Online learning-based mutually-aided state estimation and forecasting for PV

Traditional photovoltaic (PV) forecasting algorithms rely heavily on historical generation data or physical models, both of which are frequently unavailable or unreliable in practice. This paper addresses this challenge by proposing an online learning-based mutually-aided state estimation and forecasting (MASEF) algorithm, which integrates state estimation (SE) and PV forecasting into a coupled stochastic system and operates effectively without site-specific historical PV generation data or physical models. Inspired by the Kalman filter (KF), a mutually-aided online learning loop is established in the MASEF algorithm: a Bayesian neural network (BNN) approximates PV generation (prediction stage), which allows SE to refine the system state (update stage). These estimates are then fed back to the forecaster for online learning. The BNN further enhances reliability by providing probabilistic outputs and quantifying uncertainty. The forecasted values serve as pseudo measurements, which are critical in scenarios with limited observability or noisy data. Case studies demonstrate that the MASEF algorithm achieves performance competitive with state-of-the-art algorithms even in the complete absence of PV generation data or physical models. Furthermore, the results confirm the robustness of the MASEF algorithm to measurement noise and parameter inaccuracies, highlighting its adaptability and practical applicability in diverse power grid environments.