Study on Short-Term Power Load Forecasting Based on T-MobileNet-L Model

Accurate short-term power load forecasting is essential for smart grid efficiency. This study addresses the limitations of traditional algorithms in handling big data and non-linear patterns. A historical dataset (106,176 points) is preprocessed using linear interpolation and interquartile range–based outlier detection, with weather variables integrated. Baseline support vector machine (SVM) and least squares support vector machine (LS-SVM) models yield mean absolute percentage errors (MAPEs) of 12.5% and 10.3%, respectively, but fail to capture load peaks. AdaBoost reduces the MAPE to 8.7% and the root mean square error (RMSE) to 18.5 kW through iterative sampling. The lightweight T-MobileNet-L neural network—employing Leaky ReLU and transfer learning—achieves a 34× parameter reduction and 7× faster convergence, with top performance (MAPE = 6.2%, RMSE = 15.1 kW). It effectively captures load periodicity and temperature correlations. T-MobileNet-L overcomes traditional big data bottlenecks, enhances smart grid efficiency, and demonstrates the advantages of machine learning for non-linear modeling and the integration of external factors.