New Framework Boosts Continual Learning for Time Series Forecasting

Deep learning models typically require large, stationary datasets, which is a significant limitation in real-world scenarios where data distributions constantly change. Continual learning aims to overcome this by allowing models to adapt incrementally, balancing the need for stability (retaining old knowledge) and plasticity (learning new knowledge) under computational constraints. This work introduces an innovative framework for continual time series forecasting. It extends existing static forecasting models by integrating an Experience Replay strategy, which is guided by attention mechanisms. This design allows the model to dynamically adjust to new contexts while effectively preserving previously learned information, thereby mitigating the problem of catastrophic forgetting. The framework was evaluated on standard forecasting benchmarks and a piezometric dataset exhibiting diverse temporal behaviors. The results indicate that this approach successfully enhances or maintains predictive performance over time, while also significantly reducing the need for extensive retraining and large data requirements. This makes it easier to deploy and manage forecasting models in dynamic, real-world settings.