Ultra-Short-Term Forecasting of Large Distributed Solar PV Fleets Using Sparse Smart Inverter Data

Ultra-short-term power forecasting for distributed solar photovoltaic (PV) generation is a largely unaddressed, highly challenging problem due to the prohibitive real-time data collection and processing requirements for a sheer number of distributed PV units. In this paper, we propose an innovative idea of forecasting the power output of a large fleet of distributed PV units using limited real-time data of a sparsely selected set of PV units, referred to as pilot units. We develop a two-stage method to address this problem. In the planning stage, we use the K-medoids clustering algorithm to select pilot units for the installation of real-time remote monitoring infrastructure. In the operation stage, we devise a deep learning framework integrating Long Short-Term Memory, Graph Convolutional Network, Multilayer Perceptron to capture the spatio-temporal power generation patterns between pilot units and other units, and forecast the power outputs of all units in a large PV fleet using the real-time data from the few selected pilot units only. Case study results show that our proposed method outperforms all baseline methods in forecasting for power outputs of individual PV units as well as the whole PV fleet, and the forecasting time resolution is not dependent on that of weather data.