Predicting PV Power with a Multi-Stage Attention Neural Network Based on Neural Ordinary Differential Equations at Egyptian Stations

To enable the integration of photovoltaic (PV) power into electrical grids, accurate predictions are vital. This study applies the Multistage Attention Neural Ordinary Differential Equation (MANODE) model, which combines Long Short-Term Memory (LSTM) networks, Temporal Convolutional Networks (TCN), and a two-stage attention mechanism to capture complex spatiotemporal patterns for PV power forecasting. The improved MANODE model is evaluated on three real-world datasets from PV stations in Egypt. Each dataset contains 12 feature parameters and spans an entire year. Comprehensive comparisons are conducted between the improved MANODE model and other neural network models, including one-layer and two-layer FNNs and time-series models, as well as a curve-fitting method. The proposed model achieves high correlation and negligible mean squared error (MSE) in one-step-ahead hourly forecasting for three locations in Egypt. Overall, MANODE demonstrates strong performance and represents a robust deep learning framework for estimating PV system power requirements, with effective real-time modeling and feature representation under changing weather conditions.