LGTFormer: A Local–Global Transformer with Continuous State Transition for Short-Term Oilfield Electrical Load Forecasting

Electrical load reflects changes in electricity demand, and accurately forecasting trends in load variation is important for refined electricity management in oilfields. However, oilfield electrical loads exhibit more complex dynamics at finer temporal scales, placing greater demands on a model’s ability to capture temporal features. This study uses 18 months of actual operational data sampled at 5-min intervals from three distribution lines in an oilfield in China to investigate short-term load forecasting and proposes the Local–Global Transformer with Continuous State Transition (LGTFormer). Built upon a Transformer encoder, LGTFormer introduces a convolutional temporal embedding module to enhance the local temporal representation of input features and a continuous state transition module based on the closed-form continuous-time neural network (CfC) to dynamically refine high-level temporal states. Experimental results show that LGTFormer achieves the best forecasting performance across all three lines, outperforming the comparison models in terms of mean absolute error (MAE), root mean square error (RMSE) and coefficient of determination ($R^2$). Ablation experiments indicate that the convolutional temporal embedding module contributes primarily to improvements in forecasting accuracy, whilst the continuous state transition module enhances forecasting stability. Supplementary validation further confirms the forecasting accuracy advantage of LGTFormer after adjusting the test period. Model efficiency analysis indicates a reasonable balance between forecasting accuracy and training efficiency.