Conditional Multi-Step Prediction for Distributed Photovoltaic Clusters Using Implicit Neural Motion Field with Wind-Modulated Fourier Layer Under Limited Local Meteorological Sensing

Regional meteorological information does not fully resolve local microclimatic variability, while moving clouds over varied terrain cause dynamic spatial heterogeneity in distributed photovoltaic (PV) power output. This study performs retrospective evaluation adopting observed future period meteorology as known inputs. This study proposes an implicit neural motion field (INMF) for multi-step conditional prediction of distributed PV power under limited local meteorological sensing. Learnable site coordinates, site embeddings, and radial basis function interpolation map discrete site states onto a regular latent grid. A neural ordinary differential equation evolves this field through a wind-modulated Fourier layer, in which wind speed conditions spectral amplitudes and wind direction conditions spectral phases. An auxiliary irradiance-consistency term regularizes the temporal relationship between predicted outputs and irradiance. The model is evaluated on 20 rooftop PV sites in Hong Kong using observed target-period meteorology. Joint wind modulation reduces RMSE by 8.9–18.5% across the four evaluated step horizons. In the ablation experiments, phase modulation and the irradiance-consistency term yield reductions in aggregated-series RMSE. The latent coordinates represent learned intersite relationships, and the evolving field provides a representation of PV dynamics under the specified meteorological conditions.