Probabilistic prediction of photovoltaic power using multi-task learning and large language models

Solar power generation is inherently variable, making accurate forecasting crucial for integrating photovoltaic energy into the grid. A new study presents an advanced approach to predict photovoltaic power output, combining signal processing with deep learning. The researchers first utilize wavelet transform techniques to decompose complex time series data from photovoltaic systems into a more manageable form. This preprocessing step allows the subsequent models to focus on key temporal features that influence power generation.

In the core of the proposed method, the team employs a large language model (LLM) derived from Meta´s LLaMA architecture, specifically adapted for multi-task learning. By using a multi-task framework, the model simultaneously learns from multiple related prediction targets or tasks, enhancing its ability to generalize across different situations and data sets. Neural networks based on large language models are shown to harness the shared structures within the data, leading to significant improvements in probabilistic forecasting accuracy compared to traditional or single-task models.

The results indicate that this hybrid artificial intelligence system can better capture the stochastic nature of solar energy production. By outputting probabilistic forecasts, the approach provides more informative uncertainty estimates for operators and planners in the energy sector. The combination of wavelet-based signal decomposition and a meta artificial intelligence LLaMA model demonstrates a promising direction for next-generation renewable energy forecasting, helping utilities better manage supply and demand while integrating larger shares of solar power into the energy mix.