Abstract
Optimising energy use and storage from PV panels requires accurate solar forecasting. This paper compares machine learning approaches for two practical PV forecasting settings: multi-horizon prediction and distributed-system forecasting. Through this research, we aim to address the problem of implementing complex prediction techniques in micro-grid operations by comparing the performance of different approaches and levels of data availability. Experiments on a three-building Flinders University dataset show that temporal features dominate single-site accuracy, LSTM models perform best at the 1-hour horizon, and simpler horizon-conditioned or discrete strategies remain close to specialised references at longer horizons. For distributed forecasting, spatial analysis finds no exploitable lead-lag signal at hourly resolution among the three-building system. Graph-based models instead provide their strongest benefit in cold-start settings, where shared representations give strong zero-shot performance for new PV systems.

This publication has 21 references indexed in Scilit: