Method for controlling components of a building

Abstract

The invention relates to a method for controlling components of a building that have an energy consumption and provide an output. Expected output or output intervals are assigned to the components of the building, actual data is requested from individual components of the building at regular intervals, and all of the assigned and requested data is collected in a database. Control variables for the components of the building are calculated from the requested actual data while taking into account the assigned outputs or output intervals in such a way that the total energy consumption of the components of the building corresponds to a specification.

Description

[0001] 01267303-0009 April 2nd, 2026 PCT / DE0S£^|ößffiO19

[0002] Main Post Office

[0003] P05110WO

[0004] 1

[0005] Methods for controlling building components

[0006]

[0001] The invention relates to a method for controlling building components that have energy consumption and provide power, in which expected power outputs or power intervals are assigned to the building components, actual data are queried from individual building components at regular intervals, and all queried data are collected in a database.

[0007]

[0002] Such a method is known from EP 3 267629 Bl. This method is operated with a system in which an analysis program compares actual operating raw data with target operating raw data and transmits control commands to align the actual operating raw data with the target operating raw data.

[0008]

[0003] In practice, this leads to a very high, continuous data transfer and requires a large amount of computing power.

[0009]

[0004] The invention is based on the objective of using an AI-based software solution to save energy costs and CO2, particularly in commercial real estate. Care should be taken to ensure that the savings ultimately exceed the costs of operating the system.

[0010]

[0005] This problem is solved by a method according to claim 1. Advantageous further developments are the subject of the dependent claims.

[0011]

[0006] The solution lies in using an AI-based software solution to save energy costs and CO2 in real estate. This is achieved by calculating control variables for the building components from the queried actual data, taking into account the assigned services or service intervals, in such a way that the total energy consumption of the building components corresponds to a specification.

[0012]

[0007] Ultimately, software for a building should be so well trained that it adapts the control to the building in an energy-optimized manner according to requirements.

[0013]

[0008] It is advantageous if comfort is also improved by optimizing setpoint curves. By optimizing the setpoints for occupied zones, energy costs and CO2 emissions can be saved without sacrificing comfort. Currently, various manufacturers supply building data from heating, ventilation, air conditioning, room and weather data, energy meters, and energy generation systems to a computer in different formats. This is referred to as building management systems. The method according to the invention is not limited to buildings that already have an existing building management system. Buildings without a central building management system can also be integrated into the method by installing the edge component and suitable sensors.

[0014] CONFIRMATION COPY01267303-0010 02.04.2026 PCT / DE0S^S / |QßffiO19

[0015] Main Post Office

[0016] P05110WO

[0017] 2

[0018]

[0009] In particular, ventilation, cooling, and heating use different communication protocols and automation systems than shading, lighting, meters, or photovoltaics. For example, heating, ventilation, and air conditioning systems 3a are typically connected to a building automation system 4a via protocols such as BACnet or Modbus, while shading and lighting 3b are controlled via KNX or DALI with a room automation system 4b. Meters 3c often communicate with an energy management system 4c via M-Bus, and on-site power generation systems such as photovoltaic inverters 3d communicate with a separate controller 4d via Modbus. These sectors are typically operated separately and without coordination in conventional buildings.

[0019]

[0010] The invention provides data to an AI-based controller via an edge component for the control of building components that consume energy and deliver power. This edge component 6 is connected to the network, internet, and power grid in the server room. It is connected via the respective communication protocols to the existing automation systems 4a to 4d of the individual sectors 3a to 3d, as well as to the room sensors 5, and integrates them into a holistic system. The edge component 6 executes the control algorithm.

[0020]

[0011] Preferably, room sensors 5 are installed in each room to measure presence, temperature, air quality, brightness, CO2, and especially humidity in the room. The room sensors 5 are connected to the edge component 6 via a wireless communication protocol, for example LoRaWAN.

[0021]

[0012] The building components can include components from the heating / ventilation / plumbing and electrical systems. It is advantageous if they also include building components from the power generation sector, for example, photovoltaic systems, combined heat and power plants, or grid connections. These energy sources 12 provide data on prices and generation quantities as input variables for the AI ​​control algorithm 9.

[0022]

[0013] The control system can be improved by also collecting data 14 obtained outside the building as queried data in the database. This could include, for example, weather data, weather forecasts, shading of the building, and energy prices.

[0023] «

[0024]

[0014] It is also possible to determine and collect future expectation data, such as a weather forecast, in order to approximate the total energy consumption of the building components to a future-oriented specification of a total energy consumption interval.

[0025]

[0015] The control variables can be varied in a model to approximate the total energy consumption as closely as possible to the target value; this is referred to as training the AI ​​control algorithm.

[0026]

[0016] During the training of the forecast model, the control variables serve, among other things, as a feature set for the supervised learning algorithm. They are varied to capture their influence on the system behavior. Das01267303-0011 02.04.2026 PCT / DE0M^QffiO19

[0027] Main Post Office

[0028] P05110WO

[0029] The model is generated individually for each property; therefore, the feature set corresponds to the actuators, sensors, meter readings, and external influencing factors such as the weather specific to that property. The forecast model then serves as a training exercise for the control algorithm.

[0030]

[0017] It is advantageous if the variation of the data is carried out on a digital twin of the queried.

[0031]

[0018] The requirement can be a minimum of the total energy consumption or it can be variable. If it is variable, for example an energy buffer can be implemented or the energy can be obtained as peak energy demand when electricity is particularly inexpensive.

[0032]

[0019] Sometimes it is advantageous to allow higher energy consumption at a certain time in order to achieve particularly low energy consumption at another time. Therefore, it is proposed to specify the total energy consumption in a time interval as a target.

[0033]

[0020] The building components can be classified using tags. Classified building components can then be related to each other. This makes it possible, for example, to include the function of the components, spatial arrangements, and energy dependencies in the calculation.

[0034]

[0021] Relative expected performance or performance intervals can also be assigned to the building components. "Relative" here means that the expected performance or performance intervals, along with the resulting control variable interval, depend on the actual data or the expected data for the future of another building component.

[0035]

[0022] The control variables for the building components can also be calculated in real time. That is, possibly based on the model.

[0036]

[0023] Preferably, it is a reactive system. Preferably, no calendars

[0024] or static logic are stored. It is proactively controlled in real time, adapted to usage and user behavior.

[0037]

[0025] It is advantageous if all read data points automatically receive a recommendation for labeling based on a plant identification system (BIS) and are automatically transformed into a BACnet object, and current weather data in BACnet is always available for the installation location. For optimal processing, a uniform BIS is automatically assigned to the data points, and the data format is upgraded to a BACnet standard. The Edge component 6 offers interfaces to the common building automation protocols, in particular BACnet, Modbus, M-Bus, and KNX, as well as to other communication protocols such as MQTT and OPC UA.01267303-0012 02.04.2026 PCT / DE0S^S / |QßffiO19

[0038] Main Post Office

[0039] P05110WO

[0040] 4

[0041]

[0026] Preferably, the edge component has two physically separate interfaces for LAN and WAN. It is particularly advantageous if no port forwarding needs to be configured for the edge component. This allows the firewall to continue to be used fully without any configuration. The edge component should be designed to be hardware-fail-safe. This can be achieved, among other things, by installing two power supplies and four redundant, independent computing units (cluster nodes).

[0042]

[0027] With regard to the cloud component, the data connection to the cloud does not require any ports to be opened, meaning that all destinations can still be blocked by your firewall. The cloud can, for example, be hosted in Germany and comply with the FIPS 140-2 standard, which is also used by government agencies and financial institutions to ensure the security of information and data. Furthermore, regular penetration tests (SQL injection, cross-site request forgery, password cracking, broken authentication, and session management) can be performed.

[0043]

[0028] It is also advantageous to train two AI models to optimize building 10. For this purpose, a digital twin 8 of building 10 is first created once in a cloud 11, which serves as a training environment for the reinforcement learning algorithm. Within this training environment, an AI control algorithm 9 is trained and validated.

[0044]

[0029] The method is explained below with reference to the drawing. It shows

[0045] Figure 1 shows the sectors 3a to 3d in conventional buildings, which are not connected via a central system and each has separate communication protocols and automation systems 4a to 4d, which are controlled separately and not in a coordinated manner.

[0046] Figure 2 shows the building technology networked via an edge component 6, in which all sectors 3a to 3d with their automation systems 4a to 4d as well as the room sensors 5 are holistically integrated into an AI-based control system.

[0047] Figure 3, in three partial representations, shows the three-stage process of the one-time creation of a digital twin 8 in a cloud 11, the one-time training of an AI control algorithm 9 through interaction with the digital twin 8, and the continuous real-time control of the building 10 by the edge component 6 with feedback to improve the digital twin 8.

[0048] Figure 4 shows the operational data flow of the process, in which energy sources 12, storage capacities 13, room sensors 5 and external data 14 serve as input variables for the AI ​​control algorithm 9 on the edge component 6, which calculates control variables for the building components of sectors 3a and 3b from this, 01267303-0013 02.04.2026 PCT / DE0S^|ößffiO19

[0049] Main Post Office

[0050] P05110WO

[0051] 5

[0052] Figure 5 shows in two partial representations the conditioning of several zones over a day, wherein the left partial representation shows a conventional continuous conditioning of all zones during operating hours and the right partial representation shows the demand-based conditioning of individual zones according to the invention based on the occupancy detected by room sensors 5.

[0053] Figure 6 shows the structuring and processing of all state, measurement and control variable data by the forecasting service 126, the tagging service 122 and the NILM algorithm 124, as well as the storage of the results in the cloud database 106.

[0054] Figure 7 shows the structure of the digital twin 134 by linking the information of the component tree 123, the counter tree 125 and the forecasting service 126 via the relationship derivation mechanism 130 with the derived feature spaces 131 and actor spaces 132,

[0055] Figure 8 shows the training of distributed control agents 128 of the multi-agent system 127, which interact with the digital twin 134 using their feature spaces 131 and actor spaces 132 and optimize a global objective function according to the specification 129 via the mixing network 133, whereby the trained inference models 135 are transferred to the edge component.

[0056]

[0030] The method is described below using a first embodiment as an example. It begins with the creation of a digital twin of the building. Based on the building data, a model of the building is created in virtual space, which accurately reflects the unique behavior of the building at different times, weather conditions, occupancy levels, and system operations.

[0057]

[0031] This is followed by a one-time training of the AI ​​control algorithm 9. Through interaction with the digital twin 8, the AI ​​control algorithm 9 is trained, which in this process finds the optimal control strategy for the building 10 for each situation. Once the AI ​​control algorithm 9 has finished training, it is installed on the edge component 6.

[0058]

[0032] The edge component 6 enables continuous real-time control and improvement of the AI ​​control algorithm 9. The AI ​​control algorithm 9 is executed on the edge component 6 and issues commands to the building technology depending on the situation in the building 10. New building data helps to refine the digital twin 8 in the cloud 11 and, based on this, to continuously improve the AI ​​control algorithm 9.

[0059]

[0033] Since the AI ​​model is trained using reinforcement learning, it not only copies what is shown, but also independently finds better strategies. 01267303-0014 02.04.2026 PCT / DE0S^|ößffiO19

[0060] Main Post Office

[0061] P05110WO

[0062] 6

[0063]

[0034] In the following, supervised learning and reinforcement learning are explained using a chess game as an example. In supervised learning, the AI ​​model is shown historical games of master players. Each move in the games is assigned a rating indicating whether the move was advantageous or disadvantageous (labeled data sets). By recognizing patterns in the moves, the AI ​​attempts to understand which moves led to a positive rating. This results in the AI ​​model only being able to play moves in a game that it has predicted.

[0064]

[0035] In the reinforcement learning applied according to the invention, the AI ​​begins to play chess on its own without any prior knowledge. After each game, it receives feedback (reward or punishment) based on the outcome of the game. Oriented towards the reward and punishment, the AI ​​independently develops better moves and strategies with each new game. This AI model can even show strategies unknown to master players in a single game.

[0065]

[0036] The self-learning AI replaces manually programmed automation. It learns all dependencies between actuators and external factors and their effects, and controls the building based on this knowledge to achieve target values.

[0066]

[0037] The AI-based control logic learns from collected building data how the building 10 behaves in each zone: Thermal and electrical storage capacities 13, such as thermal heat storage, battery storage, or the thermal inertia of the building itself, are taken into account as input variables. The AI ​​can also make a prediction for future energy production and consumption based on meter data and weather conditions. In this way, the AI ​​can smooth the production and consumption curves.

[0067]

[0038] The presence sensors enable the AI-based algorithm to use energy only for the time when a part of the building is in use. Currently, there is no or only very limited differentiation between individual zones in the building. The entire property is continuously heated, ventilated, and lit during operating hours. In some cases, a central night setback provides initial energy savings.

[0068]

[0039] In the inventive method, room usage is recorded by sensors or other sources, e.g., calendars. The AI ​​recognizes occupancy patterns and controls zones separately according to the expected usage. The AI ​​learns the time a zone needs to heat up / cool down and controls the system so that the zone reaches the setpoint at the desired time and the pre-conditioning phases are minimal. Individual zones 15 can also exhibit permanently lower occupancy patterns, for example, due to home office use.

[0069]

[0040] Continuous monitoring and adaptive control of subsystems enable further efficiency gains. 01267303-0015 02.04.2026 PCT / DE0S^S / |QßffiO19

[0070] Main Post Office

[0071] P05110WO

[0072] 7

[0073]

[0041] Up to now, a system may fail to achieve its purpose in some cases when heating and cooling alternate rapidly and the effects of one process must be counteracted with energy-intensive measures. Now, AI predictions make it possible to avoid these cases by taking into account predicted future needs and weather fluctuations. This allows setpoint and actual value curves to be realized with as few overshoots or undershoots as possible.

[0074]

[0042] In conventional distribution systems, the heating or cooling pump is operated statically, i.e., regardless of the actual consumer demand or the actual heat demand of the heating circuit. Now, the AI ​​recognizes when the pump is actually needed to reach the target temperature. This depends on the actual heating or cooling demand, which is rarely constant.

[0075]

[0043] For example, if n building components with power outputs or power intervals L (n) and actual data I (n) are available and the target V for the total energy consumption of the n building components is a minimum energy consumption Emin, n control variables S (n) must be calculated for the n building components.

[0076]

[0044] For this purpose, all control and regulation systems connected to the building components are first switched off. Then the actual data I (n) and the expected performance values ​​or performance intervals L (n) assigned to the building components are collected in a computer.

[0077]

[0045] The computer can begin its calculation with any control variables. It is advantageous if the current control variables S(n) are transmitted to the computer. In this state, it is not yet known whether the arbitrary or current control variables S(n) lead to a minimum energy consumption Emin. Now, the control variables are varied in a model. If a control variable is, for example, variable in an interval from 1 to 10 and is at 6, it is first increased by, for example, Δ10%, that is, in this case, from 6 to 7. Then the energy consumption El is determined. Then the control variable is decreased by 10%, that is, in this case, from 6 to 5, and the energy consumption E2 is determined. In this way, all control variables are varied up and down by Δ10%, and the energy consumption E is determined for each combination of control variables.Ultimately, the combination of control variables is selected that changes the energy consumption most in the direction of the target value V for Emin.

[0078]

[0046] This determines how the expected performance or performance intervals L (n) assigned to the building components can be achieved with the lowest energy consumption by a specific combination of control variables.

[0079]

[0047] Then, the expected power outputs or power intervals L(n) assigned to the building components are changed, and for different power outputs or power intervals, those control variable combinations are calculated by varying the possible control variable values ​​that allow these power outputs or power intervals to be achieved with the lowest energy consumption. 01267303-0016 02.04.2026 PCT / DE0M^QffiO19

[0080] Main Post Office

[0081] P05110WO

[0082]

[0048] Furthermore, Delta can also be changed generally or even for each control variable to see if changing Delta can change the energy consumption in the direction of the target V for Emin.

[0083]

[0049] It will become apparent how the expected performance or performance intervals assigned to the building components can be achieved with the lowest energy consumption by special control variables.

[0084]

[0050] This optimization is not achieved through directional optimization, but rather by trying out possible combinations of control variables and subsequently evaluating the resulting total energy consumption of the building components and comparing it with the target value. Since this calculation requires a significant computational effort, the best combinations of control variables for specific expected outputs or output ranges are first determined for a particular building in order to later set the optimal combinations of control variables for a specific expected output based on this data.

[0085]

[0051] These control variable combinations, which are usually calculated in a cloud with considerable computational effort, generally also depend on external factors such as the position of the sun or the weather. Therefore, the control variable combinations should also be determined for different external factors.

[0086]

[0052] This makes it possible, for example, to determine from previously calculated values ​​whether the blinds should be opened or the heating turned up in response to a specific increase in room temperature.

[0087]

[0053] The preceding numerical example serves to simplify the illustration of the basic optimization principle. In a preferred embodiment, the calculation of the control variables is not carried out by systematic trial and error, but by a control algorithm trained using reinforcement learning, which learns optimal control strategies based on a digital twin of the building, as described in the following exemplary embodiments.

[0088]

[0054] In another embodiment, an office building with one hundred rooms is controlled by a building automation system with a central database, operated by a control room server. All components, e.g., sensors, actuators, and energy meters in the building automation system, are connected to the control room server via a fieldbus-based communication channel. This makes all steps of the method applicable to all rooms. For clarification, room 15 is described below as an example; all statements apply analogously to the other ninety-nine rooms.

[0089]

[0055] In this embodiment, the term building components refers to all actively controllable units that consume energy within the building and thereby provide usable energy. 01267303-0017 02.04.2026 PCT / DE0S^S / |QßffiO19

[0090] Main Post Office

[0091] P05110WO

[0092] 9

[0093] provide technical performance. In room 15, the building components include in particular a heating, ventilation and air conditioning unit for thermal conditioning, an LED lighting group for artificial lighting and a blind drive for daylight and glare-dependent shading.

[0094]

[0056] A sensor array continuously acquires actual data via a temperature sensor, a CO₂ sensor, a presence detector, a daylight sensor, and a humidity sensor. This data is temporarily stored as a time-measurement tuple in the control room server at ten-second intervals and then transmitted in batches to the database, thereby reducing network traffic.

[0095]

[0057] For each building component, a performance expectation profile is stored in the database, which takes into account time-dependent and usage-dependent patterns. Sources can be manufacturer data, empirical values, or Lem methods. A setpoint module generates an energy consumption target for the entire building for each time step. A control algorithm module retrieves all new actual data and the associated performance expectation profiles within the control interval and first calculates the current total consumption of all rooms. Subsequently, a factor is determined from the ratio of the target to the total consumption, which can be equal to, less than, or greater than one. This ensures that all rooms are assessed proportionally without favoring any individual rooms.

[0096]

[0058] The expected performance of each building component in room 15 is weighted by a factor, from which the following target values ​​are derived. The control variable calculator uses these values ​​to generate a target air volume change for the heating, ventilation, and air conditioning unit, a brightness target for the LED lighting group, and a slat angle for the blind drive. This ensures that the total energy consumption of the building components corresponds to the target, while the room comfort parameters remain within the permissible range.

[0097]

[0059] The calculated control parameters are immediately written to the fieldbus by the control room server, so that the actuators react within a maximum of two seconds. If the presence detector detects an unoccupied state, the system automatically reduces the control parameters to an energy-optimized minimum level.

[0098]

[0060] In a further embodiment, the end-to-end digital networking of an office building consisting of one hundred rooms is described. A control center server implemented as an edge cluster acquires, structures, and processes all status, measurement, and control variable data, synchronizes it with a cloud infrastructure, and forms the basis for a cooperative, AI-supported multi-agent system for controlling room, plant, and energy processes. In addition to connecting the field and automation levels, an event-driven extract-transform-load process is used to semantically enrich and automatically tag data. The resulting structures enable the training of distributed reinforcement learning agents, which optimize the user experience according to comfort and cost criteria.

[0099] Main Post Office

[0100] P05110WO

[0101] 10

[0102] Making control decisions. The following paragraphs explain this architecture, the data processing pipeline, and the learning and operating procedures of the control agents in detail.

[0103]

[0061] An office building 101 with one hundred rooms is fully digitally networked via a building automation system 102.

[0104]

[0062] A central component is a control center server 103, which is set up as an edge cluster and consists of several interconnected computing units. This cluster handles the local data acquisition, structuring, and preprocessing of all status, measurement, and control variable data in a database 104. The event-driven design follows a classic extract-transform-load pattern: In the extract phase, all relevant data points from the building are acquired in real time; in the transform phase, this data is normalized, enriched, formatted, and automatically annotated using a standardized asset label; finally, in the load phase, it is optionally transferred to local storage structures or forwarded to the cloud infrastructure.

[0105]

[0063] During live operation, control and regulation models are executed on the edge cluster. Communication between the control center server, field devices, plant automation, room automation, and the local management level takes place via a communication bus 105 with a standardized communication protocol.

[0106]

[0064] The processed, semantically structured, and partially disaggregated data are continuously synchronized from the local edge cluster to a database 106 on a cloud cluster 107. The use of nearly identical framework and database stacks in edge cluster and cloud environments increases reliability, simplifies maintenance, and, thanks to optimized failover mechanisms, efficient data synchronization, and streamlined deployment processes, enhances system resilience.

[0107]

[0065] Among other things, the computationally intensive training of AI models takes place on the cloud cluster. Furthermore, a web-based building automation software 108 is executed there, which allows authorized users to view the system states, perform evaluations, configure rules and settings for the control models, and configure bus participants.

[0108]

[0066] The connection between the edge cluster and the cloud cluster is encrypted via an SSL connection without port forwarding in the building network.

[0109]

[0067] The room automation is implemented via local room DDCs 109 (Direct Digital Controllers), which independently handle the regulation and control in the rooms and are simultaneously integrated into the building management system.

[0110]

[0068] Room 17 110 is representative of a basic equipped office space in the building. 01267303-0019 02.04.2026 PCT / DE0S^S / |QßffiO19

[0111] Main Post Office

[0112] P05110WO

[0113] 11

[0114]

[0069] Various sensors 111 and actuators 112 are installed in room 17 and are connected to the room DDC. The room sensors measure the room temperature, relative humidity, CO₂ concentration, illuminance, and the presence of people at ten-second intervals. The actuators installed in the room, such as valves, flow controllers, motors, and switches, are controlled in real time by the room DDC.

[0115]

[0070] In the area of ​​primary system supply, control is ensured via system DDCs 113. These process data points from the sensors 114 and actuators 115 of the heating, ventilation and air conditioning systems (HVAC systems). These include, for example, flow and return temperatures, volume flows, differential pressures, frequency converters, pumps and actuators.

[0116]

[0071] In addition, a photovoltaic system 116 is located on the roof, the inverter 117 of which is directly connected to the building automation system via the communication bus. The inverter provides real-time data on the generated electrical power and the feed-in to the grid. The kWh price for self-generated electricity and the price for grid-supplied electricity are manually provided to the building automation system via the user interface of the building automation software.

[0117]

[0072] Consumption meters are installed in the building and are connected to the building automation system via the communication bus. A main energy meter 118 records the total consumption of the building. In addition, HVAC-specific system meters 119 provide differentiated partial consumption readings.

[0118]

[0073] The building is additionally equipped with a local weather station 120. This continuously records the outside temperature, humidity, air pressure, wind speed and direction, rainfall, CO2 concentration, and global radiation. In addition, forecasts regarding cloud cover, sunshine duration, temperature trends, wind development, probability of rain, and amount of rain are obtained from an external weather data source 121 via an internet protocol.

[0119]

[0074] Data acquisition, a tagging service, and a component and counter tree are described in more detail below. Each data point transported by the communication bus is read by the edge cluster, transformed into a uniform data format, provided with metadata and a timestamp, and written to the local database. A rule-based tagging service 122 automatically assigns identifying Haystack labels to the data records. This is done, among other things, based on the metadata of each data point collected from the communication bus. The tagging service uses, among other things, a mapping table that determines which content in the metadata leads to the assignment of predefined tags. Derived from the plant identification system, the service generates a hierarchical component tree 123, which includes at least the levels location, building, floor, zone or room, as well as device group and specific device. 01267303-0020 02.04.2026 PCT / DE0S£^ / |ößffiO19.

[0120]

[0121] ©£-0^-20® Main Post Office

[0122] P05110WO

[0123] 12

[0124]

[0075] In parallel, a NILM (Non-Intrusive Load Monitoring) algorithm 124 is used, which extracts the individual consumption units from the total load input of the energy meters. The characteristic signatures of the load profiles are recognized, assigned to the corresponding rooms, zones, or system components, and then merged into a hierarchical meter tree 125. The resulting meter tree represents the energy consumption-related logical relationships and aggregation levels, e.g., building - ventilation system - motor. A large language model, which has been retrained using building component-specific documents such as motor datasheets, can also be used for the assignment.

[0125]

[0076] Next, a forecasting service 126 and an expectation data generation process are described in more detail. The forecasting service processes local measured values, e.g., from the room sensors, the weather station, as well as forecast data from the external weather source and historical PV yields from the inverter. From this, it generates several-hour to daily expected time series for, e.g., outdoor air conditions, room occupancy probabilities, and PV yields. This expected data is stored as separate objects in the database and continuously updated.

[0126]

[0077] In the event-based building automation system, a cooperative, decentralized multi-agent system 127 is used, whose learning control agents 128 can continuously influence the time series of actuators. Each agent makes decisions locally and in response to events, based on current sensor data (e.g., temperature, CO2, occupancy), its own state space, and a predefined setpoint 129 for energy consumption, thermal and visual comfort, and air quality.

[0127]

[0078] Energy targets can optionally be aimed at minimizing total energy consumption or variable, e.g., minimizing energy costs incurred within a billing period (e.g., 30 days). In this case, the standardized energy flows are weighted with dynamic cost indicators, e.g., cheaper self-generated PV power compared to grid power. The control agents are then trained to achieve a cost-minimal operating profile over a multi-day optimization window.

[0128]

[0079] The number and type of control agents to be trained depend on the building's characteristics. Control agents are trained individually for each controllable zone according to the controllable functional units present in that zone. In the case of room 17, one control agent each is required for temperature, blind, lighting, and ventilation control. The component tree can be used to more quickly determine the control agents for each zone and their feature spaces.

[0129]

[0080] For example, if we take the light control agent to be trained in room 17, it is to be configured in such a way that it is first defined what the target value given in the building automation system is.

[0130] Main Post Office

[0131] P05110WO

[0132] 13

[0133] The desired light level is defined, and which actuators, sensors, and other data sources influence its achievement. Specifically, these include, for example, the occupancy and brightness sensor readings in room 17, as well as the light actuator in room 17. The brightness target value only needs to be reached when a person is in the room, and the light actuator only needs to be activated if the room is not already naturally bright enough.

[0134]

[0081] To avoid having to perform the assignment manually, an automated relationship derivation mechanism 130 is used. This mechanism analyzes the hierarchies represented in the component tree as well as the metadata automatically assigned to each data point in the previous steps. Based on this, functional relationships between sensors, actuators, and setpoints are automatically recognized.

[0135]

[0082] For example, if a room is assigned a temperature sensor, a corresponding setpoint, and a heating valve as an actuator, a controllable functional unit for temperature control is automatically derived. This structure arises solely from the position and type of the data points in the component tree, without the need for manual assignment.

[0136]

[0083] The derivation mechanism generates both a feature space 131, which includes all relevant input variables, and an actor space 132, which describes the associated controllable output variables. These then form the basis for training the control agents. They define which information an agent can process during training and which actors it may influence in its decision-making.

[0137]

[0084] The agents are trained using reinforcement learning (RL), for example, with a QMIX algorithm. This is a method in which several decentralized agents are trained simultaneously to jointly optimize a global objective function. Each control agent has its own local evaluation function (Q-function), which assesses the expected benefit of different actions based on local sensor data. The agents are responsible for individual areas. The local Q-values ​​of the individual agents are combined in a central mixing component 133 (mixing network), which ensures that all agents jointly optimize a global objective function, subject to the constraint that comfort in the respective zones is maintained at all times or during occupancy.

[0138]

[0085] The training is simulation-based and takes place over many iterations using a dynamic building model or digital twin 134. This model precisely simulates the dynamic interactions between building-relevant physical processes and external influencing factors such as weather conditions and usage patterns. During the training process, each agent receives states and state transitions from this model, which are generated from sensor data, actuator actions, forecast data, and external influences. Based on this data, the agent selects actions, receives from 01267303-0022 02.04.2026 PCT / DE0S^|6ffiffiO19

[0139] Main Post Office

[0140] P05110WO

[0141] 14

[0142] The model provides feedback on state changes and resulting rewards, and iteratively updates its decision function.

[0143]

[0086] The digital twin is a data-driven model that precisely simulates building-related physical processes and external influencing factors. This model is technically implemented, for example, using neural differential equations whose parameters are initially determined from historical time series using supervised learning. The results of the relationship derivation mechanism are also used here.

[0144]

[0087] During the training phases, the control agents actively interact with the digital twin by selecting control variables as actions for their actuators and having the digital twin immediately simulate their effects. The digital twin then provides feedback in the form of updated states and resulting rewards, depending on the approach to the energy target and adherence to the comfort limits. This feedback enables the agents to iteratively improve their local decision-making functions and ultimately ensure global optimization of the target specifications.

[0145]

[0088] After completion of the training, the inference models 135 of the control agents are containerized, transferred from the cloud environment to the edge cluster and automatically deployed there to make decisions locally and in real time.

[0146]

[0089] In real-time operation, the trained multi-agent system reacts in an event-driven manner to changes in sensor data. If a sensor value changes, e.g., a change in presence, it is checked whether the changed sensor value is part of the feature space of a control agent. If this is the case, its inference model is activated to calculate a local control action.

[0147]

[0090] Each derived action is not initially output directly to the actuators, but is entered into a shadow time series 136 as a planned target value. This shadow time series allows for short-term forecasting and adjustment of the control actions. After a validation phase, in which, for example, it is checked whether the planned action is globally meaningful and does not lead to undesired interactions between different agents, these planned actions are finally written to the DDCs via the communication bus, which in turn control the actuators accordingly.

[0148]

[0091] Both the dynamic model and all control agents are regularly retrained in the cloud using current real-time data from the building as well as collected training information such as states, actions taken, and rewards received. After this periodic retraining, the updated models are automatically redeployed to the edge cluster. 01267303-0023 02.04.2026 PCT / D E0©££j© / |ÖQffiO19

[0149] Main Post Office

[0150] P05110WO

[0151] 15

[0152]

[0092] To ensure operational reliability, all programs with control logics, e.g. PI / PID, other programs located on the DDCs, are persisted on the edge cluster and automatically ported back to the device in case of error.

[0153]

[0093] A third embodiment is explained below with reference to Figures 6 to 8.

[0154]

[0094] This embodiment describes a hospital 101 with two hundred rooms, which has areas of varying criticality: operating rooms, intensive care units, patient rooms, administrative areas, and technical rooms. The building is networked via a building automation system 102, whose control center server 103 is implemented as an edge cluster with four redundant computing units. All status, measurement, and control variable data are recorded in a local database 104 and synchronized via an encrypted connection with a database 106 on a cloud cluster 107. A communication bus 105 connects all field devices, room DDCs 109, and plant DDCs 113 to the control center server 103. Web-based building automation software 108 enables the building services personnel to configure and monitor the entire system.

[0155]

[0095] The hospital has building components for heating, ventilation, plumbing, and electrical systems. In addition, there are building components for power generation: a photovoltaic system 116 on the roof with an associated inverter 117, a gas-fired combined heat and power plant for simultaneous electricity and heat generation, and an emergency power supply system with diesel generators. A main energy meter 118 records the hospital's total energy consumption, supplemented by system-specific meters 119 for the HVAC systems, lighting, medical equipment, and power generation systems. The sensors 114 and actuators 115 of the HVAC systems are controlled via the system DDCs 113.

[0156]

[0096] Sensors 111 are installed in each room, which record room temperature, relative humidity, CO₂ concentration, illuminance, and the presence of people at ten-second intervals. In the operating rooms, particle counters and differential pressure sensors are also monitored. The actuators 112 installed in the room, for example, valves, volume flow controllers, dimming actuators, and blind drives, are controlled in real time by the respective room DDC 109. The control variables for the building components are calculated in real time.

[0157]

[0097] A local weather station 120 continuously records outside temperature, humidity, air pressure, wind speed, global radiation, and rainfall. Additionally, forecasts for temperature trends, cloud cover, sunshine duration, and probability of rain are obtained from an external weather data source 121. Thus, data collected outside the building are also stored as queried data in the database 104. 01267303-0024 02.04.2026 PCT / DE0M^QffiO19

[0158] Main Post Office

[0159] P05110WO

[0160] 16

[0161]

[0098] A rule-based tagging service 122 automatically assigns identifying labels to each read data point based on the metadata collected from the communication bus 105. The data points are organized into a hierarchical component tree 123, which includes the levels location, building, floor, functional area, zone or room, device group, and specific device. In parallel, a NILM algorithm 124 extracts the individual consumption units from the total load input and organizes them into a meter tree 125. The classified building components are thus categorized with tags and related to one another.

[0162]

[0099] A forecasting service 126 processes local measured values, forecast data from the external weather data source 121, and historical generation and consumption data. From this, it generates multi-hour to daily expected data for outdoor air conditions, room occupancy probabilities, PV yields, and CHP heat output. The expected data for the future are determined and collected in the database 104 in order to approximate the total energy consumption of the building components to a future target of a total energy consumption interval. Furthermore, the forecasting service 126 incorporates the operating room occupancy plan from the hospital information system to predict the energy demand for air conditioning the operating rooms.

[0163]

[0100] In the hospital, a cooperative, decentralized multi-agent system 127 is used, whose learning control agents 128 continuously influence the control variables of the actuators. Each agent makes decisions locally and event-driven, based on current sensor data, its own state space, and a predefined setting 129.

[0164]

[0101] A special feature of this embodiment example lies in the assignment of relative expected performance or performance intervals to the building components. Relative here means that the expected performance or performance intervals of a building component, with the resulting control variable interval, depend on the actual data or the expected data for the future of another building component. Specifically, this is implemented as follows:

[0165]

[0102] The operating room, room 42 110, is primarily required to maintain a room temperature of 18-24 °C, a relative humidity of 30-65%, and a particle concentration below defined limits according to DIN 1946-4. These requirements are assigned the highest priority. The expected performance of the air handling system in the operating room is absolute: it does not depend on other building components, but solely on the hygienically required minimum air exchange rate.

[0166]

[0103] The expected power outputs of the adjacent recovery room, the adjoining patient rooms on the same floor, and the administrative area on the ground floor are allocated relative to the energy demand of the operating room. If the forecasting service 126 determines, based on the operating room occupancy plan, that the operating room will need to be operated with full air conditioning for the next two hours, the power intervals of the dependent zones are automatically adjusted: Das01267303-0025 02.04.2026 PCT / DE0S^S / |QßffiO19

[0167] Main Post Office

[0168] P05110WO

[0169] 17

[0170] The temperature range of the patient rooms is extended by up to 1.5 K, the illuminance in the administrative area is reduced, and the air volume flow in the recovery room is reduced to the minimum permissible fresh air supply. This creates a dynamic equilibrium in which the overall energy balance remains within the performance range of specification 129 without falling below the minimum requirements in any zone.

[0171]

[0104] In this embodiment, specification 129 is designed to be variable. It is not defined as a static minimum of total energy consumption, but rather as total energy consumption over a time interval corresponding to the hospital's shift schedule. For each eight-hour shift, an energy budget is specified based on the predicted occupancy profile, the planned operations, the weather data, and the expected output of the power generation facilities.

[0172]

[0105] Within the time interval of a shift, the total energy consumption can temporarily exceed the long-term optimum, provided it is compensated for by a later phase with lower consumption within the same time interval. For example, during the night shift, when only a few operating rooms are active and the PV system is not generating electricity, the combined heat and power plant is operated at increased output to charge the building's thermal storage mass. In the subsequent day shift with high PV feed-in, the heating output is reduced and the self-generated electricity share is maximized. The control agents 128 are trained to pursue this cross-shift optimization goal.

[0173]

[0106] The control variables are varied in a model to approximate the total energy consumption as closely as possible to the target value 129. The variation of the data is carried out on a digital twin 134 of the hospital. The digital twin represents the thermal inertia of the different building areas, the interactions between adjacent zones, the heat output of medical devices, and the dynamic load profiles of the power generation plants. It is modeled using neural differential equations, the parameters of which are determined from historical time series.

[0174]

[0107] An automated relationship derivation mechanism 130 analyzes the component tree 123 and the assigned tags to automatically recognize functional relationships between sensors, actuators, and setpoints. From these relationships, feature spaces 131 and actuator spaces 132 are derived for each control agent 128. The relationship derivation mechanism also recognizes the cross-zone dependencies that result from the relative power assignments: The feature space of the temperature agent for a patient room additionally contains the current operating state of the operating room on the same floor.

[0175]

[0108] The control agents 128 are trained using reinforcement learning, whereby a QMIX algorithm combines the local evaluation functions of the individual agents in a central mixing component 133. The training is simulation-based via the digital twin 134 under 01267303-0026 02.04.2026 PCT / DE0S^S / |QßffiO19

[0176] Main Post Office

[0177] P05110WO

[0178] 18

[0179] Consideration of relative power allocations: The reward function not only evaluates adherence to local comfort limits and approximation to the energy target, but also penalizes violations of relative dependencies, i.e., when a control agent of a dependent zone exceeds its power interval while the prioritized zone demands its maximum capacity.

[0180]

[0109] After training is complete, the inference models 135 of the control agents are containerized, transferred from the cloud environment to the edge cluster, and deployed there. In real-time operation, the trained multi-agent system reacts to changes in sensor data in an event-driven manner. Each derived control action is initially entered into a shadow time series 136 as a planned setpoint. In a validation phase, it is checked whether the planned action complies with the relative performance dependencies and does not lead to undesired interactions between agents. Only after successful validation are the manipulated variables written to the DDCs via the communication bus 105.

[0181]

[0110] A specific scenario illustrates the relative power allocation: In the morning, three operating rooms are active simultaneously. The forecasting service 126 predicted this based on the operating room occupancy plan and adjusted the expected data accordingly. The control agents 128 of the operating rooms demand maximum air conditioning power. Since the expected power outputs of the patient rooms, the recovery area, and the administration are allocated relative to the resulting energy demand, their control variables are dynamically adjusted: The blinds in the patient rooms are opened to utilize passive solar gains for heating; the ventilation of the unoccupied administration rooms is reduced to minimum air exchange; the lighting in corridors is controlled by occupancy. At the same time, the control agent of the combined heat and power plant increases its output to cover the increased electricity demand of the operating rooms and to utilize the waste heat for hot water preparation.

[0182]

[0111] In the afternoon, when only one operating room is active, the relative power allocations return to their normal state. The freed-up energy capacity is used to bring the thermal storage mass of the patient rooms back to the optimal setpoint. The photovoltaic system generates high yields at this time of day, so the self-generated electricity share is maximized and the combined heat and power plant is throttled back. This balancing ensures compliance with specification 129 for the eight-hour time interval of the day shift.

[0183]

[0112] Both the digital twin 134 and all control agents 128 are regularly retrained with current real-time data. By integrating the operating room scheduling plan into the forecasting service 126, the forecast quality continuously improves, as seasonal patterns, recurring surgery types, and their specific energy requirements are learned. To ensure operational reliability, all programs with control logic are persisted on the edge cluster and automatically backported to the DDCs in case of an error. 01267303-0027 02.04.2026 PCT / DE0M^QffiO19

[0184] Main Post Office

[0185] P05110WO

[0186] Reference symbol list

[0187] 3a Sector HVAC (Heating, Ventilation, Cooling) 109 Room DDC

[0188] 3b Sector Room Automation (Shading, Be¬ 110 Room 17)

[0189] (lighting) 111 sensors in the room

[0190] 3c sector counter, 112 actuators in the room

[0191] 3D Sector: Electricity generation (photovoltaics)

[0192] 4a Plant automation (HVAC) 113 Plant DDC

[0193] 4b Room automation 114 Sensors HVAC systems

[0194] 4c Energy Management System 115 Actuators HVAC Systems

[0195] 4D control of self-generated power systems 116 Photovoltaic system

[0196] 5 room sensors, 117 inverters

[0197] 6 Edge Component 118 Main Energy Meter

[0198] 8 Digital Twin 119 HVAC System Meter

[0199] 9 AI control algorithm

[0200] 10 buildings, 120 local weather stations

[0201] 11 Cloud 121 External weather data source 12 Energy sources (PV, CHP, grid electricity) 122 Tagging service

[0202] 13 Storage capacities (battery, thermal, 123 Component tree building) 124 NILM algorithm

[0203] 14 External data (weather, forecasts, 125 counter tree)

[0204] Energy prices

[0205] 126 Forecast Service

[0206] 15 zones with reduced occupancy (e.g.

[0207] Home office) 127 Multi-agent system

[0208] 128 control agents

[0209] 129 specification

[0210] 130 Relationship derivation mechanism lOl Office building 131 Feature space

[0211] 102 Building automation system 132 Actuator space

[0212] 103 Control Center Servers / Edge Clusters 133 Mixing Network

[0213] 104 Database Edge Cluster 134 Digital Twin

[0214] 105 Communication bus 135 Inference model

[0215] 106 Database Cloud Cluster 136 Shadow Time Series

[0216] 107 Cloud Clusters

[0217] 108 Building automation software

Claims

01267303-0029 April 2, 2026 PCT / DE0M^QffiO19 Main Post Office P05110WO 20 Patent claims:

1. Methods for controlling building components that consume energy and provide performance, where expected performance levels or performance intervals are assigned to the building components, Actual data is queried from individual building components at regular intervals and All assigned and queried data are collected in a database, characterized by the fact that From the queried actual data, taking into account the assigned services or service intervals, control variables for the building components are calculated in such a way that the total energy consumption of the building components corresponds to a specification.

2. Method according to claim 1, characterized in that the building components comprise components from the areas of heating / ventilation / sanitary and electrical systems.

3. Method according to claim 2, characterized in that the building components also include building components from the field of power generation.

4. Method according to one of the preceding claims, characterized in that data determined outside the building are also collected as queried data in the computer.

5. Method according to one of the preceding claims, characterized in that expected data for the future are determined and collected in the computer in order to approximate the total energy consumption of the building components to a future-oriented specification of a total energy consumption interval.

6. Method according to one of the preceding claims, characterized in that the control variables in a model are varied in order to approximate the total energy consumption as closely as possible to the target value.

7. Method according to claim 6, characterized in that the variation of the data is carried out on a digital twin of the queried person.

8. Method according to one of the preceding claims, characterized in that the target is a minimum of the total energy consumption. 01267303-0030 02.04.2026 PCT / DE0S^|ößffiO19 Main Post Office P05110WO 21 9. Method according to any one of the preceding claims 1 to 7, characterized in that the setting is variable.

10. Method according to one of the preceding claims, characterized in that the specification is the total energy consumption in a time interval.

11. Method according to one of the preceding claims, characterized in that the building components are classified with tags.

12. Method according to claim 11, characterized in that classified building components are placed in relation to each other.

13. Method according to one of the preceding claims, characterized in that relative expected performances or performance intervals are assigned to the building components.

14. Method according to one of the preceding claims, characterized in that the control variables for the building components are calculated in real time.

15. Method according to one of the preceding claims, characterized in that the control variables are calculated by a decentralized multi-agent system (127), wherein for each controllable zone of the building, control agents (128) are individually trained according to the controllable functional units present in the zone, each making decisions locally and based on its own state space.

16. Method according to claim 15, characterized in that each control agent ( 128) has its own local evaluation function and the local evaluation functions of the individual control agents are combined in a central mixing component (133) to jointly optimize a global objective function.

17. Method according to one of the preceding claims, characterized in that individual consumption units are extracted from the total load feed-in of energy meters by means of a NILM algorithm (124), assigned to the corresponding rooms, zones or system components and placed in a hierarchical meter tree (125).

18. A method according to one of the preceding claims, characterized in that an automated relationship derivation mechanism (130) analyzes the hierarchies represented in a component tree (123) as well as the automatically assigned meta-information and recognizes functional relationships between sensors, actuators, and setpoints, wherein a feature space (131) and an actuator space (132) are derived from the recognized relationships for each control agent (128). 01267303-0031 02.04.2026 PCT / DE0M^QffiO19 Main Post Office P05110WO 22 19. Method according to claim 6 or 7, characterized in that the model and the control agents (128) are regularly retrained with current real-time data from the building and the updated inference models (135) are automatically transferred to the control center server (103) after retraining.

20. Method according to claim 9, characterized in that the variable setting is geared towards minimizing the energy costs incurred within a billing period, wherein the normalized energy flows are weighted with dynamic cost indicators.

21. System for controlling building components that consume energy and provide performance, comprising: a plurality of building components, each with associated sensors (111, 114) and actuators (112, 115), a control center server (103) implemented as an edge cluster, which is connected to the building components via a communication bus (105), a database (104) located on the control center server (103) for collecting actual data from the sensors and expected performance or performance intervals assigned to the building components, at least one control algorithm executed on the control center server (103) which calculates control variables for the building components from the actual data, taking into account the assigned powers or power intervals, such that the total energy consumption of the building components corresponds to a specification (129), and wherein the calculated control variables are transmitted to the actuators via the communication bus (105).

22. System according to claim 21, characterized in that the control center server (103) has several independent computing units and includes physically separate interfaces for a local network and a wide area network.

23. System according to claim 21 or 22, characterized in that the system further comprises a cloud cluster (107) with a database (106), wherein the data of the control center server (103) are synchronized with the cloud cluster (107). 01267303-0032 02.04.2026 PCT / DE0M^QffiO19 Main Post Office P05110WO 23 24. System according to one of claims 21 to 23, characterized in that the synchronization between control center server (103) and cloud cluster (107) takes place via an encrypted connection without port forwarding in the building network.

25. System according to one of claims 21 to 24, characterized in that AI models are trained on the cloud cluster (107) and the trained inference models (135) are transferred to the control center server (103) after completion of the training and executed there.

26. System according to one of claims 21 to 25, characterized in that the inference models (135) are containerized, transferred to the control center server (103) and automatically deployed there.

Images