Genetic optimization of time-delay embedding boosts recurrent neural network accuracy for photovoltaic time series forecasting.
Recurrent Neural Networks and Evolutionary Algorithms for Reliable PV Power Prediction
The growing diffusion of photovoltaic plants calls for increasingly reliable forecasting tools, capable of handling complex, nonlinear time series strongly influenced by seasonal and environmental factors. In this context, a forecasting system has been developed that combines recurrent neural models with evolutionary techniques, aiming to improve estimation accuracy while reducing the complexity of the input data.
Modelling Temporal Dynamics with Reservoir Neural Networks
At the core of the approach is the use of reservoir-based recurrent neural networks, designed to model complex temporal dynamics while maintaining an efficient learning structure. These networks are coupled with an intelligent mechanism for selecting historical information: not all past samples contribute equally to the prediction, and identifying a relevant subset becomes a crucial task.
Genetic Selection of Informative Time Delays
To address this challenge, a strategy based on genetic algorithms is introduced and used to automatically select the most informative time delays. This process enables the construction of compact representations of historical time series, reducing input dimensionality and enhancing the generalisation capability of predictive models.
Validation on Real Photovoltaic Production Data
The approach has been applied to real photovoltaic production data, analysed across multiple periods of the year to account for seasonal variability. The results show a systematic performance improvement compared to traditional forecasting techniques and other regression-based models, highlighting how evolutionary selection of past information leads to more stable and robust predictions.
Toward Adaptive and Intelligent Energy Management Systems
Overall, this integration of artificial intelligence and evolutionary optimisation demonstrates an effective way to tackle the complexity of energy time series, paving the way for smarter and more adaptive management systems in future power grids.
Authors
A. Rosato, R. Araneo, M. Panella
