From solar variability to useful forecasts
My research at Mines Paris–PSL / O.I.E. examines what can be predicted, which information a model needs, and how to evaluate a forecast for its intended use. The work connects statistical physics, time series, machine learning and energy systems, with a continuing strand in medical physics.
A strong daily cycle can conceal the variability caused by clouds. The stochastic coefficient of variation (sCV) separates these effects; the forecastability index F studies how this variability relates to prediction error. The published evaluation includes 68 Spanish stations.
Read the sCV and forecastability story · Renewable Energy paper
I study how signal structure and model complexity affect forecast quality. This includes autoregressive models, Extreme Learning Machines, clear-sky preprocessing and cyclostationary persistence. BLEND uses periodically correlated structure to improve a persistence reference without training a forecasting model; the separate MIMO-ELM study forecasts seven energy signals jointly.
BLEND explained · MIMO-ELM results and limits · ELM implementation guide
NICE metrics measure errors relative to a persistence baseline across several error norms. The γ-index evaluates agreement within explicit spatial, temporal and intensity tolerances. Both approaches complement conventional metrics; their interpretation depends on the baseline, sampling and tolerances.
NICE explained · The γ-index explained · Methods and public code
Forecasts support the operation of photovoltaic systems, storage and microgrids. My project work connects prediction to energy management and control, particularly in island systems. In medical physics, my interests include radiobiological dose equivalence, radiotherapy modelling and radioiodine kinetics.
LQL-Equiv and radiobiological dose equivalence · Teaching these methods
sCV and forecastability F: definitions, figures and MATLAB code · NICE metrics and the open implementation.
How much spatial information helps satellite irradiance forecasting? Per-cell, patch and whole-field comparisons; manuscript in preparation, with a public MATLAB/Python benchmark and SoDa/HelioClim-3 data attribution.
I contribute to collaborative programmes in forecasting, renewable integration and energy management. The portfolio below separates current programmes from earlier collaborations. Official project records are linked where available.
Fine4Cast · ANR / PEPR TASE · 2023–2028
Forecasting renewable production and consumption from minutes to days at fine geographic resolution. Programme overview.
ELECTRE · Mines Carnot · since 2025
Scientific coordination of the components addressing AI for energy, forecasting and multi-source systems.
Mines Paris–PSL / University of Corsica / CNRS agreement · 2024–2027
Intermittent solar forecasting and energy-management algorithms for smart microgrids.
SAPHIR · ANR · project started January 2022
Sensor-augmented weather prediction at high resolution, combining observations and numerical modelling. ANR project record (initial duration: 48 months).
HyLES · ANR · 2021–2024
Hydrogen integration in weakly interconnected networks, with FEMTO-ST, LE2P, UPF and UCPP-SPE.
INTELPOLY · COST Action · 2021–2025
Diagnosis, control and cybersecurity of polygeneration and hybrid-storage installations.
TILOS · Horizon 2020
Local integration of battery storage and renewable generation in an island demonstrator. The funded project ran from February 2015 to January 2019. CORDIS project record.
MYRTE — short-term solar forecasting for hydrogen storage and coupled production–load management.
RESINTER / PAGLIA-ORBA — predictive control of a renewable electricity network.
ISERBATURB — integration of renewable systems into buildings and urban environments.
Franco-Polish cooperation — exchanges on renewable integration in island microgrids.
Technical University of Sofia / Erasmus — co-construction of a renewable-energy summer school.
Academic CV · Full publication record · cyril.voyant@minesparis.psl.eu