Real-time ultra short-term irradiance forecasting using a novel R-GRU model for optimizing PV controller dynamics
N. B. Sushmi, D. Subbulekshmi
Abstract
• A new R-GRU model with dropouts and L2 regularization was proposed for ultra short-term irradiance forecasting of rooftop solar panels at VIT Chennai. • Efficient feature selection and hyperparameter tuning enhanced the prediction accuracy of the model. • The model outperformed ANN, DNN, RNN, LSTM, and GRU in all key metrics offering faster training and lower computational demands. • Real time setup demonstrates the efficient working of the model on test samples, compared to benchmark models. • The power potential and optimal duty cycle of a 250 W PV panel were calibrated and analysed against other models based on predicted irradiance, providing a reference framework for PV controller optimization application. Solar radiation intermittency poses significant challenges for photovoltaic (PV) power generation, especially in grid integration and trading. Therefore, real-time solar irradiance forecasting is crucial for optimizing PV system performance. This study proposes a novel R-GRU-based ultra-short-term irradiance forecasting model that incorporates key environmental variables. To mitigate overfitting, the model utilizes dropout and L2 regularization, while feature selection and hyperparameter tuning further enhance prediction accuracy. The performance of the model is validated on rooftop panels at VIT Chennai, where it outperforms SNN, DNN, RNN, LSTM, and GRU models, achieving RMSE values of 0.0959 and 0.0731, R-squared values of 95.65 % and 98.01 %, MAE values of 0.05225 and 0.0519, NSE values of 90.82 % and 94.57 %, and MBE values of -0.00000392 and 0.0000051 for training and real-time testing, respectively. With a training time of 37 s and 16 epochs, the model ensures a rapid build process and precise predictions. Furthermore, the power output and optimal duty cycle corresponding to the predicted irradiance were estimated, and serve as an effective reference for the application of optimizing PV controller dynamics.