Post-processing of seasonal forecasts in a semi-arid river basin through artificial intelligence

The use of artificial intelligence is growing in many science areas boosted by an unprecedent increase in data availability and the improvements in computer hardware. Its application in Earth system sciences is particularly relevant due to the existence of complex behavioural patterns whose reproduction with traditional methods is challenging. Its use in seasonal forecasting is favoured by the existence of large amounts of open meteorological data.

This study showcases the potential of artificial intelligence to downscale and post-process seasonal meteorological forecasts in semiarid river basin. The artificial intelligence methodology used is fuzzy logic. Daily raw seasonal forecasts correspond to the ECMWF SEAS5 seasonal forecasts from the Copernicus Climate Change Service (CS3), available at a 1º grid, while daily ERA5 reanalysis, available at a 0.25º grid through the C3S, is employed as observational data. The meteorological variables used are precipitation; 2-meter mean, minimum and maximum temperatures; incident shortwave solar radiation and wind speed.

The artificial intelligence algorithm is coded in a Python script. The script requires the coordinates of a target grid (that may coincide or not with the grid of observational data). For each point it performs the post-processing with the following process: 1) extracts the observational and forecast data for the closest point available; 2) computes the cdf’s of both datasets per month; 3) builds and trains fuzzy logic systems to match the forecasts cdfs to the observational cdfs; and 4) obtains the post-processed forecasts for the target grid provided. The script admits any meteorological variable, seasonal forecasting system from the C3S and observational dataset (it has been successfully tested with ERA5Land and the Spain02 gridded dataset).

The script has been applied to the semi-arid upper Tagus and Segura river basins. The Segura river basin suffers a severe overexploitation alleviated by a water transfer from the upper Tagus. The fuzzy logic systems chosen were Sugeno of order 1, with two inputs: the raw meteorological forecasts and the month of the year. The same grid as ERA5 was considered, and for each point the fuzzy logic systems were trained so that the forecasts monthly cdfs match the ones from ERA5. The training process took on average 20 minutes per point and variable with a standard computer, and results show that the post-processed cdfs closely match the ERA5 cdfs. Furthermore, the skill of the post-processed forecasts was evaluated using the Mean Absolute Error (MAE) and compared to the skill of raw forecasts to assess the adequacy of the post-processing.

Acknowledgements:  This study has received funding from the European Union’s Horizon 2020 research and innovation programme under the GoNEXUS project (grant agreement No 101003722); the eGROUNDWATER project (GA n. 1921), part of the PRIMA programme supported by the European Union’s Horizon 2020 research and innovation programme; and the subvencions del Programa per a la promoció de la investigación científica, el desenvolupament tecnològic i la innovació a la Comunitat Valenciana (PROMETEO) under the WATER4CAST project.