Very short-term production prediction for photovoltaic plants using Temporal Convolutional Networks

Area of Science:

• Renewable Energy Systems
• Artificial Intelligence
• Meteorological Forecasting

Background:

• National-scale solar photovoltaic (PV) energy forecasting presents challenges due to high generation variability and spatial aggregation.
• Accurate very short-term forecasting is crucial for integrating fluctuating renewable energy sources into the grid.

Purpose of the Study:

• To evaluate Temporal Convolutional Networks (TCNs) for national-level solar production nowcasting.
• To compare TCN performance against linear regression baselines using historical weather data.
• To determine the impact of TCN architecture and weather features on forecasting accuracy at different temporal resolutions.

Main Methods:

• Utilized a deep learning architecture, Temporal Convolutional Networks (TCNs), employing causal and dilated convolutions.
• Developed multivariate models incorporating past solar irradiance and sun height observations.
• Compared TCN models with linear regression baselines using Spanish national solar production data.

Main Results:

• The multivariate TCN significantly outperformed linear regression for one-hour solar production nowcasting.
• Consistent accuracy improvements were observed with TCNs at the fifteen-minute horizon.
• Past weather inputs alone proved sufficient for accurate forecasting, negating the need for future meteorological predictions.

Conclusions:

• TCN-based models demonstrate practical applicability for national solar energy integration.
• The TCN architecture is the primary driver of accuracy gains at hourly resolutions, while weather features are key at finer scales.
• The study provides a data-driven method for feature selection in renewable energy forecasting.