Domain Generalization for Time-Series Forecasting via Extended Domain-Invariant Representations

Time-series forecasting is crucial for IoT applications, but generalizing across domains is challenging due to distinct data distributions and dynamics. Most domain generalization methods work well for image processing and classification, but they struggle with time-series forecasting. This is because they solely learn domain-invariant representations of input data, ignoring variations in the output space across domains. This oversight can lead to inaccurate forecasts when outputs in new domains exhibit different distributions or temporal patterns. In this work, we present a new approach to improve the time-series forecasting model generalization by extracting domain-invariant representations from both input and output data. Experiments demonstrate the effectiveness of our approach, achieving significant improvements in forecasting accuracy across multiple test domains. Compared to state-of-the-art methods, our approach delivers up to an 8% increase in accuracy.