadratic programming solver algorithm to find the minimum value of the global dynamic objective function.

7. The intelligent power plant concentrate treatment system according to claim 1, characterized in that: The central control module reads the apparent viscosity value of the slurry contained in the global system state vector in real time at continuous time steps; When the apparent viscosity of the slurry is greater than or equal to the safety rheological critical constant pre-cured in the controller memory, the central control module outputs a high-level trigger control word to activate the reverse safety constraint and online dilution control logic within the system, thereby imposing physical operation boundary restrictions on the optimization solution space of the model predictive control algorithm.

8. The intelligent power plant concentrate treatment system according to claim 7, characterized in that: The specific execution steps of the reverse safety constraint and online dilution control logic include: Within the control time domain, the upper limit of the steam compressor speed of the MVR crystallization module is limited, and a forced derating command to reduce the frequency of the high-pressure variable frequency pump is simultaneously issued to the membrane concentration module. The membrane concentration module continuously generates low-concentration dilute and concentrated water under low-frequency operation and discharges it into the concentrate buffer module. The concentrate buffer module pumps the low-concentration dilute and concentrated water after physical displacement mixing into the thermal circulation pipeline of the MVR crystallization module for rheological gradation dilution until the apparent viscosity of the crystal slurry is lower than the lower limit reset value for multiple consecutive time steps, at which point the reverse safety constraint command is revoked.

9. The intelligent power plant concentrate treatment system according to claim 1, characterized in that: When the membrane fouling degradation rate factor is greater than or equal to the set degradation rate warning threshold and the inorganic salt scaling probability value is greater than or equal to the set inorganic salt scaling confidence threshold, the central control module activates the anti-scaling self-healing intervention dual feedforward trigger flag in the system's underlying program. The central control module truncates the optimization iterative output path of the model predictive control objective function and sends a pulse dosing command sequence to the pretreatment dosing module, specifically stopping the dosing of alkaline agents and adding acidic agents and scale inhibitors at the maximum rated dosage.

10. The intelligent power plant concentrate treatment system according to claim 9, characterized in that: The time during which the pretreatment dosing module executes the pulse dosing command sequence is the intervention time window. The length threshold of the intervention time window is dynamically calculated and calibrated based on the hydraulic retention time obtained by dividing the total internal volume of the membrane concentration module by the current influent flow rate and multiplying it by a specific dimensionless constant ratio. Once the cumulative intervention time reaches the calculated threshold of the intervention time window length, the central control module clears the physical activation state of the anti-scaling and self-healing intervention dual feedforward trigger flag and resumes issuing the optimal control command sequence generated based on the model predictive control objective function.

Images