CN115936227A - Machine learning-based ion permeation power generation prediction system and method - Google Patents

Abstract

Disclosed is a prediction system and method for ion permeation power generation based on machine learning. In the prediction system, the data acquisition module is configured to obtain sample data for ion permeation power generation, and the neural network module is configured to construct a prediction neural network for ion permeation power generation based on artificial neural networks. Network and neural network training, the neural network module connected to the data acquisition module includes multi-layer network layers, its substructure includes an input layer, an output layer and four fully connected layers, the input layer neurons receive training data, hidden Each neuron in the containing layer and the output layer is fully connected with all neurons in the adjacent layer, and the neural network training is performed based on the sample data to form a predictive neural network for ion osmosis power generation; the model test module configuration adjusts ion osmosis power generation The hyperparameters in the prediction neural network to obtain the ion permeation power generation prediction model. The accuracy of the predictive model reached 97.6%.

Description

Prediction system and method of ion permeation power generation based on machine learning

Technical Field

The present invention relates to the field of ion permeation power generation, and in particular to a prediction system and method for ion permeation power generation based on machine learning.

Background Art

Ion permeation power generation is an energy conversion method that uses concentration gradient to drive the migration of ion carriers and selectively direct them through porous nano-membranes to form ion currents, directly converting salt difference energy into electrical energy. The energy conversion efficiency of the directional ion migration process is high (theoretical efficiency can reach 50%), which can provide a new way to utilize low-grade energy (salinity difference energy, solar energy, etc.). The physical essence of ion permeation power generation is the coupled ion migration, heat transfer and energy conversion process with electrolyte ions and nanoporous media as energy carriers and heat and mass transfer carriers respectively.

At present, the research on ion permeation power generation is mainly focused on new nanochannel materials, focusing on the performance of porous nano-ion selective membrane materials under salinity differences, and changing the concentration ratio and channel geometry to obtain better power generation capabilities. However, there is still a lack of a unified temperature-related performance theory, nor a comprehensive empirical formula to link these factors to conversion efficiency. It is urgent to explore the ion permeation energy conversion mechanism, clarify the physical nature of internal heat and mass transport and energy conversion processes, and ultimately form an ion permeation energy prediction system to accurately evaluate the ion permeation power generation efficiency.

The information disclosed in this Background section is only for enhancement of understanding of the background of the invention and therefore it may contain information that does not form the prior art that is already known to a person of ordinary skill in the art.

Summary of the invention

In view of the shortcomings or defects of the prior art, a prediction system for ion permeation power generation based on machine learning is provided. A simplified dimensionless control parameter group for ion permeation power generation is obtained through similarity principle, and sample expansion is performed to obtain a large number of experimental samples as a data set for machine learning. The machine learning artificial neural network is used to take the dimensional variables of the dominant factor physical parameters of the normalized preprocessing as input, and the output power and energy conversion efficiency as output parameters to establish a prediction model between the nanochannel geometry, concentration difference, ion properties and energy conversion performance. After the model parameters are trained, a high-precision, high-acceleration machine learning-based ion permeation power generation prediction model is finally obtained, which can guide the judgment of the dominant parameters of the permeation power generation system and the estimation of the efficiency.

The purpose of the present invention is achieved through the following technical solutions.

A prediction system for ion permeation power generation based on machine learning includes:

A data acquisition module, configured to acquire sample data of ion permeation power generation, the data acquisition module comprising:

a measuring unit that measures a plurality of physical parameters of ion-permeation power generation;

A dimensionless conversion unit connected to the measuring unit and generating a plurality of dimensionless parameters based on the plurality of physical parameters and forming a dimensionless array based on the dimensionless parameters;

A sample expansion unit connected to the dimensionless conversion unit and expanding the dimensionless array using a similar principle to obtain an expanded sample, wherein the expanded sample and the dimensionless array constitute sample data, and the sample data is divided into training data and test data according to a predetermined ratio;

a normalization processing unit connected to the sample expansion unit and normalizing the sample data;

A neural network module, which is configured to construct a prediction neural network and neural network training for ion permeation power generation based on an artificial neural network ANN, the neural network module connected to the data acquisition module includes multiple network layers, and its substructure includes an input layer, an output layer and four fully connected layers, the input layer neurons receive training data, each neuron in the hidden layer and the output layer is fully connected to all neurons in the adjacent layers, and the neural network training is performed based on the sample data to form a prediction neural network for ion permeation power generation;

A model testing module, configured to adjust hyperparameters in a prediction neural network of ion permeation power generation to obtain an ion permeation power generation prediction model, the model testing module comprising:

A hyperparameter optimization unit connected to the neural network module, the hyperparameter optimization unit adopts a combination of GPU parallel acceleration, activation function nonlinearization and back propagation algorithm optimization to update the weight w and bias b to optimize the prediction neural network of ion permeation power generation;

A testing unit is connected to the data acquisition module and the neural network module to test the prediction neural network of the ion permeation power generation based on the test data until the error reaches the expectation, and the model is saved as the ion permeation power generation prediction model.

In the prediction system of ion permeation power generation based on machine learning, the measurement unit includes a first measurement part for obtaining dielectric constant, concentration, ion diffusion coefficient, nanochannel length, radius and surface charge density, a temperature sensor for measuring temperature data, a power level for measuring output power P and a second measurement part for measuring energy conversion efficiency η.

A prediction method using the ion permeation power generation prediction system based on machine learning includes:

S100: a measuring unit measures a plurality of physical parameters of ion permeation power generation, constructs a dimensioned equation for ion permeation power generation based on the physical parameters, and a dimensionless conversion unit generates a dimensionless equation for ion permeation power generation and its corresponding dimensionless array based on the dimensioned equation for ion permeation power generation;

S200: The sample expansion unit expands the dimensionless array by adopting a similarity principle to obtain an expanded sample, wherein the expanded sample and the dimensionless array constitute sample data, and the normalization processing unit normalizes the sample data;

S300: the sample data is divided into training data and test data according to a predetermined ratio, the neural network module is configured to construct a prediction neural network for ion permeation power generation based on an artificial neural network, and a neural network training is performed based on the sample data to construct a prediction neural network for ion permeation power generation, wherein the output power P and the energy conversion efficiency η are output parameters, and the concentration C, the high temperature end temperature _Th , the low temperature end temperature _Tl , the ion diffusion coefficient D, the dielectric constant ε, the nanochannel length L, the nanochannel radius R and the surface charge density σ are dimensional variables as inputs;

S400: The hyperparameter optimization unit uses GPU acceleration, activation function nonlinearization, and a combination of back-propagation algorithms to update weights w and bias b to optimize the prediction neural network for ion permeation power generation.

S500: The testing unit tests the optimized ion permeation power generation prediction neural network based on the test data, adjusts the hyperparameters in the network according to the training results, repeats the iteration until the error reaches the expected level, and saves the model as the ion permeation power generation prediction model.

In the prediction method, the dimensionless control equation and its corresponding dimensionless array are:

Dimensionless Poisson's equation:

The dimensionless Nernst–Planck equation:

Dimensionless continuity equation:

The dimensionless Navier-Stokes equations:

Dimensionless energy equation:

Dimensionless array:

为偏微分算子，ε为介电常数，φ为电势，F为法拉第常数，c_i为第i种离子的离子浓度，z_i为第i种离子的价电荷数，D_i为第i种离子扩散系数，α_i为第i种离子的简化Soret系数(其中i＝1表示阳离子，i＝2表示阴离子)，σ为纳米通道表面电荷密度，L和R分别表示纳米通道长度和半径，u为速度，R_g为通用气体常数，T为温度，p为压力，μ为动力粘度，ρ为密度，C_p为比热，k_f和k_s分别为流体和固体导热系数，σ_f为电导率，ΔT为温度差，C_h为高浓度侧浓度，C_l为低浓度侧浓度。带*上标的参数为无量纲参数，ε^*，

φ^*，c_i ^*，u^*，D_i ^*，p^*，μ^*，σ_f ^*，Θ分别为无量纲介电常数、无量纲偏微分算子、无量纲电势、无量纲浓度、无量纲速度、无量纲扩散系数、无量纲压力、无量纲粘度、无量纲电导率、无量纲过余温度，带m下标的参数为特征参数，D_i，m为特征扩散系数(其中i＝1表示阳离子，i＝2表示阴离子)，C_m，μ_m，ε_m，T_m，σ_f，m分别为特征浓度、特征粘度、特征介电常数、特征温度、特征电导率。特别地，特征浓度C_m取为低浓度值，特征温度T_m取为平均温度，特征扩散系数D_i，m、特征粘度μ_m、特征介电常数ε_m、特征电导率σ_f，m取为特征温度下的物性值。in,

is the partial differential operator, ε is the dielectric constant, φ is the electric potential, F is the Faraday constant, _ci is the ion concentration of the i-th ion, _zi is the valence charge number of the i-th ion, _Di is the diffusion coefficient of the i-th ion, _αi is the simplified Soret coefficient of the i-th ion (where i = 1 represents a cation and i = 2 represents an anion), σ is the surface charge density of the nanochannel, L and R represent the length and radius of the nanochannel, u is the velocity, _Rg is the universal gas constant, T is the temperature, p is the pressure, μ is the dynamic viscosity, ρ is the density, _Cp is the specific heat, _kf and _ks are the thermal conductivity of the fluid and solid, _σf is the electrical conductivity, ΔT is the temperature difference, _Ch is the concentration on the high concentration side, and _Cl is the concentration on the low concentration side. Parameters with superscript * are dimensionless parameters, ε ^* ,

φ ^* , c _i ^* , u ^* , D _i ^* , p ^* , μ ^* , σ _f ^* , Θ are dimensionless dielectric constant, dimensionless partial differential operator, dimensionless potential, dimensionless concentration, dimensionless velocity, dimensionless diffusion coefficient, dimensionless pressure, dimensionless viscosity, dimensionless conductivity, dimensionless excess temperature, respectively. Parameters with subscript m are characteristic parameters. D _i,m is characteristic diffusion coefficient (where i=1 represents cation, i=2 represents anion), C _m , μ _m , ε _m , T _m , σ _f,m are characteristic concentration, characteristic viscosity, characteristic dielectric constant, characteristic temperature, characteristic conductivity, respectively. In particular, the characteristic concentration C _m is taken as the low concentration value, the characteristic temperature T _m is taken as the average temperature, the characteristic diffusion coefficient D _i,m , the characteristic viscosity μ _m , the characteristic dielectric constant ε _m , and the characteristic conductivity σ _f,m are taken as the physical property values at the characteristic temperature.

其中，P为功率，η为效率，P^*为无量纲功率。In the prediction method described, the dimensionless array is simplified to:

Where P is power, η is efficiency, and P ^* is dimensionless power.

In the prediction method, the similarity principle is used to ensure that the dimensionless array remains unchanged, and the parameter values therein are changed to obtain the corresponding output performance under different working conditions. Through a limited number of experimental samples, the sample expansion is achieved, which meets the large number of sample numbers of the original data set required by machine learning, as shown in the following formula:

η _j =η ₀

Among them, subscript j represents the sample to be expanded, subscript 0 represents the known sample, the input and output parameters of the known sample are known, and the input parameters of the expanded sample such as diffusion coefficient D _j , concentration C _j , surface charge density σ _j , nanochannel radius R _j , and dielectric constant ε _j are known. According to the above formula, the power and efficiency of the sample to be expanded are obtained to achieve sample expansion.

In the prediction method, the dimensional variables of the input parameters are normalized to the interval (0, 1), and the formula is:

Among them, x represents the original value of the sample data, x _min represents the minimum value of the sample data, and x _max represents the maximum value of the sample data.

In the prediction method, the parallel acceleration of the graphics processor is adopted, the activation function adopts the ReLU function to realize nonlinearity, the weight w and the bias b are updated by the adaptive moment estimation and the back propagation algorithm, the structural parameter optimization is realized, the loss function adopts the mean square error MSE, and the error function adopts the mean relative error MRE, wherein the calculation formulas of ReLU, MSE, and MRE are respectively expressed as:

Among them, x is the input variable value, n is the number of samples in each batch, y _target is the sample label value, and y _output is the predicted value of the network model.

In the prediction method, hyperparameter optimization includes optimizing the number of network layers, the number of neurons in each layer, the optimizer learning rate, and the size of the mini-batch.

Beneficial effects:

The present invention proposes a dimensionless array of the physical process of ion permeation power generation by dimensionlessly transforming the control equation of ion permeation power generation, while ignoring minor issues, reducing the physical parameters in the control equation, and obtaining a simplified dimensionless array. The simplified dimensionless array also takes into account the influence of temperature on ion permeation power generation. By using the similarity principle, under the condition that the dimensionless control parameters remain unchanged, a physical situation is extended to other different situations by multiplying the corresponding factors, and a set of dimensionless samples is converted into a large number of experimental samples, thereby expanding the physical information database and solving the sample requirements of the data set for neural network training. Through ANN, the dimensional variables of the dominant factor physical parameters are used as input, and the output power and energy conversion efficiency are used as output parameters. The nonlinear relationship between the nanochannel geometry, concentration difference, ion properties and energy conversion performance is established, and finally a machine learning-based ion permeation power generation prediction model with high precision and high acceleration ratio is obtained.

The above description is only an overview of the technical solution of the present invention. In order to make the technical means of the present invention clearer and to achieve the extent that those skilled in the art can implement it according to the contents of the specification, and in order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are exemplified below.

BRIEF DESCRIPTION OF THE DRAWINGS

By reading the detailed description of the preferred specific embodiments below, various other advantages and benefits of the present invention will become clear to those of ordinary skill in the art. The drawings in the specification are only for the purpose of illustrating the preferred embodiments and are not considered to be limitations of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without creative work. Moreover, the same reference numerals are used to represent the same components throughout the drawings.

In the attached picture:

FIG1 is a connection diagram of an ion permeation power generation prediction system based on machine learning provided by one embodiment of the present disclosure;

FIG2 is a flow chart of a prediction method of an ion permeation power generation prediction system based on machine learning provided by another embodiment of the present disclosure;

FIG3 is a schematic diagram of the structure of an ANN model provided by another embodiment of the present disclosure;

FIG4 is a performance comparison diagram of predictions and simulation results of a model test module of an ion permeation power generation prediction system based on machine learning provided by another embodiment of the present disclosure.

The present invention is further explained below in conjunction with the accompanying drawings and embodiments.

DETAILED DESCRIPTION

The specific embodiments of the present invention will be described in more detail below with reference to Figures 1 to 4. Although the specific embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

It should be noted that certain words are used in the specification and claims to refer to specific components. Those skilled in the art should understand that technicians may use different nouns to refer to the same component. This specification and claims do not use the difference in nouns as a way to distinguish components, but use the difference in the functions of the components as the criterion for distinction. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, so it should be interpreted as "including but not limited to". The subsequent description of the specification is a preferred embodiment of the present invention, but the description is based on the general principles of the specification and is not intended to limit the scope of the present invention. The scope of protection of the present invention shall be determined by the attached claims.

To facilitate understanding of the embodiments of the present invention, several specific embodiments will be further explained below with reference to the accompanying drawings, and each of the accompanying drawings does not constitute a limitation on the embodiments of the present invention.

In one embodiment, as shown in FIG. 1 to FIG. 3 , an ion permeation power generation prediction system based on machine learning is disclosed, including a data acquisition module, a neural network module and a model testing module;

The data acquisition module is used to establish an input and output database of ion permeation power generation and perform certain preprocessing for use by the neural network module.

The neural network module is used to construct a prediction neural network for ion permeation power generation, and to train the neural network using the database in the data acquisition module.

The model testing module is used to adjust the hyperparameters in the neural network prediction model and obtain a high-precision, high-speed-up ratio ion permeation power generation prediction model.

The prediction system, wherein the data acquisition module includes ion permeation power generation dimensioned equations, dimensionless equations, dimensionless arrays, simplified dimensionless arrays, sample expansion database, and normalization preprocessing.

Further, according to the prediction system, the dimensionally controlled equation.

Poisson's equation:

Nernst-Planck equation:

Continuity equation:

Navier-Stokes equations:

Energy equation:

为偏微分算子，ε为介电常数，φ为电势，F为法拉第常数，c_i为第i种离子的离子浓度，z_i为第i种离子的价电荷数，D_i为第i种离子扩散系数，α_i为第i种离子的简化Soret系数(其中i＝1表示阳离子，i＝2表示阴离子)。u为速度，R_g为通用气体常数，T为温度，p为压力，μ为动力粘度，ρ为密度，c_p为比热，k_f为流体导热系数，σ_f为流体的电导率，E为电场强度。in,

is the partial differential operator, ε is the dielectric constant, φ is the electric potential, F is the Faraday constant, _ci is the ion concentration of the i-th ion, _zi is the valence charge number of the i-th ion, _Di is the i-th ion diffusion coefficient, _αi is the simplified Soret coefficient of the i-th ion (where i = 1 represents a cation and i = 2 represents an anion). u is the velocity, _Rg is the universal gas constant, T is the temperature, p is the pressure, μ is the dynamic viscosity, ρ is the density, _cp is the specific heat, _kf is the thermal conductivity of the fluid, _σf is the electrical conductivity of the fluid, and E is the electric field strength.

Furthermore, according to the prediction system, a dimensionless analysis is performed on ion permeation energy conversion to unify its multi-physical parameters. By introducing dimensionless variables, the dimensionless control equation and its corresponding dimensionless array are obtained:

Dimensionless Poisson's equation:

The dimensionless Nernst–Planck equation:

Dimensionless continuity equation:

The dimensionless Navier-Stokes equations:

Dimensionless energy equation:

φ^*，c_i ^*，u^*，D_i ^*，p^*，μ^*，σ_f ^*分别为无量纲介电常数、无量纲偏微分算子、无量纲电势、第i种离子的无量纲浓度、无量纲速度、无量纲扩散系数、无量纲压力、无量纲粘度、流体的无量纲电导率。无量纲参数均定义为有量纲参数和特征参数的比。Θ为无量纲过余温度，定义为(T-T_m)/ΔT。带m下标的参数为特征参数。C_m，D_i，m，μ_m，ε_m，T_m，σ_fm分别为特征浓度、第i种离子的特征扩散系数、特征粘度、特征介电常数、特征温度、流体的特征电导率。特征浓度C_m可以取为低浓度值，特征温度T_m可以取平均温度。特征扩散系数D_i，m、特征粘度μ_m、特征介电常数ε_m、特征电导率σ_f，m为在特征温度下的物性值。Where σ is the surface charge density of the nanochannel, and ΔT is the temperature difference. L and R represent the length and radius of the nanochannel, respectively. _kf and _ks are the thermal conductivity of the fluid and solid, respectively. Parameters with superscript * are dimensionless parameters. ε ^* ,

φ ^* , c _i ^* , u ^* , D _i ^* , p ^* , μ ^* , σ _f ^* are the dimensionless dielectric constant, dimensionless partial differential operator, dimensionless potential, dimensionless concentration of the i-th ion, dimensionless velocity, dimensionless diffusion coefficient, dimensionless pressure, dimensionless viscosity, and dimensionless conductivity of the fluid, respectively. The dimensionless parameters are defined as the ratio of the dimensional parameter to the characteristic parameter. Θ is the dimensionless excess temperature, defined as (TT _m )/ΔT. Parameters with the subscript m are characteristic parameters. _{C m} , D _{i, m} , μ _m , ε _m , T _m , σ _fm are the characteristic concentration, characteristic diffusion coefficient of the i-th ion, characteristic viscosity, characteristic dielectric constant, characteristic temperature, and characteristic conductivity of the fluid, respectively. The characteristic concentration C _m can be taken as a low concentration value, and the characteristic temperature T _m can be taken as the average temperature. The characteristic diffusion coefficient D _i,m , the characteristic viscosity μ _m , the characteristic dielectric constant ε _m , and the characteristic conductivity σ _f,m are physical property values at the characteristic temperature.

Based on the similarity principle, the dimensionless number of the ion permeation power generation physical process is proposed, and the initial dimensionless array consisting of 12 dimensionless parameters is obtained:

Here, _Ch and _Cl represent high ion concentration and low ion concentration, respectively.

Furthermore, based on the basic assumptions of the physical problem, we have:

Thus, the influence of convection and Joule heat can be ignored, the dimensionless physical parameters in the control equation can be reduced, and the simplified dimensionless array can be obtained by taking the monovalent electrolyte solution as a reference, thereby achieving the unification of physical parameters in the ion permeation power generation process, which can be used to guide modeling experiments. After simplification, a dimensionless array consisting of 6 dimensionless parameters is obtained:

In a specific embodiment, according to the similarity principle, when the same dimensionless array is used, it is ensured to be unchanged. Under different physical conditions, the dimensional output power P and efficiency η can be obtained by reverse deduction through the following formula:

η _j =η ₀

Among them, subscript j represents the sample to be expanded, subscript 0 represents the known sample, and the input and output parameters of the known sample are all known. The input parameters of the expanded sample, such as diffusion coefficient D _j , concentration C _j , surface charge density σ _j , nanochannel radius R _j , and dielectric constant ε _j, are known. Under the premise that the dimensionless number remains unchanged, the power and efficiency of the sample to be expanded can be obtained according to the above formula, and sample expansion can be achieved to meet the large demand for the number of samples in the original data set for machine learning.

In this embodiment, according to the prediction system, the dimensional variables of the input parameters are normalized to the interval (0, 1), thereby eliminating the adverse effects caused by singular sample data. The specific formula is:

Among them, x represents the original value of the sample data, x _min represents the minimum value of the sample data, and x _max represents the maximum value of the sample data.

After normalization, the speed of gradient descent optimal solution can be accelerated and the convergence of the training network can be improved.

According to the prediction system, the neural network module includes a training set input, a fully connected layer (hidden layer), and structural parameter optimization.

Further, according to the prediction system, an artificial neural network ANN model is used to construct a training model, which consists of an input layer, an output layer and a hidden layer, wherein the output power P and the energy conversion efficiency η are output parameters, and the remaining physical parameters, such as concentration C, high temperature end temperature _Th , low temperature end temperature _T1 , ion diffusion coefficient D, dielectric constant ε, nanochannel length L, nanochannel radius R, and surface charge density σ are dimensional variables as input to achieve nonlinear regression from input to output.

Further, according to the prediction system, by adopting GPU parallel acceleration, the activation function uses ReLU to achieve nonlinearity and avoid gradient disappearance. The weight w and bias b are updated by Adam optimizer and back propagation BP algorithm to achieve structural parameter optimization. The loss function uses mean square error MSE, and the error function uses mean relative error MRE. The calculation formulas of MSE and MRE are respectively expressed as:

Among them, x is the input variable value, n is the number of samples in each batch, y _target is the sample label value, and y _output is the predicted value of the network model. The error is back-propagated, and the parameter weights of each layer are gradually updated iteratively to finally obtain the minimum Loss. The training is completed until convergence, indicating that the initial training of the model is completed.

Further, according to the prediction system, the model testing module includes hyperparameter optimization and ion permeation power generation prediction models.

Furthermore, according to the prediction system, the hyperparameter optimization mainly includes optimization of the number of network layers, the number of neurons in each layer, the optimizer learning rate and the size of mini-batch.

Use the test set to test the trained network and modify the model according to the test results. If overfitting occurs, adjust the network structure parameters and model hyperparameters until the accuracy meets the requirements. If the test results are good, save the neural network parameter data and its output results.

Furthermore, according to the prediction system, the ion permeation power generation prediction model can achieve high-precision and high-speed-up ratio results to guide the actual process of ion permeation power generation.

In another example, as shown in FIG2 , the present disclosure provides an ion permeation power generation prediction system and construction method based on machine learning, comprising the following steps:

S100: the measuring unit measures a plurality of physical parameters of ion permeation power generation, constructs a dimensioned equation for ion permeation power generation based on the physical parameters, and the dimensionless conversion unit generates a dimensionless equation for ion permeation power generation and its corresponding dimensionless array based on the dimensioned equation for ion permeation power generation;

S200: The sample expansion unit expands the dimensionless array by adopting a similarity principle to obtain an expanded sample, wherein the expanded sample and the dimensionless array constitute sample data, and the normalization processing unit normalizes the sample data;

S300: the sample data is divided into training data and test data according to a predetermined ratio, the neural network module is configured to construct a prediction neural network for ion permeation power generation based on an artificial neural network, and a neural network training is performed based on the sample data to construct a prediction neural network for ion permeation power generation, wherein the output power P and the energy conversion efficiency η are output parameters, and the remaining physical parameters, such as concentration C, high temperature end temperature _Th , low temperature end temperature _T1 , ion diffusion coefficient D, dielectric constant ε, nanochannel length L, nanochannel radius R, and dimensional variables of surface charge density σ are used as inputs;

S400: The hyperparameter optimization unit uses GPU acceleration, activation function nonlinearization, and a combination of back-propagation algorithms to update weights w and bias b to optimize the prediction neural network for ion permeation power generation. The data processing results are evaluated using the loss function MSE.

S500: The testing unit tests the optimized ion permeation power generation prediction neural network based on the test data, adjusts the hyperparameters in the network according to the training results, repeats the iteration until the error reaches the expected level, and saves the model as the ion permeation power generation prediction model.

In another example, as shown in FIG3 , a schematic diagram of the ANN model structure provided by an embodiment of the present disclosure is shown.

The calculation example of the present invention can directly read eight dominant characteristic parameters, namely, concentration C, high temperature end temperature _Th , low temperature end temperature _Tl , ion diffusion coefficient D, dielectric constant ε, nanochannel length L, nanochannel radius R, and surface charge density σ, as inputs, and simultaneously use the corresponding output power P and energy conversion efficiency η as label values of training samples.

Before model training, the input and output parameters of the data set are normalized to the interval (0, 1) to avoid numerical problems and enable the network to converge quickly.

After the network successfully reads the data set, the neural network is trained under the ANN deep learning framework to achieve nonlinear regression tasks. The network has 6 layers, and its substructure includes an input layer, an output layer, and four hidden layers. The dimension of the input parameter is 8, the dimension of the output parameter is 2, and the parameters of the intermediate layer are 1000, 500, 200, and 20 respectively. The input layer neurons receive input signals, and each neuron in the hidden layer and the output layer is fully connected to all neurons in the adjacent layers.

The input of each neuron in the hidden layer and the output layer is the weighted sum of the output values of all neurons in the previous layer. Assuming x _m ⁿ is the input value of the mth neuron in the nth layer of the neural network, y _m ⁿ and b _m ⁿ are the output value and bias of the neuron respectively, and w _im ^n-1 is the connection weight between the neuron and the i-th neuron in the n-1th layer, the following formula is obtained:

Among them, σ() represents the activation function.

The activation function uses ReLU to avoid gradient disappearance and realize nonlinear network. The formula of ReLU is expressed as:

Where x represents the input variable value. ReLU will make the output of some neurons 0, making the network sparse and reducing the interdependence of parameters, alleviating the occurrence of overfitting problems. After layer-by-layer feature extraction, the final output is 2 parameters, which is the regression result of the neural network.

The loss function uses the mean square error (MSE), and the error function uses the mean relative error (MRE). The convergence error of the sample on the training data set and the test data set is calculated respectively. The network is trained by the Adam optimizer, and the learning rate is set to 0.001/(1+epoch×β), where epoch is the number of training times and β=0.001. The error is back-propagated, and the parameter weights of each layer are updated step by step. Finally, the minimum loss is obtained. The training is completed until convergence, indicating that the initial training of the model is completed.

In another embodiment, as shown in FIG4 , a performance comparison diagram of the model test module prediction and simulation results of an ion permeation power generation prediction system based on machine learning provided by an embodiment of the present disclosure is provided. In order to quantitatively measure the prediction effect of the ion permeation power generation prediction model, so as to achieve high precision and high acceleration ratio and effectively guide the actual process of ion permeation power generation, five experimental samples (Case 1-5) are randomly selected from the data set to compare the prediction results of the ion permeation power generation prediction system based on machine learning with the performance of the experimental simulation results, and the results are shown in FIG4 and Table 1:

Table 1 Performance comparison between model prediction results and experimental simulation results

As can be seen from Figure 4 and Table 1, for the five randomly selected experimental samples, for the output power results, the total time consumption of the prediction model is much less than the total time consumption of the experimental simulation, saving 2047 seconds, the speedup ratio is as high as 683, and the maximum error of the prediction model is only 2.39%. The results fully prove that the prediction model has a good prediction effect, has certain practicality, and effectively guides the actual process of ion permeation power generation.

The prediction method describes the physical and mathematical characteristics of ion permeation power generation energy conversion. By non-dimensionalizing the control equation, a simplified dimensionless array of the physical process of ion permeation power generation is proposed. Under the guidance of dimensionless numbers, a set of dimensionless samples is expanded into a large number of groups of samples, thereby expanding the physical information database and solving the sample requirements of the neural network training data set. On this basis, based on machine learning, the dimensional variables of the dominant factor physical parameters that have been normalized and pre-processed are used as input, and the output power and energy conversion efficiency are used as output parameters to construct an ANN prediction model. By using GPU acceleration, activation function, optimizer and back-propagation algorithm combined optimization, the accuracy of the prediction model reaches 97.6%. The hyperparameters in the network are adjusted according to the training results, and finally a machine learning-based ion permeation power generation prediction model with high precision and high acceleration ratio is obtained, which can guide the judgment of the dominant parameters of the permeation power generation system and the estimation of efficiency.

Although the embodiments of the present invention are described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described, and the specific embodiments described are only illustrative and instructive, rather than restrictive. A person skilled in the art can make many forms under the guidance of this specification and without departing from the scope of protection of the claims of the present invention, all of which belong to the protection of the present invention.

Claims (9)

1. A prediction system of ion permeation power generation based on machine learning is characterized in that: which comprises the steps of preparing a mixture of a plurality of raw materials,

a data acquisition module configured to acquire sample data of ion permeation power generation, the data acquisition module comprising;

a measurement unit that measures a plurality of physical parameters of the ion permeation power generation;

a dimensionless transformation unit connected to the measurement unit and generating a plurality of dimensionless parameters based on the plurality of physical parameters and forming a dimensionless array based on the dimensionless parameters;

the sample expansion unit is connected with the dimensionless conversion unit and is expanded by adopting a similarity principle based on the dimensionless array to obtain an expansion sample, the expansion sample and the dimensionless array form sample data, and the sample data is divided into training data and test data according to a preset proportion;

a normalization processing unit connected to the sample expansion unit and normalizing the sample data;

the neural network module is configured to construct a prediction neural network and neural network training of ion permeation power generation based on an artificial neural network ANN, the neural network module connected with the data acquisition module comprises a plurality of network layers, the substructure of the neural network module comprises an input layer, an output layer and four full-connection layers, neurons of the input layer receive training data, each neuron of the hidden layer and the output layer is in full connection with all neurons of the adjacent layer, and the neural network training is carried out based on sample data to form the prediction neural network of ion permeation power generation;

a model testing module configured to adjust hyper-parameters in a predictive neural network of ion osmotic power generation to obtain an ion osmotic power generation predictive model, the model testing module comprising;

the hyper-parameter optimization unit is connected with the neural network module, adopts a Graphic Processing Unit (GPU) to perform parallel acceleration, activates the combination optimization of function nonlinearity and back propagation algorithm, and updates the weight w and the bias b to optimize the prediction neural network of the ion permeation power generation;

and the testing unit is connected with the data acquisition module and the neural network module so as to test the prediction neural network of the ion permeation power generation based on the test data until the error reaches the expectation, and the model is saved as the ion permeation power generation prediction model.

2. The machine learning-based prediction system for ion-osmotic power generation of claim 1, wherein preferably: the measuring unit includes a first measuring portion that obtains a dielectric constant, a concentration, an ion diffusion coefficient, a nanochannel length, a radius, and a surface charge density, a temperature sensor that measures temperature data, a power level that measures output power P, and a second measuring portion that measures energy conversion efficiency η.

3. A prediction method using the prediction system for machine learning-based ion osmotic power generation according to claim 1 or 2, characterized in that it comprises,

s100: the measuring unit measures a plurality of physical parameters of the ion permeation power generation, an ion permeation power generation dimensional equation is constructed based on the physical parameters, and the dimensionless conversion unit generates an ion permeation power generation dimensionless equation and a corresponding dimensionless array thereof based on the ion permeation power generation dimensional equation;

s200: the sample expansion unit is used for expanding based on the dimensionless array by adopting a similarity principle to obtain an expansion sample, the expansion sample and the dimensionless array form sample data, and the normalization processing unit normalizes the sample data;

S300: the sample data is divided into training data and test data according to a preset proportion, the neural network module is configured into a prediction neural network for constructing the ion permeation power generation based on the artificial neural network, the neural network training is carried out based on the sample data to form the prediction neural network for the ion permeation power generation, wherein the output power P and the energy conversion efficiency eta are output parameters, the concentration C and the temperature T at the high-temperature end are used as output parameters, and the temperature T at the high-temperature end is used as a temperature parameter _h Low temperature end temperature T _l Dimensional variables of an ion diffusion coefficient D, a dielectric constant epsilon, a nano-channel length L, a nano-channel radius R and a surface charge density sigma are taken as input;

s400: the hyper-parameter optimization unit adopts a Graphic Processing Unit (GPU) for acceleration, activates the combination optimization of function nonlinearity and back propagation algorithm, and updates the weight w and the bias b to optimize the prediction neural network of the ion permeation power generation;

s500: and the test unit tests the optimized prediction neural network of the ion permeation power generation based on the test data, adjusts the hyper-parameters in the network according to the training result, repeats iteration until the error reaches the expectation, and saves the model as the prediction model of the ion permeation power generation.

4. The prediction method according to claim 3, characterized in that: the dimensionless control equation and the dimensionless array corresponding to it are:

dimensionless poisson equation:

dimensionless nernst-planck equation:

dimensionless continuity equation:

dimensionless navier-stokes equation:

dimensionless energy equation:

dimensionless array:

wherein,

is a partial differential operator,. Epsilon.is the dielectric constant,. Phi.is the potential,. F is the Faraday constant, c _i Is the ion concentration of the i-th ion, z _i Is the number of valence charges of the i-th ion, D _i Is diffusion coefficient of i-th ion, alpha _i Simplified Soret coefficient for the ith ion, where i =1 represents a cation, i =2 represents an anion, σ is the nanochannel surface charge density, L and R represent the nanochannel length and radius, respectively, u is the velocity, R _g Is the general gas constant, T is the temperature, p is the pressure, μ is the dynamic viscosity, ρ is the density, C _p Is specific heat, k _f And k _s Thermal conductivity coefficient of fluid and solid respectively, sigma _f Is the electrical conductivity,. DELTA.T is the temperature difference, C _h Is high side concentration, C _l For low concentration side concentrations, the parameters marked with a letter are dimensionless parameters, epsilon ^* ，

φ ^* ，c _i ^* ，u ^* ，D _i ^* ，p ^* ，μ ^* ，σ _f ^* Theta is the dimensionless dielectric constant, the dimensionless partial differential operator, the dimensionless potential, the dimensionless concentration, the dimensionless speed, and the dimensionless spread respectivelyScattering coefficient, dimensionless pressure, dimensionless viscosity, dimensionless conductivity, dimensionless excess temperature, parameter with m index as characteristic parameter, D _i，m Is a characteristic diffusion coefficient, C _m ，μ _m ，ε _m ，T _m ，σ _f，m Respectively, characteristic concentration, characteristic viscosity, characteristic dielectric constant, characteristic temperature and characteristic conductivity.

5. The prediction method according to claim 4, wherein: the dimensionless array is simplified as:

where P is power, η is efficiency, P ^* Is a dimensionless power.

6. The prediction method of claim 5, wherein a similarity principle is used to ensure that dimensionless arrays are unchanged, parameter values therein are changed to obtain corresponding output performance under different working conditions, and the samples are expanded by a limited number of experimental samples, so as to satisfy the requirement of machine learning on the number of samples in the original data set, as follows:

η _j ＝η ₀

wherein, the subscript j represents the sample to be extended, the subscript 0 represents the known sample, the input and output parameters of the known sample are known, and the input parameter of the extended sample is, for example, the diffusion coefficient D _j Concentration C of _j Surface charge density σ _j Nano channel radius R _j Dielectric constant ε _j It is known to obtain the power and efficiency of the samples to be extended according to the above formula, and to implement the sample extension.

7. The prediction method according to claim 3, characterized in that: normalizing the dimension variables of the input parameters, wherein the normalization is an interval (0,1), and the formula is as follows:

wherein x represents the original value of sample data, x _min Represents the minimum value of the sample data, x _max Representing the maximum value of the sample data.

8. The prediction method according to claim 3, characterized in that: the image processor is adopted for parallel acceleration, the ReLU function is adopted as an activation function to realize nonlinearity, updating of weight w and bias b is realized through adaptive moment estimation and a back propagation algorithm, structural parameter optimization is realized, mean Square Error (MSE) is adopted as a loss function, and average relative error (MRE) is adopted as an error function, wherein the calculation formulas of ReLU, MSE and MRE are respectively expressed as follows:

wherein x is the value of the input variable, n is the number of samples per batch, y _target Is a sample tag value, y _output And predicting the value of the network model.

9. The prediction method according to claim 3, characterized in that: the hyper-parameter optimization comprises optimization of the number of network layers, the number of neurons in each layer, the learning rate of an optimizer and the size of the mini-batch.

Images