基于大语言模型的风光出力时序预测方法

Time-series Forecasting Method for Wind and Photovoltaic Power Output Based on Large Language Models

摘要:

随着高比例可再生能源并网，风光发电出力预测愈发重要。然而，现有基于统计和机器学习的风电预测方法存在复杂非线性时序特征挖掘难、领域知识融合不充分等问题。基于大语言模型的强大特征提取能力及固有的自然语言交互能力，提出了一种面向风光出力的时序预测方法，主要创新包括：一是通过领域监督微调使大语言模型适配风光出力时序预测任务，支持自然语言指令输入，简化了领域知识融合过程；二是提出时序推理机制，包括电力时序数据与自然语言模态的跨模态表征，以及基于中间变量引导的因果推理方法，增强数据特征挖掘与推理能力。实验验证了领域适配微调和时序推理模块的有效性，且所提方法在风力发电和光伏发电出力预测任务中均表现出较优性能。

Abstract:

With the increasing integration of high-penetration renewable energy, the forecasting of wind and photovoltaic power generation has become increasingly critical. However, the existing wind power prediction methods based on statistics and machine learning have problems such as difficulties in mining complex nonlinear time-series features and insufficient integration of domain knowledge. Currently, large language models (LLMs) have attracted much attention in various fields due to their strong feature extraction capabilities and inherent natural language interaction capabilities. Therefore, a time-series forecasting method based on LLMs for the task of wind and photovoltaic power output prediction is proposed. The main innovations include: Firstly, through domain-supervised fine-tuning, the LLM is adapted to wind and photovoltaic power prediction, which supports natural language instruction inputs and simplifies the integration of domain knowledge. Secondly, a temporal reasoning mechanism is proposed, including cross-modal representation of power time series data and natural language, as well as a causal reasoning method guided by intermediate variables, which enhances the ability of data feature mining and reasoning. The effectiveness of domain-adaptive fine-tuning and temporal reasoning module is verified through experiments, and the proposed method achieves superior performance in both wind and photovoltaic power generation prediction.