Abstract
Predicting multivariate time series at arbitrary time points remains a challenge for traditional discrete-time models. We present EgPDE-Net, a novel continuous-time neural network that frames time series forecasting as a Partial Differential Equation (PDE) problem. By reducing the PDE to a tractable, exogenous-guided Ordinary Differential Equation (ODE), EgPDE-Net explicitly models the influence of external factors on target variables. Experimental results across four diverse datasets demonstrate that EgPDE-Net outperforms state-of-the-art continuous models, reducing RMSE by approximately 10% and MAE by 14% for arbitrary-step predictions.
Introduction
Most deep learning architectures for time series, such as Recurrent Neural Networks (RNNs), are inherently discrete. They are optimized for fixed-interval forecasting (e.g., $T+1, T+2$). However, many real-world applications require "arbitrary-step" prediction, where forecasts are needed at irregular or continuous time points.
While Neural Ordinary Differential Equations (Neural ODEs) provide a framework for continuous-time modeling, they often struggle with multivariate datasets where exogenous variables (external drivers like weather or power load) significantly impact the target series. Existing models frequently fail to capture the complex, time-dependent correlations between these external factors and the system's dynamics.
Methodology: The EgPDE-Net Architecture
The core innovation of EgPDE-Net is treating the dynamical system as a complex PDE. This approach is more robust than standard ODEs for modeling systems influenced by multiple independent variables. To ensure mathematical tractability, the authors utilize a lemma to reduce the PDE into a regularized, exogenous-guided ODE.
The architecture is divided into two primary pipelines:
- Exogenous Pipeline: This branch processes external variables through a self-attention block to capture both global and local correlations. An ODE solver then generates "partial differential weights," representing the time-varying influence of these external factors.
- Target Pipeline: The target series is encoded using a Gated Recurrent Unit (GRU). A second ODE solver uses the weights from the exogenous pipeline to guide the trajectory of the hidden states, enabling continuous prediction at any point $t$.
By parameterizing the derivative of the hidden state with these two components, the model effectively "guides" the forecast based on environmental changes.
Experimental Evaluation
The model was evaluated on four multivariate datasets: SML2010 (indoor temperature), Electricity (homestead consumption), and ETTh1/ETTh2 (transformer temperature). To test arbitrary-step prediction, the datasets were re-sampled so the model trained on integral steps but was tested on mid-step points (e.g., $T+1.5$).
Key Results
- Superior Accuracy: EgPDE-Net reduced RMSE by 9.85% and MAE by 13.98% compared to the best continuous baseline, ETN-ODE.
- Capturing Dynamics: In periodic datasets like SML2010, the model was notably better at capturing peaks and troughs, whereas competing models often over-smoothed these critical transitions.
- Efficiency: Despite its mathematical complexity, EgPDE-Net maintains a small footprint (~50k parameters) and competitive inference times, making it practical for real-world deployment.
Conclusion
EgPDE-Net advances continuous-time forecasting by integrating exogenous variables through a PDE-based framework. By using external data to guide the derivative of the target state, the model achieves higher precision in arbitrary-step prediction. Future work will explore fully continuous encoding networks and methods to handle informative missingness in historical records.
Summary updated