TWI787954B - Method and computer system for predicting temperature of molten steel - Google Patents

Method and computer system for predicting temperature of molten steel Download PDF

Info

Publication number
TWI787954B
TWI787954B TW110129850A TW110129850A TWI787954B TW I787954 B TWI787954 B TW I787954B TW 110129850 A TW110129850 A TW 110129850A TW 110129850 A TW110129850 A TW 110129850A TW I787954 B TWI787954 B TW I787954B
Authority
TW
Taiwan
Prior art keywords
molten steel
data
machine learning
learning model
production data
Prior art date
Application number
TW110129850A
Other languages
Chinese (zh)
Other versions
TW202307744A (en
Inventor
蘇育德
林守謙
戴字庭
黃俊奎
洪子頡
Original Assignee
中國鋼鐵股份有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 中國鋼鐵股份有限公司 filed Critical 中國鋼鐵股份有限公司
Priority to TW110129850A priority Critical patent/TWI787954B/en
Application granted granted Critical
Publication of TWI787954B publication Critical patent/TWI787954B/en
Publication of TW202307744A publication Critical patent/TW202307744A/en

Links

Images

Landscapes

  • Treatment Of Steel In Its Molten State (AREA)
  • Control Of Temperature (AREA)
  • Investigating And Analyzing Materials By Characteristic Methods (AREA)
  • Carbon Steel Or Casting Steel Manufacturing (AREA)

Abstract

A method for predicting temperature of molten steel includes: training a first machine learning model according to history production data and corresponding first temperature data of molten steel; obtaining new production data and corresponding second temperature data of molten steel, and training a second machine learning model according to the history production data, the first temperature data of molten steel, the new production data, and the second temperature data of molten steel; determining a first weight of the first machine learning model and a second weight of the second machine learning model according to an optimal algorithm; and combing the first machine learning model and the second machine learning model according to the first weight and the second weight to generate an assembly model so as to predict the temperature of molten steel.

Description

鋼液溫度預測方法與電腦系統Molten Steel Temperature Prediction Method and Computer System

本揭露是關於煉鋼廠中鋼液的溫度預測方法。The present disclosure is about a method for predicting the temperature of molten steel in a steelmaking plant.

鋼液溫度的控制是各國煉鋼廠存在已久的課題，合適的溫度對於鋼品品質有很大的影響外，對於製程的穩定也有絕對的影響。在一些習知技術中是根據現場操作人員將製程條件帶入物理反應公式來預測鋼液的溫度，但這樣的做法除了有人員差異容易產生誤判外，也容易因為設備或環境因素的改變使得預測的準確度降低，因此如何提出一個自動化且適應性的做法為此領域技術人員所關心的議題。The temperature control of molten steel has been a long-standing issue in steel factories in various countries. The appropriate temperature not only has a great impact on the quality of steel products, but also has an absolute impact on the stability of the manufacturing process. In some conventional technologies, the temperature of the molten steel is predicted based on the on-site operators bringing the process conditions into the physical reaction formula, but this method is not only prone to misjudgments due to personnel differences, but also easily leads to predictions due to changes in equipment or environmental factors. The accuracy decreases, so how to propose an automatic and adaptive approach is a topic of concern to those skilled in the art.

本揭露的實施例提出一種鋼液溫度預測方法，適用於電腦系統，此鋼液溫度預測方法包括：根據歷史生產資料以及對應的第一鋼液溫度資料訓練第一機器學習模型；取得新生產資料以及對應的第二鋼液溫度資料，根據至少部分的歷史生產資料、新生產資料、至少部分的第一鋼液溫度資料以及第二鋼液溫度資料訓練第二機器學習模型；根據最佳化演算法決定對應第一機器學習模型的第一權重以及對應第二機器學習模型的第二權重；以及根據第一權重以及第二權重結合第一機器學習模型以及第二機器學習模型以產生集成式模型，藉此預測鋼液溫度。The embodiment of the present disclosure proposes a method for predicting molten steel temperature, which is suitable for computer systems. The method for predicting molten steel temperature includes: training the first machine learning model according to historical production data and corresponding first molten steel temperature data; obtaining new production data And the corresponding second molten steel temperature data, according to at least part of the historical production data, new production data, at least part of the first molten steel temperature data and the second molten steel temperature data to train the second machine learning model; according to the optimization calculation The method determines the first weight corresponding to the first machine learning model and the second weight corresponding to the second machine learning model; and combines the first machine learning model and the second machine learning model according to the first weight and the second weight to generate an integrated model , so as to predict the molten steel temperature.

在一些實施例中，歷史生產資料與新生產資料包括多個特徵參數，這些特徵參數包括鋼液量、鋼液初始溫度、合金含量以及吹氧量。第一鋼液溫度資料以及第二鋼液溫度資料是關於從轉爐被倒入接鋼桶的鋼液。In some embodiments, the historical production data and the new production data include multiple characteristic parameters, and these characteristic parameters include molten steel volume, initial temperature of molten steel, alloy content, and oxygen blowing volume. The first molten steel temperature data and the second molten steel temperature data are about the molten steel poured into the steel ladle from the converter.

在一些實施例中，第二機器學習模型是根據新生產資料以及預設期間內的歷史生產資料所訓練，預設期間是根據轉爐的維修周期所決定。In some embodiments, the second machine learning model is trained according to new production data and historical production data within a predetermined period, and the predetermined period is determined according to the maintenance cycle of the converter.

在一些實施例中，鋼液溫度預測方法還包括在訓練第二機器學習模型之前，刪除新生產資料中的無效值或缺值；以及根據第二鋼液溫度資料將新生產資料分為多個群組，並執行數據增生程序以增加其中一個群組的資料量。In some embodiments, the molten steel temperature prediction method further includes deleting invalid or missing values in the new production data before training the second machine learning model; and dividing the new production data into multiple groups, and perform a data augmentation procedure to increase the amount of data in one of the groups.

在一些實施例中，上述的最佳化演算法表示為以下數學式。

Figure 02_image001
In some embodiments, the above-mentioned optimization algorithm is expressed as the following mathematical formula.
Figure 02_image001

其中i為正整數，

Figure 02_image003
為第二權重，
Figure 02_image005
為第二機器學習模型，
Figure 02_image007
為第一權重，
Figure 02_image009
為第一機器學習模型，
Figure 02_image011
為第二鋼液溫度資料。 where i is a positive integer,
Figure 02_image003
is the second weight,
Figure 02_image005
is the second machine learning model,
Figure 02_image007
is the first weight,
Figure 02_image009
is the first machine learning model,
Figure 02_image011
is the temperature data of the second molten steel.

在一些實施例中，集成式模型表示為以下數學式，其中

Figure 02_image013
為集成式模型。
Figure 02_image015
In some embodiments, the integrated model is expressed as the following mathematical formula, where
Figure 02_image013
is an integrated model.
Figure 02_image015

以另外一個角度來說，本揭露的實施例提出一種電腦系統，包括記憶體以及處理器。記憶體儲存有多個指令，處理器通訊連接至記憶體，用以執行指令以完成多個步驟。這些步驟包括：根據歷史生產資料以及對應的第一鋼液溫度資料訓練第一機器學習模型；取得新生產資料以及對應的第二鋼液溫度資料，根據至少部分的歷史生產資料、新生產資料、至少部分的第一鋼液溫度資料以及第二鋼液溫度資料訓練第二機器學習模型；根據最佳化演算法決定對應第一機器學習模型的第一權重以及對應第二機器學習模型的第二權重；以及根據第一權重以及第二權重結合第一機器學習模型以及第二機器學習模型以產生集成式模型，藉此預測鋼液溫度。From another perspective, the embodiments of the present disclosure provide a computer system including a memory and a processor. The memory stores multiple instructions, and the processor communicates with the memory to execute the instructions to complete multiple steps. These steps include: training the first machine learning model according to historical production data and corresponding first molten steel temperature data; obtaining new production data and corresponding second molten steel temperature data, according to at least part of historical production data, new production data, At least part of the first molten steel temperature data and the second molten steel temperature data train the second machine learning model; determine the first weight corresponding to the first machine learning model and the second weight corresponding to the second machine learning model according to the optimization algorithm weight; and combining the first machine learning model and the second machine learning model according to the first weight and the second weight to generate an integrated model, thereby predicting the molten steel temperature.

關於本文中所使用之「第一」、「第二」等，並非特別指次序或順位的意思，其僅為了區別以相同技術用語描述的元件或操作。The terms "first", "second" and the like used herein do not specifically refer to a sequence or sequence, but are only used to distinguish elements or operations described with the same technical terms.

圖1是描述煉鋼過程的鋼液溫度變化示意圖，在此僅繪示煉鋼過程中的幾個階段，其他階段(例如軋延)則省略。在階段110中，由轉爐101進行吹練，在此階段鋼液102的溫度上升。在進入階段120時，鋼液102由轉爐101被倒入接鋼桶103，此時鋼液102的溫度第一次急速下降，接著接鋼桶103被運送至精煉站，此時鋼液102的熱量持續被接鋼桶103吸收也從上方散失，因此溫度逐漸下降。在階段130由脫氧機(degasser)104進行精煉，此時可藉由供電或投鋁吹氣升溫，鋼液102的溫度上升。在階段140時鋼液102被運送至連鑄站，此時鋼液102的熱量逐漸散失，溫度逐漸下降。在進入階段150時，鋼液102被注入至分配器(tundish)105，此時鋼液102的溫度會在短時間顯著下降，是鋼液102第二次的急速降溫。FIG. 1 is a schematic diagram describing the temperature change of molten steel in the steelmaking process. Only a few stages in the steelmaking process are shown here, and other stages (such as rolling) are omitted. In stage 110, blowing is carried out by the converter 101, during which the temperature of the molten steel 102 rises. When entering the stage 120, the molten steel 102 is poured into the receiving ladle 103 from the converter 101, and the temperature of the molten steel 102 drops rapidly for the first time at this time, and then the ladle 103 is transported to the refining station, at this time, the temperature of the molten steel 102 Heat continues to be absorbed by the steel drum 103 and also dissipates from above, so the temperature gradually drops. In stage 130, the deoxidizer (degasser) 104 is used for refining. At this time, the temperature of the molten steel 102 can be raised by supplying electricity or blowing aluminum. In stage 140, the molten steel 102 is transported to the continuous casting station, at this time, the heat of the molten steel 102 is gradually dissipated, and the temperature gradually drops. When entering the stage 150, the molten steel 102 is injected into the distributor (tundish) 105, at this time, the temperature of the molten steel 102 will drop significantly in a short time, which is the second rapid cooling of the molten steel 102.

電腦系統160通訊連接至各階段中的設備，用以收集各個階段的生產資料以及鋼液溫度資料。電腦系統160可以實作為個人電腦、伺服器、工業電腦、中央控制台或具有計算能力的各種電子裝置。電腦系統160包括處理器161與記憶體162，處理器161用以通訊連接至記憶體162，在此通訊連接可以透過任意有線或無線的通訊手段來達成，或者也可透過互聯網來達成。處理器161可為中央處理器、微處理器、微控制器、或特殊應用積體電路等，記憶體162可為隨機存取記憶體、唯讀記憶體、快閃記憶體、軟碟、硬碟、光碟、隨身碟、磁帶或是可透過網際網路存取之資料庫，其中儲存有多個指令，處理器161會執行這些指令來完成一個鋼液溫度預測方法，以下將詳細說明此方法。The computer system 160 is communicatively connected to the equipment in each stage to collect the production data and molten steel temperature data of each stage. The computer system 160 can be implemented as a personal computer, a server, an industrial computer, a central console, or various electronic devices with computing capabilities. The computer system 160 includes a processor 161 and a memory 162. The processor 161 is used to communicate with the memory 162. The communication connection can be achieved through any wired or wireless communication means, or can also be achieved through the Internet. The processor 161 can be a central processing unit, a microprocessor, a microcontroller, or an application-specific integrated circuit, etc., and the memory 162 can be random access memory, read-only memory, flash memory, floppy disk, hard disk, etc. Disc, CD, flash drive, magnetic tape, or a database that can be accessed through the Internet, wherein a plurality of instructions are stored, and the processor 161 will execute these instructions to complete a molten steel temperature prediction method, which will be described in detail below .

大致上來說，本揭露是每隔一段時間(例如一個月、一季或是幾個禮拜，本揭露不限制其長度)就會產生一個機器學習模型，過去產生的機器學習模型與新產生的機器學習模型會被結合在一起成為一個集成式模型，用以預測鋼液溫度。在此實施例中是要預測階段120內的鋼液溫度，但此揭露的方法也可以用於其他任意階段。Generally speaking, this disclosure generates a machine learning model at regular intervals (such as a month, a season, or several weeks, and the length of this disclosure is not limited). The models are combined into an ensemble model to predict the molten steel temperature. In this embodiment, the temperature of molten steel in stage 120 is predicted, but the disclosed method can also be used in any other stage.

首先，收集生產資料，這些生產資料包括最近一期發生(例如最近一個月)的新生產資料以及在更早之前發生的歷史生產資料。這些生產資料包括多個特徵參數，例如鋼液量、鋼液初始溫度、合金含量(此合金可以是錳、鉻或任意其他金屬)以及吹氧量等，其中鋼液初始溫度為階段120剛開始時的鋼液溫度，本揭露不限於上述特徵參數，其他任意可能會影響鋼液溫度的特徵參數都可以被納入。另外，也會收集每筆生產資料所對應的鋼液溫度資料，在此實施例中，鋼液溫度資料是關於從轉爐101被倒入接鋼桶103的鋼液102。在一些實施例中是要預測鋼液在一段時間內的溫度變化，因此每筆鋼液溫度資料可以有多溫度數值，在一些實施例中也可以預測鋼液在一特定時間點的溫度，因此每筆鋼液溫度資料可以包含一個溫度數值。舉例來說，每個產品都有對應的生產資料，當產品的中間產物(即鋼液)被倒入接鋼桶103以後可以透過溫度感測器量測鋼液的溫度，在階段120內可量測一或多次，藉此產生對應的溫度數值，這些溫度數值組成鋼液溫度資料。每個產品的生產資料與對應的鋼液溫度資料則形成一筆訓練樣本，在此可以收集多個訓練樣本。First, production data are collected, which include new production data that occurred in the latest period (for example, the last month) and historical production data that occurred earlier. These production materials include a plurality of characteristic parameters, such as molten steel amount, initial temperature of molten steel, alloy content (this alloy can be manganese, chromium or any other metal) and oxygen blowing amount etc., wherein the initial temperature of molten steel is stage 120 just beginning The molten steel temperature at the time, the present disclosure is not limited to the above characteristic parameters, and any other characteristic parameters that may affect the molten steel temperature can be included. In addition, the molten steel temperature data corresponding to each piece of production data will also be collected. In this embodiment, the molten steel temperature data is about the molten steel 102 poured from the converter 101 into the steel ladle 103 . In some embodiments, it is to predict the temperature change of molten steel in a period of time, so each molten steel temperature data can have multiple temperature values, and in some embodiments, the temperature of molten steel at a specific point in time can also be predicted, so Each molten steel temperature data can contain a temperature value. For example, each product has corresponding production data. When the intermediate product of the product (i.e. molten steel) is poured into the steel drum 103, the temperature of the molten steel can be measured through a temperature sensor. Measure one or more times to generate corresponding temperature values, and these temperature values constitute molten steel temperature data. The production data of each product and the corresponding molten steel temperature data form a training sample, where multiple training samples can be collected.

根據歷史生產資料與對應的鋼液溫度資料可以訓練出至少一個機器學習模型，以下表示為

Figure 02_image009
、
Figure 02_image017
…、
Figure 02_image019
，其中N為正整數，其中
Figure 02_image009
表示上一期所訓練出的機器學習模型，
Figure 02_image017
表示前兩期所訓練出的機器學習模型，以此類推。由於新產生的機器學習模型在下一期就會變成舊的機器學習模型
Figure 02_image009
，在此說明如何產生新的機器學習模型。 According to the historical production data and the corresponding molten steel temperature data, at least one machine learning model can be trained, which is expressed as
Figure 02_image009
,
Figure 02_image017
...,
Figure 02_image019
, where N is a positive integer, where
Figure 02_image009
Indicates the machine learning model trained in the previous period,
Figure 02_image017
Indicates the machine learning model trained in the previous two phases, and so on. Since the newly generated machine learning model will become the old machine learning model in the next period
Figure 02_image009
, which shows how to generate a new machine learning model.

在一些實施例中，可以先對新生產資料執行一些前處理，例如去除離群值(outlier)、刪除新生產資料中的無效值或缺值等。此無效值與缺值可能是因為感測器故障或是電腦系統沒有收集到相關數據所導致。在一些實施例中也可以對類別項目編碼，例如給予製程代號、精煉代號或接鋼桶狀態等等。此外，由於不同的產品可能會導致不同的鋼液溫度，但產品的數量不一定相同，如果直接把所有的新生產資料拿來訓練可能會偏向某個產品的數據，因此在一些實施例中可以根據鋼液溫度資料將新生產資料分為多個群組，例如800~1000度為一個群組，1000~1200度為一個群組，如果這些群組的資料量(即訓練樣本的個數)相差太多，可以執行一個數據增生程序以增加其中一或多個群組的資料量，使得各個群組中的資料量相當。舉例來說，如果新生產資料中缺少800~1000度這一個群組的訓練樣本，可以從更早之前產生的訓練樣本中擷取相同溫度群組的訓練樣本以加入至新生產資料中。In some embodiments, some pre-processing may be performed on the new production data, such as removing outliers, deleting invalid or missing values in the new production data, and the like. This invalid value and missing value may be caused by sensor failure or computer system not collecting relevant data. In some embodiments, category items can also be coded, for example, a process code, a refining code, or a status of a steel drum, etc. are given. In addition, since different products may lead to different molten steel temperatures, but the number of products is not necessarily the same, if all new production materials are directly used for training, it may be biased towards the data of a certain product, so in some embodiments it can be Divide the new production data into multiple groups according to the molten steel temperature data, for example, 800~1000 degrees is a group, 1000~1200 degrees is a group, if the amount of data in these groups (that is, the number of training samples) If the difference is too large, a data augmentation procedure can be performed to increase the amount of data in one or more groups, so that the amount of data in each group is equal. For example, if the training samples of the group of 800-1000 degrees are missing in the new production data, the training samples of the same temperature group can be extracted from the training samples generated earlier to add to the new production data.

接下來，根據至少部分的歷史生產資料、新生產資料以及對應的鋼液溫度資料來訓練新的機器學習模型。此機器學習模型可以是線性迴歸模型、嶺迴歸模型、KNN(k-nearest neighbors)迴歸器、XGBoost或LGBM(light gradient boosting algorithm)、支持向量機、各種類型的神經網路等，本揭露並不在此限。在一些實施例中，共會蒐集k個月的生產資料，其中k為正整數(例如k=9)，這k個月的生產資料包括最近一個月的新生產資料以及更早k-1個月的歷史生產資料。正整數k與轉爐101的維修周期有關，在轉爐101經過維修以後可能會改變一些生產條件，因此舊的生產資料並不適合用來預測新的鋼液溫度。但在一些實驗中發現當k從1增加到9時會增加預測準確度，但k＞9時並不會得到更佳的預測準確度，因此在此實施例中設定k=9。以另一個角度來說，新產生的機器學習模型是根據新生產資料以及一預設期間(即k-1個月)內的歷史生產資料所訓練出，此預設期間是根據轉爐101的維修周期所決定。新產生的機器學模型表示為

Figure 02_image005
。 Next, a new machine learning model is trained according to at least part of the historical production data, the new production data and the corresponding molten steel temperature data. This machine learning model can be linear regression model, ridge regression model, KNN (k-nearest neighbors) regressor, XGBoost or LGBM (light gradient boosting algorithm), support vector machine, various types of neural networks, etc. This disclosure is not in This limit. In some embodiments, a total of k months of production data will be collected, where k is a positive integer (for example, k=9), and these k months of production data include new production data of the latest month and earlier k-1 Monthly historical means of production. The positive integer k is related to the maintenance cycle of the converter 101, and some production conditions may change after the converter 101 is maintained, so the old production data is not suitable for predicting the new molten steel temperature. However, in some experiments, it is found that when k increases from 1 to 9, the prediction accuracy will increase, but when k>9, no better prediction accuracy will be obtained, so k=9 is set in this embodiment. From another point of view, the newly generated machine learning model is trained based on new production data and historical production data within a preset period (ie k-1 months), and the preset period is based on the maintenance of the converter 101 cycle is determined. The newly generated machine learning model is expressed as
Figure 02_image005
.

接下來要結合過去產生的機器學習模型

Figure 02_image009
、
Figure 02_image017
…、
Figure 02_image019
以及新產生的機器學習模型
Figure 02_image005
，其中N為正整數，本揭露並不限制正整數N的數值。在一些實施例中是將每個機器學習模型輸出的數值乘上一個權重後再相加以得到一個集成式模型，因此要先決定每個機器學習模型的權重，在此實施例中可根據一最佳化演算法來決定這些權重，此最佳化演算法表示為以下數學式1。 [數學式1]
Figure 02_image001
The next step is to combine the machine learning models generated in the past
Figure 02_image009
,
Figure 02_image017
...,
Figure 02_image019
and the newly generated machine learning model
Figure 02_image005
, where N is a positive integer, and the present disclosure does not limit the value of the positive integer N. In some embodiments, the value output by each machine learning model is multiplied by a weight and then added to obtain an integrated model. Therefore, the weight of each machine learning model must be determined first. In this embodiment, an optimal These weights are determined by an optimization algorithm, and this optimization algorithm is represented by the following mathematical formula 1. [mathematical formula 1]
Figure 02_image001

其中i為正整數，

Figure 02_image003
、
Figure 02_image007
…為權重，
Figure 02_image011
為最近一期的鋼液溫度資料。值得注意的是，在數學式1中的
Figure 02_image005
、
Figure 02_image009
等指的是機器學習模型所輸出的鋼液溫度，更具體來說，在此是將最近一期的生產資料輸入至各個機器學習模型
Figure 02_image009
、
Figure 02_image017
…、
Figure 02_image019
、
Figure 02_image005
，並將這些機器學習模型的輸出乘上權重後相加再與最近一期的鋼液溫度資料
Figure 02_image011
相減。數學式1又稱為一個目標函數(objective function)，最佳化演算法是要找到一組權重
Figure 02_image003
、
Figure 02_image007
…
Figure 02_image021
，使得此目標函數有最小值，在此可以用任意的搜尋方法(例如基因演算法)來找到權重
Figure 02_image003
、
Figure 02_image007
…
Figure 02_image021
，本揭露並不限制採用何種方法來求解權重
Figure 02_image003
、
Figure 02_image007
…
Figure 02_image021
。 where i is a positive integer,
Figure 02_image003
,
Figure 02_image007
…is the weight,
Figure 02_image011
It is the latest molten steel temperature data. It is worth noting that in Mathematical Equation 1 the
Figure 02_image005
,
Figure 02_image009
etc. refers to the molten steel temperature output by the machine learning model, and more specifically, here is the input of the latest production data into each machine learning model
Figure 02_image009
,
Figure 02_image017
...,
Figure 02_image019
,
Figure 02_image005
, and the output of these machine learning models is multiplied by weights and added together with the latest molten steel temperature data
Figure 02_image011
Subtract. Mathematical formula 1 is also called an objective function, and the optimization algorithm is to find a set of weights
Figure 02_image003
,
Figure 02_image007
…
Figure 02_image021
, so that this objective function has a minimum value, where any search method (such as genetic algorithm) can be used to find the weight
Figure 02_image003
,
Figure 02_image007
…
Figure 02_image021
, this disclosure does not limit the method used to solve the weight
Figure 02_image003
,
Figure 02_image007
…
Figure 02_image021
.

圖2是根據一實施例繪示集成式模型的示意圖。請參照圖2，透過權重

Figure 02_image003
、
Figure 02_image007
…
Figure 02_image021
將機器學習模型
Figure 02_image009
、
Figure 02_image017
…、
Figure 02_image019
、
Figure 02_image005
結合可以得到集成式模型
Figure 02_image013
，可表示為以下數學式2。 [數學式2]
Figure 02_image015
FIG. 2 is a schematic diagram illustrating an ensemble model according to an embodiment. Please refer to Figure 2, through the weight
Figure 02_image003
,
Figure 02_image007
…
Figure 02_image021
machine learning model
Figure 02_image009
,
Figure 02_image017
...,
Figure 02_image019
,
Figure 02_image005
Combined to get an ensemble model
Figure 02_image013
, can be expressed as the following mathematical formula 2. [mathematical formula 2]
Figure 02_image015

在此是以線性的方式來結合過去訓練好的機器學習模型以及新的機器學習模型，如此一來可以累積過去的訓練模型知識來補足目前訓練資料量的不足，也就是透過儲存過去訓練好的模型，合成出最符合現況的模型。值得注意的是，如果生產的環境或設備有很大的變動，則過去的機器學習模型可能無法準確地預測鋼液溫度，透過上述的方式會降低這些舊機器學習模型的權重(甚至有可能權重為0)，如此一來集成式模型可更適應於新的環境與設備。另一方面，如果只根據新生產資料來預測鋼液溫度，可能會面臨資料量不足的問題。或者，有可能因為設備進行維修或者是某些月份的產品較少而導致訓練樣本較少，在習知技術中可能會沒有保留住過去正確的數據而被強迫更新了模型，反而使得預測的準確度下降。透過上述手段可以平衡適應性以及資料量不足的問題。Here, the machine learning model trained in the past and the new machine learning model are combined in a linear manner, so that the past training model knowledge can be accumulated to make up for the lack of current training data, that is, by storing the past trained model, and synthesize the model that best fits the situation. It is worth noting that if the production environment or equipment has changed greatly, the past machine learning models may not be able to accurately predict the molten steel temperature, and the weight of these old machine learning models will be reduced through the above-mentioned method (there may even be a weight of is 0), so that the integrated model can be more adaptable to new environments and devices. On the other hand, if the molten steel temperature is predicted only based on new production data, it may face the problem of insufficient data. Or, there may be fewer training samples due to equipment maintenance or fewer products in certain months. In conventional technology, the correct data in the past may not be retained and the model may be forced to update, which makes the prediction more accurate. degree drops. The problems of adaptability and insufficient amount of data can be balanced through the above means.

圖3是根據一實施例繪示鋼液溫度預測方法的流程圖。請參照圖3，在步驟301，根據歷史生產資料以及對應的第一鋼液溫度資料訓練第一機器學習模型。在步驟302，取得新生產資料以及對應的第二鋼液溫度資料，根據至少部分的歷史生產資料、新生產資料、至少部分的第一鋼液溫度資料以及第二鋼液溫度資料訓練第二機器學習模型。在步驟303，根據最佳化演算法決定對應第一機器學習模型的第一權重以及對應第二機器學習模型的第二權重。在步驟304，根據第一權重以及第二權重結合第一機器學習模型以及第二機器學習模型以產生集成式模型，藉此預測鋼液溫度。然而，圖3中各步驟已詳細說明如上，在此便不再贅述。值得注意的是，圖3中各步驟可以實作為多個程式碼或是電路，本發明並不在此限。此外，圖3的方法可以搭配以上實施例使用，也可以單獨使用。換言之，圖3的各步驟之間也可以加入其他的步驟。以另外一個角度來說，本發明也提出了一電腦程式產品，此產品可由任意的程式語言及/或平台所撰寫，當此電腦程式產品被載入至電腦系統並執行時，可執行上述的方法。FIG. 3 is a flowchart illustrating a method for predicting molten steel temperature according to an embodiment. Referring to FIG. 3 , in step 301 , a first machine learning model is trained according to historical production data and corresponding first molten steel temperature data. In step 302, the new production data and the corresponding second molten steel temperature data are obtained, and the second machine is trained according to at least part of the historical production data, new production data, at least part of the first molten steel temperature data and the second molten steel temperature data learning model. In step 303, a first weight corresponding to the first machine learning model and a second weight corresponding to the second machine learning model are determined according to an optimization algorithm. In step 304, the first machine learning model and the second machine learning model are combined according to the first weight and the second weight to generate an integrated model, thereby predicting the molten steel temperature. However, each step in FIG. 3 has been described in detail above, and will not be repeated here. It should be noted that each step in FIG. 3 can be implemented as a plurality of program codes or circuits, and the present invention is not limited thereto. In addition, the method in FIG. 3 can be used in conjunction with the above embodiments, or can be used alone. In other words, other steps may also be added between the steps in FIG. 3 . From another point of view, the present invention also proposes a computer program product, which can be written by any programming language and/or platform. When the computer program product is loaded into the computer system and executed, the above-mentioned method.

雖然本發明已以實施例揭露如上，然其並非用以限定本發明，任何所屬技術領域中具有通常知識者，在不脫離本發明的精神和範圍內，當可作些許的更動與潤飾，故本發明的保護範圍當視後附的申請專利範圍所界定者為準。Although the present invention has been disclosed above with the embodiments, it is not intended to limit the present invention. Anyone with ordinary knowledge in the technical field may make some changes and modifications without departing from the spirit and scope of the present invention. The scope of protection of the present invention should be defined by the scope of the appended patent application.

110,120,130,140,150:階段 101:轉爐 102:鋼液 103:接鋼桶 104:脫氧機 105:分配器 160:電腦系統 161:處理器 162:記憶體 w 1、w 2、w 3、w N+1:權重 M now-1、M now-2、M now-N 、M now :機器學習模型 M syn :集成式模型 301~304:步驟 110, 120, 130, 140, 150: stage 101: converter 102: molten steel 103: connecting steel drum 104: deoxidizer 105: distributor 160: computer system 161: processor 162: memory w 1 , w 2 , w 3 , w N +1 : weight M now -1 , M now -2 , M now-N , M now : machine learning model M syn : integrated model 301~304: steps

為讓本發明的上述特徵和優點能更明顯易懂，下文特舉實施例，並配合所附圖式作詳細說明如下。 圖1是描述煉鋼過程的鋼液溫度變化示意圖。 圖2是根據一實施例繪示集成式模型的示意圖。 圖3是根據一實施例繪示鋼液溫度預測方法的流程圖。 In order to make the above-mentioned features and advantages of the present invention more comprehensible, the following specific embodiments are described in detail together with the accompanying drawings. Figure 1 is a schematic diagram describing the temperature change of molten steel in the steelmaking process. FIG. 2 is a schematic diagram illustrating an ensemble model according to an embodiment. FIG. 3 is a flowchart illustrating a method for predicting molten steel temperature according to an embodiment.

301~304:步驟 301~304: Steps

Claims (8)

一種鋼液溫度預測方法，適用於一電腦系統，該鋼液溫度預測方法包括：根據一歷史生產資料以及對應該歷史生產資料的一第一鋼液溫度資料訓練至少一第一機器學習模型；取得一新生產資料以及對應該新生產資料的一第二鋼液溫度資料，根據至少部分的該歷史生產資料、該新生產資料、至少部分的該第一鋼液溫度資料以及該第二鋼液溫度資料訓練一第二機器學習模型，其中該第二機器學習模型是根據該新生產資料以及一預設期間內的該歷史生產資料所訓練，該預設期間是根據該轉爐的維修周期所決定，該至少一第一機器學習模型不是根據該新生產資料所訓練；根據一最佳化演算法決定對應該至少一第一機器學習模型的至少一第一權重以及對應該第二機器學習模型的一第二權重；以及根據該至少一第一權重以及該第二權重結合該至少一第一機器學習模型以及該第二機器學習模型以產生一集成式模型。 A method for predicting molten steel temperature, suitable for a computer system, the method for predicting molten steel temperature includes: training at least one first machine learning model according to a historical production data and a first molten steel temperature data corresponding to the historical production data; obtaining A new production data and a second molten steel temperature data corresponding to the new production data, according to at least part of the historical production data, the new production data, at least part of the first molten steel temperature data and the second molten steel temperature data training a second machine learning model, wherein the second machine learning model is trained according to the new production data and the historical production data within a preset period, the preset period is determined according to the maintenance cycle of the converter, The at least one first machine learning model is not trained according to the new production data; at least one first weight corresponding to the at least one first machine learning model and one corresponding to the second machine learning model are determined according to an optimization algorithm a second weight; and combining the at least one first machine learning model and the second machine learning model according to the at least one first weight and the second weight to generate an integrated model. 如請求項1所述之鋼液溫度預測方法，其中該歷史生產資料與該新生產資料包括多個特徵參數，該些特徵參數包括鋼液量、鋼液初始溫度、合金含量以及吹氧量，其中該第一鋼液溫度資料以及該第二鋼液溫度資料是關 於從一轉爐被倒入一接鋼桶的鋼液。 The molten steel temperature prediction method as described in Claim 1, wherein the historical production data and the new production data include a plurality of characteristic parameters, and these characteristic parameters include the amount of molten steel, the initial temperature of molten steel, the alloy content and the amount of oxygen blowing, Wherein the first molten steel temperature data and the second molten steel temperature data are related For molten steel poured from a converter into a steel drum. 如請求項1所述之鋼液溫度預測方法，還包括：在訓練該第二機器學習模型之前，刪除該新生產資料中的無效值或缺值；以及根據該第二鋼液溫度資料將該新生產資料分為多個群組，並執行一數據增生程序以增加該些群組的其中之一群組的資料量。 The molten steel temperature prediction method as described in claim 1, further comprising: before training the second machine learning model, deleting invalid or missing values in the new production data; and according to the second molten steel temperature data, the The new production data is divided into multiple groups, and a data augmentation procedure is performed to increase the data volume of one of the groups. 如請求項1所述之鋼液溫度預測方法，其中該最佳化演算法表示為以下數學式：min wi [w 1×M now +w 2×M now-1+…-Data now ]2其中i為正整數，w 1為該第二權重，M now 為該第二機器學習模型，w 2為該至少一第一權重，M now-1為該至少一第一機器學習模型，Data now 為該第二鋼液溫度資料。 The molten steel temperature prediction method as described in Claim 1, wherein the optimization algorithm is expressed as the following mathematical formula: min wi [ w 1 × M now + w 2 × M now -1 + ...- Data now ] 2 where i is a positive integer, w 1 is the second weight, M now is the second machine learning model, w 2 is the at least one first weight, M now -1 is the at least one first machine learning model, Data now is The second molten steel temperature data. 如請求項4所述之鋼液溫度預測方法，其中該集成式模型表示為以下數學式：M syn =Σ i w i ×M now+1-i 其中M syn 為該集成式模型。 The molten steel temperature prediction method as described in Claim 4, wherein the integrated model is expressed as the following mathematical formula: M syn =Σ i w i × M now +1- i where M syn is the integrated model. 一種電腦系統，包括：一記憶體，儲存有多個指令；以及 一處理器，通訊連接至該記憶體，用以執行該些指令以完成多個步驟：根據一歷史生產資料以及對應該歷史生產資料的一第一鋼液溫度資料訓練至少一第一機器學習模型；取得一新生產資料以及對應該新生產資料的一第二鋼液溫度資料，根據至少部分的該歷史生產資料、該新生產資料、至少部分的該第一鋼液溫度資料以及該第二鋼液溫度資料訓練一第二機器學習模型，其中該第二機器學習模型是根據該新生產資料以及一預設期間內的該歷史生產資料所訓練，該預設期間是根據該轉爐的維修周期所決定，該至少一第一機器學習模型不是根據該新生產資料所訓練；根據一最佳化演算法決定對應該至少一第一機器學習模型的至少一第一權重以及對應該第二機器學習模型的一第二權重；以及根據該至少一第一權重以及該第二權重結合該至少一第一機器學習模型以及該第二機器學習模型以產生一集成式模型。 A computer system comprising: a memory storing a plurality of instructions; and A processor, connected to the memory by communication, for executing the instructions to complete a plurality of steps: training at least one first machine learning model according to a historical production data and a first molten steel temperature data corresponding to the historical production data ; Obtain a new production data and a second molten steel temperature data corresponding to the new production data, based on at least part of the historical production data, the new production data, at least part of the first molten steel temperature data and the second steel training a second machine learning model based on the liquid temperature data, wherein the second machine learning model is trained based on the new production data and the historical production data within a preset period based on the maintenance cycle of the converter Determining that the at least one first machine learning model is not trained according to the new production data; determining at least one first weight corresponding to the at least one first machine learning model and corresponding to the second machine learning model according to an optimization algorithm a second weight; and combine the at least one first machine learning model and the second machine learning model according to the at least one first weight and the second weight to generate an integrated model. 如請求項6所述之電腦系統，其中該歷史生產資料與該新生產資料包括多個特徵參數，該些特徵參數包括鋼液量、鋼液初始溫度、合金含量以及吹氧量，其中該第一鋼液溫度資料以及該第二鋼液溫度資料是關於從一轉爐被倒入一接鋼桶的鋼液。 The computer system as described in claim 6, wherein the historical production data and the new production data include a plurality of characteristic parameters, these characteristic parameters include molten steel volume, initial temperature of molten steel, alloy content and oxygen blowing amount, wherein the first The first molten steel temperature data and the second molten steel temperature data relate to molten steel poured into a steel ladle from a converter. 如請求項6所述之電腦系統，其中該最佳化演算法表示為以下數學式：min wi [w 1×M now +w 2×M now-1+…-Data now ]2其中i為正整數，w 1為該第二權重，M now 為該第二機器學習模型，w 2為該至少一第一權重，M now-1為該至少一第一機器學習模型，Data now 為該第二鋼液溫度資料。 The computer system as described in Claim 6, wherein the optimization algorithm is expressed as the following mathematical formula: min wi [ w 1 × M now + w 2 × M now -1 + ...- Data now ] 2 where i is positive Integer, w 1 is the second weight, M now is the second machine learning model, w 2 is the at least one first weight, M now -1 is the at least one first machine learning model, Data now is the second Molten steel temperature data.
TW110129850A 2021-08-12 2021-08-12 Method and computer system for predicting temperature of molten steel TWI787954B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
TW110129850A TWI787954B (en) 2021-08-12 2021-08-12 Method and computer system for predicting temperature of molten steel

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
TW110129850A TWI787954B (en) 2021-08-12 2021-08-12 Method and computer system for predicting temperature of molten steel

Publications (2)

Publication Number Publication Date
TWI787954B true TWI787954B (en) 2022-12-21
TW202307744A TW202307744A (en) 2023-02-16

Family

ID=85795138

Family Applications (1)

Application Number Title Priority Date Filing Date
TW110129850A TWI787954B (en) 2021-08-12 2021-08-12 Method and computer system for predicting temperature of molten steel

Country Status (1)

Country Link
TW (1) TWI787954B (en)

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101504544A (en) * 2008-02-01 2009-08-12 霍尼韦尔国际公司 Methods and apparatus for an oxygen furnace quality control system
TWI521449B (en) * 2015-02-03 2016-02-11 中國鋼鐵股份有限公司 Coke dry quench system and controlling method thereof
CN110795846A (en) * 2019-10-29 2020-02-14 东北财经大学 Construction method of boundary forest model, updating method of multi-working-condition soft computing model for complex industrial process and application of updating method

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN101504544A (en) * 2008-02-01 2009-08-12 霍尼韦尔国际公司 Methods and apparatus for an oxygen furnace quality control system
TWI521449B (en) * 2015-02-03 2016-02-11 中國鋼鐵股份有限公司 Coke dry quench system and controlling method thereof
CN110795846A (en) * 2019-10-29 2020-02-14 东北财经大学 Construction method of boundary forest model, updating method of multi-working-condition soft computing model for complex industrial process and application of updating method

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
網路文獻 M. Shahhosseini, et. al., "Optimizing Ensemble Weights and Hyperparameters of Machine Learning Models for Regression Problems" , 1908.05287, 31 Oct., 2020. [https://doi.org/10.48550/arXiv.1908.05287] *
網路文獻 Q. Yang, et. al., "Molten steel temperature prediction using a hybrid model based on information interaction-enhanced cuckoo search" , Neural Computing and Applications, 23 Oct.,2020. [https://doi.org/10.1007/s00521-020-05413-5]; *

Also Published As

Publication number Publication date
TW202307744A (en) 2023-02-16

Similar Documents

Publication Publication Date Title
CN107464023A (en) The on-line study method and apparatus of life cycle forecast model
JP6439780B2 (en) Magnetic property prediction device and magnetic property control device for electrical steel sheet
JP7503150B2 (en) System and method for controlling a production facility consisting of several facility parts, in particular for producing industrial products such as metal semi-finished products - Patents.com
US11803165B2 (en) Manufacturing support system for predicting property of alloy material, method for generating prediction model, and computer program
JP5462750B2 (en) Molten steel temperature management method, apparatus and program
JP2009265699A (en) Processing time prediction device, method, program, and computer-readable storage medium
JP2011039762A (en) Output value prediction method and device, and program for the method
JP5853723B2 (en) Phosphorus concentration prediction device and blowing control method
JP2021035693A (en) Steel plate temperature prediction device and steel plate temperature prediction method, and learning device and learning method
JP2007052739A (en) Method and device for generating model, method and device for predicting state, and method and system for adjusting state
JP2014059592A (en) Output value prediction device, output value prediction method and program for output value prediction method
TWI787954B (en) Method and computer system for predicting temperature of molten steel
JP6885321B2 (en) Process status diagnosis method and status diagnosis device
CN116258087B (en) Matte grade soft measurement method and device, electronic equipment and storage medium
JP5821656B2 (en) Quick lime concentration prediction device and blowing control method
JP6098553B2 (en) Rejuvenated phosphorus amount prediction device, recovered phosphorus amount prediction method, and converter dephosphorization control method
JP7485918B2 (en) Reactor state evaluation device, reactor state evaluation method, reactor state evaluation program, and trained model generation method
JP7353804B2 (en) Model predictive control system, information processing device, program, and model predictive control method
CN114066050A (en) Product index prediction method, product index prediction device and readable storage medium
JP5402621B2 (en) Manufacturing load prediction apparatus, method, computer program, and computer-readable storage medium
Shigemori Desulphurization control system through locally weighted regression model
JP2005174334A (en) Method and device for suboptimal control by means of search strategy, and method and device for gas separation, particularly for cryogenic air separation
JP2020197944A (en) Model prediction control system, information processing device, program, and model prediction control method
EP4283424A1 (en) Remote-control apparatus, local-control apparatus, learning processing apparatus, method, and program
US11876021B2 (en) Test circuit and test method