Large language model (LLM) has demonstrated substantial application potential in power systems. However

conventional text-based data input struggles to accurately capture the inherent structural characteristics and numerical precision of power system data

thereby limiting their effectiveness in power-related tasks. To address this issue

this paper proposes an adaptation method for pre-trained LLMs tailored to power system tasks

integrating multi-source heterogeneous data representation learning with supervised fine-tuning. First

power system data are categorized into four typical forms

snapshot measurements

time series

graph structures

and textual descriptions

and the characteristic attributes of each type are analyzed. Then

a unified representation framework for multi-source heterogeneous data is constructed

in which feature extraction and semantic alignment techniques are employed to achieve efficient and lossless mapping of heterogeneous data into the semantic space of the LLM. Finally

experimental results from load forecasting and distribution network state estimation demonstrate that effectively leveraging pre-trained LLMs significantly improves task accuracy. In the load forecasting task

the proposed method reduces prediction error by 8.0% on average compared to traditional methods. In the distribution network state estimation task

the proposed method with only 8% measurement ratio achieves the same estimation accuracy as baseline methods with 16% measurement ratio

while the training process can be completed on consumer-grade graphics processing units.