Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems

arXiv:2608.15076 · eess.SY, cs.SY · Submitted 2026-08-15 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems".

Dev: Here's the gist of the paper. The dissertation addresses the challenge of integrating industrial loads, which constitute over 60% of electricity consumption, into flexible power systems by overcoming barriers related to complex process constraints, private equipment parameters, and model complexity for optimization at scale.

Rosa: The first major contribution involves reforming existing industrial process models—the State Task Network (STN) and Resource Task Network (RTN)—into the Linearized State Task Network (LSTN) and Continuous Resource Task Network (cRTN).

Dev: This reformulation reduces the number of integer variables while preserving essential process constraints, significantly shortening the solution time for standard grid-interaction cases from 24 hours to just 30 minutes and enabling coordinated optimization across 2,000 industrial users.

Rosa: To estimate private equipment parameters, a Production Scheduling Identification (PSI) method is developed for hourly smart meter data.

Title and authors: Dev: So we started by talking about the title and the authors of this paper. Rosa, what's the main takeaway from just reading that?

Rosa: Well, it’s clear that they are focused on overcoming three major hurdles: complex process constraints, private equipment parameters, and the sheer model complexity when you try to optimize for thousands of users at once. They set out to solve those barriers from the perspective of a load aggregator.

Dev: That sounds like a lot to tackle all at once. What do you think the title "Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems" really mean in plain English?

Rosa: It means they are figuring out how industrial users, who have specific ways they operate, can participate in electricity markets effectively. It’s about bridging the gap between how a factory actually runs its production and how the power grid handles that flexible demand.

Taro: I wonder if the focus on "market-based interaction" implies that this is less about just technical modeling and more about making money through those interactions?

Dev: That's a fair point, Taro. It suggests they are not just building a simulation; they're building a system that lets aggregators maximize profit by coordinating these industrial users in the market.

Rosa: Right. And that’s what we saw earlier when we talked about the final part of their work—the bidding and disaggregation framework—it all ties back to making decisions for market income and real-time control.

The paper's summary: Rosa: Now, let's look at the actual summary of "Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems." They propose two key model reformulations: the Linearized State Task Network or LSTN, and the Continuous Resource Task Network or cRTN.

Dev: So they are taking these existing models, which are called STN and RTN, and transforming them into LSTN and cRTN to make things faster. How does that actually help in a practical sense?

Rosa: By reducing the number of integer variables they have to deal with while still keeping all the essential process constraints. This reduction is what allows them to cut the solution time for a standard grid interaction case from twenty-four hours down to just thirty minutes, and it lets them handle coordinating two thousand industrial users.

Taro: Thirty minutes is fast for optimization, but I have to ask, when you're dealing with real industrial processes like cement or steel powder—which are inherently discrete—does this continuous modeling really capture everything?

Dev: The authors say that treating continuous industrial processes as a degenerate case of discrete ones helps. They argue that this reformulation makes the problem tractable enough to solve quickly for coordinated optimization, which is a big deal when you have so many users involved.

Rosa: And it’s not just about speed; they also tackle the problem of private equipment parameters. They introduce a method called Production Scheduling Identification or PSI to estimate these private details using hourly smart meter data and cost-minimizing scheduling behavior.

Taro: So, the paper says that by using actual process mechanisms and how people schedule things to save money, you can get those private equipment parameters relatively accurately?

Rosa: Yes. Testing on cement and steel powder process models using actual prices in PJM showed load-model errors of five point two percent and eight point five percent over twenty-one training days. That’s much better than the comparison machine learning models they tested, which had errors ranging from thirteen point four to nineteen point two percent.

The paper's improvements: Dev: So we've seen the modeling reformulations and the parameter identification method. What's next? What are some of the specific improvements they suggest for these models?

Rosa: The biggest improvement seems to be in how they reduce constraints during aggregation. They introduce a Data-Driven Dimension Reduction or D3R framework that learns low-dimensional linear constraints directly from operating samples generated by the original process models.

Taro: That sounds like using the data itself to simplify the math, which is smart, but how much simplification are we actually talking about here?

Rosa: For the steelmaking case specifically, they reduced a model that had over ten thousand two hundred eight integer variables down to only twenty-four to forty-eight continuous variables. That level of reduction is lower than what current demand side participation rules allow, which is significant.

Dev: And this D3R framework isn't just for single factories; the authors extend it to aggregate distributed resource fleets and claim that it reduces the approximation error by over forty percent when coupling constraints are present.

Taro: That sounds like a real win for handling complexity at scale, but what’s the limitation on that D3R? Where does this data-driven reduction stop working?

Rosa: The paper notes that while D3R is highly accurate across cement and steelmaking, its effectiveness depends on having good operating samples. The authors also flag that when constraints are coupled in a more complicated way, the error can still be substantial.

Conclusion: Dev: So we've covered the modeling speed-up with LSTN and cRTN, how they found private parameters with PSI, and how they shrink huge models using D3R. What’s the final word on what this whole approach changes for load aggregators?

Rosa: The final piece is the "dimension-reduced bidding and precise disaggregation" framework. They split the optimization into an upper layer for long-term profit maximization during bidding, using those reduced constraints, and a lower layer for real-time power disaggregation.

Taro: So it’s not just one big model running everything at once; it’s a two-tiered approach where you plan the bidding with the simplified model first, and then use shadow prices to allocate real-time commands millisecond by millisecond?

Dev: Exactly. And they show that this strategy can increase average market income by fifty-three point three percent and cut operating costs by eleven point four percent compared to just proportional disaggregation.

Rosa: That level of control—millisecond level arithmetic—is what makes it powerful for real-time power management, and they say it achieves a four-fold reduction in interaction costs compared to simpler strategies. This whole "Industrial Load Modeling and Optimization for Market-Based Interaction with Power Systems" paper shows a comprehensive pipeline connecting process constraints directly to aggregator decisions.

Taro: For me, the real implication is that this moves the conversation from just building bigger models to building smarter, more adaptive models that can handle the noise of real-world operation while still hitting those market targets.

Dev: It certainly provides a concrete path for how an aggregator can actually translate complex industrial process data into actionable power system control commands.

Tsinghua University · North China Electric Power University · Zhejiang University

eess.SY, cs.SY

Submitted: 2026-08-15

Updated: 2026-11-10

Comments: PhD thesis, Tsinghua University, June 2026, 176 pages

Code: https://github.com/Rick10119/Efficient-Scheduling-of-Discrete-Industrial-Processes

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 100/100

The gist: The dissertation addresses the challenge of integrating industrial loads, which constitute over 60% of electricity consumption…

Key concepts

Linearized State Task Network (LSTN)
This reformulates the original process model by replacing integer start-up variables with continuous operating time variables. This eliminates integer variables entirely, allowing complex optimization problems involving thousands of users to be solved very quickly, typically within minutes.
Production Scheduling Identification (PSI)
This method estimates unknown private equipment parameters by analyzing how industrial processes and cost-minimizing schedules behave. By incorporating these behaviors into the estimation, it produces highly accurate load models with low error rates compared to other machine learning approaches.
Data-Driven Dimension Reduction (D3R)
D3R learns essential low-dimensional linear constraints directly from operating samples generated by the process models. This framework successfully reduces massive models, such as those for steelmaking, from thousands of integer variables down to just a few dozen continuous variables, meeting strict market participation rules.
Dimension-Reduced Bidding and Disaggregation
This is a two-layer control system. The upper layer uses simplified constraints for long-term profit maximization during bidding. The lower layer uses shadow prices calculated in the first layer to allocate real-time commands to thousands of resources using fast arithmetic, boosting market income significantly.

Terminology

Summary

The dissertation addresses the challenge of integrating industrial loads, which constitute over 60% of electricity consumption, into flexible power systems by overcoming barriers related to complex process constraints, private equipment parameters, and model complexity for optimization at scale. p. 2

The first major contribution involves reforming existing industrial process models—the State Task Network (STN) and Resource Task Network (RTN)—into the Linearized State Task Network (LSTN) and Continuous Resource Task Network (cRTN). p. 2 This reformulation reduces the number of integer variables while preserving essential process constraints, significantly shortening the solution time for standard grid-interaction cases from 24 hours to just 30 minutes and enabling coordinated optimization across 2,000 industrial users. p. 2

To estimate private equipment parameters, a Production Scheduling Identification (PSI) method is developed for hourly smart meter data. p. 2 PSI leverages industrial process mechanisms and cost-minimizing scheduling behavior to narrow the parameter search space. p. 2 Testing on cement and steel powder process models using actual PJM prices demonstrated load-model errors of 5. p. 22% and 8.5% over 21 training days, outperforming comparison machine-learning models which yielded errors of 13. p. 24%–19.2%.

For aggregator-level optimization, a Data-Driven Dimension Reduction (D3R) framework is proposed to learn low-dimensional linear constraints directly from operating samples generated by the original process models. p. 2 D3R has shown high accuracy across cement, steelmaking, and steel-powder cases, achieving errors of 3. p. 26%–10.3%, which falls within or near the 10%–20% tolerances specified in current demand-side participation rules. p. 2 Notably, for the steelmaking case, D3R successfully reduces a model with over 10,208 integer variables down to only 24–48 continuous variables. p. 2 This framework is extended to aggregate distributed resource fleets, reducing approximation error by over 40% relative to conventional outer approximation when coupling constraints are present. p. 8

Finally, a dimension-reduced bidding and precise disaggregation framework is developed for load aggregators. This approach separates the optimization into an upper layer for long-term profit maximization during bidding (using dimension-reduced constraints) and a lower layer for real-time power disaggregation. p. 9 Shadow prices calculated during the bidding phase are then used to allocate real-time commands among tens of thousands of resources through millisecond-level arithmetic. p. 3 This strategy has been shown to increase average market income by 53. p. 93% and reduce operating costs by 11. p. 24% compared to proportional disaggregation, while maintaining the solution quality of the joint optimization model. p. 3 The fast disaggregation algorithm achieves millisecond-level control, proving optimal under certain assumptions and demonstrating a four-fold reduction in interaction costs relative to simplified strategies. p. 3

In summary, the dissertation establishes a comprehensive four-step pipeline: (1) Efficient Load Modeling via LSTN/cRTN, (2) Load Parameter Identification via PSI, (3) Model Constraint Reduction via D3R, and (4) Optimized Decision-Making through the joint bidding and disaggregation framework. p. 6 This integrated approach effectively links industrial process constraints to aggregator decision-making for coordinated optimization of market bidding and real-time power disaggregation. p. 3

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Variable Reformulation for Speed and Tractability: The Linearized State Task Network (LSTN) model replaces integer start-up variables with continuous operating time variables, thereby completely eliminating integer variables and allowing for solving a coordinated case with 2,000 industrial users within minutes.

  2. Continuous Modeling of Discrete Processes: The Continuous Resource Task Network (cRTN) model achieves higher computational efficiency by treating tasks and resources as continuous variables, showing that continuous industrial processes are a degenerate case of discrete industrial processes and reducing the number of integer variables from 4bST to at most 4ST.

  3. Parameter Identification via Cost-Minimizing Scheduling: The Production Scheduling Identification (PSI) method estimates private equipment parameters by incorporat[ing] industrial process mechanisms and cost-minimizing scheduling behavior, producing load-model errors of 5.2% and 8.5% on test datasets, which is less than half the errors of comparison machine learning models.

  4. Data-Driven Constraint Reduction: The Data-Driven Dimension Reduction (D3R) framework learns low-dimensional linear constraints from operating samples, reducing a model containing 10,208 integer variables for steelmaking to 24–48 continuous variables, which is lower than current demand-side participation rules tolerances.

  5. Hierarchical Control Architecture for Aggregators: The system can implement a hierarchical control architecture where the VPP first computes aggregated demand response capacity using parameters from the factory set, and then uses shadow prices to allocate real-time commands through millisecond-level arithmetic, reducing interaction costs by 40% relative to a simplified strategy.

Sources

Related papers