6th International Conference Energy & Meteorology: Abstract Submission

C3S-Energy: new hydropower datasets for European electricity mixes assessments (799)

Laurent Dubus 1 , Linh Ho Tran 2 , Salomon Obahoundje 1 , Alberto Troccoli 2 3
  1. OSIRIS, EDF R&D, Palaiseau, France
  2. World Energy & Meteorology Council, Norwich, UK
  3. University of East Anglia, Norwich, UK

Objective & Background

The C3S Energy [C3SE] operational service aims to deliver key information on historical, seasonal forecast and projection periods for climate-related indicators relevant to the European energy sector. This presentation will be focused on hydropower generation information, required in energy models for energy mixes assessments.

 

Method

C3SE builds on pre-operational demonstrators developed by the two proof-of-concept Energy Sectoral Information Systems CLIM4ENERGY and the European Climate Energy Mixes (ECEM). The new ERA5 reanalysis is used in replacement of ERA-Interim for the historical period (1979-2018), and to adjust biases in seasonal forecasts and climate projections. Hydropower generation data come from the ENTSO-E Transparency Platform.

Based on previous work in ECEM, hydropower generation from run-of-river (HRO, for 12 countries) and reservoirs (HRE, for 10 countries) are modeled using a simple approach. It consists in modeling HRO and HRE using Random Forests (RnF) models, with country average temperature (T2m) and total precipitation (TP) data. In addition to instantaneous T2m and TP, the models also use cumulated TP on optimized period lengths for each country. The initial model used only one accumulation period, determined by maximizing the correlation between the lagged precipitation and HRE/HRO. The methodology is however not robust when the training dataset’s length changes. We then improved the methodology by using a two-step approach:

First, RnF is run using TP accumulated on periods of 5 to 200 days, with 5 days increments. The variables importance is determined using the drop-out loss valuesand among the top 15 predictors, only the ones with more than 30 days difference are kept. RnF is then run a second time with this optimal subset of predictors, plus instantaneous T2m and TP.

 

Results

The two-steps approach allows to automatize the predictors choice, and is then more robust when new training data become available. Out-of-bag validation shows that our simple approach is quite efficient to estimate HRO and HRE, with correlation coefficients of the order of [0.50-0.98] and [0.40-0.92] respectively. Additional tests have been done, in particular to split the year into low/high generation periods, with different models. This “seasonal” approach improves the correlations for some countries, but at this stage, we think the training dataset is not long enough to implement it. Hence, the two-steps single model approach is kept as a reference. Once the models are set up, the full ERA5 period is used to reconstruct generation on 1979-2018.

 

Discussion

Of course, HRO and, more importantly, HRE generation are influenced not only by climate variables, but also by reservoirs management decisions. A new approach is under development, which consists in modeling the water incomes into reservoirs, rather than generation itself. This is feasible due to the availability of additional data in the ENTSO-E database. Sub-country level models are also under investigation, where relevant data is available.

The models will also be used with C3S multi-model seasonal forecasts and climate projections from EURO-CORDEX simulations.