Home Energy Management under Tiered Peak Power Charges

This paper proposes a model predictive control policy for home battery management under tiered peak power charges that, using simple forecasts, achieves electricity costs within 1.7% of the theoretical minimum derived from perfect foresight.

The Gist

Imagine your home is a busy restaurant, and the electricity grid is the kitchen supplying the ingredients. In the past, you only paid for the total amount of food (energy) you ate. But now, the restaurant owner (the utility company) has added a new, tricky rule: You also pay a "rush hour fee."

This fee isn't just based on how much you eat in a month; it's based on your three busiest days. If you order a massive feast on three specific days, your bill skyrockets, even if the rest of the month you ate very little. This is the "Tiered Peak Power Charge" the paper discusses.

The authors of this paper asked: "How can we use a home battery to cheat this system and save money?"

Here is the breakdown of their solution, explained simply:

1. The Problem: The "Three Worst Days" Trap

Imagine you have a battery (like a giant power bank for your house).

• Without a battery: If you turn on your oven, heater, and AC all at once on a cold winter day, you draw a huge amount of power. If this happens on three separate days in a month, you get hit with a massive penalty fee.
• The Goal: You want to use the battery to "shave" those peaks. Instead of drawing power from the grid when you need a lot, you draw it from the battery. This keeps your grid usage low, avoiding the penalty.

2. The "God Mode" Solution (The Prescient Policy)

First, the authors imagined a scenario where we have perfect foresight. We know exactly how much power your house will need every hour for the next year, and we know exactly what the electricity prices will be.

• The Analogy: This is like having a crystal ball. You know exactly when the "rush hour" is coming three days from now, so you fill your battery up two days ago to be ready.
• The Result: Using this "God Mode" math, they calculated the absolute lowest possible bill. It's the "gold standard" or the theoretical limit of how much you can save. In their test, this saved them 15.4% compared to having no battery at all.

3. The Real-World Solution (Model Predictive Control)

Obviously, we don't have crystal balls. We can't know the future perfectly. So, the authors built a smart robot brain called Model Predictive Control (MPC).

• How it works:

  1. Look Ahead: Every hour, the robot looks at the next 30 days (the "planning horizon").
  2. Guess: It uses simple forecasts to guess what your load will be and what prices will be.
  3. Plan: It solves a complex math puzzle to decide: "Should I charge the battery now? Should I save it for later? Should I risk a peak today to save money next week?"
  4. Act: It only executes the first step of that plan (e.g., "Charge the battery now").
  5. Repeat: An hour later, it gets new information, updates its guesses, and solves the puzzle again.

• The Analogy: Think of it like playing a game of chess against a computer. You don't know your opponent's next move perfectly, but you look 10 moves ahead, make the best move you can, and then re-evaluate when the opponent actually moves.

4. The Results: Almost Perfect

The team tested this on a real house in Norway with a 40 kWh battery (enough to power a small home for a day).

• No Battery: The bill was high.
• Simple Rules: If you just told the battery to "charge at night and discharge during the day" (Energy Arbitrage), you actually lost money. Why? Because you might accidentally create a peak on a day you didn't expect, triggering the penalty fee.
• The Smart Robot (MPC): By using their smart algorithm, the battery saved 13.9% of the total cost.
• The Gap: The smart robot's bill was only 1.7% higher than the "God Mode" perfect future bill.

5. The Secret Sauce: The Forecast

How does the robot guess the future? It uses a "Baseline-Residual" method.

• The Baseline: It knows the "rhythm" of your house. It knows you usually cook dinner at 6 PM and sleep at 11 PM (like a circadian rhythm).
• The Residual: It looks at the weird stuff. Did you just turn on the dryer unexpectedly? The robot learns from recent history to adjust for these surprises.

The Big Takeaway

This paper proves that you don't need a supercomputer or a crystal ball to save money on electricity. By using a smart, adaptive algorithm that looks ahead and constantly updates its plan, a home battery can navigate complex, tiered electricity fees almost as well as if it knew the future perfectly.

In short: The battery is your "time machine" for electricity. It lets you buy cheap power now and use it later, but only if you have a smart guide (the MPC) to tell you exactly when to use it so you don't get caught by the "rush hour" penalty.

Technical Summary

Here is a detailed technical summary of the paper "Home Energy Management under Tiered Peak Power Charges."

1. Problem Statement

The paper addresses the optimization of battery storage operation in a grid-connected home to minimize total electricity costs under a specific tariff structure. The tariff consists of two components:

1. Energy Charge: Based on consumption, incorporating time-of-use (TOU) prices and day-ahead wholesale prices.
2. Tiered Peak Power Charge: Unlike traditional demand charges based on a single maximum power value, the Norwegian residential tariff (introduced in 2022) calculates the charge based on the average of the $N$ largest daily peak powers (typically $N=3$) over a billing month. This average is mapped to a piecewise-constant cost function with multiple tiers.

The Challenge:

• Nonconvexity: The tiered peak power charge introduces nonconvexity into the optimization problem because the cost function is piecewise constant.
• Uncertainty: Real-time operation requires decisions based on forecasts of future loads and prices, whereas the "perfect foresight" scenario assumes all future data is known.
• Objective: Determine a control policy that minimizes the sum of energy and peak power costs while respecting battery constraints (capacity, charge/discharge rates, efficiency).

2. Methodology

The authors propose a two-pronged approach: establishing a theoretical lower bound and developing a real-time implementable policy.

A. The Prescient Problem (Lower Bound)

To determine the theoretical minimum cost, the authors formulate a Mixed-Integer Linear Program (MILP) assuming perfect knowledge of all future loads and prices.

• Formulation: The problem minimizes the total cost subject to power balance, battery dynamics, and capacity constraints.
• Handling the Tiered Charge: The piecewise-constant peak cost function $\phi(z_k)$ is modeled using binary variables ($s_{lk}$) that assign the monthly peak average $z_k$ to a specific cost tier $l$.
• Implementation: The problem is solved using CVXPY (a domain-specific language for convex optimization) and the Gurobi solver. The non-linear max and sum-largest functions are automatically transformed into linear constraints.

B. Model Predictive Control (MPC) Policy

For real-time operation, the authors develop an MPC policy that solves a similar optimization problem at every time step using forecasts.

• Horizon: A 30-day planning horizon ($H=720$ hours) is used to capture the full billing cycle, ensuring the controller anticipates monthly peak resets.
• Forecasting: A Baseline-Residual forecasting method is employed:
  • Baseline: Captures seasonal, weekly, and daily periodicities using a sum of sinusoids (Fourier series).
  • Residual: An Autoregressive (AR) model predicts short-term deviations from the baseline.
  • Loss Function: The models are fitted using pinball (quantile) loss with $\ell_2$ regularization to reduce sensitivity to outliers and bias forecasts appropriately.
• Optimization: At each step, the MPC solves a small MILP (or an equivalent LP enumeration scheme) over the horizon. The first control action is applied, and the process repeats with updated information.
• Alternative Formulation: The authors note that since the number of tiers ($L$) is small, the MILP can be solved by enumerating all possible tier assignments for the months in the horizon and solving a sequence of Linear Programs (LPs), which is computationally efficient.

3. Key Contributions

1. Novel Tariff Modeling: The paper is the first to address the specific structure of the Norwegian residential tariff, which uses the average of the $N$ largest daily peaks rather than a single maximum, combined with a tiered cost structure.
2. MILP Formulation for Tiered Peaks: The authors successfully formulate the nonconvex tiered peak charge problem as an MILP, providing a rigorous mathematical framework for this specific cost structure.
3. Near-Optimal Real-Time Control: They demonstrate that an MPC policy using simple forecasts can achieve costs within 1.7% of the theoretical prescient lower bound, proving that complex integer programming is feasible for real-time home energy management.
4. Quantification of Value: The study quantifies the value of jointly optimizing energy and peak costs versus optimizing them separately or using naive rule-based policies.

4. Results

The methods were evaluated using real data from a home in Trondheim, Norway, over the years 2020–2022, with a 40 kWh battery system.

• Cost Savings:

  • No Storage Baseline: 25,052 NOK/year.
  • Prescient (Perfect Foresight): 21,204 NOK/year (15.4% savings).
  • MPC Policy: 21,568 NOK/year (13.9% savings).
  • Gap: The MPC policy is only 1.7% away from the optimal prescient bound.

• Comparison with Simple Policies:

  • Peak Shaving (Rule-based): Achieved 5.2% savings. It reduced peak charges but missed energy arbitrage opportunities.
  • Energy Arbitrage (Rule-based): Reduced energy costs but drove peak power into the highest tier, resulting in a negative saving (-3.3%) compared to no storage. This highlights the danger of optimizing only one cost component.
  • MPC (Energy-only): Ignoring peak charges in the MPC objective led to only 1.6% savings, confirming the necessity of joint optimization.

• Sensitivity Analysis:

  • Horizon Length: A 30-day horizon is critical. Shorter horizons (e.g., 1 day) failed to anticipate monthly peak resets, leading to significantly higher costs.
  • Parameter $N$: Changing $N$ from 1 to 3 (the actual tariff definition) had a negligible impact on total cost, suggesting the controller is robust.
  • Forecasting: Adding the AR residual component to the baseline forecast improved results by only 0.4%, primarily by better anticipating short-term load peaks for peak shaving.

5. Significance

• Practical Applicability: The study proves that complex, nonconvex energy management problems involving integer variables can be solved in real-time (approx. 0.2 seconds per step) on standard hardware.
• Policy Relevance: As more countries adopt residential peak power tariffs to manage grid congestion, this work provides a blueprint for how consumers can utilize battery storage to minimize costs without needing perfect foresight.
• Behavioral Insight: The results demonstrate that naive battery control strategies (like simple peak shaving or pure arbitrage) can be counterproductive under tiered tariffs. A holistic optimization approach is essential to avoid increasing costs.
• Scalability: The use of MILP solvers and LP enumeration makes the approach scalable for residential applications, bridging the gap between theoretical optimization and practical implementation.