Experimental Setups and Results for "BiasCast: Learning and adjusting real time biases from meteorological forecasts to enhance runoff predictions"

Version 1.0: Please be aware that this version of the dataset is NOT the final revised one! The paper is currently still in progress. This record contains the complete experimental setup and test results for the paper "BiasCast: Learning and adjusting real time biases from meteorological forecasts to enhance runoff predictions". Data Structure 1. Baseline - Reference models trained on a standard CUDA LSTM Forecast - models using forecast data only Reanalysis- models using reanalysis data only ForecastAndReanalysis - models combinining forecast and reanalysis data sources 2. CrossDomain - Cross domain evaluation experiments based on CUDA LSTM CrossDomain_Reanalysis - model trained on 5 reanalysis variables CrossDomain_Forecast - model from the CrossDomain_Reanalysis infered with same variables from forecast data source 3. Encoder_Decoder_LSTM - Handoff forecasting architecture Variants with/ without discharge (q) as input in the hindcast model phase Simple and Complex embedding configurations RA (Reanalysis) and FCRA (Forecast+Reanalysis) input combinations in the hindcast phase of the model 4. Sequential_Forecast_LSTM - Sequential LSTM architecture Variants with/ without discharge (q) as input in the hindcast model phase Simple and Complex embedding configurations RA (Reanalysis) and FCRA (Forecast+Reanalysis) input combinations in the hindcast phase of the model 5. TransferLearning - Domain adaption experiments TL_AllWeights - Transfer of all model weights TL_EmbeddingNet - Retraining only embedding net while all other weights are frozen File Types and Contents config.yml - Model configuration and hyperparameters best_model.pt - Trained PyTorch model weights train_data/train_data_scaler.yml - Data normalization parameters test/best_model/ - Test results test_metrics.csv - Per-basin performance metrics (NSE, KGE for 451 basins) test_results.p - Detailed prediction results (pickle format) Usage notes For a detailled description on how to run the experiments, we refer to the BiasCast GitHub repo Data requirements Models require access to the Extended LamaH-CE dataset with meteorological forcings and catchment attributes Citation When using this data or a pre-trained model, we kindly ask you to cite the associated paper: "BiasCast: Learning and adjusting real time biases from meteorological forecasts to enhance runoff predictions"