Short-Term Wind Speed Forecasting Based on a Novel Multiresolution Neural Network

As the penetration of wind power grows at an unprecedented pace, short-term wind speed forecasting has become indispensable. However, in practical power system applications, the temporal resolution of wind speed measurements often exceeds that required for system dispatch. To leverage these high-resolution measurements for generating lower‐resolution wind speed prediction compatible with dispatch requirements, a novel multi-resolution neural network is first proposed. This method consists of three stages: (1) Multi-resolution time series are constructed from the original wind speed series. (2) A low-resolution feature extraction module based on an improved scalar long short-term memory (sLSTM) network is employed to capture the trend patterns in the low-resolution data, while a high-resolution feature extraction module based on an improved modern temporal convolutional network (modern TCN) is designed to extract fluctuation features from the high-resolution data. (3) Both high-resolution and low-resolution features are fused through a sparse cross-attention mechanism. In the test case, three datasets are used for effectiveness verification, while the main test results indicate that the proposed model well improves the short-term wind speed forecasting accuracy.