Short-Term Power Load Forecasting Based on SDTW-IPAM and Informer
Short-term load forecasting faces escalating volatility and nonlinearity due to high renewable penetration. This study proposes a hybrid framework integrating Soft Dynamic Time Warping-Improved Partitioning Around Medoids (SDTW-IPAM) clustering with an Informer model. The SDTW distance metric captures local temporal deformations in load curves, while Gap statistics and K-means++ initialization optimize PAM clustering to adaptively determine cluster count and initial medoids. Load profiles are partitioned into double-peak, high-peak, and smooth patterns. Maximum Information Coefficient (MIC) selects differential features for each cluster, and dedicated Informer models are trained per pattern. Validation on real load data from Urumqi, Xinjiang, demonstrates that the combined model outperforms benchmark models across EMAE, ERMSE, and R², particularly for highly volatile load patterns. The method enhances forecasting accuracy and robustness, offering practical value for power system scheduling under renewable uncertainty. Limitations include exclusion of direct renewable generation, price signals, and storage states; future work will incorporate multi-variable inputs and extreme weather scenarios.