Multiscale deep network based multistep prediction of high‐dimensional time series from power transmission systems

Abstract

AbstractInternet of energy makes the future power and energy network a more complicated and intelligent system. With the development of energy industry, the sample data of such system is high dimensional, dynamic, correlative, and complex. In order to meet people's needs and reduce the power redundancy, predicting the future energy demand and production is an essential approach. It is necessary for us to predict the later hours' or days' data, which means multistep prediction. However, the common one‐step prediction model cannot forecast the power demand or production to make adequate preparation and the data have thousands of dimensions, which makes the problem challenging. In addition, the changeable pattern makes the common prediction algorithm do not perform good enough. In this article, we propose a sequence to sequence model to make multistep prediction with a baseline mean squared error (MSE) of 1.49×10−5. In addition, we improve the model to be a multiscale deep network and decrease the MSE to 1.23×10−5 through adding extra information to match different patterns. Furthermore, the multitask learning trick makes the MSE decrease to 1.18×10−5.