Multiscale convolutional neural-based transformer network for time series prediction

Time series prediction is tough resulting from the lack of multiple time-scale dependencies and the correlation among input concomitant variables. A novel method has been developed for time series prediction by leveraging a multiscale convolutional neural-based transformer network (MCTNet). It is composed of multiscale extraction (ME) and multidimensional fusion (MF) frameworks. The original ME has been designed to mine different time-scale dependencies. It contains a multiscale convolutional feature extractor and a temporal attention-based representator, following a transformer encoder layer for high-dimensional encoding representation. In order to use the correlation among variables sufficiently, a novel MF framework has been designed to capture the relationship among inputs by utilizing a spatial attention-based highway mechanism. The linear elements of the input sequence are effectively preserved in MF, which helps MCTNet make more efficient predictions. Experimental results show that MCTNet has excellent performance for time series prediction in comparison with some state-of-the-art approaches on challenging datasets.