A Sum-of-Squares Optimization Method for Learning and Controlling Photovoltaic Systems

This paper outlines a combination of two data-driven approaches leveraging sum-of-squares (SoS) optimization to: i) learn the power-voltage (p-v) characteristic of photovoltaic (PV) arrays, and ii) rapidly regulate operation of the companion PV inverter to a desired power setpoint. Estimation of available headroom in PV systems is critical to the task of providing ancillary services, and the proposed method puts forth a computationally tractable solution with minimal data needs for the same. In addition to providing this key contribution to application, from an algorithmic vantage point, we present an interior-point method to solve a linear regression reformulation of the original polynomial fitting problem with SoS constraints. We validate the proposed algorithms through time-domain numerical simulations (incorporating the PV source and a 15-th order inverter model) for a variety of large-signal disturbances (step changes in real-power demand, rapid changes in irradiance) and demonstrate that the method provides an effective strategy to concomitantly discover the p-v curve and seamlessly regulate operation to a desired setpoint.