Short-term photovoltaic power forecasting using meta-learning and numerical weather prediction independent Long Short-Term Memory models

Short-term photovoltaic (PV) power forecasting is essential for integrating renewable energy sources into the grid as it provides accurate and timely information on the expected output of PV systems. Deep learning (DL) networks have shown promising results in this area, but depending on the weather conditions and the particularities of each PV system, different DL architectures may perform best. This paper proposes a meta-learning method to improve one-hour-ahead deterministic forecasts of PV systems by dynamically blending the base forecasts of multiple DL models to learn under what conditions each model performs best. Four base models of different long short-term memory architectures are used to produce PV production forecasts without using numerical weather predictions, with the objective to enhance the generalizability of the proposed solution. The accuracy of the meta-learner is evaluated using three rooftop PV systems in Lisbon, Portugal. Results indicate that different base models perform best at different PV plants, and meta-learning can improve accuracy by up to 5% over the most accurate base model per plant and up to 4.5% over the equal-weighted combination of the base forecasts. These improvements are statistically significant and even larger during peak production hours.