Photovoltaic Power Prediction Using a Hybrid Method Combining FFT and ANN

A crucial element in the worldwide shift towards cleaner energy, renewable sources present a lasting alternative to traditional energy, significantly reducing carbon emissions and mitigating climate change. Solar energy, in particular, stands out as a highly advantageous resource, fostering technological progress, economic growth, and energy self-sufficiency. However, in order to accurately predict power generation, reliable and robust predictive models are required for improving accuracy and model performance in real-world practical solar applications. Sophisticated forecasting techniques must be developed due to the inherent intermittency of solar radiation and the growing demand for renewable energy. This study proposes a hybrid method for solar power prediction that combines Artificial Neural Networks (ANN) and the Fast Fourier Transform (FFT). The dataset used comprises 30 days of measurements, with 15 sampled values of global irradiance, ambient temperature (AT), and module temperature (MT) per day. To extract dominant periodic components, daily power profiles are first converted into the frequency domain using the FFT. In order to minimize noise while maintaining the crucial daily patterns of solar production, a frequency filtering procedure is then used to retain important low-frequency components. The input vector of the ANN model is then created by combining these extracted frequency features with meteorological data. The nonlinear relationship between environmental factors and photovoltaic power output is captured by the ANN through supervised learning. Statistical indicators such as RMSE (0.092), MAE (0.067), and the coefficient of determination (R² = 99.44%) are used to evaluate the model using independent test data. An improvement over a traditional ANN model is demonstrated by a comparative analysis (R² = 98.67%). Nevertheless, because of the comparatively small dataset and the lack of thorough testing under various operating conditions, the validation of the suggested approach is still restricted despite these encouraging results. To properly illustrate the robustness and generalization potential of the suggested approach, more research on bigger and more diverse datasets is required.