Towards efficient model recommendation: An innovative hybrid graph neural network approach integrating multisignature analysis of electrical time series

In the energy sector, it is important to meticulously choose an accurate forecasting model because making informed decisions is crucial for optimal grid operation. This article proposes a hybrid graph neural network (GNN) that successfully captures complex patterns by combining interactions based on features and time. The proposed architecture uses diverse decomposition methods, such as statistical, dynamic, and spectral, to uncover hidden patterns. The integration of attention and graph convolution layers improves the flow of information, and the cross-modal fusion layer competently combines nodes and edges. This configuration can efficaciously assimilate disparate features, providing an advantage over other approaches. The hybrid GNN with an analytic hierarchy process (AHP) performs better than the existing models when tested exhaustively on diverse regional energy demand and pricing time series. It captures complex patterns more effectively than K-nearest neighbour (KNN), random forest regressor (RFR), and GNN-based model suggestion methods. For the nonstationary price data from New South Wales and the Punjab energy demand time series, the Kendall's Tau coefficient is 0.73 and 0.81 and the Spearman's Rank coefficient is 0.74 and 0.91, respectively. This paper advances the field of time series forecasting by offering a novel strategy for improving model proposals by efficiently merging hybrid GNN with multiple feature modalities.