A forecasting method for non-equal interval time series based on recurrent neural network

Unknown data can be forecast by learning the patterns of change from the historical data at regular intervals. However, when samples are not available at a regular interval, the forecasting task becomes very challenging. This paper proposes an improved Gated Recurrent Unit model for non-equal interval time series, abbreviated as NITS-GRU, to model the data of the series without resampling and data imputation. NITS-GRU includes GRU-ODE module, multivariate-based sample correlation calculation module, and information fusion module. The GRU-ODE module generates forecast information of adjacent samples by calculating their hidden layer states, as well as the accumulated difference information between each adjacent sample and forecasting sample over time intervals. Meanwhile, we calculate the correlations between each adjacent sample and forecasting sample based on multiple variables, and adopt attention mechanism to obtain correlation weights of these adjacent samples. The information fusion module applies the correlation weights on forecast information of the adjacent samples generated by the GRU-ODE module, to obtain the final forecast results. Experiments on different datasets demonstrate that NITS-GRU outperforms the state-of-the-art baselines for forecasting unknown data in non-equal interval time series.