A Coke Quality Prediction Method Combining Production Dynamic Data

Abstract

The present invention relates to coke quality prediction, and particularly to a coke quality prediction method combining production dynamic data, constructing a deep neural network DNN model, and using the deep neural network DNN model to predict coke quality based on coal blending data so as to realize the quality prediction from coal to coke; constructing a dynamic data prediction model based on the deep Q network DQN, and simultaneously constructing a static data prediction model. In the initial stage, the static data prediction model uses the KNN algorithm to perform prediction based on static data to obtain a static prediction result; as time goes by and production dynamic data accumulates, more production dynamic data is introduced to perform prediction using the dynamic data prediction model to obtain a dynamic prediction result so as to realize real-time quality prediction during the coke oven production process; the technical solution provided by the present invention can effectively overcome the defect that the prior art cannot accurately predict coke quality.

Description

Technical Field

[0001] The present invention relates to coke quality prediction, and specifically to a coke quality prediction method combining production dynamic data. Background Art

[0002] Currently, the world's iron production mainly comes from blast furnace ironmaking. Compared with existing non-blast furnace ironmaking technologies, blast furnace ironmaking has the characteristics of good economic indicators, simple process, large production volume, and low energy consumption. It is predicted that it will still be the main ironmaking process in the future for some time.

[0003] Coke is an important raw fuel for blast furnace ironmaking, and it plays important roles such as heat source, reducing agent, burden column skeleton, and carburization of hot metal. Coke quality has an important impact on blast furnace coke ratio, production efficiency, hot metal quality, and economic benefits. With the development of large-scale blast furnaces and coal injection technology, the coal ratio has been continuously increased, and the load of coke in the blast furnace has increased. Therefore, the requirements for coke quality have become increasingly strict. Therefore, effectively predicting coke quality is of great significance for optimizing the coal blending scheme.

[0004] At present, a variety of coke quality prediction models have been established based on the basic properties of raw coal, mainly using the proximate analysis, ultimate analysis, and maceral components of raw coal as basic information to construct the linear relationship between coke quality and the basic properties of raw coal. However, the existing coke quality prediction models ignore the coking process, which is a complex production process with characteristics such as multi-parameters, time-varying, non-linear, and uncertain, resulting in the inability of the existing coke quality prediction models to accurately predict coke quality and making it difficult to effectively guide enterprises to construct the coal blending scheme. Summary of the Invention

[0005] (1) Technical Problems to be Solved

[0006] Aiming at the above-mentioned disadvantages of the existing technology, the present invention provides a coke quality prediction method combining production dynamic data, which can effectively overcome the defect that the existing technology cannot accurately predict coke quality.

[0007] (2) Technical Solutions

[0008] To achieve the above object, the present invention is realized through the following technical solutions:

[0009] A coke quality prediction method combining production dynamic data constructs a deep neural network DNN model, and uses the deep neural network DNN model to predict coke quality based on coal blending data to achieve the quality prediction from coal to coke;

[0010] A dynamic data prediction model is constructed based on the Deep Q-Network (DQN), and a static data prediction model is also constructed. In the initial stage, the static data prediction model uses the KNN algorithm to make predictions based on static data, obtaining static prediction results. As time goes by and production dynamic data accumulates, more production dynamic data is introduced and the dynamic data prediction model is used for prediction to obtain dynamic prediction results, so as to achieve real-time quality prediction of the coke oven production process.

[0011] Preferably, the Deep Neural Network (DNN) model includes an input layer, a hidden layer, and an output layer;

[0012] The input layer receives the input coal blending data, and each input node corresponds to a feature;

[0013] The hidden layer processes information and extracts features through multiple hidden layers. Each hidden layer consists of multiple neuron nodes, and each neuron node is connected to all neuron nodes in the previous layer. The input coal blending data is weighted and summed using weights and activation functions and the result is output;

[0014] The output layer outputs coke quality parameters based on the output results of the hidden layer;

[0015] Among them, the coal blending data includes assay information such as the ash content Ad value, sulfur content St.d value, volatile matter, G value, and Y value of each coal type, as well as the blending ratio of each coal type;

[0016] The coke quality parameters include the ash content Ad value, sulfur content St.d value, shatter strength M25, abrasion resistance M10, coke reactivity index CRI, and post-reaction strength CSR.

[0017] Preferably, the model training process of the Deep Neural Network (DNN) model includes:

[0018] S1. Set the first loss function and the first optimizer of the Deep Neural Network (DNN) model, and input the training set into the Deep Neural Network (DNN) model for model training;

[0019] S2. Calculate the loss value based on the first loss function, and the first optimizer updates the model parameters according to the loss value and network gradient information;

[0020] S3. If the loss value is less than the first preset threshold, the model training ends, and the current Deep Neural Network (DNN) model is the trained Deep Neural Network (DNN) model; otherwise, return to S1 and continue to use the training set for model training.

[0021] Preferably, when the Deep Neural Network (DNN) model is trained on the training set, the generalization ability of the model is evaluated by observing the performance on the validation set, and the hyperparameters and structure of the model are tuned;

[0022] After obtaining the trained deep neural network DNN model, use the test set to evaluate the performance of the model.

[0023] Preferably, the dynamic data prediction model constructed based on the deep Q-network DQN includes:

[0024] Model the coke oven production process into a Markov decision process MDP, define the state space, action space, and reward function, and guide the deep Q-network DQN to learn the optimal prediction strategy through the reward function;

[0025] Among them, the state space includes static data and production dynamic data. The static data includes the ash content Ad value and sulfur content St.d value of each coal type; the production dynamic data includes the furnace temperature of the coke oven, coking time, charging time, coke pushing plan, production maintenance, and the production data of the dry quenching furnace. The production data of the dry quenching furnace includes the material level, steam index, coke discharging load, coke discharging temperature, and steam heat exchange efficiency;

[0026] The action space includes dynamic prediction results and decision-making suggestions.

[0027] Preferably, the deep Q-network DQN learns the mapping relationship between states and actions through model training, takes the state as input, and uses the Q-value function to output the expected value of the cumulative reward that can be obtained by taking each possible action in this state, that is, the Q-value;

[0028] The deep Q-network DQN includes two target networks. One target network is the behavior network for selecting actions, and the other target network is the target Q-value network for calculating the target Q-value. The network parameters of the target Q-value network are copied from the behavior network every once in a while.

[0029] Preferably, the deep Q-network DQN uses a neural network to approximate the Q-value function, and the update process of the Q-value is represented by the following formula:

[0030]

[0031] Where s t is the state at time t, a t is the action taken at time t, Q(s t , a t ) is the Q-value of taking action a t in state s t at time t, Q’(s t , a t ) is the updated Q-value of taking action a t in state s t at time t, α is the learning rate used to represent the speed at which new information covers old information, and r is the state s at time t tTake action a next t The immediate reward obtained, where δ is the discount factor used to weigh the importance of future rewards, and α, δ ∈ [0, 1], is the state s at time t+1 t+1 is the maximum value among the Q-values of all possible actions a taken next.

[0032] Preferably, the deep Q-network DQN implements an experience replay mechanism during model training, stores the experience data of interacting with the environment in the experience replay buffer, and randomly samples from the experience replay buffer for model training to reduce the correlation between data. At the same time, the deep Q-network DQN continuously updates the network parameters using the interaction with the environment and gradually learns the optimal prediction strategy.

[0033] Preferably, the model training process of the deep Q-network DQN includes:

[0034] S1. Set the second loss function and the second optimizer of the deep Q-network DQN, and input the training set into the deep Q-network DQN for model training;

[0035] S2. Calculate the loss value based on the second loss function, and the second optimizer updates the model parameters according to the loss value and the network gradient information;

[0036] S3. If the loss value is less than the second preset threshold, the model training ends, and the current deep Q-network DQN is the trained deep Q-network DQN; otherwise, return to S1 and continue to use the training set for model training;

[0037] Among them, the second loss function uses the mean square error loss function to calculate the difference between the Q-value output by the deep Q-network DQN and the target Q-value.

[0038] Preferably, when the deep Q-network DQN conducts model training on the training set, it evaluates the generalization ability of the model by observing the performance on the validation set and tunes the hyperparameters and structure of the model;

[0039] After obtaining the trained deep Q-network DQN, the performance of the model is evaluated using the test set.

[0040] (III) Beneficial effects

[0041] Compared with the prior art, the coke quality prediction method combining production dynamic data provided by the present invention has the following beneficial effects:

[0042] 1) Construct a deep neural network DNN model. Using this efficient prediction model can accurately predict the coke quality under different coal blending schemes, helping enterprises to anticipate the changes in coke quality in advance so as to take corresponding measures to ensure that the coke quality meets the requirements;

[0043] 2) In the initial stage, the static data prediction model uses the KNN algorithm to make predictions based on static data, obtaining static prediction results. As time goes by and production dynamic data accumulates, more production dynamic data is introduced and the dynamic data prediction model is used for prediction, obtaining dynamic prediction results. As the system gradually transitions to the dynamic data prediction model, it can perform real-time prediction on the coke oven production process. By introducing production dynamic data, the dynamic prediction results can more accurately reflect the actual situation of the coke oven production process, improve the accuracy of coke quality prediction, provide effective guidance for the enterprise to construct a coal blending plan, and ensure the stability of product quality. Description of the Drawings

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0045] Figure 1 It is a flow diagram of the present invention. Detailed Embodiments

[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts fall within the scope of protection of the present invention.

[0047] A method for predicting coke quality by combining production dynamic data, as Figure 1 shown, constructs a deep neural network DNN model, and uses the deep neural network DNN model to predict coke quality based on coal blending data to achieve the quality prediction from coal to coke.

[0048] ① The deep neural network DNN model includes an input layer, a hidden layer, and an output layer;

[0049] The input layer receives the input coal blending data, and each input node corresponds to a feature;

[0050] The hidden layer performs information processing and feature extraction through multiple hidden layers. Each hidden layer is composed of multiple neuron nodes, and each neuron node is connected to all neuron nodes in the previous layer. The input coal blending data is weighted and summed using weights and activation functions and the result is output;

[0051] An output layer that outputs coke quality parameters based on the output results of the hidden layer;

[0052] Among them, the coal blending data includes the assay information such as the ash content Ad value, sulfur content St.d value, volatile matter, G value, and Y value of each coal type, as well as the blending ratio of each coal type;

[0053] The coke quality parameters include the ash content Ad value, sulfur content St.d value, shatter strength M25, abrasion resistance M10, coke reactivity index CRI, and post-reaction strength CSR.

[0054] ② The model training process of the deep neural network DNN model includes:

[0055] S1. Set the first loss function and the first optimizer of the deep neural network DNN model, and input the training set into the deep neural network DNN model for model training;

[0056] S2. Calculate the loss value based on the first loss function, and the first optimizer updates the model parameters according to the loss value and the network gradient information;

[0057] S3. If the loss value is less than the first preset threshold, the model training ends, and the current deep neural network DNN model is the trained deep neural network DNN model; otherwise, return to S1 and continue to use the training set for model training.

[0058] Specifically, when the deep neural network DNN model conducts model training on the training set, it evaluates the generalization ability of the model by observing the performance on the validation set, and tunes the hyperparameters and structure of the model;

[0059] After obtaining the trained deep neural network DNN model, use the test set to evaluate the performance of the model.

[0060] In the above technical solution, the backpropagation algorithm is used in the model training process to optimize the model parameters by minimizing the loss function, so that the prediction result of the model is as close as possible to the true value. Finally, the trained deep neural network DNN model is deployed to the actual production environment to realize the quality prediction from coal to coke. The deep neural network DNN model can help optimize the coke production process and improve product quality and production efficiency.

[0061] As Figure 1 shown, a dynamic data prediction model is constructed based on the deep Q-network DQN, and a static data prediction model is also constructed. In the initial stage, the static data prediction model uses the KNN algorithm to make predictions based on static data to obtain static prediction results; as time goes by and production dynamic data accumulates, more production dynamic data is introduced to make predictions using the dynamic data prediction model to obtain dynamic prediction results, so as to realize the real-time quality prediction of the coke oven production process.

[0062] ①Construct a dynamic data prediction model based on the Deep Q-Network (DQN), including:

[0063] Model the coke oven production process as a Markov Decision Process (MDP), define the state space, action space, and reward function, and guide the Deep Q-Network (DQN) to learn the optimal prediction strategy through the reward function;

[0064] Among them, the state space includes static data and production dynamic data. The static data includes the ash content (Ad value) and sulfur content (St.d value) of each coal type; the production dynamic data includes the coke oven temperature, coking time, charging time, coke pushing plan, production maintenance, and the production data of the dry quenching furnace. The production data of the dry quenching furnace includes the material level, steam index, coke discharging load, coke discharging temperature, and steam heat exchange efficiency;

[0065] The action space includes dynamic prediction results and decision-making suggestions.

[0066] The Deep Q-Network (DQN) learns the mapping relationship between states and actions through model training. Taking the state as the input, it uses the Q-value function to output the expected value of the cumulative reward that can be obtained by taking each possible action in this state, that is, the Q-value;

[0067] The Deep Q-Network (DQN) includes two target networks. One target network is the behavior network for selecting actions, and the other target network is the target Q-value network for calculating the target Q-value. The network parameters of the target Q-value network are copied from the behavior network every once in a while.

[0068] The Deep Q-Network (DQN) uses a neural network to approximate the Q-value function. The update process of the Q-value is represented by the following formula:

[0069]

[0070] Among them, s t is the state at time t, a t is the action taken at time t, Q(s t , a t ) is the Q-value of taking action a t in state s t at time t, Q’(s t , a t ) is the updated Q-value of taking action a t in state s t at time t, α is the learning rate used to represent the speed at which new information covers old information, r is the immediate reward obtained by taking action a t in state s t at time t, δ is the discount factor used to weigh the importance of future rewards, α, δ ∈ [0, 1], The state s at time t+1 t+1 The maximum value among the Q-values of all possible actions a taken below.

[0071] ② The model training process of the Deep Q-Network (DQN) includes:

[0072] S1. Set the second loss function and the second optimizer of the Deep Q-Network (DQN), and input the training set into the Deep Q-Network (DQN) for model training;

[0073] S2. Calculate the loss value based on the second loss function, and the second optimizer updates the model parameters according to the loss value and the network gradient information;

[0074] S3. If the loss value is less than the second preset threshold, the model training ends, and the current Deep Q-Network (DQN) is the trained Deep Q-Network (DQN); otherwise, return to S1 and continue to use the training set for model training;

[0075] Among them, the second loss function uses the mean squared error loss function to calculate the difference between the Q-value output by the Deep Q-Network (DQN) and the target Q-value.

[0076] Specifically, the Deep Q-Network (DQN) implements an experience replay mechanism during the model training process (to solve the problems of data correlation and non-stationary distribution), stores the experience data (state transitions, actions, rewards, etc.) of interacting with the environment in the experience replay buffer, and randomly samples from the experience replay buffer for model training to reduce the correlation between data. At the same time, the Deep Q-Network (DQN) continuously updates the network parameters using the interaction with the environment and gradually learns the optimal prediction strategy.

[0077] Specifically, when the Deep Q-Network (DQN) conducts model training on the training set, it evaluates the generalization ability of the model by observing the performance on the validation set, and tunes the hyperparameters and structure of the model;

[0078] After obtaining the trained Deep Q-Network (DQN), the performance of the model is evaluated using the test set.

[0079] In the above technical solution, during the model training stage, historical data is used to train the Deep Q-Network (DQN) to enable it to learn to achieve real-time quality prediction of the coke oven production process based on different static data and production dynamic data. The model receives a series of state information, including static data and production dynamic data, and then outputs dynamic prediction results and decision-making suggestions.

[0080] In the actual prediction stage, the continuously collected real-time production dynamic data is input into the trained Deep Q-Network (DQN) to obtain the real-time quality prediction results under the current coke oven production process. These prediction results can help production personnel make real-time decisions, optimize the coke oven production process, and improve the coke quality and production efficiency.

[0081] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements will not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A coke quality prediction method combining production dynamic data, characterized in that: Build a deep neural network DNN model, and use the deep neural network DNN model to predict the coke quality based on the blended coal data, so as to achieve the quality prediction from coal to coke; Build a dynamic data prediction model based on the deep Q-network DQN, and at the same time build a static data prediction model. In the initial stage, use the static data prediction model to perform predictions based on static data using the KNN algorithm to obtain static prediction results; as time goes by and production dynamic data accumulates, introduce more production dynamic data and use the dynamic data prediction model to perform predictions to obtain dynamic prediction results, so as to achieve real-time quality prediction of the coke oven production process; The building of the dynamic data prediction model based on the deep Q-network DQN includes: Model the coke oven production process into a Markov decision process MDP, define the state space, action space and reward function, and guide the deep Q-network DQN to learn the optimal prediction strategy through the reward function; Among them, the state space includes static data and production dynamic data. The static data includes the ash content Ad value and sulfur content St.d value of each coal type; the production dynamic data includes the furnace temperature, coking time, pushing plan, production maintenance of the coke oven, and the production data of the dry quenching furnace. The production data of the dry quenching furnace includes the material level, steam index, coke discharge load, coke discharge temperature and steam heat exchange efficiency; The action space includes dynamic prediction results and decision-making suggestions; The deep Q-network DQN learns the mapping relationship between states and actions through model training. Taking the state as the input, it uses the Q-value function to output the expected value of the cumulative reward that can be obtained by taking each possible action in this state, that is, the Q-value; The deep Q-network DQN includes two target networks. One target network is the behavior network used to select actions, and the other target network is the target Q-value network used to calculate the target Q-value. The network parameters of the target Q-value network are copied from the behavior network every once in a while; The deep Q-network DQN implements an experience replay mechanism during the model training process. Store the experience data of interacting with the environment in the experience replay buffer, and randomly sample from the experience replay buffer for model training to reduce the correlation between data. At the same time, the deep Q-network DQN continuously updates the network parameters using the interaction with the environment, and gradually learns the optimal prediction strategy.

2. The coke quality prediction method combined with production dynamic data according to claim 1, characterized in that: The deep neural network DNN model includes an input layer, a hidden layer and an output layer; The input layer receives the input blended coal data, and each input node corresponds to a feature; The hidden layer processes information and extracts features through multiple hidden layers. Each hidden layer is composed of multiple neuron nodes. Each neuron node is connected to all neuron nodes in the previous layer, and uses weights and activation functions to perform weighted summation on the input blended coal data and output the result; The output layer outputs the coke quality parameters based on the output result of the hidden layer; Among them, the blended coal data includes assay information such as the ash content Ad value, sulfur content St.d value, volatile matter, G value and Y value of each coal type, as well as the blending ratio of each coal type; The coke quality parameters include ash content Ad value, sulfur content St.d value, shatter strength M25, abrasion resistance strength M10, coke reactivity index CRI, and coke strength after reaction CSR.

3. The coke quality prediction method combining production dynamic data according to claim 2, characterized in that: The model training process of the deep neural network DNN model includes: S1. Set the first loss function and the first optimizer of the deep neural network DNN model, and input the training set into the deep neural network DNN model for model training; S2. Calculate the loss value based on the first loss function, and the first optimizer updates the model parameters according to the loss value and the network gradient information; S3. If the loss value is less than the first preset threshold, the model training ends, and the current deep neural network DNN model is the trained deep neural network DNN model; otherwise, return to S1 and continue to use the training set for model training.

4. The coke quality prediction method combining production dynamic data according to claim 3, characterized in that: When the deep neural network DNN model is trained on the training set, the generalization ability of the model is evaluated by observing the performance on the validation set, and the hyperparameters and structure of the model are tuned; After obtaining the trained deep neural network DNN model, the performance of the model is evaluated using the test set.

5. The coke quality prediction method combining production dynamic data according to claim 1, characterized in that: The deep Q-network DQN uses a neural network to approximate the Q-value function, and the update process of the Q-value is represented by the following formula: where s t is the state at time t, a t is the action taken at time t, Q(s t , a t ) is the Q-value of taking action a t under state s t at time t, Q’(s t , a t ) is the updated Q-value of taking action a t under state s t at time t, α is the learning rate used to represent the speed at which new information overrides old information, r is the immediate reward obtained by taking action a t under state s t at time t, δ is the discount factor used to weigh the importance of future rewards, α, δ ∈ [0, 1], is the maximum value among the Q-values of taking all possible actions a under state s t+1 at time t + 1.

6. The coke quality prediction method combining production dynamic data according to claim 1, characterized in that: The model training process of the deep Q-network DQN includes: S1. Set the second loss function and the second optimizer of the deep Q-network DQN, and input the training set into the deep Q-network DQN for model training; S2. Calculate the loss value based on the second loss function, and the second optimizer updates the model parameters according to the loss value and the network gradient information; S3. If the loss value is less than the second preset threshold, the model training ends, and the current deep Q-network DQN is the trained deep Q-network DQN; otherwise, return to S1 and continue to use the training set for model training; Among them, the second loss function uses the mean square error loss function to calculate the difference between the Q-value output by the deep Q-network DQN and the target Q-value.

7. The coke quality prediction method combining production dynamic data according to claim 6, wherein: When the deep Q-network DQN is trained on the training set, the generalization ability of the model is evaluated by observing the performance on the validation set, and the hyperparameters and structure of the model are tuned; After obtaining the trained deep Q-network DQN, the performance of the model is evaluated using the test set.

Images