GPT4PF：一种微调预训练LLM的光伏功率预测模型

PDF下载HTML

Title:

GPT4PF: fine-tuning pre-trained LLM for photovoltaic power prediction

作者:

王德文^1,2, 苗庆健¹, 李成浩¹, 孙瑗¹, 赵文清^1,3

1. 华北电力大学 计算机系, 河北 保定 071003;
2. 河北省能源电力知识计算重点实验室, 河北 保定 071003;
3. 复杂能源系统智能计算教育部工程研究中心, 河北 保定 071003

Author(s):

WANG Dewen^1,2, MIAO Qingjian¹, LI Chenghao¹, SUN Yuan¹, ZHAO Wenqing^1,3

1. Department of Computer, North China Electric Power University, Baoding 071003, China;
2. Hebei Key Laboratory of Knowledge Computing for Energy & Power, Baoding 071003, China;
3. Engineering Research Center of Intelligent Computing for Complex Energy Systems Ministry of Education, Baoding 071003, China

关键词:

预训练大语言模型; 光伏功率预测; 大语言模型; 时间编码; 多阶段微调; 时序切片; 输入编码; 线性探测

Keywords:

pre-trained large language model; photovoltaic forecast; large language model; time encoding; multi-stage fine-tuning; time-series segmentation; input encoding; linear probing

分类号:

TP183; TM615

DOI:

10.11992/tis.202507023

摘要:

准确的光伏功率预测对于保障电力系统安全、提高太阳能利用效率至关重要。本文提出一种基于微调预训练大语言模型的光伏功率预测模型。设计了一种高适应性且高效的三阶段微调策略，在提升预测精度的同时，训练参数量仅占总参数量的0.49%。构建了一种嵌入时间信息的输入编码层，以增强大语言模型对时间信息的建模能力，消融实验表明平均绝对误差下降约1%。提出了一种光伏数据切片标记方式，解决单时间点信息有限、难反映光伏数据整体特征规律的问题。实验结果表明，本模型相较DLiner、PatchTST、LLM4TS，平均绝对误差分别下降18.9%、4.1%、5.4%。

Abstract:

Accurate photovoltaic power forecasting is crucial for ensuring the security of power systems and improving the efficiency of solar energy utilization.This study proposed a photovoltaic power forecasting model named GPT4PF, which was developed by fine-tuning a pre-trained large language model (LLM). A highly adaptive and efficient three-stage fine-tuning strategy was proposed, which enhanced forecasting accuracy while limiting the number of training parameters to merely 0.49% of the total model parameters. An input encoding layer embedded with temporal information was constructed to strengthen the LLM’s capability of modeling temporal features. Ablation experiments showed that the mean absolute error (MAE) was reduced by approximately 1%. In addition, the issue that a single time point carries limited information and struggles to reflect the overall characteristics and patterns of photovoltaic data was addressed through time-series segmentation technology. Experimental results showed that compared with DLiner, PatchTST, and LLM4TS, the proposed model achieved a reduction in MAE by 18.9%, 4.1%, and 5.4% respectively.