In our recent work, we developed a forecasting method for residential PV generation that leverages the similarity of direct radiation profiles over two years of historical data. For each forecasted day, the method identifies the 10 most similar days, then calculates hourly prediction intervals using percentiles.

Applied to a real 4.8 kWp PV installation in Spain, the method achieved high accuracy and robustness under various weather conditions, outperforming more complex approaches in terms of interpretability and computational simplicity.

Sometimes, simplicity is the real power.

I’m currently finalizing a new paper where this forecasting approach is integrated into smart home energy management for optimal scheduling of shiftable loads and battery storage in hybrid systems, including a practical demonstration with smart plugs.

Figures show the parameter selection process and the PV generation forecast method workflow.

What are your thoughts on similarity-based methods compared to traditional ML models for short-term PV forecasting?

Article Interval-Based Solar Photovoltaic Energy Predictions: A Sing...