A Medium and Long-term Forecasting Method for Electric Load Curves Integrating Dynamic Variable Selection and Sparse Attention Mechanism

Accurate forecasting of electric load curves is critical for stable power system operation, efficient dispatch, and scientific planning, with medium- and long-term forecasting providing key support for grid expansion, energy allocation, and electricity market decisions. Existing methods struggle with insufficient use of load-related multi-variable data and inefficient extraction of long-term temporal features from load measurements, limiting forecasting accuracy and efficiency. To address these issues, this paper proposes a medium- and long-term electric load curve forecasting model integrating dynamic variable selection and a sparse attention mechanism, with a focus on load measurement correlation. The model takes static variables and temporal variables related to load changes as inputs. Static variables include industry type and location, and temporal variables cover historical load and meteorological data. It uses a gated feedforward network to adaptively weight variables based on their correlation with actual load measurements and filters redundant information. Meanwhile, it adopts a dual-layer encoding structure with sparse attention to prioritize key features from long load measurement sequences, reducing computational complexity while enhancing capture of long-term dependencies between historical load measurements and future loads. Experiments on three datasets show the model outperforms baselines including Informer, Autoformer and TFT. The proposed model reduces average relative error by 14.3%–55.1% and improves computational efficiency, providing reliable technical support for power system medium- and long-term planning by closely linking load measurements to forecasting performance.