Time series forecasting is pivotal for applications such as production planning and optimization, particularly in modern power plants, where precise load forecasting informs critical operational decisions. Traditional models often struggle with the intricacies of high-dimensional and complex datasets, which is especially challenging when dealing with the extended time horizons required for long-term forecasts. These models may fail to capture nuanced patterns and relationships within the data, leading to less accurate predictions and suboptimal decision making in critical applications such as power plant operations. This paper introduces Adaptformer, an innovative forecasting model specifically designed for power consumption data. Adaptformer integrates multimodal data through a deep decomposition architecture, which disaggregates sequences into periodic, temperature, and special event components. The model’s advanced encoder-decoder architecture processes historical data using series decomposition and auto-correlation mechanisms to refine predictions. Experimental results demonstrate that Adaptformer excels in predicting power consumption across various lengths (96, 192, 336, 720), consistently achieving lower MSE and MAE compared to state-of-the-art models. This superior accuracy and reliability underscore Adaptformer’s potential to significantly enhance energy management and decision-making processes.