Managing the power grid is like solving a giant puzzle.
Grid operators must ensure that the right amount of power flows to the right areas when it is needed, and they must do this in a way that minimizes costs without overloading the physical infrastructure. Moreover, this complex problem must be solved iteratively and as quickly as possible to meet ever-changing demands.
To solve this persistent challenge, researchers at MIT have developed a problem-solving tool that finds optimal solutions much faster than traditional approaches, while ensuring that the solution does not violate system constraints. In the power grid, constraints include the capacity of generators and transmission lines.
This new tool incorporates a feasibility step into a powerful machine learning model trained to solve your problem. The feasibility step uses the model’s predictions as a starting point and iteratively refines the solution until the best achievable answer is found.
The MIT system can solve complex problems many times faster than traditional solvers, with strong guarantees of success. For some very complex problems, you may find a better solution than proven tools. The technique also outperformed pure machine learning approaches, which are fast but cannot always find viable solutions.
In addition to helping schedule power generation on the grid, this new tool has the potential to be applied to many types of complex problems, such as designing new products, managing investment portfolios, and planning production to meet consumer demand.
“Successfully solving these particularly vexing problems requires combining the tools of machine learning, optimization, and electrical engineering to develop methods that achieve the right trade-offs in delivering value to the domain while meeting its requirements. We need to consider the needs of the application and design methods in a way that actually meets those needs,” said Silverman Family, School of Electrical Engineering and Computer Science (EECS). said Priya Donti, professor of career development and principal investigator at the Information Science Institute. Decision-Making Systems (LIDS).
Donti, senior author of an open access paper on this new tool called FSNet, is joined by lead author Hoang Nguyen, an EECS graduate student. This paper will be presented at the Neural Information Processing Systems Conference.
combine approaches
Ensuring optimal power flow within the power grid is a very difficult problem that is becoming increasingly difficult for operators to solve quickly.
“As we try to integrate more renewable energy into the grid, operators have to deal with the fact that the amount of power generation changes from moment to moment, and at the same time there are more distributed devices to coordinate,” Donti explains.
Grid operators often rely on traditional solvers that mathematically guarantee that the optimal solution does not violate the problem constraints. However, if the problem is particularly complex, these tools can take hours or even days to reach a solution.
Deep learning models, on the other hand, can solve even the most difficult problems in a fraction of the time, but their solutions may ignore some important constraints. For power grid operators, this can lead to problems such as dangerous voltage levels and grid outages.
“Machine learning models have a hard time meeting all constraints because they make a lot of errors during the training process,” Nguyen explains.
In the case of FSNet, researchers combined the strengths of both approaches to create a two-step problem-solving framework.
Emphasis on feasibility
In the first step, the neural network predicts the solution to the optimization problem. Very loosely inspired by the neurons in the human brain, neural networks are deep learning models that excel at recognizing patterns in data.
A traditional solver built into FSNet then performs the feasibility steps. This optimization algorithm iteratively refines the initial prediction while ensuring that the solution does not violate any constraints.
The feasibility step is based on a mathematical model of the problem, ensuring that the solution is deployable.
“This step is very important. With FSNet, you can get the strict guarantees you actually need,” says Hoang.
The researchers designed FSNet to be able to address both major types of constraints (equality and inequality) simultaneously. This makes it easier to use than other approaches that require customizing neural networks or solving each type of constraint separately.
“Here you can plug and play different optimization solvers,” says Donti.
By thinking differently about how neural networks solve complex optimization problems, researchers were able to uncover new techniques that work more effectively, she added.
They compared FSNet to traditional solvers and pure machine learning approaches for a variety of difficult problems, including power grid optimization. Their system reduced solution time by orders of magnitude compared to baseline approaches while respecting all problem constraints.
FSNet also discovered better solutions to some of the most vexing problems.
“This was a surprise to us, but not surprising: our neural network is uniquely able to recognize additional structure in the data that the original optimization solver was not designed to take advantage of,” Donti explains.
In the future, the researchers hope to reduce FSNet’s memory consumption, incorporate more efficient optimization algorithms, and scale it up to tackle more realistic problems.
“Finding near-optimal optimization problems is paramount to finding viable solutions to difficult optimization problems. Especially for physical systems such as power grids, near-optimality means nothing without feasibility. This work is an important step toward enabling deep learning models to explicitly guarantee constraint enforcement and generate predictions that satisfy the constraints,” said Kairi Baker, an associate professor at the University of Colorado Boulder, who was not involved in the study.
