Application and analysis of innovative hybrid model for short-term photovoltaic power forecasting

In today's era of surging energy demand, it is imperative for current and future generations that energy production is both clean and sustainable. Embracing renewable energy resources aligns with environmental awareness and sustainable development goals. Among these resources, solar energy stands out as a rapidly expanding source. Integrating energy generated from photovoltaic panels (PV) into the power grid necessitates accurate energy estimation to ensure grid reliability and stability. However, the intermittent nature of solar radiation presents challenges for precise prediction models. This thesis introduces a hybrid forecasting model for short-term solar energy power prediction and evaluates various decomposition methods. Utilizing data from the BSEU renewable energy laboratory up to 2021, we first assess the performance of existing decomposition models, including EMD, EEMD, IEMD, CEEMDAN, and SWD, when combined with multi-layer neural network (MLP) structures. In the second phase, we propose a hybrid approach combining the Kernel-ELM model with CEEMDAN. We decompose the data into its subcomponents using CEEMDAN, resulting in more stable data for our forecasting model. This approach yields short-term forecasts with greater accuracy. We compare the performance of our proposed model with different error metrics and discuss its superiority over conventional hybrid and deep learning-based models. In terms of the Root Mean Square Error performance metric, the proposed hybrid model dicreased performance ranging from 22.69% to 44.56% when compared to the highest model performance, across all seasonal months. Our results demonstrate that the CEEMDAN-Kernel-ELM approach outperforms other hybrid and deep learning-based models in short-term solar energy forecasting, as confirmed by statistical analysis and the Taylor Diagram.