Nico Voigt
Back to projects

automation data - 2019

Optimizing Prosumer Battery Storage with Machine Learning

IEEE research project on ANN-based household load forecasting and linear optimization for battery control under dynamic electricity pricing.

machine-learningenergyoptimizationforecastingresearch

Key Result

Annual revenue per household battery — averaged across 38 representative households

Rule-based117 EUR / year
84%
ML-supported130 EUR / year
93%
Perfect foresight140 EUR / year
100%

The ML controller improved revenue in every one of the 38 evaluated households, with gains between 3% and 21% over the rule-based baseline.

Improvement over baseline

+12%

average across all households

Households evaluated

38

selected from 2,486 load profiles

Gap to perfect foresight

10 EUR/yr

remaining optimization potential

How it works

1

Forecast demand

An ANN predicts household load for the next 24 hours using historical patterns.

2

Optimize scheduling

A linear optimizer plans charge/discharge decisions using price signals and the forecast.

3

Execute with feedback

Hourly battery control adjusts for the real state of charge at execution time.

4

Dynamic pricing signal

Day-ahead spot prices guide when to store vs. export — the rule-based controller ignores this.

Research Question

How much better can residential battery storage be managed when the controller does not only react to the current moment, but also anticipates household load and dynamic prices?

The paper compares a classic rule-based battery controller with an ML-supported control stack and a perfect-foresight benchmark.

System Design

The ML approach forecasts household demand with an artificial neural network and feeds the forecast into a 24-hour linear optimization model.

That optimization proposes charge, discharge, grid import, and feed-in decisions which are then converted into hourly battery control using rule logic and the real state of charge.

Dataset And Model

The evaluation combines 2,486 cleaned ISSDA household load profiles, dynamic day-ahead pricing, Osnabrueck PV generation data from an 8 kW rooftop system, and a 5 kWh battery with 81% round-trip efficiency.

For evaluation, 38 representative households were selected across the cluster structure of the dataset and simulated over a full year of hourly decisions.

Outcome

Average annual revenue increased from 117 EUR/year with rule-based control to 130 EUR/year with the ML-supported controller, while the perfect-foresight benchmark reached 140 EUR/year.

Across the 38 evaluated households, the ML controller improved revenue in every case and delivered gains between 3% and 21%, with a 12% average uplift.