Multi-Step Photovoltaic Power Forecasting Based on PatchTST_LSTM_CrossAttention

In the global transition towards a clean, low-carbon energy structure, photovoltaic (PV) power generation emerges as a critical renewable energy source, facing challenges such as high volatility, uncertainty, and multivariable coupling. To address these issues, this paper proposes a multi-step prediction model that integrates a PatchTST encoder, a cross-attention mechanism, and LSTM structure. Through feature analysis, key variables are selected for effective dimensionality reduction and suppression of feature redundancy. The model employs PatchTST for local temporal feature modeling, uses the cross-attention mechanism to capture dynamic coupling between primary and exogenous variables, and leverages the LSTM structure to enhance the capture of long-term dependencies. Experimental results indicate that the proposed model surpasses mainstream methods like GRU, Informer, KAN, and TSMixer in terms of multi-step prediction accuracy and stability on the Australian DKASC dataset, especially showing significant advantages in medium to long-term forecasting tasks. This research provides a new approach to improving the accuracy and generalization capabilities of PV power prediction, offering valuable applications for new energy integration and energy dispatch management.