Day-ahead photovoltaic power forecasting with multi-source temporal-feature convolutional networks

Abstract

Photovoltaic (PV) power forecasting technology enhances the absorption capacity of renewable energy. However, the PV power generation process is highly sensitive to fluctuations in weather conditions, making accurate forecasting challenging. In this paper, we propose a composite data augmentation method and a model that can effectively utilize the augmented data. The PV power generation process has a fluctuating nature over time, so an augmented sample set with temporal correlation was created. This was achieved by reconstructing meteorological features and screening measurements similar to historical meteorological conditions. To improve the feature extraction capability for multi-source heterogeneous data and the temporal modeling capability for fine-grained periods, a multi-source temporal-feature convolutional networks (MSTFCN) model is proposed. MSTFCN employs parallel convolution to capture local temporal patterns and improves global feature representation via a channel attention mechanism. Based on this, redundant information is suppressed by a cascading channel compression approach, and a temporal segmentation strategy is applied to model fine-grained temporal features. We conducted experiments on two publicly available datasets, and the results demonstrate that the proposed data augmentation method effectively improves the forecasting performance of the deep learning model. Moreover, MSTFCN achieves higher forecasting accuracy and exhibits stronger environmental adaptability than the compared models.