基于多尺度分解融合的并行双支路短期电力负荷预测研究

Research on Parallel Dual-Branch Short-Term Electric Load Forecasting Based on Multi-Scale Decomposition Fusion

摘要:

针对电力负荷数据的周期性与趋势性特征难捕捉,影响预测精度的问题,提出一种基于多尺度分解融合的MSDF-Informer-LSTM并行双支路负荷预测混合模型；该模型通过动态下采样机制对负荷数据进行多尺度时序解耦,采用季节趋势分解方法,分解不同时间尺度的季节项和趋势项,引入注意力机制驱动的跨尺度特征融合策略,实现全局趋势和局部波动特征的自适应融合；采用Informer与LSTM并行双支路预测,通过双层加权融合策略实现多尺度时序预测生成短期负荷预测值；实验结果显示,该模型较Informer在RMSE和MAE指标上分别降低21.2%和23.9%,较LSTM在RMSE和MAE指标上分别降低7.7%和11.5%,预测精度优于其他对比模型,验证了该模型可实现精准捕捉电力负荷数据的周期性与趋势性特征,有效提高模型预测精度。

Abstract:

Addressing the challenge of capturing the periodic and trending characteristics of electric load data, which affects prediction accuracy, a parallel dual-branch load forecasting hybrid model based on multi-scale decomposition fusion, MSDF-Informer-LSTM, is proposed. This model employs a dynamic downsampling mechanism to perform multi-scale temporal decoupling on load data. It utilizes a seasonal trend decomposition method to separate seasonal and trend components across different time scales. Additionally, it incorporates an attention mechanism-driven cross-scale feature fusion strategy to achieve adaptive fusion of global trends and local fluctuation features. By utilizing parallel dual-branch prediction with Informer and LSTM, and employing a two-layer weighted fusion strategy, multi-scale temporal predictions are generated to produce short-term load forecasting values. Experimental results show that this model outperforms Informer in terms of RMSE and MAE by 21.2% and 23.9%, respectively, and LSTM by 7.7% and 11.5%, respectively. Its prediction accuracy surpasses that of other comparative models, validating its ability to accurately capture the periodic and trending characteristics of electric load data and effectively enhance model prediction accuracy.