Day ahead solar forecast using long short term memory network augmented with Fast Fourier transform-assisted decomposition technique

Abstract

This work aims to develop a hybrid model for multistep PV power forecasting. The model comprises of decomposition (Noise Assisted Multivariate Empirical Mode Decomposition: NA-MEMD), dimensionality reduction (Fast Fourier Transform: FFT), and advanced deep learning (Attention mechanism-based Long short-term memory: AM-LSTM) methods. NA-MEMD addresses the non-stationary and nonlinear characteristics of complex multivariate time series data by splitting them into a number of subseries known as Intrinsic Mode Functions (IMFs). A large pool of IMFs is reduced to five sets of subseries using the Fast Fourier Transform (FFT). Finally, the model incorporates the advanced AM-LSTM technique, where the attention mechanism focuses on essential features while disregarding the irrelevant ones. The proposed N-FFT-AM-LSTM model demonstrates superior performance across multiple locations, with an average RMSE (W/m²) | nRMSE (%) | R-value of 62.97 | 6.33 | 0.9721. The proposed model surpasses both the AM-LSTM and N-AM-LSTM models, showcasing % mean RMSE (nRMSE) reduction of 36.86 % (35.25 %) and 12.98 % (11.56 %), respectively. These findings highlight the effectiveness of our approach, that is the N-FFT-AM-LSTM model, in accurately predicting solar irradiance levels across varied geographical regions.