Abstract: To address the need for improving the accuracy and efficiency of electricity price forecasting in power markets, a regional electricity price prediction method based on S-Mamba2 model was proposed. This method aims to resolve high computational resource consumption and low efficiency in complex electricity price forecasting scenarios. Building upon Mamba-2, the proposed algorithm incorporates a bidirectional Mamba-2 module and a feedforward neural network encoding layer. These enhancements effectively capture spiky characteristics, seasonal patterns, inter-variate correlations (VC), and temporal dependencies (TD) in historical electricity price data. Experiments were conducted using real-world electricity price datasets from Australian Energy Market Operator (AEMO) and 2014 Global Energy Forecasting Competition (GEFCom2014). Results demonstrate that S-Mamba2 outperforms benchmark models including iTransformer and TimeDiffusion in prediction performance, achieving a maximum prediction accuracy of 97.88%. The proposed method provides robust technical support for enhancing market efficiency, reducing trading risks, and optimizing resource allocation in power markets.

Key words: Mamba model, state space model, deep learning, electricity price forecasting

摘要： 针对电力市场中电价预测精度和效率提升的需求，提出了一种基于S-Mamba2模型的区域电价预测方法，以解决复杂场景中进行电价预测时存在的计算资源消耗高、效率低等问题。所提方法在Mamba-2的基础上，引入双向Mamba-2模块和前馈神经网络编码层，有效捕捉了电价历史数据中的尖峰特性、季节性规律以及变量的内在互相关特性(Inter-Variate Correlations,VC)和电价的时序依赖特性(Temporal Dependency,TD)。在澳大利亚电力市场运营商(Australian Energy Market Operator, AEMO)提供的电价实测数据集和2014年全球电力能源预测竞赛(2014 Global Energy Forecasting Competition,GEFCom2014)数据集上进行了电价预测实验，实验结果表明，S-Mamba2相较于iTransformer、TimeDiffusion等模型，提升了预测性能，预测准确率最高达到97.88%，为电力市场的效率提升、交易风险降低以及资源配置优化提供了有力的技术支持。

关键词: Mamba模型, 状态空间模型, 深度学习, 电价预测