Bi-Mamba-Load: A bidirectional structured state space model with sequence relationship adaptive perception for electricity load forecasting

The accurate prediction of electricity load is a critical component for smart-grid management, energy allocation, and grid stability. Traditional statistical methods and Convolutional Neural Networks (CNNs) often struggle with nonlinear relationships or long-range temporal dependencies, whereas Transformer architectures incur quadratic complexity in the sequence length. The Structured State Space Model Mamba provides linear sequence modeling but uses a unidirectional scan and a fixed treatment of multivariate dependencies. To address these limitations, we propose Bi-Mamba-Load, which combines a bidirectional Mamba encoder, a custom forget gate, and a Sequence Relationship Adaptive Perception (SRAP) module. Both encoder branches operate exclusively on the observed historical window, so no observations from the forecasting horizon are used as inputs. SRAP computes Spearman correlation statistics from the training split and performs dataset-level selection between channel-independent and channel-mixing tokenization; the selected route remains fixed for all samples in that dataset. Experiments across eight benchmark datasets indicate competitive or superior forecasting accuracy against the evaluated baselines while retaining linear scaling with the input length, subject to the constant-factor overhead of bidirectional scanning.