Half-hourly electricity price prediction model with explainable-decomposition hybrid deep learning approach

Abstract

Accurate prediction of electricity price ($EP$) is crucial for energy utilities and grid operators for enhancing the energy trading, grid stability studies, resource allocations and pricing strategies, thereby improving the overall grid reliability, efficiency, and cost-effectiveness. This study introduces a novel D3Net model for half-hourly $EP$ prediction, integrating Seasonal-Trend decomposition using LOESS (STL) and Variational Mode Decomposition (VMD) with Multi-Layer Perceptron (MLP), Random Forest Regression (RFR), and Tabular Neural Network (TabNet). The methodology involves applying STL to the $EP$ time-series to extract trend, seasonal, and residual components. The trend is predicted using an MLP model, the seasonal component is further decomposed with VMD into 20 Variational Mode Functions (VMFs) and predicted using an RFR model, and the residual component is decomposed with VMD and predicted using the TabNet model. Input features are identified using the Partial Autocorrelation Function , and models are optimized using the Optuna algorithm. The final prediction combines the trend, seasonal, and residual components’ predictions. Explainable Artificial Intelligence (xAI) methods were used to enhance model interpretability and trustworthiness, with optimization via the Optuna algorithm. Comparative analysis with seven standalone and seven decomposition-based models confirmed the superior performance and statistical significance of the D3Net model. The D3Net achieved the highest global performance indicator for South Australia ($GPI\approx 11.068$) and Tasmania ($GPI\approx 12.206$). These results validate the efficacy and statistical significance of the D3Net model, demonstrating the viability of integrating STL and VMD decomposition approaches with MLP, RFR, and TabNet for $EP$ prediction.