Short-term load forecasting based on transfer learning and TCN-BiGRU

Electricity load forecasting is of great significance to the stable operation of power systems. For load forecasting, traditional short-term forecasting techniques frequently employ linear regression models, which have low forecasting accuracy due to the models’ inability to incorporate complicated load changes. A temporal convolutional network-bidirectional gated recurrent unit (TCN-BiGRU) model based on transfer learning (TL) is proposed. Highly relevant information is moved into the experimental model using a transfer learning strategy; the data is clustered and analyzed using a K-medoids clustering algorithm; features at various TCN scales are extracted using a parallel convolution strategy; pertinent information is captured using temporal attention (TA); and the TCN training is further extracted using a BiGRU. The non-linear features of the output are further extracted using the dynamic multigroup particle swarm optimization (DMS-PSO) algorithm to optimize and tune the hyperparameters of the network training in order to find the best combination of hyperparameters. The experimental results show that the proposed TL-TCN-BiGRU algorithm reduces mean absolute error (MAE) by 38.6%, root mean square error (RMSE) by 40.7%, mean absolute percentage error (MAPE) by 30.4%, and R² by 5.3% relative to the gated recurrent unit (GRU).