An Improved CNN Approach for Short-Term Day-Ahead New Energy Output Prediction

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.