Double-membrane mixing direct drinking water system and operation method

Abstract

The invention relates to the field of direct drinking water of a fusion calculation model, and discloses a double-membrane mixing direct drinking water system and an operation method.The method comprises the steps that online data of a water supply end and a water quality monitoring end are collected for cleaning and time alignment, and a basic working condition data set is formed; the method comprises the following steps: establishing an operation characteristic model comprising water demand prediction, raw water quality prediction and membrane flux characteristic sub-models, calculating double-membrane target water yield and water quality indexes based on the model to generate a control parameter set, evaluating membrane pollution to generate an online cleaning task queue, performing joint simulation to generate a pressure stability control instruction set, and dynamically correcting the model. The conversion from passive response to active pre-judgment of the double-membrane mixing direct drinking water system is realized, and the operation stability, the water quality guarantee capability and the water-saving and energy-saving effects are improved.

Description

Technical Field

[0001] This invention relates to the field of direct drinking water using integrated computational models, and more particularly to a dual-membrane blending direct drinking water system and its operation method. Background Technology

[0002] With increasing demand for healthy drinking water, point-of-use water supply systems that meet direct drinking water standards are becoming a development trend. Ultrafiltration (UF) can effectively remove particles and bacteria, but it cannot desalinate, and the produced water cannot directly meet direct drinking water standards. While nanofiltration (NF) or RO (RO) can deeply desalinate, the mineral content of the produced water is too low, raising health concerns for long-term consumption, and the system's low recovery rate leads to significant water waste. Centralized direct drinking water systems on the market often store the deeply treated water in a purification tank, posing a risk of secondary contamination due to inadequate sealing. Adding continuous ultraviolet or ozone disinfection measures significantly increases operating costs and complexity, making widespread adoption difficult. Membrane systems require regular cleaning to maintain performance, but traditional cleaning processes easily cause system pressure fluctuations or water supply interruptions. How to efficiently clean the ultrafiltration section of the pretreatment unit without affecting a stable water supply is crucial for ensuring the long-term stable operation of the system. A new type of direct drinking water system and method that balances water quality safety, health attributes, operational economy, and water supply stability is urgently needed. Summary of the Invention

[0003] A method for operating a dual-membrane blended direct drinking water system includes:

[0004] S1. Collect online data from the water supply end and water quality monitoring end, clean and time-align the data to form a basic operating condition dataset. Based on the basic operating condition dataset, establish an operating characteristic model for dual-membrane operation prediction and constraint analysis.

[0005] Furthermore, the process of forming a basic operating condition dataset and establishing an operational feature model includes: collecting multi-source online monitoring data and constructing an original monitoring data set; performing outlier removal and missing value imputation processing on the original monitoring data set to obtain a cleaned monitoring data set; performing time alignment processing on the cleaned monitoring data set to obtain a basic operating condition dataset; and establishing an operational feature model based on the basic operating condition dataset.

[0006] Furthermore, the process of establishing the operational characteristic model is as follows: First, extract user-side water flow sequences and their corresponding time labels from the basic operating condition dataset to construct a water demand training sample set. Use a long short-term memory network as the basic structure of the water demand prediction sub-model to obtain the trained water demand prediction sub-model. Second, extract municipal water turbidity sequences and municipal water conductivity sequences and their corresponding time labels from the basic operating condition dataset to construct a raw water quality training sample set. Use an autoregressive moving average model with seasonal decomposition as the basic structure of the raw water quality prediction sub-model to obtain the trained raw water quality prediction sub-model. Third, extract ultrafiltration-related sequences from the basic operating condition dataset, calculate the ultrafiltration transmembrane pressure difference sequence and ultrafiltration specific flux sequence, and establish an ultrafiltration membrane flux characteristic sub-model using a multiple linear regression method. Similarly, establish a nanofiltration membrane flux characteristic sub-model, and merge the two to obtain the membrane flux characteristic sub-model. Finally, encapsulate the water demand prediction sub-model, the raw water quality prediction sub-model, and the membrane flux characteristic sub-model into a unified operational characteristic model, and define the input and output interfaces of the operational characteristic model.

[0007] S2, calculate the target effluent and water quality indicators of ultrafiltration and nanofiltration based on the operating characteristic model, generate a set of dual-membrane blending control parameters, and set the initial operating conditions of each membrane component and valve based on the set of dual-membrane blending control parameters.

[0008] Furthermore, the process of generating a dual-membrane blending control parameter set and setting initial operating conditions includes: calling the operating characteristic model to obtain the water demand prediction curve and raw water quality prediction curve within the predicted time span; determining the target water quality range of the blended effluent based on the water demand prediction curve and raw water quality prediction curve; constructing and solving the dual-membrane blending ratio optimization problem based on the target water quality range of the blended effluent to obtain the ultrafiltration permeate setting sequence, nanofiltration permeate setting sequence, and nanofiltration recovery rate setting sequence for each time period; and generating equipment-level control instructions based on the ultrafiltration permeate setting sequence, nanofiltration permeate setting sequence, and nanofiltration recovery rate setting sequence to constitute the dual-membrane blending control parameter set.

[0009] Furthermore, the process for constructing and solving the dual-membrane blending ratio optimization problem is as follows: Define the decision variables for the dual-membrane blending ratio optimization problem, including the ultrafiltration permeate flow rate, nanofiltration permeate flow rate, and nanofiltration recovery rate for each time period; define the constraints for the dual-membrane blending ratio optimization problem, including the water quality constraint that the water quality after blending falls within the target water quality range of the blended effluent, the supply and demand balance constraint that the sum of the ultrafiltration permeate flow rate and the nanofiltration permeate flow rate meets the water demand prediction curve, the equipment capacity constraint that the permeate flow rate of each membrane module does not exceed its rated flux, and the lower limit constraint that the nanofiltration recovery rate is not lower than the set minimum value; define the objective function for the dual-membrane blending ratio optimization problem, and perform a weighted summation of minimizing the nanofiltration concentrate discharge and minimizing the total system energy consumption as dual objectives; use a sequential quadratic programming algorithm to solve the dual-membrane blending ratio optimization problem to obtain the set sequences of ultrafiltration permeate flow rate, nanofiltration permeate flow rate, and nanofiltration recovery rate for each time period.

[0010] S3, based on the dual-membrane blending control parameter set combined with real-time membrane pressure and flux data, evaluates the fouling degree of each ultrafiltration membrane module, generates an online cleaning task queue, and adjusts the available membrane module combination and allowable operating redundancy based on the online cleaning task queue;

[0011] Furthermore, the process of generating an online cleaning task queue and adjusting operational redundancy includes: reading the requirements of the total ultrafiltration permeate volume for each time period from the dual-membrane blending control parameter set, and collecting the real-time operating parameters of each ultrafiltration membrane module; constructing a membrane fouling diagnostic feature vector based on the real-time operating parameters of each ultrafiltration membrane module; classifying the fouling level of each ultrafiltration membrane module and matching cleaning strategies based on the membrane fouling diagnostic feature vector set using a membrane fouling discrimination model; generating an online cleaning task queue based on the membrane module fouling status table and water demand prediction curve; and adjusting the system's operational redundancy distribution based on the online cleaning task queue.

[0012] Furthermore, the process for constructing the membrane fouling diagnostic feature vector is as follows: For each ultrafiltration membrane module, extract the baseline specific flux value of the ultrafiltration membrane module in the initial commissioning stage from the basic operating condition dataset; calculate the current specific flux value of the ultrafiltration membrane module, which is equal to the current instantaneous flux divided by the current transmembrane pressure difference, and then divided by the temperature correction factor; calculate the specific flux decay rate of the ultrafiltration membrane module, which is equal to the difference between the baseline specific flux value and the current specific flux value, and then divided by the baseline specific flux value; calculate the transmembrane pressure difference growth rate of the ultrafiltration membrane module, which is equal to the difference between the current transmembrane pressure difference and the initial transmembrane pressure difference, and then divided by the initial transmembrane pressure difference; combine the specific flux decay rate, transmembrane pressure difference growth rate, cumulative operating time, and cumulative permeate volume of the ultrafiltration membrane module into a four-dimensional numerical vector, which constitutes the membrane fouling diagnostic feature vector of the ultrafiltration membrane module.

[0013] S4, based on the online cleaning task queue, performs joint simulation of the water supply pump set and the piston-type pressure stabilizing tank, generates a pressure stabilization control instruction set, and determines the pump speed curve and valve opening plan for each cleaning stage based on the pressure stabilization control instruction set;

[0014] Furthermore, the process of conducting joint simulation and generating a pressure stability control instruction set includes: establishing a simplified pipeline network model; inputting cleaning tasks from the online cleaning task queue into the simplified pipeline network model for time-discrete simulation; evaluating whether the pressure response trajectory meets the pressure stability constraints, and if not, adjusting the control strategy and resimulating; and solidifying the control strategy that meets the pressure stability constraints into a pressure stability control instruction set.

[0015] Furthermore, the time-discrete simulation process is as follows: The first cleaning task record is read from the online cleaning task queue, and the membrane module number, estimated start time, cleaning step sequence, and estimated end time involved in the task are obtained. The time interval from the estimated start time to the estimated end time is discretized according to the simulation time step to obtain the simulation time point sequence. Each cleaning step in the cleaning step sequence is traversed to determine the corresponding valve action and pump action. For each simulation time point in the simulation time point sequence, the state parameters of each component in the simplified pipeline model are updated according to the valve opening and pump frequency corresponding to that time point. The Newton-Raphson iteration method is used to solve the pipeline hydraulic balance equations to obtain the pressure value of each node and the flow rate value of each pipe segment at that time point. The pressure value of the mixed effluent node and the transmembrane pressure difference value of each membrane module at each time point in the simulation time point sequence are recorded to form the pressure response trajectory under the baseline control strategy.

[0016] S5, based on the pressure stability control instruction set, drives the controlled valves and variable frequency pumps of the ultrafiltration and nanofiltration units to perform water production and online cleaning processes, and dynamically corrects the basic operating condition dataset and operating characteristic model by combining online water quality detection results.

[0017] A dual-membrane blending direct drinking water system is provided for implementing the aforementioned dual-membrane blending direct drinking water operation method. The system includes:

[0018] Data acquisition and model building module: used to collect online data from the water supply end and water quality monitoring end, clean and time-align the data to form a basic operating condition dataset, and build an operating characteristic model for dual-membrane operation prediction and constraint analysis based on the basic operating condition dataset;

[0019] Dual-membrane blending control parameter generation module: used to calculate the target effluent and water quality indicators of ultrafiltration and nanofiltration based on the operating characteristic model, generate a dual-membrane blending control parameter set, and set the initial operating conditions of each membrane component and valve based on the dual-membrane blending control parameter set;

[0020] Online cleaning task queue generation module: It is used to evaluate the fouling degree of each ultrafiltration membrane module based on the dual membrane blending control parameter set combined with real-time membrane pressure and flux data, generate an online cleaning task queue, and adjust the available membrane module combination and allowable operating redundancy based on the online cleaning task queue;

[0021] Pressure stabilization control instruction set generation module: used to perform joint simulation of water supply pump group and piston-type pressure stabilizing tank based on online cleaning task queue, generate pressure stabilization control instruction set, and determine pump speed curve and valve opening plan for each cleaning stage based on pressure stabilization control instruction set;

[0022] Water production and cleaning execution and model correction module: It is used to drive the controlled valves and variable frequency pumps of the ultrafiltration unit and nanofiltration unit to perform water production and online cleaning processes based on the pressure stability control instruction set, and dynamically correct the basic operating condition dataset and operating characteristic model by combining the online water quality detection results.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0024] This invention effectively solves the problem of the difficulty in directly using multi-source heterogeneous data in dual-membrane blended drinking water systems for operational control and predictive analysis by constructing an operational characteristic model that integrates long short-term memory networks, seasonal decomposition autoregressive moving average models, and multiple linear regression. This model unifies sensor data scattered across the municipal water inlet, ultrafiltration / nanofiltration inlet / outlet, and blended water outlet to the same time base. A water demand prediction sub-model captures the daily and weekly cycles of user water usage, a raw water quality prediction sub-model identifies the seasonal fluctuations in raw water turbidity and conductivity, and a membrane flux characteristic sub-model establishes a quantitative relationship between transmembrane pressure difference and permeate output. This achieves a shift from traditional static setpoint control to dynamic predictive control, enabling the system to adjust the dual-membrane blending ratio in advance to cope with load and water quality changes.

[0025] Based on the dual-membrane blending control parameter set generated by the operational characteristic model, the optimal water production ratio is solved by a sequential quadratic programming algorithm under the constraints of drinking water safety standards and healthy mineral target ranges. This method retains the beneficial minerals in ultrafiltration water while ensuring water quality safety by utilizing nanofiltration water. It solves the defects of simple deep desalination leading to excessively low mineral content or the inability to meet standards using only ultrafiltration. At the same time, it aims to minimize concentrate discharge and system energy consumption, maximizing nanofiltration recovery rate under water quality constraints and reducing water waste.

[0026] By using membrane fouling diagnostic feature vectors (multi-dimensional features such as specific flux decay rate and transmembrane pressure difference growth rate) and membrane fouling discrimination models, the fouling level of ultrafiltration membrane modules can be accurately classified. Combined with water demand prediction curves, an online cleaning task queue is generated during off-peak hours to avoid over-cleaning or untimely cleaning caused by traditional fixed cycles or pressure difference threshold triggering. At the same time, by simplifying the pipeline network model and co-simulating to adjust the valve action rate and pump frequency, the fluctuation of mixed water pressure during the cleaning process is controlled within the set range, eliminating the risk of water outage at the user end.

[0027] The system dynamically corrects the basic operating condition dataset and operational feature model through online water quality testing results, and updates the sub-model parameters using incremental learning methods. This enables the predictive ability to adapt to the long-term drift of raw water quality and the aging of membrane components, achieving continuous direct supply through a fully enclosed pipeline. This eliminates the risk of secondary pollution from traditional water purification tanks, fundamentally improving the system's operational stability, water quality assurance capabilities, and long-term economic efficiency. Attached Figure Description

[0028] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0029] Figure 1 This is a flowchart of a dual-membrane blending direct drinking water operation method according to the present invention;

[0030] Figure 2 This is a schematic diagram of the multi-source online monitoring equipment layout in an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram of the composite structure of the running feature model in an embodiment of the present invention;

[0032] Figure 4 This is a schematic diagram illustrating the principle of dual-membrane blending for water quality adjustment in an embodiment of the present invention;

[0033] Figure 5 This is a schematic diagram of membrane fouling diagnostic feature vectors in an embodiment of the present invention;

[0034] Figure 6 This is a schematic diagram of the buffer principle of the piston-type pressure stabilizing tank in an embodiment of the present invention;

[0035] Figure 7 This is a schematic diagram of the arrangement of the online water quality testing device in an embodiment of the present invention;

[0036] Figure 8 This is a functional block diagram of a dual-membrane blending direct drinking water system according to the present invention. Detailed Implementation

[0037] The technical solutions of 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.

[0038] Example 1:

[0039] Please see Figure 1 As shown, this embodiment provides a method for operating a dual-membrane blended direct drinking water system, including:

[0040] S1: Collect online data from the water supply end and water quality monitoring end, clean and time-align the data to form a basic operating condition dataset, and establish an operating characteristic model for dual-membrane operation prediction and constraint analysis based on the basic operating condition dataset.

[0041] This step addresses the challenge of directly using multi-source heterogeneous data in dual-membrane blended drinking water systems for operational control and predictive analysis. It involves collecting, cleaning, and time-aligning sensor signals distributed at the municipal water inlet, ultrafiltration inlet / outlet, nanofiltration inlet / outlet, and blended water outlet to create a basic operating condition dataset with a unified time reference. Based on this dataset, an operational characteristic model is established to describe the changing patterns of water demand, raw water quality fluctuations, and membrane flux characteristics. This operational characteristic model, as the final output of S1, will be used in S2 to calculate the dual-membrane blending control parameter set.

[0042] Specifically, the process of forming a basic working condition dataset and establishing a running feature model includes:

[0043] S11: Collect multi-source online monitoring data and construct the original monitoring data set.

[0044] In this step, online monitoring equipment such as pressure transmitters, flow meters, temperature sensors, turbidity meters, conductivity meters, and hardness analyzers installed on the dual-membrane blended direct drinking water system are used as data sources. The following sequences are collected synchronously: municipal water pressure, municipal water flow, municipal water temperature, municipal water turbidity, municipal water conductivity, ultrafiltration inlet pressure, ultrafiltration outlet pressure, ultrafiltration permeate flow, ultrafiltration permeate turbidity, nanofiltration inlet pressure, nanofiltration outlet pressure, nanofiltration permeate flow, nanofiltration permeate conductivity, nanofiltration concentrate flow, blended outlet pressure, blended outlet flow, blended outlet conductivity, blended outlet hardness, user-side water flow, and original timestamp sequences. Based on the above-mentioned collected sequences, an original monitoring data set is constructed. The original monitoring data set is a structured data table indexed by sensor identifier, signal type, and sampling time. Each row records the measurement value of a specific sensor at a given sampling time, and each column corresponds to the time series of a physical quantity. The municipal water supply pressure series refers to an ordered list of discrete pressure readings output by pressure transmitters installed at municipal water supply access points at fixed sampling periods. The user-side water flow series refers to an ordered list of discrete flow readings output by flow meters installed at the end of the mixed water outlet main, reflecting the real-time water demand of end users. Organizing these collection sequences by device number, physical quantity type, and time sequence yields the original monitoring data set with time as the primary key, providing a foundational data source for subsequent data cleaning and time alignment. See also Figure 2 This is a schematic diagram of the multi-source online monitoring equipment layout provided in an embodiment of this application. For example... Figure 2 As shown in the diagram, this figure illustrates the overall process flow and physical layout of key sensors in a dual-membrane blending direct drinking water system. Water enters from the municipal water inlet on the left, is pressurized by the pump set, and then splits into two paths, one entering the upper ultrafiltration (UF) membrane module and the other the lower nanofiltration (NF) membrane module. The produced water from both ultimately merges in the main pipe on the right to form blended effluent, which is then delivered to the user. The diagram clearly identifies the locations of data collection points using "P" (pressure), "F" (flow rate), and "Q" (water quality, including turbidity, conductivity, etc.), covering key nodes such as the municipal inlet, pre-membrane, post-membrane, and mixing point. In actual operation, the synchronous acquisition of multi-source heterogeneous data is fundamental to achieving refined control. By deploying high-frequency sensors at the aforementioned locations, the system can detect fluctuations in municipal pipeline pressure, sudden changes in raw water quality, and changes in the operating resistance of each membrane module in real time. This end-to-end monitoring layout breaks the limitations of traditional systems that only focus on effluent indicators. It provides a complete physical data source for data cleaning, time alignment, and the construction of a basic operating condition dataset that reflects the dynamic characteristics of the system in step S1, ensuring the input accuracy of subsequent prediction models.

[0045] S12: Based on the original monitoring data set, outlier removal and missing value imputation are performed to obtain a cleaned monitoring data set.

[0046] Specifically, the outlier removal and missing value imputation process is as follows:

[0047] S121: Traverse each acquisition sequence in the original monitoring data set. For each acquisition sequence, calculate the local mean and local standard deviation within a sliding window. The length of the sliding window is determined based on the typical variation period of the physical quantity; a minute-level window is used for pressure signals, and an hour-level window is used for water quality signals. Mark sampling points in the sequence that deviate from the local mean by more than a set multiple of the standard deviation as abnormal sampling points. The set multiple is determined through statistical analysis based on the fluctuation range of the physical quantity under normal operating conditions. Remove the measured values ​​marked as abnormal sampling points from the acquisition sequence to obtain the acquisition sequence after removing outliers.

[0048] S122: For missing sampling points in the acquired sequence after outlier removal due to sensor malfunction or communication interruption, a piecewise linear interpolation method is used to fill in the missing points. This piecewise linear interpolation method refers to calculating the estimated value of the missing sampling point by using the adjacent valid sampling points before and after the missing sampling point as endpoints, according to a linear ratio of the time distance. When the number of consecutive missing sampling points exceeds the maximum allowable interpolation length for the physical quantity, this time period is marked as an unusable time period and no interpolation processing is performed. The maximum interpolation length is determined through autocorrelation analysis based on the time correlation decay characteristics of the physical quantity.

[0049] S123: The acquisition sequence after removing outliers is merged with the acquisition sequence after interpolation to form a cleaned monitoring data set. The cleaned monitoring data set inherits the structured data table format of the original monitoring data set, but does not contain the measurement values ​​marked as outlier sampling points, and missing sampling points have been filled with estimated values.

[0050] S13: Time alignment processing is performed on the cleaned monitoring data set to obtain the basic working condition dataset.

[0051] Specifically, the time alignment process is as follows:

[0052] S131: Extract the original timestamp sequence of each acquisition sequence from the cleaned monitoring data set, analyze the sampling interval distribution of each acquisition sequence, and determine the unified sampling interval of the system. The unified sampling interval of the system is taken as an integer multiple of the minimum sampling interval among all acquisition sequences to ensure that the dynamic information of high-frequency signals is not lost.

[0053] S132: Construct a unified time axis based on the system's unified sampling interval. The unified time axis is a set of time points that increase uniformly according to the system's unified sampling interval starting from the system startup time, covering the entire time span of the cleaned monitoring data set.

[0054] S133: For each acquisition sequence in the cleaned monitoring data set, perform a resampling operation on the acquisition sequence according to a unified time axis. For acquisition sequences with a sampling interval smaller than the system's unified sampling interval, downsampling is performed using a moving average filter; for acquisition sequences with a sampling interval larger than the system's unified sampling interval, upsampling is performed using cubic spline interpolation. Through the above resampling operation, an aligned acquisition sequence mapped onto a unified time axis is obtained.

[0055] S134: The aligned acquisition sequences are horizontally stitched together according to their time points on a unified time axis to generate a basic operating condition dataset with time as the row index and physical quantity as the column index. The basic operating condition dataset is a two-dimensional numerical matrix, with the number of rows equal to the number of time points included in the unified time axis, and the number of columns equal to the number of physical quantities included in the original monitoring data set. The basic operating condition dataset serves as the input for S14 and the update object for model correction in S5.

[0056] S14: Establish an operational feature model based on the basic operating condition dataset.

[0057] The operational characteristic model is a composite model structure that encapsulates a water demand prediction sub-model, a raw water quality prediction sub-model, and a membrane flux characteristic sub-model. It provides predictive capabilities and constraint boundaries for subsequent dual-membrane blending control and online cleaning scheduling. Specifically, the process for establishing the operational characteristic model is as follows:

[0058] S141: Extract user-side water flow sequences and their corresponding time labels from the basic operating condition dataset to construct a water demand training sample set. The time labels include the hour, week, and month number of the time, used to characterize the periodicity of water demand. A Long Short-Term Memory (LSTM) network is used as the basic structure of the water demand prediction sub-model. It takes historical water flow sequences and time labels as input and future water flow sequences as output. Parameter optimization is performed by minimizing the mean square error between the predicted and actual flow rates to obtain the trained water demand prediction sub-model. The LTM network is a recurrent neural network with a gating mechanism, capable of capturing long-term dependencies in time series data.

[0059] S142: Extract the municipal water influent turbidity sequence, municipal water influent conductivity sequence, and their corresponding time labels from the basic operating condition dataset to construct a raw water quality training sample set. An autoregressive moving average model with seasonal decomposition is used as the basic structure of the raw water quality prediction sub-model. The raw water quality sequence is decomposed into trend components, seasonal components, and residual components. Each component is modeled separately to obtain the trained raw water quality prediction sub-model. The seasonal decomposition refers to separating the time series components according to a fixed period, enabling the model to handle periodic fluctuations and aperiodic drifts independently.

[0060] S143: Extract the ultrafiltration influent pressure sequence, ultrafiltration effluent pressure sequence, ultrafiltration permeate flow rate sequence, and ultrafiltration influent temperature sequence from the basic operating condition dataset, and calculate the ultrafiltration transmembrane pressure difference sequence and the ultrafiltration specific flux sequence. Each element in the ultrafiltration transmembrane pressure difference sequence equals the difference between the ultrafiltration influent pressure and the ultrafiltration effluent pressure at the same moment. Each element in the ultrafiltration specific flux sequence equals the ratio of the ultrafiltration permeate flow rate to the ultrafiltration transmembrane pressure difference at the same moment, and after temperature correction, a standardized specific flux value is obtained. An empirical relationship between ultrafiltration specific flux, ultrafiltration transmembrane pressure difference, and influent temperature is established using a multiple linear regression method, resulting in an ultrafiltration membrane flux characteristic sub-model. Similarly, extract nanofiltration-related sequences from the basic operating condition dataset to establish a nanofiltration membrane flux characteristic sub-model. Combine the ultrafiltration membrane flux characteristic sub-model and the nanofiltration membrane flux characteristic sub-model to obtain a membrane flux characteristic sub-model.

[0061] S144: Encapsulate the water demand prediction sub-model, raw water quality prediction sub-model, and membrane flux characteristic sub-model into a unified operational characteristic model, and define the input and output interfaces of the operational characteristic model. The input of the operational characteristic model is a slice of the basic operating condition dataset at the current moment and the prediction time span. The output is the water demand prediction curve, the raw water quality prediction curve, and the predicted permeate flow rate of each membrane module under given transmembrane pressure differential and temperature conditions within the prediction time span. See also... Figure 3 This is a schematic diagram of the composite structure of the runtime feature model provided in the embodiments of this application. Figure 3As shown in the diagram, this visually illustrates the internal logical architecture of the operational feature model. The three boxes on the left represent three independently trained sub-models: the "Water Demand Prediction Sub-model" is responsible for capturing the temporal patterns of user water usage behavior; the "Raw Water Quality Prediction Sub-model" is responsible for predicting the seasonal changes in turbidity and conductivity of the water source; and the "Membrane Flux Characteristics Sub-model" is responsible for quantifying the physical relationship between transmembrane pressure difference and permeate output. The outputs of these three sub-models converge via arrows into the central gear-shaped "Operational Feature Model," symbolizing the fusion and collaborative computation of multi-dimensional information. After comprehensive processing by the model, the final output to the right is the "Dual Membrane Blending Control Parameter Set." In the control of direct drinking water systems, single-dimensional predictions often struggle to balance supply and demand with equipment safety. This composite structure organically combines "external demand" (water consumption), "external constraints" (raw water quality), and "internal capacity" (membrane flux), transforming the control system from a passive "post-occurrence response" to an active "predictive regulation," providing core algorithmic support for generating the optimal permeate ratio and equipment commands in step S2.

[0062] Specifically, traditional direct drinking water systems typically rely solely on instantaneous measurement data for control decisions, failing to perceive the overall trends in water quality, quantity, and membrane operating status. This results in static, set-point control strategies that struggle to cope with seasonal fluctuations in raw water quality and periodic changes in user water load. By constructing a basic operating condition dataset and establishing an operational characteristic model, heterogeneous data scattered across various monitoring points are unified under a common time reference, enabling subsequent control strategies to be based on predictions of load fluctuations and raw water quality changes. The water demand prediction sub-model captures daily and weekly patterns of user water usage behavior, providing a basis for advance adjustment of the dual-membrane blending ratio. The raw water quality prediction sub-model identifies seasonal fluctuations in municipal water quality, providing constraints for dynamic optimization of nanofiltration recovery rates. The membrane flux characteristic sub-model establishes a quantitative relationship between transmembrane pressure difference and permeate output, providing a benchmark for determining the timing of online cleaning. The synergistic effect of these three sub-models allows the system to shift from passive response to proactive prediction, fundamentally improving the operational stability and water quality assurance capabilities of the dual-membrane blended direct drinking water system.

[0063] S2: Calculate the target effluent and water quality indicators for ultrafiltration and nanofiltration based on the operating characteristic model, generate a set of dual-membrane blending control parameters, and set the initial operating conditions of each membrane module and valve based on the dual-membrane blending control parameter set.

[0064] This step addresses the issue that simply pursuing deep desalination or using ultrafiltration alone cannot simultaneously ensure drinking water safety and mineral retention. Based on the predictive capabilities of the operational characteristic model output from S1, and under constraints of drinking water standards, target ranges for healthy minerals, and system recovery rate, multi-objective optimization is used to find the optimal permeate ratio of ultrafiltration and nanofiltration. This generates a dual-membrane blending control parameter set containing permeate setpoints, recovery rate setpoints, and equipment-level control commands for each time period. This dual-membrane blending control parameter set, as the final output of S2, will be invoked in S3 to constrain the generation of the online cleaning task queue.

[0065] Specifically, the process of generating the dual-film blending control parameter set and setting the initial operating conditions includes:

[0066] S21: Call the running feature model to obtain the water demand prediction curve and raw water quality prediction curve within the prediction time span.

[0067] The current basic operating condition dataset slices and the set prediction time span are input into the running feature model. The running feature model sequentially calls the water demand prediction sub-model and the raw water quality prediction sub-model, outputting the water demand prediction curve and the raw water quality prediction curve. The water demand prediction curve is a discrete numerical sequence with a unified time axis within the prediction time span as the abscissa and the predicted water flow rate as the ordinate, representing the expected water consumption at each future time. The raw water quality prediction curve is a discrete numerical sequence with a unified time axis within the prediction time span as the abscissa and the predicted raw water conductivity and predicted raw water turbidity as the ordinates, representing the expected raw water quality state at each future time.

[0068] S22: Determine the target water quality range for blended effluent based on the water demand prediction curve and the raw water quality prediction curve.

[0069] Specifically, the process for determining the target water quality range of the blended effluent is as follows:

[0070] S221: Extract the safety constraint boundaries of the mixed water from the direct drinking water industry standard, including the upper limit of turbidity, the upper limit of total bacteria count, the upper limit of total organic carbon, and the lower limit of residual chlorine, which constitute a set of safety constraints.

[0071] S222: Extract health constraint boundaries for blended water from healthy drinking water guidelines, including lower and upper limits for conductivity, lower and upper limits for hardness, and lower limit for total dissolved solids, forming a set of health constraints. These health constraint boundaries are used to ensure that blended water retains an appropriate amount of beneficial minerals and avoids the health risks of long-term consumption due to excessive desalination.

[0072] S223: Merge the safety constraint set and the health constraint set to form the target water quality range for the blended effluent. This target water quality range is a multi-dimensional constraint space composed of the upper and lower limits of various water quality indicators. Any feasible solution in subsequent optimization calculations must fall within this constraint space. See also... Figure 4 This is a schematic diagram illustrating the principle of dual-membrane blending water quality regulation provided in an embodiment of this application. Figure 4 As shown in the diagram, this visual comparison reveals the core advantages of the dual-membrane blending process. The water flow on the left represents ultrafiltration permeate; the dense black dots in the cup symbolize the retention of beneficial minerals such as calcium and magnesium in the original water, but its ability to remove dissolved micro-pollutants is limited. The water flow on the right represents nanofiltration permeate; the cup contains almost no dots, symbolizing its high purity after deep desalination, but long-term consumption may lead to insufficient mineral intake. The large cup below receives both water flows, and the density of mineral dots in the mixed water is moderate. This diagram vividly explains the technical path of this invention in resolving the contradiction between "safety" and "health" in traditional direct drinking water: by accurately calculating the blending ratio of the two, nanofiltration technology removes potential heavy metals and organic pollutants, establishing a safety baseline for water quality, while ultrafiltration permeate replenishes the trace elements needed by the human body, avoiding the health risks of pure water. The final blended direct drinking water falls within the predetermined target water quality range, achieving an optimal balance between safety standards and health needs.

[0073] S23: Based on the target water quality range of the blended effluent, construct and solve the dual-membrane blending ratio optimization problem to obtain the ultrafiltration permeate setting sequence, nanofiltration permeate setting sequence, and nanofiltration recovery rate setting sequence for each time period.

[0074] Specifically, the process for constructing and solving the dual-film blending ratio optimization problem is as follows:

[0075] S231: Define the decision variables for the dual-membrane blending ratio optimization problem, including ultrafiltration permeate flow rate, nanofiltration permeate flow rate, and nanofiltration recovery rate at each time period. The nanofiltration recovery rate is defined as the ratio of nanofiltration permeate flow rate to nanofiltration feed flow rate, and its value range is determined based on the design parameters of the nanofiltration membrane module.

[0076] S232: Define the constraints for the dual-membrane blending ratio optimization problem, including the water quality constraint that the blended water quality falls within the target water quality range of the blended effluent, the supply-demand balance constraint that the sum of the ultrafiltration permeate and nanofiltration permeate meets the water demand prediction curve, the equipment capacity constraint that the permeate of each membrane module does not exceed its rated flux, and the lower limit constraint that the nanofiltration recovery rate is not lower than a set minimum value. The blended water quality is calculated using the principle of mass conservation. The conductivity of the blended effluent is equal to the product of the conductivity of the ultrafiltration permeate and the ultrafiltration permeate, plus the product of the conductivity of the nanofiltration permeate and the nanofiltration permeate, divided by the total permeate. Since ultrafiltration cannot desalinate, the conductivity of the ultrafiltration permeate is approximately equal to the conductivity of the raw water, obtained from the raw water quality prediction curve.

[0077] S233: Define the objective function for the dual-membrane blending ratio optimization problem, using the minimization of nanofiltration concentrate discharge and the minimization of total system energy consumption as weighted sums as dual objectives. The nanofiltration concentrate discharge equals the nanofiltration feed water flow rate minus the nanofiltration permeate flow rate, and is negatively correlated with the nanofiltration recovery rate. The total system energy consumption equals the sum of the energy consumption of the ultrafiltration water supply pump and the nanofiltration water supply pump, calculated using a membrane flux characteristic sub-model under given permeate flow rate and transmembrane pressure difference. The weighting coefficients for nanofiltration concentrate discharge and total system energy consumption in the objective function are determined using the analytic hierarchy process (AHP) based on the system operator's water-saving or energy-saving priority strategy.

[0078] S234: A sequential quadratic programming algorithm is used to solve the dual-membrane blending ratio optimization problem. The sequential quadratic programming algorithm is a numerical solution method applicable to constrained nonlinear optimization problems. It gradually approaches the optimal solution by constructing a quadratic programming subproblem at the current iteration point and solving for the search direction. Using the predicted water consumption for each time period in the water demand prediction curve as the right-hand side of the supply-demand balance constraint, and the predicted conductivity for each time period in the raw water quality prediction curve as the input parameter for water quality constraint calculation, the algorithm iteratively solves until the convergence condition is met, obtaining the ultrafiltration permeate setting sequence, nanofiltration permeate setting sequence, and nanofiltration recovery rate setting sequence for each time period.

[0079] S24: Generate equipment-level control instructions based on the ultrafiltration permeate flow rate setting sequence, nanofiltration permeate flow rate setting sequence, and nanofiltration recovery rate setting sequence, forming a dual-membrane blending control parameter set.

[0080] Specifically, the process for generating device-level control commands is as follows:

[0081] S241: Based on the ultrafiltration permeate flow rate setting sequence and membrane flux characteristic sub-model, calculate the target flow rate and target head that the ultrafiltration water supply pump needs to maintain at each time period, and map the target flow rate and target head to the target frequency sequence of the ultrafiltration water supply pump through the pump performance curve.

[0082] S242: Based on the nanofiltration permeate flow rate setting sequence, the nanofiltration recovery rate setting sequence, and the membrane flux characteristic sub-model, the target flow rate and target pressure that the nanofiltration water supply pump needs to maintain at each time period are calculated in reverse. The target flow rate and target pressure are then mapped to the target frequency sequence of the nanofiltration water supply pump through the pump performance curve. At the same time, the target opening sequence of the nanofiltration concentrate valve is calculated based on the nanofiltration recovery rate setting sequence. The target opening of the nanofiltration concentrate valve is negatively correlated with the nanofiltration recovery rate; the higher the recovery rate, the smaller the opening of the concentrate valve.

[0083] S243: Based on the raw water quality prediction curve, determine whether the post-mineralization unit needs to be activated. When the predicted raw water conductivity is lower than the set mineralization activation threshold, generate a mineralization unit activation command; when the predicted raw water conductivity is higher than the set mineralization deactivation threshold, generate a mineralization unit deactivation command. The mineralization activation and deactivation thresholds are determined through sensitivity analysis based on the relationship between the lower limit of conductivity and the proportion of ultrafiltration permeate in the target water quality range of the blended effluent.

[0084] S244: The ultrafiltration permeate flow rate setting sequence, nanofiltration permeate flow rate setting sequence, nanofiltration recovery rate setting sequence, ultrafiltration water supply pump target frequency sequence, nanofiltration water supply pump target frequency sequence, nanofiltration concentrate valve target opening sequence, and mineralization unit commissioning command are integrated along a time axis to form a dual-membrane blending control parameter set. This dual-membrane blending control parameter set is a structured data table indexed by time and with each control variable's set value as a column. Each row corresponds to the target state of all controlled equipment within a control cycle.

[0085] Specifically, traditional direct drinking water systems either employ nanofiltration or reverse osmosis for deep desalination, resulting in excessively low mineral content in the permeate, or rely solely on ultrafiltration, leading to permeate that fails to meet direct drinking water standards. Both approaches have significant drawbacks. This paper addresses this by constructing a dual-membrane blending ratio optimization problem and solving it under the constraint of a target water quality range for the blended effluent. This approach integrates water safety, health attributes, system recovery rate, and energy consumption into a unified optimization framework. Ultrafiltration permeate retains beneficial minerals from the raw water but cannot remove dissolved contaminants, while nanofiltration permeate has high desalination capabilities but excessively low mineral content. Blending the two in an optimized ratio satisfies both direct drinking water safety standards and maintains an appropriate mineral concentration. Simultaneously, the optimization process aims to minimize nanofiltration concentrate discharge, maximizing nanofiltration recovery rate while meeting water quality constraints, thus reducing water waste at the source. The dual-membrane blending control parameter set transforms the optimization results into directly executable device-level commands. This allows the system to operate automatically, adjusting to changes in raw water quality and water load, rather than relying on empirically set fixed ratios, achieving refined operational control.

[0086] S3: Based on the dual-membrane blending control parameter set combined with real-time membrane pressure and flux data, assess the fouling degree of each ultrafiltration membrane module, generate an online cleaning task queue, and adjust the available membrane module combination and allowable operational redundancy based on the online cleaning task queue.

[0087] This step addresses the problem of inappropriate cleaning timing and water supply interruption risks caused by traditional ultrafiltration online cleaning relying on fixed time intervals or simple differential pressure thresholds. Based on the dual-membrane blending control parameter set output from S2, which specifies the total ultrafiltration permeate requirements for each time period, and combined with real-time collected membrane pressure and flux data, a membrane fouling diagnostic feature vector is constructed. A membrane fouling discrimination model is used to classify the fouling level of each ultrafiltration membrane module, and cleaning time windows are allocated to membrane modules requiring cleaning during periods of low water usage, generating an online cleaning task queue with time attributes. This online cleaning task queue, as the final output of S3, will be invoked in S4 for joint hydraulic simulation.

[0088] Specifically, the process of generating an online cleaning task queue and adjusting runtime redundancy includes:

[0089] S31: Read the requirements of each time period for the total water production of ultrafiltration in the dual-membrane blending control parameter set, and collect the real-time operating parameters of each ultrafiltration membrane module.

[0090] The ultrafiltration permeate production set sequence is extracted from the dual-membrane blending control parameter set. This sequence represents the total permeate production volume required by the ultrafiltration unit in each control cycle. Simultaneously, real-time operating parameters are collected from sensors installed on each ultrafiltration membrane module, including the inlet pressure, outlet pressure, instantaneous flux, cumulative operating time, and cumulative permeate production volume. The inlet pressure refers to the pressure measurement value on the inlet side of the ultrafiltration membrane module, and the outlet pressure refers to the pressure measurement value on the outlet side of the ultrafiltration membrane module; the difference between the two is the transmembrane pressure difference for that ultrafiltration membrane module. The instantaneous flux refers to the permeate production volume per unit membrane area per unit time, reflecting the current filtration capacity of the membrane module.

[0091] S32: Construct a membrane fouling diagnostic feature vector based on the real-time operating parameters of each ultrafiltration membrane module.

[0092] Specifically, the process for constructing the membrane fouling diagnostic feature vector is as follows:

[0093] S321: For each ultrafiltration membrane module, extract the baseline specific flux value of the ultrafiltration membrane module in the initial commissioning stage from the basic operating condition dataset. The baseline specific flux value is the permeate flux corresponding to the unit transmembrane pressure difference of the membrane module in the clean state.

[0094] S322: Calculate the current specific flux value of the ultrafiltration membrane module, which is equal to the current instantaneous flux divided by the current transmembrane pressure difference, and then divided by the temperature correction factor. The temperature correction factor is used to eliminate the influence of feed water temperature changes on membrane flux and is calculated using an exponential relationship based on the temperature sensitivity parameters of the membrane material.

[0095] S323: Calculate the specific flux decay rate of the ultrafiltration membrane module, which is equal to the difference between the reference specific flux value and the current specific flux value, and then divided by the reference specific flux value. The specific flux decay rate reflects the degree of performance degradation of the membrane module relative to the clean state, and its value ranges from zero to one. The larger the value, the more severe the pollution.

[0096] S324: Calculate the growth rate of the transmembrane pressure difference of the ultrafiltration membrane module, which is equal to the difference between the current transmembrane pressure difference and the initial transmembrane pressure difference, divided by the initial transmembrane pressure difference. The growth rate of the transmembrane pressure difference reflects the degree of increase in the resistance of the membrane module and is positively correlated with the degree of membrane fouling.

[0097] S325: Combine the specific flux decline rate, transmembrane pressure differential growth rate, cumulative operating time, and cumulative permeate flow rate of the ultrafiltration membrane module into a four-dimensional numerical vector to constitute the membrane fouling diagnostic feature vector for that ultrafiltration membrane module. Repeat the above steps for all ultrafiltration membrane modules in the system to obtain the set of membrane fouling diagnostic feature vectors. See [link to documentation] Figure 5 This is a schematic diagram of the membrane fouling diagnostic feature vector provided in the embodiments of this application. For example... Figure 5 As shown, this diagram uses a radar chart to illustrate a multi-dimensional assessment system for membrane module fouling status. Four axes radiate outward from the center point, representing four key dimensions: "specific flux decline rate," "transmembrane pressure differential growth rate," "cumulative operating time," and "cumulative permeate volume." Green, yellow, and red diamond-shaped areas are distributed from the inside out, corresponding to the boundaries of "light fouling," "moderate fouling," and "severe fouling," respectively. The dark polygonal areas represent the current measured status of a particular membrane module, with the black dots visually reflecting the degree of deviation from each indicator. In actual operation and maintenance, relying solely on a single pressure differential threshold is often insufficient to accurately determine the true cause of membrane fouling (e.g., whether it's temporary resistance due to excessive flux or substantial fouling due to pore blockage). By constructing this four-dimensional feature vector, the system can comprehensively consider hydraulic performance degradation and historical operating loads to achieve a precise profile of the membrane fouling level, thereby guiding the intelligent sequencing of the online cleaning task queue in step S3 and avoiding over-cleaning or delayed cleaning.

[0098] S33: Based on the set of membrane fouling diagnostic feature vectors, the fouling level of each ultrafiltration membrane module is classified and cleaning strategy is matched by a membrane fouling discrimination model.

[0099] The membrane fouling discrimination model is a pre-trained classifier that takes the membrane fouling diagnostic feature vector as input and outputs a fouling level label and a recommended cleaning strategy type for the ultrafiltration membrane module. Specifically, the process of fouling level classification and cleaning strategy matching is as follows:

[0100] S331: Input each membrane fouling diagnostic feature vector from the set of membrane fouling diagnostic feature vectors into the membrane fouling discrimination model. The membrane fouling discrimination model outputs a fouling level label based on the position of the feature vector in the feature space. The fouling level label takes three discrete categories: light fouling, moderate fouling, and heavy fouling. The boundary between these categories is determined by training a support vector machine classifier based on historical cleaning effect data.

[0101] S332: Match the appropriate cleaning strategy type according to the contamination level label. Light contamination is matched with a short-time backwash strategy, with a short cleaning time and no chemical agents added; moderate contamination is matched with a forward high-flow-rate flush strategy, with a medium cleaning time and the selective addition of low-concentration cleaning agents; heavy contamination is matched with a chemically enhanced cleaning strategy, with a longer cleaning time and the addition of high-concentration cleaning agents and soaking.

[0102] S333: Combine the membrane module number, fouling level label and cleaning strategy type of each membrane module into a membrane module fouling status record, forming a membrane module fouling status table.

[0103] S34: Generate an online cleaning task queue based on the membrane module fouling status table and water demand prediction curve.

[0104] Specifically, the process for generating the online cleaning task queue is as follows:

[0105] S341: Call the water demand forecasting sub-model from the running feature model to obtain the water demand forecasting curve for the future forecast time span. Traverse the water demand forecasting curve to identify consecutive periods where the predicted water flow is lower than the water trough threshold, and mark these periods as available cleaning windows. The water trough threshold is determined by taking the lower quartile of the statistical distribution of the water demand forecasting curve.

[0106] S342: Read the ultrafiltration permeate flow rate setting sequence for each time period from the dual-membrane blending control parameter set, and calculate the maximum number of membrane modules that the system is allowed to shut down within each available cleaning window. The maximum number of membrane modules is equal to the difference between the total ultrafiltration permeate flow rate of the system and the ultrafiltration permeate flow rate setting value for that time period, divided by the rated permeate flow rate of a single membrane module, and rounded down to the integer part.

[0107] S343: Select membrane modules with a pollution level label of moderate or heavy pollution from the membrane module pollution status table, and sort them in descending order according to the severity of the pollution level label and the magnitude of the specific flux decline rate to obtain a priority list of membrane modules to be cleaned.

[0108] S344: Traverse the available cleaning windows. For each available cleaning window, sequentially retrieve membrane modules from the priority list of membrane modules to be cleaned. Determine whether the cleaning duration corresponding to the cleaning strategy type of the ultrafiltration membrane module does not exceed the duration of the available cleaning window. Simultaneously, determine whether the number of membrane modules currently allocated to the window is less than the maximum number of membrane modules in the window. If both conditions are met, allocate the ultrafiltration membrane module to the available cleaning window and record the estimated start time, cleaning step sequence, and estimated end time of the ultrafiltration membrane module. The cleaning step sequence is based on the cleaning strategy type. The cleaning step sequence corresponding to the short-term backwash strategy includes seven steps: stop permeate, open the backwash valve, start the backwash pump, continue backwashing, stop the backwash pump, close the backwash valve, and restore permeate.

[0109] S345: Arrange all assigned membrane module cleaning records in ascending order of their expected start time to form an online cleaning task queue. The online cleaning task queue is an ordered list, and each record contains five fields: membrane module number, expected start time, cleaning step sequence, expected end time, and cleaning strategy type.

[0110] S35: Adjust the redundancy distribution of the system based on the online cleaning task queue.

[0111] Specifically, the process for adjusting the distribution of operational redundancy is as follows:

[0112] S351: Iterate through each control cycle within the prediction time span, count the number of membrane modules in the cleaning state in the online cleaning task queue, calculate the number of available membrane modules for that control cycle, which is equal to the total number of membrane modules in the system minus the number of membrane modules in the cleaning state.

[0113] S352: Calculate the operational redundancy for this control cycle. The calculation formula is: Operational Redundancy = (Number of available membrane modules × Rated permeate flow rate per single membrane module − Ultrafiltration permeate flow rate setpoint for this control cycle) ÷ Ultrafiltration permeate flow rate setpoint for this control cycle. The operational redundancy represents the proportion of excess permeate flow capacity that the system can reserve within this control cycle, which is used to provide boundary conditions for subsequent pressure stabilization control.

[0114] S353: Organize the operational redundancy of each control cycle in chronological order to form an operational redundancy sequence. Append the operational redundancy sequence to the metadata of the online cleaning task queue as a constraint input for the S4 co-simulation.

[0115] Specifically, traditional ultrafiltration systems typically use fixed-cycle triggering or simple differential pressure threshold triggering for online cleaning. The former may lead to over-cleaning of clean membrane modules, shortening their lifespan, while the latter may result in severely fouled membrane modules not being cleaned in time, affecting the system's water production capacity. By introducing membrane fouling diagnostic feature vectors and a membrane fouling discrimination model, the fouling degree of membrane modules is expanded from a single indicator to a multi-dimensional feature characterization, making fouling assessment more accurate. The cleaning task allocation is combined with water demand forecasting curves, prioritizing cleaning operations during off-peak water usage periods to ensure that the cleaning process does not affect normal water use. By calculating the maximum number of cleanable membrane modules in each available cleaning window, too many membrane modules are simultaneously shut down, preventing overload of remaining membrane modules. The generation of the online cleaning task queue enables intelligent scheduling for simultaneous water supply and regeneration, allowing membrane modules to complete cleaning operations without interrupting water supply. The calculation of the operational redundancy sequence provides a clear capability boundary for subsequent pressure stability control.

[0116] S4: Based on the online cleaning task queue, perform joint simulation of the water supply pump set and the piston-type pressure stabilizing tank to generate a pressure stabilization control instruction set. Based on the pressure stabilization control instruction set, determine the pump speed curve and valve opening plan for each cleaning stage.

[0117] This step addresses the problem of drastic system pressure fluctuations and even temporary water outages at the user end caused by relying solely on experience to start and stop valves and pumps during traditional ultrafiltration online cleaning. Based on the online cleaning task queue output from S3, a simplified pipeline network model is established, including ultrafiltration pipelines, nanofiltration pipelines, blending mains, and piston-type pressure stabilizing tanks. Pressure response trajectories under different control strategies are calculated through time-discrete simulation to select control schemes that meet pressure stability constraints, generating a pressure stabilization control command set that can be directly issued and executed. This pressure stabilization control command set, as the final output of S4, will be invoked in S5 to drive the actual equipment to perform the water production and cleaning process.

[0118] Specifically, the process of performing co-simulation and generating a pressure stabilization control instruction set includes:

[0119] S41: Establish a simplified pipeline network model.

[0120] The simplified pipe network model is a mathematical description of the hydraulic characteristics of a dual-membrane blended direct drinking water system, including five types of hydraulic components: pipe sections, nodes, pump sets, valves, and piston-type pressure stabilizing tanks. Specifically, the process for establishing the simplified pipe network model is as follows:

[0121] S411: Determine the network topology based on the system pipeline layout diagram. Abstract the ultrafiltration inlet pipe, ultrafiltration outlet pipe, nanofiltration inlet pipe, nanofiltration outlet pipe, mixing manifold, and user water supply main pipe into pipe segment elements. Each pipe segment element includes three attribute parameters: pipe segment length, pipe segment inner diameter, and pipe segment roughness coefficient.

[0122] S412: The connection points of each pipe section are abstracted as node elements, and each node element contains two attribute parameters: node elevation and node water demand. The ultrafiltration water supply pump set and nanofiltration water supply pump set are abstracted as pump set elements, and each pump set element contains three attribute parameters: pump performance curve parameter, rated speed, and current operating frequency.

[0123] S413: The ultrafiltration permeate valve, nanofiltration permeate valve, nanofiltration concentrate valve, backwash valve, and blending control valve are abstracted as valve elements. Each valve element contains two attribute parameters: valve characteristic curve parameters and current opening degree.

[0124] S414: The piston-type pressure stabilizing tank is abstracted as a pressure stabilizing tank element. The piston-type pressure stabilizing tank is a pressure regulating device that uses a piston to separate an air chamber and a water chamber. When the pipeline pressure increases, the piston compresses the air chamber and absorbs water; when the pipeline pressure decreases, the air chamber expands and releases water, thereby buffering pressure fluctuations. The pressure stabilizing tank element includes four attribute parameters: tank volume, initial piston position, air chamber pre-charge pressure, and piston motion damping coefficient. See also... Figure 6 This is a schematic diagram of the buffer principle of the piston-type pressure stabilizing tank provided in the embodiments of this application. Figure 6 As shown in the figure, this diagram illustrates the physical mechanism of a piston-type pressure stabilizing tank in maintaining stable pipeline pressure. The tank's interior is divided into an upper "air chamber" and a lower "water chamber" by a vertically movable piston. The air chamber is filled with compressed gas at a preset pressure, while the water chamber is connected to the lower "mixed water outlet main pipe" via a connecting pipe. The two reverse arrows in the diagram clearly describe its working process: when the membrane module shuts down or a valve switch causes a decrease in the main pipe pressure, the air chamber expands, pushing the piston downwards and forcing the water in the water chamber into the main pipe (releasing the stored water), compensating for the flow loss; conversely, when the pressure is too high, the water flow pushes the piston up, compressing the air chamber (storing excess water) and absorbing excess energy. Compared to traditional air tanks, the piston isolation design avoids the problem of gas dissolving in water. In the co-simulation of S4 and the actual execution of S5, this device, acting as the system's "hydraulic flywheel," effectively smooths pressure fluctuations caused by online cleaning operations, eliminates water hammer effects, ensures a stable and continuous water experience for users, and solves the industry problem of having to interrupt water supply for ultrafiltration cleaning.

[0125] S415: Combine the above-mentioned hydraulic components and their connection relationships into a simplified pipeline network model, and define the boundary conditions of the model, including the municipal water inflow pressure as the pressure boundary of the inlet node and the user-side water flow rate as the flow boundary of the outlet node.

[0126] S42: Input the cleaning tasks in the online cleaning task queue into the simplified pipeline network model for time-discrete simulation.

[0127] Specifically, the process of the time-discrete simulation is as follows:

[0128] S421: Read the first cleaning task record from the online cleaning task queue and obtain the membrane module number, estimated start time, cleaning step sequence and estimated end time involved in the task.

[0129] S422: Discretize the time interval from the expected start time to the expected end time according to the simulation time step to obtain the simulation time point sequence. The simulation time step is determined by characteristic time analysis based on the time constant of the system's hydraulic response to ensure that the dynamic process of pressure fluctuations can be captured.

[0130] S423: Traverse each cleaning step in the cleaning step sequence and determine the corresponding valve action and pump unit action. Taking the backflushing valve opening step as an example, the corresponding action for this step is to gradually increase the opening degree of the backflushing valve involving the membrane module from zero to the fully open state. The rate of change of the opening degree is determined according to the action time of the valve actuator.

[0131] S424: For each simulation time point in the simulation time point sequence, update the state parameters of each component in the simplified pipeline network model according to the valve opening and pump frequency corresponding to that time point. Use the Newton-Raphson iteration method to solve the pipeline network hydraulic balance equations to obtain the pressure value of each node and the flow rate value of each pipe segment at that time point. The pipeline network hydraulic balance equations include the nodal flow continuity equation and the pipe segment energy conservation equation, which are solved simultaneously to determine the steady-state hydraulic state.

[0132] S425: Records the pressure values ​​of the mixed effluent nodes and the transmembrane pressure difference values ​​of each membrane module at each time point in the simulation time sequence, forming the pressure response trajectory under the baseline control strategy.

[0133] S43: Evaluate whether the pressure response trajectory meets the pressure stability constraints. If not, adjust the control strategy and re-simulate.

[0134] Specifically, the evaluation and adjustment process is as follows:

[0135] S431: Read the minimum setpoint of the blended effluent pressure for each time period from the dual-membrane blending control parameter setpoint, and use it as the lower limit of the pressure stability constraint. Define the allowable pressure fluctuation range as a percentage of the minimum setpoint of the blended effluent pressure, which is determined through engineering experience based on the pressure sensitivity of the user-side water supply equipment.

[0136] S432: Traverse each time point in the pressure response trajectory and determine whether the pressure value of the mixed effluent node at that time point is lower than the minimum pressure setting value of the mixed effluent minus the allowable pressure fluctuation range. If there is any time point that does not meet this condition, it is determined that the current control strategy does not meet the pressure stability constraint.

[0137] S433: If the current control strategy does not meet the pressure stability constraint, adjust the execution parameters of the cleaning step, including reducing the valve opening rate to mitigate pressure changes, increasing the water supply pump frequency to compensate for flow loss, and adjusting the release water volume of the piston-type pressure stabilizing tank to buffer the pressure drop. Write the adjusted control parameters into the simplified pipeline model, return to S424 to re-perform the simulation calculation, and obtain the pressure response trajectory under the adjusted control strategy.

[0138] S434: Repeat the evaluation and adjustment process until the pressure response trajectory meets the pressure stability constraint, or the maximum number of adjustment iterations is reached. If the pressure stability constraint cannot be met even after reaching the maximum number of adjustment iterations, remove the cleaning task from the online cleaning task queue and postpone it to the next available cleaning window.

[0139] S44: Solidify the control strategy that satisfies the pressure stability constraint into a pressure stability control instruction set.

[0140] Specifically, the process of the curing pressure stabilization control instruction set is as follows:

[0141] S441: Repeat the simulation and adjustment process of S42 and S43 for each cleaning task in the online cleaning task queue to obtain the final control strategy corresponding to each cleaning task.

[0142] S442: The final control strategies corresponding to each cleaning task are merged along the time axis, and the target frequency of the ultrafiltration water supply pump, the target frequency of the nanofiltration water supply pump, the target opening degree of each valve, and the allowable water release range of the piston-type pressure stabilizing tank at each simulation time point are extracted to form a pressure stabilization control instruction set. The pressure stabilization control instruction set is a structured data table indexed by the simulation time point and with the set values ​​of each control variable as columns, which can be directly read and executed by the controller.

[0143] S443: Adds an execution status flag to the pressure stabilization control instruction set. The initial state is set to pending execution, and it is used for status tracking when S5 is actually executed.

[0144] Specifically, traditional online ultrafiltration cleaning typically relies on empirical valve opening and closing sequences and pump start-stop logic, neglecting the coupling relationship between valve action rates, pump frequency changes, and the hydraulic response of the pipeline network. This can easily lead to significant pressure drops or drastic fluctuations in the pipeline network during cleaning, affecting the stability of the user-side water supply. By establishing a simplified pipeline network model and performing time-discrete simulations, the valve actions and pump adjustments during the cleaning process are incorporated into a unified hydraulic calculation framework, enabling the prediction of the pressure response trajectory before actual execution. A piston-type pressure stabilizing tank, acting as a pressure buffer element, is included in the simulation model. Its dynamic characteristics of releasing and absorbing water can be accurately simulated, providing a clear physical basis for adjusting the control strategy. By iteratively adjusting control parameters until the pressure response trajectory meets stability constraints, the generated pressure stability control instruction set is predictable and verifiable. This allows the impact of the cleaning process on water supply stability to be assessed and eliminated before execution, fundamentally avoiding the risk of water outages at the user end due to cleaning operations.

[0145] S5: Based on the pressure stability control instruction set, the controlled valves and variable frequency pumps of the ultrafiltration and nanofiltration units are driven to perform water production and online cleaning processes. The basic operating condition dataset and operating characteristic model are dynamically corrected in combination with the online water quality detection results.

[0146] This step addresses the problem in traditional systems where online cleaning and normal water supply control are disconnected and lack adaptive operation. It uses the pressure stabilization control command set output from the S4 controller to drive the actual actuators to complete water production and cleaning operations. Simultaneously, an online water quality monitoring device continuously monitors the quality of the mixed effluent. When the monitoring results deviate from the predictions of the operational characteristic model, model correction is triggered, enabling the system to learn and evolve independently. The output of this step is the completed execution status and the corrected operational characteristic model, forming a complete closed-loop control.

[0147] Specifically, the process of executing water production and online cleaning and dynamically correcting the model includes:

[0148] S51: Download the pressure stabilization control instruction set to the programmable logic controller for real-time execution.

[0149] The programmable logic controller (PLC) is the core of the field control for the dual-membrane blended direct drinking water system. It is connected to the control interfaces of the ultrafiltration water supply pump group, nanofiltration water supply pump group, the actuators of each controlled valve, and the piston-type pressure stabilizing tank via an industrial communication bus. Specifically, the download and execution process is as follows:

[0150] S511: The pressure stabilization control instruction set is transmitted from the host computer to the instruction buffer of the programmable logic controller (PLC) through the communication interface. The PLC performs format verification and integrity checks on the received instructions.

[0151] S512: The programmable logic controller reads the first time index in the pressure stabilization control instruction set, compares the time index with the system real-time clock, and when the system real-time clock reaches the time index, extracts the corresponding target frequency of the ultrafiltration water supply pump, the target frequency of the nanofiltration water supply pump, and the target opening degree of each valve.

[0152] S513: The programmable logic controller (PLC) sends frequency setting commands to the ultrafiltration and nanofiltration water supply pump sets via the inverter drive interface, and sends opening setting commands to each valve actuator via the analog output interface. The inverter adjusts the motor speed according to the received frequency setting commands, and the valve actuator adjusts the valve core position according to the received opening setting commands.

[0153] S514: The programmable logic controller continuously reads the system's real-time clock and matches it with the time index in the pressure stabilization control instruction set, executing the control instructions corresponding to each time point in chronological order until all instructions in the pressure stabilization control instruction set have been executed.

[0154] S52: During the water production and cleaning process, the piston-type pressure stabilizing tank is used for pressure buffering.

[0155] Specifically, the process for scheduling the piston-type pressure stabilizing tank is as follows:

[0156] S521: The programmable logic controller (PLC) acquires the pressure sensor signal of the mixed water outlet main in real time and compares the acquired real-time pressure of the mixed water outlet with the expected pressure value at the corresponding time point in the pressure stabilization control instruction set.

[0157] S522: When the real-time pressure of the mixed effluent is lower than the expected pressure value minus the set pressure deviation dead zone, the programmable logic controller sends an opening command to the outlet valve of the piston-type pressure stabilizing tank. Under the action of the air chamber pressure, the piston-type pressure stabilizing tank releases the stored water to the mixed effluent main pipe to compensate for the flow loss caused by the membrane module's shutdown, so that the real-time pressure of the mixed effluent recovers.

[0158] S523: When the real-time pressure of the mixed effluent is higher than the expected pressure value plus the set pressure deviation dead zone, the programmable logic controller sends an opening command to the inlet valve of the piston-type pressure stabilizing tank. The water in the mixed effluent main pipe flows into the piston-type pressure stabilizing tank and compresses the air chamber, absorbing the excess flow caused by the membrane module resuming operation, thus reducing the real-time pressure of the mixed effluent.

[0159] S524: The programmable logic controller continuously monitors the piston position sensor signal of the piston-type pressure stabilizing tank. When the piston position approaches the upper or lower limit, it adjusts the filling and discharging rate of the piston-type pressure stabilizing tank to prevent the piston from exceeding the limit and causing the pressure stabilizing function to fail.

[0160] S53: Continuously monitor the water quality indicators of the blended effluent through an online water quality testing device.

[0161] The online water quality monitoring device is installed downstream of the main outlet pipe for the mixed water and includes an online conductivity meter, an online turbidity meter, an online residual chlorine analyzer, and an online microbial indicator detector. Specifically, the continuous monitoring process is as follows:

[0162] S531: The online conductivity meter measures the conductivity of the mixed effluent flowing through the detection tank according to the set detection cycle, and outputs the measured value as the conductivity detection value of the mixed effluent.

[0163] S532: The online turbidity meter measures the turbidity of the mixed effluent flowing through the detection tank according to the set detection cycle, and outputs the measured value as the turbidity detection value of the mixed effluent.

[0164] S533: The online residual chlorine analyzer measures the residual chlorine concentration of the mixed effluent flowing through the detection tank according to the set detection cycle, and outputs the measured value as the residual chlorine detection value of the mixed effluent.

[0165] S534: The online microbial indicator detector performs indirect measurement of the total bacterial count of the mixed effluent flowing through the detection tank using fluorescence method according to the set detection cycle, and outputs the measured value as the microbial indicator detection value of the mixed effluent.

[0166] S535: Combine the measured values ​​of conductivity, turbidity, residual chlorine, and microbial indicator indicators of the blended effluent according to the detection time to form a water quality test record for the blended effluent, and write it into the online water quality test result storage area. See also Figure 7 This is a schematic diagram of the arrangement of the online water quality testing device provided in the embodiments of this application. Figure 7 As shown in the diagram, this illustrates the closed-loop water quality monitoring system located on the main outlet pipe of the blended water system. Using a bypass sampling method, the water flows sequentially through four closely arranged detection modules: a conductivity meter, a turbidity meter, a residual chlorine analyzer, and a microbial detector. These high-precision instruments analyze the physicochemical properties of the water sample in real time, and the data is compiled into a "water quality detection record." The dashed arrows in the diagram point to the data records, indicating that the data flows to the system's feedback end. This device is not only the last line of defense for water quality safety but also the "eyes" for the system's self-evolution. In step S5, when there is a persistent deviation between the measured water quality data and the model's predicted values, these records serve as truth labels, triggering updates to the basic operating condition dataset and corrections to the operating characteristic model. This online feedback mechanism endows the system with adaptive capabilities, enabling it to cope with the effects of slow variables such as membrane module aging and long-term drift of raw water quality, ensuring that the dual-membrane blending control strategy is always in an optimal state.

[0167] S54: Determine whether model correction needs to be triggered based on the water quality test records of the mixed effluent.

[0168] Specifically, the process for determining and triggering correction is as follows:

[0169] S541: Call the raw water quality prediction sub-model and the blending ratio from the dual-membrane blending control parameter set from the operating characteristic model, and calculate the expected conductivity and turbidity values ​​of the blended effluent at the current moment.

[0170] S542: Compare the measured conductivity value of the blended effluent in the water quality test record with the expected conductivity value, and calculate the absolute value of the conductivity deviation. Compare the measured turbidity value of the blended effluent with the expected turbidity value, and calculate the absolute value of the turbidity deviation.

[0171] S543: When the absolute value of the conductivity deviation exceeds the set conductivity deviation threshold, or the absolute value of the turbidity deviation exceeds the set turbidity deviation threshold, and the above deviation conditions are continuously met for a duration exceeding the set deviation duration, it is determined that the operating characteristic model has a prediction deviation, triggering the model correction process. The conductivity deviation threshold and turbidity deviation threshold are determined through statistical analysis based on the measurement accuracy of the online water quality detection device and the width of the target water quality range of the mixed effluent. The deviation duration is used to filter false triggers caused by short-term disturbances and is determined based on the system hydraulic response time constant.

[0172] S55: Perform basic operating condition dataset updates and run feature model corrections.

[0173] Specifically, the update and correction process is as follows:

[0174] S551: The mixed effluent water quality test records within a set time range before and after the triggering time of model correction, the measured values ​​of each sensor in the original monitoring data set of the corresponding time period, and the actual operating parameters of each device in the dual-membrane blending control parameter set of the corresponding time period are aligned and integrated according to a unified time axis to form a model correction data package.

[0175] S552: Append the model correction data package to the corresponding time interval of the basic working condition dataset, replace or supplement the original data records, and complete the update of the basic working condition dataset.

[0176] S553: ​​Based on the updated basic operating condition dataset, training samples for the water demand prediction sub-model, raw water quality prediction sub-model, and membrane flux characteristic sub-model are re-extracted, and the parameters of each sub-model are updated using an incremental learning method. The incremental learning method refers to making minor parameter adjustments using only newly added data while retaining the original model parameters, thus avoiding the computational overhead of full retraining.

[0177] S554: The updated water demand prediction sub-model, raw water quality prediction sub-model, and membrane flux characteristic sub-model are repackaged into a revised operating characteristic model, replacing the original operating characteristic model, so that the subsequent calculation of the dual membrane blending control parameter set in S2 is based on the revised prediction capability.

[0178] S555: Update the execution status flag in the pressure stabilization control instruction set to "completed", record the actual execution parameters and water quality monitoring results of this cleaning task, and use them for continuous optimization of the membrane fouling discrimination model.

[0179] Specifically, in traditional direct drinking water systems, the online cleaning process and the normal water production process often employ independent control logic. After cleaning, manual intervention is required for system reset and parameter adjustment, making continuous automatic operation difficult. By using a pressure stabilization control command set as a unified interface for the execution layer, the water production and cleaning processes are coordinated under the same control framework. Real-time scheduling of the piston-type pressure stabilizing tank ensures stable pipeline pressure throughout the process. Continuous monitoring by the online water quality detection device provides the system with closed-loop feedback capability. When there is a continuous deviation between the detection results and model predictions, model correction is automatically triggered, enabling the operating characteristic model to self-learn and evolve according to changes in actual operating conditions. The dynamic updating of the basic operating condition dataset and the incremental correction of the operating characteristic model work together to allow the dual-membrane blending ratio and cleaning strategy to adaptively adjust to the long-term drift of raw water quality and the aging process of membrane modules. The entire system achieves continuous direct supply in a fully enclosed pipeline, eliminating the need for a water purification tank, which is susceptible to secondary pollution in traditional systems. This fundamentally eliminates the pollution risk in the water storage stage, while simultaneously considering water quality safety, health attributes, operational economy, and long-term water supply stability.

[0180] For example, the workflow of the above method is illustrated using a centralized direct drinking water system in a residential community as an application scenario. The system is configured with several ultrafiltration membrane modules, several nanofiltration membrane modules, one piston-type pressure stabilizing tank, one ultrafiltration water supply pump set, one nanofiltration water supply pump set, and one online water quality monitoring device. After the system starts, S1 is executed first, collecting parameters such as pressure, flow rate, temperature, conductivity, and turbidity from the municipal water inlet and the inlet and outlet of each membrane module. After outlier removal, missing value imputation, and time alignment, a basic operating condition dataset is formed. Based on this dataset, a water demand prediction sub-model, a raw water quality prediction sub-model, and a membrane flux characteristic sub-model are trained and encapsulated into an operational feature model. Subsequently, S2 is executed, and the operational feature model outputs the water demand prediction curve and raw water quality prediction curve for future periods. Under the constraints of direct drinking water safety standards and healthy mineral target ranges, the optimal water production ratio of ultrafiltration and nanofiltration for each period is solved using a sequential quadratic programming algorithm, generating a dual-membrane blending control parameter set and converting it into equipment-level control commands. After the system enters steady-state operation, step S3 diagnoses the fouling level of each ultrafiltration membrane module, scheduling cleaning operations for severely fouled modules during off-peak water usage periods, and generating an online cleaning task queue. Step S4 establishes a simplified pipeline network model to perform co-simulation of the cleaning process, adjusting valve opening rates and pump frequency change curves until the pressure response trajectory meets stability constraints, generating a pressure stabilization control instruction set. Step S5 sends control instructions to the programmable logic controller (PLC), driving the equipment to perform water production and cleaning operations. The piston-type pressure stabilizing tank adjusts water filling and discharging in real time according to the pipeline network pressure. The online water quality monitoring device continuously monitors the conductivity and turbidity of the mixed effluent. When there is a continuous deviation between the detection results and the predicted values, it triggers an update of the basic operating condition dataset and a correction of the operating characteristic model, enabling the system to gradually improve prediction accuracy and control precision over long-term operation.

[0181] Example 2:

[0182] This embodiment, based on Embodiment 1, provides a dual-membrane blending direct drinking water system, such as... Figure 8 As shown, it includes:

[0183] Data acquisition and model building module: used to collect online data from the water supply end and water quality monitoring end, clean and time-align the data to form a basic operating condition dataset, and build an operating characteristic model for dual-membrane operation prediction and constraint analysis based on the basic operating condition dataset;

[0184] Dual-membrane blending control parameter generation module: used to calculate the target effluent and water quality indicators of ultrafiltration and nanofiltration based on the operating characteristic model, generate a dual-membrane blending control parameter set, and set the initial operating conditions of each membrane component and valve based on the dual-membrane blending control parameter set;

[0185] Online cleaning task queue generation module: It is used to evaluate the fouling degree of each ultrafiltration membrane module based on the dual membrane blending control parameter set combined with real-time membrane pressure and flux data, generate an online cleaning task queue, and adjust the available membrane module combination and allowable operating redundancy based on the online cleaning task queue;

[0186] Pressure stabilization control instruction set generation module: used to perform joint simulation of water supply pump group and piston-type pressure stabilizing tank based on online cleaning task queue, generate pressure stabilization control instruction set, and determine pump speed curve and valve opening plan for each cleaning stage based on pressure stabilization control instruction set;

[0187] Water production and cleaning execution and model correction module: It is used to drive the controlled valves and variable frequency pumps of the ultrafiltration unit and nanofiltration unit to perform water production and online cleaning processes based on the pressure stability control instruction set, and dynamically correct the basic operating condition dataset and operating characteristic model by combining the online water quality detection results.

Claims

1. A method for operating a dual-membrane blended direct drinking water system, characterized in that, The method includes: S1: Collect online data from the water supply end and water quality monitoring end, clean and time-align the data to form a basic operating condition dataset, and establish an operating characteristic model for dual-membrane operation prediction and constraint analysis based on the basic operating condition dataset; S2: Calculate the target effluent and water quality indicators of ultrafiltration and nanofiltration based on the operating characteristic model, generate a set of dual-membrane blending control parameters, and set the initial operating conditions of each membrane component and valve based on the set of dual-membrane blending control parameters. S3: Based on the dual-membrane blending control parameter set combined with real-time membrane pressure and flux data, evaluate the fouling degree of each ultrafiltration membrane module, generate an online cleaning task queue, and adjust the available membrane module combination and allowable operational redundancy based on the online cleaning task queue; S4: Based on the online cleaning task queue, perform joint simulation of the water supply pump set and the piston-type pressure stabilizing tank to generate a pressure stabilization control instruction set, and determine the pump speed curve and valve opening plan for each cleaning stage based on the pressure stabilization control instruction set. S5: Based on the pressure stability control instruction set, the controlled valves and variable frequency pumps of the ultrafiltration and nanofiltration units are driven to perform water production and online cleaning processes. The basic operating condition dataset and operating characteristic model are dynamically corrected in combination with the online water quality detection results.

2. The method for operating a dual-membrane blended direct drinking water system according to claim 1, characterized in that, The process of creating a basic working condition dataset and establishing a running feature model includes: Collect multi-source online monitoring data and construct a raw monitoring data set; The original monitoring data set was processed by outlier removal and missing value imputation to obtain a cleaned monitoring data set. A basic working condition dataset is obtained by performing time alignment processing on the cleaned monitoring data set. A runtime characteristic model is established based on a basic operating condition dataset.

3. The method for operating a dual-membrane blended direct drinking water system according to claim 2, characterized in that, The process for establishing and running a feature model is as follows: Extract user-side water flow sequences and their corresponding time labels from the basic operating condition dataset, construct a water demand training sample set, and use a long short-term memory network as the basic structure of the water demand prediction sub-model to obtain the trained water demand prediction sub-model. The turbidity sequence and conductivity sequence of municipal water inflow and their corresponding time labels are extracted from the basic working condition dataset to construct a raw water quality training sample set. An autoregressive moving average model with seasonal decomposition is used as the basic structure of the raw water quality prediction sub-model to obtain the trained raw water quality prediction sub-model. Ultrafiltration-related sequences were extracted from the basic operating condition dataset. The ultrafiltration transmembrane pressure difference sequence and ultrafiltration specific flux sequence were calculated. A sub-model of ultrafiltration membrane flux characteristics was established using the multiple linear regression method. The nanofiltration transmembrane pressure difference sequence and specific flux sequence were extracted using the same method as the ultrafiltration membrane flux characteristics sub-model. A sub-model of nanofiltration membrane flux characteristics was established using multiple linear regression. The two were then combined to obtain the membrane flux characteristics sub-model. The water demand prediction sub-model, raw water quality prediction sub-model, and membrane flux characteristic sub-model are encapsulated into a unified operational characteristic model, and the input and output interfaces of the operational characteristic model are defined.

4. The method for operating a dual-membrane blended direct drinking water system according to claim 1, characterized in that, The process of generating a set of control parameters for dual-film blending and setting initial operating conditions includes: The feature model is invoked to obtain the water demand prediction curve and raw water quality prediction curve within the prediction time span; The target water quality range for blended effluent is determined based on the water demand forecast curve and the raw water quality forecast curve. Based on the target water quality range of the blended effluent, a problem of optimizing the blending ratio of the two membranes was constructed and solved, resulting in the setting sequence of ultrafiltration permeate, nanofiltration permeate, and nanofiltration recovery rate for each time period. Based on the ultrafiltration permeate flow rate setting sequence, nanofiltration permeate flow rate setting sequence, and nanofiltration recovery rate setting sequence, equipment-level control commands are generated, forming a dual-membrane blending control parameter set.

5. The method for operating a dual-membrane blended direct drinking water system according to claim 4, characterized in that, The process for constructing and solving the dual-film blending ratio optimization problem is as follows: Define the decision variables for the dual-membrane blending ratio optimization problem, including ultrafiltration permeate, nanofiltration permeate, and nanofiltration recovery rate at each time period; The constraints for the dual-membrane blending ratio optimization problem are defined, including the water quality constraint that the water quality after blending falls within the target water quality range of the blended effluent, the supply and demand balance constraint that the sum of the ultrafiltration permeate and nanofiltration permeate meets the water demand prediction curve, the equipment capacity constraint that the permeate of each membrane module does not exceed its rated flux, and the lower limit constraint that the nanofiltration recovery rate is not lower than the set minimum value. Define the objective function for the dual-membrane blending ratio optimization problem, and perform a weighted summation of minimizing nanofiltration concentrate discharge and minimizing total system energy consumption as dual objectives. The problem of optimizing the blending ratio of the two membranes was solved by using a sequential quadratic programming algorithm, and the ultrafiltration permeate flow rate setting sequence, nanofiltration permeate flow rate setting sequence, and nanofiltration recovery rate setting sequence were obtained for each time period.

6. The method for operating a dual-membrane blended direct drinking water system according to claim 1, characterized in that, The process of generating an online cleaning task queue and adjusting runtime redundancy includes: Read the requirements of the total ultrafiltration water production for each time period in the dual-membrane blending control parameter set, and collect the real-time operating parameters of each ultrafiltration membrane module; A membrane fouling diagnostic feature vector is constructed based on the real-time operating parameters of each ultrafiltration membrane module; Based on the set of membrane fouling diagnostic feature vectors, the fouling level of each ultrafiltration membrane module is classified and cleaning strategy is matched by a membrane fouling discrimination model. An online cleaning task queue is generated based on the membrane module fouling status table and water demand prediction curve. Adjust the redundancy distribution of the system based on the online cleaning task queue.

7. The method for operating a dual-membrane blended direct drinking water system according to claim 6, characterized in that, The process for constructing the diagnostic feature vector for membrane fouling is as follows: For each ultrafiltration membrane module, the baseline specific flux value of the ultrafiltration membrane module in the initial commissioning stage is extracted from the basic operating condition dataset; Calculate the current specific flux value of the ultrafiltration membrane module. The current specific flux value is equal to the current instantaneous flux divided by the current transmembrane pressure difference, and then divided by the temperature correction factor. Calculate the specific flux decay rate of the ultrafiltration membrane module. The specific flux decay rate is equal to the difference between the reference specific flux value and the current specific flux value, and then divided by the reference specific flux value. Calculate the transmembrane pressure difference growth rate of the ultrafiltration membrane module. The transmembrane pressure difference growth rate is equal to the difference between the current transmembrane pressure difference and the initial transmembrane pressure difference, and then divided by the initial transmembrane pressure difference. The specific flux decay rate, transmembrane pressure differential growth rate, cumulative operating time, and cumulative permeate volume of the ultrafiltration membrane module are combined into a four-dimensional numerical vector to form the membrane fouling diagnostic feature vector of the ultrafiltration membrane module.

8. The method for operating a dual-membrane blended direct drinking water system according to claim 1, characterized in that, The process of performing co-simulation and generating a pressure stabilization control instruction set includes: Establish a simplified pipeline network model; The cleaning tasks in the online cleaning task queue are input into a simplified pipeline model for time-discrete simulation. Evaluate whether the pressure response trajectory meets the pressure stability constraints. If it does not, adjust the control strategy and resimulate. The control strategy that satisfies the pressure stability constraint is solidified into a pressure stability control instruction set.

9. The method for operating a dual-membrane blended direct drinking water system according to claim 8, characterized in that, The process of time-discrete simulation is as follows: Read the first cleaning task record from the online cleaning task queue and obtain the membrane module number, estimated start time, cleaning step sequence and estimated end time involved in the task record; Discretize the time interval from the expected start time to the expected end time according to the simulation time step to obtain the simulation time point sequence; Traverse each cleaning step in the cleaning step sequence and determine the corresponding valve action and pump unit action for that cleaning step. For each simulation time point in the simulation time point sequence, the state parameters of each component in the simplified pipeline network model are updated according to the valve opening and pump frequency corresponding to that simulation time point. The Newton iteration method is used to solve the hydraulic balance equations of the pipeline network to obtain the pressure value of each node and the flow rate value of each pipe segment at that simulation time point. The pressure values ​​of the mixed effluent nodes and the transmembrane pressure difference values ​​of each membrane module are recorded at each time point in the simulation time sequence to form the pressure response trajectory under the baseline control strategy.

10. A dual-membrane blending direct drinking water system, used to implement the dual-membrane blending direct drinking water operation method according to any one of claims 1-9, characterized in that, The system includes: Data acquisition and model building module: used to collect online data from the water supply end and water quality monitoring end, clean and time-align the data to form a basic operating condition dataset, and build an operating characteristic model for dual-membrane operation prediction and constraint analysis based on the basic operating condition dataset; Dual-membrane blending control parameter generation module: used to calculate the target effluent and water quality indicators of ultrafiltration and nanofiltration based on the operating characteristic model, generate a dual-membrane blending control parameter set, and set the initial operating conditions of each membrane component and valve based on the dual-membrane blending control parameter set; Online cleaning task queue generation module: It is used to evaluate the fouling degree of each ultrafiltration membrane module based on the dual membrane blending control parameter set combined with real-time membrane pressure and flux data, generate an online cleaning task queue, and adjust the available membrane module combination and allowable operating redundancy based on the online cleaning task queue; Pressure stabilization control instruction set generation module: used to perform joint simulation of water supply pump group and piston-type pressure stabilizing tank based on online cleaning task queue, generate pressure stabilization control instruction set, and determine pump speed curve and valve opening plan for each cleaning stage based on pressure stabilization control instruction set; Water production and cleaning execution and model correction module: It is used to drive the controlled valves and variable frequency pumps of the ultrafiltration unit and nanofiltration unit to perform water production and online cleaning processes based on the pressure stability control instruction set, and dynamically correct the basic operating condition dataset and operating characteristic model by combining the online water quality detection results.

Images