Machine Learning Informed Optimization Applied to Pumped Hydro Energy Storage

The increased contribution of uncertain and fluctuating renewable generation, originating mainly from wind and photovoltaic sources, is substantially impacting the operation of power systems. In order to efficiently hedge against these uncertainties, there is a growing need for flexibility that can be provided by Pumped Hydro Energy Storage (PHES) plants due to their ability to quickly and cost-effectively respond to mismatches between generation and consumption. PHES systems generally use the pumping and release of water between two reservoirs at different elevations, respectively to store water when the load demand is low (typically at night) and generate electricity when the demand is high.
Accurately modeling the PHES operation is a challenging problem, arising from the fact that the PHES operation inherently couples electrical and water constraints through a non-convex and non-concave relationship. Including these characteristics in optimization models is thus associated with high computational requirements.
In this Master Thesis, we propose a new data-driven paradigm to encode the operating curves of PHES systems. Practically, we leverage regression-based supervised machine learning (ML) to learn the intricate relationship between electrical and water variables.
Firstly, multiple linear regression is studied due to its modeling simplicity but, by imposing a simple linear form, this approach suffers from a limited explanatory power. Then, the modelling power of neural networks (NN) is leveraged. Different architectures and activation functions are studied. Both methods achieve better ex-post profits on average than the state-of-the-art. The foremost advantage is the increased reliability of the two developed data-driven methods. The NN technique outperforms the linear regression approach in all the tested cases and can have a shorter solving time.