Short-term Gas Load Prediction with STL and Hybrid Neural Network

Abstract

Accurate prediction of gas load is critically important to ensure stable gas usage and accurate dispatch. In the literature currently available, the extensively used model-based methods are limited by their low prediction accuracy. In addition, data-driven approaches have low prediction accuracy because they are difficult to extract time series features directly from the original data. To resolve the above issues, a seasonal trend decomposition procedure based on loess (STL) has been developed to realize short-term gas load prediction with a hybrid neural network (HNN). Firstly, the original time series is divided into seasonal, trend, and residual components using STL. Especially, the trend is fitted with the least squares method, then both the new trend and residual components can be obtained. Secondly, the hybrid neural network is composed of a one-dimensional convolutional neutral network, a bidirectional gated recurrent unit, and a dilate gated recurrent unit, which are foused to extract temporal features for new residual component prediction. Finally, gas load prediction experiments are chosen to verify that STL-HNN outperforms the state-of-the-art methods.