Revolutionizing solar-hydro-wind power forecasts in regional power systems with hybrid machine learning models

Abstract

As renewable energy’s share in the global energy mix increases, accurate forecasting of power output is crucial for grid stability and effective energy management. In multi-step-ahead forecasts, non-stationary signals and noise often lead to error accumulation and propagation. To address this, we propose a hybrid model integrating Variational Mode Decomposition (VMD) with Long Short-Term Memory (LSTM) neural networks (VMD-LSTM). This model leverages VMD’s signal decomposition strengths and LSTM’s temporal forecasting capabilities, improving input stationarity and noise resistance. Focusing on a regional grid in Hunan province, China, we used hourly power output data from hydro, wind, and solar stations (2017–2022), divided into training and testing sets. The VMD-LSTM model, trained on decomposed power output data, generated forecasts up to 24 h ahead. Compared to the LSTM, VMD-LSTM demonstrated superior performance, with improvements in Nash-Sutcliffe efficiency coefficient (NSE) by 6.7%, 9.1%, and 8.2%, and reductions in root mean squared error (RMSE) by 9.2%, 4.7%, and 16.8% for hydropower, wind, and solar forecasts at the 24-hour horizon. This study provides a valuable tool for enhancing renewable energy integration, contributing to the stability and efficiency of power grid management systems in regions with diverse energy sources.