一种自适应补丁的长时序预测模型OA

Multivariate time series forecasting has been widely used in various real-world scenarios,such as weather forecasting,fault detection in industrial equipment,and traffic flow prediction.Currently,transformer-based models have shown great potential in time series prediction tasks.However,when the original data contains complex periodic patterns,transformer models struggle to accurately capture the local periodic patterns of the raw data,leading to the loss of temporal information.To address this problem,a long-term time series prediction model is proposed,which is based on adaptive patches.First,a patch embedding module with adaptive length is designed to transform 1D time series data into 2D time series data based on the periodic characteristics of the dataset.Second,multilayer perceptrons are utilized to capture the temporal dependencies within and between patches of the 2D time series data,achieving the capture of both local and global temporal features.Finally,by combining a channel-independent strategy,adaptive length patch embedding,and a time dependency capture module,the problems of extracting multiple periodic patterns and preventing the loss of temporal information during the training process are effectively solved.Through experiments on seven real-world datasets,the accuracy and effectiveness of the proposed model for prediction have been verified.