Douglas C. Hittle

MIMO Robust Control for HVAC Systems

Potential improvements in heating, ventilating, and air conditioning (HVAC) system performance are investigated through the application of multiple-input-multiple-output (MIMO) robust controllers. This approach differs dramatically from todays prevalent method of building HVAC controllers using multiple single-input-single-output control loops. A simulation model of an experimental HVAC system is used in the design and simulation testing of controllers. While simulation can be insightful, the only way to truly verify the performance provided by different HVAC controller designs is by actually using them to control an HVAC system. Thus, an experimental HVAC system was built for testing a wide range of advanced HVAC controllers. The design and testing of MIMO robust controllers provides valuable insight into potential improvements in performance, as well as constraints, associated with applying this control methodology to HVAC systems. Test results on the physical system demonstrate achievable performance gains (reductions in discharge air temperature settle time) in excess of 300%. Furthermore, these performance gains may be achieved without significant impact to current HVAC system architecture (interconnection).

Synthesis of reinforcement learning, neural networks and PI control applied to a simulated heating coil

An accurate simulation of a heating coil is used to compare the performance of a proportional plus integral (PI) controller to the following schemes for learning improved control: a neural network trained to predict the steady-state output of the PI controller, a neutral network trained to minimize the n-step ahead error between the coil output and the set point, and a reinforcement learning agent trained to minimize the sum of the squared error over time. Although the PI controller works very well for this task, the neural networks showed improved performance. The reinforcement learning agent, when combined with a PI controller, learned to augment the PI control output for the subset of states for which control can be improved.

Optimization of autonomous village electrification systems by simulated annealing

Abstract The electrification of remote villages by autonomous renewable power systems is often more economical than the extension of a utility electrical grid. Accurate cost comparisons between the two alternatives, however, have historically been hindered by an inability to simulate a lowest cost autonomous system. In this article, a computational tool is presented which employs the techniques of combinatorial optimization to design a near-optimal autonomous power system for a given set of demand points. The optimum design of village electrification systems is approached as a two-level optimization problem. The upper level procedure attempts to design the optimum arrangement of transformers, which are connected to each other and to the electricity source with medium voltage wire. The lower level procedure attempts to design the optimum distribution grid for a specified transformer arrangement, with consideration of isolated sources. Optimization procedures based on the simulated annealing algorithm were developed for both the upper and lower level processes. An approximate lower level procedure based on the minimum spanning tree algorithm was also implemented in order to provide a reasonable starting point for the slower simulated annealing algorithm. Exhaustive enumeration algorithms for both the upper and lower level processes were developed in order to verify the accuracy of the other methods.

Robust Reinforcement Learning Control Using Integral Quadratic Constraints for Recurrent Neural Networks

The applicability of machine learning techniques for feedback control systems is limited by a lack of stability guarantees. Robust control theory offers a framework for analyzing the stability of feedback control loops, but for the integral quadratic constraint (IQC) framework used here, all components are required to be represented as linear, time-invariant systems plus uncertainties with, for IQCs used here, bounded gain. In this paper, the stability of a control loop including a recurrent neural network (NN) is analyzed by replacing the nonlinear and time-varying components of the NN with IQCs on their gain. As a result, a range of the NNs weights is found within which stability is guaranteed. An algorithm is demonstrated for training the recurrent NN using reinforcement learning and guaranteeing stability while learning.

Energy simulation in buildings : overview and BLAST example

Abstract Energy consumption to provide thermal comfort conditions, especially in hot and humid areas is tremendous. Ways to reduce and save energy are indeed possible and needed. Simulation of HVAC energy consumption in buildings is of considerable interest and benefit for engineers and architects. Energy simulation programs can be used to analyze cost effective energy conservation measures before the building is built or modified. There are both open type simulation programs and proprietary programs. The main components of an open type simulation program are presented, and its development is reviewed. Some of the difficulties encountered by first users of these programs are pinpointed with some suggested remedies. An annual weather data file for Jeddah, Saudi Arabia, was prepared. A three story, three zone typical school in Jeddah is considered. Building loads analysis and system thermodynamics input and output files are generated to show the capabilities of the program. Three case studies were considered.

MIMO robust control for heating, ventilating and air conditioning (HVAC) systems

Potential improvements in heating ventilation and air conditioning (HVAC) system performance are investigated through the application of multiple-input, multiple-output (MIMO) robust controllers. This approach differs dramatically from todays prevalent method of building HVAC controllers using multiple single-input, single-output control loops.

Modeling Phase Change Materials With Conduction Transfer Functions for Passive Solar Applications

The use of passive solar design in our homes and buildings is one way to offset the ever-increasing dependence on fossil fuels and the resulting pollution to our air, our land, and our waters. A well-designed sunroom has the potential to reduce the annual heating loads by one-third or more. By integrating phase change materials (PCMs) into building elements, such as floor tile and wallboard, the benefits of the sunroom can be further enhanced by providing enhanced energy storage. To maximize benefits from PCMs, an engineering analysis tool is needed to provide insight into the most efficient use of this developing technology. Thus far, modeling of the PCMs has been restricted to finite difference and finite element methods, which are not well suited to inclusion in a comprehensive annual building simulation program such as BLAST or EnergyPlus . Conduction transfer functions (CTFs) have long been used to predict transient heat conduction in such programs. Phase changes often do not occur at a single temperature, but do so over a range of temperatures. The phase change energy can be represented by an elevated heat capacity over the temperature range during which the phase change occurs. By calculating an extra set(s) of CTFs for the phase change properties, the CTF method can be extended to include the energy of phase transitions by switching between the two (or more) sets of CTFs. This method can be used to accurately predict the internal and external temperatures of PCM-containing building elements during transient heat conduction. The amount of energy storage and release during a phase transition can also be modeled with this method, although there may be some degree of inaccuracy due to switching between two or more sets of CTFs. CTFs have the potential to provide an efficient method of modeling PCMs in annual building simulation programs, but more work is needed to reduce errors associated with their use.

Exact Solution to the Governing PDE of a Hot Water-to-Air Finned Tube Cross-Flow Heat Exchanger

A new dynamic coil model is presented. This model is developed via the exact solution of a previously unsolved partial differential equation that governs the coil dynamics for a step change in water flow rate. This new model is the first step toward developing a future model that can accurately predict the coil dynamics for several varying coil inlet conditions expected to occur under MIMO control. The new model is compared with previously published simplified PDE coil models, which used an approximation to this exact solution, and against actual measured coil dynamics. The new coil model is shown to have superior performance in predicting the actual coil behavior.

Robust reinforcement learning control

Robust control theory is used to design stable controllers in the presence of uncertainties. By replacing nonlinear and time-varying aspects of a neural network with uncertainties, a robust reinforcement learning procedure results that is guaranteed to remain stable even as the neural network is being trained. The behavior of this procedure is demonstrated and analyzed on a simple control task. Reinforcement learning with and without robust constraints results in the same control performance, but at intermediate stages the system without robust constraints may go through a period of unstable behavior that is avoided when the robust constraints are included.

Dynamic Response and Tuning

HVAC system control systems can be especially difficult to adjust or tune because the process gain of the control loop varies with the control point and with a range of other variables. For example, if the air flow rate over a hot-water heating coil is cut in half (as in a multizone application), the gain of the coil (the ratio of the change in outlet temperature to the change in valve position) will increase significantly.