Open Access

ARTICLE

Grid-Supplied Load Prediction under Extreme Weather Conditions Based on CNN-BiLSTM-Attention Model with Transfer Learning

Qingliang Wang¹, Chengkai Liu¹, Zhaohui Zhou¹, Ye Han¹, Luebin Fang², Moxuan Zhao³, Xiao Cao^3,*

1 Power Dispatch Control Center, State Grid Fuzhou Power Supply Company, State Grid Fujian Electric Power Co., Ltd., Fuzhou, 350013, China
2 Science and Technology Office, State Grid Fuzhou Power Supply Company, State Grid Fujian Electric Power Co., Ltd., Fuzhou, 350013, China
3 New Energy Research Center, China Electric Power Research Institute Co., Ltd., Nanjing, 530004, China

* Corresponding Author: Xiao Cao. Email:

(This article belongs to the Special Issue: AI-Driven Innovations in Sustainable Energy Systems: Advances in Optimization, Storage, and Conversion)

Energy Engineering 2025, 122(11), 4715-4732. https://doi.org/10.32604/ee.2025.068105

Received 21 May 2025; Accepted 11 July 2025; Issue published 27 October 2025

Abstract

Grid-supplied load is the traditional load minus new energy generation, so grid-supplied load forecasting is challenged by uncertainties associated with the total energy demand and the energy generated off-grid. In addition, with the expansion of the power system and the increase in the frequency of extreme weather events, the difficulty of grid-supplied load forecasting is further exacerbated. Traditional statistical methods struggle to capture the dynamic characteristics of grid-supplied load, especially under extreme weather conditions. This paper proposes a novel grid-supplied load prediction model based on Convolutional Neural Network-Bidirectional LSTM-Attention mechanism (CNN-BiLSTM-Attention). The model utilizes transfer learning by pre-training on regular weather data and fine-tuning on extreme weather samples, aiming to improve prediction accuracy and robustness. Experimental results demonstrate that the proposed model outperforms traditional statistical methods and existing machine learning models. Through comprehensive experimental validation, the attention mechanism demonstrates exceptional capability in identifying and weighting critical temporal features across different timescales, which significantly contributes to enhanced prediction performance and stability under diverse weather conditions. Moreover, the proposed approach consistently exhibits strong generalization capabilities across multiple test cases when applied to different regional power grids with distinct operational patterns and varying load characteristics, showcasing its practical adaptability to real-world scenarios. This study provides a practical solution for enhancing grid-supplied load forecasting capabilities in the face of increasingly complex and unpredictable weather patterns.

---

Keywords

---