PhD CIFRE RTE - MINES PARIS : Artificial Intelligence for Renewable Energy Forecasting

Mines Paris - PSL, Centre PERSEE & RTE; Valbonne; Sur place; EUR 23 000 - 30 000; Plein temps

Résumé du poste

Mines Paris - PSL et RTE offrent un doctorat sur des méthodes avancées basées sur l'IA pour améliorer la prévisibilité de la production d'énergie renouvelable. Le projet se concentre sur la collecte et le traitement de grandes quantités de données pour concevoir des modèles de prévision plus précis. Une approche spatiotemporelle sera adoptée pour développer des solutions innovantes, adaptées aux contextes changeants des systèmes d'énergie renouvelable.

Prestations

Financement Cifre

Qualifications

• Forte expérience en Deep Learning et IA.
• Compétences en traitement de données massives et en prévisions.

Responsabilités

• Améliorer la prévisibilité de la production d'énergie renouvelable.
• Développer des modèles de prévision pour différents niveaux géographiques.
• Analyser les erreurs de prévision en fonction des données d'entrée et des conditions météorologiques.

Connaissances

Deep Learning

Méthodologies IA

Traitement de données massives

Prévisions météorologiques

Formation

Doctorat en informatique, data science ou domaine connexe

Description du poste

Title : "Advanced AI-based methods to exploit massive data for improved predictability of renewable energy generation. "

Context and background :

Short-term energy forecasting for the next minutes to days ahead, is a prerequisite for the economic and safe operation of modern power systems and electricity markets, especially under high penetration of renewable energy sources (RES). Given the diverse application contexts, end-users require models that have a wide range of properties especially when deployed in operational settings.

Although RES forecasting models have been used operationally for years, numerous challenges remain unresolved, continuing to drive research efforts worldwide. This PhD project focuses on RES forecasting across various geographical scales : local, regional, and national. Improving forecast accuracy is a more cost-effective strategy that permits to reduce investments in hedging solutions such as storage.

At time scales of a few minutes to a few hours ahead, there is a significant potential for improving accuracy by leveraging data from geographically distributed RES plants. These plants can act as sensors, capturing the propagation of weather phenomena in space and time. This information can then be used to produce more precise predictions for individual RES sites.

The growing amount of installed RES plants results in massive amounts of data that have to be efficiently processed by the forecasting models. Additionally, data may be missing or affected by non-weather-related factors. For example, wind turbines might be slowed down due to noise reduction regulations, or may be stopped to avoid bird collision or due to electricity grid constraints.

Forecasts must also maintain consistency across different geographical levels; from individual wind farms to regional and national scale (i.e. wind farm / region / national). At present, RTE, the French Transmission System Operator, uses RES forecasts for the short-term balancing of electricity supply and demand and to estimate the conditions at the electricity grid (power flow). Even a few percentages of error in RES production at the national scale are translated to very high financial costs (several millions euros per day) and may jeopardise the grid operator’s obligation to ensure security of supply. Therefore, improving RES forecasting accuracy is a top priority for grid operators / en / uninterrupted-flow-current / balancing-supply-and-demand). Modern IA-based methods offer a great potential towards that direction.

Scientific objectives :

The aim of this thesis is to improve predictability of RES production and net load, with a focus on challenging situations due to extreme weather conditions, missing or corrupted data and non-weather-related factors that affect RES production. The research will explore the application of AI-based methodologies, including foundational models, deep learning architectures, and more conventional AI approaches as reference, to tackle these challenges. A core objective is to assess the ability of such models to process large-scale datasets, integrate contextual information, and adapt to changing system configurations (i.e. evolution in the number of installed RES plants). Furthermore, the thesis aims to develop forecasting models capable of generating consistent prediction across different geographical scales.

Methodology and expected results :

Initially the state of the art will be analysed.Given the vast volume of existing literature—often characterized by incremental advancements—this task presents a significant challenge. To efficiently identify high-potential innovations, a smart and possibly AI-assisted approach will be developed to filter, classify, and extract the most impactful contributions. The integration of additional data sources, such as alternative Numerical Weather Predictions (NWPs), radar observations, and data from meteorological stations, will also be explored to enhance forecast performance.

Benchmarks will be established based on existing forecasts used by RTE, as well as state-of-the-art methods. The prediction errors will be analyzed in relation to input data, weather conditions, and other contextual variables, in order to identify situations more prone to improvements. Building on these insights, novel forecasting approaches will be proposed with a focus on the spatiotemporal approach and / or AI-based approaches. The aim is to develop forecasting solutions that are not only more accurate but also more resilient and context-aware.

• 10-01

Funding category

Cifre