Abstract: To address the problem that the prediction accuracy of photovoltaic power is insufficient due to the lack of historical data of newly-built photovoltaic power stations, an instance transfer learning-based short-term prediction method is proposed for few-shot photovoltaic power generation. Firstly, a set of rich long-term operation photovoltaic data is used as the source domain, and then the multi-kernel maximum mean discrepancy was employed to estimate the matching similarity of photovoltaic data between source domain and target domain, and the migration source domain with high similarity was screened out. Then, a weighted adversarial bi-directional long-short time memory network was established. The photovoltaic samples in the source domain were weighted to adjust their data distribution by adversarial learning, and the adjusted source domain data was enriched to the target domain dataset. The bi-directional long-short time memory network was used to mine the bi-directional time sequence correlation of photovoltaic power sequence and meteorological data in the public knowledge domain, so as to achieve accurate prediction of few-shot photovoltaic power. The results show that the proposed method can effectively improve the prediction accuracy of photovoltaic power under the limited historical data compared with the traditional deep learning and model transfer methods.