Stream networks with reservoirs provide a particularly hard modeling challenge because reservoirs can decouple physical processes (e.g., water temperature dynamics in streams) from atmospheric signals. Including observed reservoir releases as inputs to models can improve water temperature predictions below reservoirs, but many reservoirs are not well-observed. This data release contains predictions from stream temperature models described in Jia et al. 2022, which describes different deep learning and process-guided deep learning model architectures that were developed to handle scenarios of missing reservoir releases. The spatial extent of this modeling effort was restricted to two spatially disjointed regions in the Delaware River Basin. The first region included streams above the Delaware River at Lordville, NY, and included the West Branch of the Delaware River above and below the Cannonsville Reservoir and the East Branch of the Delaware River above and below the Pepacton Reservoir. Additionally, the Neversink River which flows into the Delaware River at Port Jervis, New York, was included and contains river reaches above and below the Neversink Reservoir. For each model, there are test period predictions from 2006-12-26 through 2020-06-22. Model input, training, and validation data can be found in Oliver et al. (2021).

The publication associated with this data release is Jia X., Chen S., Xie Y., Yang H., Appling A., Oliver S., Jiang Z. 2022. Modeling reservoir release in stream temperature prediction using pseudo-prospective learning and physical simulations, SIAM International Conference on Data Mining (SDM). DOI: https://doi.org/10.1137/1.9781611977172.11

Several models were used to improve water temperature prediction in the Delaware River Basin.

PRMS-SNTemp was used to predict daily temperatures at 456 stream reaches in the Delaware River Basin. Daily stream temperature predictions for inflow and outflow reaches for Cannonsville and Pepacton reservoirs were pulled aside into a separate csv to be used as inputs to the General Lake Model (GLM). Reservoir outflow predictions and in-reservoir temperature predictions were generated with calibrated models built using GLM v3.1. We calculated a decay rate based on the modeled reservoir outflow temperatures and observed downstream river temperature to estimate the decay of the reservoir influence on stream temperature as a function of distance downstream of a reservoir. These decay rates from Pepacton and Cannonsville were used to weight the predictions from GLM and PRMS-SNTemp on the East and West Branch of the Delaware River, respectively, to represent the mix of stream and reservoir processes that affect temperature dynamics. These weighted values were then used to pre-train the deep learning models that were used to forecast temperature. More details on these methods can be found in Zwart and others 2021.

Code that generated these results can be found in Zwart and others 2021 (https://doi.org/10.5281/zenodo.5164910).

Several models were used to improve water temperature prediction in the Delaware River Basin.

PRMS-SNTemp was used to predict daily temperatures at 456 stream reaches in the Delaware River Basin. Daily stream temperature predictions for inflow and outflow reaches for Cannonsville and Pepacton reservoirs were pulled aside into a separate csv to be used as inputs to the General Lake Model (GLM). Reservoir outflow predictions and in-reservoir temperature predictions were generated with calibrated models built using GLM v3.1. We calculated a decay rate based on the modeled reservoir outflow temperatures and observed downstream river temperature to estimate the decay of the reservoir influence on stream temperature as a function of distance downstream of a reservoir. These decay rates from Pepacton and Cannonsville were used to weight the predictions from GLM and PRMS-SNTemp on the East and West Branch of the Delaware River, respectively, to represent the mix of stream and reservoir processes that affect temperature dynamics. These weighted values were then used to pre-train the deep learning models that were used to forecast temperature. More details on these methods can be found in Zwart and others 2021.

Code that generated these results can be found in Zwart and others 2021 (https://doi.org/10.5281/zenodo.5164910).

Observations related to water and thermal budgets in the Delaware River Basin. Data from reservoirs in the basin include reservoir characteristics (e.g., bathymetry), daily water levels, daily depth-resolved water temperature observations, and daily inflows, diversions, and releases. Data from streams in the basin include daily flow and temperature observations. Data were compiled from a variety of sources to cover the modeling period (1980-2021), including the National Water Inventory System, Water Quality Portal, EcoSHEDS stream water temperature database, ReaLSAT, and the New York Department of Environmental Conservation. The data are formatted as a single csv (comma separated values) or zipped csv.

For modeling purposes, we sought to create a test set of flow and temperature observations that were representative of dynamics throughout the Delaware River basin from water year 1980-present. Test holdouts are documented in the flow and temperature files. To minimize the possibility of the correlation between sites and temporal autocorrelation at single sites causing artificially high test performance, we created temporal holdouts (time periods where data from all sites were reserved for testing), and spatial holdouts (sites where all data were reserved for testing). In all, this resulted in a train/test split of 66.2%/33.8% for observed temperature reach days, and 71.4%/28.6% for observed flow reach days.

For temporal holdouts: All data in the water years 1980-84, 2011-15, and 2021 were reserved for the test set. These windows were chosen to attempt to balance the ability to test on the most recent data (critical to assess performance in an operational setting) and historical periods, while still training on a sufficient amount of modern continuous data. For spatial holdouts: We chose eight reaches of the PRMS network to reserve all data for testing, based on representing key parts of the Delaware basin (mainstem, headwaters, reservoir-adjacent reaches), representing the distribution catchment attributes (e.g. fraction of impervious surfaces) and minimizing the number of observations within a 20 km distance along the network ('fish radius').