Search

Recent studies have demonstrated the potential of flexible loads in providing frequency response services. However, uncertainty and variability in various weather-related and end-use behavioral factors often affect the demand-side control performance. This work addresses this problem with the design of a demand-side control to achieve frequency response under load uncertainties. Our approach involves modeling the load uncertainties via stochastic processes that appear as both multiplicative and additive to the system states in closed-loop power system dynamics. Extending the recently developed mean square exponential stability (MSES) results for stochastic systems, we formulate multi-objective linear matrix inequality (LMI)-based optimal control synthesis problems to not only guarantee stochastic stability, but also promote sparsity, enhance closed-loop transient performance, and maximize allowable uncertainties. The fundamental trade-off between the maximum allowable (\textit{critical}) uncertainty levels and the optimal stochastic stabilizing control efforts is established. Moreover, the sparse control synthesis problem is generalized to the realistic power systems scenario in which only partial-state measurements are available. Detailed numerical studies are carried out on IEEE 39-bus system to demonstrate the closed-loop stochastic stabilizing performance of the sparse controllers in enhancing frequency response under load uncertainties; as well as illustrate the fundamental trade-off between the allowable uncertainties and optimal control efforts., accepted for publication at the IEEE Transactions on Power Sysems

Increasing connectivity to the Internet for remote monitoring and control has made cyber-physical systems more vulnerable to deliberate attacks; purely cyber attacks can thereby have physical consequences. Long-term, stealthy attacks such as Stuxnet can be described as Advanced Persistent Threats (APTs). Here, we extend our previous work on hypergames and APTs to develop hypergame-based defender strategies that are robust to deception and do not rely on attack detection. These strategies provide provable bounds—and provably optimal bounds—on the attacker payoff. Strategies based on Bayesian priors do not provide such bounds. We then numerically demonstrate our approach on a building control subsystem and discuss next steps in extending this approach toward an operational capability.

As buildings account for approximately 40% of global energy consumption and associated greenhouse gas emissions, their role in decarbonizing the power grid is crucial. The increased integration of variable energy sources, such as renewables, introduces uncertainties and unprecedented flexibilities, necessitating buildings to adapt their energy demand to enhance grid resiliency. Consequently, buildings must transition from passive energy consumers to active grid assets, providing demand flexibility and energy elasticity while maintaining occupant comfort and health. This fundamental shift demands advanced optimal control methods to manage escalating energy demand and avert power outages. Reinforcement learning (RL) emerges as a promising method to address these challenges. In this paper, we explore ten questions related to the application of RL in buildings, specifically targeting flexible energy management. We consider the growing availability of data, advancements in machine learning algorithms, open-source tools, and the practical deployment aspects associated with software and hardware requirements. Our objective is to deliver a comprehensive introduction to RL, present an overview of existing research and accomplishments, underscore the challenges and opportunities, and propose potential future research directions to expedite the adoption of RL for building energy management.

We present a differentiable predictive control (DPC) methodology for learning constrained control laws for unknown nonlinear systems. DPC poses an approximate solution to multiparametric programming problems emerging from explicit nonlinear model predictive control (MPC). Contrary to approximate MPC, DPC does not require supervision by an expert controller. Instead, a system dynamics model is learned from the observed system’s dynamics, and the neural control law is optimized offline by leveraging the differentiable closed-loop system model. The combination of a differentiable closed-loop system and penalty methods for constraint handling of system outputs and inputs allows us to optimize the control law’s parameters directly by backpropagating economic MPC loss through the learned system model. The control performance of the proposed DPC method is demonstrated in simulation using learned model of multi-zone building thermal dynamics.

In recent years data-driven analysis of dynamical systems has attracted a lot of attention and transfer operator techniques, namely, Perron-Frobenius and Koopman operators are being used almost ubiquitously. Since data is always obtained in discrete-time, in this paper, we propose a purely data-driven approach for the design of a stabilizing feedback control law for a general class of discrete-time control-affine non-linear systems. In particular, we use the Koopman operator to lift a control-affine system to a higher-dimensional space, where the control system's evolution is bilinear. We analyze the controllability of the lifted bilinear system and relate it to the controllability of the underlying non-linear system. We then leverage the concept of Control Lyapunov Function (CLF) to design a state feedback law that stabilizes the origin. Furthermore, we demonstrate the efficacy of the proposed method to stabilize the origin of the Van der Pol oscillator and the chaotic Henon map from the time-series data.

The dynamics-aware economic dispatch (DED) problem embeds low-level generator dynamics and operational constraints to enable near real-time scheduling of generation units in a power network. DED produces a more dynamic supervisory control policy than traditional economic dispatch (T-ED) that leads to reduced overall generation costs. However, the incorporation of differential equations that govern the system dynamics makes DED an optimization problem that is computationally prohibitive to solve. In this work, we present a new data-driven approach based on differentiable programming to efficiently obtain parametric solutions to the underlying DED problem. In particular, we employ the recently proposed differentiable predictive control (DPC) for offline learning of explicit neural control policies using an identified Koopman operator (KO) model of the power system dynamics. We demonstrate the high solution quality and five orders of magnitude computational-time savings of the DPC method over the original online optimization-based DED approach on a 9-bus test power grid network., The code for producing this work is available in the repo: https://github.com/pnnl/neuromancer/tree/DED_DPC

Cyber-physical systems are vulnerable to cyber attacks that can produce serious physical consequences. Our previous work showed how hypergame analysis, an extension of game theory for situations with information asymmetries and player misperceptions, could be applied to control systems subject to deception-based attacks. Here, we build on that research to study a repeated, stochastic context. We consider detection of an attacker attempting to manipulate the control system while remaining undetected. We discuss different monitoring approaches that can be used to do this and define a learning scheme for the defender. In our numerical experiments, we find that the attacker impact and time to detection depend strongly on the cost incurred by the defender in removing an attacker from the system. We also show that the defender learning scheme enforces a strong tradeoff, for the attacker, between remaining undetected, and having an impact on the system.

The objective of traditional phasor measurement unit (PMU) placement approaches is to find the minimal number of PMUs to achieve full system observability. Very few consider the probability and the consequence severity of observability loss. For example, following cascading failures, lines are tripped, and the topology of the grid changes, which may result in observability loss for parts of the grid. With limited number of PMUs, it is desired that the PMUs should still sustain observability for the power grid after top-severe cascading failures, other than after those cascading failures of low severity. To address the above issues related with the consequence severity of observability loss, this study proposes a risk-oriented PMU placement approach. The proposed approach uses two indices, namely the probability of observability loss and the severity of its corresponding consequence to design the optimal PMU placement strategy. The authors show that the proposed approach is able to capture the critical events, while traditional approaches fail to do so. Extensive simulation results carried out on the IEEE 39-bus and 118-bus systems demonstrate the effectiveness of the proposed risk-oriented approach.

Building operations rely heavily on control systems and sensors. This paper provides a sophisticated literature review on sensor systems in building/HVAC systems, particularly in the context of controls, and their impacts on energy efficiency and thermal comfort. This study aims to understand the previous and current research and identify future research opportunities on this subject. The reviewed sensor systems were used for heuristic rule-based controls, local controls, and advanced optimal controls for existing and emerging technologies. Five major aspects of sensors are reviewed here: control loops for sensors, sensor types, sensor locations, sensor data, and a sensor impact evaluation framework. To augment the literature review, comprehensive standardized interviews were also conducted with relevant industry experts and practitioners. These interviews were designed and performed to (1) identify significant factors for selecting sensor sets and current undergoing issues, (2) identify potential improvements in sensor configuration/deployment, and (3) integrate expert (e.g., researcher, building operation practitioner) knowledge and experiences to develop structured use-case scenarios. Researchers collected and analyzed 31 interview responses for this paper.

Sequential approaches to system and control design produce sub-optimal solutions due to unidirectional coupling between the system and control variables, i.e., the system design prescribes the control approach but not vice versa. A critical challenge to co-design is the complex coupling between heterogeneous variables: the time-independent system design variables and the time-varying control parameters. Traditional optimization-based co-design methods are unsuitable for large-scale engineering systems where the plant design is complex and accurate control-oriented, white-box models may not be available. These challenges have led to the widespread adoption of industry guidelines that often oversize commercial HVAC systems by 25–50% to satisfy conservative estimates of building peak-loads and redundancy margins. Significant energy, peak-demand, and capital cost savings can be realized with properly sized HVAC units and optimized control operation. In this paper, we develop a data-driven, black-box, co-design framework using Bayesian Optimization (BO) to co-optimize the system and control design variables for commercial building HVAC system. Using real chiller data, a detailed economic assessment and numerical study is conducted to illustrate the energy, peakload and capital cost savings against existing baseline strategies for chiller sizing and plant configuration.

For effective integration of building operations into the evolving demand response programs of the power grid, real-time decisions concerning the use of building appliances for grid services must excel on multiple criteria, ranging from the added value to occupants' comfort to the quality of the grid services. In this paper, we present a data-driven decision-support framework to dynamically rank load control alternatives in a commercial building, addressing the needs of multiple decision criteria (e.g. occupant comfort, grid service quality) under uncertainties in occupancy patterns. We adopt a stochastic multi-criteria decision algorithm recently applied to prioritize residential on/off loads, and extend it to i) complex load control decisions (e.g. dimming of lights, changing zone temperature set-points) in a commercial building; and ii) systematic integration of zonal occupancy patterns to accurately identify short-term grid service opportunities. We evaluate the performance of the framework for curtailment of air-conditioning, lighting, and plug-loads in a multi-zone office building for a range of design choices. With the help of a prototype system that integrates an interactive \textit{Data Analytics and Visualization} frontend, we demonstrate a way for the building operators to monitor the flexibility in energy consumption and to develop trust in the decision recommendations by interpreting the rationale behind the ranking., Comment: accepted in the 2021 American Control Conference

Increasing deployment of advanced sensing, controls, and communication infrastructure enables buildings to provide services to the power grid, leading to the concept of grid-interactive efficient buildings. Since occupant activities and preferences primarily drive the availability and operational flexibility of building devices, there is a critical need to develop occupant-centric approaches that prioritize devices for providing grid services, while maintaining the desired end-use quality of service. In this paper, we present a decision-making framework that facilitates a building owner/operator to effectively prioritize loads for curtailment service under uncertainties, while minimizing any adverse impact on the occupants. The proposed framework uses a stochastic (Markov) model to represent the probabilistic behavior of device usage from power consumption data, and a load prioritization algorithm that dynamically ranks building loads using a stochastic multi-criteria decision-making algorithm. The proposed load prioritization framework is illustrated via numerical simulations in a residential building use-case, including plug-loads, air-conditioners, and plug-in electric vehicle chargers, in the context of load curtailment as a grid service. Suitable metrics are proposed to evaluate the closed-loop performance of the proposed prioritization algorithm under various scenarios and design choices. Scalability of the proposed algorithm is established via computational analysis, while time-series plots are used for intuitive explanation of the ranking choices.

Neural network modules conditioned by known priors can be effectively trained and combined to represent systems with nonlinear dynamics. This work explores a novel formulation for data-efficient learning of deep control-oriented nonlinear dynamical models by embedding local model structure and constraints. The proposed method consists of neural network blocks that represent input, state, and output dynamics with constraints placed on the network weights and system variables. For handling partially observable dynamical systems, we utilize a state observer neural network to estimate the states of the system's latent dynamics. We evaluate the performance of the proposed architecture and training methods on system identification tasks for three nonlinear systems: a continuous stirred tank reactor, a two tank interacting system, and an aerodynamics body. Models optimized with a few thousand system state observations accurately represent system dynamics in open loop simulation over thousands of time steps from a single set of initial conditions. Experimental results demonstrate an order of magnitude reduction in open-loop simulation mean squared error for our constrained, block-structured neural models when compared to traditional unstructured and unconstrained neural network models., Comment: 10 pages. Submitted to American Control Conference ACC 2020. Under review

In this paper, we provide sufficient conditions for dissipativity and local asymptotic stability of discrete-time dynamical systems parametrized by deep neural networks. We leverage the representation of neural networks as pointwise affine maps, thus exposing their local linear operators and making them accessible to classical system analytic and design methods. This allows us to "crack open the black box" of the neural dynamical system's behavior by evaluating their dissipativity, and estimating their stationary points and state-space partitioning. We relate the norms of these local linear operators to the energy stored in the dissipative system with supply rates represented by their aggregate bias terms. Empirically, we analyze the variance in dynamical behavior and eigenvalue spectra of these local linear operators with varying weight factorizations, activation functions, bias terms, and depths., Under review at IEEE Open Journal of Control Systems

This paper presents a methodology for enhancing community resilience through optimal renewable resource allocation and load scheduling in order to minimize unserved load and thermal discomfort. The proposed control architecture distributes the computational effort and is easier to be scaled up than traditional centralized control. The decentralized control architecture consists of two layers: The community operator layer (COL) allocates the limited amount of renewable energy resource according to the power flexibility of each building. The building agent layer (BAL) addresses the optimal load scheduling problem for each building with the allowable load determined by the COL. Both layers are formulated as a model predictive control (MPC) based optimization. Simulation scenarios are designed to compare different combinations of building weighting methods and objective functions to provide guidance for real-world deployment by community and microgrid operators. The results indicate that the impact of power flexibility is more prominent than the weighting factor to the resource allocation process. Allocation based purely on occupancy status could lead to an increase of PV curtailment. Further, it is necessary for the building agent to have multi-objective optimization to minimize unserved load ratio and maximize comfort simultaneously.

Advanced control strategies are becoming increasingly necessary in buildings in order to meet and balance requirements for energy efficiency, demand flexibility, and occupant comfort. Additional development and widespread adoption of emerging control strategies, however, ultimately require low implementation costs to reduce payback period and verified performance to gain control vendor, building owner, and operator trust. This is difficult in an already first-cost driven and risk-averse industry. Recent innovations in building simulation can significantly aid in meeting these requirements and spurring innovation at early stages of development by evaluating performance, comparing state-of-the-art to new strategies, providing installation experience, and testing controller implementations. This paper presents the development of a simulation framework consisting of test cases and software platform for the testing of advanced control strategies (BOPTEST - Building Optimization Performance Test). The objectives and requirements of the framework, components of a test case, and proposed software platform architecture are described, and the framework is demonstrated with a prototype implementation and example test case.

Data-driven control strategies suffer from the necessity of training and its associated computational time and cost. This motivates the continuous increase of popularity of transfer learning based methods. In this paper, we have investigated transfer learning in the context of controller for dynamical systems. First, we have derived conditions under which controller can be transferred between two linear dynamical systems. Furthermore, we have introduced a novel algorithm to design the transferable and non-transferable control components for linear systems. Secondly, we have derived conditions under which a stabilizing controller can be transferred between two different nonlinear dynamical systems. Furthermore, we have used the notion of zero dynamics to transform the nonlinear dynamics into global normal form, to derive the transferability conditions on the controller. Finally, we have numerically evaluated the performance of our transfer learning based methods for two sets of example linear and nonlinear dynamical systems and consequently shows the benefit of transferring a pre-trained controller from a Source to Target system in the context of computational complexity and cost benefit.

Author(s): Drgoňa, J; Arroyo, J; Cupeiro Figueroa, I; Blum, D; Arendt, K; Kim, D; Olle, EP; Oravec, J; Wetter, M; Vrabie, DL; Helsen, L | Abstract: It has been proven that advanced building control, like model predictive control (MPC), can notably reduce the energy use and mitigate greenhouse gas emissions. However, despite intensive research efforts, the practical applications are still in the early stages. There is a growing need for multidisciplinary education on advanced control methods in the built environment to be accessible for a broad range of researchers and practitioners with different engineering backgrounds. This paper provides a unified framework for model predictive building control technology with focus on the real-world applications. From a theoretical point of view, this paper presents an overview of MPC formulations for building control, modeling paradigms and model types, together with algorithms necessary for real-life implementation. The paper categorizes the most notable MPC problem classes, links them with corresponding solution techniques, and provides an overview of methods for mitigation of the uncertainties for increased performance and robustness of MPC. From a practical point of view, this paper delivers an elaborate classification of the most important modeling, co-simulation, optimal control design, and optimization techniques, tools, and solvers suitable to tackle the MPC problems in the context of building climate control. On top of this, the paper presents the essential components of a practical implementation of MPC such as different control architectures and nuances of communication infrastructures within supervisory control and data acquisition (SCADA) systems. The paper draws practical guidelines with a generic workflow for implementation of MPC in real buildings aimed for contemporary adopters of this technology. Finally, the importance of standardized performance assessment and methodology for comparison of different building control algorithms is discussed.

A fault identification and estimation approach that brings together the differential geometric concept of observability codistribution with data-driven concurrent learning, is presented in this paper. In order to identify faults in the presence of unknown disturbances, we use the differential geometric approach to design a coordinate transformation, in order to find a subspace in which the effects of disturbances and system faults can be segregated. We then use concurrent learning to estimate the magnitude of the constant fault. We illustrate the approach to fault isolation for a spherical pendulum dynamics. We use Lyapunov stability analysis to show that the fault estimate by concurrent learning converges to the actual fault value, and then illustrate the design of a recovery controller.

Net-Zero Energy Communities (NZECs) are critical to assuring the sustainability and resilience of modernized power systems. System modeling helps overcome technical challenges in designing and operating NZECs. In this paper, we present an open-source NZEC virtual testbed in Modelica based on a real NZEC in Florida, USA. This testbed consists of two sets of models: (1) higher-fidelity physics-based models that consider the interaction between subsystems of the studied NZEC and capture fast dynamics, and (2) lower-fidelity data-driven models that require fewer resources to establish and/or run. All models are validated against measurements from this real NZEC. In addition, this testbed includes a simulation framework that streamlines the processes for simulation and thus allows the use of developed models to form a virtual testbed. To demonstrate the usage of the virtual testbed, a case study is conducted where a building-to-grid integration control is evaluated via simulation. The evaluation results suggest that the tested control significantly smooths the power draw of the studied community and does not sacrifice thermal comfort to a great extent.

The load scheduling of resilient communities in the islanded mode is subject to many uncertainties such as weather forecast errors and occupant behavior stochasticity. To date, it remains unclear how occupant preferences affect the effectiveness of the load scheduling of resilient communities. This paper proposes an occupant preference-aware load scheduler for resilient communities operating in the islanded mode. The load scheduling framework is formulated as a model predictive control problem. Based on this framework, a deterministic load scheduler is adopted as the baseline. Then, a chance-constrained scheduler is proposed to address the occupant-induced uncertainty in room temperature setpoints. Key resilience indicators are selected to quantify the impacts of the uncertainties on community load scheduling. Finally, the proposed preference-aware scheduler is compared with the deterministic scheduler on a virtual testbed based on a real-world net-zero energy community in Florida, USA. Results show that the proposed scheduler performs better in terms of serving the occupants’ thermal preference and reducing the required battery size, given the presence of the assumed stochastic occupant behavior. This work indicates that it is necessary to consider the stochasticity of occupant behavior when designing optimal load schedulers for resilient communities.

Dynamic modelling of the faulty operation of chiller plants and boiler plants can help identify their impacts and support the development of fault detection methods. However, adequate models are seldom reported in the literature. In this study, we aim to develop high-fidelity models for approximating the dynamic behaviors of chiller plants and boiler plants under control-related faults. Specifically, we first designed a typical configuration of the chiller plants and the boiler plants; we then modeled both the physical systems and controllers of these typical plants with Modelica. When developing the Modelica models, we created a hierarchical model structure while modules in each layer can be redeclared and parameterized at upper layers. This model structure facilitates the implementation of fault scenarios through intuitive model modifiers. At last, we applied the proposed models in a comprehensive fault impact evaluation of the thirteen control-related faults of chiller and boiler plants. In this evaluation, the proposed models are coupled with the EnergyPlus TM thermal load model to study the impact of various faulty scenarios. Based on the fault impact evaluation results, we identified the faults that have the most significant impacts on the operation of the chiller and boiler plants, respectively. We also found that the relationship between the impacts of the studied faults and the severity level of the faults can be highly non-linear. This study contributes to the literature by providing the first dynamic models of chiller plants and boiler plants which can be used to study control-related faults on a large-scale.

The current HVAC design practice is sequential or iterative at best. This conservative approach does not consider the increased potential enabled by advanced sensing and control technologies that may be installed at a later date. Selection and optimization of the building control system is mostly an independent step in a sequential design process. The optimal design and operation of HVAC systems must account for the interconnected controls between the subsystems. In this paper, we develop a data-driven, simulation-based, black-box optimization approach based on Bayesian optimization (BO) to efficiently explore the design space and jointly optimize both the system and control-design parameters of a commercial building chiller plant. The co-design optimization determines the optimal number/configuration and size of the chillers, and the optimal chiller sequencing control variables that minimize the overall energy consumption, peak-load, operating, and capital costs incurred over a design horizon, subject to cooling load constraints. The problem is formulated as a mixed-integer programming model and solved using BO that leverages a high-fidelity commercial chiller plant emulator to evaluate different candidate designs. Based on real chiller data, we conducted a detailed economic assessment and numerical study that highlighted that, compared to current design practices, the co-optimization approach resulted in capital cost savings of about 0.7 million US dollars, annual energy savings of nearly 33%, and a 56% reduction in peak power demand while delivering the same level of cooling to the building due to the chiller’s optimized sizing and operation (switching thresholds, chiller staging).

Sensors are the key information source for fault detection and diagnostics (FDD) in buildings. However, sensors are often not properly designed, installed, calibrated, located, and maintained, which negatively impacts FDD performance. Several sensor-related FDD topics have been widely studied, covering a wide range of fault types and applications. However, it is difficult to get a clear picture of the technical development of sensor-related topics in FDD. A systematic review of sensor topics is needed to summarize the existing research in a logical way, draw conclusions on the current development, and predict the future development of sensors in building FDD. To address this gap, we conducted a comprehensive literature review of more than 100 FDD-sensor-related papers. In this article, we subdivide the FDD tasks into building-level, system-level, and component-level FDD, and review sensor-related topics in each category. Our major conclusions are: (a) current data-driven FDD research focuses more on FDD algorithms than sensors, (b) sensor “hardware” research topics are less studied than sensor “software” topics, (c) very few papers focus on sensor engineering as an integral aspect of FDD development, and (d) some important sensor topics, such as sensor cost-effectiveness and sensor schema/layout/location, are not well studied. Finally, we discuss the need for a systematic framework of FDD sensors and models to integrate sensor design/selection, sensor data analysis/mining, feature selection, physics-based or data-driven algorithm development, sensor fault detection, sensor calibration, and sensor maintenance. Finally, expert interviews are conducted to validate the above findings and conclusions.

We present a physics-constrained deep learning method to develop control-oriented models of building thermal dynamics. The proposed method uses systematic encoding of physics-based prior knowledge into a structured recurrent neural architecture. Specifically, our method incorporates structural prior knowledge from traditional physics-based building modeling into the architecture of the deep neural network model. Further, we also use penalty methods to provide inequality constraints, thereby bounding predictions within physically realistic and safe operating ranges. We observe that stable eigenvalues accurately characterize the dissipativeness of the system, and use a constrained matrix parameterization based on the Perron-Frobenius theorem to bound the dominant eigenvalues of the building thermal model parameter matrices. We demonstrate the effectiveness and physical interpretability of the proposed data-driven modeling approach on a real-world dataset obtained from an office building with 20 thermal zones. The proposed data-driven method can learn interpretable dynamical models that achieve high accuracy and generalization over long-term prediction horizons. We show that using only 10 days’ measurements for training, our method is capable of generalizing over 20 consecutive days. We demonstrate that the proposed modeling methodology is achieving state-of-the-art performance by significantly improving the accuracy and generalization compared to classical system identification methods and prior advanced methods reported in the literature. compared to prior state-of-the-art methods reported in the literature.

Deception techniques are a useful defensive mechanism that could be very valuable in control systems where updates and patches can be difficult. However, to deceive a motivated and targeted attacker it is necessary to enhance deception platforms with capabilities to model and simulate physical processes to achieve necessary fidelity. In this paper, attributes of cyber-physical decoys are discussed and a design of a system with these attributes is presented. A boiler model use case is provided to demonstrate how this novel deception capability can be integrated into and used to defend a system.

In this paper, we present a novel sequential framework containing two deep network architectures for baseline energy prediction of a building. The proposed framework utilizes convolution layers to extract features from the input data space without changing spatial relations between variables. These features are memorized by tensor train based gated recurrent units for an accurate long-term prediction. We observe that this sequential framework helps to improve long term prediction accuracy, thereby mitigating prediction error accumulation over time. Although architectures comprising the amalgamation of convolution layers and memory cell have shown promising results in domains such as social media analysis, language modeling, video frame prediction, and image recognition, this paper extends its scope to the context of energy applications. Furthermore, the addition of tensor train based gated recurrent units is motivated by the necessity of computational time reduction during the training process, thereby making this framework suitable for field deployment. Results on a commercial building dataset show that the current framework outperforms other existing machine learning based methods, in both short-term and long-term prediction categories.

There is increasing interest in designing optimization-based techniques for the control of building heating, ventilation, and air-conditioning (HVAC) systems for either improving the energy efficiency of buildings or providing ancillary services to the electric grid. The performance of such prediction-based control techniques relies heavily on models of a building's thermal dynamics. However, the development of high-fidelity building thermal dynamic models is challenging, given the presence of large uncertainties that affect thermal loads in buildings, such as building envelope performance, thermal mass, internal heat gains, and occupant behavior. In this paper, we propose a method to identify both a resistive-capacitive parametric model and nonparametric load uncertainties using measured input-output data. The parametric model is obtained using semiparametric regression, whereas the nonparametric terms are based on the Random Forest algorithm in which regression trees are used to derive the dependency of nonparametric terms on both building operation parameters and ambient temperature. The effectiveness of the method is evaluated using experimental data collected from an office building at the Pacific Northwest National Laboratory (PNNL) campus. The proposed methodology was observed to provide improved accuracy over appropriate baseline strategies in predicting indoor air temperatures.