Physics-Constrained Feature Fusion for Photovoltaic Construction Progress Monitoring Using WaveletKAN-GRU

Source abstract

Large ground-mounted photovoltaic power stations contain concurrent civil, mechanical, and electrical work fronts whose visible appearance does not always correspond to accepted engineering progress. This study proposes a WaveletKAN- GRU model with physics-constrained feature fusion for joint construction-stage recognition and continuous progress estimation. Multisource observations are represented by spectral, texture, geometric, schedule, and environmental descriptors. A soft fusion gate weights this evidence using terrain slope, row geometry, expected shadow behavior, installation order, irradiance consistency, and work-front capacity before temporal inference. Learnable wavelet mappings emphasize local frequency changes caused by emerging pile, bracket, and module rows, while a gated recurrent unit models irregular observation intervals and delayed dependencies between civil and electrical activities. Experiments use a simulated but engineering-realistic benchmark representing 18 project zones, six construction stages, irregular acquisition gaps, and terrain-dependent visibility disturbances; the site-independent split prevents observations from one zone from crossing training, validation, and test partitions. The proposed model achieved 94.8% stage accuracy, a macro F1-score of 0.936, a progress mean absolute error of 2.7 percentage points, a root mean square error of 3.8 percentage points, and a physical violation rate of 1.9%. The closest KAN-GRU baseline achieved 92.9% accuracy and a 3.8-point mean absolute error. Ablation results show complementary roles for wavelet mapping and physically constrained fusion. These controlled results support the method's potential for progress screening and targeted inspection, but independently acquired multi- project data are required before operational deployment can be established.