CN-121980166-A - Screening method and system for blast furnace heat prediction model

CN121980166ACN 121980166 ACN121980166 ACN 121980166ACN-121980166-A

Abstract

The application provides a screening method and a screening system for a blast furnace heat prediction model, and relates to the technical field of blast furnace smelting. The method comprises the steps of training a plurality of candidate prediction models, calculating a combined evaluation index of each model facing furnace thermal anomaly early warning, wherein the index comprises three indexes for measuring prediction accuracy, anomaly capturing capability and early warning advance, and determining an optimal prediction model according to an evaluation result of the combined evaluation index. According to the application, by constructing the combined evaluation index oriented to abnormal early warning, the model evaluation standard is aligned with the actual production requirement, and the screened optimal model has the advanced and accurate early warning capability on the thermal abnormality of the key furnace, so that precious reaction time can be strived for on-site operators, and the method has higher practical value.

Inventors

YANG YIRU
ZHANG YANG
QIAN WEIDONG
ZHOU YIFAN
LI LIANG
LI HONGKUAN
Zhu Chengle
ZHU JUN
HAN XIAOFEI
YUAN LEI
HUI XUEJUN

Assignees

上海宝信软件股份有限公司

Dates

Publication Date: 20260505
Application Date: 20251222

Claims (10)

1. A screening method of a blast furnace heat prediction model is characterized by comprising the following steps: Training to generate a plurality of candidate prediction models based on historical operation data of the blast furnace; calculating combined evaluation indexes corresponding to the candidate prediction models and facing the blast furnace thermal anomaly early warning, wherein the combined evaluation indexes at least comprise a first index for measuring the prediction accuracy of the models on the furnace thermal anomaly event, a second index for measuring the capturing capability of the models on the actually-occurring furnace thermal anomaly event and a third index for measuring the effective advance of the model early warning; and evaluating the candidate prediction models by using the combined evaluation index, and determining an optimal prediction model according to the evaluation result.
2. The method for screening a blast furnace heat prediction model according to claim 1, wherein the first index is an abnormality early warning accuracy rate, the second index is an abnormality early warning recall rate, and the third index is an early warning average time.
3. The method for screening a blast furnace heat prediction model according to claim 2, wherein, Counting the predicted number of hits of the same type of furnace thermal anomalies in a preset time window which takes the time point of the model prediction of the furnace thermal anomalies as the center, and dividing the predicted number of hits by the total number of times of the model prediction of the furnace thermal anomalies; The calculation mode of the abnormal early warning recall rate comprises that in a preset time window taking a time point of actual furnace thermal abnormality as a center, the statistical model also predicts the abnormal recall number of the same type of furnace thermal abnormality successfully, and the abnormal recall number is divided by the total number of actual furnace thermal abnormality.
4. A method of screening a blast furnace heat prediction model according to any one of claims 1 to 3, wherein the step of determining an optimal prediction model based on the evaluation result comprises: Respectively calculating the ranks of each candidate prediction model on the first index, the second index and the third index; adding the three ranks of each candidate prediction model to obtain a total ranking score; And selecting the candidate prediction model with the smallest total ranking score as the optimal prediction model.
5. A method of screening a blast furnace heat prediction model according to any one of claims 1 to 3, wherein the step of determining an optimal prediction model based on the evaluation result comprises: setting weight coefficients for the ranks of the first index, the second index and the third index respectively; respectively calculating the ranks of each candidate prediction model on the three indexes; carrying out weighted summation on the three ranks of each candidate prediction model to obtain weighted total score; and selecting the candidate prediction model with the minimum weighted total score as the optimal prediction model.
6. The method according to any one of claims 1 to 5, wherein in the step of training to generate a plurality of candidate prediction models, the model training module adopts a modified loss function that is a sum of losses of all samples in the training set, and the loss calculation method is that for any one sample i in the training set: Wherein, the Representing the predicted value of each sample, Representing the actual value of each sample, Representative of the number of samples of the training set, And Represented as the mean and standard deviation of the actual values of all samples in the training set.
7. The method for screening a blast furnace heat prediction model according to any one of claims 1 to 5, wherein before the step of training to generate a plurality of candidate prediction models, further comprising: and carrying out oversampling treatment on the furnace heat abnormal samples in the historical operation data.
8. A screening system for a blast furnace heat prediction model, comprising: the training module is used for training and generating a plurality of candidate prediction models based on historical operation data of the blast furnace; The index calculation module is used for calculating combined evaluation indexes corresponding to the candidate prediction models and facing the early warning of the furnace thermal anomaly of the blast furnace, wherein the combined evaluation indexes at least comprise a first index for measuring the prediction accuracy of the model on the furnace thermal anomaly event, a second index for measuring the capture capacity of the model on the actually-occurring furnace thermal anomaly event and a third index for measuring the effective advance of the model early warning; and the evaluation screening module is used for evaluating the candidate prediction models by using the combined evaluation index and determining an optimal prediction model according to the evaluation result.
9. The screening system of a blast furnace thermal prediction model according to claim 8, wherein the first index is an abnormal early warning accuracy rate, the second index is an abnormal early warning recall rate, and the third index is an early warning average time.
10. The screening system for a blast furnace heat prediction model according to claim 9, wherein, Counting the predicted number of hits of the same type of furnace thermal anomalies in a preset time window which takes the time point of the model prediction of the furnace thermal anomalies as the center, and dividing the predicted number of hits by the total number of times of the model prediction of the furnace thermal anomalies; The calculation mode of the abnormal early warning recall rate comprises that in a preset time window taking a time point of actual furnace thermal abnormality as a center, the statistical model also predicts the abnormal recall number of the same type of furnace thermal abnormality successfully, and the abnormal recall number is divided by the total number of actual furnace thermal abnormality.

Description

Screening method and system for blast furnace heat prediction model Technical Field The invention relates to the technical field of blast furnace smelting, in particular to a screening method and a screening system of a blast furnace heat prediction model. Background The blast furnace heat prediction refers to predicting the change trend of molten iron temperature or silicon content index by analyzing blast furnace operation parameters, and optimizing a regulation strategy to realize advanced control of furnace temperature and stable production. The core module comprises furnace temperature prediction, early warning and regulation, but the prediction model usually only focuses on the accuracy of prediction, so that the accuracy is high, and the situation that the furnace temperature cannot be predicted in advance when abnormality occurs often occurs. Therefore, the invention redesigns the index evaluation system of the furnace temperature prediction model, and screens the optimal model in the multi-algorithm integration, thereby improving the overall performance of the furnace temperature prediction model. From the use value of the optimization regulation of the furnace thermal prediction model, the key point is whether the abnormality can be predicted in advance, and the traditional prediction model indexes (such as R 2 and RMSE) usually only pay attention to the prediction of most stable values, so that the prediction of extreme values is ignored. Therefore, according to the key points of the actual triggering abnormal threshold, three evaluation indexes of abnormal early warning accuracy, abnormal early warning recall rate and early warning average time are required to be constructed and used for super-parameter tuning cross-validation evaluation and multi-algorithm screening methods, and a multi-algorithm model is constructed for comprehensive evaluation so as to select an optimal model. In the prior art, one common method is to train a plurality of candidate models by using a plurality of machine learning algorithms, then evaluate the candidate models by using some traditional overall statistical indexes, such as overall prediction accuracy, root mean square error or average absolute error, and the like, and select a model with the index performing optimally as a final prediction model. However, the blast furnace operation is in a steady state most of the time, and true furnace heat abnormality (such as cooling or heating of the furnace) is a minority of cases. The above conventional evaluation criteria aim at minimizing the overall prediction error of the model over all data points, which results in the screened model, while excellent over the overall data set, often sacrificing the prediction ability for a few outlier data in order to fit most stationary data. Therefore, at the key moment of the most need of early warning, namely when the furnace heat is about to be abnormal, the model screened by the standard often has the problems of inaccurate prediction, early warning lag and even complete failure. The practical value of the prediction model is severely limited by the dislocation between the model evaluation standard and the accurate prediction requirement for abnormal events in actual production regulation. Disclosure of Invention Aiming at the defects in the prior art, the invention aims to provide a screening method and a screening system for a blast furnace heat prediction model. The screening method of the blast furnace heat prediction model provided by the invention comprises the following steps: Training to generate a plurality of candidate prediction models based on historical operation data of the blast furnace; calculating combined evaluation indexes corresponding to the candidate prediction models and facing the blast furnace thermal anomaly early warning, wherein the combined evaluation indexes at least comprise a first index for measuring the prediction accuracy of the models on the furnace thermal anomaly event, a second index for measuring the capturing capability of the models on the actually-occurring furnace thermal anomaly event and a third index for measuring the effective advance of the model early warning; and evaluating the candidate prediction models by using the combined evaluation index, and determining an optimal prediction model according to the evaluation result. Preferably, the first index is an abnormal early warning accuracy rate, the second index is an abnormal early warning recall rate, and the third index is an early warning average time. Preferably, the calculation mode of the abnormality early warning accuracy comprises the steps of counting the predicted number of hits of the same type of furnace thermal abnormality actually occurring in a preset time window taking the time point of the model prediction of the furnace thermal abnormality as the center, and dividing the predicted number of hits by the total number of times of the model prediction of the furnac