Deep Learning from Model Errors
Abstract
In general, inversion and data fusion methods improve a solution by optimizing an objective function. In the optimization, the objective function is commonly based on a solution or prediction error that is the mismatch with respect to some reference information. In hydrology, a major challenge is the error of precipitation predictions, which directly propagates into the hydrologic forecasts. Common correction approaches utilize statistical bias correction methods based on simplifying assumptions with respect to the precipitation error structure. In this study, we apply deep learning (convolution neural networks, CNNs) to relax these assumptions to improve estimates of the error structure of precipitation from fully integrated groundwater-to-atmosphere simulations. The ultimate goal is to arrive at a general corrector method to be used online while the integrated simulations are running to improve precipitation estimates and, thus, indirectly the hydrologic state. In the construction of the corrector method, the simulation and reference data sets were from simulations over Europe using the Terrestrial Systems Modeling Platform (TSMP) and COSMO-REA6 reanalysis, respectively. TSMP generally overestimates precipitation, when compared to COSMO-REA6. A UNet CNN was used to learn the model-reference error structure based on the increments between TSMP and COSMO-REA6. Once the UNet is trained, it is no longer dependent on the reference data, and is able to independently produce the error structure for correction. There is a generally good agreement between the predicted and actual errors with an average correlation of about 0.7. However, for convective precipitation events in summer, the quality of the error structure produced by Unet decreases. In general, the UNet is able to produce valuable error structures useful for correcting precipitation predictions yielding improvements of some 50% in terms of the mean error, and 20 and 40% in case of the mean squared error and correlation coefficient, respectively. Based on our results, we speculate that an implementation of the trained network as an online correction method in integrated simulations will lead to significant improvements of hydrologic variables also in case of convective precipitation events, which is important for e.g. flood forecasting.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2022
- Bibcode:
- 2022AGUFM.H22D..02K