Dual-channel interpretable time series prediction model and its applications

Channel-mixing strategies are used to capture complex patterns by incorporating information across the channels. However, this strategy may distract the predictive model’s focus on the relevant features, resulting in low prediction accuracy. Additionally, it may cause the model to have low tolerance to noise and a high risk of overfitting. A Dual-channel interpretable time series prediction model is proposed to improve the prediction accuracy and interpretability of time series forecasting. Latent space vectors are extracted from target sequence and exogenous sequences using a dual-channel strategy. Exogenous channel captures features that influence changes in the target sequence through a simple and efficient linear model, while the target variable, with significant periodic characteristics, is captured temporal dependencies by GRU. The encoder segments the target sequence to reduce the number of iterations. An interpretable block (InBlock) translates the latent space vectors into seasonal, trend and exogenous components. Multiple InBlocks is adopted to enhance the model’s adaptability to fine-grained changes in the predicted sequence. Extensive experiments are conducted on five public electricity price datasets and a wastewater treatment dataset. The experimental results demonstrate an improvement in prediction accuracy with the DC-ITSPM compared to contemporary popular models.