DLinear

DLinear conducts direct multi-step forecasting by decomposing the time series into a trend and a remainder series and employs two one-layer linear networks to model these two series for the forecasting task. Then, two one-layer linear networks are applied to the two series.

Figure 1. Overall architecture of DLinear model.

Although DLinear is simple, it has some compelling characteristics:

• An $O(1)$ maximum signal traversing path length: The shorter the path, the better the dependencies are captured [18], making DLinear capable of capturing both short-range and long-range temporal relations.
• High-efficiency: As each branch has only one linear layer, it costs much lower memory and fewer parameters and has a faster inference speed than existing Transformers.
• Interpretability: After training, we can visualize weights from the seasonality and trend branches to have some insights on the predicted values [8].
• Easy-to-use: DLinear can be obtained easily without tuning model hyper-parameters.

More information can be found in: https://arxiv.org/abs/2205.13504

Use cases

1. Electricity demand dataset
2. Air-Quality dataset (CO & NO$_2$)

Get Started

1. Install Python >= 3.7
2. Install requirements.txt
3. Run DLinear - UC:AirQuality.ipynb or DLinear - UC:Electricity.ipynb

Requirements

• tqdm==4.62.3
• matplotlib==3.5.1
• numpy==1.21.2
• torch==1.7.0
• pandas==1.3.5
• scikit-learn==1.0.2

Versions

• Version 1.0
  • Basic implementation
  • Use cases: Electricity demand & Air-Quality forecasting
  • Evaluation per value of forecasting horizon
  • Model Evaluation
    • Performance evaluation metrics: MAE, RMSE, MAPE, R2
    • Examine AutoCorrelation
  • Examples