Sequence to Image Transform Based Convolutional Neural Network for Load Forecasting

Convolutional neural networks (CNNs) are known as powerful tools for image processing. Although some works have used the CNN for processing of sequences such as time series, but they usually apply this type of data in the form of sequence that is not consistent with the CNN nature which receives inputs in the image (matrix) form. To deal with this problem, sequence to image transform based CNN (STI-CNN) is proposed in this work which transforms the load sequence to several images and feed them to the CNN. The proposed STI-CNN method is used for load forecasting. Transforming load sequence to load images results in some advantages. The main profit is that the lagged load values of load are located in a two-dimensional grid and CNN can extract informative features from the neighboring load variables. While in the sequence form, each load value just has two neighbors, each load value has 8 neighbors in the image form. The experiments implemented on the ISSDA dataset (an electrical load data from Ireland) show the superior performance of STI-CNN in terms of different forecasting measures.