基于EWT-TOPSIS融合和波动性感知注意力的短期电力负荷预测OA

Accurate power load forecasting serves as a critical technology for ensuring the safe and stable operation of power grids and optimizing energy dispatch.Addressing the nonlinearity,non-stationarity,and complex temporal variation characteri-stics of power load sequences,this study proposed a prediction model integrating empirical wavelet transform(EWT),tech-nique for order preference by similarity to ideal solution(TOPSIS)decision-making,and volatility-aware attention mecha-nism.The model firstly employed EWT to dynamically decompose load sequences and generate five intrinsic mode functions.It then designed a multi-scale volatility analyzer to extract time-domain statistical features.Subsequently,it constructed a volatility-aware attention mechanism to dynamically modulate sequence weights.Finally,it introduced a two-stage TOPSIS feature fusion layer to achieve optimal integration of multi-source heterogeneous information.Experimental results demonstrate that on the Panama dataset,the model achieves a mean absolute error(MAE)of 39.87 MW,root mean squared error(RMSE)of 57.25 MW,and mean absolute percentage error(MAPE)of 3.21％.Compared to baseline models such as Crossformer and PatchTST,MAE reduces by 10.73％～67.09％and RMSE reduces by 5.81％～61.25％.On the Australian dataset,the model obtains MAE of 191.11 MW,RMSE of 244.96 MW,and MAPE of 2.06％,with MAE reduction of 2.31％～54.57％.Ablation experiments validate the effectiveness of the frequency-time-statistical feature collaborative modeling and volatility-aware adaptive mechanism.