Diffusion-driven adaptive radiance refinement with PPO-based optimization for robust solar irradiance forecasting

• Introduces a novel fusion of diffusion-based refinement and reinforcement learning to enhance solar irradiance forecasting and PV control. • PPO-based controller continuously self-corrects prediction errors in real time, ensuring adaptive and resilient PV system performance. • Achieves around 8% improvement in prediction accuracy and notable reduction in forecast volatility compared to advanced deep learning and hybrid models. • Delivers a scalable, real-time, and autonomous forecasting-and-control framework suitable for smart grid and distributed solar energy applications. Rapid global transition toward renewable energy sources has amplified the importance of accurate solar irradiance forecasting for reliable PV system operation. Conventional models such as LSTM, ANN, and optimization-based hybrids face challenges in capturing nonlinear irradiance variations, mitigating atmospheric noise, and adapting to sudden weather changes. To address these limitations, this study introduces a novel Diffusion-Based Adaptive Radiance Refinement Layer (ARRL) integrated with Proximal Policy Optimization (PPO) to enhance denoising, adaptivity, and real-time error correction. The ARRL suppresses stochastic irradiance noise, while PPO continuously refines prediction outputs through reinforcement-guided optimization. The proposed framework was implemented in Python using TensorFlow 2.15 and evaluated on the Kaggle Solar Irradiance and Weather Forecasting Dataset, consisting of 1000 temporal records sampled at 30-minute intervals. Experimental results reveal that the proposed Diffusion and PPO model achieved a MAE of 10.28 W/m², RMSE of 14.17 W/m², MAPE of 2.25 and an R² score of 0.97, representing an approximate 8% improvement in prediction accuracy compared to advanced LSTM and hybrid optimization benchmarks. Moreover, the framework demonstrated robust generalization under cloud-induced fluctuations and significant reductions in forecast volatility. The synergistic fusion of diffusion denoising and reinforcement-based adaptive learning delivers a self-correcting and scalable solution for solar forecasting. In conclusion, the proposed model establishes a highly interpretable and resilient architecture, setting a new direction for intelligent, autonomous, and data-driven solar irradiance prediction systems in dynamic environmental conditions.