Multi-Scale Dilated Convolution Network for Long-Term Time Series Forecasting

TL;DR

The paper tackles long-term time series forecasting by introducing MSDCN, a CNN-based framework that uses shallow, multi-scale dilated convolutions to capture long-range dependencies while preserving efficiency. It combines two parallel convolution branches with exponentially growing dilation and varying kernel sizes, fused with learnable weights, and complements them with an autoregressive module to model linear dynamics, with the final prediction as the sum of both paths. Empirical results on eight benchmark datasets show state-of-the-art accuracy for both multivariate and univariate forecasting and substantial speed advantages over Transformer-, CNN-, and MLP-based baselines. The approach yields robust performance across varying input lengths and demonstrates practical viability for real-time forecasting tasks by maintaining high accuracy with reduced computational cost.

Abstract

Accurate forecasting of long-term time series has important applications for decision making and planning. However, it remains challenging to capture the long-term dependencies in time series data. To better extract long-term dependencies, We propose Multi Scale Dilated Convolution Network (MSDCN), a method that utilizes a shallow dilated convolution architecture to capture the period and trend characteristics of long time series. We design different convolution blocks with exponentially growing dilations and varying kernel sizes to sample time series data at different scales. Furthermore, we utilize traditional autoregressive model to capture the linear relationships within the data. To validate the effectiveness of the proposed approach, we conduct experiments on eight challenging long-term time series forecasting benchmark datasets. The experimental results show that our approach outperforms the prior state-of-the-art approaches and shows significant inference speed improvements compared to several strong baseline methods.

Figures (5)

• Figure 1: A time series can be considered as the linear or nonlinear superposition of components such as weekly trends, monthly trends, and yearly trends.
• Figure 2: The overall architecture of MSDCN. The green box represents convolution module, and the yellow box represents autoregressive module.
• Figure 3: Ablation results for the autoregressive module in MSDCN.
• Figure 4: Visualization of different convolutional block output representation. The prediction length is 336.
• Figure 5: The MSE performance when input length is increasing. The prediction sequence length is 24 or 720.