Volatility-aware sample re-weighting framework for short-term photovoltaic power forecasting

• To the best of our knowledge, this work is the first to quantify weather-type imbalance in PV power datasets based on the intrinsic volatility of PV power, rather than relying on external parameters. • We observe that samples with high PV power volatility account for most of the training loss, which significantly reduces forecasting performance. • We design a novel volatility-aware re-weighting framework (ReMAV) that adjusts the importance of training samples based on their volatility levels, thereby improving model accuracy under imbalanced PV power datasets. • We validate the proposed framework on three benchmark datasets and demonstrate that our proposed ReMAV framework effectively handles weather-type imbalance in PV power datasets and consistently outperforms existing baseline models in forecasting accuracy. Recent short-term photovoltaic (PV) power forecasting methods have primarily focused on improving model architectures to enhance forecasting accuracy, they often overlook the issue of weather-type imbalance in PV power datasets. To this end, we first introduce a new metric, Mean Accumulated Volatility (MAV) , which quantifies the volatility of each sample. By translating unquantified weather-type imbalance into a measurable form of volatility imbalance, we observe that high-MAV samples account for most of the training loss, thereby harming the model’s forecasting accuracy. Then, we further propose ReMAV , a volatility-aware Re -weighting framework that down-weights the losses of high-MAV samples and up-weights those of low-MAV samples based on the MAV -based density. Extensive experiments on eleven baseline forecasting models across three real-world PV power datasets demonstrate that our proposed ReMAV framework effectively handles PV power with weather-type imbalance and consistently outperforms existing baseline models in forecasting accuracy. For example, on the Alice Springs dataset, ReMAV reduces average MAE by 8.53% over baselines, while on the PVOD dataset, MAE drops by 5.46% on average.