Long and short term memory network model prediction control method and device based on data driving

Abstract

The invention provides a long-short-term memory network model prediction control method and device based on data driving, and belongs to the technical field of industrial automation control, and the method specifically comprises the steps: inputting the current state information and historical working information of a ball mill at a tk-1 moment into a long-short-term memory network model, and predicting an expected state operation value at a tk moment; the expected state operation value at the tk moment and the current state information actually measured at the tk-1 moment of the ball mill are input into a control optimization solver, a control variable sequence is obtained, and the control variable sequence is used for being output to an executing mechanism of the ball mill to be controlled; when it is judged that the output condition is met, the control variable sequence is output to an execution mechanism; when it is judged that the output condition is not met, a long-short-term memory network prediction module is iteratively called based on the current state information of the ball mill equipment at the tk moment to predict the state at the tk + 1 moment, and the state is output to a control optimization solver. By means of the treatment scheme, precise regulation and control over the running state of the ball mill are improved.

Description

Technical Field

[0001] The present invention relates to the field of industrial automation control technology, and in particular to a data-driven long short-term memory network model predictive control method and device. Background Art

[0002] In the field of industrial automation control, ball mill is a key equipment, and its operating status directly affects production efficiency and product quality. However, due to the high nonlinearity, strong coupling and time-varying nature of the ball mill system, traditional control methods, such as PID control, often fail to achieve ideal control effects. In particular, when dealing with multi-variable and multi-constrained control problems, traditional methods often have problems such as low control accuracy and poor robustness.

[0003] In recent years, with the rapid development of artificial intelligence technology, deep learning technology has been widely used in the field of industrial automation control. As a special recurrent neural network (RNN), the long short-term memory network (LSTM) has powerful time series processing capabilities and can capture long-term dependencies in data, so it is very suitable for model construction in predictive control.

[0004] However, there are still some challenges in applying LSTM to ball mill control systems. For example, the training of LSTM models requires a large amount of high-quality data, while actual industrial environments often have problems such as data scarcity and low data quality. In addition, the prediction accuracy and generalization ability of LSTM models are also affected by factors such as model structure and parameter settings.

[0005] Therefore, how to design a data-driven improved long short-term memory (LSTM) network model predictive control method to handle multi-input and multi-output situations based on the actual operating characteristics of the ball mill in order to improve control accuracy and robustness has become an urgent problem to be solved in the current field of ball mill automation control. Summary of the invention

[0006] Therefore, in order to overcome the shortcomings of the above-mentioned prior art, the present invention provides a data-driven long short-term memory network model predictive control method and device, which improves the accuracy and effect of the control system by optimizing the MPC prediction model.

[0007] In order to achieve the above object, the present invention provides a data-driven long short-term memory network model predictive control method, comprising:

[0008] t k-1 The current state information and historical working information of the ball mill at the moment are input into the long short-term memory network model, and t is predicted. kThe expected state calculation value at the time, the historical working information includes at least one of the feed rate, the water supply before the ball mill, the water added to the pump pool, the cyclone overflow flow, and the frequency of the east-stage pump, and the expected state calculation value includes at least the ball mill power and the cyclone overflow fineness;

[0009] The t k The expected state calculation value at time t k-1 The current state information actually measured at the moment is input into the control optimization solver to obtain a control variable sequence, and the control variable sequence is used to output to the actuator of the ball mill for control;

[0010] When it is determined that the output condition is met, the control variable sequence is output to the actuator; when it is determined that the output condition is not met, based on t k The current state information of the ball mill equipment at time t is iteratively called to predict the long short-term memory network prediction module k+1 The state at the moment is output to the control optimization solver.

[0011] In one embodiment, when determining t k The expected state calculation value at time t k When the difference between the current state information actually measured at the moment is within a preset range, the control variable sequence is output to the actuator;

[0012] When determining t k The expected state calculation value at time t k When the difference between the current state information actually measured at the time exceeds the preset range, the long short-term memory network prediction module is obtained from t k-1 The number of calls since the moment;

[0013] When it is determined that the number of calls is not less than the preset upper limit, the control variable sequence is output to the actuator; when it is determined that the number of calls is less than the preset upper limit, based on t k The current state information of the ball mill equipment at time t is used again to call the long short-term memory network prediction module to predict t k+1 The state at the moment is output to the control optimization solver.

[0014] In one embodiment, the long short-term memory network model adopts a state encoder-state decoder structure, and the long short-term memory network model includes at least one state encoder and a state decoder.

[0015] The state encoder processes an input sequence and generates an encoded sequence, wherein the input sequence includes the t k-1 The current state information and historical working information of the ball mill at each moment, wherein the coding sequence summarizes the information in the input sequence;

[0016] The state decoder generates the sequence of control variables using the encoding sequence.

[0017] In one embodiment, the state encoder is composed of T long short-term memory units; the state decoder is composed of 3 fully connected network layers,

[0018] The input sequence of the state encoder is [t 1 , tτ] the current state information and historical working information of the ball mill at the time; the input sequence of the state decoder is the encoding sequence of the state encoder.

[0019] In one embodiment, the control optimization solver is obtained by iterative training, wherein the control quantity constraint during the training process is

[0020] u min ≤u(k)≤u max

[0021] Δu min ≤Δu(k)≤Δu max

[0022] Δu(k+j)=u(k+j)-u(k+j-1)

[0023] in, is the virtual control variable to be solved in model predictive control, Δu min and Δu max are the minimum and maximum values ​​allowed for the control sequence variation, u min and u max They are the minimum and maximum values ​​of the range of variation allowed by the control sequence.

[0024] In one embodiment, the control optimization solver is obtained by iterative training, wherein the loss function during the training process is Among them, y m (k+1)=f(u(k), u(k-1), u(k-2), ..., y(k), y(k-1)), y(k-2), ...), except for u(k) at time k, which is an unknown quantity, all other quantities before time k are known. y(k) is the actual reading of the instrument, y r (k+1) is a reference value, which is a known quantity.

[0025] Δu(k+j)=u(k+j)-u(k+j-1)

[0026] Δy m (k+j)=y m (k+j)-y m (k+j-1)

[0027]

[0028] y m (k)=f(u(k-1),u(k-2),...,u(kn u +1),y(k-1)),y(k-2),...,y(kn y +1)), n u 、n y are the delay orders of the input control variable time series and the controlled object output time series respectively.

[0029] A data-driven long short-term memory network model predictive control device, the device comprising:

[0030] The data acquisition module is used to collect the real-time data of the corresponding sensors of the ball mill and store them in correspondence with the current moment;

[0031] The LSTM prediction module is used to input the current state information and historical working information of the ball mill at time tk-1 into the LSTM model and predict t k The expected state calculation value at the time, the historical working information includes at least one of the feed rate, the water supply before the ball mill, the water added to the pump pool, the cyclone overflow flow, and the frequency of the east-stage pump, and the expected state calculation value includes at least the ball mill power and the cyclone overflow fineness;

[0032] A control optimization solver is used to compare the desired state calculation value at time tk with the ball mill tk- 1 The current state information actually measured at the moment is input into the control optimization solver to obtain a control variable sequence, and the control variable sequence is used to output to the actuator of the ball mill for control;

[0033] The judgment module is used to output the control variable sequence to the actuator when it is determined that the output condition is met; when it is determined that the output condition is not met, iteratively call the long short-term memory network prediction module based on the current state information of the ball mill equipment at time tk to predict tk+ 1 The state at the moment is output to the control optimization solver.

[0034] In one embodiment, the LSTM prediction module includes a state encoder and a state decoder, wherein the state encoder is composed of τ LSTM units; the state decoder is composed of 3 fully connected network layers, and the input sequence of the state encoder is [t 1 , tτ] the current state information and historical working information of the ball mill at the time; the input sequence of the state decoder is the encoding sequence of the state encoder.

[0035] Compared with the prior art, the advantages of the present invention are: based on the improved LSTM MPC prediction model, the prediction accuracy and generalization ability are improved by optimizing the model structure and parameter settings. A control optimization solver is designed, and a near-optimal control effect is obtained through approximate iterative training of the Adam optimization algorithm and reasonable control quantity constraints. In addition, combined with the actual operating characteristics of the ball mill, a complete control system architecture of data acquisition, predictive control, judgment and optimization solution is constructed, which improves the precise regulation of the operating state of the ball mill. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0037] Figure 1 is a flow chart of a data-driven long short-term memory network model predictive control method in an embodiment of the present invention;

[0038] Figure 2 It is a structural diagram of a predictive control method based on a data-driven long short-term memory network model in an embodiment of the present invention;

[0039] Figure 3 is a sequence prediction graph based on a long short-term memory network in an embodiment of the present invention;

[0040] Figure 4 A sequence prediction based on a long short-term memory network in an embodiment of the present invention - a long short-term memory network training sample generation diagram;

[0041] Figure 5 It is a structural diagram of a long short-term memory network auxiliary prediction module in an embodiment of the present invention;

[0042] Figure 6 It is a structural block diagram of a data-driven long short-term memory network model predictive control device in an embodiment of the present invention;

[0043] Figure 7 The present invention is a flow chart of a data-driven long short-term memory network model predictive control method in an embodiment of the present invention. DETAILED DESCRIPTION

[0044] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0045] The following describes the implementation methods of the present application through specific examples, and those skilled in the art can easily understand other advantages and effects of the present application from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. The present application can also be implemented or applied through other different specific implementation methods, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present application. It should be noted that, in the absence of conflict, the following embodiments and the features in the embodiments can be combined with each other. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work belong to the scope of protection of the present application.

[0046] It should be noted that various aspects of the embodiments within the scope of protection of the present invention are described below. It should be apparent that the aspects described herein can be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present application, it should be understood by those skilled in the art that an aspect described herein can be implemented independently of any other aspect, and two or more of these aspects can be combined in various ways. For example, any number and aspect described herein can be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein can be used to implement this device and / or practice this method.

[0047] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present application. The drawings only show components related to the present application rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0048] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0049] As a special type of recurrent neural network (RNN), the long short-term memory network (LSTM) has a strong ability to process time series and can capture long-term dependencies in data, so it is very suitable for model building in predictive control. Model predictive control (MPC) is an advanced control strategy formed in engineering applications. Model predictive control has three basic characteristics: prediction model, rolling optimization and rolling implementation control, and feedback correction.

[0050] The function of the prediction model is to predict the future output of the target value based on the historical information and current state of the controlled object. The prediction model has many forms, and the accuracy of the model directly affects the actual effect of the model predictive control.

[0051] Model predictive control abandons the static global optimization goal and instead adopts a dynamic rolling optimization strategy that is usually limited to a limited time range. Specifically, at each sampling moment, the optimization goal only focuses on the performance within a limited period of time in the future; and when entering the next sampling moment, this optimization time window will move forward synchronously to achieve a continuous online optimization process, the so-called "rolling optimization". In addition, the rolling control effect means that the future control instruction sequence obtained by optimization is not all executed on the controlled object, only the first few items of the sequence actually act on the object, which constitutes the fundamental characteristic of predictive control that distinguishes it from other control schemes. This strategy can flexibly respond to uncertain factors such as model time-varying and external interference, and achieve instant adjustment and compensation.

[0052] Model predictive control uses negative feedback to form a closed loop and improve the performance of optimal predictive control.

[0053] Therefore, model predictive control has advantages in dealing with complex system dynamics and multiple constraints, and ball mill operation control happens to have these characteristics. Therefore, model predictive control has good application value in the field of ball mill control.

[0054] The design idea of ​​this application is: the complex generation environment causes the ball mill prediction control optimization effect to fluctuate violently, which often requires the intervention of technicians and can no longer meet the needs of advanced control optimization of the ball mill. Therefore, this application designs an improved long short-term memory network prediction module to realize the ball mill power y at the next moment according to the current and historical status of the ball mill input. 1 and cyclone overflow fineness y 2 Make accurate predictions.

[0055] Based on the above ideas, the present application embodiment provides a data-driven long short-term memory network model predictive control method, such as Figure 1 As shown, the following steps are included:

[0056] Step 101: k-1 The current state information and historical working information of the ball mill at the moment are input into the long short-term memory network model, and t is predicted. k The expected state operation value at the moment, the historical working information includes at least one of the feed rate, water supply before the ball mill, water replenishment in the pump pool, cyclone overflow flow, and frequency of the east-stage pump, and the expected state operation value includes at least the ball mill power and the cyclone overflow fineness.

[0057] The long short-term memory network prediction module is used to analyze the current state information of the input ball mill equipment and historical working information (such as feed volume, pre-mill water supply, pump pool water replenishment, cyclone overflow flow, east-stage pump frequency, etc.), predict the two key target variables of ball mill power and cyclone overflow fineness, and pass the prediction results to the control optimization solver. In order to improve the prediction accuracy, the LSTM model is improved, including optimizing the model structure, selecting the appropriate activation function, adjusting the parameter settings, etc. The historical working information includes the feed volume u 1 , water supply before ball mill 2 , pump pool to add water 3 , cyclone overflow flow u 4 , East grading pump frequency u 5 , ball mill power 1 , cyclone overflow fineness y 2 and any combination of deformation n.

[0058] For example, the training samples used to train the LSTM prediction module can be: 1 , water supply before ball mill 2 , pump pool to add water 3 , cyclone overflow flow u 4 , East grading pump frequency u 5 , ball mill power 1 and cyclone overflow fineness y 2 The input state of the LSTM prediction module is: the amount of ore u 1 , water supply before ball mill 2 , pump pool to add water 3 , cyclone overflow flow u 4 , East grading pump frequency u 5 , ball mill power 1 and cyclone overflow fineness y 2 From the initial moment k 0 Start, collect n I samples with a feature dimension of 5 as input, n o The feature dimension is 2 as the target output, and the sliding window method is used to collect training samples until the remaining data no longer meets the window size.

[0059] Step 102: t k The expected state calculation value at time t k-1 The current state information actually measured at each moment is input into the control optimization solver to obtain a control variable sequence, which is used to output to the actuator of the ball mill for control.

[0060] The control optimization solver generates the control variable sequence of the actuator action according to the actual state and expected state of the ball mill and the boundary constraints. These control variable sequences are output to the actuator to implement control optimization. In order to obtain better control effect, the iterative training method is used to train the control optimization solver, and reasonable control quantity constraints and loss functions are set.

[0061] Step 103: When it is determined that the output condition is met, the control variable sequence is output to the actuator; when it is determined that the output condition is not met, based on t k The current state information of the ball mill equipment at time t is iteratively called by the long short-term memory network prediction module to predict t k+1 The state at that moment is output to the control optimization solver.

[0062] The judgment module can first judge whether the state at the next moment meets the set requirements. If the set requirements are met, the control optimization solver outputs the current control sequence to the ball mill actuator; if the set requirements are not met, the long short-term memory network prediction module is called to predict the state at the next moment, and output it to the control optimization solver for further optimization. In another embodiment, the judgment module also judges whether the long short-term memory network prediction module is from t k-1 Whether the number of calls from the moment reaches the set upper limit. When it is determined that the state at the next moment does not meet the set requirements, but the number of calls reaches the set upper limit, the control optimization solver will also output the current control sequence to the actuator of the ball mill to avoid an infinite loop.

[0063] The above device is based on the improved LSTM MPC prediction model, and improves the prediction accuracy and generalization ability by optimizing the model structure and parameter settings. A control optimization solver is designed, and the approximate optimal control effect is obtained through the Adam optimization algorithm approximate iterative training and reasonable control quantity constraints. In addition, combined with the actual operating characteristics of the ball mill, a complete control system architecture of data acquisition, predictive control, judgment and optimization solution is constructed, which improves the precise control of the ball mill operating status.

[0064] In one embodiment, when determining t k The expected state calculation value at time t k When the difference between the current state information actually measured at each moment is within a preset range, the control variable sequence is output to the actuator;

[0065] When determining t k The expected state calculation value at time t k When the difference between the current state information actually measured at the moment exceeds the preset range, the long short-term memory network prediction module is obtained from t k-1 The number of calls since the moment;

[0066] When it is determined that the number of calls is not less than the preset upper limit, the control variable sequence is output to the actuator; when it is determined that the number of calls is less than the preset upper limit, based on t k The current state information of the ball mill equipment at time t is used again to call the long short-term memory network prediction module to predict t k+1 The state at that moment is output to the control optimization solver.

[0067] In one embodiment, the long short-term memory network model adopts a state encoder-state decoder structure. The long short-term memory network model includes at least one state encoder and a state decoder. The state encoder processes the input sequence and generates a coding sequence. The input sequence includes t k-1 The current state information and historical working information of the ball mill at each moment, the encoding sequence summarizes the information in the input sequence; the state decoder uses the encoding sequence to generate a control variable sequence.

[0068] In one embodiment, the state encoder is composed of τ long short-term memory units; the state decoder is composed of 3 layers of fully connected network layers, and the input sequence of the state encoder is [t 1 ,t τ ] The current state information and historical working information of the ball mill at the current moment; the input sequence of the state decoder is the encoding sequence of the state encoder.

[0069] The long short-term memory network model is used to realize the ball milling power y of the ball mill at the next moment according to the current and historical status of the ball mill input. 1 and cyclone overflow fineness y 2 Make accurate predictions. Build a prediction strategy based on LSTM, determine the input and output parameters of the network, and complete the learning and training of the network. Input parameter selection of the long short-term memory network model: feed volume u 1 , water supply before ball mill 2 , pump pool to add water 3 , cyclone overflow flow u 4 , East grading pump frequency u 5 ; Output target parameter: ball milling power y 1 and cyclone overflow fineness y 2 .

[0070] Long Short-Term Memory (LSTM) is a variant of Recurrent Neural Network (RNN). It can learn long-term dependencies in data and has a very good effect in dealing with time series analysis problems. The network diagram is shown below. Figure 3 As shown in the figure, the LSTM introduces a forget gate, an input gate, and an output gate, which can solve the problem of gradient disappearance and gradient explosion to a certain extent.

[0071] This embodiment uses an encoder-decoder structure to use LSTM for sequence-to-sequence prediction. The LSTM encoder-decoder consists of 1 LSTM and a fully connected neural network. The first LSTM (encoder) processes the input sequence and generates an encoding state that summarizes the information in the input sequence; the second fully connected neural network (decoder) uses the encoding state to generate the output sequence. The input and output sequences can have different lengths.

[0072] The LSTM-based sequence prediction scheme mainly includes training sample generation, LSTM encoder construction, and fully connected neural network decoder training.

[0073] (1) Training sample generation

[0074] In order to train the LSTM encoder-decoder, the time series needs to be divided into several input value sequences of specified lengths and corresponding target value sequences. The present invention generates training samples that meet the above requirements through a time sliding window. The specific steps are as follows Figure 4 shown.

[0075] like Figure 4 As shown, starting from the initial time k, collect n I input values ​​as input, n o The output value is taken as the target. The sliding window method is adopted to slide the window to the second (step length step = 1) or third (step length step = 2) moment to continue collecting until the remaining data no longer meets the window size.

[0076] After the above data collection process, the input set and target set required for training are organized into a three-dimensional matrix. Output Set Among them, the first dimension is the length of the input / output data; the second dimension is the number of samples generated by the sliding window method; the third dimension is the number of input / output features.

[0077] (2) LSTM encoder-decoder construction

[0078] like Figure 5 As shown in Figure 1, the deep learning model compensation module consists of a state encoder and a state decoder. The state encoder consists of r LSIM units, and the state decoder consists of a fully connected network layer. The state encoder is [t 1 ,t τ ] as input, and use several LSTM units to generate the encoded state; the state decoder receives the encoded state transmitted by the state encoder, and takes the ball mill state x at the current moment as input, and generates the encoded state for t (τ+1)Predict the state at a given moment.

[0079] (3) LSTM encoder-decoder training

[0080] After the LSTM decoder generates an output sequence of the required length, the predicted value is compared with the actual target value, and the error mean square error is used as the loss function for network training. Backpropagation is then used to adjust the weights of the LSTM gates to minimize the loss function. After the LSTM encoder-decoder training is completed, its prediction accuracy can be tested using the data in the test set. The prediction of all test data is performed using a recursive method.

[0081] In one embodiment, the LSTM prediction module uses an improved LSTM algorithm for prediction. The LSTM network consists of an input layer, a hidden layer, and an output layer. The input layer receives the current state and historical information of the ball mill; the hidden layer is composed of multiple LSTM units to capture long-term dependencies in the data; and the output layer outputs the prediction results.

[0082] In order to improve the prediction accuracy, this embodiment makes the following improvements to the LSTM model:

[0083] (1) Optimizing the model structure: According to the actual operating characteristics of the ball mill, we selected 1 LSTM layer, 64 units, and the activation function (ReLU).

[0084] (2) Adjust parameter settings: The optimal learning rate, batch size, number of training rounds and other parameters were selected through random search and other methods.

[0085] (3) Introducing regularization technology: In order to prevent overfitting, L2 regularization technology is introduced to constrain model parameters.

[0086] In one embodiment, the control optimization solver is obtained by iterative training, wherein the control quantity constraint during the training process is

[0087] u min ≤u(k)≤u max

[0088] Δu min ≤Δu(k)≤Δu max

[0089] Δu(k+j)=u(k+j)-u(k+j-1)

[0090] in, is the virtual control variable to be solved in model predictive control, Δu min and Δu max are the minimum and maximum values ​​allowed for the control sequence variation, u minand u max They are the minimum and maximum values ​​of the range of variation allowed by the control sequence.

[0091] In one embodiment, the control optimization solver is obtained by iterative training, where the loss function during the training process is Among them, y m (k+1)=f(u(k), u(k-1), u(k-2), ..., y(k), y(k-1)), y(k-2), ...), except for u(k) at time k, which is an unknown quantity, all other quantities before time k are known. y(k) is the actual reading of the instrument, y r (k+1) is a reference value, which is a known quantity.

[0092] Δu(k+j)=u(k+j)-u(k+j-1)

[0093] Δy m (k+j)=y m (k+j)-y m (k+j-1)

[0094]

[0095] y m (k)=f(u(k-1),u(k-2),...,u(kn u +1),y(k-1)),y(k-2),...,y(kn y +1)), n u 、n y are the delay orders of the input control variable time series and the controlled object output time series respectively.

[0096] The control optimization solver is a model predictive control (MPC) optimization solver, which is used to generate an actuator action control sequence based on the actual state and expected state of the input ball mill, and output it to the LSTM prediction module. The MPC algorithm is a prior art. In this embodiment, based on the existing MPC algorithm, the objective function and constraints of the MPC algorithm are designed according to the operation characteristics of the ball mill, so that it is suitable for ball mill control optimization. The input of the MPC algorithm is: the actual state and expected state of the ball mill at the current moment; the output is: the action sequence of the actuator of the ball mill at the current moment.

[0097] In one embodiment, the control optimization solver uses Ada mThe optimization algorithm is trained iteratively for approximate optimization. During the training process, reasonable control quantity constraints and loss functions are set. The control quantity constraints include the range and rate of change of the control sequence to ensure the rationality and feasibility of the control quantity. The loss function is calculated based on the error between the predicted result and the actual result to guide the optimization direction of the model.

[0098] Among them, the loss function Loss in the training process is

[0099]

[0100] Among them, y m (k+1)=f(u(k), u(k-1), u(k-2), ..., y(k), y(k-1)), y(k-2), ...), except for u(k) at time k, which is an unknown quantity, all other quantities before time k are known. y(k) is the actual reading of the instrument, y r (k+1) is a reference value and is a known quantity.

[0101] Δu(k+j)=u(k+j)-u(k+j-1)

[0102] Δy m (k+j)=y m (k+j)-y m (k+j-1)

[0103]

[0104] y m (k)=f(u(k-1),u(k-2),...,u(kn u +1),y(k-1)),y(k-2),...,y(kn y +1)), n u 、n y are the delay orders of the input control variable time series and the controlled object output time series respectively.

[0105] Specifically, a training process of an approximately feasible control optimization solver is provided as follows:

[0106] Step 1), initialize model parameters and control sequence;

[0107] Step 2), calculating the prediction result according to the current control sequence and the actual state of the ball mill;

[0108] Step 3), calculate the loss function value according to the error between the predicted result and the actual result;

[0109] Step 4), update the model parameters and control sequence through the back propagation algorithm of the Adam optimization algorithm;

[0110] Step 5), check the boundary conditions, if exceeded, limit them to the constraint range;

[0111] Step 6), repeat steps 2) to 5) until the set number of training rounds is reached or the loss function value converges.

[0112] like Figure 2 and Figure 6 A data-driven long short-term memory network model predictive control device is shown, which includes a data acquisition module 601, a long short-term memory network prediction module 602, a control optimization solver 603 and a judgment module 604.

[0113] The data acquisition module 601 is used to collect the real-time data of the corresponding sensors of the ball mill and store it in correspondence with the current moment. The data acquisition module 601 is responsible for collecting the real-time data of the corresponding sensors of the ball mill and storing it in the database. The frequency and accuracy of data acquisition can be adjusted according to actual needs. At the same time, the module also combines historical process data and historical information, and comprehensively transmits them to the LSTM prediction module and the control optimization solver for processing and utilization.

[0114] The long short-term memory network prediction module 602 is used to convert t k-1 The current state information and historical working information of the ball mill at the moment are input into the long short-term memory network model, and t is predicted. k The expected state operation value at the moment, the historical working information includes at least one of the feed rate, water supply before the ball mill, water replenishment in the pump pool, cyclone overflow flow, and frequency of the east-stage pump, and the expected state operation value includes at least the ball mill power and the cyclone overflow fineness.

[0115] The control optimization solver 603 is used to convert t k The expected state calculation value at time t k-1 The current state information actually measured at each moment is input into the control optimization solver to obtain a control variable sequence, which is used to output to the actuator of the ball mill for control.

[0116] The judgment module 604 is used to output the control variable sequence to the actuator when it is determined that the output condition is met; when it is determined that the output condition is not met, based on t k The current state information of the ball mill equipment at time t is iteratively called by the long short-term memory network prediction module to predict t k+1The state at the moment is output to the control optimization solver. The judgment module 604 is used to judge whether the state at the next moment meets the set requirements. Specifically, when the state at the next moment meets the set requirements, for example, the error between the predicted result and the actual result is less than the set threshold, the control optimization solver outputs the current control sequence to the ball mill actuator; when the state at the next moment does not meet the set requirements, the improved LSTM prediction module is called to predict the state at the next moment, and output it to the control optimization solver for further optimization. In addition, when the number of cycles reaches the set upper limit, for example, the maximum number of iterations or the time limit is reached, the control optimization solver will also output the current control sequence to the actuator of the ball mill. When the judgment module determines that the state at the next moment does not meet the set requirements, but the number of cycles reaches the set upper limit, the control optimization solver 603 outputs the current control sequence to the actuator of the ball mill.

[0117] In one embodiment, the LSTM prediction module includes a state encoder and a state decoder.

[0118] The state encoder consists of τ long short-term memory units; the state decoder consists of 3 fully connected network layers.

[0119] The input sequence of the state encoder is [t 1 ,t τ ] The current state information and historical working information of the ball mill at the current moment; the input sequence of the state decoder is the encoding sequence of the state encoder.

[0120] like Figure 7 As shown, the above-mentioned data-driven long short-term memory (LSTM) network model predictive control device takes one control cycle as an example, and the specific control process is as follows:

[0121] Step 0: ball mill initialization, data acquisition module 601 acquires real-time data of corresponding sensors of the ball mill, and stores it in correspondence with the current time.

[0122] Step 1: The long short-term memory network prediction module predicts the ball mill power and cyclone overflow fineness at time t_k according to the ball mill status input at time t(k-1) (such as ore feed, water supply before ball milling, water replenishment in the pump pool, cyclone overflow flow, east-stage pump frequency, etc.), and transmits the prediction results to the model predictive control optimization solver.

[0123] Step 2: The model predictive control optimization solver solves the first step control sequence of the actuator action according to the input actual state and expected state of the ball mill (such as the set ball mill power and cyclone overflow fineness target value) and boundary constraints (such as the range and rate of change of the control quantity, etc.), and outputs the control sequence to the ball mill actuator to generate control action.

[0124] Step 3: The data acquisition module collects the real-time data of the corresponding sensors of the ball mill at time t_k (such as actual ball mill power and cyclone overflow fineness, etc.) and stores it in the database. Combined with historical process data and historical information, it is comprehensively transmitted to the long short-term memory network prediction module and the control optimization solver for processing and utilization.

[0125] Step 4: Repeat steps 2 and 3 to perform iterative optimization (also known as feedback correction). In each iteration, the long short-term memory network prediction module will make predictions based on the latest ball mill state data and pass the prediction results to the control optimization solver for optimization solution.

[0126] Step 5: The judgment module makes judgments and decisions based on the error between the predicted result and the actual result and the number of cycles to ensure the stability and convergence of the control system. When it is determined that the control accuracy is met, the control optimization solver outputs the current control sequence to the actuator of the ball mill.

[0127] The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by any technician familiar with the technical field within the technical scope disclosed in the present application should be included in the protection scope of the present application.

Claims

1. A data-driven long short-term memory network model predictive control method, characterized in that: include: t k-1 The current state information and historical working information of the ball mill at the moment are input into the long short-term memory network model, and t is predicted. k The expected state calculation value at the time, the historical working information includes at least one of the feed rate, the water supply before the ball mill, the water added to the pump pool, the cyclone overflow flow, and the frequency of the east-stage pump, and the expected state calculation value includes at least the ball mill power and the cyclone overflow fineness; The t k The expected state calculation value at time t k-1 The current state information actually measured at the moment is input into the control optimization solver to obtain a control variable sequence, and the control variable sequence is used to output to the actuator of the ball mill for control; When it is determined that the output condition is met, the control variable sequence is output to the actuator; when it is determined that the output condition is not met, based on t k The current state information of the ball mill equipment at time t is iteratively called to predict the long short-term memory network prediction module k+1 The state at the moment is output to the control optimization solver.

2. The long short-term memory network model predictive control method according to claim 1, characterized in that: When judging t k The expected state calculation value at time t k When the difference between the current state information actually measured at the moment is within a preset range, the control variable sequence is output to the actuator; When judging t k The expected state calculation value at time t k When the difference between the current state information actually measured at the time exceeds the preset range, the long short-term memory network prediction module is obtained from t k-1 The number of calls since the moment; When it is determined that the number of calls is not less than the preset upper limit, the control variable sequence is output to the actuator; when it is determined that the number of calls is less than the preset upper limit, based on t k The current state information of the ball mill equipment at time t is used again to call the long short-term memory network prediction module to predict t k+1 The state at the moment is output to the control optimization solver.

3. The long short-term memory network model predictive control method according to claim 1, characterized in that: The long short-term memory network model adopts a state encoder-state decoder structure, and the long short-term memory network model includes at least one state encoder and a state decoder. The state encoder processes an input sequence and generates an encoded sequence, wherein the input sequence includes the t k-1 The current state information and historical working information of the ball mill at each moment, wherein the coding sequence summarizes the information in the input sequence; The state decoder generates the sequence of control variables using the encoding sequence.

4. The long short-term memory network model predictive control method according to claim 3, characterized in that: The state encoder is composed of τ long short-term memory units; the state decoder is composed of 3 fully connected network layers. The input sequence of the state encoder is [t1,t τ ] the current state information and historical working information of the ball mill at the current moment; the input sequence of the state decoder is the encoding sequence of the state encoder.

5. The long short-term memory network model predictive control method according to claim 1, characterized in that: The control optimization solver is obtained through iterative training, where the control quantity constraint during the training process is u min ≤u(k)≤u max Δu min ≤Δu(k)≤Δu max Δu(k+j)=u(k+j)-u(k+j-1) in, is the virtual control variable to be solved in model predictive control, Δu min and Δu max are the minimum and maximum values ​​allowed for the control sequence variation, u min and u max They are the minimum and maximum values ​​of the range of variation allowed by the control sequence.

6. The long short-term memory network model predictive control method according to claim 1, characterized in that: The control optimization solver is obtained through iterative training, where the loss function in the training process is Among them, y m (k+1)=f(u(k),u(k-1),u(k-2),…,y(k),y(k-1)),y(k-2),…) Except for u(k) at time k, which is an unknown quantity, all other quantities before time k are known. y(k) is the actual reading of the instrument, y r (k+1) is a reference value, which is a known quantity. Δu(k+j)=u(k+j)-u(k+j-1) Δy m (k+j)=y m (k+j)-y m (k+j-1) y m (k)=f(u(k-1),u(k-2),...,u(kn u +1), (k-1)), y(k-2),…, y(kn y +1)), n u 、n y are the delay orders of the input control variable time series and the controlled object output time series respectively.

7. A data-driven long short-term memory network model predictive control device, characterized in that: The device comprises: The data acquisition module is used to collect the real-time data of the corresponding sensors of the ball mill and store them in correspondence with the current moment; The long short-term memory network prediction module is used to convert t k-1 The current state information and historical working information of the ball mill at the moment are input into the long short-term memory network model, and t is predicted. k The expected state calculation value at the time, the historical working information includes at least one of the feed rate, the water supply before the ball mill, the water added to the pump pool, the cyclone overflow flow, and the frequency of the east-stage pump, and the expected state calculation value includes at least the ball mill power and the cyclone overflow fineness; Control optimization solver, used to convert the t k The expected state calculation value at time t k-1 The current state information actually measured at the moment is input into the control optimization solver to obtain a control variable sequence, and the control variable sequence is used to output to the actuator of the ball mill for control; The judgment module is used to output the control variable sequence to the actuator when it is determined that the output condition is met; when it is determined that the output condition is not met, based on t k The current state information of the ball mill equipment at time t is iteratively called to predict the long short-term memory network prediction module k+1 The state at the moment is output to the control optimization solver.

8. The long short-term memory network model predictive control device according to claim 7, characterized in that: The long short-term memory network prediction module includes a state encoder and a state decoder, The state encoder is composed of τ long short-term memory units; the state decoder is composed of 3 fully connected network layers. The input sequence of the state encoder is [t1,t τ ] the current state information and historical working information of the ball mill at the current moment; the input sequence of the state decoder is the encoding sequence of the state encoder.

Images