Graph Convolutions with Wavelets for Stream Temperature Forecasting

Abstract:

Temperature of freshwater systems is a well-studied phenomenon with a number of important physical, biological, and ecological responses. The ability to estimate and predict temperature across larger hydrologic systems is critical for the management of these systems. For example, the Delaware River Basin Commission regulates freshwater withdraws from reservoirs to maintain a maximum river temperature during summer months to support aquatic wildlife habitat. The ability to estimate temperature is impacted by the complexity of interactions between stream and river networks, and environmental forcing factors. A number of attempts have been made to mechanistically predict temperature with some success, but these methods require consistent measurement and an understanding of the direct relationship between weather and its impacts on stream conditions. In this talk, we present a non-mechanistical machine learning model based on the spatial configuration of streams and their historical temperature relationships given external factors. Specifically, our approach relies on a new stateful architecture that combines the signal processing power of wavelet analysis and multiple types of neural networks, including cross and graph convolutional network layers. Via spectral clustering, we extract structural properties of the geographical area being analyzed, and use cross layers to extract correlations between the time-varying signals and the area geography. The result is a temperature forecast tailored to the particular geography in question. We show results in comparison to existing models, and discuss issues scaling this approach to other watersheds.