Advection-diffusion spatiotemporal recurrent network for regional wind speed prediction

Regional wind speed prediction is an important spatiotemporal prediction task for optimizing wind power utilization. Existing spatiotemporal recurrent networks update the information of the memory cell by relying on gated mechanisms that capture only pointwise information flow, thereby limiting their ability to capture spatial information transport within the memory cell. To address this limitation and improve the accuracy of regional wind speed prediction, we propose a novel advection-diffusion spatiotemporal recurrent network, termed ADNet, which integrates the advection-diffusion equation into a spatiotemporal recurrent network to guide the information updating process. Specifically, we introduce a new advection-diffusion LSTM (AD-LSTM) block with advection and diffusion modules to capture the spatial information transport within the memory cell. To effectively integrate the advection and diffusion modules into the information updating process of the memory cell, we have designed a new multi-scale channel attention (MSCA) unit. This unit leverages both global and local information across the channel dimension of the memory cell to generate attention coefficients that adaptively emphasize advection and diffusion. To evaluate ADNet's predictive performance, we conduct extensive experiments on three real-world wind speed datasets. The results demonstrate that ADNet significantly outperforms baseline methods. Our code is available at https://github.com/ShidongC/ADNet.