Multi-modal Prediction of Solar Power Generation in Various Provinces of China

This study pioneers a forecasting model for solar power generation tailored specifically for the provinces of China, employing a multi-modal data fusion strategy. Diverging from conventional models that rely solely on single-source data, this approach integrates a diverse array of data types including time-series power generation, stock market prices, and textual information from policy documents. This integration enhances the predictive capabilities of our model by leveraging the unique information provided by each data type. We propose a novel approach called Multi-modal Information Fusion (Multiformer), which utilizes an encoder-decoder architecture augmented with a self-attention layer to effectively capture and synthesize the interdependencies among the varied data inputs. This architectural innovation significantly reduces error propagation and improves model accuracy. Empirical validation of our model was conducted through monthly forecasts of solar power generation across various provinces, revealing that Multiformer outperforms existing state-of-the-art methods, including ITransformer and Informer, with a substantial 35% improvement in performance for a six-month forecast period. The success of Multiformer underscores its potential to generate accurate predictions that are crucial for energy policy-making and strategic planning.