Who pulls the strings in industrial control systems?

Large industrial plants use LP‑MPC, an algorithm that decides in real time how to adjust hundreds of variables. Imagine conducting an orchestra with hundreds of instruments: each instrument (manipulated variable, MV) must play at the right moment to keep the harmony (controlled variables, CV) and respect the concert rules (constraints). The issue is that, even though the model is transparent, the controller's behavior can sometimes seem mysterious. Researchers have uncovered a way to demystify this using "shadow prices" – values that indicate how costly it is to violate a constraint. Instead of treating them as abstract numbers, they reinterpret them as the sum of cost sensitivities contributed by each MV. Then, using a linear assignment algorithm, they turn these contributions into MV‑CV pairs, showing who is responsible for which constraint. The result is a post‑hoc explainability tool: engineers can instantly see which variables are pushing the limits, diagnose inefficiencies, and verify that the controller follows the original design. In practice, it’s like having a digital magnifying glass that highlights hidden relationships between the system’s levers and the outcomes we aim to keep under control.

LP‑MPC (Linear Program‑Model Predictive Control) merges linear programming with short‑term prediction to optimize complex processes. In a linear model, the optimal solution features basic and non‑basic variables, and shadow prices (Lagrange multipliers) correspond to constraints. Traditionally, these values indicate the marginal change in the objective for an infinitesimal constraint variation, but they do not reveal which manipulated variable (MV) is responsible for the pressure on the constraint. The proposed approach reconstructs the MV‑CV sensitivity matrix by computing the partial derivative of the dual value with respect to each MV, thus attributing a cost contribution to each MV for each CV. This generally dense matrix is then solved using a linear assignment algorithm (e.g., the Hungarian algorithm) to find an optimal one‑to‑one matching that minimizes the sum of residual costs. The outcome is an MV‑CV map highlighting the most influential pairs. Practically, this map quickly identifies operational bottlenecks: if a critical CV shows high shadow prices, the map reveals which MVs are pushing that constraint toward its limit. This facilitates diagnosis of sub‑optimal behavior, set‑point recalibration, and verification of design consistency. Moreover, the methodology scales well: computational complexity is dominated by the linear assignment, O(n^3), where n is the number of variables, making it suitable for systems with hundreds of MVs and CVs in real time. Embedding this explainability tool in industrial control systems can boost operational transparency, shorten intervention times, and support more informed decisions by process engineers.