Evaporation process regulation and control instruction generation method, system, equipment and medium

Abstract

The invention provides an evaporation process regulation and control instruction generation method and system, computer equipment and a medium, and belongs to the field of industrial process process index prediction.The method comprises the steps that an aluminum oxide evaporation process historical time sequence working condition sample is collected, and a nonlinear time sequence correlation feature vector is extracted through an encoder; the samples are divided into K subsets by combining a clustering algorithm and expert rules, the subsets are labeled, and the number is counted; generating an adversarial network expansion sample by using structural constraint for the category of which the sample size is less than a threshold value; training an integrated model by using each subset, outputting an initial prediction vector, splicing the initial prediction vector with the feature vector, and obtaining a prediction value through a GBDT element learner; and establishing a working condition switching cost function by using the predicted value and the label, optimizing to obtain a minimum cost path, and generating a working condition regulation and control instruction of the automatic control system. According to the method, joint optimization of the prediction result and the working condition switching path is realized, and the problems of large concentration measurement error, large working condition jump disturbance and unsafe processing process in the aluminum oxide process are solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of industrial process index prediction, and particularly relates to an evaporation process regulation instruction generation method, system, device and medium. BACKGROUND

[0002] Alumina is a key intermediate product for the production of basic non-ferrous metal materials, and is also an important basic raw material in many industrial fields such as aerospace, machinery manufacturing, chemical industry, energy and the like. As the method for preparing alumina widely used in the industry at present, the evaporation process plays a key role in the whole process, mainly bearing the functions of adjusting the concentration of mother liquor, maintaining the balance of circulating water volume and removing impurity salts and the like. Although modern enterprises have gradually introduced automatic measuring devices such as concentration meters and density meters, under the working conditions of high viscosity, high corrosive medium and high temperature and high pressure, the existing manual measuring means still has problems such as measurement lag and large error, resulting in a long parameter regulation feedback cycle. At the same time, the alumina evaporation process is faced with various complex working conditions, such as fluctuation of raw material composition, change of operating conditions, change of environmental temperature and dynamic switching of load, etc., so that the concentration of sodium aluminate at the outlet presents a highly nonlinear and strongly coupled characteristic.

[0003] In recent years, the modeling and prediction method based on data driving can realize accurate prediction of key indexes of the evaporation system. Among them, the ensemble learning method is widely concerned due to its high modeling ability and strong generalization performance for complex nonlinear relationship. However, most of the traditional ensemble models are developed for stable working conditions, and it is difficult to fully adapt to the frequent change of complex working conditions and sudden disturbance in the alumina evaporation process. There is generally a lack of dynamic sensing ability and cross-model coordination mechanism, and the prediction model is unstable between different working conditions, which is difficult to meet the robustness and generalization requirements of industrial-level prediction. SUMMARY

[0004] In order to solve the problems of large concentration measurement error and large working condition jump disturbance in the alumina process, the application provides an evaporation process regulation instruction generation method, system, computer device and medium.

[0005] In order to achieve the above purpose, the application provides the following technical scheme, an evaporation process regulation instruction generation method, comprising: Respectively collect energy consumption parameters, temperature parameters and time sequence working condition samples in the alumina evaporation process; extract a nonlinear time sequence correlation feature vector of the time sequence working condition sample.

[0006] Input the time sequence working condition sample into a pre-trained base learner, output K initial sodium aluminate solution concentration prediction vectors, input the K initial sodium aluminate solution concentration prediction vectors and the nonlinear time sequence correlation feature vector into a pre-trained GBDT meta-learner after splicing, and predict the sodium aluminate solution concentration value.

[0007] based on energy consumption parameters, temperature parameters and predicted sodium aluminate solution concentration values in an alumina evaporation process, The predicted sodium aluminate solution concentration values are adjusted by preset optimization rules to obtain an optimal regulation path adapted to the current process, and then a working condition regulation instruction for a self-control system is generated.

[0008] Preferably, the time series working condition samples are processed by a stacked ensemble prediction model based on multi-learner fusion to obtain sodium aluminate solution concentration prediction values; the stacked ensemble prediction model based on multi-learner fusion comprises K base learners and a GBDT meta-learner; the K base learners are obtained by training base learners from K working condition sample subsets.

[0009] Preferably, before training, the stacked ensemble prediction model based on multi-learner fusion also needs to be sample augmented, specifically including: sample augmentation of working condition categories with a sample number less than a set threshold in the working condition sample subset by a structure constraint generative adversarial network; the structure constraint generative adversarial network is composed of a conditional input design module, a generator and a discriminator connected in sequence; wherein the conditional input design module combines and encodes working condition labels and nonlinear time series correlation feature vectors to form a conditional input vector; the conditional input vector is input into the generator network for transposed convolution and up-sampling to generate simulated samples; the discriminator network is used to extract features of the simulated samples, and the nonlinear time series correlation feature vectors are used to judge the authenticity of the simulated samples, and the real samples are retained to obtain augmented samples.

[0010] Preferably, when training, the stacked ensemble prediction model based on multi-learner fusion also includes: optimizing network parameters of the base learners and the GBDT meta-learner by a Bayesian hyperparameter algorithm.

[0011] Preferably, the time series working condition samples are divided into K working condition sample subsets by an embedding space clustering algorithm and expert rules, specifically including: based on the nonlinear time series correlation feature vectors, the time series working condition samples are divided into K working condition sample subsets.

[0012] Preferably, the energy consumption parameter, temperature parameter and predicted sodium aluminate solution concentration value in the alumina evaporation process are used to establish a working condition switching cost function, the predicted sodium aluminate solution concentration value is optimized by using the working condition switching cost function, the optimal regulation path suitable for the current process is obtained, and then the working condition regulation instruction of the automatic control system is generated, specifically including: using the energy consumption parameter, temperature parameter and predicted sodium aluminate solution concentration value in the alumina evaporation process to establish a working condition switching cost function; obtaining the sodium aluminate solution concentration prediction value of a plurality of continuous time sequence working condition samples; substituting the sodium aluminate solution concentration prediction value of the plurality of continuous time sequence working condition samples into the working condition switching cost function, calculating the single-step cost of adjacent working condition switching; accumulating the single-step cost by using a dynamic programming algorithm; solving the minimum total cost path from the current working condition to the target working condition from the accumulated single-step cost; the prediction value corresponding to the minimum total cost path is the optimized concentration regulation target, and the working condition change path is the optimal regulation path.

[0013] Preferably, the working condition switching cost function expression is as follows: ; Wherein, And represents the prediction value of the model j at time i ; And the prediction value of the model k at time i +1 ; The membership degree of the first i data point of the non-cluster center to the first j working condition cluster, The membership degree of the first i +1 data point of the non-cluster center to the first k working condition cluster, is the influence of the ambient temperature at time i +1 on the prediction value; T is the time, is used to adjust the dynamic amplification or attenuation of the error size, β decides the maximum amplification multiple, α controls the attenuation speed.

[0014] Also provided is an evaporation process regulation instruction generation system, comprising: A sample collection module is used to collect energy consumption parameters, temperature parameters and time sequence working condition samples in an alumina evaporation process, and extract nonlinear time sequence correlation feature vectors of the time sequence working condition samples.

[0015] A model training module is used to input the time sequence working condition samples into a pre-trained base learner, and output Kan initial sodium aluminate solution concentration prediction vector, inputting the initial sodium aluminate solution concentration prediction vector and a nonlinear time series correlation feature vector into a pre-trained GBDT meta-learner, and predicting a sodium aluminate solution concentration value. K an initial sodium aluminate solution concentration prediction vector, inputting the initial sodium aluminate solution concentration prediction vector and a nonlinear time series correlation feature vector into a pre-trained GBDT meta-learner, and predicting a sodium aluminate solution concentration value.

[0016] a model application module, configured to establish a working condition switching cost function based on energy consumption parameters, temperature parameters and the predicted sodium aluminate solution concentration value in the alumina evaporation process, optimize the predicted sodium aluminate solution concentration value by using the working condition switching cost function, obtain an optimal regulation and control path adapted to the current process, and further generate a working condition regulation and control instruction for a self-control system.

[0017] The application further provides a computer device, including a memory, a processor and a computer program stored in the memory, and the processor executes the computer program to implement the steps of any one of the industrial migration cost set methods.

[0018] The application further provides a computer readable storage medium, and the storage medium stores a computer program, and the computer program can execute the steps of any one of the industrial migration cost set methods when loaded by a processor.

[0019] The evaporation process regulation and control instruction generation method provided by the application has the following beneficial effects: The energy consumption, temperature parameters and time series working condition samples are collected respectively, so that the complex characteristics of the multi-factor coupling in the alumina evaporation process can be fully captured, and a data foundation for accurate prediction is laid; the nonlinear time series correlation feature vector is extracted, so that the strong nonlinearity and dynamic correlation between parameters can be effectively mined, and the feature representation capability is improved; the initial prediction vector output by the pre-trained base learner is spliced with the feature vector to input the GBDT meta-learner, so that the complementarity of multiple models is fully utilized, and the prediction accuracy and generalization ability of the sodium aluminate solution concentration are greatly improved; the working condition switching cost function is established based on the energy consumption, temperature parameters and predicted concentration, and is optimized, so that unreasonable working condition jumps can be avoided, the path stability and system controllability are ensured, the generated regulation and control instruction is adapted to the actual process, and the intelligent operation level of the evaporation system is significantly improved, thereby providing strong support for stable production, energy saving and emission reduction. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the embodiments of the application and the design scheme thereof, the following will briefly introduce the drawings required by the embodiments. The drawings in the following description are only some embodiments of the application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0021] Figure 1 A flowchart of an evaporation process regulation and control instruction generation method of an embodiment of the application; Figure 2 A prediction flowchart for the evaporation process regulation instruction generation in the embodiment of the present application facing complex working conditions is provided. Figure 3 A comparison chart of the prediction value and the real value of the base learner and the integrated cost migration model in the embodiment of the present application is provided. Figure 4 A model prediction error distribution box chart in the embodiment of the present application is provided. DETAILED DESCRIPTION

[0022] In order to make the skilled in the art better understand the technical solutions of the present application and can be implemented, the present application is described in detail below in conjunction with the drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and cannot limit the protection scope of the present application.

[0023] The present application provides an evaporation process regulation instruction generation method, specifically as shown in Figure 1 , comprising: S1, respectively collecting energy consumption parameters, temperature parameters and time sequence working condition samples in the alumina evaporation process; extracting the nonlinear time sequence correlation feature vector of the time sequence working condition sample.

[0024] In the alumina evaporation process, the field process working condition sample is collected according to the time dimension, and the basic data support is prepared for model establishment.

[0025] S2, inputting the time sequence working condition sample into the pre-trained base learner, outputting K An initial sodium aluminate solution concentration prediction vector, inputting the K An initial sodium aluminate solution concentration prediction vector and the nonlinear time sequence correlation feature vector into the pre-trained GBDT meta-learner, and predicting the sodium aluminate solution concentration value.

[0026] Intelligent identification and classification of working conditions are used to realize intelligent identification and label acquisition of multi-effect evaporation system under complex operating conditions. An improved deep embedding clustering (DEC) method based on convolutional neural network embedding is proposed. The initial clustering result is judged and corrected combined with expert knowledge, and the fine classification of historical operating state is realized. This step specifically includes the following sub-steps.

[0027] S2.1, construct a CNN encoder network structure as a deep embedding module, perform feature compression and abstraction on the original multi-dimensional time sequence process data, and form a low-dimensional semantic embedding space. The input data is set as X∈ℝ, wherein n is the number of samples, T is the time length, d is the feature dimension. Feature encoding is obtained through multi-layer one-dimensional convolution and pooling operation: non-supervised clustering analysis is performed on the historical process operation data combined with the expert knowledge on site:

[0028] ; where, Z is the embedding representation of sample features, n=800 is the number of samples, m is the compressed embedding dimension for subsequent clustering tasks. The CNN encoder consists of two layers of one-dimensional convolution: the first layer Conv1D(filters=64, kernel_size=3), the second layer Conv1D(filters=128, kernel_size=3), each layer is followed by BatchNormalization and ReLU activation, and after the second layer, Dropout(0.2) is added. Then use GlobalAveragePooling1D to compress the time dimension, m=128-dimensional embedding vector.

[0029] S2.2, introduce an improved DEC clustering module in the embedding space, and perform soft clustering by presetting the number of categories K. The similarity between sample embedding points and cluster centers is measured by student t distribution, and the soft label distribution is defined as:

[0030] ; is the soft label distribution, which represents the probability that sample i belongs to cluster j , and the sum of the probabilities of all clusters is 1. Then construct the auxiliary target distribution pᵢⱼ , and optimize the embedding space by KL divergence to make samples of the same class closer and samples of different classes farther apart: ; is the auxiliary target distribution, which is based on further strengthened distribution, so that the probability of samples of the same class is higher and the probability of samples of different classes is lower. The loss between and is measured, so that gradually approaches , so that the embedding points of samples of the same class are closer to the cluster centers, resulting in the concentration of soft label probabilities; the embedding points of samples of different classes are farther apart, achieving the purpose of classification. Iteratively update the network weights and cluster centers to obtain structured and separable embedding clustering results.

[0031] S2.3, introduce on-site expert experience to correct the preliminary clustering results of the model. By setting expert rules (such as: under the same raw liquid temperature + low load + high steam temperature working condition, it should be classified as “strong evaporation” category), label correction is performed on the boundary fuzzy samples, and the final working condition label matrix is generated:

[0032] ; where, Repr representing the expert rule set, Y DEC outputting labels for deep clustering model, expert is a rule fusion function. The label will be used as the basis for the working condition division of subsequent prediction sub-models.

[0033] To solve the problem that some working conditions have low frequency in actual operation, insufficient data sampling, and difficulty in effectively training stable prediction models, an improved generative adversarial network (GAN-g) method based on structural prior constraints is proposed to enhance the generation of small sample working condition data, improve the generalization ability and robustness of the model under a small amount of samples. This step includes the following sub-steps.

[0034] S2.4, combine the divided working condition labels with their corresponding original nonlinear time series correlation feature vectors to encode and construct a conditional input vector.

[0035] ; wherein, is the working condition sample feature, is the working condition category label vector. The conditional input serves as the input of the generator network, which helps to generate synthetic sample distribution under controllable working conditions.

[0036] The generator and the discriminator are trained in an adversarial manner (GAN-gFramework) to construct a conditional adversarial network structure of the generator G and the discriminator D . During training, the generator attempts to synthesize a small sample data that approximates the real data, and the discriminator distinguishes between real and fake samples, thereby optimizing the network through game learning.

[0037] Generator loss function: ; wherein, represents the loss function of the generator, which is used to guide the generator to optimize the parameters through backpropagation; represents a random noise variable in the latent space, represents a latent disturbance variable sampled from a normal distribution generator working condition sample generated by the conditional generator under the specified working condition label c; the discriminator outputs the probability that the input data is judged to be real data.

[0038] Discriminator loss function: ; wherein, is the real working condition feature from the actual six-effect countercurrent four-effect flash distillation system, The discriminator determines the probability of real data being "true", and the original multi-dimensional process data is subjected to Min-Max normalization processing to map each feature to the interval [0, 1]. Then, a generative network based on WGAN-GP (WassersteinGAN with gradient penalty) is used to generate synthetic feature data. The generator (Generator) inputs a random noise vector z e R64, passes it through a two-layer fully connected network (hidden layer dimension 128, activation function ReLU), and finally maps it to the same space as the original feature dimension through a Tanh activation. The WGAN-GP gradient penalty λ = 10; before each Generator update, the Critic is updated for 5 steps; the batch size is 64; the number of training rounds is 5000; and x is the number of real samples, 800. By continuously optimizing the above loss function, the generator can generate more credible working condition samples.

[0039] After training, the similarity of the distribution of generated samples and real samples in the embedding space is evaluated, and the indicators Fréchet Inception Distance (FID) and Wasserstein Distance (WD) are used for evaluation. When the indicators reach the preset standard, the generated samples and the original samples are fused in proportion to construct an enhanced training dataset, effectively improving the performance and robustness of the model in small sample scenarios.

[0040] A stacked ensemble model based on multi-learner fusion is constructed, which fuses BiLSTM, XGBoost, and Autoformer three types of heterogeneous models as base learners. All individual multi-learner fusion stacked ensemble models are trained K Respectively, the real-time collected time series working condition samples are input into K individual multi-learner fusion stacked ensemble models, and K individual initial concentration prediction vectors are output K The and nonlinear time series correlation feature vectors Z are spliced and input into the GBDT meta-learner, and the predicted values of the working condition samples are output .

[0041] Z-score standardization was applied to the original process features, and the data was divided into training and test sets (70% / 30%) in chronological order. The models used were: XGBoost (100 trees, maximum depth 4, learning rate 0.1, random seed 42, mean squared error of the objective function); early stopping on the validation set could be incorporated during training. BiLSTM (2-layer bidirectional LSTM, 64 hidden units per layer (128-dimensional output after bidirectional concatenation), Dropout 0.2, linear mapping of the regression output layer from 128 to 1; Adam optimizer, learning rate 10%. -3 The system uses a random seed of 42, 100 training epochs, a batch size of 32, a loss function of MSE, and an early stopping strategy. The Autoformer model has the following characteristics: hidden dimension d_model=64, multi-head attention heads n_heads=4, encoder layers e_layers=2, decoder layers d_layers=1, prediction length pred_len=10; optimizer Adam, learning rate 1e-3, and random seed 42.

[0042] The gradient boosting decision tree is selected as the meta-learner of the stacked ensemble learning model. The steps for outputting the predicted concentration value of sodium aluminate solution are as follows: The prediction results of the base learners and the features of the dataset are used as input. The meta-learner is used to initialize the model as a simple predictor: ; in, The initial model output values, No. i The true sodium aluminate concentration value of each sample The current model for the first i The predicted output for each sample.

[0043] Calculate the residuals between the predicted values ​​and the actual values ​​of the current model. These residuals reflect the prediction error of the current model: ; in, In the mth round, the... i The pseudo residual values ​​of each sample. For the first m The predicted value obtained from -1 iterations. The selected loss function.

[0044] Training a new model involves training a new weak learner on the current residuals, with the goal of fitting the residuals: ; in, This represents the regression tree model for the current round. The input features corresponding to the sample are used to build a sub-model that can minimize the residual squared error of all samples. By gradually approaching the error direction, the convergence speed and fitting accuracy of the concentration modeling are effectively improved, which is suitable for small sample modeling situations such as caustic disturbance and new steam load change.

[0045] The update model weights the prediction results of the new model and adds them to the current model. The weighting factor is called the learning rate, which controls the degree of contribution of each new model to the final prediction:

[0046] The prediction value of the final model is the weighted sum of the model prediction values in all iterations: The prediction results of the base learner and the features of the data set are used as input, and the gradient boosting decision tree is used as the meta-learner. By training the meta-learner, the advantages of each base learner are utilized to obtain more accurate and stable prediction results for the evaporation process.

[0047] S3, based on the energy consumption parameters, temperature parameters and predicted sodium aluminate solution concentration values in the alumina evaporation process, a working condition switching cost function is established, and the predicted sodium aluminate solution concentration values are optimized using the working condition switching cost function to obtain the optimal control path that adapts to the current process, and then generate the working condition control instruction for the automatic control system.

[0048] To improve the adaptability of the model to continuous working condition evolution, a dynamic migration cost function based on concentration fluctuation is introduced, and combined with the path optimization mechanism, global optimization control of the historical to current prediction path is realized.

[0049] S3.1, the working condition migration cost function is constructed.

[0050] Where, and represent the prediction value of the model j at time i ; and the prediction value of the model k at time i +1 ; is the membership of the i th data point of the non-cluster center to the j th working condition cluster, is the membership of the i +1th data point of the non-cluster center to the k th working condition cluster, is the influence of the ambient temperature at time i +1 T ​​​For time, Adjustment for dynamic amplification or attenuation of error size, β Decide the maximum amplification, α Control the attenuation speed.

[0051] Based on multi-condition data, after obtaining the prediction results of each model on the test set, define the single-point cost function point_cost, comprehensively consider the prediction error absolute value, condition stability index H , temperature gradient R_T Optimization parameters α=0.010472, β =0.298795, γ=0.521254, ϕ=2.671286; H [i] Mapping from "original liquid aluminum oxide" index; Comprehensively consider the multi-dimensional influence factors such as state amplitude, energy consumption response, time sequence disturbance and temperature difference mutation, which is a scalable, controllable and engineering interpretation dynamic condition switching cost model.

[0052] S3.2, in the whole prediction time window t =1, 2,..., T, the optimal condition path is constructed by minimizing the total cost using dynamic programming or heuristic path search algorithm: ; Among them, Indicates a complete condition selected from time t =1 to T; Indicates the optimal condition path finally solved, with the minimum cost; Indicates the t th condition state node at time i ; Indicates the cost of transferring from the current state To the next time state ; In the state path reasoning process, the model not only adjusts the prediction value at each time, but also ensures the overall stability and system controllability of the path at the condition migration level, avoiding the system risks caused by violent fluctuations and frequent jumps.

[0053] S3.3, select the concentration prediction sequence corresponding to the migration path with the lowest cost as the final output result of the model, realize the prediction result correction under physical constraints, and significantly improve the continuity, smoothness and actual condition consistency of the prediction curve.

[0054] To further illustrate the effectiveness and feasibility of the present embodiment, the production data of the evaporation process of an alumina plant are selected, specifically including: the temperature of the raw solution, the temperature of the five-effect steam, the concentration of the four-flash caustic alkali, the outlet steam flow of the six-effect, the evaporation temperature difference of each stage, and the concentration of the outlet sodium aluminate solution, etc. The data sampling period is 1 minute, covering multiple typical working condition periods of continuous operation, with strong representativeness, large volatility, and complex coupling relationship between variables. The data are divided into training set and validation set according to 7:3.

[0055] The mean absolute percentage error (MAPE), mean absolute error (MAE), and root mean square error (RMSE) of the prediction model described in the present embodiment, the coefficient of determination (R²), and different prediction models such as BiLSTM, XGBoost, and Autoformer are compared, and the results are shown in Table 1. The cost transfer integrated model proposed in the present application has a significant performance advantage in the outlet concentration prediction task compared to the comparative models. As can be seen from Table 1, the model achieves the minimum values in the three key precision indicators of MAE, MSE, and RMSE, which are 0.32, 0.44, and 0.56, respectively, indicating that the deviation of the prediction results from the actual values is the smallest and the stability is the highest. In terms of the coefficient of determination (R²) for measuring the goodness of fit of the model, the cost transfer integrated model reaches 0.81, which is much higher than XGBoost (0.44), BiLSTM (0.52), and standard Autoformer (0.63). The model proposed in the present application not only has stronger prediction accuracy and robustness, but also exhibits obvious technical innovation and practical value in handling complex working condition transfer and information fusion, which can significantly improve the intelligent prediction level of the outlet concentration in industrial processes.

[0056] Table 1 Comparison of error indicators of outlet concentration prediction by four methods As Figure 2As shown, the embodiments of the present application include intelligent identification and classification of working conditions; using deep embedding clustering method to combine the expert knowledge of the scene to perform unsupervised clustering analysis on the historical process operation data, and divide the multi-effect evaporation system running process into multiple typical working condition categories. Small sample working condition data enhancement; introduce structure constraint generative adversarial network, combine the physical process characteristics of various working conditions, enhance the data of insufficient training samples under part of rare working conditions, and generate representative extended samples. Feature extraction and integrated prediction model construction; construct a stack integrated model based on multi-learner fusion, introduce GBDT as a meta-learner, and optimize the parameters of each base learner based on the Bayesian optimization algorithm. Working condition migration cost modeling and optimal path reasoning; introduce a working condition migration cost function based on the concentration fluctuation amount of sodium aluminate solution, measure the prediction risk and energy consumption change of switching between different working conditions. Combined with the dynamic path optimization algorithm, the optimal working condition migration trajectory from the history to the current time is inferred, and the concentration prediction result is combined for overall optimization, thereby improving the whole process control ability of the model under dynamic complex working conditions.

[0057] As shown, Figure 3 The prediction curves of different base learner models and cost migration integrated models on the whole sample data are shown in the embodiments of the present application, and are compared with the true value. It is observed that the prediction of a single model in some time periods has obvious deviation, especially in the interval with large local fluctuation, the prediction results of BiLSTM and XGBoost have large fluctuation, and high-frequency oscillation or deviation from the actual value is easy to occur. Compared with the cost migration integrated model, the cost migration integrated model can better track the change trend of the true value in most time periods, and has stronger smoothness and tracking ability. At the same time, after absorbing the prediction information of multiple base learners, the cost migration integrated model effectively suppresses the local error amplification problem easy to occur in a single model, maintains the same overall trend while reducing the sharp fluctuation of the prediction curve. Combined with error distribution analysis and visualization results, it is shown that the cost migration integrated model has obvious advantages in improving prediction accuracy, reducing abnormal error and enhancing robustness, and is the preferred strategy for solving the key variable prediction task under complex working conditions.

[0058] As shown, Figure 4As shown, the embodiments of the present application show the prediction error distribution of the cost migration integrated model and different base learning models. From the box plot, it can be seen that the error distribution of the cost migration integrated model is the most concentrated, the abnormal value is the least, and the overall error range is the smallest, which shows that it performs best in error control and model stability. Although the median error of the BiLSTM model is small, there are a small number of large negative abnormal values, reflecting that the prediction fluctuation on part of the samples is large. The error distribution of the XGBoost model is relatively dispersed, and there are multiple extreme values, indicating that its robustness is poor. The Autoformer model is slightly better than XGBoost in error control, but still has certain fluctuations. In summary, the cost migration integrated model has obvious advantages in reducing prediction error, suppressing outliers and improving model generalization ability, and is a better choice to solve this kind of complex time series prediction problem.

[0059] Based on the same inventive concept, the present application also provides an evaporation process regulation instruction generation system, comprising: a sample collection module, configured to collect energy consumption parameters, temperature parameters and time sequence working condition samples in an alumina evaporation process respectively, and extract a nonlinear time sequence correlation feature vector of the time sequence working condition samples; a model training module, configured to input the time sequence working condition samples into a pre-trained base learning device, and output K an initial sodium aluminate solution concentration prediction vector, and input the K initial sodium aluminate solution concentration prediction vector and the nonlinear time sequence correlation feature vector into a pre-trained GBDT meta-learner to predict a sodium aluminate solution concentration value; a model application module, configured to establish a working condition switching cost function based on the energy consumption parameters, the temperature parameters and the predicted sodium aluminate solution concentration value in the alumina evaporation process, optimize the predicted sodium aluminate solution concentration value by using the working condition switching cost function, obtain an optimal regulation path adapted to the current process, and then generate a working condition regulation instruction for a self-control system.

[0060] The present application also provides a computer device, which, at the hardware level, comprises a processor, an internal bus, a network interface, a memory and a non-volatile memory, and of course can also comprise other hardware required by other services. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs to implement the above-provided evaporation process regulation instruction generation method.

[0061] The present application also provides a computer readable storage medium, which stores a computer program, and the computer program can be used to execute the above-provided evaporation process regulation instruction generation method.

[0062] The specific limitation of the evaporation process regulation instruction generation method calculation system can refer to the limitation of the evaporation process regulation instruction generation method in the above, which will not be repeated here. Each module in the above evaporation process regulation instruction generation system can be realized by software, hardware and their combination in whole or in part. The above modules can be embedded in or independent of the processor in the computer device in hardware form, or can be stored in the memory in the computer device in software form, so as to call and execute the operation corresponding to each module by the processor.

[0063] It should be noted that the above specific embodiments can enable those skilled in the art to have a more comprehensive understanding of the present application, but in no way limit the present application. Therefore, although the present application has been described in detail in the specification and examples, those skilled in the art should understand that the present application can still be modified or replaced by equivalents; all technical solutions and improvements which do not deviate from the spirit and scope of the present application are covered in the protection scope of the present application. Any reference signs in the claims should not be regarded as limiting the claims involved.

Claims

1. A method for generating evaporation process regulation instructions, characterized in that, The method comprises: Respectively collect energy consumption parameters, temperature parameters and time sequence working condition samples in the alumina evaporation process; extract the nonlinear time sequence correlation feature vector of the time sequence working condition sample; Input the time sequence working condition sample into the pre-trained base learner, output K initial sodium aluminate solution concentration prediction vectors, input the K initial sodium aluminate solution concentration prediction vectors and the nonlinear time sequence correlation feature vector into the pre-trained GBDT meta-learner after splicing, and predict the sodium aluminate solution concentration value; Based on the energy consumption parameters, temperature parameters and predicted sodium aluminate solution concentration value in the alumina evaporation process, a working condition switching cost function is established, the predicted sodium aluminate solution concentration value is optimized using the working condition switching cost function, the optimal regulation and control path adapting to the current process is obtained, and then the working condition regulation and control instruction of the automatic control system is generated.

2. The method of claim 1, wherein, The time sequence working condition sample is processed by the stacked ensemble prediction model based on multi-learner fusion to obtain the sodium aluminate solution concentration prediction value. The stacked ensemble prediction model based on multi-learner fusion comprises K base learners and a GBDT meta-learner. The K base learners are trained by K working condition sample subsets.

3. The method of claim 2, wherein, Before training, the stacked ensemble prediction model based on multi-learner fusion also needs to expand the sample, which specifically includes: using the structure constraint generative adversarial network to expand the sample of the working condition class whose sample number is less than the set threshold; the structure constraint generative adversarial network is composed of a conditional input design module, a generator and a discriminator connected in turn; wherein the conditional input design module combines and encodes the working condition label and the nonlinear time sequence correlation feature vector to form a conditional input vector; the conditional input vector is input into the generator network for transposed convolution and up-sampling to generate simulated samples; the features of the simulated samples are extracted by the discriminator network, and the simulated samples are judged to be true or false combined with the nonlinear time sequence correlation feature vector, the real samples are retained, and the expanded samples are obtained.

4. The method of claim 2, wherein, When training, the stacked ensemble prediction model based on multi-learner fusion also includes: optimizing the network parameters of the base learner and the GBDT meta-learner using the Bayesian hyperparameter algorithm.

5. The method of claim 2, wherein, The time sequence working condition sample is divided into K working condition sample subsets by using the embedding space clustering algorithm and expert rules, specifically including: based on the nonlinear time sequence correlation feature vector, the time sequence working condition sample is divided into K working condition sample subsets.

6. The method of claim 1, wherein, The energy consumption parameter, temperature parameter and predicted sodium aluminate solution concentration value in the alumina evaporation process are used to establish a working condition switching cost function, the predicted sodium aluminate solution concentration value is optimized by using the working condition switching cost function, and an optimal regulation and control path suitable for the current process is obtained, and the method specifically comprises the following steps: establishing a working condition switching cost function by using the energy consumption parameter, temperature parameter and predicted sodium aluminate solution concentration value in the alumina evaporation process; obtaining sodium aluminate solution concentration prediction values of a plurality of continuous time sequence working condition samples; substituting the sodium aluminate solution concentration prediction values of the plurality of continuous time sequence working condition samples into the working condition switching cost function, and calculating a single-step cost of adjacent working condition switching; accumulating the single-step cost by using a dynamic programming algorithm; solving a minimum total cost path from a current working condition to a target working condition from the accumulated single-step cost; the prediction value corresponding to the minimum total cost path is the optimized concentration regulation and control target, and the working condition change path is the optimal regulation and control path.

7. The method of claim 6, wherein, The working condition switching cost function expression is as follows: ; wherein, and respectively represent the model j at time i the predicted value at time k +1 of the model i ; ; is the membership of the i th data point that is not a cluster center to the j th working condition cluster, is the membership of the i +1th data point that is not a cluster center to the k th working condition cluster, is the influence of the ambient temperature at time i +1 on the predicted value; T is the time, is the adjustment for dynamic amplification or attenuation of the error size, β determines the maximum amplification factor, α controls the attenuation speed.

8. An evaporation process regulation instruction generation system, characterized by The method comprises the following steps: A sample collection module is used to collect energy consumption parameters, temperature parameters and time sequence working condition samples in the alumina evaporation process respectively; A nonlinear time sequence correlation feature vector of the time sequence working condition sample is extracted; The model training module is configured to input the time series working condition sample into a pre-trained base learner to output an initial sodium aluminate solution concentration prediction vector. K The initial sodium aluminate solution concentration prediction vector and a nonlinear time sequence correlation feature vector are spliced and input into a pre-trained GBDT meta-learner to predict a sodium aluminate solution concentration value. K The initial sodium aluminate solution concentration prediction vector and a nonlinear time sequence correlation feature vector are spliced and input into a pre-trained GBDT meta-learner to predict a sodium aluminate solution concentration value. A model application module is used to establish a working condition switching cost function based on the energy consumption parameter, temperature parameter and predicted sodium aluminate solution concentration value in the alumina evaporation process, optimize the predicted sodium aluminate solution concentration value by using the working condition switching cost function, obtain an optimal regulation and control path suitable for the current process, and then generate a working condition regulation and control instruction for the automatic control system.

9. A computer device, wherein a memory is stored with a computer program, and the computer device comprises a processor, wherein the computer device is configured to execute the computer program to perform the method according to any one of claims 1-8. The processor executes the computer program to realize the method steps of any one of claims 1 to 7.

10. A computer readable storage medium storing a computer program, characterized in that, The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 7. The computer program is executed by the processor to realize the steps of the method of any one of claims 1 to 7.

Images