The electricity market is a complex and dynamic environment characterized by a multitude of factors that influence electricity prices. Accurate and reliable electricity price forecasting (EPF) is crucial for market participants, including power generators, consumers, and policymakers. Electricity prices are influenced by temporal dependencies and electricity consumption patterns. Therefore, dependencies across different feature dimensions (cross-dimensional dependencies) and temporal trend information are essential. To address the aforementioned issues, we propose Multimodal Integration and Linear Enhanced Transformer (MILET), which combines cross-dimensional dependencies with single-dimensional modal features. First, we decompose electricity price data into three regular modals using Variational Mode Decomposition and Sample Entropy. This approach enables us to uncover the intrinsic patterns within the variable, thereby simplifying the complexity of the data series. Then integrate these three modals and the original dataset into a five-channel encoder (Modal Integration Encoder, MIE) with both single and multi-dimensional information. MIE is composed of Overall Two-Stage Attention (OTSA) and Long Short-Term Memory (LSTM), where OTSA handles cross-dimensional dependencies, and LSTM addresses long-term dependencies. Additionally, we capture trend information in electricity consumption features through linear layers and linearly integrate the data to obtain the forecasting results. Extensive experimental results on five electricity price datasets demonstrate the effectiveness of MILET compared to state-of-the-art techniques. Our code is available at
https://github.com/Lisen-Zhao/MILET/tree/master
....

MILET: multimodal integration and linear enhanced transformer for electricity price forecasting

10.1080/21642583.2024.2313862