Robust multi-scale LSTM for reliable and intelligent photovoltaic power forecasting

Photovoltaic (PV) power generation, which playing a crucial role in promoting sustainable development and mitigating climate change, has become a key catalyst in the worldwide energy shift. However, its extensive integration into power grids presents issues for stability and efficient utilization, as PV output is highly dependent on meteorological conditions. The inherent variability and randomness of PV power time series often degrade forecasting performance. To address these challenges, a prediction approach based on a Robust Multi-scale Long Short-Term Memory Network (RMS-LSTM) is proposed in this study. We aim to construct a model which is able to capture both short-term and long-term trends. Specifically, we also employs correlation analysis and Robust Principal Component Analysis (RPCA) for feature selection and dimensionality reduction. On this basis, the refined training data inputs the multi-scale LSTM framework, aiming to derive the potential features. Then, we can employ the trained model to attain high-precision predictions of power output. The PV power data from Guoneng serves as a validation dataset to evaluate the performance of RMS-LSTM. It is shown that the Root Mean Square Error (RMSE) is 4.9482 and Mean Absolute Error (MAE) is 2.5246. Compared with benchmark methods, the proposed framework has advantages in prediction accuracy and robustness, verifying its effectiveness in PV power forecasting.