Academic Journal of Science and Technology ISSN: 2771-3032 | Vol. 9, No. 2, 2024 133 Optimizing Energy Efficiency in EAF Steel Production: A Data‐Driven Approach Yubin Zhao, Yize Qi*, Jose Diogo Franco Viveiros School of Engineering, Cardiff University, The Parade, Cardiff, Wales, CF24 3AA, UK * Corresponding author Abstract: The article’s main research mission is making predictions about energy usage within an EAF steel production operation and being able to reliably predict energy consumption. Besides, we analysed the relationship between energy consumption in electric arc furnace (EAF) production processes by the method of data mining, so we could obtain accurate forecasts to optimize production costs and maximize energy efficiency. Last but not least, by using decision tree, random tree algorithms and multilayer perceptron models, we can easily achieve the intended goals of efficient use of energy and provide a better plan for electric arc furnace (EAF). Keywords: EAF, Weka, Data analysis, Data modelling. 1. Introduction Nowadays, with the development of modern industry, we can’t ignore that the steel industry is still playing a very important role in the industry of any country around whole world. Britain and other European countries are the earliest batch of countries to develop the steel during the First Industrial Revolution, so over the past century, scientist developed many ways in order to do the steel production, such as Smelting Reduction, Gas-based Direct Reduction and Direct Reduction Ironmaking, or DRI. Besides, one of the most widely used is the Electric Arc Furnace. Although the Electric Arc Furnace is the most widely used way of steel production, there are still many problems occurred during the process, which can cause a series of problems. Like Furnace process control, Slag control and improvement of energy efficiency. In order to solve the problems, we must take the full advantages of date, which, in that case, a suitable type of data mining formulation has to be considered, and it not only can deal with the problems we meet during the steel production, but also have potential opportunity, which can lead the steel industry into a brighter future. 2. Data Understanding 2.1. Data Exploration and Description We can analyse the data given in appendix( Fig 1, Fig2), this way making any change to the data, if needed, and get a brief introduction about dataset. The dataset includes 21 variables and 3499 data points, including precise measures of energy usage, the manufacture of recycled steel, and the utilization of metal materials. When having a preliminary investigation of the values imported into WEKA, it is notable cases of unclear and missing data. Instead of having a focus on production, the main objective of this data mining project for steel production in EAF’s is to optimize the energy use. In terms of optimizing energy use, if “Billet Tons” represent the amount of steel produced per batch, then this parameter may won’t have a significant impact on optimizing the energy use. This attribute (“Billet Tons”) didn’t have a significant contribution on predicting the energy consumption and was considered for us as unimportant for the feature selection in the pre-processing stage, so the decision was to remove it. Furthermore, each sample was assigned a serial number called “Heat Number”, which was optional and didn’t have any significance importance in aiding the optimization. 2.2. Data processing With the intention of further reducing the attribute and simplifying the model even more, the option “CorrelationAttributeEval” was selected in the attribute evaluator panel and the cut-off value was set to 0.02. Figure 1. Select attributes tab on Weka. The properties with values between -0.02 and 0.02 showed a low relevance to the experiment, so they were deleted together with the attribute “Argon (kg)”. We also discovered during the data filtering process that the Steel Grade attribute has two outliers of 00/01/1900 and four consecutive outliers of 0. We made the decision to include a 134 filter so the outliers can be managed so we can keep them from throwing off the algorithm. Figure 2. GenericObjectEditor Tab on Weka. Thus, the missing data can now be checked and handled, which can be judged by the values displayed by “Missing” in the selected attribute. The “ReplaceMissingValues” function in Weka can be responsible for handling the missing data. In conclusion, those things which can get from the dataset are as following: 1) There is a moderate positive correlation between Heat Number and EAF (MWh) about 0.56, which may show that an increase in energy consumption as production continues. 2) The Carbon injected (Kg) and EAF (MWh) have moderate positive correlation about 0.39, which may suggested that the amount of carbon used may have an impact on energy consumption. 3) As for Total Scrap Mix and EAF (MWh), their moderate positive correlation is only 0.34, so it’s goes without saying that total waste may be a key factor affecting energy consumption. So in order to improve energy efficiency. The following elements need to be optimized, such as The time of the heat, amount of carbon injected and scrap mix. Besides, Focusing on carbon usage and waste mix optimization may help reduce energy consumption, thereby lowering costs and increasing production efficiency. Figure 3. Process tab in Weka. 3. Methodology The algorithms that we decided to choose to analyse the data were the J48 for Decision Trees and RandomTree. We choose these two models because they both have great advantages, such as high accuracy when dealing with small and medium sized datasets. Also, the model Multiplayer Percepton was used. This model has its unique advantages when dealing with complex energy systems, where the model can adjust to the number of layers and neurons to better adapt to the more complex non-linear relationships. These three models were chosen because they are very distinct learning and pattern recognition approaches. Furthermore, this will allow us to have a more comprehensive understanding of the relationship between energy consumption and attributes. - J48 Decision Tree: This is a decision tree algorithm, which it is often used the divide-and-conquer method in order to divide the instances, this way according to the similarity. The J48 algorithm is an application of the C4.5 algorithm, which is a very stable and reliable algorithm, as we can determine through long-term verification. -Multiplayer Perceptron: This is a very common multi- layer feed-forward artificial neural network model. It’s an algorithm often used for classification and prediction of data, and it includes one or more hidden layers that allows it to learn non-linear patterns in the data, being a acceptable tool for deep learning. -RandomTree: This algorithm can be used for both classification and regression problems. It is an important algorithm in trees. Each of its internal nodes represents a test of a characteristics, each branch represents the result of the test, and each leaf node represents a prediction. When writing up the methodology, we always tried to maintain the importance of validating the results of this experiment to ensure the accuracy and stability of the results. In addition, we also paid attention to the iterative improvement of this experiment. The algorithm was replaced several times to find the most suitable algorithm for this experiment. 4. Model Building 4.1. MultilayerPerceptron The MultiplayerPercepton is a neutral model in Weka, commonly used in data mining and machine learning. This algorithm has its own benefits when working with small to 135 medium sized data, even though it is not as accurate as deep learning algorithms when working with larger datasets. We determined that it was best to split the data into ten sections, nine for training and one for testing, based on our experience. With this said, all the data shown below was acquired using a cross-validation of 10. Note that the target attribute must be changed to (Num)EAF (MWh) Table 1. Model training results after data segmentation. Learning Rate 0.3 0.3 0.3 0.3 0.4 0.5 0.3 0.3 0.3 momentum 0.2 0.2 0.2 0.2 0.2 0.2 0.3 0.4 0.5 percent split 60 66 76 86 66 66 66 66 66 Correlation coefficient 0.8081 0.7924 0.8166 0.8142 0.7857 0.7644 0.7738 0.7921 0.7911 mean absolute error 2.0465 1.7199 1.9788 1.8565 1.7555 1.7918 1.7573 1.681 1.6878 Root mean squared error 2.6707 2.4518 2.54 2.4501 2.5537 2.5789 2.5564 2.4446 2.4931 Relative absolute error 70.321 58.078 66.7905 62.2219 59.2815 60.5079 59.34 56.766 56.9932 Root relative squared error 70.8842 63.3657 67.0566 64.2228 65.9995 66.6507 66.07 63.1805 64.4343 Figure 4. Classify tab on Weka. Looking at the data above in fig.6, we can see that when the values of the Learning Rate is 0.3, the Momentum is 0.4 and the Percentage split is 66%. The values of Mean absolute error, Root mean squared error, Relative absolute error and Root relative squared error in the data are minimum, so it can easily get that the error between the actual value and the predicted value is minimum in this data. Figure 5. Program execution process and regression diagram. 4.2. J48 The J48 algorithm offers a high degree of accuracy when analysing small to medium sized datasets, a clear structure and, results that are simple to read and comprehend. The algorithm can also handle attributes with values missing and has a high fault tolerance. It usually struggles to manage 136 characteristics that often require constant performance, tough. When using the default parameter, we must first use the option Discretize in order to change the attribute from numeric to nominal. Alternatively, we can also use the NumericalToNominal attribute to change the attribute from numeric to nominal. Figure 6. EAF data type conversion process. We then turned Cross-validation to 10 and used the algorithm J48 to run different percentage split values. Table 2. Model training results after data segmentation. percent split 56 66 76 86 96 Correctly Classified Instances 68.2911 69.6883 68.2956 70.2041 69.2857 Incorrectly Classified Instances 31.7089 30.3117 31.7044 29.7959 30.7143 Mean absolute error 0.0808 0.0806 0.0808 0.0795 0.0816 Root mean squared error 0.2239 0.2157 0.2225 0.2169 0.2164 Relative absolute error 69.8711 69.8533 69.1381 67.8949 68.4355 Root relative squared error 92.7772 89.8878 91.0512 88.7971 87.1663 Figure 6. Classify tab on Weka. We observed the following values through the window Classifier on output. We can deduct from the size of the values in the above table that the accuracy is the highest when the percentage split value is 56%.In Weka, selecting Visualise offers a more user- friendly way to present the data so that we can see how various attributes relate to one another. Moreover, users can also spot outliers by examining the image. 137 Figure 7. Program execution process and regression diagram 4.3. RandomTree In Weka, the RandomTree algorithm is often used for machine learning and also data mining. It produces a decision tree model with an easy-to-understand structure. It can handle a variety of data types, these including nominal and numerical data, which makes the operation simpler, although there is some randomness present and the accuracy in not the highest. The cross-validation is set, Folds 10 and then the RandomTree algorithm is used to classify the data. The arithmetic resulting from varying the percentage split values produces the following data. Table 3. Model learning training results after data segmentation percent split 56% 66% 76% 86% 96% Correctly Classified Instances 64.9773 63.3305 62.5745 66.3265 68.5714 Incorrectly Classified Instances 35.0227 36.6695 37.4255 33.6735 31.4286 Mean absolute error 0.0728 0.0755 0.076 0.068 0.0657 Root mean squared error 0.2559 0.2635 0.2633 0.2488 0.2376 Relative absolute error 62.9523 64.9629 65.0131 58.0774 55.0952 Root relative squared error 106.0335 108.4655 107.7565 101.8892 95.6949 From the data shown in the above table we can conclude that when the value of the percentage split is 96%, the accuracy of this data is the highest and the error is the least. Figure 8. Classify tab with the Percentage split set to 96 on Weka. Thus, in order to be able to observe the data in a better visionally way, the graphical decision tree can also be opened. 138 Figure 9. Visualize tree. 5. Results, Performance and Evaluation In this part, the performance results and evaluation outcomes of the models will be discussed, the ones we used on our group’s case study. In order to accomplish the target of energy consumption prediction in the EAF production process, we took into account the next ways: Performance Metrics Rationale: - Accuracy and Error Metrics: For the target of providing a comprehensive picture of the model’s prediction accuracy and error size. We took the metrics such as Mean Absolute Error, Root Mean Squared Error, and Correctly Classified Instances. - Relative Errors: For the Relative Errors, we choose Relative Absolute Error and Root Relative Squared Error, so we could know how our model would perform against baseline predictors, so we could get a better result. Performance Results: - Multiplayer Perceptron: I believe our model performs well in terms of accuracy, but we can still see from the error rate that there is some variability in the predictions. Besides, we should also notice that the correlation coefficient is significantly high, which means there is a relationship between the predicted and actual values. - J48 Decision Tree: Based on the database we have, there is even a high tolerance for missing attribute values, we think it’s still beneficial to the nature of our dataset. - RandomTree: With the help of the RandomTree, it’s easier to deal with various data types, including nominal and numerical. After we conducted a comprehensive data analysis by Weka, we can get the performance metrics. In addition, the error rates, accuracy percentages, and the correlation coefficients were derived from tests and validations. As for the pictures and graphs which are referenced in the Methodology section, they provide a representative and quantitative evidence of each model’s performance. Last but not least, from the presentation analysis of the three different models, it goes without saying that every one of them have their advantages and disadvantages. We believe that in order to do a good prediction of the energy consumption, we should take all of them into account, which would be the most effective way. 6. Conclusion and Reflection This study has as its main aims to reveal the complex relationships of energy consumption in the EAF production process and provide more accurate predictions to optimize production costs and maximize energy efficiency. Besides, this article emphasized the importance of using different models in Weka. This not only upgrades the existing production model, but also helps the new insights for the development of smart manufacturing in the era of Industry 4.0. Throughout the hole development of this case study in Weka, different models were used, each one being a different method of solving from each other, to make sure the prediction of the energy consumption would be the best. But there are still several things which need to be consideration. 1) In order to get the best prediction model, different software should be taken into use. 2) In the section of Data understanding, mathematical knowledge should be used as a tool to assist analysis Finally, this case study showed that with the right data pre- processing, processing and with the right models, it is possible to get a high accuracy. References [1] Huang, F. 2021. Network Activities Recognition and Analysis Based on Supervised Machine Learning Classification Methods Using J48 and Na\"ive Bayes Algorithm. Available at: https://arxiv.org/abs/2105.13698 [Accessed: 20 November 2023]. [2] IOPscience. 2017. IOP Conference Series: Materials Science and Engineering - IOPscience. Available at: https://iopscience.iop.org/journal/1757-899X [3] Raheli, B., Aalami, M.T., El-Shafie, A., Ghorbani, M.A. and Deo, R.C. 2017. Uncertainty assessment of the multilayer perceptron (MLP) neural network model with implementation of the novel hybrid MLP-FFA method for prediction of biochemical oxygen demand and dissolved oxygen: a case study of Langat River. Environmental Earth Sciences 76(14). doi: https://doi.org/10.1007/s12665-017-6842-z 139 Appendix Fig. 1 Dataset in an Excel file. Fig. 2 Dataset opened in Weka.