This dataset includes evaluation data ("test" data) and performance metrics for water temperature predictions from multiple modeling frameworks. Process-Based (PB) models were configured and calibrated with training data to reduce root-mean squared error. Uncalibrated models used default configurations (PB0; see Winslow et al. 2016 for details) and no parameters were adjusted according to model fit with observations. Deep Learning (DL) models were Long Short-Term Memory artificial recurrent neural network models which used training data to adjust model structure and weights for temperature predictions (Jia et al. 2019). Process-Guided Deep Learning (PGDL) models were DL models with an added physical constraint for energy conservation as a loss term. These models were pre-trained with uncalibrated Process-Based model outputs (PB0) before training on actual temperature observations. Performance was measured as root-mean squared errors relative to temperature observations during the test period. Test data include compiled water temperature data from a variety of sources, including the Water Quality Portal (Read et al. 2017), the North Temperate Lakes Long-TERM Ecological Research Program (https://lter.limnology.wisc.edu/), the Minnesota department of Natural Resources, and the Global Lake Ecological Observatory Network (gleon.org). This dataset is part of a larger data release of lake temperature model inputs and outputs for 68 lakes in the U.S. states of Minnesota and Wisconsin (http://dx.doi.org/10.5066/P9AQPIVD).

This dataset includes evaluation data ("test" data) and performance metrics for water temperature predictions from multiple modeling frameworks. Process-Based (PB) models were configured and calibrated with training data to reduce root-mean squared error. Uncalibrated models used default configurations (PB0; see Winslow et al. 2016 for details) and no parameters were adjusted according to model fit with observations. Deep Learning (DL) models were Long Short-Term Memory artificial recurrent neural network models which used training data to adjust model structure and weights for temperature predictions (Jia et al. 2019). Process-Guided Deep Learning (PGDL) models were DL models with an added physical constraint for energy conservation as a loss term. These models were pre-trained with uncalibrated Process-Based model outputs (PB0) before training on actual temperature observations. Performance was measured as root-mean squared errors relative to temperature observations during the test period. Test data include compiled water temperature data from a variety of sources, including the Water Quality Portal (Read et al. 2017), the North Temperate Lakes Long-TERM Ecological Research Program (https://lter.limnology.wisc.edu/), the Minnesota department of Natural Resources, and the Global Lake Ecological Observatory Network (gleon.org). This dataset is part of a larger data release of lake temperature model inputs and outputs for 68 lakes in the U.S. states of Minnesota and Wisconsin (http://dx.doi.org/10.5066/P9AQPIVD).

Multiple modeling frameworks were used to predict daily temperatures at 0.5m depth intervals for a set of diverse lakes in the U.S. states of Minnesota and Wisconsin. Process-Based (PB) models were configured and calibrated with training data to reduce root-mean squared error. Uncalibrated models used default configurations (PB0; see Winslow et al. 2016 for details) and no parameters were adjusted according to model fit with observations. Deep Learning (DL) models were Long Short-Term Memory artificial recurrent neural network models which used training data to adjust model structure and weights for temperature predictions (Jia et al. 2019). Process-Guided Deep Learning (PGDL) models were DL models with an added physical constraint for energy conservation as a loss term. These models were pre-trained with uncalibrated Process-Based model outputs (PB0) before training on actual temperature observations. Zip files for each lake contain four files, one for each of PB, PB0, DL, and PGDL.

Multiple modeling frameworks were used to predict daily temperatures at 0.5m depth intervals for a set of diverse lakes in the U.S. states of Minnesota and Wisconsin. Process-Based (PB) models were configured and calibrated with training data to reduce root-mean squared error. Uncalibrated models used default configurations (PB0; see Winslow et al. 2016 for details) and no parameters were adjusted according to model fit with observations. Deep Learning (DL) models were Long Short-Term Memory artificial recurrent neural network models which used training data to adjust model structure and weights for temperature predictions (Jia et al. 2019). Process-Guided Deep Learning (PGDL) models were DL models with an added physical constraint for energy conservation as a loss term. These models were pre-trained with uncalibrated Process-Based model outputs (PB0) before training on actual temperature observations. Zip files for each lake contain four files, one for each of PB, PB0, DL, and PGDL.

Multiple modeling frameworks were used to predict daily temperatures at 0.5m depth intervals for a set of diverse lakes in the U.S. states of Minnesota and Wisconsin. Process-Based (PB) models were configured and calibrated with training data to reduce root-mean squared error. Uncalibrated models used default configurations (PB0; see Winslow et al. 2016 for details) and no parameters were adjusted according to model fit with observations. Deep Learning (DL) models were Long Short-Term Memory artificial recurrent neural network models which used training data to adjust model structure and weights for temperature predictions (Jia et al. 2019). Process-Guided Deep Learning (PGDL) models were DL models with an added physical constraint for energy conservation as a loss term. These models were pre-trained with uncalibrated Process-Based model outputs (PB0) before training on actual temperature observations. Zip files for each lake contain four files, one for each of PB, PB0, DL, and PGDL.

Observed water temperatures from 1980-2018 were compiled for 68 lakes in Minnesota and Wisconsin (USA). These data were used as training data for process-guided deep learning models and deep learning models, and calibration data for process-based models. The data are formatted as a single csv (comma separated values) file with attributes corresponding to the unique combination of lake identifier, time, and depth. Data came from a variety of sources, including the Water Quality Portal, the North Temperate Lakes Long-Term Ecological Research Project, and digitized temperature records from the MN Department of Natural Resources.

Observed water temperatures from 1980-2018 were compiled for 68 lakes in Minnesota and Wisconsin (USA). These data were used as training data for process-guided deep learning models and deep learning models, and calibration data for process-based models. The data are formatted as a single csv (comma separated values) file with attributes corresponding to the unique combination of lake identifier, time, and depth. Data came from a variety of sources, including the Water Quality Portal, the North Temperate Lakes Long-Term Ecological Research Project, and digitized temperature records from the MN Department of Natural Resources.

Observed water temperatures from 1980-2018 were compiled for 68 lakes in Minnesota and Wisconsin (USA). These data were used as training data for process-guided deep learning models and deep learning models, and calibration data for process-based models. The data are formatted as a single csv (comma separated values) file with attributes corresponding to the unique combination of lake identifier, time, and depth. Data came from a variety of sources, including the Water Quality Portal, the North Temperate Lakes Long-Term Ecological Research Project, and digitized temperature records from the MN Department of Natural Resources.

This dataset includes model inputs that describe weather conditions for the 68 lakes included in this study. Weather data comes from gridded estimates (Mitchell et al. 2004). There are two comma-separated files, one for weather data (one row per model timestep) and one for ice-flags, which are used by the process-guided deep learning model to determine whether to apply the energy conservation constraint (the constraint is not applied when the lake is presumed to be ice-covered). The ice-cover flag is a modeled output and therefore not a true measurement (see "Predictions" and "pb0" model type for the source of this prediction). This dataset is part of a larger data release of lake temperature model inputs and outputs for 68 lakes in the U.S. states of Minnesota and Wisconsin (http://dx.doi.org/10.5066/P9AQPIVD).

This dataset includes model inputs that describe weather conditions for the 68 lakes included in this study. Weather data comes from gridded estimates (Mitchell et al. 2004). There are two comma-separated files, one for weather data (one row per model timestep) and one for ice-flags, which are used by the process-guided deep learning model to determine whether to apply the energy conservation constraint (the constraint is not applied when the lake is presumed to be ice-covered). The ice-cover flag is a modeled output and therefore not a true measurement (see "Predictions" and "pb0" model type for the source of this prediction). This dataset is part of a larger data release of lake temperature model inputs and outputs for 68 lakes in the U.S. states of Minnesota and Wisconsin (http://dx.doi.org/10.5066/P9AQPIVD).