LoRA-diag: enhancing low-rank adaptation via diagonal modulation for power system modeling code generation

Code generation for power system modeling encounters challenges due to complex engineering constraints and multi-level topological structures. While low-rank adaptation techniques perform well in code generation tasks, their fixed low-rank constraints inherently limit expressive power in handling complex power systems. This paper introduces LoRA-Diag, which improves model adaptability for PandaPower modeling by incorporating learnable diagonal modulation parameters into the low-rank decomposition's intermediate representation. Experimental results demonstrate that LoRA-Diag outperforms standard LoRA in code correctness, functional completeness, and execution success rate while maintaining high parameter efficiency.