METHOD OF OPERATION OF A SEAWATER PROCESSING PLANT

Abstract

Digital seawater processing. Operating methods for membrane processes, such as seawater purification processes, are described in this document.The methods generally include: flowing seawater into a membrane purification process; recovering purified seawater and concentrated seawater from the membrane purification process; detecting operational parameters of the membrane purification process; determining a state of the membrane purification process from the operational parameters using a physical model; using a statistical model to resolve a time dependence of the membrane purification process state; using a recursive statistical process to update the statistical model; using the updated statistical model to predict a future state of the membrane purification process; comparing the predicted future state of the membrane purification process with a limit state; and predicting a remaining time before the membrane purification process reaches the limit state.

Description

METHOD OF OPERATION OF A SEAWATER PROCESSING PLANT CROSS-REFERENCE TO RELATED REQUESTS

[0001] This document is based on and claims priority for U.S. Provisional Application Serial No. 62 / 837910, filed on April 24, 2019, which is incorporated herein by reference in its entirety. FIELD OF THE INVENTION

[0002] The embodiments described in this document generally refer to the operation of membrane separation processes. Specifically, this application describes the advanced control and operation of seawater purification processes that are applicable to any membrane separation process. FUNDAMENTALS

[0003] Membrane processes are common in process industries. Membranes are commonly used to separate materials according to differential permeability. As the filter material of membranes, a portion normally permeates the membrane and a portion does not. Over time, the membrane surface can become coated, fouled, or otherwise impacted by material collection on the surface, leading to a decline in membrane separation efficiency, primarily due to reduced permeability. Chemical treatments can be used to reduce fouling or membrane fouling. In some cases, continuous dosing can prolong the process. Petition 870240083556, dated 09 / 30 / 2024, page 13 / 58 2 / 35 is the time between membrane cleanings. Additionally, some membranes have specific materials for specific filtrations, and these materials can be sensitive to certain components. Chemical treatments can be used to neutralize components that may damage the membrane chemicals. Chemicals generally improve membrane performance up to a certain point and degrade performance above that point. Furthermore, this point can change over time as the membrane ages.

[0004] One area where membrane separation is used is in seawater processing. Seawater may have undergone some type of extraction process, such as hydrocarbon or metal extraction, or it may be desired to increase the purity of the seawater to a potable state. In one case, membranes are used to remove certain components from seawater. For example, membrane filters can remove suspended solids and / or sulfates from seawater. In some cases, membranes are sensitive to certain components of seawater, such as chlorides. In these cases, chemical treatment is often used to remove the components that give rise to the sensitivity. Additionally, chemical treatments can be used to remove filtered material from the membranes. In many cases, continuous chemical treatment is employed, both to remove sensitive components and to clean membranes. Managing the operation of such Petition 870240083556, dated 09 / 30 / 2024, page 14 / 58 3 / 35 processes are sensitive to changes in environmental conditions, including seawater composition, ambient temperature, and seawater temperature, among others. Chemicals can be wasted and membranes damaged by improper chemical management. Methods and systems for managing membrane processes are necessary. SUMMARY

[0005] The embodiments described in this document provide a method comprising: flowing seawater into a membrane purification process; recovering purified seawater and concentrated seawater from the membrane purification process; detecting operational parameters of the membrane purification process; determining a state of the membrane purification process from the operational parameters using a physical model; using a statistical model to resolve a time dependence of the state of the membrane purification process; using a recursive statistical process to update the statistical model; using the updated statistical model to predict a future state of the membrane purification process; comparing the predicted future state of the membrane purification process with a limit state; and predicting a remaining time before the membrane purification process reaches the limit state.

[0006] Other modalities described in this document provide a method, comprising the flow of seawater in a Petition 870240083556, dated 09 / 30 / 2024, page 15 / 58 4 / 35 membrane purification process; recover purified seawater and concentrated seawater from the membrane purification process; detection of operational parameters of the membrane purification process; determine a state of the membrane purification process from the operational parameters using a physical model; use a statistical model to solve a time dependence of the state of the membrane purification process; use a recursive statistical process to update the statistical model; use the updated statistical model to predict a future state of the membrane purification process; compare the predicted future state of the membrane purification process with a limit state; predict a remaining time before the membrane purification process reaches the limit state; and adjust a chemical treatment flow rate to a membrane of the membrane purification process based on the remaining time.

[0007] Other embodiments provide a method, comprising the flow of seawater in a membrane purification process; recovering purified seawater and concentrated seawater from the membrane purification process; detecting operational parameters of the membrane purification process; determining a state of the membrane purification process from the operational parameters using a linear physical model; using an exponential statistical model to resolve a time dependence of the process state. Petition 870240083556, dated 09 / 30 / 2024, page 16 / 58 5 / 35 membrane purification; use a recursive Bayesian statistical process to update the exponential statistical model; use the updated exponential statistical model to predict a future state of the membrane purification process; compare the predicted future state of the membrane purification process with a limiting state; and predict a remaining time before the membrane purification process reaches the limiting state; and adjust a chemical treatment flow rate for a membrane of the membrane purification process based on the remaining time. BRIEF DESCRIPTION OF THE FIGURES

[0008] In order for the resources of this disclosure mentioned above to be understood in detail, a more specific description of the disclosure, briefly summarized above, may be taken as a reference for the embodiments, some of which are illustrated in the attached figures. It should be noted, however, that the attached figures illustrate only exemplary embodiments and, therefore, should not be considered limiting of its scope, and may admit other equally effective embodiments.

[0009] Fig. 1 is a process diagram of a device according to an embodiment.

[0010] Fig. 2 is a process diagram of a device according to another embodiment.

[0011] Fig. 3 is a flow diagram that summarizes a method Petition 870240083556, dated 09 / 30 / 2024, p. 17 / 58 6 / 35 according to another modality.

[0012] To facilitate understanding, identical reference numbers have been used, whenever possible, to designate identical elements common to the figures. It is contemplated that the elements and characteristics of one modality may be beneficially incorporated into other modalities without additional recitation. DETAILED DESCRIPTION

[0013] Fig. 1 is a process flow diagram of a seawater processing plant 100 according to one embodiment. The seawater processing plant 100 has a seawater inlet 102 to an initial coarse separation vessel 104, where any large solids are settled, deformed, or otherwise removed. An initially purified seawater stream 106 is sent from the initial settling vessel 104 to a chemical treatment unit 108. The chemical treatment unit 108 adds chemical treatment 109 to the seawater to remove any components that may cause problems downstream. For example, in this process there will be downstream membrane filtration units, which may be damaged by specific bacteria and / or ions in the seawater, for example, by contamination of the membranes. The chemical treatment unit 108 may add chemicals to remove or manage them.For example, hypochlorite. Petition 870240083556, dated 09 / 30 / 2024, page 18 / 58 7 / 35 Chlorinated isocyanurates, ozone, and / or other biocides may be added to remove or neutralize bacteria. UV treatment may also be used to control bacteria. Ferric chloride may be used to collect particles. Bisulfite may remove chlorine. Scale inhibitors may be used to minimize scale formation on membranes. The application of any of the above chemicals may be continuous for a period of time or intermittent. The flow rate of chemical treatment 109 may be controlled using flow controller 128.

[0014] Chemically treated seawater is routed to a first membrane unit 110. The first membrane unit 110 filters impurities from the seawater. The first membrane unit may be one or more ultrafiltration units, nanofiltration units, or a combination thereof, arranged in any convenient series, parallel, or combination arrangement. A first permeate stream 112 is a single permeate stream for a membrane or a combination of permeate streams for a plurality of membranes. A first reject stream 114 is a single reject stream for a membrane or a combination of reject streams for a plurality of membranes. The membrane or membranes of the first membrane unit 110 may collect fouling, reducing membrane performance or potentially rendering the membranes unusable, after a Petition 870240083556, dated 09 / 30 / 2024, page 19 / 58 8 / 35 operating period. To prolong the operation of the membranes before they become inoperable due to fouling or other occlusion, the membrane can be rinsed with a first cleaning treatment 111, intermittently or continuously, to remove fouling. Here, the first cleaning treatment 111 is represented as a single stream, but it can be multiple streams with individual flow control if the first membrane unit 110 has multiple membranes. The first cleaning treatment 111 can be added to the first membrane unit 110 at a location upstream of the first membrane unit 110, so that the first cleaning treatment 111 encounters one or more membrane surfaces and removes at least partially fouling.The flow rate of the first cleaning treatment 111 to the first membrane unit 110 is controlled by the first cleaning flow controller 113, which can control multiple streams of the first cleaning treatment 111.

[0015] Temperatures and pressures can be measured at the inlet to the first membrane unit 110 and in the first reject and permeate streams 114 and 112. The flow rates of each stream can also be measured. The purity and impurity content of any or all streams can also be measured. From this information, various methods can be used to infer specific permeabilities of the first membrane unit 110. A set of sensors of Petition 870240083556, dated 09 / 30 / 2024, page 20 / 58 9 / 35 inlet 116 measures inlet flow parameters. A first set of reject sensors 118 measures first reject flow parameters 114. A first set of permeate sensors 120 measures permeate flow parameters 112. Each sensor set includes temperature, pressure, and flow sensors. In addition, the first set of permeate sensors 120 and one or more of the inlet and reject sensor sets 116 and 120 include composition sensors. The first permeate and reject sensor clusters 120 and 118 may include multiple sensor clusters for multiple permeate and reject flows, where the first membrane unit 110 includes multiple membranes. An inlet flow controller 122 controls the flow of the inlet stream to the first membrane unit 110.A controller 130 is operatively coupled to the input sensor array 116, the first permeate sensor array 118, and the first rejection sensor array 120 to receive signals from the sensor arrays representing the physical quantities detected by the various sensors in each array. The controller interprets the signals to resolve estimates of temperature, pressure, flow rate, and composition, as may be relevant for each sensor array.

[0016] The first permeate stream 112 is routed to a second membrane unit 150 which removes different Petition 870240083556, dated 09 / 30 / 2024, page 21 / 58 10 / 35 impurities from the first permeate stream 112. The second membrane unit 150 has one or more sulfate removal membranes and may also have ultrafiltration and / or nanofiltration membranes. A second reject stream 152 and a second permeate stream 154 are emitted from the second membrane unit 150. The second membrane unit 150 may include multiple membranes in any operating ratio. A second set of permeate sensors 156 detects conditions (temperature, pressure, composition) of the second permeate stream 154 and a second set of reject sensors 158 detects conditions (temperature, pressure, optionally composition) of the second reject stream 152.Each set of sensors 156 and 158 is operationally coupled to controller 130, such that controller 130 receives signals from sensor sets 156 and 158 representing physical parameters of the respective second permeate and reject streams 154 and 152 and interprets the signals to resolve the parameters detected by the sensor sets. The second permeate stream 154 is a purified seawater stream that can be routed for further processing.

[0017] A second cleaning treatment 155 is coupled to the second membrane unit 150, with flow controlled by the flow controller 132. Each of the first and second cleaning treatments 111 and 155 may include more than one Petition 870240083556, dated 09 / 30 / 2024, page 22 / 58 11 / 35 chemical product to improve membrane performance. Any number of membranes can be arranged in one processing plant, or in a plurality of processing plants. If each membrane receives a chemical treatment, the chemical treatment of each membrane can be controlled based on the membrane performance indicated by the sensor arrays. As with the first membrane unit, the second membrane unit can feature multiple permeate and reject streams that are combined, and the sensor arrays can include multiple sensors for the multiple permeate and reject streams. Furthermore, the second cleaning treatment 155 can be multiple streams, each with individual flow control, whereas the second membrane unit 150 includes multiple membranes.

[0018] A controller 130 is operatively coupled to control valves 122, 124, 126, 128, and 132 and sensor assemblies 116, 118, 120, 156, and 158 to control the membrane process 100. The controller 130 receives signals from sensor assemblies 116, 118, 120, 156, and 158 representing temperature, pressure, and composition and resolves these operational parameters from the signals. The controller 130 also sends control signals to control valves 122, 124, 126, 128, and 132 based on signals from the sensor assemblies and, optionally, from other calculation instances. Petition 870240083556, dated 09 / 30 / 2024, page 23 / 58 12 / 35 to control the membrane process 100.

[0019] The flow of any component through a membrane, or indeed any medium, is proportional to the pressure drop across the membrane. q = k(âp — Δπ). The proportionality constant k is the membrane permeability with respect to the permeate species. Permeability can be decomposed into an intrinsic permeability K, a temperature compensation Θ(T), and a fouling factor φ. The temperature compensation factor increases with temperature to reflect the increase in permeability with temperature and may be different for different components of the feed stream. The fouling factor φ decreases with decreasing permeability and may have different behavior for different components of the feed stream. Thus, for the complete equation, q = κΘ(T)φ(Δρ — Δπ) gives the permeate flow rate taking into account temperature and fouling. Here, Δρ is the average pressure drop across the membrane. Δρ Pf + PrPv where Pf is the feed pressure, Pr is the reject flow pressure, and Pp is the permeate flow pressure. The osmotic pressure drop can be represented as follows: Petition 870240083556, dated 09 / 30 / 2024, page 24 / 58 13 / 35 Δπ = πρ= [—1.12 (273 + Τ)^mf\ 2Cp + e7YCf + !7YC( 2Cf according to a model. In this model, mf is the molal concentration of ion species in the feed, and Cp, Cf, and Cc are the salt concentrations in the permeate, feed, and concentrate (reject) streams, respectively. Thus, we can calculate a fouling factor from process observations, as follows: Q ω = —7---:—κθ(Δρ — Δπ) where κ is a constant, Δρ is calculated from the process pressures, and Δπ is calculated from the process temperatures and compositions. In the equations above, Y is the permeate recovery from the feed, permeate flow rate divided by feed flow rate. The temperature correction θ can be calculated, in a model, as follows: 1 1θ(T) = e&^298 273+. where A is 2640 if T < 25 °C and 3020 if T < 25 °C. With sensor sets 116, 118, 120, 156, and 158 returning temperatures, pressures, flow rates, and compositions, the equations above can be used to calculate the fouling factor with each reading from the sensor sets. The concentration in the permeate stream, for example of salt, can be estimated, if the sensor accuracy is insufficient, as Petition 870240083556, dated 09 / 30 / 2024, page 25 / 58 14 / 35 follows: _ BS0(T)(Cf + Cr)L= 2Q where B is the mass transfer coefficient of the solute in question.

[0020] Flow rates, temperatures, pressures, and compositions are archived in time series, typically in a digital processing system that may be local to a membrane, centralized in a processing facility, or remotely integrating multiple processing facilities. Physical data are stored with date and time stamps so that time-dependent analysis can be performed. Using the equations above, fouling factors and / or fouling factor parameters can be calculated for each vector of historical data. Fouling factors can be time-related according to any convenient model. Linear models, power models, hyperbolic models, and exponential models, and combinations thereof, can be used. Namely, the fouling factor can be calculated according to any of the following: φ' = 1 — a't (1) 1π(φ2) = 1 — a2t (2) φ3 = 1 — ^tl (3) a^ φ* = ~ (4) φ = Σ=ιφι (5) Petition 870240083556, de 30 / 09 / 2024, pág. 26 / 58 15 / 35 Other models, such as compositions of the simple functions above and piecewise models, can also be used. Supervised and unsupervised machine learning algorithms, such as artificial neural networks, can be used to determine the best shape and fit of the model to historical data. The machine learning system determines a model that predicts the fouling factor as a function of time, fitting and / or combining mathematical operations to historical fouling factor data. Time-dependent parameters are determined from the fit, and dependence on other factors can also be verified by machine learning techniques. For example, the fouling factor may decrease more rapidly over time under certain circumstances, such as low temperature or high concentration, than under other circumstances.In any case, historical fouling factors can be calculated from the physical equations above, and decay parameters and combination parameters, if any, can be determined from statistical treatment. The fouling factor, or membrane performance based on the fouling factor, can be projected into the future to determine when membrane performance will reach a critical point where some intervention is necessary. Namely, if the fouling factor at an endpoint is φ>, the remaining membrane life can be calculated by finding the zero of φ^ί)—φΩ. Petition 870240083556, dated 09 / 30 / 2024, page 27 / 58 16 / 35

[0021] If the remaining lifespan of a membrane is calculated on successive occasions, the chemical treatment can be adjusted according to the movement in the remaining lifespan. For example, if the remaining lifespan decreases more than the time elapsed since the last determination of the remaining lifespan, the chemical treatment can be increased to intensify the remediation of fouling. If the remaining lifespan increases, the chemical treatment can be decreased to save costs. Regression, or principal component analysis, can be used to determine, from historical data, which combination of measures most affects the remaining lifespan of a membrane.In a dataset of flow rates, temperatures, and pressures for a membrane operation, standard transformations can be applied to specific dimensions of the dataset, and regression or principal component analysis can be performed to determine any unknown strong relationships in the dataset, and these relationships can be used to construct control algorithms.

[0022] Fig. 2 is a flow diagram summarizing method 200 according to an embodiment. Method 200 is a method of operation of a seawater purification process by membrane. In 202, a dataset containing operational parameters – temperatures, pressures, flow rates and compositions – of the membranes of the water purification process Petition 870240083556, dated 09 / 30 / 2024, page 28 / 58 17 / 35 of seawater is provided to a digital processing system, for example, by a controller such as controller 130 in Fig. 1, receiving signals from sensor sets such as sensor sets 116, 118, 120, 156, and 158. In 204, the fouling factors and, optionally, the fouling factor decay parameters, are calculated by the digital processing system for each grouping of temperature, pressure, flow rate, and composition, based on the operating parameters using, for example, the equations above and added to the dataset. In 206, the expanded dataset is provided to a machine learning system to determine patterns in the dataset. The machine learning system maintains a model of the dataset. The dataset includes a plurality of parameters that are measured and calculated and that represent the operation of the seawater purification process.The parameters include temperatures, pressures, flow rates, and compositions, and may include fouling factors calculated from the primary parameters. The model may include the same parameters as the dataset. Alternatively, the model may include only principal components of the dataset after performing principal component analysis on the dataset. In this case, the dataset may be transformed on a principal component basis.

[0023] It should be noted that historical factors of fouling Petition 870240083556, dated 09 / 30 / 2024, page 29 / 58 18 / 35 and decay parameters need to be calculated only once and stored with the historical process observations. It is not necessary to recalculate the historical fouling factors from process observations each time method 200 is practiced, provided that the historical fouling factors are available in the dataset.

[0024] The dataset typically contains a temperature for each membrane in the facility, a pressure for feed, permeate, and reject flows for each membrane in the facility, feed, permeate, and reject flow rates for each membrane, and feed compositions and at least one permeate and reject composition for each membrane. The dataset also contains chemical treatment flow rates for each membrane. These parameters constitute the parameter vector. In the case of the dataset being transformed into a principal component basis, a reduced parameter vector can be used. The model is typically based on the same parameter vector or reduced parameter vector.

[0025] Each entry in the model is a calculation based on the data in the dataset and, potentially, on previous model results. Thus, each model entry is based on previous entries in the dataset and in the model. The function of each model is constructed from one or more mathematical operations selected based on the desired model type. The model, in this case, resides in the system of Petition 870240083556, dated 09 / 30 / 2024, page 30 / 58 19 / 35 machine learning, and is a machine learning model that seeks the best calculation of each parameter based on previous data. During each iteration of the model, a plurality of candidate values ​​for each parameter is calculated from a plurality of functions. These functions can be simple individual functions or combinations of functions. A library of functions, such as linear, exponential, periodic, logistic, statistical, and other functions, can be maintained and permuted in the parameter vector. Each function can also be combined with other functions by a list of defined operators, such as addition, multiplication, convolution, etc., and the model can be configured to permute combinations of functions and operators in the search for the best mathematical representation of historical data.

[0026] The data in the database can be updated at different times. The data is typically aggregated by a controller configured to collect data from the processing plant and send control signals to controlled units in the processing plant. For example, the controller might receive an updated value for a first parameter, such as a temperature, after a first duration, while a second parameter, such as a flow rate, might be updated after a second duration different from the first duration. Thus, a record in the dataset might have only one updated parameter value, more than one updated parameter value, or more than one updated parameter value. Petition 870240083556, dated 09 / 30 / 2024, page 31 / 58 20 / 35 parameter updated or all parameters may have updated values. In order to calculate a model input for the same timestamp, input processing may be necessary to construct an input dataset for the model. The controller may perform input processing before invoking the machine learning system, or an input processor may receive data from the controller, prepare the input for the machine learning system, and send the input to the machine learning system. If the model is configured to accept a vector of parameter values ​​as input, the input vector can be constructed from the dataset in any convenient way. For example, the last three available values ​​of each parameter can be calculated, or the last available value of each parameter can be used to form the input vector.If the model is configured to accept a time series dataset, the dataset can be regularized by filling in any missing values ​​in the time series by any convenient means.

[0027] In 208, the machine learning system updates a process model using the new data. The model parameters are recalculated based on the new data, and statistical methods are applied to determine the best model parameters given the new data. When the machine learning system converges on the new model parameters, Petition 870240083556, dated 09 / 30 / 2024, page 32 / 58 21 / 35 The parameters of the old model are replaced by the new model parameters. The parameters of the old model can be archived for trend analysis, if desired. For each model parameter, the machine learning system searches for the best mathematical representation of the parameter's dependence on time or any other variable. Any suitable data mining technique can be used to determine the best result. The machine learning system can reconstruct each predictive function of the model from scratch each cycle, or it can check the validity of the current model using the new data and compare the result to a tolerance to determine whether to keep, update, or rebuild each function. Alternatively, predictive functions can be configured to update on individual schedules so that a predictive function is not updated before its scheduled update time.In such cases, when an update time arrives for a predictive function, more than one new data record may be available to update the predictive function. The machine learning system can be configured to run update cycles periodically and update predictive functions whose scheduled update times have arrived or that are not configured with a schedule. Alternatively, each time a scheduled update time arrives for a predictive function, the machine learning system can update that predictive function. Petition 870240083556, dated 09 / 30 / 2024, page 33 / 58 22 / 35

[0028] In 210, the machine learning system projects the future evolution of fouling factors for each membrane in the processing plant. The projection can be strictly linear, or it can be based on the current model or a combination of current and previous models, potentially weighted according to age.

[0029] In 212, the machine learning system determines a point in time when the future projection of fouling factors for each membrane reaches the tolerance point for the membrane fouling factor.

[0030] In 214, the machine learning system calculates the remaining lifespan for each membrane. The current time is subtracted from the future time at which the model has determined that the fouling factor for each membrane will reach the tolerance point.

[0031] In 216, the calculated remaining lifespan is compared to previously calculated remaining lifespan instances. This activity can be performed by the controller or the machine learning system. In either case, a controlled variable, or plurality of controlled variables, is selected to adjust based on the comparison. A threshold value can be applied to determine whether to change any controlled variables. For example, if the change in the remaining lifespan of a membrane is substantially equal to the time elapsed between Petition 870240083556, dated 09 / 30 / 2024, page 34 / 58 After 23 / 35 iterations of the model, the limit value may prevent any change to the controlled variable.

[0032] If the conditions are met to change a controlled variable to manage fouling factors of any membranes, the fouling factor model function can be used to determine a controlled variable, or plurality of controlled variables, that have the greatest effect on the fouling factor. For example, the machine learning system or the controller can be configured to differentiate the fouling factor model function and select those controlled variables that impact the fouling factor by more than a threshold amount.

[0033] In 218, one or more controlled variables of the processing plant, such as chemical treatment flow rate, temperature and / or feed flow rate, and permeate fluxes, are adjusted to beneficially impact the fouling factor of one or more membranes. The controlled variables may be those determined above in connection with operation 216 to have the greatest impact on the fouling factor. The fouling factor model function can be used to calculate changes in the controlled variables. A constrained optimization can be performed, based on the current values ​​of the controlled variables and any limits thereof, to define new setpoints for the controlled variables that achieve a predetermined change in the factor. Petition 870240083556, dated 09 / 30 / 2024, page 35 / 58 24 / 35 fouling time for the membrane. The controlled variables can be adjusted inversely to the time remaining before the process reaches the limit condition. For example, an adjustment value for one or more controlled variables can be calculated as an inverse ratio to the remaining time. Other inverse relationships can be designed by those skilled in the art.

[0034] Method 200 can be performed locally by a digital processing system located at the processing facility where the membrane in question operates, either in a control center or a calculation center of the processing facility, or method 200 can be performed by a combination of local and remote units. For example, the machine learning system can be remote and can perform machine learning functions for a plurality of processing facilities.

[0035] The machine learning system can be a neural network type system. In a neural network system, a plurality of calculations is performed to define a model relating variables. Here, the neural network can relate process observations, such as temperature, pressure, flow rate, and composition, and equipment parameters, such as thickness, flow area, surface area, and porosity, to the process performance variable, such as fouling factor. The calculations can be arranged as nodes, or simulator nodes, in a neural network, where each node is a calculation that Petition 870240083556, dated 09 / 30 / 2024, page 36 / 58 25 / 35 predicts a set of variables from another set of variables. The calculations can all be based on the same mathematical form, such as exponential, logarithmic, linear, polynomial, power law, etc., or any distribution of mathematical forms can be used in the calculations. The mathematical forms are driven by one or more parameters or coefficients and can be combined using one or more combination parameters. Namely, in a topology, each node will calculate a model, as follows: fi(x) = Aι(χ)Πα2(@)ΠA3(@) - □Ac(@) where □ represents any mathematical operation to combine the mathematical forms 0j(x). The forms 0j(x) each include parameters γ; the operator □ can also include combination parameters F. The input vector x, in this case, is the history of process observations and equipment parameters, or a subset thereof, and the output can be one or more predicted quantities, which may be predictions of certain process attributes. For example, the model may predict permeate flow rate and / or composition, or the model may predict fouling factors and / or fouling factor decay parameters. In other cases, the model may predict process performance improvements due to chemical discharge, if chemical discharge attributes are included in the input vector.

[0036] In the topology above, where each computation node has the Petition 870240083556, dated 09 / 30 / 2024, page 37 / 58 26 / 35 With the same objective of calculating a desired output, refinement nodes can synthesize the output of calculation nodes according to a score determined for each calculation node that represents a level of confidence in the result of the calculation by the calculation node. The refinement node can take, as input, the output of a series of nodes and their scores, and can produce a synthesized version of the calculations performed by the nodes, weighted by their respective scores, essentially providing a synthesized mathematical model.

[0037] In general, each of the nodes in the neural network has a different approach to deriving a relationship between the variables in the dataset, and the output of the nodes is combined and / or selected according to a confidence metric in the result of each node calculation and according to a combination process defined for a group of nodes. The combination processes themselves can be nodes in the neural network, with their own confidence metrics.

[0038] In some cases, a plurality of sites may have membrane processes monitored and controlled by a neural network process, and a collective neural network may aggregate data and results from the local neural networks and produce a combined model. In this scenario, the collective neural network and the local neural networks will have the same topology, and the collective neural network may combine the parameters of the local neural networks. The combination may be a weighted average of the parameters of the Petition 870240083556, dated 09 / 30 / 2024, pages 38 / 58 27 / 35 local neural networks, with weighting provided by a statistical confidence measure, such as the number of data points or other uncertainty metrics or process performance metrics.

[0039] The machine learning system described above can apply direct statistical techniques such as regression and principal component analysis. Other statistical methods can be used to improve the result, taking into account process noise, measurement noise, and other sources of uncertainty. Fig. 3 is a flowchart summarizing a 300 method according to another embodiment. The 300 method is a way to perform the 208 update operation of Fig. 2. Fig. 3 is an example of a recursive Bayesian method that can be used to effectively weight the results based on noise contributions from various sources.

[0040] Method 300 will be explained using an example problem space, where the fouling factor of a membrane separation unit is modeled above from process observations. In the example above, the fouling factor φ is determined independently as a function of temperature, pressure, flow rate, and process composition observations and as a function of time. There is a real fouling factor, which is not known precisely due to noise in the process observations. The real fouling factor, along with the real decay parameters and others Petition 870240083556, dated 09 / 30 / 2024, pp. 39 / 58 28 / 35 time-dependent parameters constitute a state vector Φ = {φ, a, A,...}. Each component of the state vector Φ varies with time, therefore Φ = Φ(3) = {φ(t), a(t), A(t),...}. For method 300, there is a model that calculates Φ at a future time. This model is a state transition function F that maps Φ from one time to another. Furthermore, the membrane process described above is controlled by the operation of the various control valves 122, 124, 126, 128, and 132. Method 300 generally conforms to the Kalman filter application process.

[0041] In 302, a next value of the state function is estimated, as follows: Φ<= FΦt_1+But where U is a vector of control values ​​most recently applied to the process (at time t) and B is a function that maps the control values ​​to their impact on the state vector.

[0042] In 304, the error covariance of the above estimate is also estimated. The error covariance of the previous time point t-1 is projected using the same state transition function, and the process noise contribution is added, as follows: Pt = FPt_'FT+ Q It should be noted here that the state transition function, control application function, and process noise are shown as time-invariant, but they may vary over time. Petition 870240083556, dated 09 / 30 / 2024, pages 40 / 58 29 / 35 in which case these parameters in the equations above will have values ​​at the times indicated t-1. This is a mistaken assumption for the purposes of clear explanation.

[0043] These two calculated parameters are then refined based on new data obtained in the process. As noted above, the fouling factor can be calculated from new observations of the process. This calculated fouling factor will be different from the fouling factor estimated in the state vector. A recursive Bayesian technique can be used to appropriately evaluate the noise in the observations and, in general, in the process to find the estimate of the fouling factor, decay parameter, and combination factors with greater statistical confidence.

[0044] In 306, a residual is calculated that represents the difference between the detected process observations and the process observations implied by the predicted state vector. The calculation is as follows: Ot =zt — QGt Where z is a vector of the latest process observations (temperature, pressure, flow rate, composition) and H is a function that maps (reverse maps) the state vector to the process observations. Since the state vector includes the fouling factor, along with decay parameters and combination parameters that may not be observed as dependent on the process observations, H may have... Petition 870240083556, dated 09 / 30 / 2024, pp. 41 / 58 30 / 35 only one line that is not zero.

[0045] In 308, the covariance of the residual above is calculated as follows: St = R + HPtHT Here, R is the noise covariance of process Q.

[0046] A correction factor is calculated at 310. The correction factor represents a residual treatment calculated above that is used to correct the state vector prediction. In many cases, the Kalman gain Kt = PtHTS+' is used, but other correction factors can also be used. The Kalman gain takes a ratio of the estimated state vector covariance to the residual covariance and uses the process observation reverse mapping function to express this ratio in a way that can be used to refine the estimated state vector. In any case, the correction factor is used to apply some weighting to the residual resulting from the deviation of the new process observations from the process observations that would be recorded if the estimated state vector were 100% accurate to update the estimated state vector. In this way, the estimated state vector is partially shifted towards the new process observations.

[0047] In 312, the correction factor is applied to refine the estimated state vector, as follows: Petition 870240083556, dated 09 / 30 / 2024, pp. 42 / 58 31 / 35 G*t = Gt + xtyt where the correction factor Xt can be the Kalman gain, as described above, or another convenient correction factor. The correction factor Xt has the form of a linear map that transforms the residual observations of the process into fits to the components of the state vector. Of course, the correction factor Xt can be a matrix of numbers, as in the Kalman gain, or it can be a set of linear or non-linear functions that perform similar transformations. The correction factor can be structured to emphasize and de-emphasize any factors to any desired extent.

[0048] In 314, a refined covariance estimate of the refined state vector is calculated as follows: P*t= (I- XtH)Pt(I - XQ + X^X. expressing that the covariance of the estimated state vector is refined essentially by applying the same correction factor to incorporate the covariance of the process noise.

[0049] Optionally, in 316, the final (refined) observation residual can be calculated as follows: yt=Zt-HGt where the refined state vector is transformed into the observation space by the reverse map function of the state vector and compared with the most recent measurements. This is the part of these measurements that is not accounted for by the updated model.

[0050] Where nonlinear state transition models Petition 870240083556, dated 09 / 30 / 2024, pp. 43 / 58 32 / 35 (fouling factor time prediction and decay parameters) are used, versions of the above process can be used where the state transition function is locally linearized, such as extended or unscented Kalman filters. Other variations may include outlier detection in historical data before prediction, smoothing and / or dimensioning of historical data, and time-weighting or supporting historical data. For example, when forming or updating the statistical model of time-dependent state variables, data older than a predetermined threshold can be ignored. Alternatively, the contribution of the data to the statistical analysis can be time-weighted. If, for example, equation (4) above is used to fit historical fouling factors to time records, a simple regression seeks to minimize — atj)2. Instead, weighting factors can be included in the objective function to de-emphasize older data. In one example, y2 Σ(1ηφι — ati)2 (t- 3d The quantity itself (t — ti) can be weighted to increase or decrease its effect on the regression.

[0051] Methods 200 and 300 can be practiced using observational data from different membrane processes. Petition 870240083556, dated 09 / 30 / 2024, pages 44 / 58 33 / 35 performing generally similar separations. For example, a plurality of membrane seawater purification processes may operate at one location, or at more than one location. Each process records process observations, stores the observations locally, and has a local controller to control the process. A centrally located digital processing system can receive process observations from all membrane seawater purification processes into a database for processing as an aggregated dataset. If the source of each data record is recorded with the data record, the modeling and prediction processes of methods 200 and 300 can be performed for the entire aggregated dataset or on subsets based on location and / or region, which can be named according to any geographic indicator (latitude, longitude, altitude, etc.).For example, a first statistical model can be solved from a first subset of the aggregated dataset based on a first data attribute, while a second statistical model is solved from a second substrate of the aggregated dataset based on a second data attribute. The first and second data attributes can be numerical or non-numerical attributes, such as latitude, longitude, altitude, ambient temperature, operational team, weather (rain, snow), static or dynamic foundation slope, variation in parameters of... Petition 870240083556, dated 09 / 30 / 2024, pages 45 / 58 34 / 35 process, type of personnel scheduling (number of different operational teams per day), chemical suppliers, or any conceivable attribute of the data. In this way, statistical models describing the evolution of membrane process operation over time can be compared to discover potential causes of variation in membrane life between different processes.

[0052] It should be noted that the models described in this document are models for calculating and predicting fouling factors and decay parameters for membrane processing units. It is possible to use other models with recursive Bayesian techniques. For example, fouling factor data before and after a chemical discharge, along with chemical discharge flow rate, temperature, and time, can be provided to an artificial neural network to model the effect of chemical discharge on the fouling factor. In another example, fouling factor decay parameters before and after chemical discharge, along with chemical discharge rate, time, and temperature, can be provided to an artificial neural network to model the effect of chemical discharge on fouling factor decay parameters.In such cases, a shallow neural network can be constructed in which the input is the fouling factor, decay parameter, and process observations immediately before and after the chemical discharge, as well as the... Petition 870240083556, dated 09 / 30 / 2024, pages 46 / 58 35 / 35 parameters of the chemical discharge. Alternatively, a convolutional neural network can be additionally constructed, accepting as input the time series evolution of the fouling factor and membrane conditions as a two-dimensional input set. The convolutional layers of the neural network are then used to extract spatial correlations in the membrane history and to predict the conditions for maximum effectiveness of a future chemical treatment.

[0053] Although the above refers to embodiments of the present invention, other and additional embodiments of the present disclosure may be conceived without departing from the scope of the present instrument, and the scope thereof is determined by the claims below.

Claims

CLAIMS 1. A method of operating a seawater processing plant, characterized in that it comprises: directing a seawater feed stream, via an inlet of a seawater processing plant, into a separation vessel configured to perform membrane purification and comprising one or more membrane units, wherein the one or more membrane units comprise one or more membrane filters; recovering, via the one or more membrane filters, purified seawater and concentrated seawater; determining, via one or more sensors disposed in the one or more membrane units, operational parameters of the one or more membrane units, wherein the one or more sensors are configured to generate sensor data related to the one or more operational parameters;To determine an indicative fouling factor of the condition of one or more membrane filters based on determined operating parameters, wherein the determined operating parameters comprise one or more parameters related to seawater composition, an ambient temperature of the separation vessel, a seawater temperature, a feed stream flow rate, a permeate stream, or both, a pressure across the filter, or any combination thereof; Petition 870240083556, dated 09 / 30 / 2024, p. 48 / 58 2 / 4; to use a statistical model to resolve a time dependence of the condition of one or more membrane filters; to use a recursive statistical process to update the statistical model; to use the updated statistical model to predict a future condition of one or more membrane filters; to compare the predicted future condition of one or more membrane filters with a limiting condition;To predict a remaining time before one or more membrane filters reach the limit condition; and to adjust at least one of the following: ambient temperature of the separation vessel, flow rate of the feed stream, or flow rate of a permeate stream based on the remaining time.

2. Method, according to claim 1, characterized in that the recursive statistical process is a Bayesian process.

3. Method, according to claim 1, characterized in that the recursive statistical process uses the Kalman filter.

4. Method, according to claim 1, characterized in that the statistical model is non-linear.

5. Method, according to claim 4, characterized in that the recursive statistical process Petition 870240083556, dated 09 / 30 / 2024, page 49 / 58 3 / 4 is a Bayesian process.

6. A method according to claim 1, characterized in that it further comprises: flowing chemical treatment to a membrane of the membrane purification process; and adjusting a flow rate of the chemical treatment based on the remaining time.

7. Method according to claim 6, characterized in that the adjustment of the flow rate of the chemical treatment comprises calculating an adjustment amount for the flow rate as an inverse relationship to the remaining time and applying the adjustment amount to the flow rate of the chemical treatment.

8. Method, according to claim 1, characterized in that the statistical model is exponential.

9. A method according to claim 1, characterized in that using a statistical model to resolve the time dependence of the condition of one or more membrane filters comprises aggregating process observations from the one or more membrane filters and solving the statistical model from the aggregated observations.

10. Method, according to claim 9, characterized in that solving the statistical model from the aggregated observations comprises solving a first statistical model based on a first subset of the aggregated observations based on a first attribute of the aggregated observations and a second statistical model based on a second subset of the aggregated observations based on a second attribute of the aggregated observations.

Images