Sodium aluminate solution concentration prediction method and system based on dynamic weight, processor and storage medium

Abstract

The invention discloses a sodium aluminate solution concentration prediction method and system based on dynamic weight, a processor and a storage medium, and belongs to the field of alumina production, S1, industrial process parameters are input into an improved Transform-based data-driven sodium aluminate solution concentration prediction sub-model, and a preliminary concentration prediction value is output; s2, outputting an error compensation value by using the sequence error prediction sub-model, and taking the error compensation value as a correction value of the initial concentration prediction value to obtain a corrected concentration prediction value; s3, utilizing the mechanism prediction sub-model to output a concentration prediction value under the physical constraint; and S4, carrying out dynamic weight fusion on the corrected concentration predicted value and the concentration predicted value under the physical constraint to obtain a final concentration predicted value. By the adoption of the sodium aluminate solution concentration prediction method and system based on the dynamic weight, the processor and the storage medium, high precision, high robustness and working condition adaptability of sodium aluminate solution concentration prediction are achieved.

Description

Technical Field

[0001] This invention relates to the field of alumina production technology, and in particular to a method, system, processor, and storage medium for predicting the concentration of sodium aluminate solution based on dynamic weights. Background Technology

[0002] In alumina production, the evaporation process is a crucial core step ensuring the stability of subsequent processes and product quality. Multi-effect evaporation and multi-stage flash evaporation systems, as core equipment, primarily undertake the key tasks of mother liquor concentration and cascaded utilization of heat energy. Their operating efficiency directly impacts the energy consumption level and resource utilization rate of the entire alumina production line. The concentration of sodium aluminate solution at the discharge port of this process is not only a core indicator for measuring product quality in the evaporation process but also a vital basis for subsequent operational control (such as setting process parameters in the decomposition section). Specifically, precise control of the sodium aluminate discharge port concentration directly affects the scaling rate of heat transfer equipment, the energy intensity of the production process, the supersaturation of the sodium aluminate solution in the decomposition section, and the particle size distribution of the final product. It also plays a decisive role in maintaining the alkali balance and the efficiency of secondary steam recovery and utilization in the production line.

[0003] However, in actual industrial production scenarios, multi-effect evaporation and multi-stage flash evaporation systems present a complex structure of multiple cascaded devices, resulting in significant strong coupling and lag characteristics in the entire system: on the one hand, the material transfer and energy exchange between the devices are closely related, and a small change in the parameter of a single device can trigger a chain reaction; on the other hand, the detection signal of the sodium aluminate solution concentration at the outlet needs to be transmitted and processed by multiple devices, resulting in significant detection lag, making it difficult to provide real-time feedback to the operation end to guide production control, which may lead to the concentration deviating from the target range and affecting the stability of subsequent processes.

[0004] To address the aforementioned concentration detection lag issue, soft sensing technology, a virtual sensing technique that indirectly acquires difficult / unmeasurable key parameters through measurable variables, has been widely applied in industrial process monitoring. Existing soft sensing modeling methods are mainly divided into three categories: First, mechanistic modeling methods. These methods construct mathematical expressions based on prior knowledge in the field (such as physical conservation laws and chemical reaction laws), offering advantages in strong interpretability and clear physical meaning. However, in the complex environment of the alumina evaporation process (such as operating condition fluctuations and equipment aging), they struggle to fully cover all nonlinear factors and interferences, resulting in limited prediction accuracy. Second, data-driven modeling methods. These methods mine implicit correlations between parameters by learning from large-scale industrial operating data, possessing high prediction accuracy and adaptability to operating conditions. However, because model construction relies on data distribution and lacks integration of physical laws, they suffer from poor interpretability and poor robustness to data quality fluctuations. The third is the hybrid modeling method, which combines the advantages of mechanism modeling and data-driven modeling through parallel, series, or weighted fusion to alleviate the shortcomings of single modeling methods to some extent. However, existing hybrid modeling schemes still have significant defects. Their fusion methods mostly adopt static weight allocation or fixed weights based on experience. That is, the contribution ratio of mechanism model and data-driven model remains fixed throughout the entire production cycle. It cannot adaptively adjust according to real-time changes in the production environment (such as fluctuations in raw liquid composition), sudden disturbances in the operating conditions (such as sudden changes in steam pressure), or equipment aging (such as a decrease in heat transfer coefficient). It is very easy to have prediction deviations under fluctuating operating conditions or non-steady-state operating conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, processor, and storage medium for predicting the concentration of sodium aluminate solution based on dynamic weights, thereby solving the aforementioned technical problems.

[0006] To achieve the above objectives, this invention provides a method for predicting the concentration of sodium aluminate solution based on dynamic weights, comprising the following steps: S1. Input the industrial process parameters into the data-driven sodium aluminate solution concentration prediction sub-model based on the improved Transformer to achieve preliminary prediction of sodium aluminate solution concentration and output the preliminary concentration prediction value. S2. Construct a sequence-to-sequence error prediction sub-model based on BiLSTM-attention. Take the industrial process parameters as the input of the sequence error prediction sub-model, take the difference between the actual concentration value and the preliminary concentration prediction value as the prediction target of the sequence error prediction sub-model, output the error compensation value, and then take the error compensation value as the correction value of the preliminary concentration prediction value to obtain the corrected concentration prediction value. S3. Utilize a mechanism prediction sub-model based on domain knowledge to output the concentration prediction value under physical constraints; S4. Construct a dynamic adaptive weight network, and dynamically integrate the corrected concentration prediction value output from step S2 and the concentration prediction value under physical constraints output from step S3 to obtain the final concentration prediction value.

[0007] Preferably, the improved Transformer described in step S1 includes an input layer, an improved encoder, a multi-scale convolutional module, a decoder, and an output layer arranged sequentially. The improved encoder uses temporal convolutional multi-head attention instead of a multi-head attention mechanism, and the improved encoder model expression is as follows: ; In the formula, Indicates the length of the encoder input sequence; Represents the hidden dimension of the encoder; The number of heads indicating multi-head attention; Indicates a single-head dimension; This indicates the kernel size of the causal convolution; Indicates the number of causal convolutional layers; Indicates the number of layers in the encoder; Indicates the first One point of attention; Indicates the kernel index; Represents the input sequence; , These represent the weights and biases of the linear projection of the input into the hidden dimension, respectively. Indicates the projected sequence; This indicates the sine or cosine position encoding of the encoder; This represents the initial representation after adding location information; Indicates entering the first Input before the layer encoder; Indicates the first The dilation rate of each causal convolutional sublayer is taken as... ; , They represent the first Layer encoder, first Each causal convolutional sublayer has a kernel offset. Linear mapping weights and biases at the point; Indicates the first Sub-layer at time The convolution output; Represents a non-linear activation function; Presentation layer normalization operation; The output after normalization of the representation layer; , and They represent the first Layer encoder The query, key, and value matrix of layered causal convolution; , and They represent the first The first layer encoder The projection matrix of query, key, and value moments in layered causal convolution; , and They represent the first Layer encoder Layer causal convolution The query submatrix, key matrix transpose, and value submatrix of each attention head; Indicates the first Layer encoder Layer causal convolution The output of each attention head; Indicates the first Layer encoder The output matrix of the multi-head attention mechanism of layered causal convolution; Indicates the first Layer encoder The linear projection matrix of the multi-head attention output of the layered causal convolution; This indicates splicing based on feature dimensions; Indicates the first The cross-layer attention feature concatenation matrix of the layer encoder; Indicates the first The cross-layer attention mapping matrix of the layer encoder; Indicates the random deactivation operator; Indicates the first The output of the layer encoder after layer normalization; Indicates a feedforward network. and They represent the first The two-layer linear transformation weight matrix of the layer encoder feedforward network; and These represent the two bias terms of the feedforward network, respectively. Indicates the first The improved layer encoder outputs a normalized feature matrix. Indicates the first The output matrix of the layer encoder feedforward network; The expression for the multi-scale convolution module is as follows: ; In the formula, These represent coarse-grained features, medium-grained features, and fine-grained features, respectively. Represents the activation function of the rectified linear unit; Indicates that the convolution kernel is Convolution operations; Indicates that the convolution kernel is Fill with Convolution operations; Indicates will Perform feature splicing; Indicates will Features after feature concatenation, and ; This represents the multi-scale fused features output by the multi-scale convolution module; The decoder model expression is as follows: ; In the formula, Indicates the starting marker; This represents the final step in retrieving the memory after multi-scale convolution; This indicates the starting point, up to the [number]th [marker]. The sequence of all markers up to step 1, concatenated chronologically, is used to perform the first step. Prediction of steps; Indicates to Add decoder sine or cosine position encoding The target sequence after; , and These represent the query matrix, key matrix, and value matrix, respectively. , , These represent the projection weight matrices of the query, key, and value in the self-attention mechanism, respectively. and These represent self-attention and cross-attention mechanisms, respectively. A mask matrix representing self-attention; This represents the output feature matrix of the self-attention mechanism; , and These represent the query matrix, key matrix, and value matrix of the cross-attention matrix, respectively. , and They represent the first In the layer cross-attention mechanism, the projection weight matrix of query, key, and value; This represents the transpose in cross attention; This represents the feature matrix of values ​​after weighting by attention weights in cross-attention; and They represent the first The two-layer linear transformation weight matrix of the layer decoder feedforward network; and They represent the first Two layers of bias terms in the feedforward network of the layer decoder; Indicates the first Layer decoder layer normalized output; Indicates the output sequence The last vector; This indicates the preliminary concentration forecast; This represents the transpose of the weight matrix of the final output layer; This represents the bias term of the final output layer; The data-driven sub-model expression for predicting sodium aluminate solution concentration is as follows: ; In the formula, This represents the function for the temporal convolutional multi-head attention mechanism. This represents the feature matrix after processing by the multi-scale convolution module; This represents the industrial process parameters after preprocessing and location coding; This represents a multi-scale convolution module function; This represents the decoder module function of the Transformer.

[0008] Preferably, step S1 specifically includes the following steps: S11. Collect industrial process parameters and input them into the improved Transformer for preprocessing and location encoding. The industrial process parameters include the inlet liquid volumetric flow rate. Liquid density Liquid concentration and liquid temperature and the density of the liquid at the discharge port. and concentration The steam flow rate, evaporator temperature, and evaporator pressure of each evaporator are pre-processed to obtain standardized industrial process parameters. And the position encoding expression is as follows: ; In the formula, Indicates the position encoding matrix of the first... Layer, First The encoded value of the dimension; Indicates the position encoding matrix of the first... Layer, First The encoded value; S12. Use an improved encoder based on the improved Transformer to extract features and output a feature matrix. ; S13, Feature Matrix Input a multi-scale convolution module, and use the multi-scale convolution module to process the feature matrix. Perform grouped convolution and multi-scale fusion to output multi-scale fused features. ; S14. Utilizing the decoder to fuse multi-scale features Decoding was performed to obtain preliminary concentration prediction values. .

[0009] Preferably, step S2 specifically includes the following steps: S21. Construct the following sequence error prediction sub-model: ; In the formula, This represents the prediction error of the sequence error prediction sub-model; Represents the decoder function; Represents the encoder function; Represents the attention mechanism function; S22. Using the industrial process parameters as input to the sequence error prediction sub-model constructed in step S21, the difference between the actual concentration value and the preliminary concentration prediction value is used as the prediction target of the sequence error prediction sub-model, the prediction error is output, and then the prediction error is used as the correction value of the preliminary concentration prediction value to obtain the corrected error compensation value: ; In the formula, express The concentration prediction value after time correction, express Time-based improved Transformer predictions.

[0010] Preferably, step S3 specifically includes the following steps: S31. Set the following assumptions for mechanism modeling: the change in solute mass caused by scaling is 0; the production steam is saturated steam and does not contain non-condensable gases; the material and steam are uniformly distributed in the heating tube; the heat transfer efficiency in the preheater is higher than the set threshold. S32. Constructing the equilibrium equation: Based on the principles of mass conservation and thermal balance, constructing a mechanism prediction sub-model: ; In the formula, This indicates the total mass flow rate of water evaporated during the evaporation process; express The total volumetric flow rate of secondary steam generated by the efficient evaporator; Indicates the first Liquid density in an efficient evaporator; S33. Model Initialization: Based on historical industrial process parameters, set the steam quantity error threshold. Concentration error threshold Correlation coefficient and initial estimate Initialize the given inlet feed parameters. And determine the physical property models describing the density, concentration, and temperature of the feed liquid. ; S34. Evaporation System Property Calculation: Based on the cascade relationship of the evaporation system, determine the density, concentration, and temperature of the liquid at the outlet of each evaporator, as well as the volumetric flow rate of the secondary steam generated by each evaporator. ; S35. Solve the equilibrium equations; S351, Using the heat balance formula Calculate the secondary steam volumetric flow rate of each effect evaporator ;in, , , and They represent the first The heat transfer coefficient, heat transfer area, effective temperature difference, and latent heat of vaporization of secondary steam of the effective evaporator; S352, Secondary steam volumetric flow rate for each effect evaporator Summing these values ​​yields the total mass flow rate of water evaporated during the evaporation process. : ; S353. The total mass flow rate of water evaporated during the evaporation process, obtained from step S352. Substitute the solutions into the equilibrium equations constructed in step S32 and solve. ; S354, Determine if the condition is met. and If so, the mechanism prediction sub-model is determined to have converged; otherwise, return to step S34 for updating. and They represent the first The second iteration and the first The concentration calculation value for the next iteration; S36. Calculate the total mass flow rate of evaporated water during the evaporation process using the converged mechanism prediction sub-model. Based on the principle of mass conservation, the volumetric flow rate of the liquid at the outlet is determined. Then, based on the principle of solute mass conservation, the concentration prediction value under physical constraints is obtained. .

[0011] Preferably, in step S353, if rely Then adopt Iteration, where, This indicates the temperature of the liquid at the outlet.

[0012] Preferably, the expression for the dynamic adaptive weight network in step S4 is as follows: ; in, ; In the formula, express Real-time dynamic weighting; Indicates the activation function; This represents a neural network used to learn feature maps; This represents the final predicted concentration value; express The concentration prediction value after time correction; express time, Time-corrected concentration prediction Concentration prediction under physical constraints The absolute difference between them; express The Monte Carlo random inactivation method is used to continuously perform sequence error prediction sub-modeling. The standard deviation of the residuals is obtained from the first forward propagation; express The mean error of the time series error prediction sub-model.

[0013] The system based on dynamic weights for predicting sodium aluminate solution concentration includes a data-driven sodium aluminate solution concentration prediction sub-model based on an improved Transformer, a sequence-to-sequence error prediction sub-model, a mechanism prediction sub-model, and a dynamic adaptive weight network. Among them, the data-driven sodium aluminate solution concentration prediction sub-model based on the improved Transformer is used to obtain preliminary concentration prediction values ​​based on industrial process parameters; A sequence-to-sequence error prediction sub-model is used to correct the initial concentration prediction value based on the actual concentration value, and obtain the corrected concentration prediction value. The mechanism prediction sub-model is used to output concentration prediction values ​​under physical constraints based on domain knowledge. A dynamic adaptive weighted network is used to dynamically weight and correct the concentration prediction value and the concentration prediction value under physical constraints to obtain the final concentration prediction value.

[0014] A processor is configured to execute a method for predicting the concentration of sodium aluminate solution based on dynamic weights.

[0015] A storage medium storing computer program instructions, characterized in that the computer program instructions, when executed by a processor, implement a method for predicting the concentration of sodium aluminate solution based on dynamic weights.

[0016] Therefore, the present invention employs the above-mentioned method, system, processor, and storage medium for predicting sodium aluminate solution concentration based on dynamic weights, and has the following beneficial effects: 1. The multi-scale convolution module can extract and fuse features of different granularities (coarse, medium, and fine), which enhances the ability to express the multi-scale temporal changes of sodium aluminate solution concentration (such as long-term trends, medium-term fluctuations, and short-term disturbances), and provides more comprehensive feature support for subsequent predictions. 2. The improved Transformer encoder, combined with the temporal convolutional multi-head attention mechanism, not only ensures the causality of temporal prediction (using only current and historical data), but also expands the receptive field, improving the ability to extract temporal dependencies and process-related features during the evaporation of sodium aluminate solution. 3. The dynamic adaptive weight network can dynamically adjust the weight ratio of the data-driven sub-model and the mechanism sub-model in the final prediction based on the prediction difference and error uncertainty, etc., which solves the limitations of a single model under different working conditions and improves the accuracy and robustness of concentration prediction. 4. The mechanism model constructed based on the principles of mass conservation and thermal balance provides physical constraints for the prediction of sodium aluminate solution concentration. Even when there are data anomalies, the physical rationality of the prediction results can still be guaranteed, thus enhancing the reliability of the model.

[0017] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0018] Figure 1 This is a flowchart of the sodium aluminate solution concentration prediction method based on dynamic weights of the present invention. Figure 2 The prediction results of the present invention as described in the simulation experiment are shown in the figure. Figure 3 This is a comparison chart of the prediction results of the present invention with other single models in the simulation experiment. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely illustrative of the embodiments of the present invention and are not intended to limit the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of this application. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout.

[0020] It should be noted that the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, such as a process, method, system, product, or server that includes a series of steps or units, not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such process, method, product, or device.

[0021] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0022] like Figure 1 As shown, the method for predicting the concentration of sodium aluminate solution based on dynamic weights includes the following steps: S1. Input the industrial process parameters into the data-driven sodium aluminate solution concentration prediction sub-model based on the improved Transformer to achieve preliminary prediction of sodium aluminate solution concentration and output the preliminary concentration prediction value. The improved Transformer described in step S1 includes an input layer, an improved encoder, a multi-scale convolutional module, a decoder, and an output layer arranged sequentially. The improved encoder uses temporal convolutional multi-head attention instead of a multi-head attention mechanism, and the improved encoder model expression is as follows: ; In the formula, Indicates the length of the encoder input sequence; Represents the hidden dimension of the encoder; The number of heads indicating multi-head attention; Indicates a single-head dimension; This indicates the kernel size of the causal convolution; Indicates the number of causal convolutional layers; Indicates the number of layers in the encoder; Indicates the first One point of attention; Indicates the kernel index; Represents the input sequence; , These represent the weights and biases of the linear projection of the input into the hidden dimension, respectively. Indicates the projected sequence; This indicates the sine or cosine position encoding of the encoder; This represents the initial representation after adding location information; Indicates entering the first Input before the layer encoder; Indicates the first The dilation rate of each causal convolutional sublayer is taken as... ; , They represent the first Layer encoder, first Each causal convolutional sublayer has a kernel offset. Linear mapping weights and biases at the point; Indicates the first Sub-layer at time The convolution output; Represents a non-linear activation function; Presentation layer normalization operation; The output after normalization of the representation layer; , and They represent the first Layer encoder The query, key, and value matrix of layered causal convolution; , and They represent the first The first layer encoder The projection matrix of query, key, and value moments in layered causal convolution; , and They represent the first Layer encoder Layer causal convolution The query submatrix, key matrix transpose, and value submatrix of each attention head; Indicates the first Layer encoder Layer causal convolution The output of each attention head; Indicates the first Layer encoder The output matrix of the multi-head attention mechanism of layered causal convolution; Indicates the first Layer encoder The linear projection matrix of the multi-head attention output of the layered causal convolution; This indicates splicing based on feature dimensions; Indicates the first The cross-layer attention feature concatenation matrix of the layer encoder; Indicates the first The cross-layer attention mapping matrix of the layer encoder; Indicates the random deactivation operator; Indicates the first The output of the layer encoder after layer normalization; Indicates a feedforward network. and They represent the first The two-layer linear transformation weight matrix of the layer encoder feedforward network; and These represent the two bias terms of the feedforward network, respectively. Indicates the first The improved layer encoder outputs a normalized feature matrix. Indicates the first The output matrix of the layer encoder feedforward network; The expression for the multi-scale convolution module is as follows: ; In the formula, These represent coarse-grained features, medium-grained features, and fine-grained features, respectively. Represents the activation function of the rectified linear unit; Indicates that the convolution kernel is Convolution operations; Indicates that the convolution kernel is Fill with Convolution operations; Indicates will Perform feature splicing; Indicates will Features after feature concatenation, and ; This represents the multi-scale fused features output by the multi-scale convolution module; The decoder model expression is as follows: ; In the formula, Indicates the starting marker; This represents the final step in retrieving the memory after multi-scale convolution; This indicates the starting point, up to the [number]th [marker]. The sequence of all markers up to step 1, concatenated chronologically, is used to perform the first step. Prediction of steps; Indicates to Add decoder sine or cosine position encoding The target sequence after; , and These represent the query matrix, key matrix, and value matrix, respectively. , , These represent the projection weight matrices of the query, key, and value in the self-attention mechanism, respectively. and These represent self-attention and cross-attention mechanisms, respectively. A mask matrix representing self-attention; This represents the output feature matrix of the self-attention mechanism; , and These represent the query matrix, key matrix, and value matrix of the cross-attention matrix, respectively. , and They represent the first In the layer cross-attention mechanism, the projection weight matrix of query, key, and value; This represents the transpose in cross attention; This represents the feature matrix of values ​​after weighting by attention weights in cross-attention; and They represent the first The two-layer linear transformation weight matrix of the layer decoder feedforward network; and They represent the first Two layers of bias terms in the feedforward network of the layer decoder; Indicates the first Layer decoder layer normalized output; Indicates the output sequence The last vector; This indicates the preliminary concentration forecast; This represents the transpose of the weight matrix of the final output layer; This represents the bias term of the final output layer; The data-driven sub-model expression for predicting sodium aluminate solution concentration is as follows: ; In the formula, This represents the function for the temporal convolutional multi-head attention mechanism. This represents the feature matrix after processing by the multi-scale convolution module; This represents the industrial process parameters after preprocessing and location coding; This represents a multi-scale convolution module function; This represents the decoder module function of the Transformer.

[0023] Step S1 specifically includes the following steps: S11. Collect industrial process parameters and input them into the improved Transformer for preprocessing and location encoding. The industrial process parameters include the inlet liquid volumetric flow rate. Liquid density Liquid concentration and liquid temperature and the density of the liquid at the discharge port. and concentration The steam flow rate, evaporator temperature, and evaporator pressure of each evaporator are pre-processed to obtain standardized industrial process parameters. And the position encoding expression is as follows: ; In the formula, Indicates the position encoding matrix of the first... Layer, First The encoded value of the dimension; Indicates the position encoding matrix of the first... Layer, First The encoded value; S12. Use an improved encoder based on the improved Transformer to extract features and output a feature matrix. ; S13, Feature Matrix Input a multi-scale convolution module, and use the multi-scale convolution module to process the feature matrix. Perform grouped convolution and multi-scale fusion to output multi-scale fused features. ; S14. Utilizing the decoder to fuse multi-scale features Decoding was performed to obtain preliminary concentration prediction values. .

[0024] S2. Construct a sequence-to-sequence error prediction sub-model based on BiLSTM-attention. Take the industrial process parameters as the input of the sequence error prediction sub-model, take the difference between the actual concentration value and the preliminary concentration prediction value as the prediction target of the sequence error prediction sub-model, output the error compensation value, and then take the error compensation value as the correction value of the preliminary concentration prediction value to obtain the corrected concentration prediction value. Step S2 specifically includes the following steps: S21. Construct the following sequence error prediction sub-model: ; In the formula, This represents the prediction error of the sequence error prediction sub-model; Represents the decoder function; Represents the encoder function; Represents the attention mechanism function; S22. Constructing the residual sequence: Using the industrial process parameters as input to the sequence error prediction sub-model constructed in step S21, the difference between the actual concentration value and the preliminary concentration prediction value is used as the prediction target of the sequence error prediction sub-model. The prediction error is output, and then the prediction error is used as the correction value of the preliminary concentration prediction value to obtain the corrected error compensation value. ; In the formula, express The concentration prediction value after time correction, express Time-based improved Transformer predictions.

[0025] S3. Utilize a mechanism prediction sub-model based on domain knowledge to output the concentration prediction value under physical constraints; Step S3 specifically includes the following steps: S31. Set the following assumptions for mechanism modeling: the change in solute mass caused by scaling is 0; the production steam is saturated steam and does not contain non-condensable gases; the material and steam are uniformly distributed in the heating tube; the heat transfer efficiency in the preheater is higher than the set threshold. S32. Constructing the equilibrium equation: Based on the principles of mass conservation and thermal balance, constructing a mechanism prediction sub-model: ; In the formula, This indicates the total mass flow rate of water evaporated during the evaporation process; express The total volumetric flow rate of secondary steam generated by the efficient evaporator; Indicates the first Liquid density in an efficient evaporator; S33. Model Initialization: Based on historical industrial process parameters, set the steam quantity error threshold. Concentration error threshold Correlation coefficient and initial estimate Initialize the given inlet feed parameters. And determine the physical property models describing the density, concentration, and temperature of the feed liquid. ; S34. Evaporation System Property Calculation: Based on the cascade relationship of the evaporation system, determine the density, concentration, and temperature of the liquid at the outlet of each evaporator, as well as the volumetric flow rate of the secondary steam generated by each evaporator. ; S35. Solve the equilibrium equations; S351, Using the heat balance formula Calculate the secondary steam volumetric flow rate of each effect evaporator ;in, , , and They represent the first The heat transfer coefficient, heat transfer area, effective temperature difference, and latent heat of vaporization of secondary steam of the effective evaporator; S352, Secondary steam volumetric flow rate for each effect evaporator Summing these values ​​yields the total mass flow rate of water evaporated during the evaporation process. : ; S353. The total mass flow rate of water evaporated during the evaporation process, obtained from step S352. Substitute the solutions into the equilibrium equations constructed in step S32 and solve. ; S354, Determine if the condition is met. and If so, the mechanism prediction sub-model is determined to have converged; otherwise, return to step S34 for updating. and They represent the first The second iteration and the first The concentration calculation value for the next iteration; S36. Calculate the total mass flow rate of evaporated water during the evaporation process using the converged mechanism prediction sub-model. Based on the principle of mass conservation, the volumetric flow rate of the liquid at the outlet is determined. Then, based on the principle of solute mass conservation, the concentration prediction value under physical constraints is obtained. .

[0026] Preferably, in step S353, if rely Then adopt Iteration, where, This indicates the temperature of the liquid at the outlet.

[0027] S4. Construct a dynamic adaptive weight network, and dynamically integrate the corrected concentration prediction value output from step S2 and the concentration prediction value under physical constraints output from step S3 to obtain the final concentration prediction value.

[0028] The expression for the dynamic adaptive weight network described in step S4 is as follows: ; in, ; In the formula, express Real-time dynamic weighting; Indicates the activation function; This represents a neural network used to learn feature maps; This represents the final predicted concentration value; express The concentration prediction value after time correction; express time, Time-corrected concentration prediction Concentration prediction under physical constraints The absolute difference between them; express The Monte Carlo random inactivation method is used to continuously perform sequence error prediction sub-modeling. The residual standard deviation obtained from the second forward propagation represents the uncertainty of the sequence error prediction sub-model, reflecting the uncertainty of the current sample. express The mean error of the time series error prediction sub-model means that the sub-model estimates the magnitude of the residuals of the main model, reflecting the magnitude of the main model's bias.

[0029] Simulation Experiment A typical piece of equipment from an alumina plant, as shown in Table 1, was selected as the test object. 800 continuously varying sodium aluminate solution outlet concentration data points from its evaporation process were collected as the data source for the implementation case. The first 560 sets of data were used for model training, the middle 80 sets for model validation, and the last 160 sets for model testing.

[0030] Table 1 Mechanism Models of Typical Equipment

[0031] In Table 1, This indicates the volumetric flow rate of the feed liquid; the subscript 0 indicates the feed liquid entering the six-effect reactor. Indicates the first Effective export; subscript This indicates the liquid phase at the flash evaporator outlet. Indicates the density of the liquid; This indicates the density of the corresponding steam. Indicates specific enthalpy of vapor, subscript This is secondary steam for primary effect. For the sake of efficiency, come to the car. For the first Flash steam; Indicates the temperature of the liquid feed; It is a specific heat at constant pressure; For heat transfer, This indicates the heat loss of the corresponding equipment; Indicates the overall heat transfer coefficient; Indicates the heat transfer area; Indicates the temperature of the heating steam; This indicates the temperature of the condensate.

[0032] To further ensure the effectiveness and feasibility of the present invention, prediction comparison experiments were conducted using the model described in this invention against traditional Transformer, Temporal Convolutional Network (TCN), Long Short-Term Memory (BiLSTM), and Support Vector Regression (SVR) models.

[0033] like Figure 2 As shown, in the process of the present invention from the improved Transformer to error compensation, and then to the adaptive fusion with the mechanism model, the difference between the predicted value and the true value obtained in each step gradually decreases, indicating that the model proposed in this invention improves the prediction accuracy of the sodium aluminate solution outlet concentration during the evaporation process.

[0034] from Figure 3It is evident that, compared to single prediction models such as SVR, Transformer, TCN, and BiLSTM, the model described in this invention performs superiorly in concentration change prediction tasks. Its predicted concentration changes show a higher degree of consistency with the actual concentration trends, and all prediction errors are significantly reduced, resulting in a substantial improvement in prediction accuracy. Conversely, single prediction models such as SVR, Transformer, TCN, and BiLSTM struggle to keep pace with actual concentration fluctuations, exhibiting significant deviations from the true values ​​and failing to adequately meet the demand for high-precision tracking of concentration changes.

[0035] To further evaluate the accuracy of the model described in this invention, the mean squared error (MSE), mean absolute error (MAE), root mean square error (RMSE), and coefficient of determination (R²) of the model described in this invention were compared with other prediction models, and the results are shown in Table 2.

[0036] Table 2

[0037] As can be seen from Table 2, the RMSE, MAE, MSE, and R of the prediction model described in this invention are... 2 All three indicators are optimal among all models. Furthermore, the MSE of the model designed in this invention is only 0.2033, lower than BiLSTM's 0.5321, TCN's 0.7680, Transformer's 0.8825, and SVR's 1.6923; the MAE is 0.3419, also significantly better than other models such as BiLSTM's 0.5729, TCN's 0.6862, Transformer's 0.7421, and SVR's 1.0432; the RMSE is only 0.4509, a reduction of over 30% compared to BiLSTM (0.7295), TCN (0.8763), Transformer (0.9394), and SVR (1.3009), indicating that the predicted values ​​obtained by this invention have a smaller deviation from the actual concentration values. Meanwhile, the R... 2 The value of 0.8825 is much higher than that of BiLSTM (0.6923), TCN (0.5560), Transformer (0.4898), and SVR (0.0216), which means that the model involved in this invention can accurately capture the dynamic change pattern of actual concentration data and avoid the problem of smooth fluctuation in prediction results common in other models.

[0038] In summary, the model described in this invention significantly outperforms single-data-driven prediction models such as SVR, Transformer, TCN, and BiLSTM in all core performance indicators for export concentration prediction. In terms of prediction accuracy, the model exhibits lower error and a smaller deviation between the predicted results and actual export concentration values, thus reflecting the true concentration situation more accurately. It also demonstrates outstanding sensitivity to concentration changes and trend tracking capabilities, accurately capturing the dynamic fluctuations in export concentration data and effectively avoiding the problem of predictive results tending to flatten and failing to keep pace with actual concentration changes, a common issue with other single-data-driven models. Therefore, the model described in this invention is not only a reliable export concentration prediction method but also, due to its high accuracy and strong sensitivity, can adapt to concentration prediction needs in more scenarios, significantly broadening its applicability.

[0039] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for predicting the concentration of sodium aluminate solution based on dynamic weights, characterized in that: Includes the following steps: S1. Input the industrial process parameters into the data-driven sodium aluminate solution concentration prediction sub-model based on the improved Transformer to achieve preliminary prediction of sodium aluminate solution concentration and output the preliminary concentration prediction value. S2. Construct a sequence-to-sequence error prediction sub-model based on BiLSTM-attention. Take the industrial process parameters as the input of the sequence error prediction sub-model, take the difference between the actual concentration value and the preliminary concentration prediction value as the prediction target of the sequence error prediction sub-model, output the error compensation value, and then take the error compensation value as the correction value of the preliminary concentration prediction value to obtain the corrected concentration prediction value. S3. Utilize a mechanism prediction sub-model based on domain knowledge to output the concentration prediction value under physical constraints; S4. Construct a dynamic adaptive weight network, and dynamically integrate the corrected concentration prediction value output from step S2 and the concentration prediction value under physical constraints output from step S3 to obtain the final concentration prediction value.

2. The method for predicting sodium aluminate solution concentration based on dynamic weights according to claim 1, characterized in that: The improved Transformer described in step S1 includes an input layer, an improved encoder, a multi-scale convolutional module, a decoder, and an output layer arranged sequentially. The improved encoder uses temporal convolutional multi-head attention instead of a multi-head attention mechanism, and the improved encoder model expression is as follows: ； In the formula, Indicates the length of the encoder input sequence; Represents the hidden dimension of the encoder; The number of heads indicating multi-head attention; Indicates a single-head dimension; This indicates the kernel size of the causal convolution; Indicates the number of causal convolutional layers; Represents the input sequence; , These represent the weights and biases of the linear projection of the input into the hidden dimension, respectively. Indicates the projected sequence; This indicates the sine or cosine position encoding of the encoder; This represents the initial representation after adding location information; Indicates entering the first Input before the layer encoder; Indicates the first The dilation rate of each causal convolutional sublayer is taken as... ; , They represent the first Layer encoder, first Each causal convolutional sublayer has a kernel offset. Linear mapping weights and biases at the point; Indicates the first Sub-layer at time The convolution output; Represents a non-linear activation function; Presentation layer normalization operation; The output after normalization of the representation layer; , and They represent the first Layer encoder The query, key, and value matrix of layered causal convolution; , and They represent the first The first layer encoder The projection matrix of query, key, and value moments in layered causal convolution; , and They represent the first Layer encoder Layer causal convolution The query submatrix, key matrix transpose, and value submatrix of each attention head; Indicates the first Layer encoder Layer causal convolution The output of each attention head; Indicates the first Layer encoder The output matrix of the multi-head attention mechanism of layered causal convolution; Indicates the first Layer encoder The linear projection matrix of the multi-head attention output of the layered causal convolution; This indicates splicing based on feature dimensions; Indicates the first The cross-layer attention feature concatenation matrix of the layer encoder; Indicates the first The cross-layer attention mapping matrix of the layer encoder; Indicates the random deactivation operator; Indicates the first The output of the layer encoder after layer normalization; Indicates a feedforward network. and They represent the first The two-layer linear transformation weight matrix of the layer encoder feedforward network; and These represent the two bias terms of the feedforward network, respectively. Indicates the first The improved layer encoder outputs a normalized feature matrix. Indicates the first The output matrix of the layer encoder feedforward network; The expression for the multi-scale convolution module is as follows: ； In the formula, These represent coarse-grained features, medium-grained features, and fine-grained features, respectively. Represents the activation function of the rectified linear unit; Indicates that the convolution kernel is Convolution operations; Indicates that the convolution kernel is Fill with Convolution operations; Indicates will Perform feature splicing; Indicates will Features after feature concatenation, and ; This represents the multi-scale fused features output by the multi-scale convolution module; The decoder model expression is as follows: ； In the formula, Indicates the starting marker; This represents the final step in retrieving the memory after multi-scale convolution; This indicates the starting point, up to the [number]th [marker]. The sequence of all markers up to step 1, concatenated chronologically, is used to perform the first step. Prediction of steps; Indicates to Add decoder sine or cosine position encoding The target sequence after; , and These represent the query matrix, key matrix, and value matrix, respectively. , , These represent the projection weight matrices of the query, key, and value in the self-attention mechanism, respectively. and These represent self-attention mechanism and cross-attention mechanism, respectively. A mask matrix representing self-attention; This represents the output feature matrix of the self-attention mechanism; , and These represent the query matrix, key matrix, and value matrix of the cross-attention matrix, respectively. , and They represent the first In the layer cross-attention mechanism, the projection weight matrix of query, key, and value; This represents the transpose in cross attention; This represents the feature matrix of values ​​after weighting by the attention weights in cross-attention; and They represent the first The two-layer linear transformation weight matrix of the layer decoder feedforward network; and They represent the first Two layers of bias terms in the feedforward network of the layer decoder; Indicates the first Layer decoder layer normalized output; Indicates the output sequence The last vector; This indicates the preliminary concentration forecast; This represents the transpose of the final output layer's weight matrix; This represents the bias term of the final output layer; The data-driven sub-model expression for predicting sodium aluminate solution concentration is as follows: ； In the formula, This represents the function for the temporal convolutional multi-head attention mechanism. This represents the feature matrix after processing by the multi-scale convolution module; This represents the industrial process parameters after preprocessing and location coding; This represents a multi-scale convolution module function; This represents the decoder module function of the Transformer.

3. The method for predicting sodium aluminate solution concentration based on dynamic weights according to claim 1, characterized in that: Step S1 specifically includes the following steps: S11. Collect industrial process parameters and input them into the improved Transformer for preprocessing and location encoding. The industrial process parameters include the inlet liquid volumetric flow rate. Liquid density Liquid concentration and liquid temperature and the density of the liquid at the outlet. and concentration The steam flow rate, evaporator temperature, and evaporator pressure of each evaporator are pre-processed to obtain standardized industrial process parameters. And the position encoding expression is as follows: ； In the formula, Indicates the position encoding matrix of the first... Layer, First The encoded value of the dimension; Indicates the position encoding matrix of the first... Layer, First The encoded value; S12. Use an improved encoder based on the improved Transformer to extract features and output a feature matrix. ; S13, Feature Matrix Input a multi-scale convolution module, and use the multi-scale convolution module to process the feature matrix. Perform grouped convolution and multi-scale fusion to output multi-scale fused features. ; S14. Utilizing the decoder to fuse multi-scale features Decoding was performed to obtain preliminary concentration prediction values. .

4. The method for predicting sodium aluminate solution concentration based on dynamic weights according to claim 3, characterized in that: Step S2 specifically includes the following steps: S21. Construct the following sequence error prediction sub-model: ； In the formula, This represents the prediction error of the sequence error prediction sub-model. Represents the decoder function; Represents the encoder function; Represents the attention mechanism function; S22. Constructing the residual sequence: Using the industrial process parameters as input to the sequence error prediction sub-model constructed in step S21, the difference between the actual concentration value and the preliminary concentration prediction value is used as the prediction target of the sequence error prediction sub-model. The prediction error is output, and then the prediction error is used as the correction value of the preliminary concentration prediction value to obtain the corrected error compensation value. ； In the formula, express The time-corrected concentration prediction value express Time-based improved Transformer predictions.

5. The method for predicting sodium aluminate solution concentration based on dynamic weights according to claim 4, characterized in that: Step S3 specifically includes the following steps: S31. Set the following assumptions for mechanism modeling: the change in solute mass caused by scaling is 0; the production steam is saturated steam and does not contain non-condensable gases; the material and steam are uniformly distributed in the heating tube; the heat transfer efficiency in the preheater is higher than the set threshold. S32. Constructing the equilibrium equation: Based on the principles of mass conservation and thermal balance, constructing a mechanism prediction sub-model: ； In the formula, This indicates the total mass flow rate of water evaporated during the evaporation process; express The total volumetric flow rate of secondary steam generated by the efficient evaporator; Indicates the first Liquid density in an efficient evaporator; S33. Model Initialization: Based on historical industrial process parameters, set the steam quantity error threshold. Concentration error threshold Correlation coefficient and initial estimate Initialize the given inlet feed parameters. And determine the physical property models describing the density, concentration, and temperature of the feed liquid. ; S34. Evaporation System Property Calculation: Based on the cascade relationship of the evaporation system, determine the density, concentration, and temperature of the liquid at the outlet of each evaporator, as well as the volumetric flow rate of the secondary steam generated by each evaporator. ; S35. Solve the equilibrium equations; S351, Using the heat balance formula Calculate the secondary steam volumetric flow rate of each effect evaporator ;in, , , and They represent the first The heat transfer coefficient, heat transfer area, effective temperature difference, and latent heat of vaporization of secondary steam of the effective evaporator; S352, Secondary steam volumetric flow rate for each effect evaporator Summing these values ​​yields the total mass flow rate of water evaporated during the evaporation process. : ； S353. The total mass flow rate of water evaporated during the evaporation process, obtained from step S352. Substitute the solutions into the equilibrium equations constructed in step S32 and solve. ; S354, Determine if the condition is met. and If so, the mechanism prediction sub-model is determined to have converged; otherwise, return to step S34 for updating. and They represent the first The second iteration and the first The concentration calculation value for the next iteration; S36. Calculate the total mass flow rate of evaporated water during the evaporation process using the converged mechanism prediction sub-model. Based on the principle of mass conservation, the volumetric flow rate of the liquid at the outlet is determined. Then, based on the principle of solute mass conservation, the concentration prediction value under physical constraints is obtained. .

6. The method for predicting sodium aluminate solution concentration based on dynamic weights according to claim 5, characterized in that: In step S353, if rely Then adopt Iteration, where, This indicates the temperature of the liquid at the outlet.

7. The method for predicting sodium aluminate solution concentration based on dynamic weights according to claim 6, characterized in that: The expression for the dynamic adaptive weight network described in step S4 is as follows: ； in, ； In the formula, express Real-time dynamic weighting; Indicates the activation function; This represents a neural network used to learn feature maps; This represents the final predicted concentration value; express The concentration prediction value after time correction; express time, Time-corrected concentration prediction Concentration prediction under physical constraints The absolute difference between them; express The Monte Carlo random inactivation method is used to continuously perform sequence error prediction sub-modeling. The standard deviation of the residuals is obtained from the first forward propagation; express The mean error of the time series error prediction sub-model.

8. The system of the method for predicting sodium aluminate solution concentration based on dynamic weights as described in any one of claims 1-7, characterized in that: This includes a data-driven sodium aluminate solution concentration prediction sub-model based on an improved Transformer, a sequence-to-sequence error prediction sub-model, a mechanism prediction sub-model, and a dynamic adaptive weight network; Among them, the data-driven sodium aluminate solution concentration prediction sub-model based on the improved Transformer is used to obtain preliminary concentration prediction values ​​based on industrial process parameters; A sequence-to-sequence error prediction sub-model is used to correct the initial concentration prediction value based on the actual concentration value, and obtain the corrected concentration prediction value. The mechanism prediction sub-model is used to output concentration prediction values ​​under physical constraints based on domain knowledge. A dynamic adaptive weighted network is used to dynamically weight and correct the concentration prediction value and the concentration prediction value under physical constraints to obtain the final concentration prediction value.

9. A processor, characterized in that: It is configured to perform the dynamic weight-based sodium aluminate solution concentration prediction method according to any one of claims 1-7.

10. A storage medium storing computer program instructions thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method for predicting the concentration of sodium aluminate solution based on dynamic weights as described in any one of claims 1-7.

Images