Carbon-aware Market Participation for Building Energy Management Systems

This paper proposes a unified, real-time carbon-aware building energy management system that integrates a Transformer-based forecasting model with a mixed-integer linear program to co-optimize grid imports, storage, and flexible demand across day-ahead and real-time markets, achieving a 22.5% emission reduction with only a 1.7% cost increase.

The Gist

Imagine your building is a smart, hungry traveler trying to get from Point A to Point B (getting through the day) while carrying a heavy backpack of energy.

Traditionally, this traveler only had one goal: spend the least amount of money possible. They would buy energy whenever it was cheapest, regardless of where that energy came from. If the cheapest energy happened to be generated by a dirty coal plant spewing smoke, the traveler didn't care. They just wanted the bargain.

This paper introduces a new rule for the traveler: "You must also care about the smoke you're breathing."

Here is the breakdown of the paper's big idea, using simple analogies:

1. The Problem: The "Blind" Wallet

Most building energy systems (EMS) are like a shopper who only looks at the price tag. They buy the cheapest electricity available. The problem is that electricity isn't all the same.

• The Dirty Hour: Sometimes, the grid is running on coal or gas. Buying power then is cheap but creates a lot of carbon pollution.
• The Clean Hour: Sometimes, the sun is shining or the wind is blowing. Power is cheap and clean.
• The Messy Reality: The old systems don't know the difference. They just buy the cheapest option, accidentally buying "dirty" power when it's cheap, which hurts the climate.

2. The Solution: The "Carbon-Aware" GPS

The authors built a new system called CAEMS (Carbon-Aware Energy Management System). Think of this as a super-smart GPS for your building's energy wallet.

This GPS doesn't just look at the Price of gas; it also looks at the Pollution level of the road.

• The Goal: It tries to do two things at once: keep the bill low and keep the carbon footprint low.
• The Magic Tool: It uses a "Carbon Price." Imagine the government or the market says, "Every time you buy dirty electricity, you have to pay a tiny extra tax." This makes dirty electricity look expensive on the GPS screen, even if the base price is low.

3. How It Works: The "Time-Traveling" Battery

The building has two main tools to help this GPS:

1. A Battery (The Storage Tank): Like a water tank, it can fill up when water is cheap and clean, and use it when water is expensive or dirty.
2. Flexible Demand (The Flexible Schedule): This is like a flexible meeting schedule. If the building needs to run its AC or charge its EVs, it can wait.

The Strategy:
The system says: "Hey, right now the sun is out, the wind is blowing, and the electricity is clean and cheap. Let's fill up the battery and run the heavy machines now!"
Then, at night when the grid is dirty and expensive, it says: "Don't buy power from the grid. Use the clean energy we stored earlier."

4. The "Crystal Ball" (AI Forecasting)

You can't just wait and see what happens; you have to plan ahead. The old systems assumed they knew the future perfectly (which is impossible).

This new system uses a Transformer AI (a type of advanced computer brain) to act as a crystal ball.

• It looks at the last 24 hours of weather, prices, and pollution.
• It predicts the next 24 hours: "Tomorrow at 2 PM, the sun will be strong, so electricity will be clean. At 6 PM, everyone will turn on their lights, and the grid will get dirty."
• It uses this prediction to make a plan, but it updates the plan every hour as new real-world data comes in. This is called Model Predictive Control (MPC). It's like checking your GPS every 10 minutes to adjust for traffic, rather than just setting it once and hoping for the best.

5. The Results: A Win-Win

The authors tested this in the real world using data from the PJM electricity market (a huge grid in the US).

• The Result: By adding a small "carbon tax" to their calculations, they reduced carbon emissions by 22.5%.
• The Cost: The building's energy bill only went up by 1.7%.

The Analogy:
Imagine you are driving to work.

• Old Way: You take the fastest route, even if it goes through a smoggy tunnel.
• New Way: You take a slightly longer route that avoids the smog.
• The Paper's Finding: The "smoggy route" isn't actually that much faster (only 1.7% more time/money), but the "clean route" saves you from a massive headache (22.5% less pollution).

Why This Matters

This paper proves that we don't have to choose between saving money and saving the planet. By using smart software to shift when we use electricity, we can:

1. Clean the Grid: We stop buying power when it's dirty.
2. Save Money: We buy power when it's cheap (which often overlaps with when it's clean).
3. No Magic Required: It uses existing batteries and smart thermostats; it just needs a smarter brain to tell them when to switch.

In short, this system teaches our buildings to be smart shoppers who buy "green" deals, rather than just cheap shoppers who buy whatever is on sale.

Technical Summary

1. Problem Statement

The decarbonization of electric power systems is critical for combating climate change, yet traditional Energy Management Systems (EMS) in buildings primarily focus on economic efficiency (minimizing cost) and reliability, often overlooking the environmental impact of operational decisions.

• The Gap: Existing approaches typically address supply-side mitigation or treat demand-side flexibility and carbon reduction as separate, independent problems. There is a lack of a unified, real-time framework that simultaneously co-optimizes grid imports, energy storage, and flexible demand while explicitly integrating time-varying carbon intensity signals.
• The Challenge: Conventional optimization assumes "perfect foresight" of future prices and emissions, which is unrealistic. Furthermore, integrating carbon signals into market participation (Day-Ahead and Real-Time) requires a robust method to handle forecast uncertainty without compromising system reliability.

2. Methodology

The authors propose a Carbon-Aware Energy Management System (CAEMS) that operates at the building level. The methodology consists of three core components:

A. Unified Optimization Framework (MILP)

The core of the CAEMS is a Mixed-Integer Linear Program (MILP) formulated to minimize a total cost function ($J$) comprising both monetary energy costs ($p_E$) and carbon costs ($p_C$).

• Objective: Minimize $J = p_E + p_C$.
  • $p_E$: Aggregates Day-Ahead (DA) and Real-Time (RT) electricity prices.
  • $p_C$: Aggregates marginal carbon intensities multiplied by a carbon price ($\pi_C$).
• Decision Variables: The model determines real-time grid imports ($g_t$), Energy Storage System (ESS) charging/discharging ($b_t$), and flexible demand shifting ($\Delta_t$).
• Constraints: The model enforces physical laws including:
  • Power Balance: Supply must meet inelastic demand plus shifted demand.
  • ESS Dynamics: State-of-Charge (SoC) evolution, charge/discharge limits, and mutual exclusivity (cannot charge and discharge simultaneously).
  • Demand Shifting: Flexible loads must return to baseline by the end of the horizon.
• Market Integration: The system coordinates participation in both DA and RT markets, treating DA schedules as a fixed baseline while optimizing RT deviations.

B. Model Predictive Control (MPC)

To address the unrealistic assumption of perfect foresight, the deterministic MILP is embedded within an MPC framework.

• Receding Horizon: At each time step $t$, the system solves the optimization over a future horizon $H$, implements only the first-step decision, and then re-solves using updated measurements and forecasts.
• Robustness: This approach allows the system to dynamically correct for prediction errors regarding prices and carbon intensity.

C. Transformer-Based Forecasting

A critical enabler for the MPC is a deep learning module that jointly predicts electricity prices and carbon intensity.

• Architecture: A Transformer network (utilizing self-attention mechanisms) is employed to capture long-range temporal dependencies in time-series data.
• Inputs/Outputs: The model takes historical sequences of DA/RT prices, demand, and carbon intensities to forecast the next 24 hours.
• Advantage: Unlike recurrent networks (e.g., LSTMs), the Transformer excels at parallel processing and capturing complex temporal patterns, providing the high-accuracy forecasts necessary for robust MPC.

3. Key Contributions

1. Unified Co-Optimization: Proposes a single integrated framework that simultaneously optimizes grid imports, storage, and flexible demand under dual objectives: cost minimization and carbon emission reduction.
2. Market-Ready Formulation: Develops a MILP model capable of coordinated control across both Day-Ahead and Real-Time electricity markets, explicitly incorporating time-varying marginal carbon intensity signals.
3. AI-Driven Uncertainty Handling: Introduces a Transformer-based forecaster for joint prediction of prices and carbon intensity, enabling a practical MPC extension that operates effectively without perfect foresight.

4. Simulation Results

The proposed CAEMS was validated using real-world hourly data from the PJM electricity market over a 10-day period.

• Experimental Setup:

  • Baseline: Unmanaged operation (grid imports match demand exactly).
  • Comparators: Cost-only dispatch (Carbon Price = $0/t) vs. Carbon-aware dispatch (Carbon Price = $30/t).
  • Forecasting Performance: The Transformer model achieved the lowest Root Mean Squared Error (RMSE = 0.0462) compared to other models like XGBoost, LSTM, and ARIMA.

• Key Findings:

  • Emission Reduction: Implementing a modest carbon price of $30/ton resulted in a 22.5% reduction in daily carbon emissions (dropping from 20.4 t/day to 15.8 t/day).
  • Cost Impact: This significant emission reduction came with only a 1.7% increase in daily energy costs (from $653 to $664).
  • Operational Dynamics: The carbon-aware strategy successfully shifted procurement away from high-emission periods (e.g., early morning peaks) and high-price spikes. It demonstrated the ability to curtail imports during high-carbon hours, a capability absent in cost-only dispatch.
  • Trade-off: The results illustrate a highly favorable Pareto frontier where marginal cost increases yield substantial environmental benefits.

5. Significance and Impact

• Technical Feasibility: The paper proves that carbon-aware dispatch is technically viable and can be integrated into existing building EMS without compromising reliability.
• Economic Viability: The study demonstrates that the "cost of decarbonization" for building operators is minimal (approx. 1.7%), suggesting that carbon pricing mechanisms can be adopted without prohibitive economic burdens.
• Policy and Market Implications:
  • For Market Operators: The framework supports the integration of Locational Marginal Carbon Emission (LMCE) signals into market clearing and settlement processes.
  • For Policymakers: It provides a pathway for demand-side participation in decarbonization, showing how flexible loads and storage can act as active resources to flatten emission peaks.
  • Scalability: The modular design allows for the integration of more complex policy mechanisms (e.g., cap-and-trade, tiered pricing) in future iterations.

In conclusion, this work bridges the gap between economic optimization and environmental goals, offering a practical, data-driven solution for building-level participation in a decarbonizing power grid.