Visual Dynamic Pattern Detection in LSTM-based Water Level Forecasting Model

摘要

Nowadays, deep learning models are increasingly being applied in climate prediction analysis as an alternative to computational y expensive physical models for its features of flexible data-driven learning and universality. However, besides the forecasting output, to make the overal evaluation of the target we analyze, more information that can further explain the features needs to be derived. In this paper, we focus on the water level forecasting problem by using improved long short-term memory (LSTM) network model and we are interested in better understanding how the deep learning model can capture and learn the feature of river basin from the data so that the features of river basins can be further concluded to support the evaluation and risk management. In our research, we have made LSTMVis, a visual analysis tool for recurrent neural networks, adapted to water level analysis and try detecting the precipitation and water level patterns that may have a significant influence on the forecasting by evaluating the changes in hidden state dynamics. We wil show several use cases to demonstrate how the patterns can be detected and verified by using our proposed systems so that more features of river basins can be further explained. Our proposed visual analytic can assist the domain experts better evaluate the river basins to enable better risk management so that the corresponding measures can be conducted efficiently in case of potential emergencies especial y for floods.