In this project we were tasked to design, develop, test, and ship an economic optimization algorithm for a cogenerating power plant.

The challenge involved modeling the power plant's thermodynamics, along with its physical and operational constraints, in order to define a suitable action space for optimization.

Our team's unique mix of skills in both machine learning and the physical and engineering aspects of the system were instrumental in tackling this complex problem and developing an effective solution.

HelioSwitch developed algorithms for regression and analysis of hydroelectric power plant production and river discharge data. The tool developed involves the following 3 main steps:

• processing of hydroelectric generation data which includes, on top of standard quality check and anomaly detection, an advanced frequency analysis to detect spurious signal in the data provided;
• extraction of location-specific river discharge data coupled with a latent space analysis, highlighting the physical phenomenon underlying the raw data;
• generation of nonlinear transfer functions for each hydroelectric power plant between river discharge data, and the water inflow at the power plants, inferred from their production.

Our team successfully developed a scalable algorithm that predicts power production from run-of-river hydropower plants, utilizing minimal registry information (location, size, plant type), weather data, and historical generation data. The algorithm provided hourly forecasts for a one-week prediction horizon and evaluated the possibility of extending the horizon to two weeks.

We developed and tested machine learning methodologies with Illumia to infer the hourly dynamics of PODs in the delta-PRA, starting from their monthly integral, in order to improve PRA forecast accuracy.