Wind power prediction using foundation large time series models enhanced by time series prompt in exogenous and tuning forms

Large Time Series Models (LTSMs) hold broad application prospects in the field of energy, where time series analysis plays an important role in various practical downstream tasks such as power forecasting. However, the neglect of exogenous variables and limitations of full fine-tuning have hindered their adaptation to downstream tasks. Proposing the concept of Time Series Prompt (TSP), this work develops a TSP-based scheme to integrate exogenous variables into foundation LTSMs together with Parameter-Efficient Fine-Tuning (PEFT) method, enabling more flexible and effective adaptation. Firstly, this work proposes a TSP construction method that embeds exogenous variables as prompts to guide model generation without altering the model’s backbone. Secondly, a prompt-based PEFT method is introduced, known as prompt tuning (PT). By augmenting the input with artificial prompts optimized for downstream tasks, PT allows for the training of approximately only 10 % of the model’s parameters while the model’s backbone remains frozen. The presented scheme significantly enhances the flexibility of foundation LTSMs in adapting to downstream tasks. The proposed methods are validated in wind power prediction by ablation study, using Timer as an example of foundation LTSMs and adopting accurate or noisy future wind speed as exogenous variables. The results demonstrate that introducing exogenous variables via prompts can reduce the prediction MSE by approximately 50 %, and subsequent PT can further reduce the MSE by up to an additional 50 %. The results confirm the effectiveness of the TSP in foundation LTSMs, providing a reference for efficient and flexible adaptation of foundation LTSMs to wind power prediction.