Forecasting hourly day-ahead solar photovoltaic power generation by assembling a new adaptive multivariate data analysis with a long short-term memory network

Accurate multi-step PV power forecasting is a challenging task because of complex time series and error buildup in muti-step forecasts. This work is based on developing a decomposition-based hybrid model for hourly day-ahead PV power forecasting. Noise-assisted multivariate empirical mode decom-position (NA-MEMD) is an updated version of a popular MEMD technique that breaks the original series into several subseries termed as intrinsic mode functions (IMFs). Due to IMFs' statistically independent behavior, a separate long short-term memory (LSTM) network is designed for each set of multivariate IMFs. A single LSTM model is trained to simultaneously predict the IMF of the target variable for the sunshine hours of a day. Incorporating NA-MEMD with LSTM showed an average % RMSE (% MAE) reduction of 47.65 (50.99) and 11.11 (15.76) compared to the standalone LSTM and MEMD-based LSTM model, respectively, across three locations. The proposed model reported an average RMSE of 65.08 W/m2 across three locations. This investigation indicates that decomposition-based hybrid models can potentially reduce the error of multi-step forecasts. This study also recommends using NA-MEMD instead of MEMD because it improves forecast accuracy and requires less time for data demarcation (MEMD: 251sec, NA-MEMD: 195sec). & COPY; 2023 Elsevier Ltd. All rights reserved.