Spec2LLM: A Spectral-to-Language Reprogramming Framework for Power Transformer Temperature Forecasting

Spec2LLM is a novel framework for long-horizon time series electricity forecasting that leverages spectral reprogramming to adapt frozen large language models (LLMs) to multivariate time-series inputs. The architecture integrates three key components: (i) a localized Enhanced Patch Embedding module that extracts top-k frequency magnitudes, phases, and local slope features via Fast Fourier Transform (FFT); (ii) a Spectral Prompt Builder that encodes global frequency priors into a natural language prompt; and (iii) a lightweight LLM Reprogramming Layer that projects spectral patch embeddings into the LLM’s hidden space using a learnable mini-vocabulary attention mechanism. The combined prompt and token sequence is processed through a frozen LLM, followed by a compact regression head. Experiments on four benchmark datasets show that Spec2LLM improves MSE by 5% and MAE by 3.5% over Time-LLM, underscoring its ability to unlock stronger forecasting accuracy from frozen LLMs.