Planning LV Grids by Predicting Residual Loads of Households via Methods of Machine Learning

Objectives and context

The transformation of the energy system is progressing, particularly at the distribution grid level. The policy goals of reducing CO2 emissions indicate that the trend towards decentralized and renewable energy generation will continue in the future. Furthermore, the electrification of the heating and mobility sector will lead to the integration of additional decentralized consumers at the low voltage level. The prediction of those decentralized energy resources (DER) challenges distribution system operators due to unknown future subsidy and price components. Nevertheless, DER have a strong influence on the planning of electrical grids.

To determine the influence of DER on the distribution grid, several methods are currently evaluated. One method is the optimization of a household’s self-consumption combined with the optimal investment decision in DER. Those optimizations models are mostly solved via Mixed-Integer-Linear-Programming (MILP). However, the computing time of those MILP for longer time periods, multiple investment decision and high temporal resolution can be extensive. Therefore, this paper presents a Machine Learning (ML) based approach to approximate the conventional MILP based approaches.

Approach

In this paper, supervised learning using classifications or regressions are investigated. With the help of supervised learning, it is possible to learn extensive constraints as well as the objective function of an optimization model. Due to the variety of different methods of supervised learning, e.g. neural networks (NN), these methods are applied and compared with each other in this paper.

The initial optimization model was set up as a mixed integer linear model. During the investigations it was shown that a separate consideration of the model in the form of a mixed integer part and a linear part leads to a better prognosis. For the validation of the method, the extent of the changes in the input parameters of the optimization model was first limited to a change in the load curve of the household and the component prices of photovoltaic systems and battery storage systems.

Outcomes and Conclusions

ML can be used in different forms, depending on the type of problem to be optimized. Classification algorithms such as decision tree, random forest, k-nearest-neighbors or NN are used to forecast the mixed-integer part, which represents the expansion of DER. The forecast of the linear part, which represents the DER’s operating time of the household, is performed using Recurrent NN with an additional Long short-term memory.

Methods of ML are suitable for approximating both the expansion decision for DER and the residual load curve of a household. This allows to determine the influence of different future developments on the electrical characteristics of grid participants. As a result, grid planning can be performed in more detail and future flexibility in resource utilization in the existing grid can be determined