DESCINet: A hierarchical deep convolutional neural network with skip connection for long time series forecasting

Time series forecasting is the process of predicting future values of a time series from knowledge of its past data. Although there are several models for making short-term predictions, the problem of long temporal sequences can still receive new contributions. Recent studies have applied Transformers-based solutions to the long-time series forecasting task and achieved good results. When it comes to time series modeling, the goal is to capture the temporal relationships within a sequence of ordered and continuous points. While using positional encoding and embedding sub-series as tokens in Transformers can help to preserve some ordering information, the permutation-invariant self-attention mechanism used in these models can lead to loss of temporal information. In this paper, we used convolutional networks in a binary tree structure with skip connections between the levels of the tree that allowed greater precision and efficiency in the training. The proposed model was evaluated on five real-life datasets. Experimental results show that our model significantly improves forecast accuracy relative to existing solutions.