🤖 AI

TriTS: Time Series Forecasting from a Multimodal Perspective

TriTS is a novel cross-modal disentanglement framework that achieves state-of-the-art long-term time series forecasting by projecting 1D signals into orthogonal time, frequency, and 2D-vision spaces, leveraging a Period-Aware Reshaping strategy with Visual Mamba and Multi-Resolution Wavelet Mixing to efficiently model complex dynamics while drastically reducing computational costs.

Original authors: Xiang Ao

Published 2026-04-21

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xiang Ao

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to predict the weather for the next week. You have a massive notebook of past weather data. If you just look at the numbers in a single line (one by one), it's hard to see the big picture. You might miss the fact that "every Tuesday it rains" or that "the temperature is slowly rising over the year."

This paper introduces TriTS, a new super-smart computer program designed to predict the future of time-based data (like stock prices, energy usage, or weather) by looking at that data through three different pairs of glasses at the same time.

Here is how TriTS works, explained simply:

1. The Problem: The "One-Dimensional" Blind Spot

Most old forecasting models look at time series data like a single, long string of beads. They try to guess the next bead based on the ones before it.

The Issue: Real-world data is messy. It has long-term trends (like a slow economic rise), repeating patterns (like rush hour traffic every day), and sudden surprises (like a storm). Looking at it as just a "line" makes it hard to untangle these different threads.
The Old "Vision" Fix: Some recent models tried to turn that line of numbers into a 2D picture (like folding a long scroll into a square) so a computer could "see" patterns like an image. But, these models were like trying to read a book by staring at every single pixel individually—it was incredibly slow and used up too much computer memory.

2. The Solution: The TriTS "Three-Legged Stool"

TriTS solves this by splitting the data into three specialized teams, each doing what they are best at, and then combining their answers.

👁️ Team 1: The "Vision" Team (The Pattern Spotter)

What they do: They fold the data into a 2D image to spot big, repeating patterns (like "every 7 days, sales go up").
The Innovation: Instead of using a slow, heavy camera (the old "Vision Transformer" that gets bogged down with big images), TriTS uses a Visual Mamba.
The Analogy: Imagine a traditional camera that takes a photo of every single pixel in a room, one by one. It's slow. The Visual Mamba is like a smart security guard who walks through the room, quickly scanning left-to-right and right-to-left, instantly noticing the "texture" of the room without checking every single dust mote. It sees the whole picture but runs super fast and uses very little battery.

🎵 Team 2: The "Frequency" Team (The Musician)

What they do: They listen to the data like music to separate the "beat" (trends) from the "noise" (random glitches).
The Innovation: Old methods used a tool called FFT (Fast Fourier Transform), which is like trying to analyze a whole song by looking at the average pitch of the entire album. It misses sudden changes. TriTS uses Wavelets.
The Analogy: Think of a multi-scale magnifying glass. If you look at a forest from a plane, you see the general shape. If you zoom in with a magnifying glass, you see individual leaves. The Wavelet team zooms in and out simultaneously. They can separate the "slow, steady rhythm" of the data from the "sudden, sharp cracks" of noise, ensuring the model doesn't get confused by random spikes.

📉 Team 3: The "Time" Team (The Anchor)

What they do: They keep a simple, straight line to track the basic, steady movement of the data.
The Innovation: They use a simple "Exponential Moving Average."
The Analogy: Imagine a ship sailing in a storm. The Vision and Frequency teams are the crew looking at the waves and the wind to steer. The Time Team is the heavy anchor. Even if the other teams get excited by a crazy wave, the anchor keeps the ship from drifting too far off course. It ensures the prediction stays grounded in reality and doesn't go crazy.

3. The "Chef" (The Fusion)

Once these three teams have done their analysis, they don't just average their answers. TriTS has a Smart Chef (a gating mechanism) who decides how much to trust each team based on the current situation.

If the data is very stable and predictable, the Chef might trust the Vision Team more.
If the data is full of sudden, random spikes, the Chef might trust the Frequency Team more.
If the data is just drifting slowly, the Time Team gets the most say.

Why is this a Big Deal?

Speed & Efficiency: By swapping out the slow "pixel-by-pixel" vision models for the fast "smart guard" (Visual Mamba), TriTS is much faster and uses way less computer memory. It can handle huge amounts of history without crashing.
Accuracy: Because it looks at the data from three angles (Time, Frequency, and Vision), it catches things other models miss. It's like solving a puzzle by looking at the picture, the edge pieces, and the color patterns all at once.
Real-World Ready: It works great on everything from electricity grids to stock markets, proving that you don't need a massive, slow computer to make smart predictions.

In short: TriTS is like a forecasting team that doesn't just stare at a spreadsheet. It has a fast visual scanner, a precise audio analyzer, and a steady anchor, all working together to predict the future with high accuracy and low cost.

1. Problem Statement

Long-term Time Series Forecasting (LTSF) is critical for sectors like finance, energy, and meteorology but faces significant challenges due to the entanglement of temporal dynamics in real-world signals. Current approaches suffer from specific limitations:

1D Perspective Bottleneck: Analyzing signals solely in the time domain often fails to capture complex global dependencies and non-stationary mutations.
Vision-Based Limitations: While recent methods (e.g., VisionTS) project 1D series into 2D "temporal images" to leverage vision models, they rely on Vision Transformers (ViTs). ViTs incur a prohibitive quadratic computational complexity ( $O(N^2)$ ), making them inefficient for ultra-long sequences.
Frequency-Based Limitations: Traditional frequency methods using Fast Fourier Transform (FFT) rely on global basis functions, making it difficult to localize transient mutations and non-stationary shifts (e.g., sudden trend changes) prevalent in real data.

2. Methodology: The TriTS Framework

TriTS proposes a cross-modal disentanglement framework that projects 1D time series into three orthogonal, complementary spaces: Time, Frequency, and 2D-Vision. These are dynamically fused to resolve representation bottlenecks.

A. Time-Domain Branch (The Linear Anchor)

Goal: Maintain numerical stability and capture non-stationary trends without over-smoothing.
Mechanism: Instead of standard Linear Layers or Simple Moving Averages (SMA), TriTS uses a Streaming Exponential Moving Average (EMA).
- EMA assigns exponentially decaying weights, responding sensitively to recent changes while avoiding boundary bias issues associated with SMA padding.
- The extracted trend is passed through a streaming linear layer to serve as a robust "linear anchor" for the entire system.

B. Frequency-Domain Branch (MR-WM)

Goal: Decouple non-stationary signals into trend and noise for fine-grained time-frequency localization.
Mechanism: Replaces global FFT with Multi-Resolution Wavelet Mixing (MR-WM).
- Decomposition: Uses Multi-level Discrete Wavelet Transform (DWT) to split the signal into approximation coefficients (long-term trends) and detail coefficients (local variations/noise).
- Independent Branches: Each frequency component is processed in a completely independent resolution branch with its own normalization and mixing modules. This prevents high-frequency noise from contaminating low-frequency trends.
- Dual-Stage Mixing: Within each branch, a Patch Mixer captures local temporal context, and an Embedding Mixer extracts global semantic features across channels.
- Reconstruction: Predictions from all branches are reconstructed via Inverse DWT (IDWT).

C. Vision-Domain Branch (Visual Mamba)

Goal: Efficiently model cross-period global dependencies without the $O(N^2)$ cost of ViTs.
Mechanism: Integrates Visual Mamba (Vim) with a Period-Aware Reshaping strategy.
- Reshaping: The 1D sequence is folded into a 2D "temporal image" based on the dominant period $P$ (detected via autocorrelation). Intra-period variations become row textures; inter-period evolution becomes column structures.
- Vim Encoder: Replaces the self-attention mechanism with a Bidirectional State Space Model (Bi-SSM).
  - Utilizes a Selective Scanning mechanism where parameters are dynamically generated by the input, allowing the model to adaptively "memorize" or "forget" information.
  - Bidirectional Scanning: Addresses the lack of strict causality in visual perception by scanning the sequence forward and backward, fusing the results via a gating mechanism.
- Complexity: Achieves linear computational complexity ( $O(N)$ ) while capturing global visual textures.

D. Adaptive Scale-Aware Fusion

A lightweight gating network generates time-step-level confidence weights ( $G_{time}, G_{freq}, G_{vis}$ ) based on the input representations.
The final output is a weighted sum of the three branches, allowing the model to dynamically prioritize the most relevant modality (e.g., frequency for periodic data, vision for long-range trends) for each specific dataset.

3. Key Contributions

TriTS Framework: A novel cross-modal disentanglement architecture that synergizes time, frequency, and vision modalities, effectively resolving information loss inherent in single-modality methods.
Visual Mamba Integration: Pioneers the use of Visual Mamba (Vim) in time series forecasting via a 1D-to-2D bridging strategy. This breaks the $O(N^2)$ bottleneck of traditional vision-based forecasters, enabling efficient global texture capture in ultra-long sequences with linear complexity.
MR-WM Module: Introduces a Multi-Resolution Wavelet Mixing encoder that explicitly decouples trends and noise, addressing the limitations of FFTs in handling non-stationary mutations and local anomalies.
Efficiency & Performance: Demonstrates that a multimodal approach can drastically reduce parameter counts and inference latency compared to existing vision-based models while achieving superior accuracy.

4. Experimental Results

Datasets: Evaluated on 7 benchmark datasets (ETT, Weather, Electricity, Traffic) covering diverse domains.
Baselines: Compared against 8 SOTA models, including VisionTS/VisionTS++, WPMixer, FDNet, iTransformer, PatchTST, TimeMixer, and DLinear.
Performance:
- TriTS achieved State-of-the-Art (SOTA) performance (lowest MSE/MAE) on the majority of datasets and prediction lengths (e.g., $L=720$ ).
- It outperformed VisionTS significantly, particularly in datasets with high numerical fluctuations (ETT), by compensating for vision models' lack of numerical fidelity with the time-domain linear branch.
Efficiency:
- Parameters: TriTS uses significantly fewer parameters (e.g., ~~8.82M) compared to VisionTS (~~111.91M).
- Latency: Due to the linear complexity of Vim and efficient wavelet modules, TriTS exhibits much lower inference latency and memory usage than ViT-based models.
Ablation Studies: Removing any single branch (Time, Frequency, or Vision) led to performance degradation. Notably, removing the Time Branch caused severe instability, confirming its role as the numerical backbone.

5. Significance

TriTS represents a paradigm shift in time series forecasting by treating data as a multimodal entity rather than a unidimensional signal. Its significance lies in:

Solving the Efficiency-Accuracy Trade-off: It proves that high-performance forecasting does not require the computational heaviness of ViTs; linear-complexity SSMs (Mamba) can effectively model visual patterns in time series.
Robustness to Non-Stationarity: By combining wavelet-based local decomposition with global visual modeling, TriTS handles both stable trends and sudden mutations better than previous methods.
Scalability: The linear complexity makes it feasible for ultra-long sequence forecasting tasks that were previously computationally prohibitive for vision-based approaches.

In conclusion, TriTS redefines the efficiency baseline for multimodal time series analysis, offering a scalable, accurate, and computationally efficient solution for complex real-world forecasting challenges.