关键词: 时间序列预测, 序列分解, 膨胀卷积, 编码器, 解码器, 注意力机制

Abstract: Addressing challenges in time series prediction tasks, such as high-dimensional features, large-scale data, and the demand for high prediction accuracy, a multi-scale trend-period decomposition model based on a multi-head gated dilated convolutional neural network is proposed. A multi-scale decomposition approach is employed by the model to separate the original covariate sequence and the prediction variable sequence into their respective trend and period components, thereby enabling independent prediction. For the period components, an introduction is made of a multi-head gated dilated convolutional neural network encoder to extract individual period information. In the decoder stage, channel information interaction and fusion are achieved through the utilization of a crossattention mechanism. The combined sampled and aligned period information of the prediction variable is achieved through time attention and channel fusion to enable period prediction. Regarding trend components, trend prediction is executed using an autoregressive approach, and the final prediction sequence is obtained by incorporating the trend prediction results with the period predictions. Compared with seven mainstream benchmark models such as Long Short-Term Memory (LSTM), Informer, etc., a reduction in Mean Squared Error (MSE) is observed, ranging from 19.2% to 52.8% across five datasets including ETTm1, ETTh1, etc. Additionally, a decrease in Mean Absolute Error (MAE) is noted, ranging from 12.1% to 33.8% on the mentioned datasets. The ablation experiments confirmed that the multi-scale decomposition module, multi-head gated dilation mechanism, and time attention module proposed enhance the accuracy of time series prediction.

Key words: Time Series Prediction, Sequence Decomposition, Dilated Convolution, Encoder, Decoder, Attention Mechanism