A Novel LSTM Approach for Asynchronous Multivariate Time Series Prediction

Long-short-term-memory (LSTM) recurrent neural networks have difficulty in representing temporal and non-temporal inputs simultaneously, due to the sequential emphasis of the architecture. This limits the LSTM applicability in settings where multivariate data is difficult to align. In this paper, a modified hierarchical approach is proposed where a set of univariate LSTM's is trained for asynchronous temporal sequences. The resulting representation is jointly trained with an encoding of multivariate input for prediction. The approach is generalized to combine with non-sequential inputs. The proposed architecture is verified experimentally to improve prediction performance and convergence over conventional LSTM interpolation approaches on simulated Lorenz data and pipeline flow prediction.