Multidimensional Data-Driven Ultra-Short-Term PV Power Forecasting with SVMD-WT-TPA-SRU

Ultra short term photovoltaic power forecasting is of significant importance for energy management, scheduling-optimization of the power grid, and production management of power plants. To address the challenge of predicting photovoltaic power due to data noise caused by meteorological factors, this paper introduces a multidimensional data-driven ultra-short-term photovoltaic power forecasting model with Successive Variational Mode Decomposition (SVMD), Wavelet Transform (WT), Temporal Pattern Attention (TPA), and Simple Recurrent Unit (SRU), referred to as the SVMD-WT-TPA-SRU model. Firstly, the scroll time window values are determined utilizing the Autocorrelation Function (ACF) and Partial Autocorrelation Function (PACF). Subsequently, the Pearson correlation coefficient method is employed to identify and select the primary factors that significantly influence photovoltaic power generation. Secondly, the SVMD method is used for signal decomposition and denoising of all relevant factors, followed by secondary denoising using Wavelet Transform (WT). Finally, the Simple Recurrent Unit (SRU) is applied for training and prediction, with the introduction of the Temporal Pattern Attention (TPA) mechanism to optimize the SRU model by minimizing attention to irrelevant variables, thus significantly improving both speed and accuracy. Simulated tests and comprehensive ablation analyses were conducted using the Alice Springs DKASC photovoltaic data from Australia to verify the effectiveness of the proposed method. Comparisons were made with LSTM, TRANSFORMER, and SRU models. The experimental results demonstrate that the proposed method effectively overcomes the impact of measurement noise and accurately predicts photovoltaic power, achieving higher prediction accuracy.