Averaging favors MPC: How typical evaluation setups overstate MPC performance for residential battery scheduling

TL;DR

This work investigates how time-averaging in evaluation setups biases the comparison between Model Predictive Control (MPC) and Rule-Based Control (RBC) for residential PV–battery scheduling under Net Billing. By simulating 15 buildings over five months with ideal and realistic forecasts across scheduling steps of 60, 30, and 15 minutes, the study shows that coarse averages understate total costs and dramatically inflate MPC’s apparent advantages, with shrinkage up to $0.69$ (69%) in some cases. It demonstrates that, when intrainterval fluctuations are accounted for at minute-level ground truth, MPC’s edge over RBC can vanish or reverse, and in certain buildings RBC even outperforms MPC with perfect forecasts, primarily due to how intra-interval dynamics interact with Net Billing tariffs and degradation costs. The paper argues for careful specification of settlement rules, alignment of evaluation timescales with real billing, and consideration of high-frequency fluctuations to enable fair, policy-relevant comparisons and guidance for practice.

Abstract

Residential prosumers with PV-battery systems increasingly manage their electricity exchange with the power grid to minimize costs. This study investigates the performance of Model Predictive Control (MPC) and Rule-Based Control (RBC) under 15/30/60 minute averaging commonly used in research, when Net Billing and battery degradation are considered. We simulate five consecutive months for 15 buildings in northern Germany, generating costs at up to 1-minute resolution while scheduling at 15/30/60 minutes. We find that time-averaged evaluations make MPC look consistently better than RBC, yet when costs are recomputed at minute-level ground-truth, the reported advantage shrinks by 69\% on average for hourly schedulers. For individual buildings, the finer evaluation can reverse conclusions, and simple RBC can achieve lower total costs than an MPC with perfect foresight. These findings caution against drawing conclusions from coarse averages and show how a fair assessment of battery scheduling approaches can be obtained.

Figures (6)

• Figure 1: General setting. The residential net-load $p_L$, the battery power $p_B$ and the grid power $p_G$ are in balance.
• Figure 2: Net-load $p_L$ averaged over three resolutions. High-resolution data can better approximate costs under Net Billing.
• Figure 3: Selected TOU prices for imports $c^{imp}$ and exports $c^{exp}$ with constant battery degradation costs $c^{{deg}}$.
• Figure 4: Total costs for two different ground-truth resolutions $\Delta gt$ with an MPC scheduler step length of $\Delta s = 30\text{min}$. All MPC approaches lead to comparatively high battery degradation costs when compared to RBC.
• Figure 5: Total costs for varying scheduler step lengths $\Delta s$. Left-hand blocks of a specific color indicate Fully Averaged Evaluations ($\Delta gt = \Delta s$). Right-hand blocks indicate Fine-Resolution Evaluations ($\Delta gt = 1\text{min}$). Averaging hides costs. Throughout all performed experiments, average evaluations using $\Delta gt = \Delta s$ underestimate costs. The bigger $\Delta s$, the bigger the underestimation.