Neural network-based forecasting and uncertainty analysis of new power generation capacity of electric energy

Abstract

The prediction of new energy generation is challenging due to its intermittency and uncertainty. To solve this, we propose a framework combining an optimized multiscale convolutional neural network (MSCNN) and long - short - term memory network (LSTM). MSCNN improves feature extraction with dynamic scale selection and deep residual modules. LSTM captures long - term dependencies better using bidirectional processing and attention mechanisms. We also introduce a fuzzy decision support system (FDSS) to handle prediction uncertainty. Our model outperforms ARIMA, SVM, Gradient Boosting, CNN, and RNN in hourly, daily, and weekly predictions. It also excels in uncertainty quantification and generalization, offering strong support for accurate new energy generation prediction and dispatch.