Advanced control method and related device of alumina caustic ratio under small sample conditions

Abstract

The present invention discloses an advanced control method for alumina causticity ratio under small sample conditions and related devices, wherein the method comprises: obtaining small sample data produced by an alumina plant; selecting a plurality of interference variables, a plurality of manipulated variables and a control variable that affect the causticity ratio from the small sample data, and merging the interference variables, the manipulated variables and the control variables into a training data set; extracting sequential samples of the training data set through a sliding data window to obtain a multivariate time series data set; obtaining an alumina causticity ratio network model through training based on the multivariate time series data set; freezing the self-attention block of the model during fine-tuning, and only updating the position encoding and the regularization term; inputting the production data of the alumina plant into the alumina causticity ratio network model to predict the causticity ratio value; and optimizing the manipulated variables through gradient descent rolling optimization based on multiple future causticity ratio values ​​to achieve advanced control of the alumina causticity ratio.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to an advanced control method and device for alumina causticity ratio under small sample conditions, and a computing device. Background Art

[0002] Alumina dissolution process is one of the main processes in the production of alumina. The purpose of dissolution is to dissolve the alumina hydrate in bauxite into sodium aluminate solution. The dissolution effect will directly affect the economic benefits of alumina production. The caustic ratio is an important indicator to measure the dissolution process, which determines the output, quality and alkali consumption of the product. The stable control of the caustic ratio is the key to optimizing the dissolution process.

[0003] Model predictive control (MPC) is an effective control method widely used in industrial processes. It obtains appropriate control strategies by predicting and optimizing the future response of the system. MPC aims to decompose a long time span into several shorter time spans or finite time spans to achieve optimal control. It continuously solves the optimization problem under system constraints in the rolling time domain and continuously outputs the control signal for the next moment. Due to the characteristics of the dissolution process, such as long process, multiple parameters, and strong coupling, the system reaction process fluctuates greatly, so that the control correction and disturbance have different characteristics in different working areas. The change of process variables will cause the control system to oscillate cyclically. The real-time prediction and online optimization of MPC are suitable for the alumina dissolution process with strong inertia and hysteresis. It faces two difficulties: (1) It is difficult to obtain effective training information and mark effective data; (2) MPC is a model-based control algorithm. It is crucial to establish an accurate system dynamic model. The accuracy of the model affects or determines the control performance of the closed-loop system. However, these requirements are difficult for actual factory applications. First, a large amount of data has put forward higher requirements for the intelligence of factory equipment, and most of the labels have to be manually marked or tested for a long time by experienced workers, such as high temperature and high pressure scenes, combustion scenes, and environments with toxic and harmful gases or liquids. Secondly, due to the complex coupling relationship and strong nonlinearity of industrial dynamic systems, it is difficult to establish a mechanism model. The update of neural network parameters is limited by small samples, and the traditional gradient descent method may cause the model to overfit or underfit. Finally, according to process requirements, production conditions will change periodically. More future condition information can effectively reduce control fluctuations, requiring a longer review window and prediction step size, and the long-term prediction capability of the prediction model has higher requirements. Summary of the invention

[0004] In view of the above problems, the present invention is proposed to provide an advanced control method and device for alumina causticity ratio under small sample conditions that only requires a small number of samples to fine-tune the network and improve the robustness of the model.

[0005] According to one aspect of the present invention, an advanced control method for alumina caustic ratio under small sample conditions is provided, comprising:

[0006] Get a small sample of data on alumina plant production;

[0007] Selecting a plurality of interference variables, a plurality of operating variables and a control variable that affect the caustic ratio from the small sample data, and merging the interference variables, the operating variables and the control variables into a training data set;

[0008] Extracting sequential samples of the training data set through a sliding data window to obtain a multivariate time series data set;

[0009] A network model of aluminum oxide causticity ratio is obtained by training according to the multivariate time series data set; wherein the network model of aluminum oxide causticity ratio freezes the self-attention block during fine-tuning and only updates the position encoding and regularization term;

[0010] Inputting the production data of the alumina plant into the alumina causticity ratio network model to predict the causticity ratio value;

[0011] Based on multiple future caustic ratio values, the operating variables are optimized by gradient descent rolling to achieve advanced control of alumina caustic ratio.

[0012] In an optional embodiment, the interference variables include the mother liquor caustic ratio, the mother liquor alkali concentration, the mother liquor temperature, the slurry fine particle size, the slurry solid content, the insulation retention tank temperature, the secondary steam temperature, the secondary steam flow rate and the desiliconization tank level, the operating variables include the slurry flow rate, the mother liquor flow rate and the new steam flow rate, and the controlled variable is the caustic ratio.

[0013] In an optional embodiment, the alumina caustic ratio network model uses GPT-2 as a backbone model, including:

[0014] Process variable input layer, Instance Norm+patch layer, Embedding layer, position encoder, frozen Transformer block, flatten and linear layer, and control variable output layer;

[0015] The frozen Transformer block includes a frozen multi-head attention layer and a frozen FFN layer.

[0016] In an optional embodiment, achieving advanced control of alumina caustic ratio further comprises:

[0017] Step 1, inputting the multivariate time series data set into the alumina causticity ratio network model to generate a plurality of prediction values;

[0018] Step 2, based on the future multiple caustic ratio values, optimizing the operating variables by gradient descent rolling, and controlling the caustic ratio of the alumina refinery by the operating variables, thereby improving the dissolution efficiency;

[0019] Step 3, collect the newly generated process variables, construct a multivariate time series data set, and input it into the alumina causticity ratio network model, and execute step 1 to continue training.

[0020] In an optional implementation, the objective function of the gradient descent is:

[0021]

[0022] in, , , ;

[0023] and is the loss weight of the model and control increments.

[0024] In an optional implementation, the objective function uses a specific formula of gradient descent:

[0025]

[0026]

[0027] in, Represents the learning rate.

[0028] In an optional embodiment, the change range of the operating variable The calculation formula is:

[0029] .

[0030] According to another aspect of the present invention, there is provided an advanced control device for alumina causticity ratio under small sample conditions, comprising:

[0031] Acquisition module, used to obtain small sample data produced by alumina plants;

[0032] A merging module, used for selecting several interference variables, several operating variables and one control variable that affect the caustic ratio from the small sample data, and merging the interference variables, the operating variables and the control variables into a training data set;

[0033] An extraction module, used for extracting sequential samples of the training data set through a sliding data window to obtain a multivariate time series data set;

[0034] A training module, used for training an alumina causticity ratio network model according to the multivariate time series data set; wherein the alumina causticity ratio network model freezes the self-attention block during fine-tuning and only updates the position encoding and regularization term;

[0035] A prediction module, used for inputting the production data of the alumina plant into the alumina causticity ratio network model to predict the causticity ratio value;

[0036] A control module is used to optimize the operating variables by gradient descent rolling based on multiple future caustic ratio values ​​to achieve advanced control of alumina caustic ratio.

[0037] According to another aspect of the present invention, there is provided a computing device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0038] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to execute operations corresponding to the advanced control method of alumina causticity ratio under the above-mentioned small sample working conditions.

[0039] According to the solution provided by the present invention, a small sample data of the production of an alumina plant is obtained; several interference variables, several operating variables and a control variable that affect the caustic ratio are selected from the small sample data, and the interference variables, operating variables and control variables are merged into a training data set; sequential samples of the training data set are extracted through a sliding data window to obtain a multivariate time series data set; an alumina caustic ratio network model is trained according to the multivariate time series data set; wherein the alumina caustic ratio network model freezes the self-attention block during fine-tuning, and only updates the position encoding and regularization terms; the production data of the alumina plant is input into the alumina caustic ratio network model to predict the caustic ratio value; based on multiple future caustic ratio values, the operating variables are optimized by gradient descent rolling to achieve advanced control of the alumina caustic ratio. The present invention freezes the pre-trained language model, only a small number of samples are needed to fine-tune the network, and disturbance variables are introduced and a larger prediction step size is used, thereby improving the robustness of the model.

[0040] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0042] Figure 1 A schematic flow chart of an advanced control method for alumina caustic ratio under a small sample working condition according to an embodiment of the present invention is shown;

[0043] Figure 2 A schematic diagram of an advanced control structure of alumina caustic ratio according to an embodiment of the present invention is shown;

[0044] Figure 3 A schematic diagram of a training process of alumina caustic ratio network model according to an embodiment of the present invention is shown;

[0045] Figure 4 A schematic diagram of constructing a data set according to an embodiment of the present invention is shown;

[0046] Figure 5 A schematic diagram of a single prediction sample form according to an embodiment of the present invention is shown;

[0047] Figure 6 A schematic diagram of the structure of the aluminum oxide caustic ratio network model according to an embodiment of the present invention is shown;

[0048] Figure 7 A schematic diagram showing all predicted values ​​of the aluminum oxide caustic ratio network model with 96 steps in an embodiment of the present invention is shown;

[0049] Figure 8 The advanced control effect diagram of alumina caustic ratio under a small sample working condition of an embodiment of the present invention is shown;

[0050] Fig. 9 A schematic diagram of the structure of an advanced control device for alumina caustic ratio under a small sample working condition of an embodiment of the present invention is shown;

[0051] Fig.10 A schematic diagram of the structure of a computing device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0052] The exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present invention are shown in the accompanying drawings, it should be understood that the present invention can be implemented in various forms and should not be limited by the embodiments set forth herein. On the contrary, these embodiments are provided in order to enable a more thorough understanding of the present invention and to enable the scope of the present invention to be fully communicated to those skilled in the art.

[0053] Figure 1The flowchart of the advanced control method of aluminum oxide caustic ratio under the small sample working condition of the embodiment of the present invention is shown. Figure 1 As shown, the following steps are included:

[0054] Step S101, obtaining small sample data produced by an alumina plant.

[0055] The current AI or AGI and other alumina caustic ratio network models require huge data to support the training and optimization process, while there are relatively few studies that combine small sample problems with control systems. This is because the design and optimization of control systems are usually based on precise mathematical models and rely on a large amount of data verification and adjustment. Data-driven control systems usually need to collect a large amount of operating data and establish accurate nonlinear models to describe the behavior of the system through statistical analysis and other methods. However, it is not applicable to small sample situations. Therefore, this application proposes a method that can achieve advanced control of alumina caustic ratio even in small sample situations.

[0056] Step S102: Selecting a plurality of interference variables, a plurality of operating variables and a control variable that affect the caustic ratio from the small sample data, and merging the interference variables, the operating variables and the control variable into a training data set.

[0057] like Figure 4 As shown in the figure, the data set construction in the input sequence is shown. Due to the complexity of the environment, the control variables are affected by multiple factors. It is difficult to find an accurate model to represent the behavior of the controlled system, so the introduction of disturbance variables To help find the causal relationship between variables.

[0058] In an optional embodiment, the interference variables include the mother liquor caustic ratio, the mother liquor alkali concentration, the mother liquor temperature, the slurry fine particle size, the slurry solid content, the insulation retention tank temperature, the secondary steam temperature, the secondary steam flow rate and the desiliconization tank level, the operating variables include the slurry flow rate, the mother liquor flow rate and the new steam flow rate, and the controlled variable is the caustic ratio.

[0059] For example, a small amount of alumina plant production data is input, and 9 interference variables that affect the caustic ratio (mother liquor caustic ratio, mother liquor alkali concentration, mother liquor temperature, slurry fineness, slurry solid content, insulation retention tank temperature, secondary steam temperature, secondary steam flow rate, desiliconization tank level), 3 operating variables (slurry flow rate, mother liquor flow rate, new steam flow rate) and 1 control variable (caustic ratio) are selected according to the alumina dissolution process, sensor equipment, expert experience, etc., and the interference variables, operating variables, and control variables are combined to form a data set.

[0060] Step S103, extracting sequential samples of the training data set through a sliding data window to obtain a multivariate time series data set.

[0061] like Figure 4 , Figure 5 As shown, it is necessary to vectors, in order to maintain the same Vectors with consistent shapes, at future time steps , The value is replaced by , so that the label value that needs to be trained is not included in the feature in advance. Because of the long-term characteristics of industrial dynamic systems, interference variables In time step It can be assumed that the value does not change or changes slightly. Expressed as A complete and uniformly sampled multivariate time series dataset uses a sliding data window to extract sequential samples. The window moves with a fixed step size, which is usually equal to 1 to ensure temporal continuity. The total length of each forecast sample is , including past data , the historical review length is , future data , the predicted length is , past data Used to predict future data , thereby training a large language offline model. Figure 7 The first prediction value of the 200 sliding windows in the model with a step length of 96 is shown. Benefiting from the model's huge prior knowledge, the model can accurately predict its results. Figure 7 All the prediction values ​​of the model with a 96-step length are displayed. For long time series, the next 96 steps contain more production information and operating condition changes, which is conducive to reducing control fluctuations.

[0062] Step S104, obtaining an alumina causticity ratio network model by training according to the multivariate time series data set; wherein the alumina causticity ratio network model freezes the self-attention block during fine-tuning, and only updates the position encoding and regularization term.

[0063] In this embodiment, the aluminum oxide caustic ratio network model is a generative pre-trained Transformer network, which uses a multi-head attention mechanism to capture complex dependencies in time series data. In addition, in order to accelerate the training process of the model and prevent overfitting, a multi-head attention freezing strategy is adopted. In the pre-training stage, some network layers are frozen and only some parameters are trained; in the fine-tuning stage, all layers are unfrozen and the entire network is trained.

[0064] In an optional embodiment, the alumina caustic ratio network model uses GPT-2 as a backbone model, including:

[0065] Process variable input layer, Instance Norm+patch layer, Embedding layer, position encoder, frozen Transformer block, flatten and linear layer, and control variable output layer;

[0066] The frozen Transformer block includes a frozen multi-head attention layer and a frozen FFN layer.

[0067] In this embodiment, the pre-trained GPT-2 is used as the backbone model, and the transformer as the main module contains a lot of prior knowledge. The data is input into the model after instance normalization and patching. In order to solve the problem that recursive and recurrent neural networks cannot be calculated in parallel, the transformer injects information about the relative or absolute position of the markers in the sequence. In the system, sine and cosine functions of different frequencies are used as position encoding. Figure 6 As shown in the figure, the Instance Norm layer is used to normalize the samples of each channel. The patch layer is used to divide the input data into patches of fixed size to capture local features. The Embedding layer is used to convert the input data into a vector of fixed dimension. The linear layer is used to adjust the dimension of the input data to meet the requirements of multi-head attention. The flattening layer is used to flatten the multi-dimensional tensor into one dimension to connect the linear layer. The multi-head attention layer is used to capture the information of different positions in the input sequence and calculate the attention weight between each position and other positions. Each position can also participate in the calculation of the attention weights of other positions, so that the Transformer can learn the interaction between different positions. Among them, the self-attention mechanism is an important part of the transformer model. Each token is converted into three different vectors: the query vector , key vector Sum value vector . It is obtained by multiplying the embedding vector by three different weight matrices, where the dimensions of the three matrices are also the same and their parameters are shared. In the system, a trainable fully connected layer is used to generate a learnable vector, and the multi-head attention formula is as follows:

[0068]

[0069] in, denote the query matrix, key matrix and value matrix respectively, express Matrix dimensions, It's multi-headed attention. They are The weight of .

[0070] The data after multi-head attention enters the feedforward neural network (FFN). As a fully connected feedforward neural network, it consists of two linear transformations (fully connected layers) and a nonlinear activation function (GELU). These three parts are performed in sequence. The activation function is usually inserted between the two linear transformations. The feedforward neural network performs nonlinear transformation and mapping on the embedded vector. The overall calculation process of the alumina caustic ratio network model is as follows:

[0071]

[0072] in, represents instance normalization, represents the attention mechanism, represents a feed-forward neural network, It's multi-headed attention.

[0073] Since the self-attention layer and feedforward neural network have learned a lot of knowledge from language datasets, the self-attention block of the Alumina Caustic Ratio Network model is frozen during fine-tuning, and only the position encoding and regularization terms are updated.

[0074] In an optional embodiment, the training of the alumina causticity ratio network model based on the multivariate time series data set further includes:

[0075] Step 1, inputting the multivariate time series data set into the alumina causticity ratio network model to generate a plurality of prediction values;

[0076] Step 2, optimizing the operating variables by rolling gradient descent to control the caustic ratio of the alumina refinery by the operating variables, thereby improving the dissolution efficiency;

[0077] Step 3, collect the newly generated process variables, construct a multivariate time series data set, and input it into the alumina causticity ratio network model, and execute step 1 to continue training.

[0078] like Figure 3 As shown, gradient descent is used to optimize the operating variables, and the parameters of the model are continuously adjusted in an iterative manner to minimize the error between the optimized value and the actual value. After multiple iterations, a set of optimized operating variables, namely the second operating variable set, is obtained. The first operating variable in the second operating variable set is input to the factory, the newly generated process variables are collected, the future operating variables are merged into the alumina caustic ratio network model, and step 1 is repeated to continue training the model. Each iteration uses the optimized operating variables to update the predicted value of the model, and the value of the operating variable is further adjusted by the gradient descent algorithm.

[0079] In an optional embodiment, the objective function of the alumina caustic ratio network model is:

[0080]

[0081] in, , , .

[0082] In this embodiment, the control system aims to find appropriate operating variables to make the system output as close to the set value as possible. Due to the influence of model errors and interference, the system's predicted value The actual output error needs to be corrected. Given the prediction and control range, the objective function is expressed as follows:

[0083]

[0084] The objective function is subject to the following constraints:

[0085] (1)

[0086] in, is the interference variable, is the operating variable, is the control variable, represents the prediction matrix, represents a vector;

[0087] in and is the loss weight of the alumina caustic ratio network model with the control increment, Indicates the set value, Indicates the magnitude of change in the manipulated variable.

[0088] The most likely sequence of manipulated variables is searched in the control horizon to minimize the objective function.

[0089] In an optional implementation, the objective function uses a specific formula of gradient descent:

[0090] (2)

[0091] The fine-tuned alumina causticity ratio network model is embedded into the system framework. In order to reduce the amount of calculation and quickly find the optimal result, the following gradient descent is used:

[0092] (3)

[0093] in represents the learning rate. According to formula (1), the inverse of the objective function is as follows:

[0094] (4)

[0095] In an optional embodiment, the change range of the operating variable The calculation formula is:

[0096] (5)

[0097] According to formula (2) and formula (5), it can be deduced that:

[0098] (6)

[0099] In formula (2) for:

[0100] (7)

[0101] Step S105, inputting the production data of the alumina plant into the alumina causticity ratio network model to predict the causticity ratio value.

[0102] like Figure 8 The figure shows the advanced control effect of alumina caustic ratio under small sample conditions. The system can successfully achieve the required control effect under small sample conditions.

[0103] Step S106, based on the multiple caustic ratio values, optimizing the operating variables by gradient descent to control the alumina caustic ratio network model.

[0104] like Figure 2 As shown, the prediction error is obtained based on the predicted multiple future caustic ratio values ​​and the actual value, and the operating variables are optimized by gradient descent to achieve the control of the alumina caustic ratio system.

[0105] According to the solution provided by the present invention, a small sample data of the production of an alumina plant is obtained; several interference variables, several operating variables and a control variable that affect the caustic ratio are selected from the small sample data, and the interference variables, operating variables and control variables are merged into a training data set; sequential samples of the training data set are extracted through a sliding data window to obtain a multivariate time series data set; an alumina caustic ratio network model is trained according to the multivariate time series data set; wherein the alumina caustic ratio network model freezes the self-attention block during fine-tuning, and only updates the position encoding and regularization terms; the production data of the alumina plant is input into the alumina caustic ratio network model to predict the caustic ratio value; based on multiple future caustic ratio values, the operating variables are optimized by gradient descent to realize the alumina caustic ratio control system. The present invention freezes the pre-trained language model, only a small number of samples are needed to fine-tune the network, and a disturbance variable is introduced and a larger prediction step size is used, thereby improving the robustness of the model.

[0106] It should be noted that the solution provided by the present invention has the following beneficial effects:

[0107] (1) A new advanced control method for alumina causticity ratio based on a large language model is proposed. Compared with the traditional control method, this method uses a frozen pre-trained language model and only requires a small number of samples to fine-tune the network, which reduces the labeling cost and meets the requirements of actual industrial processes.

[0108] (2) The system adds disturbance variables and uses a larger prediction step size, which improves the robustness of the model to future changes in operating conditions.

[0109] (3) The system combines the gradient descent method to derive the derivative of the attention mechanism, and proves the stability of the advanced control system of alumina caustic ratio through theoretical analysis.

[0110] The following is a proof of the system stability theory:

[0111] The network model of alumina caustic ratio is a prediction model based on gradient descent. The Jacobian matrix can be calculated to meet the gradient descent condition. In actual industrial processes, stable control methods are superior to advanced control methods. A stable control system needs to meet two requirements: (1) The loss function of the prediction model is convergent. (2) The rolling optimization can keep the control system stable and track the reference trajectory.

[0112] First, the convergence of large language models is studied. In the convergence analysis, the training loss gradually decreases and remains constant to prove the proposed method. The definition is as follows:

[0113]

[0114]

[0115]

[0116] in and is the prediction model, represents the zero-error network weight matrix, Represents the network weight matrix of the large language model during training. In the theoretical proof, the weights of industrial data and large language models are considered bounded. Then:

[0117]

[0118]

[0119] in , represents the Euclidean distance. Since the weight is bounded, there exists a positive real number α, .at the same time, Defined as a nonlinear mapping, the output of the network usually has upper and lower bounds.

[0120] The Lyapunov function is a method for analyzing the stability of control systems. Given a large language model based on transformer, the training error The proof is as follows:

[0121]

[0122] The derivative of the Lyapunov function is as follows:

[0123]

[0124] To prove is negative definite, subject to the following conditions:

[0125] (1) If , ,therefore Satisfy the conditions;

[0126] (2) If , , in the future, there will always be a solution that reaches the scope , because ,when When it is large enough, the following solution may be , therefore Satisfy the conditions.

[0127] According to the Lyapunov stability theorem. Consistently eventually bounded,convergence for large language models.

[0128] Then, the stability of the control system for aluminum oxide dissolution is analyzed. In the optimization problem of the controller based on gradient descent, the existing hyperparameters will not affect the stability of the control system, and the loss function weight and Usually positive, and the learning rate of the controller gradient descent It only determines the different optimization speeds and has nothing to do with control stability. The tracking control error can be defined as follows:

[0129]

[0130] in, Indicates the caustic ratio setting value, is the true output of the caustic ratio, therefore, , to prove stability, the Lyapunov function is as follows:

[0131]

[0132] The derivative of the Lyapunov function is as follows:

[0133]

[0134] According to the chain rule:

[0135]

[0136] According to the gradient descent and the previous derivation, the derivative of the Lyapunov function can be rewritten as:

[0137]

[0138] available , therefore the alumina caustic ratio control system based on the large language model is stable.

[0139] Fig. 9 The schematic diagram of the structure of the advanced control device for aluminum oxide caustic ratio under small sample conditions of an embodiment of the present invention is shown. The advanced control device for aluminum oxide caustic ratio under small sample conditions includes: an acquisition module 1010, a merging module 1020, an extraction module 1030, a training module 1040, a prediction module 1050 and a control module 1060.

[0140] The acquisition module 1010 is used to acquire small sample data produced by the alumina plant;

[0141] The merging module 1020 is used to select several interference variables, several operating variables and one control variable that affect the caustic ratio from the small sample data, and merge the interference variables, the operating variables and the control variables into a training data set;

[0142] The extraction module 1030 is used to extract sequential samples of the training data set through a sliding data window to obtain a multivariate time series data set;

[0143] The training module 1040 is used to train the aluminum oxide causticity ratio network model according to the multivariate time series data set; wherein the aluminum oxide causticity ratio network model freezes the self-attention block during fine-tuning and only updates the position encoding and regularization term;

[0144] The prediction module 1050 is used to input the production data of the alumina plant into the alumina causticity ratio network model to predict the causticity ratio value;

[0145] The control module 1060 is used to control the alumina caustic ratio network model by optimizing the operating variables through gradient descent rolling based on the multiple caustic ratio values.

[0146] In an optional embodiment, the interference variables include the mother liquor caustic ratio, the mother liquor alkali concentration, the mother liquor temperature, the slurry fine particle size, the slurry solid content, the insulation retention tank temperature, the secondary steam temperature, the secondary steam flow rate and the desiliconization tank level, the operating variables include the slurry flow rate, the mother liquor flow rate and the new steam flow rate, and the controlled variable is the caustic ratio.

[0147] In an optional embodiment, the alumina caustic ratio network model uses GPT-2 as a backbone model, including:

[0148] Process variable input layer, Instance Norm+patch layer, Embedding layer, position encoder, frozen Transformer block, flatten and linear layer, and control variable output layer;

[0149] The frozen Transformer block includes a frozen multi-head attention layer and a frozen FFN layer.

[0150] In an optional embodiment, the network model for controlling the alumina caustic ratio further comprises:

[0151] Step 1, inputting the multivariate time series data set into the alumina causticity ratio network model to generate a plurality of prediction values;

[0152] Step 2, based on the future multiple caustic ratio values, optimizing the operating variables by gradient descent rolling, and controlling the caustic ratio of the alumina refinery by the operating variables, thereby improving the dissolution efficiency;

[0153] Step 3, collect the newly generated process variables, construct a multivariate time series data set, and input it into the alumina causticity ratio network model, and execute step 1 to continue training.

[0154] In an optional implementation, the rolling optimization function of the gradient descent is:

[0155]

[0156] in, , , , , and is the loss weight of the model and control increment, Indicates the set value, represents the magnitude of change of the manipulated variable, is the historical input, Represents the predicted output.

[0157] In an optional implementation, the specific formula of the rolling optimization function using gradient descent is:

[0158]

[0159] in, Represents the learning rate.

[0160] In an optional embodiment, the change range of the operating variable The calculation formula is:

[0161] .

[0162] Fig.10 The schematic diagram of the structure of the computing device embodiment of the present invention is shown. The specific embodiment of the present invention does not limit the specific implementation of the computing device.

[0163] like Fig.10 As shown, the computing device may include: a processor (processor) 1102 , a communication interface (Communications Interface) 1104 , a memory (memory) 1106 , and a communication bus 1108 .

[0164] The processor 1102, the communication interface 1104, and the memory 1106 communicate with each other via a communication bus 1108. The communication interface 1104 is used to communicate with other devices such as a client or other server network elements. The processor 1102 is used to execute the program 1110, which can specifically execute the relevant steps in the above-mentioned embodiment of the advanced control method for the causticity ratio of aluminum oxide under the small sample working condition.

[0165] Specifically, the program 1110 may include program codes, which include computer operation instructions.

[0166] The processor 1102 may be a central processing unit (CPU), or an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present invention. The one or more processors included in the computing device may be processors of the same type, such as one or more CPUs; or processors of different types, such as one or more CPUs and one or more ASICs.

[0167] The memory 1106 is used to store the program 1110. The memory 1106 may include a high-speed RAM memory, and may also include a non-volatile memory (non-volatile memory), such as at least one disk memory.

[0168] According to the solution provided by the present invention, a small sample data of the production of an alumina plant is obtained; a number of interference variables, a number of operating variables and a control variable that affect the caustic ratio are selected from the small sample data, and the interference variables, operating variables and control variables are merged into a training data set; sequential samples of the training data set are extracted through a sliding data window to obtain a multivariate time series data set; an alumina caustic ratio network model is trained according to the multivariate time series data set; wherein the alumina caustic ratio network model freezes the self-attention block during fine-tuning, and only updates the position encoding and regularization terms; the production data of the alumina plant is input into the alumina caustic ratio network model to predict the caustic ratio value; based on multiple caustic ratio values, the operating variables are optimized by gradient descent rolling to control the alumina caustic ratio network model. The present invention freezes the pre-trained language model, only a small number of samples are needed to fine-tune the network, and a disturbance variable is introduced and a larger prediction step size is used, thereby improving the robustness of the model.

[0169] The algorithm or display provided herein is not inherently related to any particular computer, virtual system or other equipment. Various general purpose systems can also be used together with the teachings based on this. According to the above description, it is obvious to construct the structure required for this type of system. In addition, the embodiment of the present invention is not directed to any specific programming language yet. It should be understood that various programming languages ​​can be utilized to realize the content of the present invention described herein, and the description made to specific languages ​​above is for disclosing the best mode of the present invention.

[0170] In the description provided herein, a large number of specific details are described. However, it is understood that embodiments of the present invention can be practiced without these specific details. In some instances, well-known methods, structures and techniques are not shown in detail so as not to obscure the understanding of this description.

[0171] Similarly, it should be understood that in order to streamline the present invention and aid in understanding one or more of the various inventive aspects, in the above description of exemplary embodiments of the present invention, the various features of the embodiments of the present invention are sometimes grouped together into a single embodiment, figure, or description thereof. However, this disclosed method should not be interpreted as reflecting the following intention: that the claimed invention requires more features than the features explicitly recited in each claim. More specifically, as reflected in the claims below, inventive aspects lie in less than all the features of the individual embodiments disclosed above. Therefore, the claims that follow the specific embodiment are hereby expressly incorporated into the specific embodiment, with each claim itself serving as a separate embodiment of the present invention.

[0172] Those skilled in the art will appreciate that the modules in the devices in the embodiments may be adaptively changed and arranged in one or more devices different from the embodiments. The modules or units or components in the embodiments may be combined into one module or unit or component, and further may be divided into a plurality of submodules or subunits or subcomponents. All features disclosed in this specification (including the accompanying claims, abstracts and drawings) and all processes or units of any method or device so disclosed may be combined in any combination, except that at least some of such features and / or processes or units are mutually exclusive. Unless otherwise expressly stated, each feature disclosed in this specification (including the accompanying claims, abstracts and drawings) may be replaced by an alternative feature that provides the same, equivalent or similar purpose.

[0173] In addition, those skilled in the art will appreciate that, although some embodiments herein include certain features included in other embodiments but not other features, the combination of features of different embodiments is meant to be within the scope of the present invention and form different embodiments. For example, in the claims below, any one of the claimed embodiments may be used in any combination.

[0174] The various component embodiments of the present invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. It should be understood by those skilled in the art that a microprocessor or a digital signal processor (DSP) may be used in practice to implement some or all of the functions of some or all of the components according to the embodiments of the present invention. The present invention may also be implemented as a device or apparatus program (e.g., a computer program and a computer program product) for executing part or all of the methods described herein. Such a program implementing the present invention may be stored on a computer-readable medium, or may be in the form of one or more signals. Such a signal may be downloaded from an Internet website, or provided on a carrier signal, or provided in any other form.

[0175] It should be noted that the above embodiments illustrate the present invention rather than limit it, and that those skilled in the art may design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference symbol between brackets shall not be construed as a limitation on the claims. The word "comprising" does not exclude the presence of elements or steps not listed in the claims. The word "one" or "an" preceding an element does not exclude the presence of a plurality of such elements. The present invention may be implemented by means of hardware comprising a number of different elements and by means of a suitably programmed computer. In a unit claim that lists a number of devices, several of these devices may be embodied by the same hardware item. The use of the words first, second, and third, etc. does not indicate any order. These words may be interpreted as names. The steps in the above embodiments, unless otherwise specified, should not be understood as limitations on the order of execution.

Claims

1. An advanced control method for alumina caustic ratio under small sample conditions, characterized in that: include: Get a small sample of data on alumina plant production; Selecting several interference variables, several operating variables and one control variable that affect the caustic ratio from the small sample data, and merging the interference variables, operating variables and control variables into a training data set; the interference variables include the mother liquor caustic ratio, the mother liquor alkali concentration, the mother liquor temperature, the slurry fine particle size, the slurry solid content, the insulation retention tank temperature, the secondary steam temperature, the secondary steam flow rate and the desiliconization tank liquid level; the operating variables include the slurry flow rate, the mother liquor flow rate and the new steam flow rate; the control variable is the caustic ratio; Extracting sequential samples of the training data set through a sliding data window to obtain a multivariate time series data set; A network model of aluminum oxide causticity ratio is obtained by training according to the multivariate time series data set; wherein the network model of aluminum oxide causticity ratio freezes the self-attention block during fine-tuning and only updates the position encoding and regularization term; The alumina causticity ratio network model uses GPT-2 as the backbone model, including: Process variable input layer, Instance Norm+patch layer, Embedding layer, position encoder, frozen Transformer block, flatten and linear layer, and control variable output layer; The frozen Transformer block includes a frozen multi-head attention layer and a frozen FFN layer; The Instance Norm layer is used to normalize the samples of each channel; the patch layer is used to divide the input data into blocks of fixed size to capture local features; the Embedding layer is used to convert the input data into a vector of fixed dimension; the linear layer is used to adjust the dimension of the input data to meet the requirements of multi-head attention; the flattening layer is used to flatten the multi-dimensional tensor into one dimension to connect the linear layer; the multi-head attention layer is used to capture the information of different positions in the input sequence and calculate the attention weight between each position and other positions. Each position also participates in the calculation of the attention weights of other positions, so that the Transformer learns the interaction between different positions; Inputting the production data of the alumina plant into the alumina causticity ratio network model to predict the causticity ratio value; Based on multiple future caustic ratio values, the operating variables are optimized by gradient descent to control the alumina caustic ratio network model.

2. The advanced control method for alumina caustic ratio under small sample conditions according to claim 1, characterized in that: Achieving advanced control of alumina causticity ratio further includes: Step 1, inputting the multivariate time series data set into the alumina causticity ratio network model to generate a plurality of prediction values; Step 2, based on the future multiple caustic ratio values, optimizing the operating variables by gradient descent rolling, and controlling the caustic ratio of the alumina refinery by the operating variables, thereby improving the dissolution efficiency; Step 3, collect the newly generated process variables, construct a multivariate time series data set, and input it into the alumina causticity ratio network model, and execute step 1 to continue training.

3. The advanced control method of alumina caustic ratio under small sample conditions according to claim 1, characterized in that: The rolling optimization objective function of the gradient descent is: in, , , ; and is the loss weight of the model and control increments.

4. The advanced control method for alumina caustic ratio under small sample conditions according to claim 3, characterized in that: The specific formula of the rolling optimization using the gradient descent method is: in, Represents the learning rate.

5. The advanced control method for alumina caustic ratio under small sample conditions according to claim 4, characterized in that: The magnitude of the change in the manipulated variable The calculation formula is: 。 6. An advanced control device for alumina caustic ratio under small sample conditions, characterized in that: The method for controlling the alumina causticity ratio under the small sample condition as claimed in any one of claims 1 to 5 comprises: Acquisition module, used to obtain small sample data produced by alumina plants; A merging module, used for selecting several interference variables, several operating variables and one control variable that affect the caustic ratio from the small sample data, and merging the interference variables, the operating variables and the control variables into a training data set; An extraction module, used for extracting sequential samples of the training data set through a sliding data window to obtain a multivariate time series data set; A training module, used for training an alumina causticity ratio network model according to the multivariate time series data set; wherein the alumina causticity ratio network model freezes the self-attention block during fine-tuning and only updates the position encoding and regularization term; A prediction module, used for inputting the production data of the alumina plant into the alumina causticity ratio network model to predict the causticity ratio value; A control module is used to optimize the operating variables by gradient descent rolling based on multiple future caustic ratio values ​​to achieve advanced control of alumina caustic ratio.

7. A computing device comprising: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the advanced control method of alumina causticity ratio under small sample conditions according to any one of claims 1-5.

Images