LMCNN: A Lightweight Attention-Based Modern Convolutional Neural Network Design for Runoff Prediction

In the field of time series prediction, runoff prediction is a critical subfield. Accurate runoff forecasting is essential for optimizing water resource allocation and mitigating flood risks. Conventional methods often fail to capture the complex patterns inherent in runoff formation due to its intricate mechanisms. While convolutional neural networks (CNNs) have shown promise in runoff prediction due to their robust feature extraction capabilities, existing studies predominantly focus on increasing network complexity, with limited emphasis on optimizing convolution operations. This study proposes a lightweight attention-based modern CNN (LMCNN) that enhances the efficiency of convolutional layers. LMCNN introduces a novel convolutional strategy that employs distinct convolutional layers to capture temporal features within channels, intravariable interchannel features, and intervariable interchannel features, significantly improving feature extraction while reducing computational complexity compared to traditional CNNs. Specifically, the model integrates meteorological and runoff data using a channel-independent strategy, followed by depthwise convolution (DWConv) to extract temporal features within individual channels. Two layers of pointwise and SE-enhanced convolution (PSConv), incorporating pointwise convolution for cross-channel fusion and improved squeeze-and-excitation networks (improved SENets) for dynamic feature selection, are then applied to extract intravariable and intervariable channel features. The data are subsequently flattened and processed through a linear layer to produce predictions. The model’s performance is validated using data from three hydrological stations in the upper, middle, and lower reaches of the Yellow River Basin, China. A comparative analysis with other advanced prediction models validates the effectiveness of the proposed model.