Short-term regional wind power forecasting for small datasets with input data correction, hybrid neural network, and error analysis

The security of power systems and electrical grids can be affected by the stochastic nature of wind energy. Therefore, reliable techniques for load forecasting and planning must be developed. This paper presents a model for short-term regional wind power forecasting based on small datasets. The model comprises three steps: input data correction, a hybrid neural network, and error analysis. First, regional numerical weather predictions (NWP) are corrected by using a stacked multilevel-denoising autoencoder (SMLDAE) to generate more effective inputs; this is the first study to use SMLDAE for NWP data correction. Second, a neural network-based hybrid model is employed for regional wind power forecasting to predict the wind power in the region. The proposed hybrid model employs three processes: multiscale mathematical morphological decomposition (MMMD), k-means clustering, and a stacked denoising autoencoder. MMMD can decompose the data directly in the time domain, thus, the signal does not need to be transferred from the time domain to the frequency domain to accomplish the decomposition. Third, a long short-term memory network is used for error analysis of the preliminary forecasted data. The preliminary results and error series are aggregated to generate the final forecasting result. For small datasets, we use multi-distribution mega-trend diffusion to augment the dataset. The proposed model was validated using a dataset consisting of data generated by regional wind farms in northern China. The results show that the proposed model enables wind forecasting at both the regional and single-farm level. Moreover, whereas most benchmark models require almost one year of data, the model requires only approximately three months of NWP data to produce reliable forecasting within the next 24 h.