Short-Term Power Load Forecasting Based on Multi-Granularity Fusion Model

Individual customer load forecasting is crucial for power system operation and planning. However, due to the complexity of individual customer behavior, personal load profiles exhibit significant volatility and randomness. As a result, individual load forecasting poses substantial challenges. This paper proposes a short-term power load forecasting model based on multi-granularity fusion. To address the challenges of long output horizons, the proposed model employs different target strategies to mitigate the impact of uncertainty, instability, and nonlinearity in time series data on neural network-based forecasting models. In the proposed model, multi-scale pooling is utilized to extract the long-term trends of the raw load dataset, and a gated recurrent unit (GRU) is employed to construct the model based on the pooled dataset. Subsequently, a dual-path feature enhancement network is introduced to extract features from different sources, which are then used to build a Transformer-based forecasting model. We conducted comprehensive experiments on real-world datasets, and the proposed framework demonstrated more stable and accurate predictions compared to several classical load forecasting methods. Ablation studies reveal that the performance of the proposed model degrades when any component is removed or replaced. Overall, the proposed model exhibits superior performance and maintains robust predictive capabilities even in the presence of noise interference.