LightTimesNet — Efficient 2D Time-Series Forecasting

LightTimesNet is a lightweight redesign of TimesNet for efficient multi-horizon time-series forecasting. The project investigates whether the dynamic FFT-based period extraction used by TimesNet can be replaced by domain-knowledge-driven static spatial patching, reducing computational overhead while maintaining competitive forecasting accuracy.

Key Idea

TimesNet dynamically extracts dominant periods using FFT and reshapes 1D time series into 2D representations. LightTimesNet replaces this process with a deterministic [frequency, period] reshape based on known periodic structure. This enables:

• fully vectorized tensor construction
• reduced period-extraction overhead
• simpler hardware execution
• lightweight spatial backbones
• improved accuracy–efficiency trade-offs

Models

Model	Representation
DLinear	1D
CausalTCN	1D
TimesNet	Dynamic 2D
Fixed-Period TimesNet	Static 2D
LightTimesNet	Static 2D
LightTimesNet + Depthwise	Static 2D

Datasets

Experiments use standard forecasting benchmarks:

Dataset	Frequency	Horizons
ETTh1	1-hour	24, 48, 96
ETTm1	15-minute	24, 48, 96
Electricity	1-hour	24, 48, 96

Results

The experiments demonstrate that deterministic period folding can substantially reduce the computational cost of TimesNet while retaining competitive forecasting performance.

Comparison	Accuracy	Parameters	Training
ETTh1 H=24 — DLinear	MSE 0.337	4.7K	4.68 s
ETTh1 H=24 — LightTimesNet-DW	MSE 0.326	4.2K	10.15 s
Electricity H=48 — TimesNet	MSE 0.293	169K	51.9 s
Electricity H=48 — LightTimesNet-Group	MSE 0.302	28.9K (-83%)	20.7 s (-60%)

The ETTh1 experiment shows that LightTimesNet-DW can slightly outperform a strong lightweight 1D baseline while using fewer parameters.

On Electricity, the Group-based LightTimesNet variant reduces the model size by approximately 83% and training time by 60% compared with TimesNet, at the cost of only a 3.1% increase in MSE. Demystifying_2D_Time_Series_Forecasting__A_Targeted_Redesign_of_TimesNet_for_High_Frequency_Data.pdf

The optimal lightweight backbone is dataset-dependent: Depthwise convolution performs best on ETTh1, whereas Group convolution provides a stronger accuracy–complexity compromise on Electricity. Demystifying_2D_Time_Series_Forecasting__A_Targeted_Redesign_of_TimesNet_for_High_Frequency_Data.pdf

Analysis

The repository includes experiments on:

• dynamic vs. fixed-period modeling
• 1D vs. 2D temporal representations
• top-k period selection
• number of TimesBlocks
• alternative spatial backbones
• depthwise convolutions
• parameter count and computational cost
• accuracy–efficiency Pareto analysis

Getting Started

Requirements

• Python 3.9+
• PyTorch
• NumPy
• Pandas
• PyYAML
• Matplotlib

Installation

pip install -r requirements.txt

Run an Experiment

python run_pipeline.py --config configs/etth1_24.yaml

Other experiment configurations are available under configs/.

Repository Structure

.
├── configs/              # Experiment configurations
├── src/
│   ├── models_1d.py      # 1D baselines
│   ├── models_2d.py      # TimesNet and 2D models
│   ├── models_light.py   # LightTimesNet
│   ├── timesnet_original.py
│   ├── fixed_period_inception.py
│   └── runners/          # Experiment runners
├── src/standalone_files/ # Analysis and evaluation scripts
├── run_pipeline.py
├── requirements.txt
└── ...

Technical Focus

Time-Series Forecasting · TimesNet · PyTorch · FFT · 1D/2D Representations · Spatial Patching · CNNs · Depthwise Convolutions · Ablation Studies · Pareto Analysis · Model Efficiency

Reference

Based on:

Wu et al., “TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis.”