This application relates to a high-resolution
reconstruction method for
photovoltaic power generation data. The method includes: collecting photovoltaic data at a resolution lower than a first preset level and time-aligned
weather data; performing frequency and
time domain modeling on the photovoltaic data and
weather data using
Fourier transform to obtain input
frequency domain features; selectively enhancing the input
frequency domain features using learnable
frequency domain channel attention to obtain a frequency-modulated sequence; and modeling the frequency-modulated sequence using a Mamba state-
space network to obtain reconstructed frequency domain features; restoring the reconstructed frequency domain features to a
time domain output through inverse frequency domain transform; and fusing the
time domain output with the original time domain features to generate photovoltaic data and weather features at a resolution higher than a second preset level based on the fused features. This solves the problems in related technologies, such as the lack of an effective multimodal
feature fusion mechanism, which makes it difficult to deeply integrate key external information such as meteorological data with the power sequence, and the bias in frequency domain modeling, tending to learn low-frequency trends while ignoring high-frequency details, leading to rapid fluctuation
distortion in the reconstructed
signal.