Interpretable neural network coke hot quality prediction model driven by domain knowledge
By building a domain knowledge-driven interpretable neural network model, using SHAP algorithm and weight loss function to optimize coal parameter groups, the problems of poor interpretability and association of the existing coke thermal mass prediction model are solved, and the correlation does not conform to the coking field cognition is improved. Predictive performance and credibility.
Patent Information
- Application Number
- CN202411029309.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2044-07-30
AI Technical Summary
The existing coke thermal mass prediction model based on neural networks has problems such as poor interpretability, the correlation does not conform to the understanding of the coking field, and it is difficult to improve prediction performance.
By building a domain knowledge-driven interpretable neural network model, using SHAP algorithm to identify unreasonable coal parameter groups, optimize the weight scores of coal parameter groups, combine the importance maximization and minimize loss functions, inject knowledge in coking domains, and improve the model Interpretability and predictive performance.
It improves the interpretability and prediction performance of the model, ensures that the predicted results are in line with the understanding of the coking field, enhances the credibility of the model, and can more reasonably analyze abnormal results, and optimizes the coal mixing plan.
Smart Images

Figure CN118553326B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the fields of coking, artificial intelligence and data mining processing technology, and in particular to an interpretable neural network coke thermal quality prediction model driven by domain knowledge. Background Art
[0002] Due to its powerful feature extraction and prediction capabilities, most of the current accurate coke hot quality prediction methods are mainly based on neural network models. However, the black box characteristics of neural network models have always made them have poor interpretability. This not only limits the further improvement of the model's prediction performance, but also reduces its credibility in practical applications, bringing great challenges to the industrial deployment of existing neural network-based coke hot quality prediction models.
[0003] The traditional neural network-based coke hot quality prediction model only relies on a single regression loss function during training, so that the model can only optimize the trainable weight parameters from the perspective of data statistics, and the association between the learned coal parameters and coke quality often violates the knowledge in the coking field. This model based on unreasonable associations may mistakenly emphasize parameters that are considered weakly correlated in the field, while ignoring parameters that are strongly correlated, thereby hindering the model from mining deeper associations that are consistent with the field's cognition, limiting the further improvement of the model's prediction performance. In addition, the model's explanation of abnormal coke hot quality prediction results is often insufficient. Therefore, how to improve the interpretability of the existing neural network-based coke hot quality prediction model, correctly identify and adjust the importance of unreasonable coal parameters to conform to the coking field's cognition, and then accurately and stably predict the hot quality of coke through coal blending plans is the core issue of the current intelligent coal blending technology. Summary of the invention
[0004] The purpose of the present invention is to overcome the problems that the prior art has poor interpretability when predicting the hot quality of coke, the mined associations do not conform to the coking field cognition, and the prediction performance is difficult to further improve, and to provide a domain knowledge-driven interpretable neural network coke hot quality prediction model.
[0005] In order to achieve the above technical objectives and the above technical effects, the present invention is implemented through the following technical solutions:
[0006] A domain knowledge-driven interpretable neural network coke hot quality prediction model includes the following steps:
[0007] Step S1: constructing a coal sample data set;
[0008] Step S2: Neural network model construction and prediction, outputting the predicted coke hot quality value based on the numerical characteristics of the blended coal;
[0009] Step S3: identifying coal parameter groups, identifying coal parameter groups that do not conform to the knowledge in the coking field;
[0010] Step S4: Calculation of coal parameter group weight score;
[0011] Step S5: Optimizing the importance of coal parameter groups, using a weight loss function to optimize the weight scores of the identified coal parameter groups.
[0012] Furthermore, in the step S1, the blended coal sample data set includes multi-dimensional coal parameters and corresponding coke thermal quality values, and the coke thermal quality values include reactivity CRI and post-reaction strength CSR.
[0013] Furthermore, in step S2, the neural network is an artificial neural network model including a single input layer, multiple hidden layers and a single output layer, and the loss function used in the training process is the mean absolute error, mean square error or root mean square error function.
[0014] Furthermore, the single output layer is a fully connected layer with two neurons, one of which outputs the reactiveness CRI, and the other outputs the post-reaction strength CSR.
[0015] Furthermore, in step S3, the identification of the coal parameter group that does not conform to the knowledge in the coking field is implemented based on the SHAP algorithm, and the output SHAP value is used to measure the importance of the coal parameter group.
[0016] Furthermore, there are two groups of coal parameter groups, one group is a parameter group P1 that needs to be increased in importance and is determined based on domain knowledge, and the other group is a parameter group P2 that needs to be reduced in importance.
[0017] Furthermore, in step S4, the weight scores of the coal parameter groups are calculated by respectively passing the parameter group P1 that needs to increase its importance and the parameter group P2 that needs to decrease its importance into the neural network to obtain their respective outputs and dividing their values by the sum of the two to obtain the weight scores C1 and C2 of the two coal parameter groups.
[0018] Furthermore, the weight score is the average value of the weight scores of all coal samples in a batch during the model training process.
[0019] Furthermore, in step S5, the weight loss function includes an importance maximization loss function and a minimization loss function, wherein the importance maximization loss function is used to increase the importance of the identified coal parameter group P1, and the importance minimization loss function is used to reduce the importance of the identified coal parameter group P2.
[0020] Furthermore, the importance maximization loss function and the minimization loss function are mean absolute error, mean square error or root mean square error function.
[0021] The beneficial effects of the present invention are:
[0022] 1) The coal parameter group weight score calculation and weight loss function optimization method proposed in the present invention can inject coking field knowledge into the neural network of the black box model. This method can further enable the model to mine and learn deeper and more reasonable associations during the training process, which helps to improve the performance of coke hot quality prediction results.
[0023] 2) The method proposed in the present invention combines the SHAP value and aligns the associations learned by the model with the domain knowledge, which can make the importance of the model output more consistent with the coking field, greatly improving the interpretability of the model and the credibility in practical application. In addition, the method can also provide a more reasonable analysis of abnormal prediction results to determine the factors affecting the quality of coke, which is helpful for the timely adjustment and optimization of the coal blending plan. BRIEF DESCRIPTION OF THE DRAWINGS
[0024] Figure 1 is a block diagram of the prediction method of the present invention;
[0025] Figure 2 is a flow chart of the prediction method of the present invention;
[0026] Figure 3 is a schematic diagram of a neural network model in the prediction method of the present invention;
[0027] Figure 4 A schematic diagram showing the comparison between the predicted values and the true values of the coke CRI and CSR indicators of the artificial neural network ANN after optimizing the weights and combining the domain knowledge;
[0028] Figure 5 Schematic diagram of the comparison between the predicted values and true values of the coke CRI and CSR indicators of the artificial neural network ANN without combining domain knowledge. DETAILED DESCRIPTION
[0029] The present invention will be described in detail below with reference to the accompanying drawings and in combination with embodiments.
[0030] like Figures 1 to 3As shown, a domain knowledge-driven interpretable neural network coke hot quality prediction model includes the construction of a blended coal sample data set, the construction and prediction of a neural network model, the identification of coal parameter groups, the calculation of coal parameter group weight scores, and the optimization of coal parameter group importance. The blended coal data with moisture, ash, volatile matter, sulfur, G value, X value, and Y value parameters are input into an artificial neural network model containing a single input layer, multiple hidden layers, and a single output layer. The loss function during training can use the mean absolute error, mean square error, or root mean square error function. Based on the SHapley Additive exPlanations (SHAP) algorithm, the SHAP value is output to measure the importance of the coal parameter group, where the number of coal parameter groups is two, one is the parameter group P1 that needs to be improved according to domain knowledge, and the other is the parameter group P2 that needs to be reduced. Then, the parameter group P1 and the parameter group P2 are respectively input into the neural network to obtain their respective outputs and their values are divided by the sum of the two to obtain the weight scores C1 and C2 of the two coal parameter groups. Subsequently, the weight loss function including the importance maximization loss function and the minimization loss function is used to optimize the weight score of the identified coal parameter group, where the importance maximization loss function is used to increase the importance of the identified coal parameter group P1, and the importance minimization loss function is used to reduce the importance of the identified coal parameter group P2. Finally, the coke reactivity CRI and coke post-reaction strength CSR are predicted using the trained domain knowledge-driven interpretable neural network.
[0031] Embodiment 1:
[0032] 94 industrial production data were collected, and the mixed coal data with parameters of moisture, ash, volatile matter, sulfur, G value, X value and Y value were passed into an artificial neural network model containing a single input layer, multiple hidden layers and a single output layer. The loss function during training was the mean absolute error. Based on the SHapley Additive exPlanations (SHAP) algorithm, the SHAP value was output to measure the importance of the coal parameter group. There were two groups of coal parameter groups, one of which was the parameter group P1 that needed to be improved according to domain knowledge, and the other was the parameter group P2 that needed to be reduced in importance. Then, the parameter group P1 and the parameter group P2 were respectively passed into the neural network to obtain their respective outputs and their values were divided by the sum of the two to obtain the weight scores C1 and C2 of the two groups of coal parameter groups. Then, the weight loss function containing the importance maximization loss function and the importance minimization loss function was used to optimize the weight scores of the identified coal parameter groups, where the importance maximization loss function was used to improve the importance of the identified coal parameter group P1, and the importance minimization loss function was used to reduce the importance of the identified coal parameter group P2. Finally, the coke reactivity (CRI) and coke post-reaction strength (CSR) are predicted using the trained domain knowledge-driven interpretable neural network. Figure 4The thermal quality parameters of coke predicted by the model of the present invention and the thermal quality parameters of coke actually produced by industrial coke ovens. As can be seen from the figure, the average absolute error value between the coke CRI and CSR indicators predicted by the prediction model constructed by the present invention and the thermal quality parameters of coke produced industrially is 3.41, and the prediction result is very close to the true value.
[0033] Comparative Example 1:
[0034] 94 industrial production data used in Example 1 were selected, and the blended coal data with parameters of moisture, ash, volatile matter, sulfur, G value, X value and Y value were input into an artificial neural network model without domain knowledge, including a single input layer, multiple hidden layers and a single output layer. The loss function during training adopted the mean absolute error. Finally, the trained neural network was used to predict the coke reactivity CRI and the coke post-reaction strength CSR. Figure 5 The predicted values and true values of the coke CRI and CSR indicators of the artificial neural network ANN without combining domain knowledge. It can be seen from the figure that based on unreasonable correlation, the average absolute error between the predicted coke CRI and CSR indicators and the hot quality parameters of industrial coke is 5.45, and the prediction model cannot accurately predict the coke CRI and CSR indicators.
[0035] In summary, the method provided by the present invention can improve the interpretability and prediction accuracy of the existing neural network-based coke hot quality prediction algorithm, and can dig out deeper associations that are consistent with domain cognition, thereby meeting the needs of actual coking production and reducing production costs.
[0036] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.
Claims
1. A domain knowledge-driven interpretable neural network coke hot quality prediction model, characterized in that: The model consists of the following steps: Step S1: constructing a coal sample data set; Step S2: Neural network model construction and prediction, outputting the predicted coke hot quality value based on the numerical characteristics of the blended coal; Step S3: identifying coal parameter groups, identifying coal parameter groups that do not conform to the knowledge in the coking field; Step S4: Calculation of coal parameter group weight score; Step S5: optimizing the importance of coal parameter groups, using a weight loss function to optimize the weight scores of the identified coal parameter groups; In step S3, the identification of the coal parameter group that does not conform to the knowledge in the field of coking is implemented based on the SHAP algorithm, and the importance of the coal parameter group is measured by using the output SHAP value; The number of the coal parameter groups is two, one is a parameter group P1 that needs to be increased in importance according to domain knowledge, and the other is a parameter group P2 that needs to be decreased in importance; In step S4, the weight scores of the coal parameter groups are calculated by respectively passing the parameter group P1 that needs to increase its importance and the parameter group P2 that needs to decrease its importance into the neural network to obtain their respective outputs and dividing their values by the sum of the two to obtain the weight scores C1 and C2 of the two coal parameter groups.
2. The domain knowledge driven interpretable neural network coke hot quality prediction model according to claim 1 is characterized in that: In step S1, the blended coal sample data set includes multi-dimensional coal parameters and corresponding coke thermal quality values, wherein the coke thermal quality values include reactivity CRI and post-reaction strength CSR.
3. The domain knowledge driven interpretable neural network coke hot quality prediction model according to claim 2 is characterized in that: In step S2, the neural network is an artificial neural network model including a single input layer, multiple hidden layers and a single output layer, and the loss function used in the training process is a mean absolute error, a mean square error or a root mean square error function.
4. The domain knowledge driven interpretable neural network coke hot quality prediction model according to claim 3 is characterized in that: The single output layer is a fully connected layer with two neurons, one of which outputs the reactiveness CRI and the other outputs the post-reaction strength CSR.
5. The domain knowledge driven interpretable neural network coke hot quality prediction model according to claim 1 is characterized in that: The weight score is the average value of the weight scores of all coal samples in a batch during the model training process.
6. The domain knowledge driven interpretable neural network coke hot quality prediction model according to claim 5 is characterized in that: In step S5, the weight loss function includes an importance maximization loss function and a minimization loss function, wherein the importance maximization loss function is used to increase the importance of the identified coal parameter group P1, and the importance minimization loss function is used to reduce the importance of the identified coal parameter group P2.
7. The domain knowledge driven interpretable neural network coke hot quality prediction model according to claim 6 is characterized in that: The importance maximization loss function and the minimization loss function are mean absolute error, mean square error or root mean square error function.
