Sparse spatial–temporal attention forecasting network: A new model for time series forecasting

Time series forecasting has practical applications in many fields such as power load forecasting and traffic flow prediction. However, current time series forecasting methods focus more on temporal correlations, with less consideration given to the impact of spatial dimensions on time series predictions. Today, complex spatial–temporal coupling relationships significantly influence time series forecasting. Existing works mostly model time series predictions purely based on temporal correlations, without modeling spatial–temporal correlations, leading to insufficient prediction accuracy in time series forecasting. Inspired by the recent success of Transformers in the graph domain and their successful application in forecasting, we propose a new model for time series prediction called the Sparse Spatial–Temporal Attention Forecasting Network. This network performs adaptive spatial–temporal attention modeling on spatial–temporal graphs and incorporates a sparsity measure function, MLP, and average pooling into the traditional Transformer’s multi-head self-attention mechanism to better weight the attention scores. To improve prediction accuracy, the Sparse Spatial–Temporal Attention Forecasting Network constructs a weighted adjacency matrix based on the topological relationships between spatially adjacent nodes and information such as latitude and longitude. It then builds an adaptive graph using this weighted adjacency matrix to capture hidden spatial–temporal coupling relationships. Simultaneously, it employs a sequence decomposition algorithm based on average pooling to divide the data into two main blocks for separate prediction, thereby enhancing prediction accuracy. The Sparse Spatial–Temporal Attention Forecasting Network’s superior performance is demonstrated through experimental results on the Shanghai Pudong power load dataset and the public dataset PEMS-BAY.