Prediction of Steam Consumption in the Tobacco Industry's Primary Processing Workshop

With the continuous maturation of emerging technologies such as artificial intelligence and machine learning, the intelligent application of the primary processing production process has gradually become a major research focus in the industry [1-2]. Steam, as an essential secondary energy source in the tobacco industry, relies heavily on non-renewable energy sources, including natural gas, coal, and petroleum. Rational prediction and decision-making regarding steam production and consumption are of great significance for enterprises to save resources and improve production efficiency [3-7]. Currently, in the steam systems of tobacco enterprises, the primary processing workshop consumes a large amount of steam. When production begins in the primary processing workshop, gas boilers in the power workshop operate at high loads to provide steam to the system. To ensure the production quality of cut tobacco, the steam supply to the primary processing workshop often exceeds its actual demand. This unrefined management of steam supply results in significant steam dispersion, falling short of achieving precise supply and demand management, and causing energy waste due to excessive steam production.

In recent years, increasing attention has been paid to research on energy conservation, consumption reduction, and refined management. Commonly used energy consumption prediction methods include parametric regression, time series analysis, and artificial neural networks [8-13]. For example, Chen et al. [14] proposed an energy consumption prediction model for air conditioning systems based on a deep learning gated recurrent unit (GRU) neural network, predicting energy consumption data for the air conditioning system of a tobacco factory’s storage workshop. Zhao et al. [15] developed a prediction model for steam production and consumption in the steel industry, achieving accurate steam consumption predictions and reducing energy consumption per ton of steel. Song et al. [16] used six machine learning algorithms, including decision trees, Bayesian classifiers, neural networks, and logistic regression, to predict personal annual income. Yu et al. [17] analyzed educational big data using five major machine learning algorithms: logistic regression, decision trees, Bayesian algorithms, Random Forest, and others. While the scope of research on load prediction is broad, studies on steam load prediction in the primary processing workshops of tobacco enterprises are limited. Multiple production devices interact within the workshop, with five steam pipelines supplying high-pressure steam to each piece of equipment, greatly increasing the complexity of the steam load fluctuation patterns. The steam load in the primary processing workshop of tobacco enterprises fluctuates significantly with the start and stop of production equipment. Relying on workers experience to measure and adjust the steam supply often leads to energy waste. If an accurate prediction model is available, it would facilitate source-side control, achieving the goal of energy conservation and emission reduction.

The steam load in the primary processing workshop is affected by various factors, including production processes, equipment, and tasks, resulting in nonlinear changes in steam load data with no obvious patterns and relatively small data volumes. Due to its ensemble nature and feature randomness, Random Forest can effectively reduce the problem of model overfitting. Moreover, the OOB Predictor Importance function embedded in the TreeBagger function of the Random Forest model provides feature importance evaluation, helping identify the key factors influencing steam load. Based on the above analysis, this paper examines the production processes of tobacco enterprises and identifies the factors affecting steam load. It analyzes the relationship between production plans and the required steam load, selecting highly correlated factors as input variables for the model. The Random Forest learning method is employed to predict steam consumption in the primary processing workshop, and the prediction results are compared with those from other traditional models. This research offers a refined prediction method for forecasting steam consumption in the primary processing workshop, providing data support for the rational scheduling of the steam system in tobacco enterprises.

3.1 Data processing

The data used in this study were collected from the primary processing workshop of a tobacco factory over a period of seven months, from March 2023 to September 2023. Due to factory shutdowns or production scheduling issues, data for 138 days of production tasks and steam energy consumption were obtained during this period. First, outlier detection was performed on the dataset, and noisy data with high fluctuations in steam load were preprocessed by removing faulty data. There were no missing values in the dataset. Afterward, normalization was applied to the dataset, and the processed data were fed into the model. After preprocessing, a total of 131 data samples were obtained, which were used in the simulation tests of the prediction model described later.

3.2 Decision tree model

The decision tree is a commonly used machine learning algorithm that performs classification and regression analysis by simulating the human decision-making process. Its core idea is to partition the dataset based on features to build a tree-structured model. As a supervised learning algorithm, the decision tree is easy to understand and interpret. In this study, the decision tree extracts decision rules from the factors affecting the steam load in the primary processing workshop and the corresponding steam load to make predictions. The structure of the decision tree is similar to that of a real tree, consisting of internal nodes, a root node, and leaf nodes. The classification or partitioning of the steam load-related factors in the workshop determines the number of nodes in the decision tree. The goal of this study is to use the influencing factors of the steam load in the workshop to predict the steam load for the next time step. Since this is a regression problem, a regression decision tree has been chosen for steam load prediction. Figure 2 shows the topology of the decision tree.

2.jpg

Figure 2. Principle of decision tree

The primary algorithm used in regression decision trees is Classification and Regression Trees (CART). The CART algorithm can be applied to both classification and regression problems. For regression problems, CART predicts continuous target variables by constructing decision trees. The CART algorithm follows these steps: the model input consists of data from the training set, and it continuously calculates the size of different values under various influencing factors. The optimal splitting point is found using the least squares method to minimize the squared error, which solves the following equation:

$\min _{j, s}\left[\min _{c 1} \sum_{x_i \in R_1(j, s)}\left(y_i-c_1\right)^2+\min _{c 2} \sum_{x_i \in R_2(j, s)}\left(y_i-c_2\right)^2\right]$

where, j represents the j-th feature variable in the dataset; s represents value of the j-th feature variable; and there are: $R_1(j, s)=\left\{x \mid x^j \leq s\right\}$, and $R_2(j, s)=\left\{x \mid x^j \geq s\right\}$.

Through this process, the calculation values for different regions are obtained, and the above steps are repeated for subregions until the squared error reaches its minimum. The final result is a division into several regions, forming a regression decision tree.

3.3 Random Forest model

The Random Forest is an ensemble learning method that primarily uses decision trees as base learners. By constructing multiple decision trees and aggregating their prediction results, the model enhances both its accuracy and stability. The Random Forest model employs bootstrap sampling to train and generate multiple decision trees, which together form the Random Forest. The outputs of these trees are then aggregated using voting or weighted averaging, improving the model's robustness and generalization capability. This approach embodies the concept of “collective wisdom”—while individual models may be prone to errors, the collective prediction of multiple models is more reliable. Random Forests perform exceptionally well across various datasets, especially when the feature space is large. Compared to traditional BP neural networks and some deep learning models, Random Forests achieve greater robustness through collective predictions by building multiple decision trees. This makes them highly effective in situations involving large fluctuations in steam load in the workshop and relatively small datasets, ensuring high-performance predictions.

3.4 Correlation analysis

The Random Forest model not only solves classification or regression problems but also evaluates the importance of each factor in the dataset. The Out-of-Bag (OOB) dataset is used in the Random Forest algorithm to assess model performance and feature importance. OOB refers to the data that are not included in the training set due to the bootstrap sampling process, where approximately one-third of the data are left out. These OOB samples serve as a built-in validation set, allowing model evaluation without additional computational costs. The concept of OOBPredictorImportance is introduced in the model to assess the importance of each predictor (feature) using the OOB dataset. Each decision tree in the Random Forest is independently built, with the splitting features at each node selected from a randomly chosen subset of features. This ensures that each feature is evaluated a varying number of times across different trees, providing a natural way to compare feature importance. The importance of a feature is determined by calculating the impurity reduction (for classification) or error reduction (for regression) it achieves when splitting nodes across all trees. The cumulative contribution of each feature is averaged or weighted to obtain an overall importance score. The higher the importance score, the greater the impact of that feature on the model’s predictions. Using this method, the importance of various factors affecting the steam load in the primary processing workshop is determined, with the results presented in Table 2.

Analysis of the Table 2 shows that the factors significantly affecting the steam load in the primary processing workshop are the Vacuum Area, Leaf Blade Area, Stem Area (Before stem storage), and Stem Area (After stem storage), with correlation coefficients all above 0.5, indicating strong correlations. The Leaf Strip Areas (Phase I and Phase II) exhibit weak correlations with steam load, with coefficients around 0.4. The Blending Areas (Phase I and Phase II) have correlation coefficients below 0.2, indicating very low correlation. The low correlation in the blending areas can be explained by the fact that only the blending and flavoring machines use steam in this area, and the amount of steam used is minimal, making it negligible. When building the predictive model, it is important to select highly correlated factors as input features. The more comprehensive and relevant the input features, the higher the model's prediction accuracy. However, including too many features may lead to overfitting, negatively affecting the model's performance. The correlation analysis provides a reference for selecting input features [18]. For this study, factors with correlation coefficients above 0.5—namely, the Vacuum Area, Leaf Blade Area, Stem Area (Before stem storage), and Stem Area (After stem storage)—are chosen as input features for the prediction model.

Table 2. Correlation between influencing factors and steam load in the primary processing workshop

3.5 Hyperparameter selection

For the Random Forest model, the tree depth (dp) and number of trees (es) are generally considered the two most important hyperparameters. In this paper, the primary focus is on selecting the maximum tree depth and the number of regression trees, while other parameters are kept at their default values. The maximum depth of the tree is a crucial hyperparameter for both decision trees and Random Forest models. If the maximum depth is set too shallow, the model may fail to capture the complex relationships in the data, leading to underfitting. Conversely, if the depth is too large, the model might overfit the noise in the training data, reducing its generalization ability. Common methods for determining the optimal tree depth include cross-validation and grid search. In this study, cross-validation was employed to determine the optimal maximum depth, which was found to be 4.

3.jpg

Figure 3. Impact of the number of trees on prediction accuracy

The number of trees in a Random Forest refers to the number of decision trees within the model. Increasing the number of trees typically improves the model's performance, as more trees can better capture the features of the data and reduce variance. However, the performance gains are not linear, and after a certain point, additional trees offer diminishing returns. Too many trees may also lead to overfitting, especially when the dataset is relatively small. The Figure 3 shows the change in prediction error with different numbers of decision trees, using the OOB error as the evaluation metric. As illustrated in Figure 3 and Table 3, the prediction error decreases as the number of trees increases. The error reaches 0.01755 when the number of trees reaches 30, and further increases in the number of trees provide no significant improvement in prediction accuracy. To avoid overfitting, the optimal number of trees was set to 30.

Table 3. Errors for different numbers of decision trees

This paper analyzed the energy consumption of the steam system in a tobacco manufacturing plant's primary processing workshop, applying emerging technologies such as artificial intelligence and machine learning to predict daily steam consumption. Initially, the OOBPredictorImportance concept was introduced in the prediction model to assess the importance of each feature using the OOB dataset. Factors that significantly affect the steam load in the processing workshop were identified, including the vacuum area, leaf blade area, stem area (before storage), and stem area (after storage). Subsequently, the Random Forest method was employed to predict the daily steam load and compared with traditional BP neural network, SVM, and decision tree. The results demonstrate that the Random Forest model outperformed the others in both prediction accuracy and fitting effectiveness, providing a reliable basis for optimizing the scheduling of the enterprise's steam system. While the prediction results of the model generally achieved satisfactory outcomes, it is noted that the data exhibited significant fluctuations and the sample size was relatively small, which may limit the feature selection process. Therefore, future research should focus on exploring additional influencing factors and enriching the dataset to optimize the model and further enhance prediction accuracy, facilitating the regulation of steam supply and achieving energy-saving and emission-reduction goals.