Contextual feature fusion convolutional transformer complementation for PV power prediction

Abstract

Accurate forecasting of photovoltaic power generation can facilitate the integration of photovoltaic into modern power systems. In this paper, a Contextual Feature Fusion Convolutional Transformer Complementary for the Photovoltaic Power Generation Prediction Model is proposed. Historical photovoltaic data, historical weather, and predicted weather data are input for normalization and convolution operations. The computed positional encoding is embedded into the convolved feature information. The feature information encoded in the embedded position is fed into the Feature Complementary Module, and the local and long-dependent features are extracted using a Convolutional Neural Network and Transformer, respectively. Complementarity between features is achieved. Contextual feature fusion is utilized to enhance the correlation between different features. Finally, the final output is the predicted value of PV power generation at 24 moments of a given day. The experimental results show that compared to other prediction models on Ausgrid, OpenWeatherMap, and Solcast datasets, the proposed model reduces to 0.0059, 0.0208, and 0.2107 in terms of mean absolute error, mean square error, and weighted mean absolute percentage error.