Mixer-Pvformer: A Hybrid Model Combining Graph Convolution and Transformer for Long-Sequence Photovoltaic Power Forecasting

With the accelerated growth of the renewable energy sector, accurately forecasting the dynamic changes in photovoltaic (PV) power generation has become essential for optimizing energy allocation and maintaining stable grid operation. However, current PV forecasting methods struggle with capturing the complex correlations between PV generation and auxiliary factors (e.g., meteorological conditions), and face challenges in modeling both short- and long-term dependencies in time series data. This paper introduces Mixer-Pvformer, a novel hybrid model combining Graph Convolution Networks(GCNs) with Transformer architecture to overcome these limitations. The model segments time series data into non-overlapping patches, employing GCNs to quantify correlations between PV power and auxiliary features via an adjacency matrix. Simultaneously, a Multi-Layer Perceptron and attention mechanisms enable the capture of local and long-range temporal dependencies. Experimental results on datasets from three locations in Alice Springs, Australia, demonstrate that Mixer-Pvformer significantly outperforms state-of-the-art models in forecast accuracy and generalization, offering a promising approach for long-sequence PV power forecasting.