[Objective] Traditional time-series forecasting methods often struggle to simultaneously capture crossscale nonlinear fluctuations and long-range temporal dependencies

which leads to limited accuracy in short-term electricity price prediction for spot markets

especially when prices exhibit spikes

volatility clustering

and pronounced non-stationarity. [Methods] To address these challenges

this study proposes a short-term electricity price forecasting framework based on a fused global-residual Mamba model that combines series decomposition

dual-branch selective state-space encoding

and global residual learning to strengthen representation power and improve training stability. First

a moving average filter is applied to the normalized electricity price sequence to decouple it into a trend component and a residual component

separating relatively stable low-frequency movements from high-frequency stochastic variations. The decomposed sequences are concatenated along the temporal dimension and mapped through a high-dimensional embedding layer to obtain a richer latent representation capable of characterizing complex market dynamics. To better reflect the multi-factor formation mechanism of spot prices

the model also incorporates exogenous variables

such as regional load and weather-related information(e.g.

temperature and meteorological conditions). Building on these inputs

a parallel dual-branch Mamba encoder is designed to extract local-to-global dynamic features from complementary perspectives. The variable-correlation branch focuses on learning time-varying interdependencies between electricity prices and exogenous drivers

explicitly modeling cross-variable coupling and market co-movements. In parallel

the feature-interaction branch targets nonlinear transformations and interactions within the embedded feature space; by permuting tensor dimensions so that selective scanning operates along the embedding dimension rather than only along time

it uncovers abstract interaction patterns that conventional temporal scanning may overlook. To integrate heterogeneous information from both branches

their outputs are concatenated and passed to a global residual learning module

which performs additive fusion between the fused representations and the original embedded input. This global residual pathway provides a stable channel for information flow

alleviates gradient degradation in deeper state-space architectures

and enhances the model's ability to capture multi-scale patterns by preserving original signals while enriching them with learned cross-variable and cross-feature dynamics. For robust performance and reduced manual effort

Bayesian optimization is applied under a time-series cross-validation(TSCV) setting to tune key hyperparameters

while training further adopts learning-rate scheduling and early stopping to improve efficiency and stability. [Result]Experiments on real operational data from the Australian Energy Market Operator(AEMO) for 24-hour-ahead forecasting demonstrate clear performance gains: the proposed Residual Mamba reduces RMSE by 29.22% relative to the baseline Mamba model and by 35.33% relative to LSTM

confirming superior accuracy and robustness. [Conclusion]Ablation results further highlight the essential role of the series decomposition module

the importance of the variable-correlation branch in the dual-branch design

and the effectiveness of global residual connections in stabilizing training and improving feature expression for highly volatile spot-market price series.