Exploring the Efficacy of Feature Selection and Attention Enhanced Deep Learning in Multivariable Electricity Price Forecasting

Accurate electricity price prediction has become critical for operational decision making and strategy making. In recent years, machine learning techniques have emerged as an effective solution for the day ahead electricity market price forecasting problem. However, despite their capability to make accurate predictions, they lack transparency and often leave the researchers with a non-understandable black box impression. Electricity prices present dependance on several external factors, and deep learning techniques, with their inherent capabilities to deal with temporal dependencies and non-linearity, offer solutions to these multivariable dependent forecasting problems. This paper explores the efficacy of integrating advanced feature selection techniques and attention-enhanced deep learning models in multivariable electricity price forecasting.A robust methodology, combining Pearson correlation matrix-based feature selection Stacked Long-Short Term Memory (SLSTM) network followed by an attention mechanism, has been proposed to make the process more transparent. This study validates the efficacy of proposed methodology on real world electricity market dataset of Sweden. This comprehensive dataset encompasses multiple variables influencing electricity prices, s.a. electrical demand and weather data. The study further compares two scenarios, one inclusive of all the features and the other with selective features and attention-based approach. The results underscore the importance of effective feature selection and the incorporation of attention mechanism in enhancing the predictive capabilities of deep learning models for electricity price forecasting. The proposed methodology not only generates more accurate forecasts than the base case but also results in a reduction in computational time.