Multivariate time series forecasting is a task that aims to derive key seasonality from past patterns to predict future time series. It has been studied in various domains, such as the financial markets and electricity consumption. Besides this, time ...
Multivariate time series forecasting is a task that aims to derive key seasonality from past patterns to predict future time series. It has been studied in various domains, such as the financial markets and electricity consumption. Besides this, time series forecasting can be divided into one-step prediction and multi-step prediction. In the industrial sector, multi-step prediction is crucial because the continuous perspective leads to better decision-making. However, multi-step prediction is highly unstable because of its dependency on the previous prediction values. To solve this problem, we propose a novel encoder-decoder model based on the neural network for multivariate multi-step time series forecasting.
In our study, we propose a denoising training method that allows us to take advantage of both Teacher Forcing (TF) and without Teacher Forcing (w/o TF) methods. Moreover, we propose dual attention. One is for the time dependency between the encoder and the decoder, and the other is for the time dependency in the decoder time steps. Moreover, we stack the dual attention to stabilize the long-term dependency and multi-step prediction problem. We add positional encoding to provide the model with time information. Using this, we can more accurately recognize the seasonality in the model. Because our method is based on a non-linear neural network model, we add an autoregression component to resolve the lack of linear properties. Unlike the typical autoregression model, we propose skip-autoregression for multi-seasonality. Moreover, we adopt multi-head fully connected layers for variable-specific modeling because of our multivariate time series data.
This study performs multi-step prediction of time series with complex seasonality through the modeling method described above. To verify our approach, we compare our model with machine learning and deep learning models. Moreover, to demonstrate our study’s performance, we conduct various experiments such as an ablation study, a denoising intensity test, and a stack attention test.