Optimization condition prediction method for flocculation-assisted solidification based on SCA-LSTM neural network

Abstract

The invention provides a flocculation-assisted solidification optimization condition prediction method based on an SCA-LSTM neural network, and the method comprises the steps: adding water into an ardealite sample, stirring, and filtering to obtain an ardealite filtrate for subsequent solidification treatment; the method comprises the following steps: respectively recording the dosage of a curing agent, the stirring time, the stirring speed, the dosage of a flocculating agent and the soluble phosphorus and fluorine removal rates under corresponding stirring conditions through a single factor experiment, and generating a condition data sample set; performing parameter optimization on the LSTM neural network by adopting a sine and cosine optimization algorithm SCA to improve the prediction precision of the network, and evaluating the model through average absolute relative deviation AARD, root mean square error RMSE and decision coefficient R; according to the method, the time sequence modeling capability of the LSTM network and the global search capability of the SCA algorithm are combined, and the optimal treatment conditions of soluble phosphorus and fluorine in the flocculant-assisted curing phosphogypsum are accurately predicted and optimized.

Description

Technical Field

[0001] The present invention relates to the technical field of chemical engineering and environmental engineering, and in particular to a method for predicting optimization conditions of flocculation-assisted solidification based on a SCA-LSTM neural network. Background Art

[0002] Phosphogypsum is a byproduct of the decomposition of phosphate rock by sulfuric acid, and its main component is calcium sulfate dihydrate. For every ton of phosphoric acid produced, about 4.5-5 tons of phosphogypsum are produced. Phosphogypsum contains a large amount of soluble phosphorus and fluorine impurities, which not only reduce the mechanical strength of phosphogypsum, but also affect its recycling. In addition, soluble phosphorus and fluorine make phosphogypsum acidic, which not only increases the difficulty of storage and transportation, but also may cause pollution of groundwater and soil when phosphogypsum is piled up. Therefore, it is of great environmental significance to study effective methods for solidifying soluble phosphorus and fluorine in phosphogypsum.

[0003] The traditional treatment method is usually to add a large amount of calcium oxide or calcium chloride to the phosphogypsum, and solidify the soluble phosphorus and fluorine into calcium phosphate and calcium fluoride through chemical reactions. However, this method not only consumes a large amount of curing agent, but the solidified phosphorus and fluorine still remain in the phosphogypsum, limiting its resource utilization. Although the water washing method can effectively remove soluble phosphorus and fluorine, the resulting eluent requires secondary treatment, and the treatment cost is high. By adding lime or calcium chloride to the filtrate, although calcium fluoride precipitate can be generated and recycled, the precipitate particles are small, the sedimentation rate is slow, and separation is difficult, resulting in a high fluoride ion content in the filtrate. Therefore, the use of flocculants to assist flocculation is a common solution.

[0004] However, the flocculation effect is affected by many factors, including the type, dosage, reaction time, stirring rate and order of addition of flocculants, and the complex nonlinear relationship between the factors makes it difficult to find the optimal treatment conditions with a single experimental method. Traditional orthogonal experimental design mainly explores the main effects and ignores the interaction effects in complex systems. The response surface method has a high computational complexity when dealing with high-dimensional problems. As the number of variables increases, the number of experiments and the amount of calculations increase sharply, making the model construction and optimization process time-consuming and difficult to implement. Summary of the invention

[0005] The present invention provides a flocculation-assisted solidification optimization condition prediction method based on an SCA-LSTM neural network, and finds an optimal condition combination of soluble phosphorus and fluorine in a solidified phosphogypsum washing liquid among numerous experimental factors, so that the soluble phosphorus and soluble fluorine in the phosphogypsum can achieve a large removal rate while recovering the solidified phosphorus and fluorine. The method can effectively predict and optimize the optimal combination of various factors in the flocculation method, thereby improving the removal effect of soluble phosphorus and fluorine in the phosphogypsum.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is: The optimization condition prediction method of flocculation-assisted solidification based on SCA-LSTM neural network includes the following steps: S1. Add water to the phosphogypsum sample and stir it, and obtain the phosphogypsum filtrate by filtration for subsequent detection and solidification treatment; S2. Conduct single factor experiments on multiple experimental conditions such as the dosage of the curing agent, the stirring time and stirring rate after the curing agent is added, and the dosage of the flocculant, the stirring time and stirring rate after the flocculant is added, record the removal rates of soluble phosphorus and fluorine, and generate a conditional data sample set; S3. Build a neural network based on LSTM and define the network structure and initial parameters; S4. Use the sine-cosine optimization algorithm (SCA) to optimize the parameters of LSTM and improve the prediction accuracy of the LSTM neural network. The SCA algorithm simulates the behavior of the sine-cosine function and combines the global search capability to find the optimal parameter combination of the LSTM network. S5. Use the average absolute relative deviation AARD, root mean square error RMSE and determination coefficient to The removal rate was evaluated; S6. Use the trained SCA-LSTM model to predict the optimal treatment conditions for soluble phosphorus and fluorine in flocculation-assisted solidification of phosphogypsum.

[0007] The specific process of building the prediction model in the above S3 is: The input gate mechanism of the LSTM model is used to determine what new information should be added to the memory cell. It first generates weight updates through the Sigmoid layer, and then generates candidate input information through the Tanh function: ; ; in is the output of the input gate, which controls the amount of new information input. is a candidate memory state, generated by the tanh activation function; Forget gate: The forget gate determines which information should be forgotten and which information should be retained; the forget gate receives the hidden state of the previous moment and the current input , and outputs a scalar between [0,1] indicating how much information should be retained: ; in is the weight matrix of the forget gate, is the bias term of the forget gate, is the Sigmoid function; Cell state update: Update the state of the memory unit, i.e., the cell, based on the output of the forget gate and the output of the input gate; the forget gate determines how much old information is discarded, and the input gate determines how much new information is added: ; in Indicates the current cell status. The cell state at the previous moment, Element-wise multiplication; Output gate: The output gate determines the hidden state at the current moment , and adjust the output through Sigmoid and Tanh functions: ; ; in is the value of the output gate, which is used to control the output of the hidden state. It is the hidden state at the current moment and the output of LSTM.

[0008] In the above S4, the sine-cosine optimization algorithm SCA includes the following steps: S4.1. Initialize the individual group and randomly generate a group of individuals within the set range. Each individual represents a candidate solution. S4.2. For each individual i, calculate the mean square error to evaluate the performance of the particle: ; in is the predicted output of the model, is the label of the training sample, that is, the actual removal rate, is the number of training samples; S4.3. Update the position of the individual and update the solution of each individual according to the sine and cosine functions. The specific formula is as follows: ; in is the position of the i-th individual at the t-th iteration, is a parameter that controls the exploration step size, which decays with the number of iterations. , They are all random numbers between 0 and 1. is the learning factor; S4.4, find the optimal solution, each individual adjusts its position according to the sine or cosine formula, if the performance of an individual is better than the current global optimal solution, then update the global optimal solution; S4.5. Repeat steps S4.2-S4.4 until the minimum error condition is met, and apply the optimized conditional parameters to LSTM.

[0009] In the above S1, the mass ratio of phosphogypsum to water is 1:10-20.

[0010] In the above S2, the curing agent is selected from calcium oxide, wherein the mass ratio of the curing agent to the mass ratio of fluorine in the filtrate is 1.5 to 7.5.

[0011] The flocculant in S2 is selected from inorganic flocculants containing aluminum and iron, including polyaluminum sulfate, or polyferric sulfate, or polyaluminum chloride, or polyferric chloride, or polyaluminum sulfate ferric, or polyaluminum ferric chloride, wherein the mass ratio of the flocculant to fluorine is 0-2.7:1.

[0012] The stirring time after adding the curing agent in the above S2 is 0 to 60 minutes, and the stirring time after adding the flocculant is 5 to 60 minutes; After adding the curing agent, the stirring rate is 50 r / min~500 r / min, and after adding the flocculant, the stirring rate is 50 r / min~500 r / min.

[0013] In the above S2, the order of adding the flocculants is controlled by changing the stirring time after adding the curing agent. When the stirring time after adding the curing agent is 0, the curing agent and the flocculant are added at the same time. The order of adding the curing agent and the flocculant is: first add the curing agent, then add the flocculant; or add the curing agent and the flocculant at the same time.

[0014] In the above S4, the maximum number of iterations of the SCA algorithm is set to 50, the step size is set to 2, and the number of individuals is set to 30.

[0015] The specific formula for evaluation in S5 above is: ; ; ; in represents the predicted value of the sample, represents the true value of the sample, n is the number of samples, is the mean of the true values ​​of the samples.

[0016] The present invention provides a method for predicting the optimal conditions of flocculation-assisted solidification based on an SCA-LSTM neural network, which uses the memory characteristics of the LSTM neural network and the global search capability of the sine-cosine optimization algorithm to predict and optimize the optimal treatment conditions for soluble phosphorus and fluorine in flocculation-assisted solidification of phosphogypsum. In model construction, the LSTM network not only establishes a dynamic feature extraction mechanism for the collected data, but also verifies its fitting ability for nonlinear and multivariable coupling systems through experiments, and optimizes the network structure to adapt to the characteristics of experimental data. On the other hand, the SCA optimization algorithm efficiently searches for and optimizes network hyperparameters on a global scale based on the complex parameter space of the phosphogypsum treatment process by simulating the characteristics of the sine-cosine function, and selects the optimal process parameter combination in combination with the experimental design. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The present invention will be further described below in conjunction with the accompanying drawings and embodiments: Figure 1 It is an overall step for optimizing the conditions of the method of the present invention.

[0018] Figure 2 It is a calculation flow chart of the sine-cosine optimization algorithm of the present invention. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical solutions and advantages of the present invention clearer, the following content will systematically and completely describe the specific technical solutions of the present invention in combination with the drawings provided according to the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0020] Embodiment 1: like Figure 1 and 2 As shown in, the optimization condition prediction method of flocculation-assisted solidification based on SCA-LSTM neural network includes the following steps: S1. Add water to the phosphogypsum sample and stir it, and obtain the phosphogypsum filtrate by filtration for subsequent detection and solidification treatment; the mass ratio of phosphogypsum to water is 1:10-20; S2. Conduct single factor experiments on multiple experimental conditions such as the dosage of the curing agent, the stirring time and stirring rate after the curing agent is added, and the dosage of the flocculant, the stirring time and stirring rate after the flocculant is added, record the removal rates of soluble phosphorus and fluorine, and generate a conditional data sample set; In a preferred case, the curing agent is selected from calcium oxide, wherein the mass ratio of the curing agent to the mass ratio of fluorine in the filtrate is 1.5 to 7.5; In a preferred case, the flocculant is selected from inorganic flocculants containing aluminum and iron, including polyaluminum sulfate, or polyferric sulfate, or polyaluminum chloride, or polyferric chloride, or polyaluminum sulfate ferric, or polyaluminum ferric chloride, wherein the mass ratio of the flocculant to fluorine is 0-2.7:1.

[0021] In a preferred case, the stirring time after adding the curing agent is 0 to 60 min, and the stirring time after adding the flocculant is 5 to 60 min; After adding the curing agent, the stirring rate is 50 r / min to 500 r / min, and after adding the flocculant, the stirring rate is 50 r / min to 500 r / min; In a preferred case, in the above S2, the order of adding the flocculant is controlled by changing the stirring time after adding the curing agent. When the stirring time after adding the curing agent is 0, the curing agent and the flocculant are added at the same time. In a preferred case, the order of adding the curing agent and the flocculant is: first add the curing agent, then add the flocculant; or add the curing agent and the flocculant at the same time; The above single factor experiment is conducted and the removal rates of soluble phosphorus and fluorine are recorded as the conditional data sample set; S3, build a neural network based on LSTM, define the network structure and initial parameters; LSTM neural network consists of an input layer, multiple hidden layers and an output layer, and can effectively predict data through time series modeling; in the preferred case, a two-layer LSTM network is used, and the hidden layer unit decibels of LSTM are set to 20 and 10, and each layer is followed by a ReLU function as an activation function to improve the ability of nonlinear fitting. Finally, the output is obtained through a fully connected layer and a Sigmoid function; The input gate mechanism of the LSTM model is used to determine what new information should be added to the memory cell. It first generates weight updates through the Sigmoid layer, and then generates candidate input information through the Tanh function: ; ; in is the output of the input gate, which controls the amount of new information input. is a candidate memory state, generated by the tanh activation function; Forget gate: The forget gate determines which information should be forgotten and which information should be retained; the forget gate receives the hidden state of the previous moment and the current input , and outputs a scalar between [0,1] indicating how much information should be retained: ; in is the weight matrix of the forget gate, is the bias term of the forget gate, is the Sigmoid function; Cell state update: Update the state of the memory unit, i.e., the cell, based on the output of the forget gate and the output of the input gate; the forget gate determines how much old information is discarded, and the input gate determines how much new information is added: ; in Indicates the current cell status. The cell state at the last moment, Element-wise multiplication; Output gate: The output gate determines the hidden state at the current moment , and adjust the output through Sigmoid and Tanh functions: ; ; in is the value of the output gate, which is used to control the output of the hidden state. is the hidden state at the current moment and the output of LSTM; S4. Use the sine-cosine optimization algorithm (SCA) to optimize the parameters of LSTM to improve the prediction accuracy of the LSTM neural network. In the preferred case, the maximum number of iterations of the SCA algorithm is set to 50, the step size is set to 2, and the number of individuals is set to 30. The SCA algorithm simulates the behavior of sine and cosine functions and combines global search capabilities to find the optimal parameter combination of the LSTM network; The Sin-Cosine Optimization Algorithm (SCA) includes the following steps: S4.1. Initialize the individual group and randomly generate a group of individuals within the set range. Each individual represents a candidate solution. S4.2. For each individual i, calculate the mean square error to evaluate the performance of the particle: ; in is the predicted output of the model, is the label of the training sample, that is, the actual removal rate, is the number of training samples; S4.3. Update the position of the individual and update the solution of each individual according to the sine and cosine functions. The specific formula is as follows: ; in is the position of the i-th individual at the t-th iteration, is a parameter that controls the exploration step size, which decays with the number of iterations. , They are all random numbers between 0 and 1. is the learning factor; S4.4, find the optimal solution, each individual adjusts its position according to the sine or cosine formula, if the performance of an individual is better than the current global optimal solution, then update the global optimal solution; S4.5, repeat steps S4.2-S4.4 until the minimum error condition is met, and apply the optimized conditional parameters to LSTM; S5. Use the average absolute relative deviation AARD, root mean square error RMSE and determination coefficient to Removal efficiency was evaluated: ; ; ; in represents the predicted value of the sample, represents the true value of the sample, n is the number of samples, is the average of the true values ​​of the samples; S6. Use the trained SCA-LSTM model to predict the optimal treatment conditions for soluble phosphorus and fluorine in flocculation-assisted solidification of phosphogypsum. The specific steps are as follows: input the test sample, use the trained LSTM model to predict the removal rate, and select the experimental conditions with the best removal effect of phosphorus and fluorine in phosphogypsum based on the results.

[0022] The sample data of the single-factor experiment is divided into a training set and a test set. The prediction model is trained using the training set, and the prediction is performed using the test set. The statistical parameters of the test set are: , , . The training model is used to identify the influence of the removal conditions of soluble phosphorus and fluorine in phosphogypsum. The dosage range of the curing agent is 45-75 mg, the stirring time after adding the curing agent is 30-60 min, the stirring rate after adding the curing agent is 400-800, the flocculant dosage is 2-20 mg, the stirring time after adding the flocculant is 10-60 min, and the stirring rate after adding the flocculant is 100-500. The best conditions are found to be a curing agent dosage of 75 mg, a stirring time after adding the curing agent of 30 min, a stirring rate after adding the curing agent of 700, a flocculant dosage of 4 mg, a stirring time after adding the flocculant of 40 min, and a stirring rate after adding the flocculant of 100. The best removal rate is 99.28%. The present invention has significant advantages over traditional orthogonal experimental design and response surface method. Traditional methods are limited by the limitations of fixed model assumptions and low-order interaction effects, and cannot fully model complex multi-order interactions between high-dimensional variables. Orthogonal experimental design requires preset levels for each factor. As the number of variables and levels increases, the experimental combination grows exponentially, resulting in an overly large experimental scale that is difficult to operate in practice. The prediction model used in the present invention can handle complex nonlinear relationships of multiple factors and high-dimensional variables, thereby more accurately simulating experimental data and optimizing removal conditions. In addition, the combination of model training and experiment avoids the problem of overly large experimental scale caused by too many variables and levels in orthogonal experiments, and has higher practicality and efficiency.

Claims

1. The optimization condition prediction method of flocculation-assisted solidification based on SCA-LSTM neural network is characterized in that: The following steps are involved: S1. Add water to the phosphogypsum sample and stir it, and obtain the phosphogypsum filtrate by filtration for subsequent detection and solidification treatment; S2. Conduct single factor experiments on multiple experimental conditions such as the dosage of the curing agent, the stirring time and stirring rate after the curing agent is added, and the dosage of the flocculant, the stirring time and stirring rate after the flocculant is added, record the removal rates of soluble phosphorus and fluorine, and generate a conditional data sample set; S3. Build a neural network based on LSTM and define the network structure and initial parameters; S4. Use the sine-cosine optimization algorithm (SCA) to optimize the parameters of LSTM and improve the prediction accuracy of the LSTM neural network. The SCA algorithm simulates the behavior of the sine-cosine function and combines the global search capability to find the optimal parameter combination of the LSTM network. S5. Use the average absolute relative deviation AARD, root mean square error RMSE and determination coefficient to The removal rate was evaluated; S6. Use the trained SCA-LSTM model to predict the optimal treatment conditions for soluble phosphorus and fluorine in flocculation-assisted solidification of phosphogypsum.

2. The method for predicting the optimal conditions for flocculation-assisted solidification based on the SCA-LSTM neural network according to claim 1, characterized in that: The specific process of constructing the prediction model in S3 is as follows: The input gate mechanism of the LSTM model is used to determine what new information should be added to the memory cell. It first generates weight updates through the Sigmoid layer, and then generates candidate input information through the Tanh function: ； ； in is the output of the input gate, which controls the amount of new information input. is a candidate memory state, generated by the tanh activation function; Forget gate: The forget gate determines which information should be forgotten and which information should be retained; the forget gate receives the hidden state of the previous moment and the current input , and outputs a scalar between [0,1] indicating how much information should be retained: ； in is the weight matrix of the forget gate, is the bias term of the forget gate, is the Sigmoid function; Cell state update: Update the state of the memory unit, i.e., the cell, based on the output of the forget gate and the output of the input gate; the forget gate determines how much old information is discarded, and the input gate determines how much new information is added: ； in Indicates the current cell status. The cell state at the last moment, Element-wise multiplication; Output gate: The output gate determines the hidden state at the current moment , and adjust the output through Sigmoid and Tanh functions: ； ； in is the value of the output gate, which is used to control the output of the hidden state. It is the hidden state at the current moment and the output of LSTM.

3. The method for predicting the optimal conditions for flocculation-assisted solidification based on SCA-LSTM neural network according to claim 2, characterized in that: In S4, the sine-cosine optimization algorithm SCA comprises the following steps: S4.

1. Initialize the individual group and randomly generate a group of individuals within the set range. Each individual represents a candidate solution. S4.

2. For each individual i, calculate the mean square error to evaluate the performance of the particle: ； in is the predicted output of the model, is the label of the training sample, that is, the actual removal rate, is the number of training samples; S4.

3. Update the position of the individual and update the solution of each individual according to the sine and cosine functions. The specific formula is as follows: ； in is the position of the i-th individual at the t-th iteration, is a parameter that controls the exploration step size, which decays with the number of iterations. , They are all random numbers between 0 and 1. is the learning factor; S4.4, find the optimal solution, each individual adjusts its position according to the sine or cosine formula, if the performance of an individual is better than the current global optimal solution, then update the global optimal solution; S4.

5. Repeat steps S4.2-S4.4 until the minimum error condition is met, and apply the optimized conditional parameters to LSTM.

4. The method for predicting the optimal conditions for flocculation-assisted solidification based on SCA-LSTM neural network according to claim 1, characterized in that: In the S1, the mass ratio of phosphogypsum to water is 1:10-20.

5. The method for predicting the optimal conditions for flocculation-assisted solidification based on SCA-LSTM neural network according to claim 4, characterized in that: In the S2, the curing agent is selected from calcium oxide, wherein the mass ratio of the curing agent to the mass ratio of fluorine in the filtrate is 1.5 to 7.

5.

6. The method for predicting the optimal conditions for flocculation-assisted solidification based on SCA-LSTM neural network according to claim 5, characterized in that: The flocculant in S2 is selected from inorganic flocculants containing aluminum and iron, including polyaluminum sulfate, or polyferric sulfate, or polyaluminum chloride, or polyferric chloride, or polyaluminum sulfate ferric, or polyaluminum ferric chloride, wherein the mass ratio of the flocculant to fluorine is 0-2.7:

1.

7. The method for predicting the optimal conditions for flocculation-assisted solidification based on SCA-LSTM neural network according to claim 6, characterized in that: The stirring time after adding the curing agent in S2 is 0 to 60 min, and the stirring time after adding the flocculant is 5 to 60 min; After adding the curing agent, the stirring rate is 50 r / min~500 r / min, and after adding the flocculant, the stirring rate is 50 r / min~500 r / min.

8. The method for predicting the optimal conditions for flocculation-assisted solidification based on SCA-LSTM neural network according to claim 7, characterized in that: In the above S2, the order of adding the flocculants is controlled by changing the stirring time after adding the curing agent. When the stirring time after adding the curing agent is 0, the curing agent and the flocculant are added at the same time. The order of adding the curing agent and the flocculant is: first add the curing agent, then add the flocculant; or add the curing agent and the flocculant at the same time.

9. The method for predicting the optimal conditions for flocculation-assisted solidification based on SCA-LSTM neural network according to claim 3, characterized in that: The maximum number of iterations of the SCA algorithm in S4 is set to 50, the step size is set to 2, and the number of individuals is set to 30.

10. The method for predicting the optimal conditions for flocculation-assisted solidification based on SCA-LSTM neural network according to claim 1, characterized in that: The specific formula for evaluation in S5 is: ； ； ； in represents the predicted value of the sample, represents the true value of the sample, n is the number of samples, is the mean of the true values ​​of the samples.

Images