Short-Term Power Load Forecasting Based on Fractional-order Dynamic Pooling Convolutional Neural Network

Due to the complex nonlinear coupling between meteorological variables and dynamic consumption behavior, and the multi-scale time dependence in load series, short-term power load forecasting faces inherent challenges. Traditional approaches often fail to address key issues, including the high redundancy of multivariate weather data, obscuring essential features. The inherent conflict of the existing CNN pooling layer in simultaneously capturing sudden load peaks during extreme events and smoothing baseline trends, as well as the limited generalization caused by static hyperparameter configuration. To overcome these limitations, a hybrid SVD-FPCNN-LSTM model is proposed. A new fractional-order pooling CNN adaptively balances extreme and progressive load changes by dynamically fusing maximum and average pooling operations with learnable fractional-order parameters. The model combines long term and short term memory networks to model long term dependencies, and further uses genetic algorithm to jointly optimize fractional order and network hyperparameters, effectively reducing the risk of overfitting. The root mean square error (RMSE) and mean absolute percentage error (MAPE) are lower than other comparison models, which indicates that this model has high efficiency and accuracy in short-term power load forecasting.