A Comprehensive Analysis of Dropout Assisted Regularized Deep Learning Architectures for Dynamic Electricity Price Forecasting

As the world moves towards a sustainable future, there is a need for efficient integration of intermittent energy sources (IES) sources with intelligent energy grids. Electricity from IES (such as wind, solar, hydro, etc.) relies on various environmental factors and introduces significant uncertainties. This, in turn, impacts the financial viability of the investment plan for energy utilities and demands the deployment of efficient electricity price forecasting (EPF) approaches. Many existing EPF approaches show limited performance due to significant nonlinearity, high volatility, and continuous evolutions of exogenous variables. For such applications, data-driven approaches such as deep learning (DL) have shown promising forecasting capabilities in recent years. Therefore, this work comprehensively analyses the performance of various DL architectures for EPF across five major global deregulated electricity markets. Furthermore, to improve the performance, we propose a distinct approach by integrating Monte Carlo dropouts and weighted constraints in a DL architecture. The performance of model is assessed against regression-based models and benchmark DL architectures in terms of error matrices and computational time. The results demonstrate that the presented methodology enables the identification of the most appropriate DL architecture and demonstrates improved forecasting performance for electricity markets when compared against benchmarks.