Original title: Comparative Evaluation of MILP, MPC, and Reinforcement Learning for Commercial Battery Dispatch Under Time-of-Use Tariffs
Hafiz Majid Hussain, Wajiha Samar, Lurian Klein, Pedro Nardelli · September 2026
Businesses with rooftop solar panels and batteries want to lower their power bills. Under special pricing plans, electricity costs more at certain peak hours. Batteries help by storing solar energy and discharging it when power is expensive. But controlling when to charge and discharge is tricky. Researchers tested three control methods using a commercial site's real data from 2023. One method used perfect future knowledge to find the best possible savings. A simpler method predicted tomorrow's needs by just assuming tomorrow looks like yesterday. The third method was an artificial intelligence agent that learns by trial and error. The ideal benchmark cut energy costs by 24.6% compared to using no battery. The simple method was nearly as good, capturing 99.2% of those ideal savings using only yesterday's data. Surprisingly, the learning agent performed terribly, costing more than having no battery at all.
You have a water tank that you fill when water is cheap and use when it's expensive. You don't know future prices, so you need a strategy. A simple strategy of using the same schedule as yesterday works almost as well as knowing the future. But a complex strategy that tries to learn from past mistakes can actually cost you more than having no tank. In this paper, the water tank is a battery, the water is electricity, and the price changes are the time-of-day electricity rates.
Commercial buildings with rooftop solar and batteries can cut electricity bills by timing their energy use. This study shows that a simple forecasting method achieves nearly all possible savings, while a complex machine-learning approach can actually increase costs. For businesses investing in battery storage, the control strategy is as important as the hardware.
The results are based on a single commercial site and one time-of-use tariff. The RL agent used a specific algorithm (SAC) and may not represent all RL approaches. The persistence forecast may perform poorly in variable weather. The MILP oracle assumes perfect future knowledge, which is not achievable in practice.
Ask anything about this paper. Answers come from its abstract, and say so when the abstract doesn't cover your question.
Comparative Evaluation of MILP, MPC, and Reinforcement Learning for Commercial Battery Dispatch Under Time-of-Use Tariffs — Hafiz Majid Hussain, Wajiha Samar, Lurian Klein et al.
Abstract licence: CC0
Explanation generated by Synthesis from the abstract — it is not the original paper. This is a preprint, so it hasn't necessarily been peer reviewed.