Embedding of time series for the prediction in photovoltaic power plants

The energy produced by photovoltaic systems is valuable, but also unpredictable. Its variability, driven by weather conditions and natural cycles, poses a daily challenge for grid operators. How can we make reliable forecasts to optimize energy distribution? The answer lies in three advanced approaches that combine neural networks, fuzzy logic, and nonlinear modeling.

At the core of the method is a technique called embedding, which transforms historical energy production data into vector representations that capture the system’s evolution over time. Each data point is described using its past values, with parameters carefully selected to account for the complexity and cyclicality of solar behavior.

Three predictive models were tested on this foundation. Adaptive Neuro-Fuzzy Inference System (ANFIS): a neuro-fuzzy network that blends logical rules with machine learning. Mixture of Gaussian (MoG): which models the input-output relationships as a mixture of Gaussian distributions. Radial Basis Function (RBF): a radial basis function network ideal for capturing nonlinear structures in the data. These models were trained on real-world data from a large photovoltaic power plant in southern Italy, with precise forecasts for single days using training windows of either 7 or 30 days. Performance was assessed using standard metrics (MSE, NMSE, NSR), revealing that all models can deliver accurate predictions under normal operating conditions.

The RBF model proved particularly effective in the short term, leveraging the intrinsic periodicity of solar irradiance, while ANFIS stood out for its ability to model more complex behaviors. Surprisingly, the linear model (LSE) also performed well over short timeframes, confirming the influence of daily regularities.

Predicting how much energy a solar plant will produce isn’t just a scientific challenge, it’s a practical necessity for ensuring grid stability and maximizing the value of renewable resources. Thanks to these AI-based solutions and advanced time series analysis, we now have one more tool to face the intermittency of the sun with intelligence.