Abstract:

To improve the accuracy of day-ahead photovoltaic (PV) power forecasting and mitigate the challenge of data scarcity in newly built photovoltaic power stations, leveraging both the reasoning capabilities of large language models (LLMs) and the correlation information across various dimensions of sequences, this paper proposes a meteorological covariate attention-enhanced Time-LLM model for day-ahead PV power prediction. Firstly, the input sequence is constructed by padding and concatenating historical PV power series with meteorological covariate series. Then, a novel covariate attention module is introduced to capture the cross-dimensional dependencies between meteorological variables and PV power sequences. Finally, the Time-LLM architecture is employed to achieve modality alignment between time-series and text sequences, effectively leveraging the textual analysis capabilities of LLMs for accurate PV power forecasting. Experimental results on public PV datasets demonstrate that the proposed model achieves superior forecasting performance and exhibits remarkable zero-shot learning capability. The proposed method not only improves the accuracy of day-ahead photovoltaic power forecasting, but also provides a promising solution for newly built PV plants with limited historical data, where traditional deep learning models often fail due to data scarcity.

Key words: photovoltaic power forecasting, large language model, Time-LLM, covariate attention, zero-shot