Estimating Uncertainty for Time Series Forecasting with Deep Neural Networks

Background/Question/Methods

Over the last decade, neural networks have achieved numerous state-of-the-art time series forecasting results. Yet, to build trustable models for ecological forecasting problems, it will be critical to use models which estimate uncertainty; and, most machine learning algorithms are not designed with uncertainty estimation in mind. This talk will examine the effectiveness of deep learning approaches to predict uncertainty on the Aquatics NEON Forecast Challenge. We compare two of the most commonly used neural network architectures for sequence prediction tasks, Long Short-Term Memory (LSTM) and Transformer networks, to forecast dissolved oxygen and water temperature in different aquatic ecosystems. This talk will analyze the effectiveness of dropout and ensemble methods paired with LSTMs and Transformers to estimate uncertainty.

Results/Conclusions