PVMTF: End-to-end long-sequence time-series forecasting frameworks based on patch technique and information fusion coding for mid-term photovoltaic power forecasting

Abstract

Accurate photovoltaic power forecasting can alleviate the impact on grid stability. Most existing photovoltaic power prediction models rely on increasing model complexity or increasing the size of the look-back window to expand the amount of extracted information, but this often leads to catastrophic forgetting of learned information or the introduction of excessive redundant noise. In addition, some models predict by decomposing data and using non end-to-end learning, which may lead to inconsistent information and cumulative errors, limiting the improvement of prediction accuracy. To address the aforementioned challenges, we propose end-to-end PVMTF frameworks consisting of two models, PatchGRU and PatchGRU_h. This study is divided into two modules. In the data preprocessing module, we use Isolation Forest for outlier detection and replace outliers with window averages. Grey relational analysis is used for feature selection to reduce training complexity. In the photovoltaic power forecasting module, the PVMTF frameworks are used to directly achieve photovoltaic power forecasting. Firstly, based on the patch technique, the data is divided into independent short patches for separate learning, which can effectively preserve and learn historical information, avoiding catastrophic forgetting of important information that has already been learned as the look-back window grows. Specifically, for each patch, parameter sharing or independent parameter training Gated Recurrent Units (GRUs) are introduced to adapt to different computing needs, extract features within the patches, and achieve feature fusion. Next, a neural network-based gating mechanism is introduced to nonlinearly learn hidden states and fuse information. Finally, based on the above information fusion coding, accurate photovoltaic power forecasting is achieved by extracting the relationships between patches. Strict numerical verification indicates that PVMTF outperforms various state-of-the-art (SOTA) time series forecasting models in the three PV forecasting tasks (1-step, 384-step (4 days-ahead) and 672-step (7 days-ahead)), which provides an effective tool for PV power management and dispatch.