This library is a collaborative effort between:

kisters_logo
 deltares_logo

rtc-tools-meta

This is the documentation for rtc-tools-meta (Meta), a distributed model predictive control (DMPC) library. Meta finds the interface conditions between independent rtc-tools models. This allows users of rtc-tools to model complex systems as a network of modular submodels that can run independently, and use Meta determine how the models interact.

Overview

Meta allows you to easily aggregate existing rtc-tools models into larger ones, and then do an optimization to find good boundary conditions between those models. Once the boundary conditions are optimized and the results are imposed on the submodels, Meta runs the submodels as usual. Here is a visual overview of the process:

_images/overview.svg

Strengths and Limitations of rtc-tools-meta

Before you get started, here are some important strengths and limitations of this library to be aware of.

Strengths

  1. By aggregating the model, interface variables are by definition equal. In contrast, an iterative scheme (where submodels take turns running) has no guarantee of finding a solution with consistent interface variables between the submodels, and less still of finding a reasonably optimal one. An aggregated model can take consistent interface variables for granted, and focus on finding their best possible values.
  2. The Meta api is recursive-ready. This means you can aggregate meta-models into meta-meta-models, etc. And you can have unique meta-objectives for each meta level. As long as the largest model is small enough to run, you can connect as large a network as you like.

Limitations

  1. There is not yet any model simplification. This means that the meta problem is the same size as the sum of the subproblems. If your models are already quite large, you may find you can’t use this library.
  2. The Meta optimization problem has a single meta-objective function. This means that the models must be linear, and the meta-objective must be at least convex, or convergence will be an issue. Keep in mind that interior-point methods like IPOPT can optimize massive problems if they are convex, and Linear Programming algorithms can handle even larger problems. This is a good thing with regard to Limitation #1. (Note that integration with HomotopyMixin is planned- meta models would run with the linear equations and submodels would resolve the non-linear hydraulics)
  3. Variables that interface between two models must be collocated at identical timesteps. In practice, this means that all models are run at the same timestep for the same duration, unless you get funky with the times() method in rtc-tools.

Contents: