Ultra-Short-Term PV Power Forecasting Based on CNN-BiLSTM-Attention with Principal Component Analysis

Accurate ultra-short-term photovoltaic (PV) power forecasting is essential for maintaining the stability of power systems. We propose a model integrating Convolutional Neural Network (CNN), Bidirectional Long Short-Term Memory (BiLSTM), and an attention mechanism with Principal Component Analysis (PCA) to enhance forecasting accuracy and computational efficiency. PCA is applied to compress weakly correlated meteorological features into a single composite factor, reducing input dimensionality. CNN captures local spatial patterns. BiLSTM learns bidirectional temporal dependencies. Attention module adaptively weights important time steps. The proposed CNN-BiLSTM-Attention model with PCA was tested on a dataset collected from Alice Springs, Australia. It achieves a Mean Squared Error (MSE) of 0.0689, a Mean Absolute Error (MAE) of 0.1033, and an R2 score of 0.9778, surpassing the benchmark in all metrics. Additionally, on the test set, the use of PCA significantly reduced the overall forecasting time from 5.31 s to 4.27 s, improving the model’s efficiency in real-time deployment scenarios.