LAUSR

Time series forecasting is crucial in network traffic management, weather prediction, and traffic scheduling. Recurrent neural networks have significantly progressed time series analysis by effectively utilizing temporal dependencies through autoregressive strategies. However, the inherent nature of autoregressive models tends to accumulate errors during inference. Additionally, the complexity of temporal patterns in time series makes it challenging for models to capture dependencies reliably. This paper proposes an enhancing denoising probability neural network designed to separate noise from preliminary predictions, providing a flexible framework for probabilistic forecasting. The enhanced denoising mechanism leverages initial prediction results as conditions for a generative model, enabling it to perform sequence-level forecasting and thus mitigate cumulative errors associated with autoregressive forecasting. Specifically, we encode both historical observations and preliminary forecasts into a conditional latent distribution, where noise and uncertainty are modeled probabilistically and separated from the underlying signal via distributional inference. This process provides a versatile framework for disentangling intricate patterns via interactions within the latent space. Extensive experiments across five datasets validate the effectiveness and robustness of our model, indicating its competitive performance against state-of-the-art methods.