An incremental photovoltaic power prediction model considering online updating and catastrophic forgetting

Abstract

Accurate photovoltaic power generation forecasts provide valuable and reliable insights for power system scheduling. In real-world scenarios, forecasting models need to be frequently updated to mitigate performance degradation caused by evolving input data. However, frequent updates can lead to catastrophic forgetting of previously learned knowledge, thereby reducing the predictive accuracy of the updated model. To address this issue, this paper proposes an online-updated multivariate time series predicting model, the PTER model, which integrates PatchTST architecture with DER++ incremental learning. The model employs the patch token strategy to capture the multi-scale periodic characteristics of PV power sequences and captures multivariate dependencies through the self-attention mechanism. And it utilizes experience replay to mitigate catastrophic forgetting during online updates. Consequently, the PTER improves the accuracy of PV power generation prediction and enhances adaptability to abnormal weather conditions. The study focuses on a PV power generation unit at a power station in Xinjiang, simulating the model evolution process under real-time data updates through experimental design. Compared to PatchTST, Transformer, Informer, and Autoformer models, the PTER achieves a maximum reduction of 61.05 % in mean absolute error and a 57.29 % reduction in root mean square error, confirming its superior predictive accuracy. Furthermore, DER++ improves the RMSE by 13.71 %, 14.48 %, and 11.11 % compared to incremental learning EWC, LwF, and MAS, respectively. Under cloudy weather conditions, the PTER model exhibits the lowest mean absolute error and root mean square error among all evaluated models, indicating that it is more adaptable and generalizable to sudden changes in weather conditions.