AutoMoG 3D: Automated Data-Driven Model Generation of Multi-Energy Systems Using Hinging Hyperplanes

Abstract

The optimal operation of multi-energy systems requires optimization models that are accurate and computationally efficient. In practice, models are mostly generated manually. However, manual model generation is time-consuming, and model quality depends on the expertise of the modeler. Thus, reliable and automated model generation is highly desirable. Automated data-driven model generation seems promising due to the increasing availability of measurement data from cheap sensors and data storage. Here, we propose the method AutoMoG 3D (Automated Model Generation) to decrease the effort for data-driven generation of computationally efficient models while retaining high model quality. AutoMoG 3D automatically yields Mixed-Integer Linear Programming models of multi-energy systems enabling efficient operational optimization to global optimality using established solvers. For each component, AutoMoG 3D performs a piecewise-affine regression using hinging-hyperplane trees. Thereby, components can be modeled with an arbitrary number of independent variables. AutoMoG 3D iteratively increases the number of affine regions. Thereby, AutoMoG 3D balances the errors caused by each component in the overall model of the multi-energy system. AutoMoG 3D is applied to model a real-world pump system. Here, AutoMoG 3D drastically decreases the effort for data-driven model generation and provides an accurate and computationally efficient optimization model.

Boosting operational optimization of multi-energy systems by artificial neural nets

“…For the neural nets, we used TensorFlow. For the operational optimization problems, we used the open source toolbox AutoMoG (Kämper et al, 2021b). In the AutoMoG toolbox, you can find the implementation of every equation stated in Section 2.…”

Dynamic Ramping Constraints for Demand Response of Flat MIMO Processes

“…As, in the general case, the limit functions ρ (γ) min , ρ (γ) max are multivariate functions, multivariate regression methods, e.g., hinging hyperplanes (Breiman, 1993;Adeniran and Ferik, 2017;Kämper et al, 2021), convex region surrogates Schweidtmann et al, 2021), or artificial neural networks with ReLU activation functions (Grimstad and Andersson, 2019;Lueg et al, 2021), can be used to find piece-wise affine approximations.…”