Data predictive control using regression trees and ensemble learning

Abstract

Decisions on how to best operate large complex plants such as natural gas processing, oil refineries, and energy efficient buildings are becoming ever so complex that model-based predictive control (MPC) algorithms must play an important role. However, a key factor prohibiting the widespread adoption of MPC, is the cost, time, and effort associated with learning first-principles dynamical models of the underlying physical system. An alternative approach is to employ learning algorithms to build black-box models which rely only on real-time data from the sensors. Machine learning is widely used for regression and classification, but thus far data-driven models have not been used for closed-loop control. We present novel Data Predictive Control (DPC) algorithms that use Regression Trees and Random Forests for receding horizon control. We demonstrate the strength of our approach with a case study on a bilinear building model identified using real weather data and sensor measurements. In a one-to-one comparison, we show that DPC explains 70\% variation in the MPC controller. We further apply DPC to a large scale multi-story EnergyPlus building model to curtail total power consumption in a Demand Response setting. In such cases, when the model-based controllers fail due to modeling cost, complexity and scalability, our results show that DPC curtails the desired power usage with high confidence.

Predictive control of a water distribution system based on process historian data

“…The example is a closed-loop simulation of 120 hours with initial state x(0) = [3, 2.44, 1.58, 1.5, 0.99, 1.51], (14) and the reference used is…”

“…In this setting, a widely used approach is to derive an explicit model from data to later use it in the controller. Regression trees and ensemble learning have been used by Jain et al to obtain a prediction model from data. A major breakthrough in system identification was the developing of subspace identification methods (see the works of Van Overschee and De Moor), which have also been used in the context of data‐driven predictive control .…”

Learning and Control Using Gaussian Processes

“…Using such models for control synthesis where some of the inputs must be optimized can lead to computationally intractable optimization. Our previous work uses an adaptation of Random Forests which overcomes this problem by separation of variables to derive a local linear input-output mapping at each time step [5]. This paper uses GPs for receding horizon control where the output mean and variance are analytical functions of the inputs, albeit non-convex.…”

“…A broad range of data-driven modeling, assessment, and control methods for DR with buildings have been investigated in the literature. Regression trees were used in our previous work [5], [11], [12] to model and compute set-point schedules of buildings for DR. Neural networks were used for MPC of a residential HVAC system in [13] and Deep Reinforcement Learning for scheduling electrical devices in [14]. However, these methods require huge amount of data.…”

Data‐driven optimal predictive control of seismic induced vibrations in frame structures

“…For this reason, a recent very active research line in the literature tackles the problem of combining Machine Learning and MPC, and demonstrates its effectiveness in different application domains, e.g. in the energy and automation communities [40][41][42][43][44] just to cite a few.…”