A novel gated dual convolutional neural network model with autoregressive method and attention mechanism for probabilistic load forecasting

Accurate load forecasting is prime in the electric power industry, while the complexity and variability of the load data make it a challenging problem. Therefore, the probabilistic load forecasting is used to provide a prediction interval for the power system. However, the input scale insensitive problem has rarely been addressed in existing works when exploring the multi-load zones data, resulting in a lack of robustness. In addition, the coverage and width of the prediction interval have seldom been simultaneously considered, affecting the forecasting accuracy. Therefore, to overcome the above drawbacks and achieve reliable forecasting results, a novel gated dual convolutional neural network model with autoregressive method and attention mechanism is proposed. Firstly, the autoregressive method is adopted as a linear component of the probabilistic load forecasting model to handle the input scale insensitive problem and enhance the modeling robustness. Then, a dual convolutional neural network is extended with the gating mechanism and is employed to capture the vertical correlation of information among multi-load zones and enhance the coverage of the prediction interval. Next, the attention mechanism is extended with pooling operation to extract the significant horizontal correlation of information among multi-load zones and obtain a narrower prediction interval. Finally, the superiority of the proposed model was validated through both ablation and comparison experiments on two real-world datasets. Graphical abstract