⚛️ quantum physics

Stabilization of industrial processes with time series machine learning

This paper proposes a two-neural-network pipeline comprising an oracle predictor and an optimizer that replaces traditional point-wise optimization with network training, achieving a threefold improvement in temperature control stability for industrial processes compared to ordinary solvers.

Original authors: Matvei Anoshin, Olga Tsurkan, Vadim Lopatkin, Leonid Fedichkin

Published 2026-10-07

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Matvei Anoshin, Olga Tsurkan, Vadim Lopatkin, Leonid Fedichkin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the vast, humming world of modern industry, from the molten flows of steel mills to the delicate thermal balances of power plants, stability is everything. Factories run on time series data, which is simply a record of how things change over time: the temperature of a furnace, the pressure in a pipe, or the speed of a motor. Keeping these values steady is a constant battle against chaos. If a temperature spikes too high, equipment can break; if it drops too low, the product fails. For decades, engineers have relied on classical mathematical methods to solve this, treating the problem like a series of isolated steps. They would look at the current state, calculate the perfect adjustment for the next instant, make that change, and then repeat the process. While this works, it is often slow and computationally heavy, like trying to navigate a complex city by checking a map for every single step you take.

The challenge has always been finding a way to stabilize these processes faster and with less effort, without sacrificing precision. This is where the intersection of industrial engineering and machine learning offers a new path. Instead of calculating the perfect move for every single moment in time, researchers are exploring whether a computer can learn the general "feel" of the system. By training a machine to understand the relationships between different variables—such as how changing the current to a heater affects the room temperature hours later—it might be possible to predict the best course of action for a long stretch of time all at once. This shift from calculating individual steps to learning a continuous strategy promises to make industrial control more efficient, allowing systems to self-correct with a level of smoothness that traditional methods struggle to achieve.

In a recent study, researchers from the Moscow Institute of Physics and Technology proposed a new way to tackle this problem, focusing specifically on the stabilization of industrial processes. They approached the task by building a two-part system designed to mimic the behavior of a physical process and then learn how to control it. The first part of their system acts as a predictor, a digital twin that learns to forecast what will happen next based on past data. The second part is an optimizer, a neural network that learns the best way to adjust the controls to keep the process stable. The core idea is a clever substitution: rather than spending vast amounts of computer power trying to find the perfect value for every single moment, the system spends its time training the neural network to find the perfect set of internal settings, or weights, that produce the best results over a long period.

To test this idea, the team created a realistic simulation of a room temperature control problem. They modeled a room where the temperature is influenced by a heater, but also by external factors like the humidity in the air and the temperature outside. The goal was to keep the room at a steady twenty-three degrees Celsius. They gathered data on how the room responded to changes in the heater's current, including the natural fluctuations of the environment. Using this data, they trained their predictor model to understand the physics of the room, learning how the temperature would rise or fall given specific conditions. This model proved highly accurate, capable of predicting future temperatures with a very small margin of error, effectively capturing the complex dance of heat and air without needing to solve complex physics equations in real time.

Once the predictor was trained, the researchers turned to the optimizer. This second neural network was tasked with looking at the current state of the room and the predicted future, then deciding exactly how to adjust the heater current to keep the temperature on target. Instead of solving for one moment at a time, the optimizer learned a policy that worked across a window of time. The researchers tested this approach on data the system had never seen before, letting it run in a mode where it constantly updated its view of the room and adjusted the heater accordingly. The results were striking. The machine learning system kept the room temperature remarkably steady, hovering around the target with a deviation of just over one degree.

When compared to traditional methods, the new approach showed clear advantages in accuracy, though it required more computation time. The researchers tested their system against a standard industrial controller known as a PID, which is a widely used algorithm that adjusts controls based on the difference between the current value and the target. They also tested it against a more complex mathematical solver that tries to find the best answer by checking many possibilities one by one. The machine learning system performed slightly better than the standard PID controller, keeping the temperature more consistent by approximately 10%. However, this came at the cost of speed; the PID controller was significantly faster, making decisions in 0.05 seconds compared to the machine learning system's 1.46 seconds. In contrast, the traditional solver was vastly slower, taking hundreds of seconds to calculate a solution for a short window of time, and it was about three times less accurate in its stabilization than the machine learning approach.

The study suggests that this method of substituting the optimization of individual values with the training of a neural network offers a powerful alternative for industrial control. By learning the underlying patterns of a process, the system can anticipate changes and adjust controls proactively rather than reactively. The researchers found that this approach scales well, meaning it could potentially handle more complex systems with many more variables without becoming bogged down by the sheer amount of calculation required. While the work was conducted on a simulated dataset based on real-world physics, the results indicate that this technique could be a significant step forward for industries looking to improve the stability and efficiency of their processes. The ability to stabilize a system three times better than ordinary solvers, while using significantly less time than those solvers, points to a future where industrial processes are managed with a level of intelligence that goes beyond simple calculation.

Technical Summary: Stabilization of Industrial Processes with Time Series Machine Learning

Problem Statement
The stabilization of time series processes is a critical challenge across industrial sectors, particularly in metals and energy development. Classical optimization methods—such as finite-horizon Markov decision processes, non-linear programming, and point-wise optimization—are frequently used to improve production quality and minimize expenses. However, these approaches often suffer from high computational costs and resource intensity. Specifically, point-wise optimization, which seeks to minimize error at each individual time step, becomes exponentially difficult as the number of steps increases. This inefficiency is exacerbated in industrial settings where data spans vast numbers of sensors with high discretization steps, making real-time or near-future planning computationally prohibitive.

Methodology
The authors propose a machine learning (ML) driven pipeline that substitutes the direct optimization of controllable values with the optimization of neural network weights. The approach utilizes a two-network architecture:

The Oracle (Predictor): An enhanced Long Short-Term Memory (LSTM) model trained to predict the future state of the industrial process. It takes historical data (controlled, non-adjustable, and target features) and predicts the resulting target values over a specific time horizon ( $t_{range}$ ). In the experimental setup, this model predicts five features (including temperature and current) with high accuracy (MSE of 0.002–0.01).
The Optimizer: A neural network (implemented as a linear regression with ReLU activation) that learns the optimal control policy. Instead of solving for specific control values ( $x_{adjustable}$ ) at every step, the optimizer learns the weights ( $w$ ) that map historical data to the best sequence of control actions.

The Core Substitution
The central innovation is the substitution of the point-wise optimization problem (finding the best $x_{adjustable}$ values for a specific interval) with a weight approximation problem.

Classical Approach (Alg. 1): Iteratively optimizes control values for a specific time window using a cost function, then sets only the immediate value before moving to the next step. This requires solving an NP-hard problem repeatedly for every data snippet.
ML-Driven Approach (Alg. 2): The Optimizer network maps historical data to a control policy. The Oracle function predicts the outcome of this policy. A loss function (the difference between the predicted target and the ideal target) is backpropagated to update the Optimizer's weights via Stochastic Gradient Descent (SGD). Once trained, the Optimizer can be applied autoregressively to unseen data.

Experimental Setup
The methodology was validated using a synthetic dataset simulating room temperature control based on Newton's law of cooling with Joule heating.

Variables: Controlled features included heater current ( $I(t)$ ); non-adjustable features included absolute humidity, special humidity, and external temperature; the target was room temperature ( $T(t)$ ).
Data: The dataset was generated using a finite-difference solution to the thermal conductivity differential equation, incorporating realistic noise and environmental variations.
Training: The dataset (2764 points) was split into training and testing subsets. The Oracle was trained for 50 epochs, followed by the training of the Optimizer for 50 epochs.

Results
The proposed ML-driven approach was compared against two classical benchmarks: a Proportional-Integral-Derivative (PID) controller and a point-wise Nelder-Mead solver.

Stabilization Accuracy: The ML-driven autoregressive approach achieved a Root Mean Square Error (RMSE) of 1.15 °C. This outperformed the PID controller (RMSE: 1.20 °C) and significantly outperformed the Nelder-Mead solver (RMSE: 3.27 °C). The authors note the ML approach improved stability by approximately 3 times compared to the Nelder-Mead solver.
Computational Efficiency: The ML-driven optimization required 1.46 seconds for the optimization phase (application time), as listed in Table I. In contrast, the Nelder-Mead solver required 356.80 seconds for the same task. While the PID controller was faster (0.05 seconds), the ML approach demonstrated significantly reduced time requirements compared to ordinary solvers like Nelder-Mead, aligning with the paper's claim of requiring "less computational resources."
Scalability: The authors highlight that the ML approach offers superior scaling capabilities. The Optimizer network can be increased in size to capture more complex trends and dependencies without the exponential time cost associated with classical point-wise solvers.

Significance and Claims
The paper claims that substituting the optimization of controllable values with the optimization of neural network weights yields superior results in both work-time efficiency and optimization accuracy.

Performance Claims: The authors state that the ML-driven optimization provides "slightly better results" (approximately 10% improvement in RMSE) compared to the classical PID controller over a strictly specified physical model. Crucially, the methodology achieves this with significantly less computational resources required compared to ordinary solvers (like Nelder-Mead), rather than requiring more time.
Key Advantage: The primary contribution is the demonstration that this substitution allows for the stabilization of industrial processes with less computational resources during the optimization phase compared to ordinary solvers, while maintaining or improving stability. The approach is presented as a generalizable solution applicable to industrial problems with higher dimensions in both time and parameter scales.

Technical Summary: Stabilization of Industrial Processes with Time Series Machine Learning

More like this