Forecasting of PV Plant Output Using Interpretable Temporal Fusion Transformer Model

Authors:

Abstract
The stochastic nature of solar energy generation poses a challenge for grid operators, especially with higher penetration of solar-based renewables in the grid. This paper proposes an attention-based temporal fusion transformer (TFT) model for short-term (an hour ahead) photovoltaic (PV) power forecasting using available geographic data such as solar irradiation, temperature, and statistical features extracted from historical PV data. TFT utilizes a self-attention layer for long-term dependencies where recurrent networks are used for local processing. The model selects relevant features through a series of gating layers to achieve high performance for multi-horizon forecasting. The temporal fusion transformer model also provides interpretable insights into the temporal dynamics of different features. A real-world PV dataset has been utilized to compare the model performance with some other state-of-the-art forecasting models.

This publication has 18 references indexed in Scilit: