SDCNet 周期性
While some deep learning models discover dependencies in decomposed time series, they are not good at capturing local dynamics and long-term dependencies from time series without obvious periodicity. LSTM-MSNet [10] uses the traditional time series decomposition methods as a data pre-processing technique. The Auto-Correlation mechanism [21] based on traditional mathematical methods is more advantageous in capturing seasonal patterns that do not vary with time. However, the Auto-Correlation does not capture local time dynamics well for real-world time series with a mixture of local dynamics and long-term patterns. For time series without significant periodicity, the Auto-Correlation cannot take advantage of its merits. In addition, capturing complex combinations of time series related to forecasting series is beneficial for improving forecast accuracy [13]. Also, noise in the time series and some exogenous series with weak correlation with the forecast series may impair the forecast performance [22]. However, LSTM-MSNet and Autoformer [21] do not capture inter-time series dependencies.
To this end, we design a novel network, namely SDCNet, to improve forecasting capability, memory utility, and computational efficiency. Different from previous approaches, SDCNet uses the time series decomposition method to untangle the entangled temporal patterns and employs CNNs to extract the dependencies in the temporal and feature dimensions. SDCNet uses the seasonal and trend components of the time series to generate forecasts. In terms of model design, SDCNet is a forecasting model built entirely with CNNs, thus its better ability to capture multi-scale dependencies. Our contributions are summarized as follows:
- •
We propose a concise and efficient model, SDCNet, to improve the performance of long-term forecasting of multivariate time series. The SDCNet framework contains two types of components: CNNs and time series decomposition components. This framework validates that CNNs can effectively extract seasonal and trend patterns in time series and that combining time series decomposition with CNNs can further improve forecast accuracy.
- •
We propose temporal convolution to capture both local and long-term temporal patterns. CNNs endow SDCNet with local context awareness to reduce the negative impact of outliers. First, temporal convolution discovers and refines temporal repeating patterns at multiple scales. Then, the time series decomposition module gradually decomposes more predictable seasonal and trend-cyclical components from the time series refined by the time convolution component.
- •
We propose feature convolution to capture the complex combinatorial patterns of the time series related to the predicted time series. Feature convolution has two functions. The first is to eliminate the negative impact of uncorrelated or weakly correlated exogenous series. The second is to capture the dependencies between time series and the complex combinations between correlated time series.
- •
We evaluate the performance of SDCNet on four datasets, two with significant periodicity and two without obvious periodicity. Experimental results show that SDCNet outperforms the state-of-the-art methods with a relative performance improvement of 16.73%, as well as better memory usage and computational efficiency.
InParformer: Evolutionary Decomposition Transformers ...
