PaperlayerView exact revision on arXiv

Source-linked AI summary

Data-driven Predictive Control for Unlocking Building Energy Flexibility: A Review

Anjukan Kathirgamanathan, Mattia De Rosa, Eleni Mangina, Donal P. Finn

arXiv:2007.14866v1eess.SY

TL;DR

Variable renewable generation and heterogeneous building thermal dynamics make scalable grid integration difficult. This review analyses 115 data-driven predictive-control implementations, focusing on model development, control integration, passive thermal mass, and feature selection. It finds data-driven predictive control promising for grid-interactive buildings, while identifying feature selection and data quality as continuing research gaps.

  • Problem

    Developing computationally tractable control-oriented models for heterogeneous, nonlinear building thermal dynamics remains a major hurdle to unlocking building energy flexibility.

  • Method

    The review categorises 115 data-driven predictive-control studies for building energy management and demand-side management, with emphasis on model–control integration and feature selection.

  • Results

    Data-driven predictive control is reported as promising for grid-interactive buildings, with meaningful grid integration and carbon-footprint reduction at the single-building level.

  • Takeaways & Limitations

    Model Predictive Control and its robust and stochastic variants dominate research interest, while Reinforcement Learning has gained interest in recent years.

  • Takeaways & Limitations

    The review identifies unresolved questions about robust feature selection and the influence of data quality on predictive-control performance and harnessed flexibility.

Abstract

from arXiv · show

Managing supply and demand in the electricity grid is becoming more challenging due to the increasing penetration of variable renewable energy sources. As significant end-use consumers, and through better grid integration, buildings are expected to play an expanding role in the future smart grid. Predictive control allows buildings to better harness available energy flexibility from the building passive thermal mass. However, due to the heterogeneous nature of the building stock, developing computationally tractable control-oriented models, which adequately represent the complex and nonlinear thermal-dynamics of individual buildings, is proving to be a major hurdle. Data-driven predictive control, coupled with the "Internet of Things", holds the promise for a scalable and transferrable approach,with data-driven models replacing traditional physics-based models. This review examines recent work utilising data-driven predictive control for demand side management application with a special focus on the nexus of model development and control integration, which to date, previous reviews have not addressed. Further topics examined include the practical requirements for harnessing passive thermal mass and the issue of feature selection. Current research gaps are outlined and future research pathways are suggested to identify the most promising data-driven predictive control techniques for grid integration of buildings.

1. Introduction

Variable renewable supply increases the need to balance electricity demand, making buildings important flexibility resources through their thermal mass. The review examines scalable data-driven predictive-control approaches while focusing on model–control integration and feature selection.

  • Motivation: Buildings represent about 40% of Europe’s total primary energy consumption and can provide thermal energy storage for demand-side management.Commercial buildings are especially relevant because of their thermal mass and HVAC loads.
  • Motivation: Aggregating flexibility across unique buildings requires scalable and transferable methods for assessing and harnessing that resource.
  • Control challenge: Predictive control can account for future grid or weather disturbances and harness building thermal storage, unlike standard rule-based control.Rule-based control remains common because of its simplicity and dependence on operator expertise.
  • Data-driven control: Data-driven approaches retain predictive control capabilities without first-principles model generation, but nonlinear black-box models can increase computational complexity.Researchers have used variable separation and branch-and-bound techniques to mitigate these issues.
  • Research gaps: The review addresses gaps in systematic technique evaluation, practical passive-thermal-mass modelling, and input-feature selection for building demand-side management.
  • Contributions: Its contributions include categorising approaches, qualitatively analysing model–control coupling, and assessing feature-selection techniques for ubiquitous building data.

2. Background

Building energy management must handle nonlinear dynamics, disturbances, constraints, interacting control loops, and external grid signals. The background compares modelling frameworks and control strategies used to address these challenges.

  • Energy flexibility: Building energy flexibility encompasses demand and generation potential together with communication or coordination with grid or aggregator needs.The review covers both power and energy flexibility products.
  • Building modelling frameworks: White-box models require detailed physical knowledge and can impose substantial computational overhead that limits online building control.
  • Building modelling frameworks: Black-box models are fully empirical and require large, rich datasets spanning seasons and operational envelopes.
  • Building modelling frameworks: Grey-box models combine physics-based and empirical elements, with parameters identified from experimental data and interpretable physical meaning.The review notes their suitability for buildings that are not overly complex.
  • Control frameworks: MPC repeatedly optimises a constrained finite-horizon cost function and can use predicted states to shift consumption through thermal or electrical storage.
  • Control frameworks: Data-driven predictive control must preserve tractable optimisation while representing operating ranges absent from training data.Comparisons with classical MPC remain rare for large complex buildings.

3. Methods

The review screened studies using data-driven predictive control for building energy management and demand-side management, extending the scope to related predictive-control applications. It selected 115 studies and excluded model-accuracy and uncertainty reviews from detailed scope.

  • Selection criteria: The review covers studies published from 2010 through 01/11/19 that use forecasted variables in data-driven predictive control.
  • Selection criteria: Eligible work used building sensor data, whether measured or synthetic and simulated, and addressed defined energy-management applications.
  • Selection criteria: Applications included demand-side management, thermal comfort, MPC development or alternatives, and co-simulation.
  • Scope extension: The scope included predictive-control studies on thermal comfort, energy efficiency, and economic management because changing the objective function can support energy shifting.
  • Review process: 115 studies were selected and categorised according to the review’s points of consideration.
  • Exclusions: Model prediction accuracy and model uncertainty were excluded because existing reviews addressed them and quantitative accuracy comparisons vary with test periods, climates, and buildings.

4. Data-driven Predictive Control Approaches

The review categorises data-driven predictive-control studies by application and reports substantial use in demand-side management. Energy efficiency, cost savings, and alternatives to traditional model-based MPC are also common themes.

  • Review summary: The reviewed studies are categorised by application, building model, control technique, and energy-flexibility scope.
  • Review summary: Data-driven predictive control has been implemented in a significant number of demand-side-management studies.The review identifies related applications in energy efficiency and cost savings.

4.3. Building Type

Commercial and single-zone buildings dominate predictive-control studies, while most models use reduced-order or increasingly black-box approaches. Research commonly simplifies multi-zone buildings and rarely evaluates more than one building.

  • Commercial buildings receive significantly more predictive-control attention than residential buildings.Commercial buildings offer more training data through BEMS, more notable predictive-control benefits, and less pronounced occupancy impacts.
  • Single-zone buildings receive substantial attention, while multi-zone buildings introduce modelling challenges such as thermal interactions between zones.Many studies approximate multi-zone buildings as single zones, for example by averaging zonal temperatures.
  • Most reviewed studies employ reduced-order models, with state-space implementations the most common model type.State-space models are well established and widely used in MPC applications, particularly in the process industry.
  • Black-box building models, including tree-based methods, neural networks, Gaussian processes, and deep learning, have increased in recent years.These models are generally nonlinear and implicitly relate system inputs to outputs, requiring integration techniques for optimal control.
  • Most studies do not quantify model predictive power across the prediction horizon, leaving its effect on predictive-control performance unclear.Model pruning is also essential for some black-box models, including neural networks and tree-based methods, to reduce overfitting.

4.5. Control Approaches

MPC dominates building control research, while black-box MPC and reinforcement learning have gained attention more recently. Reviewed studies mainly optimize energy, cost, and comfort objectives using varied control formulations.

  • Control Approaches: MPC dominates the reviewed control approaches, while black-box MPC and reinforcement learning show increasing interest in recent years.Other control techniques have comparatively small sample sizes.
  • Control Approaches: Data-driven approaches trained on MPC input-output data can reduce computational and memory requirements and support deployment on low-level hardware.Their implementation still requires an original MPC model and controller to generate training data.
  • Optimization Objectives: Energy-demand minimization is the most common optimization objective, followed by cost and discomfort minimization.A cost objective requires time-varying price signals, such as a day-ahead price schedule or real-time pricing forecast.
  • Optimization Objectives: Comfort is predominantly represented by deviations of internal temperatures from setpoints or allowable temperature bands.Few studies use PMV, which describes user thermal comfort more directly but increases modelling and computational requirements.
  • Optimization Objectives: Quadratic comfort objectives commonly represent temperature-tracking problems, while linear functions are otherwise used and weighted for multi-objective problems.The quadratic form supports stability and reduced computational effort in optimization.
  • Optimization Objectives: Less common objectives include controller stability, PV curtailment, peak loads, CO2 emissions, exergy destruction, primary-energy factor, and reserve provision.This range reflects MPC's accommodation of differing stakeholder requirements.

4.7. Energy Flexibility Resources Considered

All reviewed DSM studies use building passive thermal mass as an energy-flexibility resource. Active storage, electric vehicles, and on-site generation are considered much less often.

  • All reviewed DSM applications using predictive control utilize building passive thermal mass.Passive thermal mass receives attention because of its energy-flexibility potential and the challenge of modelling building thermal dynamics.
  • Considerably fewer studies consider active thermal or electrical storage systems, including batteries.
  • Even fewer studies include electric vehicles and on-site generation as flexibility resources.
  • Only one reviewed study considers all listed flexibility resources and coordinates 15,000 residential buildings using grid and nodal pricing.

4.8. Feature Selection

The review finds limited formal feature-selection practice despite abundant building data, while forecast inputs commonly include weather, grid signals, and occupancy.

  • Feature Selection: Only six reviewed studies employed a formalised feature-selection technique in their data-driven model-development methodology.Approaches included manual feature elimination, principal component analysis, rules-based selection, and Gaussian-process relevance determination.
  • Predictive Inputs: 93% of studies included future external-weather predictions as inputs to the predictive controller.Exceptions used active thermal storage, constant ambient conditions, or reinforcement-learning approaches without predictive weather forecasts.
  • Grid Signals: Approximately 48% of studies incorporated grid signals into their control framework, with static time-of-use pricing the most common pricing signal.Static time-of-use prices do not require a forecast because rates and times are fixed in advance.
  • Occupancy: 66% of studies forecast or used occupancy as a disturbance, although constant schedules are less justifiable for stochastic residential occupancy.Commercial occupancy was often treated as more predictable because of working-hour patterns.

4.10. Validation of Approaches

Validation is dominated by surrogate simulations, while real-building deployment introduces costly integration and makes matched savings comparisons difficult.

  • Validation Settings: Most controllers were tested only on virtual surrogate simulation testbeds, commonly using physics-based white-box models.Few studies implemented predictive control in experimental settings.
  • Real-Building Testing: Integrating predictive controllers with a building automation system can be time-consuming and costly.Real-building testing also makes it difficult to quantify savings because reference control cannot usually be tested under exactly the same conditions.
  • Quantitative Performance: An average saving of 23% in cost or energy consumption was reported against rule-based-control baselines across quantified sampled studies.The summary covers 10 relevant papers, but differing buildings, disturbances, and time periods prevent meaningful direct comparison of techniques.
  • Quantitative Performance: The review presents quantitative performance comparisons in Table 3 for a sample of studies implementing data-driven predictive control.The table summarizes results from 10 relevant papers.

5. Discussion

The review identifies weak feature-selection practice, nonlinear model-integration challenges, simplified building representations, limited general-purpose environments, and narrow coverage of flexibility resources.

  • Feature Selection: Only 5% of studies considered feature selection during model development, with no standard methodology for feature selection and engineering.The review questions whether such methods are suitable and robust for large complex buildings with many sensors and features.
  • Feature Selection: Feature selection can reduce controller complexity and implementation cost while potentially improving data-driven model predictive accuracy.The review cites this as a benefit demonstrated by a simple approach in one study.
  • Model-Control Integration: Integrating highly nonlinear black-box models with MPC creates optimisation challenges, requiring either linearisation approaches or nonlinear-capable optimisers.The reviewed approaches differ in computational benefits and costs.
  • Scalability and Generalisability: Many studies simplify multi-zone buildings as single-zone models, while few implement predictive control across building clusters or groups.Different zone occupancy patterns and energy gains can produce distinct thermal behaviours whose implications are not always quantified.
  • Research Environments: General-purpose environments supporting multiple data-driven approaches remain lacking beyond specific domains such as district-level reinforcement-learning DSM.The review notes CityLearn as an example of a domain-specific environment.
  • Energy Flexibility Resources: Less than 1% of studies considered five energy-flexibility resources, while over 75% considered only one, predominantly passive thermal mass.The five resources are passive thermal mass, active thermal storage, active electric storage, electric vehicles, and on-site generation.

6. Conclusions

The review presents data-driven predictive control as promising for grid-interactive buildings, but finds unresolved challenges in feature practice, benchmarking, scalability, real-building validation, and multi-resource flexibility.

  • Conclusions: Most buildings have not closed the control loop between the grid and the building, limiting their participation in demand-response programmes and grid balancing.The review identifies data-driven predictive control as a potential framework for closing this loop where physics-based models are difficult to scale.
  • Conclusions: Data-driven predictive control is reported as promising for grid-interactive buildings and for harnessing passive thermal mass at single-building level.The literature also shows meaningful grid integration and carbon-footprint reduction at that level.
  • Research Gaps: No standard methodology is used for feature selection and engineering despite increased sensor installation and data availability.The review notes limited justification for the features selected in studies.
  • Benchmarking: Linear or quadratic building models guarantee convexity in optimal-control formulations, but varied black-box integration approaches lack sufficient comparison.Benchmarking datasets, processes, and toolkits are needed to compare efficacy and scalability across varying buildings and dynamic boundary conditions.
  • Scalability: Most passive-thermal-mass predictive-control applications target single buildings, leaving scalability to the wider building stock unresolved.The review states that scalability must be proven for widespread demand-side-management adoption.
  • Validation: Most predictive-control frameworks are validated on surrogate simulations rather than real buildings with Building Automation System integration and stochastic disturbances.The review identifies practical implementation and real-disturbance response as commonly unconsidered.
  • Multi-Resource Flexibility: Few studies combine passive thermal mass with thermal storage, batteries, electric vehicles, or onsite generation.These additional flexibility sources are increasingly integrated with building energy systems.

7. Recommendations and Future Work

The review identifies four priorities for advancing data-driven predictive control: feature selection, benchmarking, data quality, and scalability. It frames these as future challenges for improving building-grid integration and energy-flexibility applications.

  • Overall research agenda: Together, these priorities are identified as research gaps requiring further attention to accelerate market accessibility and penetration.The review presents them as future challenges for data-driven predictive control in building-grid integration.
  • Feature selection: Feature selection should identify robust, building-relevant inputs for energy-flexibility applications.The review asks which features are essential for different building types.
  • Benchmarking: Benchmarking needs a suitable dataset and process for comparing data-driven predictive-control techniques.The review specifically raises this issue for predictive optimal control in multi-energy-vector buildings.
  • Data quality: Data quality, including synthetic versus real data and training-period choices, may affect predictive-control performance and harnessed energy flexibility.The review identifies data quality as a research question spanning model performance and its consequent flexibility impact.
  • Scalability and transferability: Model and control development should become more transferable and scalable across different or multiple buildings.The review asks how a process developed for one building can be adapted to broader building applications.
Loading 2007.14866v1…