A Long Time Series Prediction Model Based on CNN and Linear Combination

In recent years, Transformer and CNN models have dominated the landscape of time series forecasting tasks. Addressing the issues of Transformer’s high computational resource demands and linear models’ limited ability to capture complex features, this paper proposes a hybrid model combining linear layers with CNNs to tackle such challenges. Our model leverages both convolutional operations and dilated convolutions in synergy to learn both local and global information from the data while enhancing local learning with the aid of a linear framework. It employs a dual-learning approach that integrates two CNNs of varying depths with linear layers, strengthening the model’s capability to learn both linear and nonlinear patterns from the data. Experimental results on seven benchmark time series forecasting datasets demonstrate that, compared to state-of-theart (SOTA) methods, our model not only ensures comparable accuracy but also reduces model complexity, thereby enhancing its runtime performance relative to SOTA models.