Implementation:Sktime Pytorch forecasting NBeatsKAN

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

Overview

NBeatsKAN is an N-BEATS variant that replaces standard MLP layers with Kolmogorov-Arnold Network (KAN) layers for improved accuracy and interpretability.

Description

NBeatsKAN extends NBeatsAdapter and implements the N-BEATS architecture with KAN-augmented blocks based on the Kolmogorov-Arnold representation theorem. Instead of using fixed MLP edge weights, the KAN blocks use learnable univariate spline functions on each edge, allowing the network to better capture complex temporal patterns with improved parameter efficiency. The model supports three stack types -- "generic", "trend", and "seasonality" -- using NBEATSGenericBlockKAN, NBEATSTrendBlockKAN, and NBEATSSeasonalBlockKAN respectively. It also includes a update_kan_grid method for updating the KAN spline grids during training.

Usage

Use NBeatsKAN when you want the interpretable structure of N-BEATS (trend/seasonality decomposition) combined with the enhanced expressiveness of KAN layers. This model is particularly useful for univariate time series forecasting where interpretability and parameter efficiency are important, or for zero-shot cross-domain transfer learning scenarios.

Code Reference

Source Location

• Repository: Sktime_Pytorch_forecasting
• File: pytorch_forecasting/models/nbeats/_nbeatskan.py
• Lines: 1-287

Signature

class NBeatsKAN(NBeatsAdapter):
    def __init__(
        self,
        stack_types: list[str] | None = None,
        num_blocks: list[int] | None = None,
        num_block_layers: list[int] | None = None,
        widths: list[int] | None = None,
        sharing: list[bool] | None = None,
        expansion_coefficient_lengths: list[int] | None = None,
        prediction_length: int = 1,
        context_length: int = 1,
        dropout: float = 0.1,
        learning_rate: float = 1e-2,
        log_interval: int = -1,
        log_gradient_flow: bool = False,
        log_val_interval: int = None,
        weight_decay: float = 1e-3,
        loss: MultiHorizonMetric = None,
        reduce_on_plateau_patience: int = 1000,
        backcast_loss_ratio: float = 0.0,
        logging_metrics: nn.ModuleList = None,
        num: int = 5,
        k: int = 3,
        noise_scale: float = 0.5,
        scale_base_mu: float = 0.0,
        scale_base_sigma: float = 1.0,
        scale_sp: float = 1.0,
        base_fun: callable = None,
        grid_eps: float = 0.02,
        grid_range: list[int] = None,
        sp_trainable: bool = True,
        sb_trainable: bool = True,
        sparse_init: bool = False,
        **kwargs,
    ):

Import

from pytorch_forecasting.models.nbeats import NBeatsKAN

I/O Contract

Inputs

Name	Type	Required	Description
stack_types	None	No	Stack types: "generic", "trend", or "seasonality". Defaults to ["trend", "seasonality"].
num_blocks	None	No	Number of blocks per stack. Defaults to [3, 3].
num_block_layers	None	No	FC layers per block. Defaults to [3, 3].
widths	None	No	FC layer widths per stack. Defaults to [32, 512].
sharing	None	No	Weight sharing per stack. Defaults to [True, True].
expansion_coefficient_lengths	None	No	Expansion coefficient or polynomial degree per stack. Defaults to [3, 7].
prediction_length	int	No	Forecast horizon. Defaults to 1.
context_length	int	No	Lookback period. Defaults to 1.
dropout	float	No	Dropout rate. Defaults to 0.1.
loss	MultiHorizonMetric	No	Loss function. Defaults to MASE().
backcast_loss_ratio	float	No	Weight of backcast loss. Defaults to 0.0.
num	int	No	KAN grid intervals (G). Defaults to 5.
k	int	No	KAN piecewise polynomial order. Defaults to 3.
noise_scale	float	No	KAN initialization noise scale. Defaults to 0.5.
scale_base_mu	float	No	KAN residual function mean. Defaults to 0.0.
scale_base_sigma	float	No	KAN residual function sigma. Defaults to 1.0.
scale_sp	float	No	KAN spline base function scale. Defaults to 1.0.
base_fun	callable	No	KAN residual function. Defaults to torch.nn.SiLU().
grid_eps	float	No	KAN grid interpolation parameter (0=percentile, 1=uniform). Defaults to 0.02.
grid_range	list[int]	No	KAN grid range. Defaults to [-1, 1].
sp_trainable	bool	No	Whether scale_sp is trainable. Defaults to True.
sb_trainable	bool	No	Whether scale_base is trainable. Defaults to True.
sparse_init	bool	No	Whether to use sparse initialization. Defaults to False.

Outputs

Name	Type	Description
prediction	torch.Tensor	Forecast output in target space.
backcast	torch.Tensor	Backcast reconstruction (when backcast_loss_ratio > 0).

Usage Examples

from pytorch_forecasting.models.nbeats import NBeatsKAN

# Interpretable mode with trend and seasonality stacks
model = NBeatsKAN(
    stack_types=["trend", "seasonality"],
    num_blocks=[3, 3],
    num_block_layers=[3, 3],
    widths=[32, 512],
    expansion_coefficient_lengths=[3, 7],
    prediction_length=24,
    context_length=96,
    dropout=0.1,
    num=5,    # KAN grid intervals
    k=3,      # KAN polynomial order
)

# Generic mode
model = NBeatsKAN(
    stack_types=["generic"],
    num_blocks=[1],
    num_block_layers=[4],
    widths=[512],
    expansion_coefficient_lengths=[32],
    prediction_length=24,
    context_length=96,
)

# Update KAN grids during training
model.update_kan_grid()

Related Pages

• Principle:Sktime_Pytorch_forecasting_KAN_Architecture
• Sktime_Pytorch_forecasting_NHiTS - Related hierarchical interpolation model that outperforms standard N-BEATS