Robust Parametric Microgrid Dispatch Under Endogenous Uncertainty of Operation- and Temperature-Dependent Battery Degradation

This paper proposes a robust parametric model predictive control framework for microgrid dispatch that addresses the endogenous uncertainty of battery degradation by integrating an XGBoost-based probabilistic degradation model with tunable penalty weights to optimize the trade-off between operational costs and long-term battery health under varying temperature conditions.

The Gist

Imagine you own a high-tech lemonade stand in a very hot city. You have a solar panel on your roof to make free electricity, but the sun is unpredictable (sometimes it's blazing, sometimes cloudy). To keep your stand running smoothly, you have a giant battery that stores extra power for when the sun isn't shining.

This paper is about how to run that battery stand smartly so you don't go broke, but also so the battery doesn't die young.

Here is the breakdown of the problem and the solution, using simple analogies:

1. The Problem: The "Treadmill" Effect

Most people think of a battery like a gas tank: you fill it up, you use it, and that's it. But batteries are more like a pair of running shoes.

• The Issue: Every time you run (charge/discharge), the shoes wear out. If you run too fast (high power) or on hot pavement (high temperature), they wear out even faster.
• The Trap: If you try to save money by running the battery hard every day to avoid buying expensive grid power, you might save a few dollars today but have to buy a brand new pair of shoes (a new battery) in two years instead of ten.
• The "Chicken and Egg" Mystery: This is the tricky part. Your decision to run the battery causes it to wear out. But as it wears out, it becomes weaker and can't run as hard tomorrow. This is called Endogenous Uncertainty. It's like a runner whose speed depends on their shoes, but their shoes' condition depends on how hard they run. You can't predict the future perfectly because your own actions change the future.

2. The Solution: A "Crystal Ball" for Wear and Tear

The authors built a smart system to solve this. They did two main things:

A. The "Crystal Ball" (The XGBoost Model)

Instead of guessing how fast the battery wears out, they trained a super-smart computer brain (called XGBoost) on data from 196 real batteries.

• How it works: They didn't just ask, "How much will it wear?" They asked, "What is the range of possibilities?"
• The Analogy: Think of it like a weather forecast. A normal forecast says, "It will be 75°F." This model says, "It will likely be between 70°F and 80°F, but there's a small chance it hits 85°F if it's really humid."
• Why it matters: This allows the system to see that hot days or aggressive charging make the "wear and tear" much more unpredictable and dangerous.

B. The "Smart Coach" (Robust Parametric Dispatch)

Now that they have a crystal ball, they need a coach to tell the battery when to run and when to rest.

• The Old Way: The coach just says, "Run as hard as possible to save money!" (This kills the shoes).
• The New Way: The coach uses a Robust Optimization strategy. This means the coach plans for the worst-case scenario.
  • Imagine: The coach thinks, "If I push the runner too hard today, and the weather gets hot, the shoes might fall apart tomorrow. So, I'll push a little less today to be safe."
• The Tuning Knob: The system has a "tuning knob" (parameters) that decides how much to worry about the battery's health versus today's electricity bill. The paper uses a simulation to find the perfect setting for this knob so that over the battery's entire life (10+ years), you spend the least amount of money total.

3. The Results: Playing the Long Game

They tested this on a virtual microgrid (a small local power grid). Here is what happened:

• The "Greedy" Strategy: Tried to save money every day. Result: The battery died in about 4.5 years. Total cost was high because they had to replace it early.
• The "Average" Strategy: Tried to be careful but didn't plan for the worst. Result: The battery lasted longer, but still not as long as it could have.
• The "Robust" Strategy (This Paper): The system was slightly more conservative. It accepted paying a tiny bit more for electricity today to protect the battery.
  • Result: The battery lasted 3,680 days (about 10 years) instead of 1,665 days.
  • The Payoff: Even though they paid a bit more for electricity daily, the total cost over 10 years was the lowest because they didn't have to buy a new battery early.

The Big Takeaway

This paper teaches us that being too cheap today can be expensive tomorrow.

By using a smart computer model that understands how batteries actually get tired (especially in the heat) and planning for the worst-case scenarios, we can run our energy systems in a way that saves the most money over the battery's entire lifetime. It's the difference between running a marathon in flip-flops to save $5 on shoes, versus wearing proper running shoes to finish the race.

Technical Summary

Here is a detailed technical summary of the paper "Robust Parametric Microgrid Dispatch Under Endogenous Uncertainty of Operation- and Temperature-Dependent Battery Degradation."

1. Problem Statement

The paper addresses the critical challenge of optimizing microgrid dispatch in systems integrating Photovoltaic (PV) generation and Energy Storage Systems (ESS). While existing research focuses on balancing renewable intermittency and load demand, it often overlooks the accelerated degradation of batteries caused by frequent charging/discharging cycles, particularly under extreme temperatures.

The core problem is defined by three specific gaps in current literature:

1. Arbitrary Penalty Coefficients: Most studies penalize degradation in the objective function but lack a rigorous method to determine the optimal weight (penalty coefficient) that balances immediate operational costs against long-term battery life.
2. Deterministic vs. Stochastic Models: Existing optimization models often use deterministic degradation assumptions, failing to capture the inherent stochastic nature of battery aging observed in real-world experiments.
3. Endogenous Uncertainty (DDU): There is a lack of frameworks that model the bidirectional, decision-dependent relationship between dispatch and degradation. Dispatch decisions influence the probability distribution of future degradation, while the resulting degradation (reduced capacity) constrains future dispatch capabilities. This creates a complex Decision-Dependent Uncertainty (DDU) problem that standard robust optimization methods struggle to solve.

2. Methodology

The authors propose a Robust Parametric Microgrid Dispatch Framework that explicitly accounts for DDU. The methodology consists of three main components:

A. Probabilistic Degradation Model (XGBoost)

• Data Source: Trained on experimental data from 196 lithium-ion batteries (Sony/Murata US18650VTC5) under varying temperatures and aging conditions (calendar and cyclic).
• Algorithm: Utilizes XGBoost (Extreme Gradient Boosting) with Quantile Regression.
• Mechanism: Instead of predicting a single degradation rate, the model predicts a vector of quantiles ($q \in (0,1)$) for the capacity degradation rate.
• Inputs: Current capacity, ambient temperature, Depth of Discharge (DOD), and maximum charging/discharging power.
• Output: A probabilistic distribution of degradation rates, allowing the system to define "worst-case" scenarios based on a chosen confidence level (quantile $q$).

B. Parametric Model Predictive Control (MPC)

• Dispatch Formulation: A rolling-horizon MPC problem minimizes the sum of operational costs (grid purchase/sell) and degradation penalty terms.
• Penalty Terms: The objective function includes weighted penalties for:
  • Total energy throughput (proxy for Equivalent Full Cycles - EFC).
  • Operational range of stored energy (proxy for DOD).
  • Maximum charging/discharging power.
• Parameters: A vector of non-negative weights $\theta = (\theta_{EFC}, \theta_{DOD}, \theta_{C}, \theta_{D})$ controls the trade-off between cost and degradation.

C. Robust Optimization (RO) with DDU

• Objective: To find the optimal parameter vector $\theta$ that minimizes the Total Life-Cycle Cost (investment + discounted operational costs) over the battery's lifespan, considering the worst-case degradation path.
• DDU Modeling: The future battery capacity is treated as an uncertain variable whose distribution depends on the current dispatch policy. The uncertainty set is defined by the $q$-quantile forecast from the XGBoost model.
• Solution Approach:
  • The problem is formulated as a nested optimization (minimizing cost over $\theta$ while maximizing cost over worst-case degradation).
  • Due to computational intractability, the authors prove that the "max" operator (worst-case) corresponds to the lower bound of the uncertainty set (minimum capacity).
  • A Simulation-based Particle Swarm Optimization (PSO) is used. The PSO searches for the optimal $\theta$, while an inner loop simulates the battery's life-cycle using the MPC dispatch and the probabilistic degradation model to evaluate the total cost.

3. Key Contributions

1. Probabilistic Degradation Modeling: Development of an XGBoost-based model that captures the complex, non-linear, and stochastic relationship between operational stress factors (temperature, DOD, power) and battery degradation, providing prediction intervals rather than point estimates.
2. Robust Parametric Framework: A novel dispatch framework that explicitly models Decision-Dependent Uncertainty (DDU). It treats the degradation process as endogenous, where dispatch decisions alter the probability distribution of future battery states.
3. Optimal Parameter Tuning: A method to automatically tune the degradation penalty weights ($\theta$) via life-cycle simulation and robust optimization, eliminating the need for arbitrary coefficient selection.
4. Comprehensive Validation: Extensive case studies demonstrating that the proposed method outperforms deterministic and risk-neutral approaches in terms of both total life-cycle cost and battery lifespan.

4. Results

The framework was validated using a PV-ESS integrated microgrid case study (Initial Capacity: 910.8 kWh, Ambient Temp: 35°C).

• Degradation Model Performance:

  • The XGBoost model achieved unbiased median predictions (centered around zero error).
  • Prediction intervals (e.g., 10th–90th percentiles) successfully enveloped the vast majority of actual test data, confirming the model's ability to quantify uncertainty.
  • Sensitivity analysis confirmed that degradation uncertainty is highly dependent on temperature and operational decisions (DDU).

• Dispatch Strategy Comparison:

  • Benchmark (No Degradation Penalty): Aggressive usage led to the shortest battery life (~1,665 days) and highest worst-case costs ($1.38M).
  • Stochastic Programming (Risk-Neutral): Extended life but failed to hedge against worst-case scenarios effectively.
  • Proposed Robust Optimization (RO):
    • Achieved the longest mean battery life (~3,680 days).
    • Resulted in the lowest total life-cycle cost in both 90% ($1.142M) and 95% ($1.178M) worst-case scenarios.
    • Demonstrated that a more conservative dispatch strategy (higher penalty weights) effectively hedges against uncertainty, extending battery life and reducing replacement costs.

• Sensitivity Analysis:

  • Quantile ($q$): Higher $q$ values (more conservative) generally increased battery life but showed a complex trade-off with total cost; $q=0.90$ offered the best balance.
  • Temperature: Extreme temperatures (5°C and 50°C) significantly increased life-cycle costs and reduced lifespan compared to the 35°C baseline, with 50°C reducing life by nearly 3 years.
  • Capacity: Larger initial ESS capacities provided operational flexibility, lowering DOD stress and reducing total life-cycle costs.

5. Significance

This paper makes a significant contribution to the field of microgrid energy management by shifting the paradigm from deterministic, static degradation models to dynamic, probabilistic, and decision-dependent frameworks.

• Economic Impact: By optimizing the trade-off between operational savings and battery replacement costs, the proposed method offers a more accurate and robust strategy for minimizing the Levelized Cost of Storage (LCOS).
• Theoretical Advancement: It provides a solvable approach to the complex class of Robust Optimization problems with Endogenous Uncertainty, which has been a major theoretical hurdle in operations research.
• Practical Application: The framework is directly applicable to real-world microgrid operators and ESS planners, offering a data-driven tool to extend asset life and ensure system reliability under varying environmental and operational conditions.

In conclusion, the study demonstrates that acknowledging and mathematically modeling the feedback loop between dispatch decisions and battery degradation is essential for achieving truly cost-effective and sustainable microgrid operations.