The Transformative Potential of Advanced Machine Learning in Building Energy Management

Many thanks to our sponsor Focus 360 Energy who helped us prepare this research report.

Abstract

The escalating global imperative to achieve sustainable development goals and mitigate climate change places significant emphasis on enhancing energy efficiency within the built environment. Buildings account for a substantial proportion of global energy consumption and greenhouse gas emissions, making them a critical focus for innovation. This comprehensive research report meticulously explores the profound impact and transformative potential of integrating advanced machine learning (ML) techniques into modern building energy management systems (BEMS). It delves into a diverse array of ML algorithms, ranging from sophisticated recurrent neural networks (RNNs) for predictive analytics and deep reinforcement learning (DRL) for adaptive control to advanced anomaly detection methods for proactive fault identification. The report provides an in-depth analysis of the intricate challenges inherent in data acquisition, rigorous preprocessing methodologies, and effective feature engineering crucial for robust model development. Furthermore, it addresses the complexities of seamless integration with existing Building Management Systems (BMS), including considerations for interoperability, real-time data processing, and scalability. Crucially, the report critically examines the ethical dimensions of AI deployment in building automation, emphasizing the imperative for transparency, data privacy, security, and the mitigation of algorithmic bias. By synthesizing current research and practical applications, this document aims to provide a holistic understanding of how ML can revolutionize building operations, drive substantial energy savings, and foster more resilient and sustainable urban infrastructures.

Many thanks to our sponsor Focus 360 Energy who helped us prepare this research report.

1. Introduction

The built environment is a cornerstone of human civilization, yet its energy footprint presents one of the most significant environmental and economic challenges of our time. Buildings are estimated to consume approximately 30-40% of global primary energy and contribute a similar proportion to global greenhouse gas emissions (IEA, 2023). This substantial consumption is driven by a complex interplay of factors, including population growth, urbanization, evolving comfort standards, and the operational inefficiencies of existing infrastructure. Traditional approaches to building energy management, often relying on fixed schedules, rule-based controls, or manual adjustments, are inherently limited in their capacity to adapt to the dynamic and often unpredictable variables that influence building performance. These variables include fluctuating occupancy levels, volatile weather conditions, changing energy tariffs, and the stochastic nature of equipment degradation.

The advent of machine learning offers a powerful paradigm shift in addressing these complexities. ML algorithms possess the unparalleled ability to learn intricate patterns from vast datasets, make sophisticated predictions, and derive optimal control strategies that transcend the capabilities of conventional methods. By leveraging data-driven insights, ML can enable building systems to operate autonomously, intelligently, and adaptively, leading to significant reductions in energy consumption, operational costs, and environmental impact, all while enhancing occupant comfort and productivity. This report provides an exhaustive examination of the methodologies, applications, challenges, and ethical considerations surrounding the integration of advanced ML techniques into the fabric of building energy management. It aims to elucidate the current state-of-the-art and chart future directions for this rapidly evolving and critically important field.

Many thanks to our sponsor Focus 360 Energy who helped us prepare this research report.

2. Machine Learning Algorithms in Building Energy Management

The application of machine learning in building energy management spans a wide spectrum, addressing various facets of building operations from predictive control to fault detection and occupant-centric optimization. The selection of an appropriate ML algorithm is contingent upon the specific problem definition, the nature of available data, and the desired outcome.

2.1 Predictive Analytics for HVAC Scheduling and Demand Response

Heating, Ventilation, and Air Conditioning (HVAC) systems are unequivocally the most energy-intensive components within commercial and often residential buildings, frequently accounting for 40-60% of total building energy consumption. Optimizing their operation is paramount for achieving substantial energy savings. Predictive analytics, powered by ML algorithms, moves beyond reactive control to anticipate future energy demand and adjust HVAC operations proactively.

Recurrent Neural Networks (RNNs), particularly their advanced variants like Long Short-Term Memory (LSTM) networks and Gated Recurrent Units (GRUs), have emerged as highly effective tools for time-series forecasting in building energy contexts. Unlike traditional feedforward neural networks, RNNs are designed to process sequential data, possessing ‘memory’ that allows them to leverage information from previous time steps to inform future predictions. In the context of HVAC scheduling, RNNs can learn the complex, non-linear relationships between:

• Historical Energy Consumption: Past electrical load profiles, thermal energy usage, and gas consumption.
• External Weather Data: Temperature (dry-bulb, wet-bulb), humidity, solar irradiance, wind speed and direction, cloud cover, and precipitation forecasts (Jain et al., 2020).
• Internal Building Conditions: Occupancy levels (derived from sensor data, Wi-Fi analytics, access control systems), indoor temperature and humidity setpoints, and CO2 concentrations.
• Building Characteristics: Thermal properties of the building envelope, internal gains from equipment and lighting, and system efficiencies.
• Operational Schedules: Planned operating hours, holiday schedules, and demand response events.

By processing these inputs, RNNs can accurately forecast future energy demand, typically for horizons ranging from minutes to days ahead. This predictive capability enables BEMS to implement strategies such as:

• Pre-cooling/Pre-heating: Initiating HVAC operations during off-peak hours or when renewable energy generation is high, thus shifting load away from peak demand periods or high-carbon intensity times.
• Optimal Setpoint Adjustment: Dynamically modifying indoor temperature setpoints within acceptable comfort ranges based on predicted occupancy and external conditions, ensuring comfort is maintained with minimal energy expenditure.
• Load Shedding and Shifting: Participating actively in utility demand response programs by accurately forecasting peak loads and strategically reducing non-critical consumption or leveraging thermal mass.
• Enhanced Fault Detection: Anomalies in predicted versus actual consumption can signal potential equipment malfunctions even before they become critical, integrating with fault detection systems.

Beyond RNNs, a broader spectrum of predictive models contribute significantly to this domain:

• Traditional Statistical Models: ARIMA (AutoRegressive Integrated Moving Average) and SARIMA models are effective for capturing linear temporal dependencies and seasonality in energy data, especially for short-term forecasting.
• Ensemble Methods: Random Forests, Gradient Boosting Machines (e.g., XGBoost, LightGBM) can handle non-linear relationships and interactions between features, offering high accuracy and robustness.
• Support Vector Machines (SVMs) and Regression Models: While less adept at sequential data without feature engineering, these can be effective for specific sub-problems or as components in hybrid models.
• Deep Learning Architectures: Convolutional Neural Networks (CNNs) can be adapted for time series by treating segments as ‘images,’ extracting local patterns. Transformer networks, initially popular in natural language processing, are increasingly being applied to time series due to their ability to model long-range dependencies efficiently.

The benefits extend beyond mere energy savings, encompassing improved grid stability through demand-side management and a reduction in greenhouse gas emissions by optimizing energy use during periods of lower grid carbon intensity.

2.2 Anomaly Detection for Equipment Faults and Performance Degradation

Continuous, proactive monitoring of building systems is essential for maintaining optimal operational efficiency and ensuring occupant safety and comfort. Equipment malfunctions, sensor drift, or control system errors can lead to significant energy waste, reduced comfort, and costly downtime. Machine learning-driven anomaly detection provides a powerful mechanism to identify these deviations from normal operational patterns swiftly and accurately, enabling timely maintenance interventions before minor issues escalate into major failures.

Anomaly detection in building systems involves learning a ‘normal’ operational profile from historical sensor data and then flagging any new data points that deviate significantly from this established norm. This ‘norm’ can be a complex, multi-variate pattern influenced by numerous factors. The types of anomalies can be broadly categorized:

• Point Anomalies: A single data point that is abnormal relative to other data, e.g., a sudden, inexplicable spike in fan power consumption.
• Contextual Anomalies: A data point that is abnormal in a specific context but not otherwise, e.g., high energy consumption during unoccupied hours is anomalous, but high consumption during peak operating hours is normal.
• Collective Anomalies: A sequence of data points that, individually, might not be anomalous but collectively represent an abnormal pattern, e.g., a gradual, sustained increase in discharge air temperature over several days indicating a failing sensor or refrigerant leak.

Various ML techniques are employed for anomaly detection:

• Statistical Methods: Simple techniques like Z-score, interquartile range (IQR), or control charts can identify univariate outliers. More advanced methods like Principal Component Analysis (PCA) can detect anomalies in multivariate data by identifying data points that deviate from the primary variance structure.
• Clustering-Based Methods: Algorithms like K-means or DBSCAN can group ‘normal’ operational states into clusters. Data points that do not belong to any cluster or are very far from existing cluster centroids are flagged as anomalies. This approach is particularly useful for identifying distinct operational modes and detecting transitions or deviations from them.
• Distance-Based Methods: K-Nearest Neighbors (K-NN) and Local Outlier Factor (LOF) assign an ‘outlier score’ based on how isolated a data point is from its neighbors. High scores indicate anomalies.
• Machine Learning-Based Methods:
  • One-Class SVM (Support Vector Machine): Learns a boundary that encapsulates the ‘normal’ data points, classifying any point outside this boundary as an anomaly.
  • Isolation Forest: Constructs an ensemble of isolation trees to isolate anomalies. Anomalies are data points that require fewer splits to be isolated compared to normal points.
  • Autoencoders (Deep Learning): A neural network trained to reconstruct its input. When trained on normal data, it learns a compressed, efficient representation. Anomalous data, being dissimilar, will result in high reconstruction errors, which can be used as an anomaly score. This is particularly effective for high-dimensional sensor data streams.
  • Generative Adversarial Networks (GANs): Can learn the distribution of normal data and then identify samples that do not fit this distribution.

Early detection of anomalies (Jarvis Contracting, 2023) translates into several tangible benefits:

• Preventive Maintenance: Shifting from reactive repairs to predictive maintenance, reducing emergency call-outs and extending equipment lifespan.
• Energy Savings: Addressing inefficiencies caused by faulty equipment (e.g., stuck dampers, refrigerant leaks, clogged filters, sensor drift) that can lead to significant energy waste.
• Improved Occupant Comfort: Preventing system failures that disrupt temperature control, air quality, or lighting.
• Reduced Operational Costs: Minimizing repair costs, reducing downtime, and optimizing maintenance schedules.

Robust anomaly detection requires clean, reliable sensor data across various building systems, including HVAC, lighting, power distribution, and even plumbing. The system must be able to adapt to seasonal changes, operational mode changes, and even the normal degradation of equipment over time to avoid excessive false positives or negatives.

2.3 Reinforcement Learning for Adaptive Energy Optimization

Reinforcement Learning (RL) represents a powerful paradigm for developing control strategies that can learn to adapt and optimize performance in complex, dynamic environments without explicit programming. Unlike supervised learning, which relies on labeled data, RL agents learn through trial and error by interacting with their environment, receiving feedback in the form of rewards or penalties.

The application of RL in building energy management frames the optimization problem as follows:

• Agent: The intelligent controller, typically an ML algorithm, responsible for making decisions.
• Environment: The building itself, encompassing its physical characteristics, HVAC systems, lighting, occupancy dynamics, external weather conditions, and energy price signals.
• State: A snapshot of the environment at a given time, comprising observable parameters such as indoor temperature, humidity, CO2 levels, occupancy, outdoor temperature, solar radiation, historical energy consumption, and current energy prices.
• Actions: The control decisions the agent can take, such as adjusting HVAC setpoints, fan speeds, damper positions, lighting levels, or scheduling energy storage discharge.
• Reward: A scalar value provided by the environment to the agent after each action, indicating the desirability of that action. In building energy management, the reward function is typically designed to maximize energy efficiency (e.g., negative energy consumption or cost) while simultaneously ensuring occupant comfort (e.g., penalty for exceeding comfort thresholds) and meeting other operational constraints.
• Policy: The strategy learned by the agent, mapping observed states to optimal actions.

Over many iterations and interactions with the environment (often simulated initially), the RL agent learns an optimal policy that balances conflicting objectives, such as minimizing energy cost while maintaining occupant comfort. Key RL algorithms applicable here include:

• Q-Learning and SARSA: Model-free value-based methods that learn an action-value function, estimating the expected return for taking a specific action in a given state.
• Deep Q-Networks (DQNs): Extend Q-learning by using deep neural networks to approximate the Q-function, enabling them to handle high-dimensional state spaces and continuous actions.
• Policy Gradient Methods: Directly learn the optimal policy, rather than a value function. Examples include REINFORCE and Actor-Critic methods (e.g., A2C, A3C, DDPG, PPO), which combine value-based and policy-based approaches for more stable and efficient learning.

Platforms like BuildingGym (Dai et al., 2025) provide open-source toolboxes that facilitate the development and testing of AI-based building energy management using RL. These platforms often incorporate detailed building simulations (e.g., EnergyPlus) to serve as the ‘environment,’ allowing RL agents to learn optimal policies without risking actual building operations. Research by Zhang et al. (2022) also explores meta-reinforcement learning frameworks for adaptive building energy management systems, highlighting the potential for systems to rapidly adapt to new buildings or changing conditions.

Challenges in implementing RL for building energy management include:

• Sample Efficiency: Training RL agents can require a vast number of interactions, which is often infeasible in real-world buildings. Simulation environments are crucial for initial training.
• Exploration vs. Exploitation: Balancing the need to explore new actions to discover better policies with exploiting known good policies.
• Reward Function Design: Crafting an effective reward function that accurately reflects all desired objectives (energy, comfort, cost, grid interaction) is complex.
• Safety Constraints: Ensuring that the learning process does not lead to unsafe or uncomfortable operational states.
• Transferability: Policies learned in one building might not directly transfer to another due due to differences in building physics, occupant behavior, or HVAC systems.

Despite these challenges, RL holds immense promise for achieving truly adaptive, self-optimizing building control that can respond intelligently to real-time changes and achieve unprecedented levels of energy efficiency and comfort.

2.4 Broader Applications of Machine Learning in Building Optimization

Beyond the core areas of predictive analytics, anomaly detection, and reinforcement learning for HVAC, ML extends its utility to various other critical aspects of building energy management and operational efficiency.

2.4.1 Occupancy Detection and Prediction

Accurate knowledge of real-time occupancy and its prediction significantly enhances energy management. ML models can fuse data from diverse sources such as passive infrared (PIR) sensors, CO2 sensors, ultrasonic sensors, Wi-Fi router pings, Bluetooth beacons, and even anonymized camera feeds to infer current occupancy. Predictive models, often leveraging RNNs or ensemble methods, can forecast future occupancy patterns based on historical data, schedules, and even external events. This enables highly granular, occupant-centric control of lighting, ventilation, and temperature, ensuring resources are only expended when and where needed. For instance, an HVAC zone could be pre-cooled only when a predicted high occupancy is expected, or lighting could be dimmed in unoccupied areas, leading to significant energy savings and personalized comfort.

2.4.2 Lighting Optimization

Lighting is another major energy consumer, especially in commercial buildings. ML can optimize lighting systems by integrating data from daylight sensors, occupancy sensors, schedules, and even weather forecasts. Algorithms can facilitate dynamic daylight harvesting, adjusting artificial lighting levels based on natural light availability. They can also implement adaptive dimming schedules that respond to real-time occupancy and task requirements, or even predict ideal lighting configurations based on user preferences and historical data, minimizing energy waste while maintaining visual comfort.

2.4.3 Renewable Energy Integration and Storage Optimization

With the increasing adoption of on-site renewable energy sources (e.g., rooftop solar PV) and battery storage, ML plays a crucial role in optimizing their integration. ML models can forecast solar or wind power generation, predict building load, and then intelligently dispatch stored energy or control grid interaction (e.g., buying when prices are low, selling or discharging when prices are high) to maximize self-consumption, minimize electricity costs, and support grid stability. This requires complex optimization that balances generation forecasts, load predictions, battery state-of-charge, market prices, and grid constraints.

2.4.4 Predictive Maintenance

While anomaly detection identifies current faults, predictive maintenance takes this a step further by forecasting when a piece of equipment is likely to fail. ML models, trained on historical failure data, sensor readings, and operational parameters, can identify degradation trends and estimate remaining useful life (RUL) for critical assets like chillers, boilers, or pumps. This allows maintenance to be scheduled precisely when needed, minimizing unforeseen breakdowns, optimizing inventory management for spare parts, and preventing catastrophic failures, thereby significantly extending asset life and reducing operational costs (Jarvis Contracting, 2023).

2.4.5 Building Performance Benchmarking and Retrofit Analysis

ML can analyze vast datasets from multiple buildings to establish performance benchmarks, identifying outliers or underperforming assets. By clustering buildings with similar characteristics (size, age, climate, usage), ML models can create peer groups for comparison. Furthermore, ML can be used to simulate the impact of various energy retrofit measures (e.g., insulation upgrades, window replacements, HVAC modernizations) on energy consumption and cost, helping building owners prioritize investments with the highest return on investment. This significantly streamlines the decision-making process for energy efficiency improvements.

2.4.6 Digital Twins and ML Synergy

The synergy between ML and digital twin technology is particularly powerful. A digital twin is a virtual representation of a physical asset, system, or process, continually updated with real-time data from its physical counterpart. ML models can be embedded within the digital twin to simulate different operational scenarios, predict future performance, detect anomalies, and test control strategies in a virtual environment before deployment in the physical building. This creates a continuous feedback loop where ML algorithms refine the digital twin’s predictive capabilities, and the digital twin provides a safe, rich environment for ML model development and validation, leading to highly optimized and resilient building operations (Jarvis Contracting, 2023).

These expanded applications underscore the pervasive potential of machine learning to transform every facet of building energy management, moving towards truly autonomous, intelligent, and sustainable buildings.

Many thanks to our sponsor Focus 360 Energy who helped us prepare this research report.

3. Data Acquisition, Preprocessing, and Feature Engineering

The efficacy and reliability of any machine learning model are fundamentally contingent upon the quality, quantity, and relevance of the data it is trained on. In the context of building energy management, where data originates from a multitude of disparate sensors, systems, and external sources, rigorous data acquisition, preprocessing, and feature engineering are not merely best practices but absolute prerequisites for successful ML deployment.

3.1 Data Acquisition Challenges

The collection of comprehensive and high-fidelity data from operational buildings is often fraught with significant challenges that can undermine subsequent ML efforts:

• Sensor Inaccuracies and Malfunctions: Sensors, the primary source of building data, are susceptible to various issues. These include calibration drift, manufacturing defects, environmental degradation, incorrect placement, and intermittent failures, leading to noisy, biased, or erroneous readings. For example, a temperature sensor exposed to direct sunlight might consistently report higher than actual ambient temperatures.
• Data Incompleteness and Gaps: Data streams are rarely continuous. Gaps can arise from communication failures (e.g., network outages, wireless interference), power interruptions to data loggers, maintenance downtime, or limited storage capacity. Missing data, if not handled properly, can introduce bias or prevent models from learning complete patterns.
• Integration with Legacy Building Management Systems (BMS): Many existing BMS are proprietary, utilize diverse communication protocols (e.g., BACnet, Modbus, LonWorks, KNX), and lack open Application Programming Interfaces (APIs). Extracting data from these siloed systems can be a complex, time-consuming, and costly undertaking, often requiring custom integration layers or middleware solutions.
• Lack of Semantic Interoperability: Even when data can be extracted, it often lacks consistent naming conventions or semantic meaning across different systems or buildings. For instance, ‘Zone_Temp_1’ in one building might correspond to ‘Room_A_Temperature’ in another, making data aggregation and model generalization difficult without manual mapping or specialized schemas.
• High Data Volume and Velocity: Modern smart buildings generate massive volumes of high-frequency data (e.g., readings every minute or second from thousands of sensors). Storing, transmitting, and processing this data in real-time presents significant infrastructure and computational challenges.
• Data Silos and Fragmentation: Data relevant for energy optimization often resides in different systems operated by different departments (e.g., BMS for HVAC, lighting control system, access control for occupancy, utility meters for consumption). Integrating these disparate sources into a unified dataset for ML analysis requires sophisticated data orchestration.
• Sampling Rate and Resolution Issues: Data collected at too low a frequency might miss critical events or rapid fluctuations, while excessively high frequencies can lead to unnecessary data burden without adding proportional value.

Ensuring data reliability, consistency, and accessibility is paramount. Without a robust data acquisition infrastructure, even the most sophisticated ML algorithms will yield suboptimal or misleading results.

3.2 Data Preprocessing Techniques

Raw sensor data is inherently ‘dirty’ and unsuitable for direct use in ML models. Data preprocessing is the crucial step of transforming raw data into a clean, structured, and usable format. This stage significantly impacts model accuracy, robustness, and interpretability.

• Handling Missing Values (Imputation): Strategies to address data gaps include:
  • Deletion: Removing rows or columns with missing values (only suitable if missingness is minimal and random).
  • Simple Imputation: Replacing missing values with the mean, median, or mode of the respective feature. This is straightforward but can reduce variance and distort relationships.
  • Last Observation Carried Forward (LOCF) / Next Observation Carried Backward (NOCB): For time-series data, filling gaps with the last observed valid value or the next valid value.
  • Interpolation: Using linear, spline, or polynomial interpolation methods to estimate missing values based on surrounding data points in a time series.
  • Model-Based Imputation: Employing regression models or K-Nearest Neighbors (K-NN) to predict missing values based on other features. More advanced techniques like Kalman filters can be used for state estimation in dynamic systems.
• Outlier Detection and Handling: Outliers, which are data points significantly different from the rest, can skew model training. Techniques include:
  • Statistical Methods: Z-score (for normally distributed data), IQR rule (for non-normally distributed data).
  • Isolation Forest or One-Class SVM: ML-based methods that identify anomalies in multivariate datasets.
  • Domain-Specific Thresholds: Identifying physically impossible or highly improbable values (e.g., negative energy consumption, temperatures far outside equipment operating ranges).
  • Handling: Outliers can be removed, capped (winsorization), or transformed depending on their nature and impact.
• Noise Reduction: Random errors or unwanted variations in data can obscure true patterns. Techniques include:
  • Smoothing Filters: Moving average, Exponential Smoothing, Gaussian filters.
  • Kalman Filters: Optimal estimators for linear dynamical systems, effectively reducing noise in time series.
  • Wavelet Transforms: Decomposing signals into different frequency components, allowing noise to be isolated and removed.
• Data Normalization and Scaling: Many ML algorithms (e.g., SVMs, neural networks, K-NN) are sensitive to the scale and range of input features. Scaling ensures that features contribute equally to the model.
  • Min-Max Scaling: Rescales features to a fixed range, typically [0, 1].
  • Standardization (Z-score normalization): Rescales features to have a mean of 0 and a standard deviation of 1, suitable for algorithms assuming normally distributed data.
  • Robust Scaling: Uses median and interquartile range, making it less susceptible to outliers.
• Time Series Specific Preprocessing:
  • Resampling: Adjusting the frequency of time series data (e.g., aggregating 1-minute data to 15-minute or hourly averages).
  • Detrending: Removing a long-term trend component from the data to focus on cyclical or seasonal patterns.
  • Deseasonalization: Removing seasonal components to reveal underlying patterns or make the data stationary.

Thorough preprocessing is critical for enhancing model accuracy, preventing algorithm convergence issues, and ensuring that models learn meaningful relationships rather than noise or artifacts from the data.

3.3 Feature Engineering Best Practices

Feature engineering is arguably the most creative and impactful stage in the ML pipeline for building energy management. It involves transforming raw data into meaningful features that better represent the underlying phenomena and improve the predictive power of ML models. This process heavily relies on domain expertise.

• Leveraging Temporal Features: Time-series data inherently contains rich temporal information. Useful features include:
  • Time of Day: Hour, minute.
  • Day of Week: Monday-Sunday (categorical or one-hot encoded).
  • Day Type: Weekend, weekday, public holiday.
  • Month, Season, Quarter: Capturing seasonal variations in energy use and weather.
  • Lagged Features: Previous hour’s energy consumption, previous day’s average temperature, or consumption at the same hour on the previous day. These capture autocorrelation and temporal dependencies.
• Integrating Weather-Related Features: Weather is a primary driver of building energy consumption.
  • Outdoor Air Temperature: Current, past few hours’ average, forecast temperature.
  • Temperature Difference: Difference between indoor setpoint and outdoor temperature (a proxy for heating/cooling load).
  • Degree Days: Heating Degree Days (HDD) and Cooling Degree Days (CDD) summarize the cumulative heating/cooling load over a period.
  • Solar Irradiance/Radiation: Direct and diffuse solar radiation, especially relevant for lighting and cooling loads.
  • Humidity, Wind Speed, Cloud Cover: All influence thermal comfort and building envelope heat transfer.
• Incorporating Occupancy and Use-Related Features:
  • Occupancy Count: From sensors, Wi-Fi data, or access logs (actual or predicted).
  • Occupancy State: Binary (occupied/unoccupied) or categorical (low, medium, high).
  • Building Type and Use Schedule: (e.g., office, retail, residential), operating hours.
• Creating Interaction Terms: Combining existing features can reveal non-linear relationships. For example, ‘temperature * occupancy’ might better explain HVAC load than either feature alone.
• Aggregations and Rolling Statistics: Creating features like rolling averages, standard deviations, or sums over specific time windows to capture trends and variability (e.g., 24-hour rolling average of power consumption).
• Domain Expertise is Paramount: An understanding of building physics, HVAC systems, occupant behavior, and local climate is crucial for identifying which features are likely to be predictive and physically meaningful. For instance, knowing that thermal mass causes a delay in temperature response can lead to creating lagged temperature features that are more predictive.
• Feature Selection: Once a rich set of features is engineered, it’s often necessary to select the most relevant ones to avoid overfitting, reduce computational complexity, and improve model interpretability. Techniques include:
  • Filter Methods: Based on statistical measures (e.g., correlation, mutual information) between features and the target variable.
  • Wrapper Methods: Use a specific ML model to evaluate subsets of features (e.g., Recursive Feature Elimination).
  • Embedded Methods: Feature selection is built into the model training process (e.g., Lasso regression, which performs regularization and feature selection simultaneously).

Effective feature engineering can often lead to greater improvements in model performance than simply using more complex algorithms, highlighting its critical role in the success of ML applications in building energy management.

Many thanks to our sponsor Focus 360 Energy who helped us prepare this research report.

4. Integration with Building Management Systems

The theoretical promise of machine learning in building energy management can only be fully realized through its seamless and robust integration with existing Building Management Systems (BMS). This integration is not merely a technical challenge but also involves addressing issues of interoperability, data flow, processing capabilities, and system scalability.

4.1 Compatibility and Interoperability

The landscape of BMS is highly fragmented, characterized by a proliferation of proprietary systems, diverse communication protocols, and varying levels of data accessibility. This heterogeneity poses significant hurdles to the widespread deployment of ML solutions.

• Communication Protocols: Buildings often employ a mix of standard and proprietary protocols:
  • BACnet (Building Automation and Control Network): The most prevalent open protocol for interoperability between different vendors’ equipment in building automation. While widely adopted, different implementations and versions can still lead to compatibility issues. Interfacing with BACnet/IP, BACnet MS/TP, or BACnet/Ethernet requires specialized gateways or software clients. (MDPI, 2024, ‘Machine Learning in Smart Buildings: A Review of Methods, Challenges, and Future Trends’)
  • Modbus: A simpler, older protocol often used for industrial control and prevalent in older HVAC equipment. Its simplicity can be a double-edged sword, as it lacks the advanced features of BACnet.
  • LonWorks (Local Operating Network): Another open control platform often used for distributed control systems.
  • KNX: A global standard for home and building control, particularly strong in Europe.
  • Proprietary Protocols: Many legacy systems or specific vendor solutions use their own unique communication methods, making data extraction difficult without vendor-specific interfaces or reverse engineering.
• Lack of Open APIs: Historically, BMS vendors have not prioritized open Application Programming Interfaces (APIs), creating ‘data silos’ where operational data is locked within the vendor’s ecosystem. This necessitates custom development for each integration, increasing costs and deployment time.
• Semantic Interoperability: Even if data can be extracted, interpreting its meaning across different systems is challenging. For example, ‘air handler unit 1 discharge air temperature’ might be represented by different identifiers, units, or data types in different BMS. This lack of a common data model requires extensive manual mapping, which is error-prone and not scalable.
• Emerging Standards: To address semantic interoperability, initiatives like Project Haystack and Brick Schema provide standardized naming conventions and data models for building operational data. These efforts aim to create a common language for building data, significantly simplifying the integration of analytical applications like ML. The Digital Twin Definition Language (DTDL) also contributes to defining asset characteristics and relationships in a standardized way.

Overcoming these challenges requires a concerted effort towards adopting open standards, developing robust middleware layers that abstract protocol complexities, and promoting vendor collaboration to ensure broader data accessibility through standardized APIs.

4.2 Real-Time Data Processing

Many advanced ML applications, such as real-time adaptive control, demand response optimization, and immediate anomaly detection, necessitate processing and acting upon data with minimal latency. This often pushes the computational requirements beyond traditional centralized cloud processing.

• Edge Computing Solutions: Edge computing involves performing computation closer to the data source (i.e., at the building or network gateway level) rather than sending all data to a centralized cloud. This architecture offers several critical advantages (MDPI, 2023, ‘Applications and Trends of Machine Learning in Building Energy Optimization’):
  • Reduced Latency: Decisions can be made and actions executed within milliseconds, which is vital for time-sensitive control loops in building systems.
  • Reduced Bandwidth Requirements: Only processed insights or aggregated data need to be sent to the cloud, reducing network traffic and associated costs.
  • Enhanced Data Privacy and Security: Sensitive operational data can be processed locally, minimizing the exposure of raw data outside the building’s perimeter.
  • Offline Operation: Systems can continue to function and make intelligent decisions even during temporary loss of internet connectivity.
• Cloud Integration: While edge computing handles real-time local processing, the cloud remains indispensable for:
  • Massive Data Storage: Long-term archival of historical data for model retraining and historical analysis.
  • High-Performance Computing: Training complex deep learning models that require significant computational resources.
  • Global Analytics and Benchmarking: Aggregating data across multiple buildings for portfolio-level insights and performance comparison.
  • Software Updates and Model Management: Centralized deployment of ML model updates and maintenance.
• Hybrid Architectures: The most effective approach often involves a hybrid architecture, combining the best of both edge and cloud. Edge devices handle real-time control and immediate analytics, while the cloud performs batch processing, model retraining, and higher-level optimization across a portfolio of buildings. Data synchronization mechanisms are crucial for maintaining consistency between edge and cloud components.
• Stream Processing Frameworks: For handling the high velocity and volume of real-time sensor data, specialized stream processing frameworks like Apache Kafka (for data ingestion and buffering), Apache Flink, or Spark Streaming are employed. These enable continuous processing of data ‘in motion,’ facilitating immediate insights and triggers for ML models.

Achieving seamless real-time data processing requires careful architectural design, robust network infrastructure, and efficient data pipelines that can handle the volume, velocity, and variety of building operational data.

4.3 Scalability Considerations

As building portfolios grow, and as more sensors and smart devices are integrated into individual buildings, the underlying ML infrastructure must be designed for scalability. This ensures that performance is maintained without degradation as data volumes increase and computational demands evolve.

• Data Storage Scalability: Traditional relational databases can struggle with the sheer volume and velocity of time-series sensor data. Modern solutions include:
  • Time-Series Databases (TSDBs): Optimized for storing and querying time-stamped data (e.g., InfluxDB, Prometheus, TimescaleDB).
  • NoSQL Databases: Flexible schema databases suitable for diverse sensor data (e.g., Cassandra, MongoDB).
  • Data Lakes: Storing raw, semi-structured, and structured data at scale, often on cloud object storage services (e.g., AWS S3, Azure Data Lake Storage).
• Computational Resource Scalability: Training and inference for ML models can be computationally intensive. Cloud-native solutions offer elasticity:
  • Serverless Computing: Automatically scales compute resources based on demand (e.g., AWS Lambda, Azure Functions).
  • Containerization (Docker) and Orchestration (Kubernetes): Enables packaging ML models and their dependencies into portable containers, which can then be deployed and scaled across clusters of machines in a highly efficient and automated manner.
  • Distributed Computing Frameworks: Apache Spark and Hadoop can distribute large-scale data processing and ML tasks across multiple nodes.
• Model Management and MLOps: Scalability extends beyond infrastructure to the lifecycle management of ML models themselves. MLOps (Machine Learning Operations) principles are crucial:
  • Model Versioning: Tracking different iterations of models and datasets.
  • Automated Deployment: Deploying new models to production environments seamlessly.
  • Continuous Monitoring: Tracking model performance (accuracy, drift) and system health in production.
  • Automated Retraining: Periodically retraining models with fresh data to maintain accuracy and adapt to changing conditions (e.g., seasonal changes, building renovations, occupant behavior shifts).
  • Experiment Tracking: Managing various experiments, hyperparameters, and results.
• Horizontal vs. Vertical Scaling: Prioritizing horizontal scaling (adding more low-cost machines) over vertical scaling (upgrading to more powerful, expensive machines) for cost-effectiveness and resilience.

Without a well-planned scalable architecture, ML solutions in building energy management risk becoming bottlenecks, costly to maintain, or obsolete as the demands on them inevitably increase. Scalability ensures that the system can grow with the needs of the building and its occupants, delivering sustained value over time.

Many thanks to our sponsor Focus 360 Energy who helped us prepare this research report.

5. Ethical Considerations and Explainability

The integration of sophisticated AI and ML into the core operational fabric of buildings is not merely a technical undertaking; it introduces a complex array of ethical considerations that demand careful scrutiny. Addressing these concerns is paramount for fostering trust, ensuring fair outcomes, and achieving broad societal acceptance of intelligent building technologies.

5.1 Transparency and Explainability in AI Decision-Making

One of the primary ethical challenges in deploying AI in building automation is the ‘black box’ problem, where complex ML models make decisions without providing clear, human-understandable justifications. For building operators, occupants, and regulators, understanding how an AI system arrives at a particular recommendation or control action is crucial.

• The Need for Explainable AI (XAI): Transparency in AI decision-making (Tzortzis et al., 2024) is vital for several reasons:
  • Trust and Acceptance: Users are more likely to trust and adopt systems they can understand. If a building’s HVAC system makes an unexpected adjustment, knowing the rationale (e.g., ‘system pre-cooled because a peak demand charge is predicted and occupancy is low’) builds confidence.
  • Debugging and Auditing: When an ML system performs unexpectedly or makes an error, explainability helps engineers identify the root cause, debug the model, and rectify issues.
  • Regulatory Compliance: Future regulations may mandate explanations for AI systems, particularly those impacting resource allocation or individual comfort.
  • Optimized Human-AI Collaboration: Operators can learn from AI’s insights and refine their own strategies, leading to a more synergistic approach.
• Techniques for XAI: While some ML models (e.g., linear regression, decision trees) are inherently interpretable, deep learning and complex ensemble methods often require post-hoc explanation techniques:
  • Model-Agnostic Methods: These work on any trained model without needing access to its internal workings.
    • SHAP (SHapley Additive exPlanations): Based on cooperative game theory, SHAP values explain the contribution of each feature to a prediction for a single instance, providing local interpretability.
    • LIME (Local Interpretable Model-agnostic Explanations): Explains individual predictions by approximating the black-box model locally with an interpretable model (e.g., linear regression).
  • Model-Specific Methods: Leverage the internal structure of certain models.
    • Feature Importance: In tree-based models (Random Forests, Gradient Boosting), this measures how much each feature contributes to the reduction in impurity across all trees.
    • Attention Mechanisms: In deep learning models (e.g., Transformers), attention layers highlight which parts of the input data were most influential in a particular prediction.
  • Surrogate Models: Training a simpler, interpretable model (e.g., a decision tree) to mimic the predictions of a complex black-box model, thereby providing an interpretable approximation.
• User Interface for Explanations: Presenting explanations in an intuitive, actionable format through dashboards or natural language interfaces is crucial for their utility. For example, a system could provide alerts like ‘Anomaly detected in Chiller A: fan speed unusually high given current load and outdoor temperature, indicating potential bearing wear.’

Developing and integrating XAI capabilities into building energy management systems is essential for building trust, facilitating better decision-making, and ensuring responsible AI deployment.

5.2 Data Privacy and Security

The intelligent operation of buildings relies on the collection and analysis of vast quantities of data, including occupancy patterns, individual comfort preferences, schedules, and even biometric data in advanced access systems. This data, if mishandled, poses significant risks to individual privacy and system security.

• Data Privacy Concerns:
  • Tracking and Profiling: Continuous monitoring of occupancy, movement, and resource usage can inadvertently reveal sensitive information about individuals’ habits, presence, and even health status.
  • Personal Preferences: Learning individual temperature or lighting preferences could be seen as an invasion of privacy if not explicitly consented to.
  • Anonymization Challenges: While anonymization aims to remove personally identifiable information, re-identification attacks are increasingly sophisticated, making truly anonymous data difficult to achieve.
• Regulatory Frameworks: Stricter data protection regulations, such as the General Data Protection Regulation (GDPR) in Europe and the California Consumer Privacy Act (CCPA) in the US, impose stringent requirements on how personal data is collected, stored, processed, and shared. Compliance is not optional and carries substantial penalties.
• Cybersecurity Threats: Intelligent buildings, with their interconnected devices and centralized control systems, present appealing targets for cyber attackers:
  • Data Breaches: Unauthorized access to building data can compromise occupant privacy and proprietary operational information.
  • System Manipulation: Attacks could aim to disrupt critical building services (e.g., turning off HVAC, manipulating fire alarms), cause physical damage, or create unsafe conditions.
  • Ransomware Attacks: Holding critical building control systems hostage.
  • Spoofing and Tampering: Malicious actors might inject false sensor data to mislead ML models or manipulate control decisions.
• Mitigation Strategies:
  • Privacy-by-Design: Integrating privacy considerations into the system’s architecture from the outset, rather than as an afterthought.
  • Data Minimization: Collecting only the necessary data for a specific purpose and deleting it when no longer needed.
  • Anonymization and Pseudonymization: Employing techniques to remove or obscure direct identifiers from data. Advanced methods like Differential Privacy add noise to data to protect individual privacy while still allowing for statistical analysis.
  • Access Controls: Implementing strict role-based access controls to ensure only authorized personnel can access sensitive data and systems.
  • Encryption: Encrypting data at rest (stored data) and in transit (data being transmitted) to prevent unauthorized access.
  • Secure Communication Protocols: Using secure protocols like TLS/SSL for data exchange between devices and servers.
  • Regular Security Audits and Penetration Testing: Proactively identifying vulnerabilities in the system.
  • Federated Learning: A privacy-preserving ML technique where models are trained locally on individual devices or buildings, and only model updates (not raw data) are shared with a central server, allowing for collaborative learning without centralizing sensitive data.
  • Secure Multi-Party Computation (SMC) and Homomorphic Encryption: Advanced cryptographic techniques that allow computations to be performed on encrypted data, further enhancing privacy.

Robust data governance policies, coupled with advanced cybersecurity measures and privacy-enhancing technologies, are essential to build trust and ensure the secure and ethical operation of ML-driven smart buildings.

5.3 Bias and Fairness

Machine learning models are only as unbiased as the data they are trained on. If training data reflects existing biases or inadequately represents certain populations or conditions, the ML model can inadvertently perpetuate and even amplify these biases, leading to unfair or suboptimal outcomes.

• Sources of Bias in Building Data:
  • Historical Data Bias: If a building’s historical operation data primarily reflects the preferences of a majority group of occupants, or if sensors were consistently faulty in certain zones, an ML model trained on this data might optimize for these skewed conditions, leading to discomfort or suboptimal performance for other groups or zones.
  • Sensor Coverage Bias: Uneven distribution or accuracy of sensors can lead to certain areas or occupants being ‘under-represented’ in the data, resulting in less accurate predictions or control for those areas.
  • Algorithm Bias: Even with unbiased data, some algorithms can inherently lead to unfair outcomes if not carefully designed and evaluated.
• Manifestations of Bias in Building Energy Management:
  • Unequal Comfort Levels: An ML-driven HVAC system might optimize for the average occupant, leading to consistently too hot or too cold conditions for minority groups within the building (e.g., women often prefer higher indoor temperatures than men).
  • Unfair Resource Allocation: If an optimization algorithm implicitly prioritizes energy savings over comfort in certain zones due to historical data patterns, it could disproportionately affect those zones.
  • Inaccurate Predictions: Models trained on data primarily from one climate zone might perform poorly when deployed in a significantly different climate, leading to inefficient operation.
• Mitigation Strategies for Bias and Fairness:
  • Diverse and Representative Data Collection: Actively seeking out and incorporating data that represents the full spectrum of building conditions, occupant demographics, and operational scenarios. This might involve deploying additional sensors in previously under-monitored zones or collecting data across diverse buildings.
  • Fairness-Aware Algorithms: Developing or adapting ML algorithms that explicitly incorporate fairness metrics during training, aiming to reduce disparities in performance across different groups or conditions.
  • Bias Detection and Auditing: Continuously monitoring model outputs for signs of bias. This involves analyzing predictions and control actions across different segments of the building (e.g., different floors, orientations, occupant groups) to identify systematic discrepancies.
  • Explainable AI (XAI) for Bias Identification: Using XAI techniques to understand why a model is making certain predictions can help uncover underlying biases in the data or algorithm.
  • Human-in-the-Loop: Incorporating human oversight and feedback mechanisms to correct biased decisions and provide real-world insights that ML models might miss.
  • Regular Model Validation and Retraining: Periodically validating models against new, diverse datasets and retraining them to adapt to evolving conditions and mitigate emerging biases.
  • Defining ‘Fairness’ Contextually: Clearly defining what ‘fairness’ means in the context of building energy management (e.g., equitable comfort, equitable energy cost burden, equitable access to resources) and aligning ML objectives with these definitions.

Addressing bias and ensuring fairness is not just an ethical imperative but also a practical one. Biased systems can lead to occupant dissatisfaction, wasted energy, and ultimately, a failure to achieve the holistic goals of smart building technologies. A continuous, proactive approach to identifying and mitigating bias is crucial for the successful and responsible deployment of ML in the built environment.

Many thanks to our sponsor Focus 360 Energy who helped us prepare this research report.

6. Future Directions

The landscape of machine learning in building energy management is characterized by rapid innovation and evolving capabilities. Future research and development are poised to push the boundaries of what is possible, transforming buildings into truly intelligent, responsive, and resilient entities within a broader smart ecosystem.

6.1 Advanced ML Architectures and Methodologies

Beyond current applications, the field will see the adoption of more sophisticated ML techniques:

• Generative AI for Building Simulation: Generative models could create highly realistic synthetic building operational data, accelerating model training and testing without reliance on extensive real-world data. They could also generate optimal building designs or retrofit strategies based on desired performance criteria.
• Graph Neural Networks (GNNs): Buildings are inherently complex, interconnected systems (e.g., HVAC zones, power networks, occupancy flows). GNNs are ideal for modeling relationships within graph-structured data, offering new ways to optimize interdependencies between building components and systems.
• Neuromorphic Computing: Hardware specialized for AI, mimicking the human brain, promises ultra-low power consumption and high processing speed for on-edge ML inference, potentially enabling highly autonomous, always-on intelligent building control at the device level.
• Physics-Informed Machine Learning (PIML): Integrating fundamental physical laws and engineering principles directly into ML models. PIML can improve model accuracy, robustness, and generalizability, especially in scenarios with limited data or when extrapolating beyond observed conditions. This combination could lead to models that are both data-driven and physically consistent.

6.2 Human-in-the-Loop and Adaptive Personalization

Future systems will move towards a more symbiotic relationship between AI and human occupants/operators:

• Enhanced Human-AI Collaboration: AI will increasingly act as an intelligent assistant, providing optimal recommendations and control options, but allowing human operators to override or refine decisions based on contextual factors that AI might not fully grasp. This involves sophisticated user interfaces and natural language processing.
• Personalized Comfort and Wellness: Beyond basic temperature control, ML will enable highly personalized environmental control, adapting lighting, air quality, acoustic levels, and even scent based on individual preferences, physiological data (e.g., wearable sensors), and activity, moving towards truly adaptive and anticipatory personalized comfort zones.

6.3 Federated Learning and Privacy-Preserving AI

Addressing privacy concerns will drive the adoption of new ML paradigms:

• Federated Learning: This approach allows ML models to be trained on decentralized datasets residing on local devices or within individual buildings, without the need to transfer raw data to a central server. Only model updates (weights or gradients) are aggregated, enhancing data privacy and security while enabling collaborative learning across a network of buildings (MDPI, 2024, ‘AI4EF: Artificial Intelligence for Energy Efficiency in the Building Sector’). This is crucial for scaling ML across large portfolios without compromising sensitive occupant data.
• Homomorphic Encryption and Secure Multi-Party Computation: Further advancements in these cryptographic techniques will enable computations and ML inference directly on encrypted data, offering an even higher level of privacy protection.

6.4 Full Integration with Digital Twins and Urban Grids

The synergy between ML, digital twins, and the broader energy ecosystem will deepen:

• Dynamic Digital Twins: Digital twins will become more dynamic and predictive, continuously updated by ML models, which then inform control strategies and simulate future performance in real-time. This provides a sandbox for testing radical energy optimization strategies without risk.
• Buildings as Active Grid Participants: ML will enable buildings to become fully active participants in smart energy grids. Through precise load forecasting, flexible demand-side management, and optimized dispatch of distributed energy resources (DERs) like solar and storage, buildings can provide grid services, enhance resilience, and facilitate higher penetration of renewable energy (Wikipedia, 2024, ‘Open energy system models’). This moves beyond simply being energy consumers to becoming ‘prosumers’ and active grid assets.

6.5 Standardization and Policy Support

Accelerating the adoption of ML in buildings requires a collaborative effort to overcome systemic barriers:

• Open Standards and Protocols: Continued development and widespread adoption of open data models (e.g., Brick Schema, Project Haystack) and communication protocols are critical for seamless interoperability and reducing integration costs.
• Policy and Regulatory Incentives: Governments and regulatory bodies will play a crucial role in promoting smart building technologies through incentives, mandates for energy performance, and support for research and development.

6.6 Resilience and Sustainability

Finally, ML will be instrumental in addressing broader challenges:

• Climate Change Adaptation: ML can help buildings adapt to more extreme weather events by optimizing envelopes, control systems, and energy storage for resilience.
• Path to Net-Zero and Carbon Neutrality: ML will be a key enabler for achieving aggressive net-zero energy and carbon goals by optimizing every aspect of energy generation, consumption, and storage, and by integrating with carbon-aware energy purchasing strategies.

The future of building energy management with machine learning is one of increasing autonomy, intelligence, personalization, and integration, promising a built environment that is significantly more efficient, sustainable, and responsive to human needs and environmental imperatives.

Many thanks to our sponsor Focus 360 Energy who helped us prepare this research report.

7. Conclusion

The integration of advanced machine learning techniques into building energy management systems represents a pivotal transformation for the built environment. As detailed in this report, ML offers an unprecedented capacity for dynamic, real-time optimization of energy consumption, moving beyond the limitations of traditional rule-based controls. Through predictive analytics, algorithms such as recurrent neural networks and ensemble methods empower buildings to anticipate energy demand, enabling proactive HVAC scheduling, load shifting, and intelligent participation in demand response programs. Anomaly detection, utilizing techniques ranging from statistical methods to deep learning autoencoders, provides a crucial layer of proactive maintenance, preventing energy waste and extending equipment lifespan by identifying malfunctions before they escalate. Furthermore, reinforcement learning offers a powerful framework for developing truly adaptive control strategies that continuously learn and optimize building operations based on real-time feedback, balancing energy efficiency with occupant comfort in complex, dynamic environments (Jain et al., 2020; Dai et al., 2025; Zhang et al., 2022).

However, realizing this transformative potential is contingent upon meticulously addressing a series of intricate challenges. Data quality remains a foundational concern, requiring robust strategies for acquisition, imputation of missing values, detection of outliers, and noise reduction. The efficacy of ML models is further enhanced by thoughtful feature engineering, which leverages domain expertise to extract meaningful insights from raw sensor data. Seamless integration with existing, often proprietary, Building Management Systems demands overcoming significant hurdles related to communication protocol compatibility, semantic interoperability, and the establishment of robust data pipelines capable of real-time processing, often through hybrid edge-cloud architectures designed for scalability (MDPI, 2024, ‘Machine Learning in Smart Buildings: A Review of Methods, Challenges, and Future Trends’; MDPI, 2023, ‘Applications and Trends of Machine Learning in Building Energy Optimization’).

Crucially, the ethical dimensions of AI deployment cannot be overlooked. The imperative for transparency and explainability in AI decision-making is vital for fostering trust and enabling effective human-AI collaboration. Robust measures for data privacy and cybersecurity are non-negotiable, particularly given the sensitive nature of occupant data and the vulnerability of interconnected control systems. Furthermore, continuous vigilance against algorithmic bias is essential to ensure fairness and equitable outcomes for all occupants and building zones.

The future trajectory of machine learning in building energy management points towards even greater sophistication, encompassing advanced AI architectures like generative models and graph neural networks, privacy-preserving techniques such as federated learning, and a deeper synergy with digital twins and smart urban grids. Ultimately, a collaborative and interdisciplinary approach, involving technologists, building managers, policymakers, and occupants, is indispensable to harness the full benefits of ML for creating a more energy-efficient, sustainable, and resilient built environment.

Many thanks to our sponsor Focus 360 Energy who helped us prepare this research report.

References

• Dai, X., Chen, R., Guan, S., Li, W.-T., & Yuen, C. (2025). BuildingGym: An open-source toolbox for AI-based building energy management using reinforcement learning. arXiv preprint arXiv:2509.11922.
• Edgebricks Inc. (n.d.). ‘Top 5 Challenges and Solutions in Building AI/ML Infrastructure.’ Retrieved from https://edgebricks.com/top-5-challenges-and-solutions-in-building-ai-ml-infrastructure/ (Accessed October 26, 2023).
• International Energy Agency (IEA). (2023). ‘Buildings Energy Consumption.’ Retrieved from https://www.iea.org/reports/tracking-buildings-2023 (Accessed October 26, 2023).
• Jain, A., Smarra, F., Reticcioli, E., D’Innocenzo, A., & Morari, M. (2020). NeurOpt: Neural network based optimization for building energy management and climate control. arXiv preprint arXiv:2001.07831.
• Jarvis Contracting. (2023). ‘Optimising Energy Performance in Buildings using Machine Learning and Digital Twins.’ Retrieved from https://www.jarvisbuild.co.uk/optimising-energy-performance-in-buildings-using-machine-learning-and-digital-twins/ (Accessed October 26, 2023).
• Lawrence Berkeley National Laboratory (LBNL). (n.d.). ‘Machine Learning for Improved Efficiency Analysis & Asset Information.’ Building Technology and Urban Systems. Retrieved from https://buildings.lbl.gov/emis/machine-learning (Accessed October 26, 2023).
• MDPI. (2023). ‘Applications and Trends of Machine Learning in Building Energy Optimization: A Bibliometric Analysis.’ Buildings, 13(7), 994. Retrieved from https://www.mdpi.com/2075-5309/15/7/994 (Accessed October 26, 2023).
• MDPI. (2024). ‘AI4EF: Artificial Intelligence for Energy Efficiency in the Building Sector.’ Buildings, 14(12), 2631. Retrieved from https://www.mdpi.com/2075-5309/15/15/2631 (Accessed October 26, 2023).
• MDPI. (2024). ‘Machine Learning in Smart Buildings: A Review of Methods, Challenges, and Future Trends.’ Applied Sciences, 14(14), 7682. Retrieved from https://www.mdpi.com/2076-3417/15/14/7682 (Accessed October 26, 2023).
• Number Analytics. (n.d.). ‘Smart Buildings with Machine Learning.’ Retrieved from https://www.numberanalytics.com/blog/smart-buildings-with-machine-learning (Accessed October 26, 2023).
• Optimise AI. (n.d.). ‘Pioneering a New Era of Energy Efficiency with Machine Learning.’ Retrieved from https://optimise-ai.com/en/blog/pioneering-a-new-era-of-energy-efficiency-with-machine-learning (Accessed October 26, 2023).
• Tzortzis, A. M., Kormpakis, G., Pelekis, S., Michalitsi-Psarrou, A., Karakolis, E., Ntanos, C., & Askounis, D. (2024). AI4EF: Artificial Intelligence for Energy Efficiency in the Building Sector. arXiv preprint arXiv:2412.04045.
• Wikipedia. (n.d.). ‘Energy-based model.’ Retrieved from https://en.wikipedia.org/wiki/Energy-based_model (Accessed October 26, 2023).
• Wikipedia. (n.d.). ‘Open energy system models.’ Retrieved from https://en.wikipedia.org/wiki/Open_energy_system_models (Accessed October 26, 2023).
• Zhang, H., Wu, D., & Boulet, B. (2022). MetaEMS: A Meta Reinforcement Learning-based Control Framework for Building Energy Management System. arXiv preprint arXiv:2210.12590.