Improving short-term photovoltaic power forecasting with an evolving neural network incorporating time-varying filtering based on empirical mode decomposition

Abstract

Accurately forecasting photovoltaic power generation is essential for the efficient integration of renewable energy into power grids. This paper presents a novel, high-accuracy framework for short-term photovoltaic productivity forecasting, tailored to the climatic conditions of the Algerian region of El-Oued. The framework automatically adapts the neural network forecast using a nature-inspired algorithm, eliminating the need for manual adjustments. It first addresses the complex, non-stationary nature of photovoltaic generation by incorporating a time-varying filter based on empirical mode decomposition, which decomposes the original photovoltaic data into multiple low-frequency components. Particle swarm optimization is then applied to enhance key elements of the framework, including the neural network structure and input variables such as the extracted components of photovoltaic data and weather parameters, along with their historical values. This optimization process efficiently identifies the near-optimal model structure, resulting in an improved forecaster whose performance is validated using real-world data measured in El-Oued. The proposed framework demonstrates impressive accuracy, with a Normalized Root Mean Squared Error ranging from 2.96% to 4.8% for annual forecasts, 2.28% for summer forecasts, and 4.97% for generalization ability. Similarly, the Normalized Mean Absolute Error ranges from 1.89% to 2.89% for annual forecasts, 1.61% for summer forecasts, and 3.76% for generalization ability. The correlation coefficient is outstanding, between 99.9% and 99.96% for annual forecasts, reaching 99.97% for summer forecasts, and 99.67% for generalization ability. The study confirms the effectiveness of the proposed framework in enhancing network stability and power distribution.