Short-term forecasting of energy production and consumption using extreme learning machine: A comprehensive MIMO based ELM approach

A multiple-input multiple-output ( MIMO ) extreme learning machine ( ELM ) is introduced for short-term forecasting of seven grid variables in Corsica (France): total demand and generation from solar, wind, hydropower, thermal, bioenergy, and imports. Based on six years of hourly data, the model integrates sliding windows and cyclic time encodings to handle non-stationarity and seasonal effects without heavy preprocessing. At a 1-hour horizon, solar and thermal achieve nRMSE of 0.179 and 0.051 with R 2 > 0.98 , while total demand forecasts remain reliable up to 5 h ahead. Wind and bioenergy remain challenging due to high intrinsic variability, but overall accuracy is robust across sources. Compared with persistence and an LSTM configured under realistic tuning budgets, MIMO − ELM consistently improves skill, offering small but stable gains over Single-Input Single-Output models ( SISO ). Beyond accuracy, the closed-form solution ensures fast training and suitability for real-time updates, enabling potential use in online learning contexts. A key advantage of the MIMO formulation is internal coherence between aggregate demand and its components, an important requirement for operators. The methodology adapts to local constraints such as grid characteristics, resource availability, and market structures, ensuring transferability beyond the Corsican case. The study shows that a parsimonious approach such as MIMO − ELM can deliver forecasts that are accurate, coherent, and computationally efficient, providing a practical decision-support tool for energy management and renewable integration. • A novel Extreme Learning Machine-based Multi-Input Multi-Output ( MIMO ) framework for short-term energy forecasting is proposed. • The model accurately predicts energy outputs from multiple renewable and non-renewable sources, achieving high accuracy up to five hours ahead. • Performance is evaluated using normalized RMSE , MAE , and R 2 , demonstrating significant improvements over the persistence SISO model and deep learning-based LSTM approaches. • Accurate forecasting can be a powerful decision-support tool, optimizing dispatch and aiding renewable integration within operational constraints.