Short-Term Load Forecasting Method Based on Prophet-CNN-Bigru

Given the strong randomness and complex influencing factors of power load data, as well as the insufficient prediction accuracy of traditional single models, this paper proposes a Prophet-CNN-BiGRU hybrid forecasting model that integrates the advantages of time series decomposition, spatial feature extraction, and bidirectional sequence modeling. In this approach, the Prophet layer is first employed to deeply process the original historical load and meteorological data, effectively capturing and decomposing key temporal patterns such as trend, seasonality, and holiday effects, thereby generating a feature matrix with significantly enhanced information. Subsequently, this enhanced feature matrix is fed into the CNN module to extract local spatial features and high-level abstract patterns. Furthermore, the BiGRU network is utilized to deeply mine the complex bidirectional long-term dependencies among features. Ultimately, the model performs end-to-end load fore-casting. Experimental results demonstrate that the proposed Prophet-CNN-BiGRU model significantly outperforms the comparative models in forecasting performance, achieving a mean absolute percentage error (MAPE), root mean square error (RMSE), and coefficient of determination (R²) of 0.167%, 0.057%, and 0.993, respectively. These results fully verify that the model possesses higher accuracy, greater stability, and significant advantages in short-term power load forecasting.