Short-term Wind and Solar Power Forecasting via Integrated Frequency-Time Feature Analysis and Multi-scale Sliding Windows

To address insufficient frequency-domain feature extraction and dynamic feature capture challenges in renewable power forecasting, a convolutional neural network-bidirectional long short-term memory network (CNN-BiLSTM) algorithm integrating frequency-time domain analysis with multi-scale sliding window feature fusion was proposed. The methodology involved three phases: 1) raw data was denoised using an isolation forest algorithm, and high-relevance input features were selected through a weighted ensemble evaluation combining Pearson correlation coefficients and gradient boosting regression trees (GBRT); 2) temporal statistics were extracted via multi-scale sliding windows to capture dynamic evolution patterns; 3)the CNN-BiLSTM was augmented with fast Fourier transform (FFT)-based frequency-domain feature extraction and self-attention (SA) mechanisms for adaptive temporal weight allocation, enabling synergistic frequency-time representation learning and improved power prediction accuracy. Case studies at wind and solar farms in northern China demonstrate significant accuracy improvements. For wind power, the proposed model achieves average reductions of 35.43% and 31.47% in mean absolute error (MAE) and root mean square error (RMSE), respectively, and an increase of 8.04% in coefficient of determination (R²). For solar power, average reductions in MAE and RMSE are 31.48% and 19.39%, respectively, and R² improves by 2.19%.