Kernel Density Estimation and Quantile Regression for Uncertainty-Aware Short-term Solar PV Power Forecasting

Accurate forecasting of solar photovoltaic (PV) power is crucial for grid stability and efficient energy management. This study conducts a comparative analysis of four neural network models—Convolutional Neural Networks (CNNs), MultiLayer Perceptrons (MLPs), Long Short-Term Memory networks (LSTMs), and Gated Recurrent Units (GRUs)—for short-term PV power prediction using historical solar irradiance and power generation data. The analysis incorporates probabilistic forecasting to quantify uncertainty through confidence intervals. Results indicate that LSTM and GRU models outperform CNN and MLP in capturing temporal dependencies and delivering reliable forecasts with well-calibrated uncertainty intervals. Probabilistic forecasts in this study utilize Kernel Density Estimation and quantile regression to generate confidence intervals. The findings highlight the importance of balancing accuracy with uncertainty quantification to address solar energy variability and enhance energy planning and management.