Abstract:Time series forecasting typically needs to address non-stationary data with evolving trend and seasonal patterns. To address the non-stationarity, reversible instance normalization has been recently proposed to alleviate impacts from the trend with certain statistical measures, e.g., mean and variance. Although they demonstrate improved predictive accuracy, they are limited to expressing basic trends and are incapable of handling seasonal patterns. To address this limitation, this paper proposes a new instance normalization solution, called frequency adaptive normalization (FAN), which extends instance normalization in handling both dynamic trend and seasonal patterns. Specifically, we employ the Fourier transform to identify instance-wise predominant frequent components that cover most non-stationary factors. Furthermore, the discrepancy of those frequency components between inputs and outputs is explicitly modeled as a prediction task with a simple MLP model. FAN is a model-agnostic method that can be applied to arbitrary predictive backbones. We instantiate FAN on four widely used forecasting models as the backbone and evaluate their prediction performance improvements on eight benchmark datasets. FAN demonstrates significant performance advancement, achieving 7.76% ~ 37.90% average improvements in MSE.

What problem does this paper attempt to address?

### Problems the paper attempts to solve This paper aims to solve the problem of non - stationary data in time - series prediction, especially those data with evolutionary trends and seasonal patterns. Existing methods usually relieve the influence of trends through reversible instance normalization, but these methods mainly rely on statistics such as mean and variance and cannot effectively handle seasonal patterns. Therefore, this paper proposes a new instance - normalization solution - Frequency - Adaptive Normalization (FAN) to expand the ability of instance normalization in dealing with dynamic trends and seasonal patterns. ### Specific problem description 1. **Challenges of non - stationary data**: - Time - series data usually contains trends and seasonal patterns, which change over time, causing covariate patterns to change between different time steps. - These dynamic changes pose a significant challenge to prediction tasks, especially in fields such as finance, transportation, and infectious diseases. 2. **Limitations of existing methods**: - Current reversible instance - normalization methods mainly rely on statistics in the time domain (such as mean and variance) to remove non - stationary signals. - Although these methods improve prediction accuracy to a certain extent, they can only extract the most significant trend components and cannot effectively capture seasonal patterns. 3. **Contributions of this paper**: - **Propose FAN**: By filtering the first K dominant components of each input instance in the frequency domain, FAN can handle trends and seasonal patterns simultaneously. - **Pattern - adaptation module**: Use a simple MLP model to predict future non - stationary information instead of assuming that these patterns remain unchanged. - **Experimental verification**: Apply FAN to four widely - used prediction models and evaluate its performance improvement on eight benchmark datasets. The results show that FAN improves by an average of 7.76% to 37.90% on the MSE metric. ### Method overview 1. **Frequency - Adaptive Normalization (FAN)**: - **Frequency Residual Learning (FRL)**: Convert the input into Fourier components in complex form through the Discrete Fourier Transform (DFT), select the K frequency components with the largest magnitudes, and restore them to the time domain through the Inverse Discrete Fourier Transform (IDFT). - **Normalization process**: Remove these non - stationary components from the input to obtain a more stationary input sequence. - **Denormalization process**: Predict future non - stationary information through an MLP model and add it back to the prediction result. 2. **Loss function**: - To assist the residual - learning process, a prior - guidance loss function is introduced to predict the main frequency components. - The final loss function is a multi - task optimization problem, ensuring the optimization of the model's overall prediction accuracy. ### Experimental results - **Datasets**: Eight popular multivariate time - series prediction benchmark datasets are used, including power - transformer temperature, power consumption, exchange rates, traffic flow, and weather, etc. - **Performance improvement**: FAN significantly improves the prediction performance on multiple datasets, especially when dealing with datasets with obvious seasonal patterns. - **Comparison with existing methods**: FAN outperforms existing reversible instance - normalization methods such as SAN, Dish - TS, and RevIN in most cases. ### Conclusion The FAN method proposed in this paper effectively solves the shortcomings of existing methods in dealing with seasonal patterns by processing non - stationary data in the frequency domain and significantly improves the accuracy of time - series prediction.

Frequency Adaptive Normalization For Non-stationary Time Series Forecasting

FredNormer: Frequency Domain Normalization for Non-stationary Time Series Forecasting

Adaptive Normalization for Non-stationary Time Series Forecasting: A Temporal Slice Perspective

Addressing Distribution Shift in Time Series Forecasting with Instance Normalization Flows

Evolving Multi-Scale Normalization for Time Series Forecasting under Distribution Shifts

SIN: Selective and Interpretable Normalization for Long-Term Time Series Forecasting

Not All Frequencies Are Created Equal:Towards a Dynamic Fusion of Frequencies in Time-Series Forecasting

Adaptive Temporal-Frequency Network for Time-Series Forecasting

Deep Adaptive Input Normalization for Time Series Forecasting

Deep Frequency Derivative Learning for Non-stationary Time Series Forecasting

GAS-Norm: Score-Driven Adaptive Normalization for Non-Stationary Time Series Forecasting in Deep Learning

FDNet: Focal Decomposed Network for Efficient, Robust and Practical Time Series Forecasting

MFF-FTNet: Multi-scale Feature Fusion across Frequency and Temporal Domains for Time Series Forecasting

MMFNet: Multi-Scale Frequency Masking Neural Network for Multivariate Time Series Forecasting

Learning Multi-Pattern Normalities in the Frequency Domain for Efficient Time Series Anomaly Detection

Self-Adaptive Scale Handling for Forecasting Time Series with Scale Heterogeneity.

HiFNet: rethinking time series forecasting models from a perspective of error power spectrum

Reversible Instance Normalization for Accurate Time-Series Forecasting against Distribution Shift

A 2-Stage Strategy for Non-Stationary Signal Prediction and Recovery Using Iterative Filtering and Neural Network

FAN: Fourier Analysis Networks

Abnormality Forecasting: Time Series Anomaly Prediction via Future Context Modeling