Abstract
The power load fluctuates sharply due to the uncertain shift of operating modes in the electrical smelting process. Forecasting the shift accurately can help implement peak shaving and valley filling, thus saving power supply capacity and reducing the electricity bill for enterprises. Each of the operating modes has different and uncertain durations, the forecasting of which has some difficulties such as multi-labels, class-imbalance, and sparse shifted data. To address this problem, it proposes a two-layer forecasting model based on Belief Rule Base (BRB) and eXtreme Gradient Boosting (XGBoost). It mines the coarse-grained knowledge of the cyclic operating mode by using BRB, which makes a binary decision whether the mode is shifted or not. The decision is regarded as a new feature and its value is added into each sample of data set. An adaptive synthetic sampling approach is utilized to process and rebalance the training set, with which a XGBoost model is trained to obtain the fine-grained multi-label classification. The field data of an electro-fused magnesium enterprise was used, and compared with other machine learning methods, our model obtained more satisfactory results in terms of prediction accuracy, precision, recall, and F1 score, indicating that the new prediction classification method is reasonable and effective.

This publication has 19 references indexed in Scilit: