Multi-variable photovoltaic power prediction method based on deep attention convolutional neural network

The uncertainty of photovoltaic (PV) power generation systems is a critical factor affecting grid integration rates, making accurate PV power forecasting essential for ensuring stable grid operation. This paper proposes a multivariate information fusion method for PV power forecasting, named CNN-BiLSTM-Attention, which integrates a Convolutional Neural Network (CNN), Attention Mechanism, and Bidirectional Long Short-Term Memory (BiLSTM). First, leveraging the feature extraction capabilities of CNN, temporal dependency features are extracted from Numerical Weather Prediction (NWP) data, including Global Horizontal Irradiance (GHI), relative humidity, temperature, wind speed, and other meteorological factors. BiLSTM is then employed to capture the long-term dependencies and bidirectional characteristics in the time-series data. The Attention Mechanism is used to achieve adaptive fusion of NWP data with historical PV power data, thereby enhancing forecasting accuracy. Finally, simulation experiments validate the effectiveness of the proposed method.