AI on the edge for a secure and autonomous distribution grid control with a high share of renewable energies. – AI4DG

The control methods developed in the first part of the project are based on predictive control approaches with a prediction step of the consumption/production profiles, and an optimization step.
The power predictions on a daily horizon and on a half-hourly basis use artificial intelligence methods with convolutional neural networks. These approaches have been implemented at LIG (UGA) and at UBI.
From these predictions, a first optimization strategy is based on a centralized approach as a reference case. This method developed by UASBI aims at controlling the equipment behind the meter in order to minimize the peak power demand at the distribution transformer (i.e. taking into account the aggregation of all the users' power profiles).
A second decentralized control method is proposed by the G2Elab (UGA) and is based on a two-step control with 1) a predictive phase: commitment at D-1 on a profile to be followed and 2) a corrective phase close to real time to take into account the prediction errors and stick to the commitment profile. The first phase is based on an alternating direction algorithm with successive optimization at the user and processor (coordinating agent) levels. There is an exchange of quantity (power) / price (Lagrangian) information between the coordinator and the users over the iterations until a convergence - i.e. balance between the users' objectives (minimum bill) and the coordinator's (power profile smoothed at the MV-LV transformer level). The real time phase consists of three steps - i) at each time step the users define their available flexibility and transmit this information to the coordinator, ii) the coordinator then chooses the power references for each user to correct the deviation of the prediction on the profile, iii) the users finally apply the references.
In parallel, a prospective study conducted at the G2Elab (UGA) analyses the adequacy and the trade-off between the users' and the coordinator's objectives, taking into account potential constraints on the distribution network (voltage and losses). This study takes the form of offline scenarios and optimizations by varying the models and weights of the different objectives.

From a simulation point of view for the control methods, the convergence of the implemented methods has been validated for about a hundred controlled users. In particular, the decentralized method was tested for different installed storage capacities and different prediction qualities. The results showed that increasing the number of users in the control perimeter allowed to reach smoother aggregate power profiles (i.e. coordinator's objective) while improving the correction of prediction errors (overrun effect).
In parallel to the simulations, the technical specifications for the integration of the methods in the platform of the industrial partner ATOS have been defined. In particular, the architecture described takes into account the messages/data to be exchanged between the actors of the system according to the chosen management strategy (centralized Vs decentralized).
From an experimental test point of view, a zone of the WWN distribution network has been identified in Germany with very high solar penetration (panels behind the meter). Four users have applied to install storage equipment that will be monitored as part of the experimental deployment of the project.

From a research/simulation point of view, the most important perspective is to identify the potential of artificial intelligence to replace/improve model-based and optimization-based control strategies. Particular attention will be paid to the explicability of methods and the compactness of controllers (e.g. rules, fuzzy logic, decision trees). In a more prospective way, the analysis of the trade-off between user and network objectives will allow to propose reward strategies for eventual system services provided to the distribution network (voltage control, power smoothing, etc).
From an integration point of view, the next (immediate) steps will be to implement the prediction and control strategies in the ATOS Codex Smart Edge platform. Before the first experimental tests are scheduled by the end of 2023 / early 2024.

R. Rigo-Mariani, V. Debusschere, “An ADMM-based Coordination Strategy for the Control of Distributed Storage at the Household Level – Impact of the End-User Settings”, ELECTRIMACS, Nancy, France, May 2022. hal.science/hal-03675572. This paper is also submitted to the MATCOM journal, Mathematics and Computers in Simulation.
• K. Hendle, K. Schulte, R. Rigo-Mariani, J. Haubrock, “Minimizing reverse power flow at the low voltage transformer by controlling solar battery storages using a linear optimization algorithm”, submitted to ISGT conference 2023.
• J. Coignard, R. Rigo-Mariani, V. Debusschere, “A Trade-off between minimizing individual costs and collective network constraints”, working paper
• V. Debusschere, «Use of artificial intelligence for the management of smart grids«, online webinar, March, 2023.

To guarantee the voltage quality at all electric grid levels, the safe and reliable operation and to avoid costly expansion of future electric
grids, volatile renewable energy sources and storage systems must be intelligently and informatively connected. The project AI4DG aims to research and develop a distributed AI on the edge platform for an autonomous and secure battery storage control system in low voltage grids with a high share of renewable energy sources using smart meter data. Due to the distributed AI on the edge approach, the system is more fail-safe and able to preprocess sensible smart meter data to ensure data protection regulations than a centralized approach. After project preparations, the partners analyze the AI requirements defined for the energy system and AI methodology for the project. In parallel, the project partners develop a detailed hardware low voltage grid simulation and a cognitive edge architecture for distributed AI to implement and validate the distributed AI system. After successful validation, the AI system will be evaluated in the field. The project results will be disseminated at scientific open access journals and conferences. The partner AtosWorldgrid will review the further development of the project results of the AI on the edge system to achieve market readiness. Due to the AIbattery storage control, costly grid expansions will be minimized in the electric grid of Stadtwerke Versmold