Dispatching Distribution Grids Using Data Analysis and Machine Learning Methods

Contact: Eleni Stai


Electrical grids are characterized by an increasing penetration of uncertain renewable energy sources. However, in electrical grids, the demand should be always in balance with the supply. Therefore, due to the uncertainty of the renewable energy generation, the need of an increasing amount of reserves emerges. However, the reserves are costly and lead to increased operational and investment costs for the power grid. To reduce the amount of reserves, energy storage, such as batteries, can absorb the uncertainty of the renewable energy sources. The batteries can e.g., charge to consume energy when there is over-generation of energy by the renewable energy sources and discharge to provide extra energy when the demand is more than the supply. In order to exploit the full potential of the batteries, it is necessary to plan their control. To do so we can apply scenario-based optimization techniques, where the scenarios represent possible realizations of the uncertain quantities obtained by historical data. In addition, by appropriately controlling the batteries, we can compute optimal consumption/generation plans for aggregated sets of prosumers, i.e., we make sure that a given cluster of prosumers follows a predetermined prosumption (namely, demand plus distributed generation).

Computing a dispatch plan for a distribution grid with stochastic resources and storage devices while accounting for operational constraints and system losses involves an AC Optimal Power Flow (AC OPF). This problem involves the non-linear power flow equations, thus, it is, as well known, nonconvex and hard to solve. There are several approaches in literature for approximating the solution of the AC OPF. In this project, we use CoDistFlow [1].

In this project: We will compute optimal dispatch plans as well as batteries trajectories by solving scenario-based optimization problems. We will use scenarios for the uncertain quantities that are obtained from historical data. However, increasing the number of scenarios, increases the run time complexity of the solution. To solve these run time complexity issues, we will combine convex optimization and data analysis/machine learning techniques (e.g., cross-validation [2]) to obtain efficient dispatch plans with affordable run time complexity.

Project Goals:

  • Obtain dispatch plans by combining convex optimization and data analysis/machine learning techniques
  • Comparative results via simulations and numerical evaluations in large-scale settings, e.g., large number of scenarios for the scenario-based optimization and large distribution grids

Student profile


  • Knowledge of Load Flow and Optimal Power Flow
  • Basics in statistics


  1. http://ieeexplore.ieee.org/document/7948761/
  2. https://en.wikipedia.org/wiki/Cross-validation_(statistics)