Evolutionary Neural Architecture Search via Enhanced Weight Sharing for Distributed Photovoltaic Output Prediction

With the advancement of smart grids and distributed energy systems, more photovoltaic (PV) units are being integrated into distribution networks, making accurate forecasting vital for grid stability. However, geographical differences lead to diverse PV output patterns, challenging traditional models with fixed architectures. To address it, this proposes an attention-based LSTM model combined with an evolutionary neural architecture search framework using supernet weight sharing. This method captures the temporal features of the signal via Variational Mode Decomposition followed by fuzzy entropy–based reconstruction, then aggregates the reconstructed outputs through an LSTM to form the final prediction. To enhance the model's adaptability, an evolutionary search strategy with enhanced weight sharing in the supernet is employed to dynamically optimize the attention-equipped LSTM architecture according to regional conditions. Experimental results show that the proposed method achieves improved accuracy and robustness compared to conventional baselines.