EGU General Assembly 2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Seasonal streamflow forecasts fostering hydro power cascade operation applying the adaptive policy search framework

Christoph Libisch-Lehner1,2, Harald Kling1, Martin Fuchs1, and Hans-Peter Nachtnebel2
Christoph Libisch-Lehner et al.
  • 1Pöyry Austria GmbH, Vienna
  • 2Institute for Hydrology and Water Management, University of Natural Resources and Life Sciences, Vienna

Hydro power assets contribute a valuable share of carbon-free energy generation worldwide. Large reservoirs are able to store energy and, combined with pump-storage capacities, they will play an important role in the future’s energy mix. In the future, the stronger integration of volatile energy sources, like solar and wind energy demands the flexibility of hydro power plants. In general, the operation of hydro power plants is a multi-stakeholder and multi-objective dynamic problem related to critical infrastructure. This requires flexible and robust reservoir operation policies, defined as closed-loop release functions where the system state is the input and turbine flows are the response of the function. Recently, Evolutionary-Multi-Objective-Direct-Policy-Search (EMODPS) yielded promising control policies for water resources systems. EMODPS is a kind of machine learning approach that relies on long records, or stochastic streamflow replicates capturing a wide range of possible conditions. A stochastic streamflow generator should actually cover all possible conditions related to the state-action-space and inflates the optimization process. Furthermore, the search procedure can implicitly identify the "most representative" states of the system and tends to approximate a better solution for these states. States that are very rarely explored but can be very important for a reliable operation have little effect on the optimized policy. In addition, artificial neuronal networks (ANN) derived from EMODPS suffer under the curse of instable sections . This is because ANN's are good at interpolating, but bad at extrapolating actions from unobserved states in the training sequence. Thus, we extend the well-known EMODPS framework by an re-optimizing approach utilizing seasonal streamflow predictions. Periodically, the reservoir policies are re-optimized based on an ensemble of streamflow predictions and the actual reservoir water levels. This adaptive policy search (APS) approach is applied to a three reservoirs cascade under Mediterranean climate, where the energy market will play an important role in the future. First results show that the hydropower operation can be improved: energy generation can slightly be increased at clearly lower cost of flood risk compared to static robust policies.

How to cite: Libisch-Lehner, C., Kling, H., Fuchs, M., and Nachtnebel, H.-P.: Seasonal streamflow forecasts fostering hydro power cascade operation applying the adaptive policy search framework, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-3578,, 2020


Display file