In the big data era, learning-based techniques have attracted more and more attention in many industry areas such as smart grid, intelligent transportation. The power load forecasting is one of the most critical issues in data analysis of smart grid. However, learning-based methods have not been widely used due to the poor data quality and computational capacity. In this paper, we propose a comprehensive learning-based model to forecast heavy and over load (HOL) accidents according to the data from various information systems. At first, we present a combined random under- and over-sampling technique for imbalanced electric data, and choose an optimal sampling rate through several experiments. Then, we reduce the attributes that have significant impact on the power load by using learning-based methods. Finally, we provide an algorithm based on the random forest method to prevent the over-fitting problem. We evaluate the proposed model and algorithms with the real-world data provided by China Grid. The experimental results show that our model works efficiently and achieves low error rates.