MFFDM-WLS: A multi-granularity feature-based coherent forecasting method for temporal hierarchical wind speed time series

Wind energy, known for its clean and sustainable characteristics, has become an integral part of the global energy system. However, the intermittency and fluctuation of wind speed introduce significant uncertainty in wind power generation, posing challenges for grid integration. Additionally, multi-granularity wind speed forecasting can provide richer information compared to single-granularity forecasting, which is more favorable for wind farm operation and planning. Therefore, to further enhance the accuracy and reliability of wind speed forecasting and to obtain multi-granularity forecasts that satisfy the hierarchical consistency, MFFDM-WLS, a multi-granularity feature-based coherent forecasting method for temporal hierarchical wind speed time series, is proposed in this study. First, a multi-granularity feature fusion-based deep model (MFFDM) is proposed to obtain the base forecasts. MFFDM employs a bottom-up self-attention module and a top-down adaptive decomposition module to interact the wind speed features at different granularities, and utilizes the squeeze-and-excitation network and residual block to obtain features at each granularity, and finally generates the deterministic and probabilistic base forecasts using three loss functions. Then, seven common reconciliation techniques are compared, and the best reconciliation technique is determined based on their ranks to obtain the final reconciled forecasts. Experimental results conducted on four real-world datasets demonstrate that the combination of quantile loss-based MFFDM and weighted least squares (WLS)-based reconciliation technique achieves the highest performance ranks both for deterministic and probabilistic forecasting of temporal hierarchical wind speed, and the reconciliation process significantly improves the forecasting results.