Causal Mechanism-Enabled Zero-Label Learning for Power Generation Forecasting of Newly-Built PV Sites

Abstract

Power forecasting of newly built photovoltaic (PV) sites faces huge challenges owing to the lack of sufficient training samples. To this end, this paper proposes an unsupervised zero-label learning method for power generation forecasting of newly built PV sites without relying on any historical power output data. The main idea is to extract invariant causal structures across different PV sites and leverage this causality to enhance the power forecasting performance on the newly built ones. In particular, a causality-enabled domain adaptation network (CEDAN) is designed to capture the causal mechanism of PV generation from the multiple fine-grain segments of time-lagged data. It relaxes the causal relationships to an associative structure which is further concretized as attention score vectors through the designed intra- and inter-variable attention mechanisms. To quantify the distribution discrepancies between source and target domain causal structures, a specific domain adaptation loss function is designed for the optimization of CEDAN. It is further extended to a domain adaptation quantile loss to handle the uncertainties of PV power output. By jointly minimizing the domain adaptation loss and power forecasting error/conditional quantile loss, an invariant power generation causal mechanism across data domains for a newly built PV site can be learned. This allows the proposed method to achieve accurate and highly generalized power generation forecasting for newly built PV sites without relying on labeled data. Extensive experiments utilizing real PV generation data demonstrate that the proposed method surpasses state-of-the-art transfer learning methods by 7.57% at least in deterministic forecasting and 8.37% at least in probabilistic forecasting.