Optimized CNN–BiGRU–BiLSTM Architecture with Variational Mode Decomposition and Attention Mechanism for Electricity Price Forecasting

Abstract With the advent of deregulation in electricity markets, price forecasting has become an important task for generation and retail companies to enhance economic efficiency and market competitiveness. Traditional statistical models are widely employed, but they struggle to learn the volatile and complex characteristics of electricity prices, which are further complicated by integration of renewable energy sources and fluctuations in demand. To address these issues, deep learning methods have gained attention due to their unique ability to effectively learn temporal and spatial dependencies. This study introduces a novel hybrid model integrating a Convolutional Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU), and Bidirectional Long Short-Term Memory (BiLSTM) networks augmented with attention mechanism for enhanced short-term electricity price forecasting. The CNN extracts detailed spatial features, BiGRU captures bidirectional short- and mid-term temporal dependencies, and BiLSTM models longer-term patterns. The attention mechanism dynamically adjusts feature weights across different time scales and spatial contexts, improving the robustness, interpretability, and accuracy of the model for non-linear stochastic electricity price data. An advanced pre-processing pipeline incorporates data cleansing, outlier detection, and imputation, while Variational Mode Decomposition (VMD) generates compact, noise-free, and variance-stable datasets that facilitate the learning of rich, multi-scale temporal features. Furthermore, Honey Badger Algorithm optimizes the model’s hyperparameters to enhance performance stability. Evaluations against state-of-the-art benchmarks demonstrate highly competitive performance in most seasons except in Spring due to increased price volatility from higher renewable generation integration during this period, further supported by statistical validation using the Diebold–Mariano test. This comprehensive framework provides a reliable and high-performing solution for capturing complex electricity price dynamics, offering significant potential for enhancing power market operations.