This study introduces a novel approach for predicting photovoltaic efficiency based on the MDC-Informer model. First, three-dimensional convolution is used to extract feature variables from photovoltaic and meteorological inputs, constructing a high-dimensional mapping feature vector. To enhance the model’s feature extraction capability, the convolution module in the MDC-Informer architecture incorporates a series of three 3D convolutional layers. The encoder and decoder components of the Informer framework process the feature maps generated by the convolutional module. Subsequently, the outputs from these components are merged with the original input features, which are then passed through a dense layer for further processing, mapping the feature space obtained from the convolution module to the sample space, thus yielding the final prediction results. The experimental findings demonstrate that the MDC-Informer model achieves superior prediction performance compared to other network architectures. Additionally, it exhibits consistently high accuracy across weather conditions with similar patterns, such as sunny, cloudy, and rainy days.