Method for controlling HVAC equipment of a building, computer system and computer program product thereof

Abstract

Computer-implemented method for controlling HVAC, heating, ventilation, and air conditioning, equipment of a building, the method comprising pretraining a machine-learning forecaster model for predicting energy consumption and indoor thermal conditions of the building, respective computer program product embodied in a non-transitory computer-readable medium and a non-transitory computer-readable medium.

Description

D E S C R I P T I O NMETHOD FOR CONTROLLING HVAC EQUIPMENT OF A BUILDING, COMPUTER SYSTEM AND COMPUTER PROGRAM PRODUCT THEREOFTECHNICAL FIELD

[0001] The disclosure relates to a computer-implemented method for controlling HVAC, heating, ventilation, and air conditioning, energy flows of a building.BACKGROUND

[0002] Most existing predictive models for controlling HVAC, heating, ventilation, and air conditioning, energy flows of a building are useful for evaluating actions that do not differ drastically from the past MO, modus-operandi, that is reflected in the historical data. Therefore, the options for actuation should be predetermined, possibly hand crafted by human domain experts. This prohibits free exploration (e.g. by a true optimizer) and, as a result, any true novelty in control strategy.

[0003] Existing solutions have several disadvantages. First, they may suffer from limited data diversity, as historical data often does not cover the full range of potential scenarios, resulting in a narrow dataset that may not account for unusual or new conditions. Second, they may be inapplicable to newly constructed building that lack historical data, making it impossible to apply this method effectively from the outset. Finally, there is a lack of novelty in the historical data, which inhibits free exploration of an improved optimizer.

[0004] Document EP2987112B1 discloses apparatus, systems, methods, and related computer program products for generating and implementing thermodynamic models of a structure. Thermostats disclosed herein are operable to control an HVAC system. In controlling the HVAC system, a need to determine an expected indoor temperature profile for a particular schedule of setpoint temperatures may arise. To make such a determination, a thermodynamic model of the structure may be used. The thermodynamic model may be generated by fitting weighting factors of a set of basis functions to a variety of historical data including time information, temperature information, and HVAC actuation state information. The set of basis functions characterize an indoor temperature trajectory of the structure inresponse to a change in HVAC actuation state, and include an inertial carryover component that characterizes a carryover of a rate of indoor temperature change that was occurring immediately prior to the change in actuation state.

[0005] Document W02018200094A1 relates to a thermostat includes an equipment controller and a model predictive controller. The equipment controller is configured to drive the temperature of a building zone to an optimal temperature setpoint by operating HVAC equipment to provide heating or cooling to the building zone. The model predictive controller is configured to determine the optimal temperature setpoint by generating a cost function that accounts for a cost operating the HVAC equipment during each of a plurality of time steps in an optimization period, using a predictive model to predict the temperature of the building zone during each of the plurality of time steps, and optimizing the cost function subject to a constraint on the predicted temperature of the building zone to determine optimal temperature setpoints for each of the time steps.

[0006] EP1866575B1 relates to methods, systems and computer program products provided for controlling a climate in a building. Sensed data is received at a local processor in the building. The sensed data is associated with the climate in the building, weather outside the building and / or occupants of the building. The received sensed data is compared at the local processor with corresponding predictive data associated with the climate in the building, weather outside the building and / or occupants of the building. One or more parameters associated with the climate of the building is adjusted at the local processor based on a result of the comparison of the received sensed data and the predictive data.

[0007] EP3286501B1 discloses a method for controlling temperature in a thermal zone within a building, comprising: using a processor, receiving a desired temperature range for the thermal zone; determining a forecast ambient temperature value for an external surface of the building proximate the thermal zone; using a predictive model for the building, determining set points for a heating, ventilating, and air conditioning ("HVAC") system associated with the thermal zone that minimize energy use by the building; the desired temperature range and the forecast ambient temperature value being inputs to the predictive model; the predictive model being trained using respective historical measured value data for at least one of the inputs; and, controlling the HVAC system with the set points to maintain an actual temperature value of the thermal zone within the desired temperature range for the thermal zone.

[0008] US2021 / 191342 discloses systems and methods for training a reinforcement learning (RL) model for HVAC control are disclosed herein. A calibrated simulation model is used to train a surrogate model of the HVAC system operating within a building. The surrogate model is used to generate simulated experience data for the HVAC system. The simulated experience data can be used to train a reinforcement learning (RL) model of the HVAC system. The RL model is used to control the HVAC system based on the current state of the system and the best predicted action to perform in the current state. The HVAC system generates real experience data based on the actual operation of the HVAC system within the building. The real experience data is used to retrain the surrogate model, and additional simulated experience data is generated using the surrogate model. The RL model can be retrained using the additional simulated experience data.

[0009] These facts are disclosed in order to illustrate the technical problem addressed by the present disclosure.GENERAL DESCRI PTION

[0010] The present disclosure relates to a computer-implemented method for predicting the indoor conditions (such as temperature, humidity, etc.) and energy consumption of HVAC (heating, ventilation, and air conditioning) systems in a building. This predictive modelling is a crucial component for any control method (including but not limited to rules, domain-expert- logic, autonomous agents, reinforcement learning algorithms, and optimization algorithms) aimed to optimizing HVAC system operations to minimize energy costs and maximize occupants comfort.

[0011] The disclosure addresses the technical problems of (i) improving the quality and (ii) enhancing the availability of building data for training predictive Machine Learning (ML) models. This issue arises due to the need to accurately predict building behavior, particularly HVAC (Heating, Ventilation, and Air Conditioning) systems and their effect on indoor conditions. These ML models are part of an automated control system that optimizes occupant comfort and the building energy costs. It should be noted that the solution is not limited to real-time data alone.

[0012] The present disclosure thus relates to a computer-implemented method for controllingHVAC, heating, ventilation, and air conditioning of a building, the method comprising traininga surrogate machine-learning forecaster model for predicting energy consumption and indoor conditions of the building by: receiving a signal comprising data from the building, including but not limited to temperature, humidity, energy consumption, air quality levels, operational settings, status variables, and alerts; developing a digital twin of the building and using it for obtaining synthetic building data; training surrogate machine-learning forecaster models using the received data and the synthetic building data; the method further comprising the use of the trained surrogate machine-learning forecaster models by a controller for: generating a plurality of control parameters sets for the building, predicting, for each generated control parameter set, energy consumption and indoor conditions of the building by using the trained surrogate machine-learning forecaster models; selecting a generated control parameter set that optimizes predetermined criteria of energy cost and occupant comfort from the predicted energy consumption and indoor conditions; sending a signal comprising the selected control parameter set to be applied to the building.

[0013] In particular, one of the primary technical problems this disclosure addresses is the "cold start" issue, which refers to the lack of sufficient training data. This availability problem occurs in buildings where little to no historical data is available. This issue arises in any building lacking historical data, with new buildings being a clear subset of this category.

[0014] Moreover, in real-world scenarios, even buildings with availability of historical data usually exhibit very consistent behaviors with minimal variation in HVAC operations. This leads to a less descriptive sample of data points and constitutes the quality problem: ML models trained with such historical equipment settings will have trouble discovering relationships and patterns outside this "normal" past behavior and will therefore have poor predictive power with novel cases.

[0015] One of the advantages of the present disclosure is their utilization of existing data, which may leverage historical data to make the approach relatively cost-effective.

[0016] Advantages of the disclosure: Digital Twin (DT) Surrogate Using Machine Learning. The present disclosure offers several significant advantages. Firstly, it enhances the quality and availability of data by generating useful and diverse datasets through the digital twin. This addresses the "cold start" issue by providing synthetic data for newly constructed buildings lacking historical data. The second advantage should be that it creates diverse datasets for buildings with consistent operation and historical data. These diverse synthetic datasets cantrain models with good performance in situations that are very different from the past; this can include different HVAC commands, building operation schedules, and environmental conditions. These models can support control strategies that explore and innovate successfully. Additionally, it enables continuous learning and adaptation, allowing the Digital Twin to evolve and improve over time. The use of real-time data minimizes the reality gap, though some gap remains, ensuring the insights are both reliable and relevant. Lastly, the innovation promotes cost-effectiveness by utilizing existing data where available and reducing the need for extensive data collection and processing. While the digital twin generates synthetic data, we consistently collect real data, either from pre-installed sensors or those we deploy ourselves. This data is essential for calibration and training the machine learning models. Here are some additional advantages that have been identified:

[0017] Comprehensive Data Generation: The digital twin can simulate a wide range of "what- if" scenarios, generating diverse and comprehensive datasets that cover potential operational conditions not presented in historical data.

[0018] Enhanced ML Training: With access to synthetic data generated by the digital twin, ML models can be trained more robustly, improving their performance in real-world scenarios.

[0019] Applicability to buildings without historical data: This solution is especially effective for new buildings lacking historical data, including new constructions. By generating buildingspecific data from scratch, it eliminates the need of extensive prior data, enabling accurate predictions and insights. Importantly, this approach is not limited to new buildings— it can be applied to any structure without existing historical data.

[0020] Adaptability: The model is effective in a variety of scenarios, even unseen events in reality.

[0021] Pre-deployment Testing: The ability to test control strategies within the digital twin before actual deployment reduces the risk of operational issues and enhances system reliability. Physics models are more reliable and provide independent testing. These models ensure that control strategies are validated against realistic and precise simulations.

[0022] The following table presents a comparison between the invention and the state of the art:

[0023] Building hardware components- such as sensors, HVAC equipment, and other devices- generate real-time data (e.g. temperature, humidity, energy consumption, air quality levels, operational settings, status variables, alerts, etc.), which is collected by on-location gateways or remote servers and passed to the software systems. The software processes this data to facilitate the creation of the digital twin, simulates scenarios to generate synthetic data, and trains machine learning models. General purpose hardware (standard computers, cloud servers) may be used to provide the necessary resources and infrastructure for executing the software instructions.

[0024] The most critical variables when creating simulated scenarios to predict energy costs and temperatures include:Outdoor conditions: Weather data such as temperature, humidity, solar radiation, and wind speed.Indoor conditions: Building occupancy patterns and internal heat gains from equipment and occupants.HVAC Control settings: Setpoints for temperature, operational schedules, control strategies, and mode selection (e.g. heat, cool, fan).

[0025] Simulating different scenarios can improve the robustness of the machine-learning predictive model by:Expanding the training dataset: Synthetic data from simulated scenarios can cover a wider range of conditions than what is available in historical data, providing the model with a more comprehensive understanding of possible situations.Enhancing model generalization: By exposing the model to a variety of scenarios, including extreme and rare conditions, it can learn to generalize better and make accurate predictions even in previously unseen situations.Improving robustness to data variability: Simulated data introduces variability and noise, helping the model become more resilient to changes and fluctuations in real-world data.

[0026] The synthetic data generated from the simulated scenarios includes:Temperature profiles: Time series data of indoor temperatures under different scenarios. Energy consumption patterns: Detailed energy usage data of the HVAC system and other building equipment.Occupancy data: Simulated patterns of how occupants use the building over time.HVAC settings: Variations in HVAC control settings and their impact on indoor conditions. Integration with real sensor data involves:Preprocessing: Normalizing and aligning the data to ensure consistency and compatibility (e.g. transforming energy J in Wh).Augmentation: Using synthetic data to augment real data, enhancing the training dataset's diversity.Validation: Validating the model's performance on real sensor data to ensure accuracy and reliability.

[0027] The metrics used to compare the simulation with real data include:Mean Absolute Error (MAE): Measures the average magnitude of errors between simulated and real values.Root Mean Square Error (RMSE): Provides a quadratic scoring rule that measures the average magnitude of errors, giving higher weight to larger errors.Coefficient of Variation of the Root Mean Square Error (CVRMSE): Standardizes the RMSE by dividing it by the mean of the observed values, useful for comparing different datasets.

[0028] The variability of simulated scenarios affects the model's ability to generalize to new situations by:Reducing overfitting: By training on a diverse dataset, the model is less likely to overfit to specific patterns in the historical data, improving its ability to generalize.Enhancing prediction accuracy: Exposure to different scenarios helps the model develop a more nuanced understanding of the relationships between variables, leading to more accurate predictions.Improving resilience to data drifts: The model becomes more robust to changes in input data distributions, such as those caused by new HVAC configurations, occupancy patterns, or climate conditions.

[0029] It is disclosed a computer-implemented method for controlling HVAC, heating, ventilation, and air conditioning, equipment of a building, the method comprising pretraining one or more machine-learning forecaster models for predicting energy consumption and indoor thermal conditions of the building, by: receiving a preexisting dataset comprising sensor data of energy consumption and indoor thermal conditions and comprising equipment operational settings, for one or more volumes of the building; simulating the building using a preexisting digital twin model for obtaining simulated data of energy consumption and indoor thermal conditions and simulated equipment operational settings, for the one or more volumes of the building, wherein the digital twin model is a physics-based model comprising explicit physical quantities and physical relationships between said physical quantities; training one or more machine-learning forecaster models using the received preexisting dataset and the simulated data and settings; the method further comprising the use of the trained one or more machine-learning forecaster models by a controller by: generating a plurality of sets of equipment operational settings for the one or more volumes of the building; predicting, for each of the generated sets of equipment operational settings, energy consumption and indoor thermal conditions using the one or more machine-learning forecaster models; selecting a generated set of equipment operational settings that maximizes one or more predetermined criteria from the predicted energy consumption and indoor thermal conditions; sending a signal comprising the selected set of equipment operational settings to be applied to the one or more volumes of the building.

[0030] Simulating the building can thus be used for generating synthetic training data by sampling operational settings for equipment of a primary device type and determining operational settings using predetermined operational rules for equipment of a dependent device type as a function of the sampled operational settings for the primary device type. This provides dependency-aware sampling, where data generation is carried out by sampling control configurations for a primary device type and then determining or deriving the state of dependent devices based on predefined operational rules, enabling the physical and operational validity of each generated data point. This synthetic data generation process systematically explores a controlled yet broadened region of the operational parameter space, thereby introducing sufficient variability to improve the model's robustness, while avoiding unrealistic scenarios that could impair predictive accuracy. This controlled variance approach can address the "cold start" and "data drift" problems more efficiently, as it guarantees that simulated scenarios remain relevant to the building's operational reality but still provide the diversity necessary for generalisation.

[0031] In an embodiment, the generation of synthetic training data by simulating the building comprises, for one or more numeric variables of said operational settings and physical quantities: receiving an equipment configuration collection comprising equipment operational settings for each unit of the HVAC, heating, ventilation, and air conditioning equipment; and uniformly drawing from a truncated normal distribution whose mean and standard deviation are determined from the received equipment configuration collection. This allocates non-negligible probability to both lower and upper ends of the truncated normal distribution to provide proportional representation of extreme and intermediate configurations, i.e to provide tail coverage; the end values of the truncated normal distribution may be the minimum and maximum of the received equipment configuration collection, or may be preferably extended in respect of the minimum and maximum of the received equipment configuration collection by an increased variance in respect of the received equipment configuration collection, see below.

[0032] In an embodiment, the generation of synthetic training data by simulating the building comprises, for Boolean variables of said operational settings of a predetermined operational settings type, uniformly drawing from a uniform distribution the total number of variables in a "true" state and then randomly setting those variables with said "true" state. This balances the dataset distribution to avoid the statistical bias introduced when Boolean device statesare sampled independently, which tends to over-represent mid-range activation counts (binomial distribution), ensuring all activation counts are equally represented, as the trained machine-learning models become more robust and generalizable, with improved accuracy in predicting outcomes for both typical and atypical configurations, optimizing HVAC performance under unusual conditions, rather than only for average operating scenarios.

[0033] In an embodiment, the predetermined operational settings type is equipment operational state, wherein the "true" state is "On" and a corresponding false state is "Off".

[0034] An embodiment further comprises simulating the building under a predictive free- floating model in the absence of HVAC actuation, over an optimization horizon, to determine a main mode of operation selected from heating and cooling based on the simulated building relative to a comfort threshold; and constraining the optimizer to generate the plurality of sets of equipment operational settings restrained to the selected main mode of operation. This restricts the optimizer's search space so that it only selects configurations corresponding to the required operation mode (e.g., only heating configurations if the predicted temperature falls below the comfort threshold), further increasing process efficiency.

[0035] An embodiment, the method further comprises the steps of: receiving a current dataset comprising sensor data of energy consumption and indoor thermal conditions for the one or more volumes of the building; making a comparison between the received dataset and the predicted energy consumption and indoor thermal conditions for the selected set of equipment operational settings; using the comparison to adjust the one or more machine-learning forecaster models.

[0036] In an embodiment, the physics-based model is a heat balance model, a heat transfer function model, a thermal network model, a response factor model, or combinations thereof.

[0037] In an embodiment, simulating the building comprises using the preexisting digital twin model for obtaining simulated data of energy consumption and indoor thermal conditions and simulated equipment operational settings, comprises applying equipment operational settings with an increased variance from 75% to 125% in respect of the equipment operational settings of the preexisting dataset.

[0038] In an embodiment, simulating the building comprises using the preexisting digital twin model for obtaining simulated data of energy consumption and indoor thermal conditions andsimulated equipment operational settings, comprises applying equipment operational settings with an increased variance from 75% to 125% in respect of the equipment operational settings of a received equipment configuration collection.

[0039] An embodiment further comprises applying equipment operational settings with an increased variance from 50% to 150% in respect of the equipment operational settings of the preexisting dataset, further in particular an increased variance from 20% to 300% in respect of the equipment operational settings of the preexisting dataset.

[0040] An embodiment further comprises applying equipment operational settings with an increased variance from 50% to 150% in respect of the equipment operational settings of the preexisting dataset, further in particular an increased variance from 20% to 300% in respect of the equipment operational settings of a received equipment configuration collection.

[0041] In an embodiment, the equipment operational settings include HVAC mode, HVAC on / off, HVAC setpoint, HVAC fan speed, HVAC operation schedules, HVAC temperature setpoints, in particular the HVAC mode comprises HVAC cooling mode, HVAC heating mode, and HVAC fan mode, further in particular the HVAC mode comprises HVAC cooling mode, HVAC heating mode, HVAC fan mode, and HVAC dehumidifier mode.

[0042] In an embodiment, the indoor thermal conditions comprise temperature for each of the one or more volumes of the building, in particular the indoor thermal conditions comprise temperature and humidity for each of the one or more volumes of the building.

[0043] In an embodiment, the one or more predetermined criteria include energy efficiency, energy cost or costs, or combinations thereof.

[0044] An embodiment further comprises generating an alert or alerts if measured energy consumption or measured indoor thermal conditions deviate from the predicted energy consumption and indoor thermal conditions.

[0045] An embodiment further comprises displaying, using a user interface, the selected set of equipment operational settings and, optionally, the predicted energy consumption and / or indoor thermal conditions.

[0046] In an embodiment, each of the one or more machine-learning forecaster models is specific to one HVAC mode.

[0047] It is also disclosed a computer program product embodied in a non-transitory computer-readable medium comprising computer program instructions, which when executed by a computer processor, cause the computer processor to carry out the computer- implemented method for controlling HVAC, heating, ventilation, and air conditioning, equipment of a building of any of the disclosed embodiments.

[0048] It is also disclosed a non-transitory computer-readable medium comprising one or more machine-learning forecaster models for predicting energy consumption and indoor thermal conditions to be used for controlling HVAC, heating, ventilation, and air conditioning, equipment of a building, wherein the one or more machine-learning forecaster models have been obtained by: receiving a preexisting dataset comprising sensor data of energy consumption and indoor thermal conditions and comprising equipment operational settings, for one or more volumes of the building; simulating the building using a preexisting digital twin model for obtaining simulated data of energy consumption and indoor thermal conditions and simulated equipment operational settings, for the one or more volumes of the building, wherein the digital twin model is a physics-based model comprising explicit physical quantities and physical relationships between said physical quantities; training one or more machine-learning forecaster models using the received preexisting dataset and the simulated data and settings.

[0049] In an embodiment, simulating the building comprises using the preexisting digital twin model for obtaining simulated data of energy consumption and indoor thermal conditions and simulated equipment operational settings, comprises applying equipment operational settings with an increased variance from 75% to 125% in respect of the equipment operational settings of the preexisting dataset, in particular applying equipment operational settings with an increased variance from 50% to 150% in respect of the equipment operational settings of the preexisting dataset, further in particular an increased variance from 20% to 300% in respect of the equipment operational settings of the preexisting dataset.BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The following figures provide preferred embodiments for illustrating the disclosure and should not be seen as limiting the scope of invention.

[0051] Figure 1: Schematic representation of an embodiment comparing the present disclosure with the Prior Art.

[0052] Figure 2: Schematic representations of an embodiment illustrating the workflow of the building optimization solution.

[0053] Figure 3: Schematic representation of an embodiment of a use-case illustration.DETAILED DESCRI PTION

[0054] The present disclosure relates to a computer-implemented method for controlling the energy flows of HVAC (heating, ventilation, and air conditioning) systems in a building. The method comprises pretraining a surrogate machine-learning forecaster model to predict the building's energy flows and temperatures. Furthermore, the pretrained surrogate machinelearning model is used by a controller to optimize and manage the HVAC system efficiently. It is important to note that the controller itself is not a machine learning model; rather, it leverages the insights from the pretrained model to make informed decisions and adjustments for improved system performance.

[0055] The present disclosure employs a digital twin surrogate to power predictive machine learning by generating synthetic data. The benefits of the present disclosure are:I) dealing with the cold start problem, i.e. building has no (or very little) historic data,II) ability to deal with changing conditions, andIII) capacity for novelty in control strategy.

[0056] A digital twin is a virtual representation of a building that mirrors its physical behaviour. The present disclosure uses real data from sensors, equipment and devices within the building— such as temperature, humidity, energy consumption, and operational settings— along with external data - such as weather - to create a digital twin that is a faithful representation of the real building. The real data that is needed for the digital twin creation can be very limited, around 2-4 weeks. This amount of data would be insufficient for the training of a meaningful machine learning prediction model, but not for the creation of a realistic digital twin; therefore, the disclosed solution can successfully handle buildings without historical data (cold start problem).

[0057] The available data might be insufficient for training a meaningful machine learning prediction model due to several limitations:Limited historical data: There may be a lack of extensive historical data covering a wide range of conditions and scenarios, reducing the model's ability to learn from diverse examples.Insufficient variability: The available data might not include enough variability in key variables such as weather conditions, occupancy patterns, or HVAC settings, limiting the model's ability to generalize to new situations, and thus, our optimizer will lack the variability needed to explore novel approaches and scenarios.Data sparsity: Data may be sparse, with missing values or gaps in time series, which can hinder the model's ability to learn accurate patterns.

[0058] Even if the data is insufficient for training a robust machine learning model, it can still be adequate for creating a realistic digital twin because:Calibration and validation: The available data is enough for the calibration and validation during the digital twin construction, ensuring it accurately represents the real-world system.Physical and engineering principles: Digital twins leverage physical and engineering principles, which can fill in gaps and provide a more complete representation of the system.Integration of expert knowledge: Expert knowledge and domain-specific insights can be integrated into the digital twin to enhance its accuracy and realism.

[0059] To compensate for the lack of sufficient data when creating a digital twin, the following techniques are used to ensure its accuracy and reliability:Integration of domain expertise: Expert knowledge and industry standards are integrated into the digital twin to guide its development and validation.Benchmark measures: Benchmarking involves comparing the digital twin's performance against known standards and performance metrics from similar buildings in the same sector. This helps in validating the accuracy of the digital twin by ensuring it performs within expected ranges.Calibration: The digital twin is calibrated using the available real-world data to fine-tune its parameters and behaviour to closely match the actual building's performance. Thisinvolves adjusting the digital twin model until its output aligns with observed data from the real building.

[0060] Data might be insufficient for machine learning but sufficient for a digital twin in scenarios such as:Limited historical data: When there is not enough historical data to train a machine learning model, but the available data can still be used to calibrate and validate a digital twin, since it only requires around two weeks of historical data for that.Consistent behaviours: If the available data mostly represents repetitive behaviours, the prediction model will perform well on familiar patterns but will fail to generalize to new or unseen situations. However, such data is still valuable for creating a digital twin. A digital twin relies on accurately modelling the physical and mechanical aspects of the building, which can be effectively calibrated using consistent data. This means that even with repetitive behaviour data, the digital twin can still provide a realistic and reliable representation of the building's operations. The calibration process focuses on aligning the digital twin with the observed physical and thermal properties, rather than needing to explore a wide range of behavioural conditions. This allows the digital twin to simulate and predict performance accurately within the scope of the available data.

[0061] The difference between DT and ML is the fundamental modelling approach. ML has domain agnostic representations that can express any kind of relationship / pattern. DT is physics-based and incorporates domain-specific elements; it can express a very specific and limited range of patterns and systems. We could say that ML is a "generalist" (ok performance on a very wide range of problems) and DT a "specialist" (great performance on a very limited type of problem).

[0062] In general, the more degrees of freedom and the more agnostic the model, the more expressive it is and the more data it requires for meaningful fitting.

[0063] Therefore, it is about the nature of the models itself, and not about specific techniques or situations.

[0064] Using the digital twin, we can then simulate a wide range of scenarios using specific generation strategies. This synthetic data generation allows for the creation of comprehensive datasets that encompass a broad array of potential actions and conditions. The generated synthetic data is then used to train machine learning models that can predict (i) comfortconditions and (ii) energy cost, given specific operation actions. By augmenting the limited (or absent) real-world data with the strategically generated synthetic data points, we can create predictive models with good performance even for input that is very different from any available historic data of the real building. This makes the models resilient to changes in conditions, such as building operation hours, equipment operational settings, weather, etc. (data drift problem). After the initial training of the ML models with the synthetic data, the (deployed) ML models are periodically re-trained using real building data that comes in over time.

[0065] A wide range of scenarios in the context of HVAC energy flow and temperature predictions includes:Weather conditions: Variations in temperature, humidity, solar radiation, wind speed, and seasonal changes.Occupancy patterns: Different levels and patterns of building occupancy, including peak times, off-peak times, and varying schedules.Equipment usage: Different usage patterns of lighting, grills, and other energy-consuming equipment.Operational strategies: Variations in HVAC control strategies, such as varying temperature setpoints, scheduling of HVAC operations, and different modes of operation.

[0066] These scenarios are defined by strategically peaking combinations of HVAC settings together with a wide range of weather, occupancy and usage setups. They are categorized based on key influencing factors and their potential variations: weather conditions, occupancy patterns, equipment usage, and operational strategies.

[0067] Examples of scenarios critical for system optimization include:Extreme weather conditions: Simulating HVAC performance during extremely hot or cold weather to ensure the system can maintain comfort and efficiency.Free-floating temperature models: Simulating the building with the HVAC always off to train free-floating temperature models. This helps determine if indoor conditions can remain within the comfort range without activating the HVAC system. This scenario is critical for maximizing energy savings.Low occupancy periods: Analysing energy-saving strategies during periods of low or no occupancy, such as nighttime or weekends.System options: Simulating scenarios where parts of the HVAC system are turned off (for several reasons) to ensure the system with the remaining HVAC devices can adapt and maintain acceptable performance.

[0068] Indicators or metrics used to assess performance across different scenarios include:Energy consumption: Measuring the total energy used by the HVAC system and other equipment.Energy costs: Calculating the energy costs taking into account different tariffs structures (e.g. seasonal and / or daily changes, peak power calculations).Thermal comfort: Assessing occupant comfort levels using metrics such as the Predicted Mean Vote (PMV) and Degree Hour (DH).Response time: Measuring how quickly the system can adapt to changing conditions and achieve the desired setpoints.

[0069] These are not scenarios in the classical sense. If we consider the N-dimensional space created by all the influential factors (model inputs), "scenarios" are various areas of that space. The "scenarios" we create are covering this space in a much better way, than historical data that would typically be concentrated in small specific areas.

[0070] Therefore, we are interested not in critical or extreme scenarios, but about combinations of factors that are comprehensive enough to train a model that can capture the patterns of all possible situations it can be queried about.

[0071] The specific operation actions will optimize the relation between thermal comfort and energy costs.

[0072] Examples of specific actions that significantly impact HVAC system efficiency include: Adjusting thermostat setpoints: Lowering or raising temperature setpoints during different times of the day or night to reduce energy use.Outdoor conditions: Taking advantage of favourable outdoor conditions to provide free cooling, thereby reducing the load on mechanical cooling systems.Optimizing ventilation rates: Adjusting ventilation based on occupancy levels and indoor air quality to minimize energy use while maintaining air quality.Scheduling HVAC operations: Turning on the HVAC only when its necessary.

[0073] Balancing the impact of specific operation actions on occupant comfort and building performance involves:Predictive control: Employing machine learning models to predict the impact of actions on both comfort and energy use, allowing for proactive adjustments.Adaptive control systems: Implementing control systems that can dynamically adjust settings based on real-time data, ensuring a balance between comfort and efficiency.User feedback: Incorporating feedback from occupants to fine-tune operations and ensure comfort needs are met without compromising efficiency.

[0074] Strategies to ensure both efficiency and comfort are maintained include:Zoning systems: Dividing the building into zones with independent controls, allowing for more precise temperature management and improved comfort.Multi-objective optimization: Search for the best HVAC settings to ensure optimal results for balanced control of energy consumption and comfort.

[0075] The trained predictive models are employed by a control strategy component that evaluates potential actions (in real time) by asking the models to predict the results of these actions. Thanks to the usage of the strategically generated synthetic data, the models can also handle actions that may be radically different from past operation. Therefore, it allows the control strategy component to freely explore potential actions (control novelty). With this freedom, we can employ powerful optimizers (either black-box or problem-tailored) that can go beyond human conventional wisdom.

[0076] As a note, the reader may here observe that the digital twin itself could be directly used by the optimizer to evaluate potential control actions. The first motivation for using machine learning models instead is the speed and efficiency: DT operations (simulation, creation using real data) are more expensive compared to machine learning operations (inferencing, training). This DT cost is more than acceptable as an overhead (considering the benefits), but not in production. Furthermore, periodic re-training in ML model is a standard and well understood practice, while continuously improving the DT would be more complex. In effect, the solution uses the DT for sophisticated and powerful data enrichment.

[0077] In summary, the present disclosure solves the technical problem of the cold start and data drift, and it facilitates novelty by leveraging a digital twin surrogate and synthetic data to enhance the training of ML models. This can enable wider applicability (e.g. new buildings) and better performance (cost, comfort) of building operation.

[0078] Figure 1 shows an embodiment comparing the present disclosure with the Prior Art.

[0079] Figure 2 illustrates the workflow of the current building optimization solution. The process begins with the real building, where data from normal or past operations is collected. This data includes multiple metrics like temperature, humidity, energy consumption and equipment operation settings. A digital twin, a virtual representation of the real building, is then created using this data. The digital twin performs strategic simulations to generate synthetic data tailored to the optimizer, allowing for the exploration of various operational scenarios and novel control strategies. The combined dataset, consisting of both real and synthetic data, is used to train the machine learning models. These models are designed to predict the building's comfort conditions and energy consumptions based on different operational actions. Within the controller, the trained model interacts with control strategy component. The control strategy component evaluates candidate control actions by predicting their expected results using the model's output. This enables the selection of the most effective actions to implement in the real building.

[0080] Figure 3 shows a use-case illustration of how a building operator interacts with the disclosure. The process begins with the building operator providing building information (such as floor plans, HVAC system characteristics, building construction specifications, and equipment schedules), which is then used to create the digital twin. Then they configure the preferences for comfort and savings. These choices can be changed at any time. Following the creation of the digital twin, we can begin generating synthetic data and training prediction models. After the models are complete, we are ready for controlling the building by looking for the ideal settings based on the HVAC system's real-time data. The optimizer objective function relies on the prediction models. Once the optimizer has completed its work, the best- found settings are submitted to the HVAC system for activation.

[0081] The present disclosure can operate autonomously, adjusting HVAC settings to achieve energy savings and improve comfort. Since the model works autonomously, is not required for building operators to perform any action, reducing building operator's workload.

[0082] The building Digital Twin model can simulate any kind of scenario, including different outdoor and indoor conditions. Outdoor refers mainly to weather, ranging from mild to extreme conditions, either historical or synthetic weather data. Indoor conditions refer to building operation. This includes occupancy patterns, equipment usage, and HVAC settings. These scenarios, where different inputs are given to the model, allow to evaluate the response of the HVAC system in terms of energy consumption and also evaluate comfort metrics.

[0083] Testing different controlling strategies using the Digital Twin model allows to access the impact of each strategy on energy costs and comfort. By doing this, we are able to choose the best control strategy for each type of building, location, and usage pattern.

[0084] The controller works autonomously to improve both HVAC performance and comfort, by searching (in real-time, in advance, or any combination of the two) for the combination of actions and equipment operational settings that optimizes energy costs and occupant's comfort. Once the search ends, the best-found solution is applied in the building automatically. By doing this, the model is helping in the process of making informed decisions regarding performance optimization and comfort.

[0085] The present disclosure can operate autonomously, adjusting HVAC settings to achieve energy savings and improve comfort. By giving priority to HVAC energy costs or comfort, the operators can fine-tune the controller. Depending on their demands, they can select the energy-saving vs. occupant-comfort balance.

[0086] Furthermore, the model is able to generate settings that otherwise would not be easily achieved even by a domain expert. This is only possible because synthetic data generation introduces novelty and allows prediction models to be trained in a wide range of unseen combinations of conditions and actions. In this way, we expand the optimizer's effective search space by ensuring that we have prediction models capable of forecasting in such a broad space.

[0087] The "what-if" scenario simulations allow for the creation of comprehensive datasets (synthetic data) that encompass a broad array of potential actions and conditions. By augmenting the limited real-world data with these synthetic data points, we enable models to forecast events that have not yet transpired in reality. As a result, we can identify previously unknown solutions and potentially achieve greater savings. This approach transcends the constraints of using only historical data, providing a rich and expansive resource for training ML models.

[0088] With the "what-if" scenario simulations, the Al controller can adapt to diverse conditions, ensuring robust performance even in scenarios where real-world data is limited or inconsistent. It can also produce novelty that goes beyond the status quo operation approach, leading to innovative solutions. All this contributes to reduce energy costs and improve comfort.

[0089] The present disclosure can reduce HVAC energy costs and enhance occupants' comfort.

[0090] If we consider the majority of modes of operation in the real world, the most advanced options would be operating equipment using predefined seasonal schedules. Compared to these cases, the advantages of the proposed solution can be very high in terms of energy costs reduction and comfort improvements. The ability to continuously and automatically act proactively and adapt to current and foreseeable conditions can outperform any static schedule.

[0091] Considering existing real-time control technologies, the best solution we are aware of is training predictive models using historical data and using them to predict if the building's equilibrium is expected to be disturbed in the near future. This detection can be used proactively; the strategies are not clear but are most likely based on well understood domain experience. In this case, the main advantage of the new solution is that it can address a whole new segment of the market (buildings with no historical data). The additional advantages are that control actions are not confined by human bias and preconceptions (and therefore can produce control novelty), and that the system can better deal with changing conditions.

[0092] The challenges addressed by the solution and the general application is relevant to any company with a similar product. Particularly, the ability to apply the technology to buildings with little to no historical data opens up a part of the market that would otherwise be unworkable. Furthermore, the novelty and adaptability characteristics can improve product performance and customer satisfaction.

[0093] The following pertains to improved optimization through problem representation according to the disclosure. To address the HVAC control optimization problem, a coding scheme can be established to bridge the original problem domain and the search space where the optimization algorithm operates. The set of all candidate and operational control solutions is hereby defined as a decoded operational schedule space. The abstract representation space, in which the optimization process is executed, is hereby defined as the encoded control representation space.

[0094] The proposed method encodes HVAC system control schedules using the encoded control representation - decoded operational schedule abstraction, such as decodedoperational schedule, which corresponds to a complete and detailed operational control schedule.

[0095] The decoded operational schedule is expressed in terms of real, executable control variables, such as setpoint temperatures, On / Off states, and operation modes, e.g., heating, cooling, ventilation, for each HVAC unit or group of units. This representation reflects a time- valid control sequence that respects all operational constraints of the system. In the case of a encoded control representation, it corresponds to a compact, abstract representation of the control schedule, used directly by the optimization algorithm.

[0096] The encoded control representation consists of a sequence of discrete configuration identifiers for each control time step and for each group of devices. Each configuration identifier, e.g., an integer value, maps to a predefined, specific set of control parameters in the decoded operational schedule space, such as illustrated in the table below (this is an exemplary implementation, the boundaries could be different, for example more distinct configurations for higher ranges of temperature settings). These configurations are designed to have a direct correlation with energy consumption, where a higher configuration value corresponds to a higher expected energy usage.

[0097] This separation between encoded control representation and decoded operational schedule allows for a computationally efficient search within a reduced dimensional parameter space and a simplified incorporation of complex operational constraints and dependencies during the encoded control representation-to-decoded operational schedule decoding process, ensuring that any solution generated by the optimizer is always operationally feasible.

[0098] Additionally, the method uses a predictive Free Floating model to determine the main mode of operation (heating or cooling) required for a given optimization horizon. This modelpredicts the evolution of the building's indoor temperature in the absence of any HVAC operation. The result of this prediction is used to restrict the optimizer's search space so that it only selects configurations (encoded control representation) corresponding to the required operation mode (e.g., only heating configurations if the predicted temperature falls below the comfort threshold), further increasing process efficiency.

[0099] This representation structure enhances optimization efficiency and guarantees the feasibility of all generated control plans.

[0100] The following pertains to the disclosed synthetic data generation. To train an accurate surrogate model that predicts the outcomes, e.g., energy consumption, thermal conditions, of a given control schedule without requiring full simulations, a synthetic data generation process is disclosed. This process ensures that the training dataset is both balanced and comprehensive.

[0101] Unlike conventional data generation methods, which may introduce bias by over- representing common operational states, the approach described here explicitly enforces system-level constraints between interconnected devices. Examples of such interconnections include Air Handling Units (AHUs) with their respective Room Indoor Units (RIUs) and Variable Refrigerant Flow (VRF) systems with their Indoor Units (lUs).

[0102] The key features of this process include a dependency-aware sampling, i.e. the data generation is carried out by sampling control configurations for a primary device type and then determining or deriving the state of dependent devices based on predefined operational rules. This procedure guarantees the physical and operational validity of each generated data point. For example: a dependency rule stipulates that if at least one Indoor Unit (IU) of a VRF system has a configuration other than "Off," the corresponding Outdoor Unit (VRF) must be in the "On" state and another rule states that if an AHU Indoor Unit (RIU) is "On," the corresponding AHU configuration cannot be "Off."

[0103] Other key features of this process include a distribution-guided balancing, i.e. during the synthetic data generation process, the primary objective is to ensure a balanced and uniform distribution of operational states across the dataset. To achieve this, the generation process is actively monitored and guided to ensure that the distributions of key aggregate metrics — such as the average configuration value for devices with multiple continuous or discrete control states or the count of On / Off states for devices with binary control states —are kept uniform. This prevents the over-sampling of specific operational patterns and guarantees that the resulting training data provides a comprehensive representation of the entire feasible operational space.

[0104] This constraint-aware and balanced approach significantly broadens the operational space covered in the training set. The result is a more robust, accurate, and generalizable surrogate model, capable of making reliable predictions even for less common control scenarios, which in turn leads to superior HVAC control optimization.

[0105] As an example, considering the on / off and each unit's on / off as a random variable (in {0,1}). If we sample uniformly each variable, then we have a binomial distribution. It has mean (n x p), i.e. the cases where half the units are on is represented heavily. On the other hand, cases where only a few, or most / all units, are on, are hardly ever sampled. That's why we sample uniformly the number of units to be on, and then select randomly which units exactly will be on. Conversely, for the configuration variables, when configuration values for each machine are independently sampled from a uniform distribution, the average of these values across all machines converges towards a normal distribution in accordance with the Central Limit Theorem. That means that we would sample a lot of cases where the combined configurations average in the middle of the range and almost no cases where all units would have the lowest or the highest configuration. Thus, in an aspect of the disclosure, the sampling strategy can be:• For boolean variables: sample uniformly the total number of units to be in the "On" state. Then, select randomly which specific units will be "On" (random drawing without replacement);• For numeric variables:• Uniformly sample the mean configuration value for the set of units under consideration (from a received equipment configuration collection);• Using this mean as the central value (p) and a randomly sampled standard deviation (o), generate individual unit configuration values from a truncated normal distribution whose bounds are adjusted (e.g., [min_value-0.49, max_value+0.49]) to ensure proportional representation of extreme values;Round the generated values to the nearest valid discrete configuration level and assign them to each unit.

[0106] The following pertains to system calibration via the disclosed digital twin modelling workflow. The foundation for generating high-fidelity training data for the surrogate models is preferably a calibrated digital twin of the building. The creation and calibration of this digital twin can be achieved through a systematic, multi-stage workflow that integrates architectural data with energy simulation. The method may comprise the following steps:3D Geometric Model Creation: An accurate and detailed three-dimensional model of the building's geometry, including all spatial zones, structures, and envelope characteristics, is created using computer-aided design software, such as SketchUp;HVAC System Definition and Zoning: The complete HVAC system is defined and configured within a building energy modelling environment, such as OpenStudio. This includes defining all mechanical components, their physical properties, the zoning of the building into distinct thermal areas, and the interconnections between all system components.Control Logic and Schedule Implementation: The operational logic, control sequences, and schedules for the HVAC system are implemented within a dynamic simulation engine, such as EnergyPlus. This step ensures that the simulation reflects realistic operational patterns.Iterative Calibration and Parametric Analysis: A calibration process is conducted, for instance by using an automated tool like OpenStudio-PAT. This process systematically and iteratively adjusts key model parameters, e.g., thermal properties of materials, equipment efficiency, infiltration rates, and compares the simulation outputs against measured, real-world performance data from the actual building. The parameters are refined until the simulation's outputs, e.g., energy consumption, zonal temperatures, achieve a predetermined level of accuracy in matching the measured data.

[0107] This comprehensive workflow is able to ensure that the surrogate models are trained using data from a validated, high-fidelity simulation environment, rather than from purely theoretical or uncalibrated models, thereby substantially increasing the accuracy and real- world applicability of the resulting optimization system.

[0108] The core of the control system is a hybrid optimization framework that couples the afore-mentioned machine learning surrogate models with a probabilistic metaheuristic algorithm optimizer. This approach is chosen over a sole reliance on other methods, such as reinforcement learning (RL), to provide enhanced transparency, controllability, and a mechanism for continuous model improvement.

[0109] An aspect of the present disclosure includes an Interpretable and Decomposable Optimization. The use of a probabilistic meta-heuristic algorithm optimizer, which evaluates solutions using the surrogate model as its objective function, allows for the decision-making process to be inspected and validated. Each candidate solution, decoded operational schedule, tested by the optimizer can be analysed, providing clear insight into the system's logic, which is not readily available in end-to-end black-box models.

[0110] Another aspect of the present disclosure includes an Adaptive Data Space Refinement via Optimizer Feedback. The framework includes a method for automatically improving the quality of the training dataset. During the optimization process, all candidate solutions evaluated by the optimizer are stored. A posterior analysis is then performed, automatically comparing the distribution of these "online" or "in-situ" data points against the distribution of the original synthetic dataset. The surrogate model is then retrained on this enriched dataset, creating a self-improving system that becomes more robust and accurate over time.

[0111] Another aspect of the present disclosure includes an Explicit Multi-Objective Trade-off Control. The framework provides a direct and explicit mechanism for balancing competing objectives, most notably energy consumption and occupant thermal comfort. This may be implemented via a user-configurable balance parameter within the optimizer's objective function. This parameter acts as a weight, allowing a building operator to directly adjust the system's priority between minimizing cost and maintaining comfort, a feature that is not straightforward to implement or control in typical reinforcement learning formulations.

[0112] This hybrid framework is able to ensure a transparent, controllable, and continuously improving optimization system for building energy management.

[0113] The term "comprising" whenever used in this document is intended to indicate the presence of stated features, integers, steps, components, but not to preclude the presence or addition of one or more other features, integers, steps, components or groups thereof.

[0114] It is to be appreciated that certain embodiments of the disclosure as described herein may be incorporated as code (e.g., a software algorithm or program) residing in firmware and / or on computer useable medium having control logic for enabling execution on a computer system having a computer processor, such as any of the servers described herein. Such a computer system typically includes memory storage configured to provide output from execution of the code which configures a processor in accordance with the execution. The code can be arranged as firmware or software, and can be organized as a set of modules, including the various modules and algorithms described herein, such as discrete code modules, function calls, procedure calls or objects in an object-oriented programming environment. If implemented using modules, the code can comprise a single module or a plurality of modules that operate in cooperation with one another to configure the machine in which it is executed to perform the associated functions, as described herein.

[0115] It will be appreciated by those of ordinary skill in the art that unless otherwise indicated herein, the particular sequence of steps described is illustrative only and can be varied without departing from the disclosure. Thus, unless otherwise stated the steps described are so unordered meaning that, when possible, the steps can be performed in any convenient or desirable order.

[0116] A hardware-based computer processor, as used herein, refers to any system, device, or apparatus capable of processing data in accordance with the methods described in this disclosure. The hardware-based computer data hardware-based processor may include one or more hardware-based processors, such as a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), or any combination thereof. These hardware-based processors may be implemented as a single chip, a multi-core processor, a distributed computing system, or any other suitable configuration. For example, this can be a central processing unit (CPU), such as an Intel® Core™ i7 processor, and memory modules, including 16 GB of DDR4 RAM. The system may include a solid-state drive (SSD) for storage, an optional GPU (e.g., NVIDIA® GeForce RTX™ 3060), and runs a standard operating system, such as Microsoft® Windows® or Linux®. For example, this can be an embedded system utilizing a microcontroller, such as the ARM® Cortex®-M4 processor, with onboard memory (e.g., 1 MB of flash storage and 256 KB of SRAM). This system operates with real-time operating system (RTOS) software and can be integrated into an industrial device. For example, this can also be a cloud-based virtual machine hosted on a server infrastructure,such as an Amazon Web Services (AWS) EC2 instance, featuring virtual CPUs (vCPUs) based on Intel® Xeon® or AMD EPYC™ processors. The instance can be configured for example with 32 GB of RAM, 1 TB of elastic block storage (EBS), and executes server-side software designed to perform the computational processes disclosed in this specification.

[0117] The hardware-based computer processor may further include memory (e.g., random access memory (RAM), read-only memory (ROM), flash memory, or other suitable storage devices) for storing instructions and data. The hardware-based processor executes instructions stored in memory to perform the functions described in this specification. The instructions may be implemented in any programming language, including but not limited to assembly language, C, C++, Python, or Java.

[0118] The hardware-based computer processor may communicate with input / output devices (e.g., a keyboard, mouse, touchscreen, or display), peripheral devices, or external systems via wired or wireless connections, such as USB, Bluetooth®, Wi-Fi®, or Ethernet. The hardware-based processor may also be integrated with or connected to a network, including a local area network (LAN), wide area network (WAN), or the internet, to receive and transmit data.

[0119] The disclosure should not be seen in any way restricted to the embodiments described and a person with ordinary skill in the art will foresee many possibilities to modifications thereof. The above-described embodiments are combinable. The following claims further set out particular embodiments of the disclosure.

Claims

C L A I M S1. Computer-implemented method for controlling HVAC, heating, ventilation, and air conditioning, equipment of a building, the method comprising pretraining one or more machine-learning forecaster models for predicting energy consumption and indoor thermal conditions of the building, by: receiving a preexisting dataset comprising sensor data of energy consumption and indoor thermal conditions and comprising equipment operational settings, for one or more volumes of the building; generating synthetic training data by simulating the building using a preexisting digital twin model for obtaining simulated data of energy consumption and indoor thermal conditions and simulated equipment operational settings, for the one or more volumes of the building, wherein the digital twin model is a physics-based model comprising explicit physical quantities and physical relationships between said physical quantities, by sampling operational settings for equipment of a primary device type and determining operational settings using predetermined operational rules for equipment of a dependent device type as a function of the sampled operational settings for the primary device type; and training one or more machine-learning forecaster models using the received preexisting dataset and the simulated data and sampled operational settings; the method further comprising the use of the trained one or more machine-learning forecaster models by a controller by: generating a plurality of sets of equipment operational settings for the one or more volumes of the building; predicting, for each of the generated sets of equipment operational settings, energy consumption and indoor thermal conditions using the one or more machine-learning forecaster models; selecting a generated set of equipment operational settings that maximizes one or more predetermined criteria from the predicted energy consumption and indoor thermal conditions; and sending a signal comprising the selected set of equipment operational settings to be applied to the one or more volumes of the building.

2. Method according to the previous claim, wherein the generation of synthetic training data by simulating the building comprises, for one or more numeric variables of said operational settings and physical quantities: receiving an equipment configuration collection comprising equipment operational settings for each unit of the HVAC, heating, ventilation, and air conditioning equipment; uniformly drawing from a truncated normal distribution whose mean and standard deviation are determined from the received equipment configuration collection.

3. Method according to any of the previous claims wherein the generation of synthetic training data by simulating the building comprises, for Boolean variables of said operational settings of a predetermined operational settings type, uniformly drawing from a uniform distribution the total number of variables in a "true" state and then randomly setting those variables with said "true" state.

4. Method according to the previous claim, wherein the predetermined operational settings type is equipment operational state, wherein the "true" state is "On" and a corresponding false state is "Off".

5. Method according to any of the previous claims, wherein the sampling operational settings for equipment of a primary device type comprises sampling operational settings from a received equipment configuration collection.

6. Method according to any of the previous claims, further comprising simulating the building under a predictive free-floating model in the absence of HVAC actuation, over an optimization horizon, to determine a main mode of operation selected from heating and cooling based on the simulated building relative to a comfort threshold; and constraining the optimizer to generate the plurality of sets of equipment operational settings restrained to the selected main mode of operation.

7. Method according to any of the previous claims, further comprising the steps of: receiving a current dataset comprising sensor data of energy consumption and indoor thermal conditions for the one or more volumes of the building;making a comparison between the received dataset and the predicted energy consumption and indoor thermal conditions for the selected set of equipment operational settings; using the comparison to adjust the one or more machine-learning forecaster models.

8. Method according to any of the previous claims wherein the physics-based model is a heat balance model, a heat transfer function model, a thermal network model, a response factor model, or combinations thereof.

9. Method according to any of the previous claims wherein simulating the building comprises using the preexisting digital twin model for obtaining simulated data of energy consumption and indoor thermal conditions and simulated equipment operational settings, comprises applying equipment operational settings with an increased variance from 75% to 125% in respect of the equipment operational settings of the preexisting dataset.

10. Method according to the previous claim, comprising applying equipment operational settings with an increased variance from 50% to 150% in respect of the equipment operational settings of the preexisting dataset, further in particular an increased variance from 20% to 300% in respect of the equipment operational settings of the preexisting dataset.

11. Method according to any of the previous claims wherein the equipment operational settings include HVAC mode, HVAC on / off, HVAC setpoint, HVAC fan speed, HVAC operation schedules, HVAC temperature setpoints, in particular the HVAC mode comprises HVAC cooling mode, HVAC heating mode, and HVAC fan mode, further in particular the HVAC mode comprises HVAC cooling mode, HVAC heating mode, HVAC fan mode, and HVAC dehumidifier mode.

12. Method according to any of the previous claims wherein the indoor thermal conditions comprise temperature for each of the one or more volumes of the building, in particular the indoor thermal conditions comprise temperature and humidity for each of the one or more volumes of the building.

13. Method according to any of the previous claims wherein the one or more predetermined criteria include energy efficiency.

14. Method according to any of the previous claims further comprising generating an alert or alerts if measured energy consumption or measured indoor thermal conditions deviate from the predicted energy consumption and indoor thermal conditions.

15. Method according to any of the previous claims, further comprises displaying, using a user interface, the selected set of equipment operational settings and, optionally, the predicted energy consumption and / or indoor thermal conditions.

16. Method according to any of the previous claims wherein each of the one or more machine-learning forecaster models is specific to one HVAC mode.

17. A computer program product embodied in a non-transitory computer-readable medium comprising computer program instructions, which when executed by a computer processor, cause the computer processor to carry out the computer-implemented method for controlling HVAC, heating, ventilation, and air conditioning, equipment of a building of any of the claims 1-16.

18. A computer system comprising a computer processor, configured to carry out the computer-implemented method for controlling HVAC, heating, ventilation, and air conditioning, equipment of a building of any of the claims 1-16.

19. A non-transitory computer-readable medium comprising one or more machine-learning forecaster models for predicting energy consumption and indoor thermal conditions to be used for controlling HVAC, heating, ventilation, and air conditioning, equipment of a building, wherein the one or more machine-learning forecaster models have been obtained by: receiving a preexisting dataset comprising sensor data of energy consumption and indoor thermal conditions and comprising equipment operational settings, for one or more volumes of the building;simulating the building using a preexisting digital twin model for obtaining simulated data of energy consumption and indoor thermal conditions and simulated equipment operational settings, for the one or more volumes of the building, wherein the digital twin model is a physics-based model comprising explicit physical quantities and physical relationships between said physical quantities; training one or more machine-learning forecaster models using the received preexisting dataset and the simulated data and settings.

20. The non-transitory computer-readable medium according to the previous claim wherein simulating the building comprises using the preexisting digital twin model for obtaining simulated data of energy consumption and indoor thermal conditions and simulated equipment operational settings, comprises applying equipment operational settings with an increased variance from 75% to 125% in respect of the equipment operational settings of the preexisting dataset, in particular applying equipment operational settings with an increased variance from 50% to 150% in respect of the equipment operational settings of the preexisting dataset, further in particular an increased variance from 20% to 300% in respect of the equipment operational settings of the preexisting dataset.

Images