Open Access

ARTICLE

DS-Kansformer: A Novel Distribution Adaptive Load Prediction Method for Air Conditioning Cooling

Cuihong Wen¹, Jingjing Wen¹, Qinyue Zhang¹, Yeting Wen², Fanyong Cheng^3,*

1 College of Information Science and Engineering, Hunan Normal University, Changsha, 410081, China
2 Research and Development Department, XiaMen INESIN Control Technology Company Limited, Xiamen, 361001, China
3 Key Laboratory of Advanced Perception and Intelligent Control of High-End Equipment, Anhui Polytechnic University, Wuhu, 241000, China

* Corresponding Author: Fanyong Cheng. Email:

Energy Engineering 2026, 123(3), 23 https://doi.org/10.32604/ee.2025.071911

Received 15 August 2025; Accepted 31 October 2025; Issue published 27 February 2026

Abstract

Air conditioning is a major energy-consuming component in buildings, and accurate air conditioning load forecasting is of great significance for maximizing energy utilization efficiency. However, the deep learning models currently used in the field of air conditioning load forecasting often suffer from issues such as distribution bias in load data and insufficient expression ability of nonlinear features in the model, which affect the accuracy of load forecasting. To address this, this paper proposes a novel load forecasting model. Firstly, the model employs the Dish-TS (DS) module to standardize the input window data through self-learning standardized parameters, thereby addressing the spatial intra-bias problem existing between data. Secondly, DS-Kansformer introduces Kolmogorov-Arnold Networks (KANs) to enhance the expression ability of nonlinear features. Finally, the output window is denormalized through the self-learning parameter of the DS module to restore the original distribution of the predicted data. In this paper, experiments were carried out based on the air-conditioning load dataset collected from a multi-functional comprehensive building, and the experimental results show that after adding the DS module, the Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and R-squared (R²) of the model are 20.46%, 34.44%, and 92.61%, respectively; after introducing KAN, the MAE, RMSE, and R² are 22.81%, 35.72%, and 92.05%, respectively; the model also exhibits high prediction accuracy after integrating the two modules (with RMSE, MAE, and R² being 19.75%, 34.05%, and 92.78%, respectively), outperforming common time series prediction models, confirming the reliability and efficiency of the model, which can provide reliable support for intelligent energy management in buildings.

---

Keywords

---