Energy Price Prediction with CEEMDAN-STL and Hybrid Neural Network using Time2Vec Attention-based CNN-biGRU

In dynamic energy markets, accurate energy price prediction is crucial for risk management, resource allocation, efficient and fast decision-making. This study proposes a novel hybrid neural network architecture combining Time2Vec coding method, convolutional neural networks (CNNs), attention mechanism (AM) and bidirectional gated recurrent units (biGRUs). This architecture is complemented by input data processing techniques using complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) and seasonal and trend decomposition using locally estimated scatterplot smoothing (STL). CEEMDAN captures nonlinear and nonstationary patterns inherent in the input data. STL is widely used to isolate seasonal components. The Time2Vec encoding method is effectively utilized to capture temporal patterns, transforming time-related features into continuous embeddings. CNN block extracts the features of the data and AM improves the interpretability of the model by focusing on significant features. Further, biGRUs read data dependencies from past and future steps. The Time2Vec-CNN- AM-biGRU architecture learns hierarchical representations of data features, exploiting spatial and temporal dependencies in the data. Experimental open-loop forecasting on real electricity price datasets demonstrates the superiority of the proposed approach over baseline models in terms of prediction accuracy.