.png){kind=icon}
.png){kind=icon}
Return
An Effective Summary Preprocessing Method for Time Series Forecasting With Multiple Temporal Granularities
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2025.3553154.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Time-series forecasting plays a pivotal role in decision-making. Recently, as deep learning models have shown exceptional performance in time-series forecasting, research in the field of time-series forecasting based on deep learning has been actively conducted. Deep learning models use a lot of past time-series data as a look-back window to effectively capture temporal information. However, increasing the size of a look-back window increases the resource demands, such as training time, GPU usage, and memory allocation, posing limitations to its extension. In response, we averaged the time-series data into summaries having multiple time granularities to accommodate a large length of time-series data in a look-back window. These constructed look-back windows have a similar effect as if the size of the look-back window were increased, even though an equal length of the look-back window was given. We employed six datasets and three time-series forecasting models to demonstrate the effectiveness of the proposed summary preprocessing method. Finally, the experimental results demonstrated an 11% improvement in forecasting accuracy along with a significant reduction in training time by 77% on average.
Keywords:
Forecasting
Time series analysis
Data models
Transformers
Predictive models
Training
Indexes
Computer architecture
Computational modeling
Arrays
Time series data
time series forecasting
summary preprocessing
transformer
look-back window
neural prediction
PatchTST