TDCN: A novel temporal depthwise convolutional network for short-term load forecasting

• A novel temporal depthwise convolutional network is proposed for load forecasting. • Depthwise separable convolution reduces the model’s the computational complexity. • Inverted bottleneck structure captures hidden features and channel correlation information. • The nonlinear representations of the model are improved by the layer normalization and GELU. • The superiority of the proposed model is verified by two real-world load datasets. Accurate and efficient short-term load forecasting (STLF) is crucial for the reliable and economic operation of the electric grid. However, with the growing integration of renewable energy sources like wind power and photovoltaics, load data have become increasingly complex and nonlinear, making accurate forecasting a challenging task in modern power systems. While numerous STLF models have been developed, many of these models are complex hybrid structures that struggle with issues such as overfitting, stacking errors, high computational costs, and suboptimal generalization. To address these challenges, this paper proposes a novel temporal depthwise convolutional network model for STLF. First, a dilated causal convolution is employed to optimize the depthwise convolution, taking advantage of its ability to exponentially increase sampling points and expand the receptive field, thereby improving the capture of temporal information within low-dimensional channels. Next, pointwise convolution networks are utilized to adjust the channel dimension. The feature map is initially expanded to higher dimension and then projected back to a low-dimensional matrix, forming an improved depthwise separable convolution model with an inverted bottleneck structure. This design minimizes information loss and leakage during the transformation of compressed feature space, leading to enhanced prediction accuracy while reducing the number of training parameters. Finally, layer normalization and Gaussian error linear unit are incorporated to further improve the model’s convergence and nonlinear representation capabilities. To evaluate the effectiveness and generalization of the proposed model, experiments are conducted using two real-world datasets. Comparative experiments with other state-of-the-art methods are also performed. The results demonstrate that the proposed model outperforms existing approaches in terms of prediction accuracy, computational efficiency, and generalization.