Abstract:Long-term time series forecasting (LTSF) is a critical task across diverse domains. Despite significant advancements in LTSF research, we identify a performance bottleneck in existing LTSF methods caused by the inadequate modeling of Temporal Dependencies within the Target (TDT). To address this issue, we propose a novel and generic temporal modeling framework, Temporal Dependency Alignment (TDAlign), that equips existing LTSF methods with TDT learning capabilities. TDAlign introduces two key innovations: 1) a loss function that aligns the change values between adjacent time steps in the predictions with those in the target, ensuring consistency with variation patterns, and 2) an adaptive loss balancing strategy that seamlessly integrates the new loss function with existing LTSF methods without introducing additional learnable parameters. As a plug-and-play framework, TDAlign enhances existing methods with minimal computational overhead, featuring only linear time complexity and constant space complexity relative to the prediction length. Extensive experiments on six strong LTSF baselines across seven real-world datasets demonstrate the effectiveness and flexibility of TDAlign. On average, TDAlign reduces baseline prediction errors by \textbf{1.47\%} to \textbf{9.19\%} and change value errors by \textbf{4.57\%} to \textbf{15.78\%}, highlighting its substantial performance improvements.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the deficiency of existing methods in modeling Temporal Dependencies within the Target (TDT) in Long - Term Time - Series Forecasting (LTSF). Specifically, the author points out that although existing LTSF research has made significant progress, these methods fail to fully capture the inherent temporal dependency relationships between adjacent time steps in the target time series, which leads to a bottleneck in prediction performance.
### Specific manifestations of the problem
1. **Error accumulation and slow inference speed**: Autoregressive decoding strategies (such as LSTM, GRU, etc.) perform well in short - term time - series forecasting, but are not suitable for long - term forecasting tasks due to error accumulation and slow step - by - step inference speed.
2. **Insufficient temporal - dependency modeling**: Many existing methods mainly focus on temporal pattern extraction in the input time series, while ignoring the temporal - dependency modeling in the target time series, resulting in inaccurate or unrealistic prediction results.
### Solutions
To solve the above problems, the author proposes a new general framework - **Temporal - Dependency Alignment (TDAlign)**. The main contributions of TDAlign include:
1. **Introducing a new loss function**: By defining a new loss function \( L_D \), ensure that the temporal changes between predicted values are consistent with those between target values, so as to better capture the dynamic patterns in the time series.
\[
L_D=\frac{1}{H} \sum_{i = 1}^{H} \ell(d_{t + i},\hat{d}_{t + i})
\]
where \( d_{t + i}=x_{t + i}-x_{t + i - 1} \) represents the change value between adjacent time steps in the target time series, and \(\hat{d}_{t + i}\) represents the corresponding change value in the predicted time series.
2. **Adaptive loss - balancing strategy**: Introduce an adaptive weight \(\rho\) to dynamically adjust the relative importance of the traditional prediction loss \( L_Y \) and the newly introduced temporal - dependency loss \( L_D \).
\[
\rho=\frac{1}{H} \sum_{i = 1}^{H} 1(\text{sgn}(d_{t + i})\neq\text{sgn}(\hat{d}_{t + i}))
\]
where \(\text{sgn}(x)\) is the sign function, and \(1(\cdot)\) is the indicator function, which is used to calculate the consistency ratio of the change directions between the predicted value and the target value.
3. **Plug - in framework**: TDAlign can be seamlessly integrated into existing LTSF methods without introducing additional learnable parameters, and has linear time complexity and constant space complexity, thus ensuring computational efficiency.
### Experimental verification
The author conducted extensive experiments on six strong baseline models (DLinear, MICN, TimesNet, PatchTST, iTransformer, SegRNN), covering seven real - world datasets (ETTh1, ETTh2, ETTm1, ETTm2, Weather, Electricity, ILI). The experimental results show that TDAlign can significantly improve the prediction performance, reducing the prediction error by an average of 1.47% - 9.19% and the change - value error by 4.57% - 15.78%.
In conclusion, this paper solves the deficiency of existing LTSF methods in temporal - dependency modeling by introducing the TDAlign framework, and significantly improves the accuracy of long - term time - series forecasting.