Process-Guided Deep Learning for Water Temperature Prediction

Abstract:

Accurate water temperature predictions are essential for managing aquatic habitats and ensuring sufficient water quality for human uses. As one of the most-observed water quality parameters, temperature is a promising target for data-intensive deep learning methods. We have applied deep learning for temperature prediction in several recent applications, including for fisheries assessment in hundreds of lakes in the Upper Midwest and for informing timed releases of cold water from reservoirs into streams of the Delaware River Basin. For these applications we have experimented with the integration of physical constraints, including energy balance, physically meaningful intermediate variables, and monotonically increasing water density with depth (where density is a function of temperature). We are finding that these physical constraints give the greatest boost to neural network accuracy in conjunction with recurrence and convolution to convey time- and space-awareness, respectively, and pretraining to initialize the model as an emulator of a process-based model. In combination, these several mechanisms of process guidance produce more accurate water temperature predictions than purely process-based models even when predicting outside the range of observations used to train the model, making them particularly useful for predictions in changing climates or land uses. Additionally, these models outperform process-based models even when data are sparse, and thus are applicable in a wide range of lakes, reservoirs, and river networks. Although the trustworthiness of data-driven models continues to be a point of discussion and investigation, process guidance greatly strengthens the reliability and utility of deep learning for the kinds of water temperature prediction problems posed by stakeholders.