Grammatical Neuroevolution for Efficient Multivariate Time Series Forecasting

Resumo

This work introduces Grammatical Neuroevolution (GNE), a novel approach that applies Grammatical Evolution (GE) to automatically design neural network architectures for multivariate time series forecasting. Traditional methods often depend on manual design or fixed templates, which may not capture the complex temporal and cross-variable dependencies in energy demand data. GNE uses a context-free grammar to define the architecture search space, covering layer types, neuron counts, activation functions, hyperparameters, and connectivity patterns. The system evolves sequences of production rules to generate diverse and potentially optimal architectures without human intervention. We evaluate GNE on a synthetic multivariate time series dataset using standard metrics such as Root Mean Squared Error, Mean Absolute Error, and R². Results show that GNE achieves superior forecasting accuracy, discovering a specialized architecture that significantly outperforms a traditional statistical model. This demonstrated ability to generate high-performance, tailored models is particularly advantageous for complex forecasting tasks and for deployment in performance-critical applications. These findings underscore the potential of grammar-based approaches for enhancing predictive modeling in complex domains.