A Review of Advancing Solar PV Forecasting with Deep State Space Models

Authors:

Dheeraj Sharma¹, Dr. Chandan Kumar¹

¹Department of Computer Science and Engineering, Career Point University, Hamirpur, H.P., India.

Corresponding author: chandansharmahmr@gmail.com

Abstract: As solar photovoltaic (PV) systems become a more significant part of modern power systems, there is a growing demand for more accurate and reliable prediction methods suited to the intermittent and uncertain nature of renewable energy generation. Many traditional statistical methods and classical machine learning models are ineffective for modeling stochastic behavior and nonlinear temporal relationships in solar PV data. With recent progress in deep learning, Deep State Space Models (DSSMs) have been developed to combine probabilistic state-space models with deep neural networks to effectively model hidden temporal dynamics and uncertainty. This review includes a detailed assessment of a state-of-the-art analysis of the DSSM approaches for solar PV forecasting, and the perspectives of these approaches in renewable energy systems. The theory that underlies the use of state-space models, the architecture of DSSM, latent state representations, observation models, and uncertainty quantification mechanisms are discussed. In addition, the data preparation methods, feature engineering methods, and the impact of exogenous variables on the performance of DSSM are systematically analyzed. Comparative analysis between DSSMs and traditional methods such as ARIMA, ANN, RNN, LSTM, and hybrid Deep learning (HDL) methods is also provided. Another area highlighted in the review includes applications in short-term forecasting, probabilistic forecasting, and multi-site solar forecasting. The existing problems, such as computational complexity, scalability problems, data quality, and interpretability, are seriously studied. Lastly, the future research directions, including emerging trends with transformers, graph neural networks, metaheuristic optimization, and explainable artificial intelligence, are discussed. Results show that DSSMs have great potential to enhance the forecasting accuracy and uncertainty management, which could facilitate the reliable integration of renewable energy and smart grid management.

Keywords: Deep Learning, Deep State Space Models, Time Series Forecasting. Renewable Energy, Probabilistic Forecasting, Uncertainty Quantification