Probabilistic prediction of photovoltaic power : a multi-task learning and large language model-based approach

Abstract

Accurate weekly probabilistic forecasting of photovoltaic power holds immense value for optimizing power generation schedules and market trading strategies. However, current research for photovoltaic power forecasting focuses on short-term prediction, and there is insufficient research on weekly probabilistic prediction, especially when data availability is limited. For this purpose, this paper proposes a weekly probabilistic photovoltaic power forecasting approach based on multi-task learning and a large language model (LLM). First, the wavelet transform is employed to decompose the photovoltaic power time series into smoother sub-frequency curves, which are predicted using a new LLM meta AI (LLaMA)-based LLM. The proposed LLM harnesses the shared feature correlations derived from multi-task learning, coupled with the robust generalization capabilities of the pre-trained LLaMA, to effectively capture intricate nonlinear characteristics of photovoltaic power under zero-shot and few-shot data. Then, to adapt to the photovoltaic power prediction task and improve the prediction accuracy, a dilated convolutional bidirectional long short-term memory-based adapter is introduced for fine-tuning the LLM. Finally, a new probabilistic forecasting approach that integrates the proposed LLM with direct probability forecasting methods is introduced to characterize uncertainties across different quantiles, and deterministic forecasting is achieved by setting the quantile to 0.5. The proposed deterministic and probabilistic forecasting performance has been validated using weekly data from two photovoltaic power stations in northwestern China, and experimental results have indicated that the proposed approach achieves an average improvement of 112.16% in the average interval sharpness metric compared with state-of-the-art benchmarks under zero-shot and few-shot data predictions.