Optimized Sequential Agentic AI-Guided Trainable Hybrid Activation Functions for Solar Irradiance Forecasting

Accurate solar irradiance forecasting is essential for seamlessly integrating renewable energy into the power grid. This study proposes a novel agentic AI framework for solar irradiance prediction, addressing the limitation of conventional static activation functions focusing on restrict adaptability and limited expressiveness when modelling non-linear relationships in time series data. To overcome these challenges, we design a trainable hybrid activation function whose weights are dynamically optimised by the Adaptive Moment Estimation (Adam) controlled by the agentic AI based on the prior performance feedback. The methodology is rigorously validated using real-world datasets from Singapore and France, offering critical insights into how agentic AI optimally adjusts the weights of hybrid activation functions. For the Singapore dataset, the GRU with attention using ReLU activation achieved a MAE of 23.56, whereas the proposed agentic activation function attained a lower MAE of 22.38 reflecting an improvement of approximately 5%. Similarly, in the France dataset, the GRU with attention paired with Leaky-ReLU yielded an MAE of 26.47, while the agentic activation function achieved 25.41, indicating a performance gain of nearly 4%. In isolated scenarios, traditional activation functions demonstrated marginally superior performance (within 1%), underscoring the nuanced dependency between activation dynamics and model architecture in specific contexts. To ensure a thorough evaluation, three distinct error metrics were employed across four state-of-the-art forecasting models, providing a robust framework for assessing the effectiveness of the guided hybrid activation function. The results underscore the function’s adaptability to varying model architectures and diverse dataset characteristics while maintaining superior prediction accuracy and reliability. This research contributes a promising agentic AI-guided hybrid activation function, advancing the field of solar irradiance forecasting by demonstrating its impact on learning efficiency and accuracy across diverse forecasting models. The findings pave the way for more intelligent, adaptive approaches to renewable energy integration.