Implementation:Sktime Pytorch forecasting DLinear V2

Knowledge Sources	Sktime_Pytorch_forecasting
Domains	Time_Series, Forecasting, Deep_Learning
Last Updated	2026-02-08 08:00 GMT

Overview

DLinear is a decomposition-linear model for long-term time series forecasting that separates trend and seasonal components and processes each with simple linear layers.

Description

DLinear extends TslibBaseModel and implements the LTSF-DLinear architecture from the Time-Series-Library. The model decomposes input time series into trend and seasonal components using a moving average filter (SeriesDecomposition), then applies separate linear layers to each component. The final prediction is the sum of the trend and seasonal projections. This is an experimental V2 implementation built on the TslibBaseModel pipeline and supports both shared and individual per-variate linear layers, as well as quantile loss for probabilistic forecasting.

Usage

Use DLinear when you need a simple, fast, and interpretable baseline for long-term time series forecasting. It is particularly effective when the underlying patterns can be well captured by linear transformations of trend and seasonal components. Note that this is an experimental V2 implementation and its API may change.

Code Reference

Source Location

• Repository: Sktime_Pytorch_forecasting
• File: pytorch_forecasting/models/dlinear/_dlinear_v2.py
• Lines: 1-307

Signature

class DLinear(TslibBaseModel):
    def __init__(
        self,
        loss: nn.Module,
        moving_avg: int = 25,
        individual: bool = False,
        logging_metrics: list[nn.Module] | None = None,
        optimizer: Optimizer | str | None = "adam",
        optimizer_params: dict | None = None,
        lr_scheduler: str | None = None,
        lr_scheduler_params: dict | None = None,
        metadata: dict | None = None,
        **kwargs: Any,
    ):

Import

from pytorch_forecasting.models.dlinear._dlinear_v2 import DLinear

I/O Contract

Inputs

Name	Type	Required	Description
loss	nn.Module	Yes	Loss function for training (e.g. MSELoss, QuantileLoss).
moving_avg	int	No	Kernel size for moving average decomposition. Defaults to 25.
individual	bool	No	Whether to use individual linear layers per variate (True) or shared (False). Defaults to False.
logging_metrics	None	No	Metrics to log during training. Defaults to None.
optimizer	str | None	No	Optimizer for training. Defaults to "adam".
optimizer_params	None	No	Parameters for the optimizer. Defaults to None.
lr_scheduler	None	No	Learning rate scheduler name. Defaults to None.
lr_scheduler_params	None	No	Parameters for the LR scheduler. Defaults to None.
metadata	None	No	Metadata from TslibDataModule with dataset shape information. Defaults to None.

Outputs

Name	Type	Description
prediction	torch.Tensor	Forecast output of shape (batch_size, prediction_length, target_dim) or (batch_size, prediction_length, n_quantiles) when using QuantileLoss.

Usage Examples

from pytorch_forecasting.models.dlinear._dlinear_v2 import DLinear
from pytorch_forecasting.metrics import QuantileLoss
import torch.nn as nn

# Instantiate with MSE loss
model = DLinear(
    loss=nn.MSELoss(),
    moving_avg=25,
    individual=False,
    optimizer="adam",
    optimizer_params={"lr": 1e-3},
    metadata=metadata,
)

# Instantiate with quantile loss for probabilistic forecasting
model = DLinear(
    loss=QuantileLoss(),
    moving_avg=25,
    individual=True,
    metadata=metadata,
)

# Forward pass
output = model(x_batch)
predictions = output["prediction"]

Related Pages

• Principle:Sktime_Pytorch_forecasting_Decomposition_Linear
• Sktime_Pytorch_forecasting_TFT_V2 - Alternative V2 model with attention mechanisms