A CNN Based Bagging Learning Approach to Short-Term Load Forecasting in Smart Grid

Short-term load forecasting in smart grid is key to electricity dispatch scheduling, reliability analysis, and maintenance planning for the generators. In this paper, we present a convolutional neural networks (CNN) based bagging model for forecasting hourly loads. We employ CNN to train forecasting models on big load data sets. Then, we segment a real industry load data set into many subsets, fine-tune the forecasting models on these subsets to learn weak forecasting models, and assemble these weak forecasting models to conduct a bagging forecasting model, where the learning and assembling procedures are implemented on Spark. Specifically, all load samples in those data sets are reorganized as images with respect to similarities between relations of pixels in images and those of features in load samples. Experimental results indicate the effectiveness of the proposed method.