Temporal Hierarchy Reconciliation Network for PV Power Forecasting

Photovoltaic (PV) stations are distributed spatially and their power forecasts play an important role in coordinated decision-making in power systems. For different decision-making requirements, power forecasts at various temporal scales are investigated. However, these forecasts are implemented independently. In this paper, we consider the inherent relation among these power forecasts at various temporal scales and propose a Temporal Hierarchy Reconciliation Network (THReN). THReN adopts a temporal hierarchical structure to generate coherent multi-site forecasts for various temporal scales. It integrates base forecasts and reconciliation processes in a trainable architecture that effectively captures both intra-level spatio-temporal correlations and inter-level temporal dependencies for a set of distributed PV stations. With a temporal hierarchy matrix S, THReN ensures prediction coherence while enhancing forecasting accuracy at multiple scales. Experiments on a synthetic dataset demonstrate that THReN consistently outperforms other competing methods.