Circulating cooling water scaling model prediction method and system
Through training and optimizing neural network algorithms, a method and system for scaling model prediction methods and systems for cyclic cooling water are established, which solves the problem of difficult and high cost of scaling judgment in the circulating cooling water system in the existing technology, and achieves fast, accurate and low-cost Scale determination effect.
Patent Information
- Application Number
- CN202411502068.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-25
- Publication Date
- 2025-02-07
AI Technical Summary
The prior art is difficult to quickly and effectively judge the scale condition of the circulating cooling water system, and the existing methods are costly and slow to respond, making it difficult to meet the real-time monitoring needs of the cooling water system.
Using training-optimized neural network algorithm, a cyclic cooling water scale model prediction method and system is established through gradient estimation and backpropagation methods, and using physical quantities such as pH, temperature and concentration multiple as inputs to predict scaling components and scaling quantity.
It realizes the rapid, accurate and low cost of scaling judgment in the circulating cooling water system, and effectively improves the scaling judgment ability of the cooling water system.
Smart Images
Figure CN119397902A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the field of theoretical models of crystal scaling growth, and in particular to a circulating cooling water scaling model prediction method and system. Background Art
[0002] During the circulating cooling process, the circulating cooling water evaporates continuously. The evaporation process does not take away the salt in the water, resulting in the concentration of salts and organic matter. When the circulating water concentration exceeds the saturated pH value, calcium carbonate is in an oversaturated state and crystallizes out of the water, depositing on the heat transfer surface to form scale, which is prone to scaling and corrosion problems, resulting in scaling and corrosion in the metal components, fillers, pools and walls of the cooling tower. Existing methods for judging scaling and corrosion, such as the detection test method and the hanging plate method, have their own defects. Although the detection test method uses sensors, it is easily affected by water quality, resulting in probe damage, and the maintenance cost is high, which limits its possibility of large-scale promotion. The hanging plate method has a long experimental cycle, usually requiring a sampling cycle of more than 1 month. For a long time, relevant researchers have conducted a lot of research on the scaling process and proposed a series of theoretical models to describe the growth of crystal scaling. However, since the growth process of scaling thermal resistance is extremely complex and the measurement process is difficult to repeat, there are still many challenges in establishing an accurate mathematical model to describe the growth process of scaling thermal resistance. Therefore, it is urgent to establish a low-cost, fast-response circulating cooling water corrosion scaling model prediction system to effectively improve the scaling judgment ability of the cooling water system. Summary of the invention
[0003] The purpose of the present invention is to overcome the problems existing in the prior art and provide a circulating cooling water scaling model prediction method and system, which can quickly and effectively improve the scaling judgment ability of the cooling water system while reducing costs.
[0004] In order to achieve the above-mentioned object, the first aspect of the present invention provides a circulating cooling water scaling model prediction method, which adopts a training optimization neural network algorithm for gradient estimation, and the neural network algorithm comprises an input layer, a hidden layer and an output layer.
[0005] Preferably, the gradient estimation is: using the gradient descent method to calculate the network parameters for updating, the network parameters include: the number of learning rounds Epoch, the number of neurons j in each hidden layer, and the number of input units of the neurons.
[0006] Preferably, three physical quantities in the form of R3 are used as the input layer, and the three physical quantities include pH, temperature and concentration multiple.
[0007] Preferably, the hidden layer includes a first hidden layer and a second hidden layer. The first hidden layer using 2-3 groups of R6 forms corresponds to physical implicit interactions. The first hidden layer can accommodate enough weights and biases to meet the nonlinear relationship between pH and temperature on the amount of scaling and the peak properties. The second hidden layer using 1 group of R12 forms corresponds to the scaling tendency of the components to be scaled. Between the first hidden layer and the second hidden layer, an exponential activation function is added. As a key component in the neural network, it plays the role of introducing nonlinear factors, thereby increasing the expression ability of the neural network and the ability to fit complex functions.
[0008] Preferably, the output layer is a 12-group scaling component of the R12 form.
[0009] Preferably, the update of the network parameters is based on the changes of the ring cooling water, including residence time, concentration multiple, pH, and temperature. The model predicts that the scaling components include CaCO3, CaCO3(*), CaSO4, CaSO4·2H2O, Ca3(PO4)2, SiO2, Fe(OH)3, FeCO3, FePO4, MgSiO3, Mg(OH)2, and hydroxyapatite.
[0010] Preferably, the model training uses the average standard deviation of the predicted scaling amount of each scaling substance as a loss function, and minimizes the loss function in a back-propagation manner.
[0011] Preferably, the model training uses 70% of the data as a training optimization set and 30% of the data as a test set until the training optimization parameters are stable and the loss function is minimized to a specified threshold.
[0012] Preferably, the method further comprises: using the test set data to evaluate the accuracy of the prediction, specifically: using the convergence of the loss function as an indicator of the training optimization process, and using the correlation between the predicted value and the experimental value as an evaluation indicator.
[0013] A second aspect of the present invention provides a circulating cooling water scaling model prediction system, which is used to execute the steps of the above method.
[0014] According to the thermodynamic principle of solution, the machine learning method based on neural network was used to collect water quality data from a certain UHV converter station, and the scaling model of the solution established according to the given data set was trained, optimized and evaluated. The main physical parameters of the solution include retention time, concentration multiple, pH, and temperature. The model training optimization and predictable scaling components include CaCO3, CaCO3(*), CaSO4, CaSO4·2H2O, Ca3(PO4)2, SiO2, Fe(OH)3, FeCO3, FePO4, MgSiO3, Mg(OH)2, and Hydroxyapatite.
[0015] The data extracted on site may have significant deviations and omissions due to some special reasons. Therefore, before selecting the data, we first eliminate and fill in the gaps using the average method. The elimination principle is to eliminate data that does not meet the water quality parameter operation indicators. The specific parameter operation indicators are: pH∈(7.2-11), Ca 2+ ≥120, total phosphorus∈(8.0-14), concentration multiple ≥3.
[0016] Since the source of water quality samples is stable and the specific concentration of each solute has little variability, within the given data range, the pH, temperature, and concentration multiple are approximately independent variables, but when it comes to scaling, there are certain implicit influencing factors. Taking into account the principles of solution thermodynamics, for multivariate solutions, the crystallinity of the solution is a function of the concentration of all solutes, and there are usually fourth-order interactions between solutes. At this time, a neural network model with three hidden layers is applicable, and the implicit influencing factors are incorporated into the neural network structure. On this basis, training and optimization are carried out to obtain a specific prediction model for samples from this source.
[0017] Python is used as the main programming language, PyTorch as the main neural network framework, and scipy, sklearn, and NumPy as the main tools for programming. The code level is finally encapsulated into classes and high-level functions, so that the data in the training optimization and analysis process are stored separately from the main structure of the code. The program is packaged as a binary executable program, and a command line interface is reserved for calculating and predicting the amount of scaling in other cases. In the model training optimization, 70% of the data will be used as the training optimization set and 30% of the data will be used as the test set. Until the parameters of the training optimization are stable and the loss function is minimized to a certain threshold. The test set data is used to evaluate the accuracy of the prediction. The convergence of the loss function is used as an indicator of the training optimization process. The correlation between the predicted value and the experimental value is used as an evaluation indicator.
[0018] Compared with the prior art, the present invention has the following beneficial effects:
[0019] (1) Traditional machine learning algorithms such as support vector regression (SVR) and adaptive boosting algorithm (Adaboost) perform poorly when processing large data model training and optimization. The present invention uses the more widely applicable BP (Backpropagation) neural network algorithm to establish a more efficient, concise and accurate cooling water scale prevention prediction model.
[0020] (2) Compared with traditional methods with long experimental cycles such as the hanging plate method, the established circulating cooling water corrosion scaling model prediction system has lower cost and faster response, and can effectively improve the scaling judgment ability of the cooling water system.
[0021] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS
[0022] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the embodiments of the present invention, but do not constitute a limitation on the embodiments of the present invention. In the accompanying drawings:
[0023] Figure 1 It is a schematic diagram of the initial design of the neural network model;
[0024] Figure 2 is a graph showing changes in the loss function during the optimization process of an embodiment of the present invention;
[0025] Figure 3 1 is a schematic diagram of the correlation between the experimental value and the predicted value of the scale amount in the embodiment of the present invention;
[0026] Figure 4 It is a schematic diagram of the correlation between the experimental values and the predicted values after grouping in an embodiment of the present invention. DETAILED DESCRIPTION
[0027] The technical solutions in the embodiments of the present invention are described clearly and completely below. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0028] The first aspect of an embodiment of the present invention provides a method for predicting a scaling model of circulating cooling water, which adopts a gradient estimation optimization method for a neural network algorithm: a BP (Backpropagation) neural network algorithm, which mainly includes an input layer, a hidden layer, and an output layer. The optimization algorithm uses a gradient descent method to calculate the update of network parameters. For a single input and output example, back propagation is an efficient algorithm for calculating the gradient of the loss function relative to the network weights. It calculates the gradient layer by layer, iterating from the output layer to the input layer to avoid the chain rule of redundant calculations of intermediate terms. The specific gradient descent method includes calculating the derivative of the loss function relative to the network weights. This is usually done using back propagation. Assuming there is an output neuron, the square function of the error is (Formula 1):
[0029] E=L(t,y) (1)
[0030] Where: L is the output loss y and the target value loss t;
[0031] t is the target output of the training optimization sample;
[0032] y is the actual output of the output neuron.
[0033] For each neuron j, its output o j Defined as (Formula 2):
[0034]
[0035] Where: The activation function φ is nonlinear and differentiable on the activation area (ReLU is not differentiable at one point). j is the weight of the output; if the neuron is in the first layer after the input layer, x k The input layer is the input k of the network. The number of neuron input units is n. The variable w kj Represents the weight between the previous layer neuron k and the current layer neuron j. The initial design of the neural network model is as follows Figure 1 shown.
[0036] (2) Three physical quantities (pH, temperature, and concentration factor) in the form of R3 are used as the input layer; 2–3 groups of the first hidden layer in the form of R6 correspond to the physical implicit interactions and can accommodate enough weights and biases to meet the nonlinear relationship between pH and temperature on the amount of scaling and the peak properties; 1 group of the second hidden layer in the form of R12 corresponds to the scaling tendency of the components to be scaled; the output layer is 12 groups of scaling components in the form of R12.
[0037] (3) Based on the thermodynamic principles of the solution, the initial structure model of the solution was established using the water quality data collected on-site from a certain UHV converter station. Then, a machine learning method based on a neural network was used to continuously optimize the scaling model. The main physical parameters of the solution include retention time, concentration multiple, pH, and temperature. The model training optimization and predictable scaling components include CaCO3, CaCO3(*), CaSO4, CaSO4·2H2O, Ca3(PO4)2, SiO2, Fe(OH)3, FeCO3, FePO4, MgSiO3, Mg(OH)2, and Hydroxyapatite. Before selecting the data, it was first eliminated and the average method was used to complete it. The elimination principle is to eliminate data that does not meet the water quality parameter operating indicators. The specific parameter operating indicators are: pH∈(7.2-11), Ca 2+ ≥120, total phosphorus∈(8.0-14), concentration multiple ≥3.
[0038] (4) Mass balance principle: The net scaling rate is the difference between the deposition rate and the erosion rate (Formula 3) as given below:
[0039]
[0040] If the degree of scaling is measured by the change in fouling thermal resistance, the scaling rate can be expressed as (Formula 4):
[0041]
[0042] Where: m f is the mass of dirt deposited on the unit heat exchange surface, kg / m 2 ;
[0043] m d is the mass deposition rate, kg / (m 2 s);
[0044] m r is the mass erosion rate;
[0045] t is time, s;
[0046] ρ f is the dirt density, kg / m 3 ;
[0047] λ f is the thermal conductivity of the dirt, W / (m·K);
[0048] R f is the fouling thermal resistance, m 2 K / W;
[0049] Φ d is the deposition rate, (m·K) / N;
[0050] (5) Diffusion reaction principle: The equation for the scaling rate of CaCO3 under different conditions is (Formula 5):
[0051]
[0052] Where: K R is the constant of crystallization rate;
[0053] K sp is the volume of CaCO3 at the solid-liquid interface;
[0054] [Ca 2+ ] i [CO3 2- ] i Ca 2+ and CO3 2- interface concentration.
[0055] (6) Surface reaction principle:
[0056] Where: R f is the fouling thermal resistance, m 2 K / W;
[0057] c1, c2 are constants;
[0058] P d is the probability function;
[0059] Ω is the characteristic factor of water;
[0060] E is the activation energy;
[0061] R is the universal gas constant;
[0062] T is temperature;
[0063] Ψ is a function of the strength of the fouling layer and is affected by flow rate and water quality.
[0064] τ is the local stress on the dirt wall;
[0065] x f is the thickness of the dirt layer.
[0066] (7) In the modeling process, Python is used as the main programming language, PyTorch is used as the main neural network framework, and scipy, sklearn, and NumPy are used as the main tools for programming. The code level is finally encapsulated into classes and high-level functions so that the data in the training optimization and analysis process are stored separately from the main structure of the code. The program is packaged as a binary executable program, and a command line interface is reserved for calculating and predicting the amount of scaling in other situations. In model optimization, 70% of the data will be used as the training optimization set, and 30% of the data will be used as the test set. Until the optimized parameters are stable and the loss function is minimized to a certain threshold, the loss function changes during the optimization process as shown in the following figure. Figure 2 As shown in the figure, the loss function is close to convergence at about the 100th step of training, and finally reaches the threshold at 2200 steps. The test set data is used to evaluate the accuracy of the prediction. The convergence of the loss function is used as an indicator of the training optimization process. The correlation between the predicted value and the experimental value is used as an evaluation indicator. The correlation between the experimental value and the predicted value of the scale is summarized as follows Figure 3 As shown in the figure, the effect is best when the scaling amount is 200-800. The overall mean absolute percentage error (MAPE) is 7.53%, the determination coefficient (R2) is 0.985, and the overall warning accuracy rate reaches 92.47%. There is a good correlation between the simulated prediction value and the experimental measurement value. It is worth noting that most scaling prediction algorithms based on traditional classical theory usually have an absolute percentage error of 10%-20%.
[0067] (8) Since the scaling components account for different proportions of the total scaling, it is necessary to analyze the prediction errors of different key components. The prediction results are divided into 6 groups, namely CaCO3, CaCO3(*), Fe(OH)3, FeCO3, MgSiO3 and other five main contributing components. The correlation between the experimental values and the predicted values after grouping is shown in the figure. Figure 4 As shown in the figure, it can be seen that for the components with larger scaling amounts, CaCO3, CaCO3(*), and Fe(OH)3, the prediction accuracy is the best, and their determination coefficients are 0.991, 0.991, and 0.986, respectively. For the components with larger scaling amounts, CaCO3, CaCO3(*), and Fe(OH)3, the prediction accuracy is the best.
[0068] like Figure 3 As shown, the overall mean absolute percentage error (MAPE) of the circulating cooling water scaling model prediction method of the present invention is 7.53%, the determination coefficient (R2) is 0.985, and the overall warning accuracy rate reaches 92.47%. A more efficient, concise and accurate cooling water scaling prediction model is established. Compared with traditional methods with long experimental cycles such as the hanging plate method, the established circulating cooling water corrosion scaling model prediction system has lower cost and faster response, and effectively improves the scaling judgment ability of the cooling water system.
[0069] Based on the same inventive concept, a second aspect of an embodiment of the present invention provides a circulating cooling water scaling model prediction system for executing the steps of the above method.
[0070] In summary, the beneficial effects of the present invention are: (1) Traditional machine learning algorithms such as support vector regression (SVR) and adaptive boosting algorithm (Adaboost) perform poorly when processing large data model training optimization. The present invention uses a more widely applicable BP (Backpropagation) neural network algorithm to establish a more efficient, concise and accurate cooling water scale prevention prediction model. (2) Compared with traditional methods with long experimental cycles such as the hanging plate method, the established circulating cooling water corrosion scaling model prediction system has lower cost and faster response, and effectively improves the scaling judgment ability of the cooling water system.
[0071] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity, or apparatus.
[0072] The above are only embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included within the scope of the claims of the present application.
Claims
1. A method for predicting scaling model of circulating cooling water, characterized in that: A training optimization neural network algorithm is used for gradient estimation, and the neural network algorithm comprises an input layer, a hidden layer and an output layer.
2. The method according to claim 1, characterized in that The gradient estimation is: using the gradient descent method to calculate the network parameters for updating, the network parameters include: the number of learning rounds Epoch, the number of neurons j in each hidden layer, and the number of input units of the neurons.
3. The method according to claim 2, characterized in that Three physical quantities in the form of R3 are used as the input layer, and the three physical quantities include pH, temperature and concentration multiple.
4. The method according to claim 2, characterized in that: The hidden layer includes a first hidden layer and a second hidden layer, wherein the first hidden layer in the form of 2-3 groups of R6 corresponds to the physical implicit interaction, and the first hidden layer can accommodate enough weights and biases to meet the nonlinear relationship between pH and temperature and the scaling amount and the peak property; A second hidden layer in the form of R12 is used to correspond to the scaling tendency of the components to be scaled; an exponential activation function is added between the first hidden layer and the second hidden layer. As a key component in the neural network, it plays the role of introducing nonlinear factors, thereby increasing the expression ability of the neural network and the ability to fit complex functions.
5. The method according to claim 2, characterized in that: The output layer is a 12-group scaling component in the form of R12.
6. The method according to any one of claims 1 to 5, characterized in that: The updating of the network parameters is based on the changes of the ring cooling water, including residence time, concentration multiple, pH, and temperature. The model predicts that the scaling components include CaCO3, CaCO3(*), CaSO4, CaSO4·2H2O, Ca3(PO4)2, SiO2, Fe(OH)3, FeCO3, FePO4, MgSiO3, Mg(OH)2, and hydroxyapatite.
7. The method according to claim 6, characterized in that The model training uses the average standard deviation of the predicted scaling amount of each scaling substance as the loss function, and minimizes the loss function in a back-propagation manner.
8. The method according to claim 7, characterized in that The model training uses 70% of the data as a training optimization set and 30% of the data as a test set until the training optimization parameters are stable and the loss function is minimized to a specified threshold.
9. The method according to claim 7, characterized in that: The method also includes: using the test set data to evaluate the accuracy of the prediction, specifically: using the convergence of the loss function as a training optimization process indicator, and using the correlation between the predicted value and the experimental value as an evaluation indicator.
10. A circulating cooling water scaling model prediction system, characterized in that: The system is used to execute the steps of the method according to any one of claims 1-9.