To  address  the  issues  of  data  noise  interference,  feature  scale  discrepancy,  and  insufficient  multiscale meteorological  pattern  modeling  in  photovoltaic  power  forecasting,  a  prediction  method  based  on  dynamic  data preprocessing  and  gated  dense  multiscale  convolutional  neural  network  (GDMS-CNN)  is  proposed.  Firstly,  an  anomaly detection  mechanism  based  on  dynamic  sliding-window  Z-score  is  established,  and  missing  values  are  processed  via covariance-weighted  multivariate  interpolation.  Secondly,  an  adaptive  piecewise  normalization  algorithm  is  adopted  to eliminate  feature  dimensional  differences,  and  cloud-cover  correction  factor  and  atmospheric  attenuation  factor  are constructed to enhance physical feature representation. Finally, a GDMS-CNN is designed, wherein the feature extraction efficiency  is  optimized  by  depthwise  separable  convolution  modules,  densely  connected  dilated  convolution  blocks  are constructed to capture multiscale spatiotemporal correlation features, and an asymmetric gated channel attention mechanism is embedded to dynamically recalibrate feature weights. Experimental results demonstrate that the proposed method reduces the root mean square error (RMSE) by 16.4% compared with the optimal baseline model genetic algorithm-variational mode decomposition-echo  state  network  (GA-VMD-ESN),  and  by  43.4%  compared  with  the  traditional  random  forest.  The proposed method provides a novel solution for photovoltaic output forecasting and effectively enhances the reliability of power grid dispatching.