Optimizing solar and wind forecasting with iHow optimization algorithm and multi-scale attention networks.

Area of Science:

• Renewable energy systems
• Artificial intelligence
• Forecasting methodologies

Background:

• Deep learning models face challenges with high-dimensional feature spaces and hyperparameter sensitivity in renewable energy forecasting.
• These limitations lead to increased computational costs and reduced model generalization and robustness.

Purpose of the Study:

• To present a hybrid deep learning-optimization framework to address dimensionality and hyperparameter sensitivity in renewable energy forecasting.
• To leverage cognitively inspired metaheuristics, specifically the Binary iHow Optimization Algorithm (biHOW) and iHOW, for feature selection and hyperparameter tuning.

Main Methods:

• Utilized the Multi-Scale Attention Network (MSAN) for time series forecasting, adept at capturing multi-scale temporal dependencies.
• Employed biHOW for efficient feature selection, reducing model complexity and improving interpretability.
• Applied iHOW for fine-tuning MSAN's architectural and training parameters to optimize forecasting performance.

Main Results:

• The hybrid framework achieved high accuracy in wind and solar generation prediction, with initial MSEs of 0.0105 (wind) and 0.0976 (solar).
• biHOW reduced the average misclassification rate to 0.3925 (wind) and 0.4161 (solar), identifying compact feature subsets.
• iHOW further decreased MSEs to [Formula: see text] (wind) and [Formula: see text] (solar), outperforming other state-of-the-art metaheuristics.

Conclusions:

• The proposed iHOW-based optimization framework significantly enhances forecasting accuracy and computational scalability for renewable energy systems.
• This hybrid approach supports adaptive forecasting, crucial for intelligent energy management in smart grids.