Controlling supply technology of a building by means of a neural network

Abstract

The invention discloses a method for training a neural network (12; 112) which is designed to be used to control supply technology (30; 50) of a building (40). The following values are provided in the method: system parameters (18) of the supply technology; building data (20) relating to the building; environmental data (22) relating to the building; and control data which can be used to control the supply technology. The system parameters (18), building data (20), environmental data (22) and control data provided are preprocessed in order to generate a training data set (26; 60, 62, 64, 66) for the neural network (12; 112), which training data set comprises the preprocessed system parameters (18), building data (20) and environmental data (22) as input data (24) and the preprocessed control data as output data. The neural network (12; 112) is trained using the training data set in order to learn a relationship between the input data and the output data.

Description

[0001] Control of a building's supply technology using a neural network

[0002] Technical area

[0003] The present invention relates to a method for training a neural network designed to be used to control the supply technology of a building. Furthermore, the invention relates to a method for using the neural network to control the supply technology of the building. The invention also relates to a computer program product and a control device for controlling the supply technology of a building using a control data set.

[0004] State of the art

[0005] The energy supply of buildings accounts for a significant share of global energy consumption. In recent years, various methods have been developed to thermally model buildings and control their supply technology in an energy-optimized manner. One such method is Model Predictive Control (MPC), which creates a discrete-time dynamic model of the process to be controlled. This approach can therefore also be referred to as "hard control." Furthermore, the control of a building's supply technology can also be achieved using a conventional control system such as a proportional-integral-derivative (PID) controller. However, the control and optimization methods used are very computationally intensive and time-consuming.For this reason, thermal modeling, for example, is only performed at isolated times during the day, which means that the supply technology is not always optimally controlled. Alternatively, a considerable computational effort is required, with an entire server required for the calculation for individual buildings. The effort involved in this approach is correspondingly high and limits the achievable efficiency gains. Description of the invention.

[0006] In a first aspect, the invention relates to a method for training a neural network designed to be used to control a building's utility system. The neural network can provide improved thermal building modeling without requiring the exact structure of the control system and / or the system to be controlled to be determined beforehand.

[0007] A neural network can, for example, be a mathematical model. The neural network can be created using a computer and designed as an artificial neural network. The neural network can have input nodes, output nodes, and a plurality of intermediate nodes arranged between the input nodes and the output nodes. The input nodes can, for example, be designed as data interfaces via which input data can be entered into the neural network. The output nodes can, for example, be designed as data interfaces via which output data can be output from the neural network. The input nodes can be connected to the intermediate nodes, and the intermediate nodes can be connected to one another. The intermediate nodes can be connected to the output nodes.

[0008] Information can be buffered, at least temporarily, on the intermediate nodes. It can be provided that at least one arithmetic operation is carried out on the intermediate nodes. The input data can be transferred from the input nodes via the intermediate nodes to the output nodes, for example, in order to generate the output data there. During this transfer, the input data can be processed, for example converted into output data. The intermediate nodes of the neural network can be arranged in one or more layers or levels. The intermediate nodes can be interconnected within a layer. In addition, the intermediate nodes of one layer can be connected to the intermediate nodes of other layers. The individual connections between the input nodes, the intermediate nodes, and the output nodes can be provided with mathematical weights.The connections between the individual nodes of the neural network can be weighted. During training of the neural network, the weights can be changed or even initially specified. By adjusting the weights of the connections between the individual nodes during training, the neural network can learn a relationship between the input data and the output data. During use of the neural network for its intended purpose, the neural network can apply the learned relationship to the input data to generate output data according to the specified purpose of the neural network.

[0009] For training the neural network, the long short-term memory (LSTM) technique can be used, for example. This involves first generating a training signal from the input nodes to the output nodes. Error adjustment is then performed in the opposite direction, i.e., from the output nodes to the input nodes, to retroactively adjust the weighting of the individual connections. For each node, the input value, the stored value, and the output value are recorded to ensure that all nodes are adequately considered during error correction.

[0010] Alternatively or additionally, the technique of generative adversarial networks (GAN) can be used to train the neural network. This involves two artificial neural networks, one of which creates a selection of possible candidates for the desired outcome. The other neural network evaluates these candidates based on the provided training data. Furthermore, a convolutional neural network (CNN), for example, can be used, in which the intermediate nodes of different layers are at least partially connected by a convolution function.

[0011] A building can be any type of structure permanently attached to the ground designed to house people and / or goods. For example, a building can be a warehouse, an office building, a factory building, a residential building, a sports hall or swimming pool, a school building, a retail building, and / or a mixed-use building consisting of a combination of the aforementioned building types.

[0012] A building's supply technology can be a system by means of which heat input into the building can be influenced. Heat input can involve heating or cooling the building. For example, the supply technology can have one or more heaters, one or more air conditioning systems and / or other temperature control devices, such as automatically opening windows and / or adjustable sun protection. The supply technology can influence the thermal state of the building, in particular to specify a thermal state. A thermal state of the building can, for example, comprise a temperature, optionally taking into account air humidity and / or air pressure within the building. In addition, the thermal state can also relate to a temperature gradient, i.e. a cooling rate or a heating rate.The supply technology can be designed to adjust additional climate parameters within the building, such as fresh air supply, floor temperature and / or sun protection. Controlling the supply technology results in instantaneous energy consumption. Energy consumption can include, for example, the consumption of electricity and / or fossil fuels. The goal of controlling the supply technology can be to minimize this energy consumption or the respective costs, in particular those caused by this energy consumption. Maintenance costs can also be taken into account. Boundary conditions, such as a desired indoor temperature range, can be taken into account during control. The boundary conditions can be fixed values ​​or determined depending on additional parameters, such as a time of day and / or the usage status of the building.

[0013] The method comprises a step of providing system parameters of the supply technology. The system parameters can have one or more values. The system parameters can define a thermal state of the supply technology. For example, the system parameters can indicate which air temperature prevailed, prevails, or should prevail inside the building at a specific point in time. The method further comprises a step of providing building data of the building. The building data can have one or more values. The building data can define a structural state of the building. For example, the building data can be a characteristic value for the insulation of the building. Suitable parameters for this are, for example, the age of the building, the respective wall thicknesses of the building and / or the size and / or orientation of the respective window areas.The method further comprises a step of providing environmental data of the building. The environmental data can have one or more values. The environmental data can define a thermal state of the building's environment. For example, the environmental data can be a characteristic value for precipitation or solar radiation on the building. In addition to weather conditions, the environmental data can also be a geographical position of the building as well as a positioning relative to other buildings and / or a floor design. For example, a building on a street can be partially shielded from wind and solar radiation, which can be taken into account by the environmental data. The environmental data can contain corresponding information.

[0014] The method further includes a step of providing control data that can be used to control the utility system. The control data can define an operating state of the utility system. For example, the control data can specify that the utility system generate a specific heating or cooling output in order to change the thermal state of the building. Likewise, the control data can specify, for example, a window position and / or a sunshade position.

[0015] During provision, those system parameters, building data, environmental data, and / or control data that optimize at least one property of the supply technology can be selected from a plurality of data sets. For example, those system parameters, building data, environmental data, and / or control data that optimize energy consumption of the supply technology can be selected from the plurality of data sets. Alternatively or additionally, those system parameters, building data, environmental data, and / or control data that optimize temperature and / or humidity in the building can be selected from the plurality of data sets.

[0016] When the system parameters, building data, environmental data, and control data are provided, at least one value of the aforementioned data can be determined. The at least one value can be recorded, measured, and / or synthetically generated. The provided system parameters, building data, and environmental data can be transmitted, for example, by means of an input device to at least one of the input nodes of the neural network. The provided control data can be transmitted, for example, by means of the input device or a further input device to at least one output node of the neural network. The input device and the further input device can be part of a control device of the building or can be in signal communication with the control device.The input device and the further input device can each have a data interface via which the provided system parameters, building data and environmental data are transmitted to at least one of the input nodes and via which the provided control data are transmitted to at least one of the output nodes of the neural network.

[0017] The method further comprises a step of preprocessing the provided system parameters, building data, environmental data, and control data to generate a training data set for the neural network. The training data set includes the preprocessed system parameters, the preprocessed building data, and the preprocessed environmental data as input data, and the preprocessed control data as output data.

[0018] The provided system parameters, building data, environmental data, and control data can be preprocessed so that they can be processed more quickly or even meaningfully at all by the neural network. Preprocessing can include selecting whether a value of the data is used as input data or output data for training the neural network. For example, a value of this data may not be suitable as input data for training the neural network if it has no or insufficient influence on the control of the utility technology. By selecting this option, the amount of data can be significantly reduced and more robust optimization can be achieved.

[0019] Alternatively or additionally, a data value may not be suitable as input data for training the neural network if, for example, it cannot be processed by the neural network due to its data structure. For example, the provided data may be in the form of a multidimensional matrix. However, it may be provided that the neural network to be trained is designed to process one-dimensional numerical values. In this case, the multidimensional matrix must be decomposed into one-dimensional numerical values ​​before it can be used to train the neural network. Furthermore, preprocessing may also include sorting the provided data and / or removing a subset of the provided data in order to reduce the amount of data used to train the neural network.

[0020] The method further comprises a step of training the neural network with the training data set to learn a relationship between the input data and the output data. While processing the input data, the neural network can change the weighting of the connections of the individual nodes. In this way, the neural network can learn the relationship between the input data and the output data.

[0021] The proposed method significantly reduces the computing time required to optimize thermal building control. This allows even short-term events, such as weather changes or energy price fluctuations, to be taken into account during optimization. Furthermore, the trained neural network can be used not only to generate individual control parameters for controlling the building's utility technology, but also to generate initial values ​​for an optimization algorithm, particularly an analytical and / or simulation-based optimization algorithm. Thus, the proposed method can be easily integrated into existing optimization routines, significantly reducing the effort required.For example, the use of a neural network eliminates the need for a high-performance data center, which was previously necessary to optimize building control systems. This also reduces building control costs. Finally, the improved optimization routine for building control systems allows the energy required for building control systems to be reduced and / or adapted to energy price trends. Therefore, the proposed method also reduces energy costs. Finally, when training the neural network, only the input and output data are taken into account. The structure of the neural network, on the other hand, can be irrelevant or can be generated during training. The neural network used can therefore be a "black box" defined solely by the input and output data.The proposed method for training a neural network can therefore be applied to a variety of different types of neural networks.

[0022] According to one embodiment of the method for training a neural network, the provided system parameters, building data, environmental data, and / or control data are historical values ​​and / or are synthetically generated. Historical values ​​can include values ​​recorded at a specific point in time in the past. The historical values ​​can have been recorded for the building, for the building's surroundings, and / or for other buildings. The synthetic data can be generated by processing recorded and / or measured data. Alternatively or additionally, the synthetic data can be generated by processing predetermined parameters of the building. Alternatively or additionally, the synthetic data can be generated, for example, by a simulation.Due to the different types of data provided, the proposed method can be adapted to a variety of building types and / or building environments.

[0023] According to one embodiment, the provided system parameters, building data, environmental data and / or control data relate to sub-areas of the building, sub-areas of the supply technology, multiple buildings and / or the respective supply technology of multiple buildings. Sub-areas of the building can, for example, be individual rooms of the building. Alternatively or additionally, sub-areas of the building can, for example, be individual floors of the building. Sub-areas of the supply technology can, for example, be individual supply elements of the supply technology, such as a heating system and / or a ventilation system. For example, one wing of the building, which forms a first sub-area, can be supplied with heat by a first heating pipe as a supply element, and another wing of the building, which forms a second sub-area, can be supplied with heat by a second heating pipe as a supply element.Alternatively or additionally, the supply technology can be subdivided into individual sub-areas of the building, for example, individual rooms of the building. The training data provided to the neural network can thus be selected according to the intended use of the neural network. Due to the large number of possible training data, a so-called super-neural network can be easily generated. This super-neural network can be trained using training data that, for example, comes from multiple buildings. The super-neural network can learn a relationship between this training data from multiple buildings and use this to determine control data for the supply technology of a differently constructed building for whose specific construction no training data, for example in the form of historical control data, is available.Analogously, the super-neural network can be trained with training data for several sub-areas of a building in order to be able to determine control data for another sub-area of ​​a building.

[0024] According to a further embodiment, the preprocessing of the provided system parameters, building data, environmental data, and control data comprises indexing and / or reducing. The indexing of the provided data can be performed by assigning at least one index to the respective provided data. Based on the at least one index, the provided data can be sorted and / or ordered. The reduction of the provided data can be performed by removing a subset and / or individual data from the provided training data set. For example, the data that cannot be processed by the neural network to be trained due to its structure can be removed from the provided training data set.Alternatively or additionally, negative values, zero values, and / or values ​​that are very far apart from the other values ​​contained in the training dataset can be removed from the provided training dataset. This allows the neural network to be trained more effectively. The provided system parameters, building data, environmental data, and control data can thus be optimized for their use as training data for the neural network through preprocessing. This allows the neural network to be trained more quickly and / or efficiently.

[0025] According to one embodiment, indexing is performed using a cycle index depending on time information of the system parameters or the respective data. The cycle index includes, for example, a month, a day of the week, a time of day, and / or information about building usage. A cycle index can be the interval between two points in time at which the neural network accesses the provided training data. Information about building usage can, for example, include information about whether there were people in the building at the time the data was collected and / or how many people were in the building at the time the data was collected.Information about the use of a building may, for example, include information about whether the building was in use, for example as office space for employees of a company, or was unused, for example due to a holiday or because the company was closed.

[0026] By indexing with a cycle index depending on time information, it can be ensured that a factual relationship can be mapped from a large number of available training data when training the neural network, which can enable better training and / or training with a smaller data set. For example, the provided system parameters, building data, environmental data, and control data can have the same or at least a similar cycle index, depending on the conditions that influence the thermal state, such as a general climate or usage status. For example, indexing purely by date has no connection to usage, since the same day in the same month can be a different day of the week in different years.In contrast, if it was a Wednesday, it can be assumed with high confidence that, for example, an office building was regularly used by employees. This type of cycle indexing is therefore useful. Furthermore, such indexing can be carried out easily and even automated with little effort, for example to pre-process historical data sets for training. This can prevent the use of data to train the neural network that, for example, leads to energy-optimal control data but which has no relationship to one another. Furthermore, the time information of the cycle index can be used to arrange the provided data in the chronological order corresponding to the data access of the neural network. The provided data is therefore already available in the temporal structure required by the neural network.This reduces the computation time required to train the neural network.

[0027] According to one embodiment, reducing comprises determining a correlation coefficient of the input data with the output data. The neural network is trained only with input data whose correlation coefficient with the output data is above a threshold. For example, the Pearson correlation coefficient can be used, which indicates the correlation between two variables as a percentage. Other methods for determining correlation coefficients can also be used. It can be provided that only those input data are used to train the neural network whose correlation with the output data is greater than 10%, greater than 5%, or greater than 1%. Alternatively or additionally, the input data can be sorted according to the level of their correlation, and then, for example, only three-quarters, half, or one-quarter of the data with the highest correlation coefficient can be used.By reducing the correlation coefficients, it can be ensured that only those input data are used to train the neural network which, due to a sufficiently high correlation with the output data, have a sufficient influence on the output data. Conversely, it can be specified that, when creating the training dataset, those input data which, due to their low correlation with the output data, do not have a sufficient influence on the output data are ignored. This ensures that the relationship between the input data and output data determined by the neural network during training is not random, but, due to the sufficiently high correlation, also exists between comparable input data and output data.In addition, reducing the size of the training dataset can be kept small, which also reduces the computation time required to carry out the procedure for training the neural network.

[0028] In a second aspect, the invention relates to a method for using a neural network to control the supply technology of a building. The neural network is used to control the supply technology of the building in such a way that the neural network executes at least one of the steps necessary for controlling the supply technology or at least prepares this step. The neural network can have been trained using the method according to the first aspect. Corresponding features, advantages, and developments can be taken from the first aspect and apply equally to the second aspect. Conversely, respective features, advantages, and developments also constitute respective features, advantages, and developments for the first aspect, where applicable.

[0029] The method comprises a step of receiving system parameters of the utility technology. The system parameters of the utility technology can be detected by sensors designed for this purpose, such as a temperature sensor and / or a humidity sensor. The corresponding sensors can be designed as part of the utility technology. For example, respective sensor signals can be transmitted wired or wirelessly to a computing device on which the neural network is implemented. The computing device can be part of a control device of the utility technology or, for example, be formed by an external server. The system parameters can define the thermal state of the building, as already explained in the first aspect.

[0030] The method according to the second aspect further comprises a step of inputting building data of the building, environmental data of the building, and the received system parameters into the neural network. The building data can represent a structural condition of the building, as already explained in relation to the first aspect. The environmental data can represent a thermal condition of the building's environment, as already explained in relation to the first aspect. Using the building data and / or environmental data, the neural network can, for example, apply relationships learned from other buildings to the control of the supply technology. The building data and the environmental data of the building can be received and / or measured by sensors designed for this purpose, analogous to the system parameters of the supply technology. For example, this data can also be retrieved from databases, such as a satellite map.Alternatively, the building data and / or the environmental data can be predefined data. For example, the building data and / or the environmental data can be based on empirical values. The building data, the environmental data, and the system parameters can be transmitted as input data from a dedicated input device to at least one input node of the neural network.

[0031] Furthermore, the method according to the second aspect comprises a step of generating a control data set that can be used to control the building's supply technology. The control data set is generated based on a relationship between the system parameters, the building data, and the environmental data with control data, learned by the neural network. The control data can be used to control the supply technology. The control data set can contain control data with which at least one supply element of the supply technology, for example a heating system or a ventilation system, can be controlled. In a simple case, the control data can also comprise merely a control value of a heater's output.Alternatively, the control data can also correspond, for example, to a desired heat input into the building, on the basis of which a control device of the supply technology specifies the respective control parameters. The control data can be transmitted directly to the respective supply element of the supply technology for its respective control. Alternatively, the control data can serve as starting values ​​for a thermal simulation model of the building. The final values ​​of this simulation model can then be used to control the supply technology. Due to the use of the neural network, the method according to the second aspect enables a faster and less complex generation of the control data set compared to conventional methods.Thanks to the accelerated calculations using the neural network, the control of the utility technology can be adapted to changing and / or externally specified conditions at shorter intervals than with conventional simulation methods, for example, every 15 minutes. For example, sudden storms, a changed weather forecast, and / or an unscheduled shutdown of the utility technology due to construction work, for example, can be responded to at short intervals by adjusting the control data set.

[0032] According to one embodiment of the method according to the second aspect, the system parameters are representative of energy consumption, energy costs, and / or actual performance of the building's utility technology. The control data set is generated by the neural network with respect to an optimization condition for energy consumption and / or energy costs or with respect to a target performance of the utility technology. Energy consumption of the utility technology can be determined, for example, based on thermal insulation of the building, an efficiency of at least one supply element of the utility technology, or an energy distribution within the building. Energy costs of the utility technology can be determined, for example, based on an electricity price, a gas price, an operating time, and / or a utilization of the building's utility technology.An actual output of the supply technology can be determined, for example, based on an actual temperature, an actual air humidity, an actual position of at least one control element and / or an actual energy consumption of the supply technology of the building. An optimization condition for energy consumption and / or energy costs can be specified for at least one of the above-mentioned parameters. Alternatively, an optimal value to be achieved can be specified for each of the above-mentioned parameters by a user and / or manufacturer of the supply technology. A target output of the supply technology can be specified for at least one of the above-mentioned parameters by a user and / or manufacturer of the supply technology. Alternatively, a target value to be achieved can be specified for each of the above-mentioned parameters by a user and / or manufacturer of the supply technology.Each parameter can be a time-dependent value and, for example, can be specified differently depending on the time of day, day of the week, and / or month. In particular, each parameter can be specified time-dependently, similar to cycle indexing. Using the proposed method, the energy consumption and / or energy costs of the utility system can be optimized, or the performance of the utility system can be adapted to a given condition.

[0033] According to a further embodiment, the method is applied to sub-areas of the building, sub-areas of the building's supply technology, multiple buildings and / or the respective supply technology of multiple buildings. Sub-areas of the building can, for example, be individual rooms of the building. Alternatively or additionally, sub-areas of the building can, for example, be individual floors and / or wings and / or other sections of the building. Sub-areas of the supply technology can, for example, be individual supply elements of the supply technology, such as a heating system and / or a ventilation system. Alternatively or additionally, the supply technology can also be subdivided into respective sub-areas with regard to the individual sub-areas of the building, for example with regard to the individual rooms of the building. The proposed method thus offers a multitude of different areas of application.

[0034] According to a further embodiment, the control data set generated by the neural network is used as an input value for a classical mathematical model for controlling the building's supply technology. A classical mathematical model can be a mathematical model calculated using classical methods, i.e., without the use of artificial intelligence and / or machine learning. The control data set can comprise at least one parameter that can be used as a starting value for the classical mathematical model. The classical mathematical model can be designed to analytically optimize control of the supply technology, for example, to optimize costs and / or energy consumption for a thermal target state. Alternatively, the control data set can comprise a plurality of parameters, all or a subset of which can be used as starting values ​​for the classical model.The mathematical model can, for example, generate an additional control data set that can be used to control the building's utility systems. Since the control data set is generated more quickly by the neural network, the control of the building's utility systems can also be carried out more quickly. This allows changes in system parameters, building data, and / or environmental data, especially those that occur at short notice, to be taken into account when controlling the building's utility systems.

[0035] According to one embodiment, the neural network used in the method according to the second aspect can be trained using the method for training a neural network according to the first aspect. The technical effects and advantages mentioned in relation to the first aspect can thus be applied to the neural network used in the method according to the second aspect.

[0036] According to an embodiment of the first or second aspect, the system parameters of the supply technology can be selected from at least one of the following: a temperature within the building, an air humidity within the building, a load of a power source of the supply technology, a number of people within the building, a number of particularly active electrical consumers within the building, a mass flow through at least one heating element of a heating circuit, a flow temperature through at least one heating element of a heating circuit, a temperature of an air flow of a ventilation system, an air humidity of an air flow of a ventilation system, a mass flow through a ventilation system, a dimensioning of at least one supply element of the supply technology, and a predetermined change in at least one of the aforementioned parameters.A mass flow rate through a heating element of a heating circuit can be a volume flow rate of an energy source, such as oil or gas, which is used to heat the building. The mass flow rate can be a characteristic value for heat input into the building. A flow temperature through a heating element can be the temperature of an energy source that flows through the heating element to heat the building, such as the temperature of heating oil or heating gas. The flow temperature can be a characteristic value for heat flow in the building. A mass flow rate through a ventilation system can be a volume flow rate of an amount of air that flows through the ventilation system to ventilate the building, i.e. for air exchange within the building. The mass flow rate can be a characteristic value for heat exchange in the building.The dimensioning of a supply element can be its maximum and / or minimum performance capacity or, for example, also relate to a surface area and optionally its arrangement for heat exchange. The dimensioning of the supply element can be a parameter for its energy consumption and its potential heat input into the building. A temperature within the building, a humidity within the building, a temperature of an air flow of a ventilation system and / or a humidity of an air flow of a ventilation system can be a parameter for a perceived temperature within the building. A load of a power source of the supply technology, a number of people within the building and / or a number of electrical consumers can be important parameters that should be taken into account for the control of the building's supply technology.

[0037] Alternatively or additionally, the building data of the building can be selected from at least one of the following: a type of building, geometric dimensions of the building, a wall thickness of the building, a material composition of the building, a number of windows of the building, and an orientation of the building with respect to a north-south axis. A type of building can be a characteristic value for an energy requirement of the supply technology. For example, the energy requirement for an office building can be higher than for a warehouse. Geometric dimensions of the building can be, for example, a height, a width and / or a volume of the building. The geometric dimensions can have an influence on a heat distribution within the building and / or an influence of the weather on the thermal condition of the building.A building's wall thickness, a building's material composition, and / or a number of windows can be a parameter for heat input into the building due to certain weather conditions, such as solar radiation and outside temperature. The building's orientation with respect to a north-south axis can influence airflow within the building. Alternatively or additionally, the building's environmental data can be selected from at least one of the following: historical weather data for the building's surroundings, historical climate data for the building's surroundings, forecast weather data for the building's surroundings, forecast climate data for the building's surroundings, and an electricity price in the building's surroundings.Historical weather data for the building's surroundings and / or historical climate data for the building's surroundings can be used to determine a predefined default setting for the utility technology at specific operating times, for example, a default setting for daytime operation and / or a default setting for nighttime operation. Forecasted weather data and / or forecasted climate data can be determined using a mathematical simulation and describe expected future weather and / or climate. The forecasted weather data and / or climate data can be used to create a weather forecast.Forecasted weather data for the area surrounding the building, forecasted climate data for the area surrounding the building and / or an electricity price in the area surrounding the building can be an indicator for the expected energy consumption and / or expected energy costs of the building.

[0038] Alternatively or additionally, the control data can be selected from at least one of the following: a setting of a control element of a heating system of the building, a setting of a control element of a cooling system of the building, a setting of a control element of an air conditioning system of the building, and a setting of a control element of a ventilation system of the building. A control element can be a mechanical control element, for example, a rotary switch, and / or an electronic control element. The setting of the control element can be a characteristic value for a thermal state of the supply technology.

[0039] The neural network can therefore be trained using a large number of training data sets. Furthermore, the control data set can be generated based on a large number of input data sets, output data sets, and optimization conditions. The proposed methods are therefore suitable for a wide variety of different applications. In a third aspect, the invention relates to a computer program product for controlling the supply technology of a building using a control data set. The control data set is generated according to the method according to the second aspect and / or by a neural network trained using the method according to the first aspect. The control data set is used in particular to control a heating system, a cooling system, an air conditioning system, and / or a ventilation system of the building.The embodiments, technical effects and advantages explained for the first and second aspects respectively therefore also apply analogously to the computer program product according to the third aspect and vice versa.

[0040] The computer program product can be stored on a data carrier. The data carrier can be a computer-readable, non-volatile storage medium on which the computer program product is present in the form of program code. The data carrier can be portable. The data carrier can be connected to a computer for signal transmission, so that the program code present on the data carrier can be executed by the computer.

[0041] In a fourth aspect, the invention relates to a control device for controlling the supply technology of a building using a control data set. The control device comprises a computer-readable, non-volatile storage medium containing computer program code that, when executed on a computer, generates the control data set according to the method according to the second aspect. In particular, the control device can be designed to implement the previously described neural network. The embodiments, technical effects, and advantages explained for the first and second aspects thus also apply analogously to the control device according to the fourth aspect.

[0042] In a fifth aspect, the invention relates to a trained neural network designed to be used to control a building's supply technology according to the second aspect. The neural network was trained according to the method according to the first aspect. The embodiments, technical effects, and advantages explained for the first and second aspects thus also apply analogously to the trained neural network according to the fifth aspect.

[0043] Short description of the drawings

[0044] Fig. 1 shows a flowchart with steps of a method for training a neural network which is designed to be used to control a supply technology of a building.

[0045] Fig. 2 shows a flowchart with steps of a method for using a neural network to control a building's utility technology.

[0046] Fig. 3 shows a schematic diagram of a control device for controlling a building's supply technology with a neural network.

[0047] Fig. 4 schematically illustrates the functioning of a super-neural network.

[0048] Detailed description of embodiments

[0049] Fig. 1 shows a flowchart with steps of a method for training a neural network designed to be used to control the supply technology of a building. In the exemplary embodiment of Fig. 1, the building is an office building in which many people work. Certain climatic conditions should be maintained within the office building, at least during the people's working hours. For example, a room temperature between 18°C ​​and 22°C and a humidity between 50% and 70% should prevail. In order to maintain these climatic conditions, it must be possible to react to, for example, temperature fluctuations, humidity fluctuations, and the number of people within the building.At the same time, depending on energy costs and the weather forecast, it may be useful for energy and / or cost optimization to achieve these climatic conditions long before the start of use or only very shortly before the start of use. This can result in different heating or cooling rates of the building's climate due to the supply technology. The start of use can, for example, be defined as a typical start time for working hours in the building. The neural network trained according to the method of the exemplary embodiment in Fig. 1 enables rapid adjustment of the control of the office building's supply technology with low computational effort in order to implement such optimization.

[0050] In a first training step TS1, system parameters of the utility technology are provided. In the exemplary embodiment shown in Fig. 1, the system parameters of the utility technology are the load of a power source of the utility technology, a number of people within the building, a mass flow rate through a heating element of a heating circuit of the utility technology, a flow temperature through the heating element of the heating circuit, a temperature of an air flow of a ventilation system of the utility technology, a humidity of the air flow of the ventilation system, and a mass flow rate through the ventilation system. Optionally, other system parameters of the building's utility technology can be considered alternatively or additionally.

[0051] In a second training step TS2, building data for the building is provided. In the embodiment of Fig. 1, the building data includes the building's wall thickness, the number of windows, and the building's orientation relative to a north-south axis. Optionally, additional building data can be considered alternatively or additionally.

[0052] In a third training step TS3, environmental data of the building is provided. In the embodiment of Fig. 1, the environmental data of the building are historical weather data for the surroundings of the building, historical climate data for the surroundings of the building, forecast weather data for the surroundings of the building, and forecast climate data for the surroundings of the building. Optionally, further environmental data of the building, for example energy costs for the surroundings of the building, can be taken into account alternatively or additionally. In a fourth training step TS4, control data that can be used to control the building's supply technology is provided. In the embodiment of Fig. 1, the control data are a setting of a control element of a heating system of the building and a setting of a control element of a ventilation system of the building.Optionally, further control data can be taken into account alternatively or additionally.

[0053] The provided system parameters, building data, environmental data, and control data are preprocessed in a fifth training step (TS5) to generate a training dataset for the neural network. The training dataset includes the preprocessed system parameters, the preprocessed building data, and the preprocessed environmental data as input data, and the preprocessed control data as output data. The input data is transmitted to the input nodes of the neural network. The output data is transmitted to the output nodes of the neural network in a similar manner.

[0054] In the embodiment shown in Fig. 1, the provided data and system parameters are indexed during preprocessing according to training step TS5 with a cycle index, which includes a month, a day of the week, a time of day, and information about building usage. The data and system parameters provided for training the neural network are thus placed in a contextual context, so that the neural network's training results are based on data collected at the same or comparable times.

[0055] Furthermore, the provided data and system parameters are reduced during training step TS5. The number of provided values ​​is reduced through selection. For this purpose, the Pearson correlation coefficient of the input data with the output data is determined. The neural network is only trained with input data whose Pearson correlation coefficient with the output data is above a threshold of 5%. This ensures that the results obtained during training of the neural network are not based on chance, but that there is a dependency between the output data and the input data. In a sixth training step TS6, the neural network is trained with the training data set in order to learn a relationship between the input data and the output data.For training, the input data transmitted to the input nodes of the neural network is transmitted to intermediate nodes of the neural network via connections mathematically weighted by the neural network. The transmitted input data is processed at the intermediate nodes to obtain the output data provided at the output nodes. The output nodes are connected to the intermediate nodes via connections mathematically weighted by the neural network. The mathematical weighting of the respective connections is adjusted by the neural network during the training process.

[0056] By preprocessing the training data, the training process of the neural network can be accelerated. The neural network trained in this way can be used to control the building's utility systems. For example, the trained neural network can be used to generate a control data set for the building's utility systems. By incorporating building data and environmental data, the trained neural network can also be used for other buildings and / or for changing environmental conditions within the building.

[0057] Fig. 2 shows a flowchart with steps of a method for using a neural network to control a supply technology of a building, according to a further embodiment of the invention.

[0058] In a first determination step BS1, system parameters of the supply technology are received. The system parameters are the same parameters with which the neural network was trained according to the embodiment shown in Fig. 1 and are recorded by sensors designed for this purpose.

[0059] In a second determination step BS2, building data, environmental data, and the received system parameters are input into the neural network. The building data and environmental data are the same parameters with which the neural network was trained according to the embodiment of Fig. 1. However, the respective values ​​of the individual parameters of the building data and environmental data correspond to currently measured values. The system parameters, building data, and environmental data are then transmitted to the input nodes of the neural network.

[0060] In a third determination step BS3, the neural network generates a control data set that can be used to control the building's supply technology. The control data set is determined based on a relationship learned by the neural network between the system parameters, the building data, and the environmental data with control data that can be used to control the supply technology. The control data set generated by the neural network is used as input for a classic mathematical model for controlling the building's supply technology. Using the classic model, optimized control data for the building's supply technology is generated based on the control data set. The control data generated based on the control data set is used to control the supply technology. The control data described with reference to Fig.The climatic conditions within the building mentioned in section 1 can thus be achieved as energy-efficiently and / or cost-effectively as possible. Furthermore, due to the accelerated calculation by the neural network, short-term weather changes or short-term changes in energy prices can be responded to by adjusting the control data set accordingly.

[0061] The neural network used to perform the method of the embodiment of Fig. 2 may have been trained according to the method of the embodiment of Fig. 1.

[0062] Fig. 3 schematically shows a control device 10 for controlling a supply technology 30 of a building by means of a neural network 12.

[0063] The control device 10 comprises a computer-readable, non-volatile storage medium 14 on which the neural network 12 is stored. The neural network 12 can access a computing unit (not shown) of the control device 14 for data processing. The control device 10 further comprises an input device 16, which is designed to receive system parameters 18 of the supply technology 30. The input device 16 is further designed to receive building data 20 and environmental data 22 of the building. The input device 16 transmits the system parameters 18, the building data 20, and the environmental data 22 as input data 24 to the neural network 12.

[0064] Based on the input data 24, the neural network generates a control data set 26. The control data set 26 is generated by the neural network 12 based on a relationship learned by the neural network 12 between the system parameters 18, the building data 20, and the environmental data 22 with control data that can be used to control the utility technology 30. The neural network 12 transmits the control data set 26 to an output device 28. The output device 28 transmits the control data set 26 to the utility technology 30 of the building (not shown).

[0065] The control data set 26 can be used to optimize energy consumption and / or energy costs of the supply technology 30. Alternatively or additionally, the control data set 26 can be used to achieve a target performance of the supply technology 30.

[0066] The control device 10 shown in Fig. 3 can be used to train the neural network according to the embodiment of Fig. 1.

[0067] Fig. 4 schematically shows a super-neural network 112. A super-neural network is a neural network that is trained with training data originating from multiple sub-areas of a building or one or more buildings whose supply technology has, for example, comparable or scalable values. Each of the sub-areas of the building can be supplied by the supply technology independently of the other sub-areas of the building, so that different climatic conditions prevail in each of the sub-areas. Alternatively or additionally, each of the sub-areas can have different building data, for example a different number and / or types of windows and / or different wall thicknesses. Analogously, each of the multiple buildings can be supplied by an independent supply technology, so that different climatic conditions and different building data prevail in each of the buildings.The super-neural network can thus be trained taking into account all sub-areas of the building and / or multiple buildings. For example, the super-neural network can be trained using data from three buildings and, with appropriately adapted building data, used to control the utility technology of another building.

[0068] The super-neural network 112 is used to control a utility system 50 of a building 40. The building 40 is the office building mentioned with reference to Fig. 1.

[0069] Building 40 is divided into several sub-areas 42, 44, 46, and 48. In Figure 4, the sub-areas are the individual offices 42, 44, 46, and 48 of building 40. Optionally, building 40 can be divided into further and / or different sub-areas, for example, multiple floors. The climatic conditions of sub-areas 42, 44, 46, and 48 are controlled by the supply technology 50. Different climatic conditions can prevail in each of the sub-areas 42, 44, 46, and 48. For example, each of the sub-areas 42, 44, 46, and 48 can be supplied by the supply system 50 via a separate heating pipe with variable temperature control assigned to the respective sub-area 42, 44, 46, 48 and / or a separate ventilation system with variable air flow assigned to the respective sub-area 42, 44, 46, 48. Furthermore, the building data can be different in each of the sub-areas 42, 44, 46, and 48 of the building 40.For example, the number of windows in each of the sub-areas 42, 44, 46 and 48 may be different.

[0070] The super-neural network 112 receives a training data set 60, 62, 64, and 66 from each of the sub-areas 42, 44, 46, and 48 of the building 40. The training data sets 60, 62, 64, and 66 are generated according to the exemplary embodiment of Fig. 1. Based on the training data sets 60, 62, 64, and 66, the super-neural network 112 generates a control data set 70, which can be used to control the supply technology 50 of the building 40 as described above. The control data set 70 is generated according to the exemplary embodiment of Fig. 2. Based on the training data sets 60, 62, 64 and 66, the super-neural network 112 can thus generate the control data set 70 for the supply technology 50, even though the super-neural network 112 was not trained with control data for the entire building 40.The super-neural network 112 can determine the control data for the entire building 40 from the control data for the individual sub-areas 42, 44, 46 and 48 of the building 40.

[0071] Using the control data set 70, the supply technology 50 can be controlled such that the climatic conditions explained with reference to Fig. 1 can be set as energy-efficiently as possible in each of the sub-areas 42, 44, 46, and 48 of the building 40. Alternatively, different climatic conditions can be generated by the supply technology 50 in each of the sub-areas 42, 44, 46, and 48 of the building 40 using the control data set 70.

[0072] In a further embodiment, the super-neural network 112 can receive a training data set from one or more additional buildings. The training data set is generated according to the embodiment of Fig. 1. Based on this training data set, the super-neural network 112 generates a control data set that can be used to control the building's utility technology as described above. The control data set is generated according to the embodiment of Fig. 2. Based on the training data set received from one or more buildings, the super-neural network 112 can thus generate the control data set for the building's utility technology, even though the super-neural network 112 was not trained with control data from this building.The super-neural network 112 may determine the control data for the building based on the control data included in the training data set obtained from one or more buildings.

[0073] The super-neural network 112 can alternatively be trained with training data originating from a plurality of sub-areas of the utility system 50 of the building 40. In this exemplary embodiment, the super-neural network 112 generates a control data set for each of the sub-areas of the utility system 50. These control data sets can be used to control the respective sub-area of ​​the utility system.

[0074] When training the super-neural network 112, the various sub-areas of the building 40 and / or the various sub-areas of the utility technology 50 can be taken into account. The trained super-neural network 112 can thus be used for each of the sub-areas of the building as well as for the entire building to control the utility technology 50.

[0075] Reference symbol

[0076] 10 Control device

[0077] 12 Neural Network

[0078] 14 Storage medium

[0079] 16 Input device

[0080] 18 System parameters of the supply technology

[0081] 20 building data

[0082] 22 Environmental data

[0083] 24 input data

[0084] 26 Control data record

[0085] 28 Dispensing device

[0086] 30 Supply technology of a building

[0087] 112 Super-neural network

[0088] 40 buildings

[0089] 42, 44, 46, 48 parts of the building

[0090] 50 Supply technology

[0091] 60, 62, 64, 64 Training data of the parts of the building

[0092] 70 Control data record

[0093] TS1 first training step

[0094] TS2 second training step

[0095] TS3 third training step

[0096] TS4 fourth training step

[0097] TS5 fifth training step

[0098] TS6 sixth training step

[0099] BS1 first determination step

[0100] BS2 second determination step

[0101] BS3 third determination step

Claims

Patent claims 1. A method for training a neural network (12; 112) which is designed to be used to control a supply technology (30; 50) of a building (40), the method comprising the following steps: - Providing (TS1) system parameters (18) of the supply technology (30; 50); - Providing (TS2) building data (20) of the building (40); - Providing (TS3) environmental data (22) of the building (40); - Providing (TS4) control data which can be used to control the supply technology (30; 50); - Preprocessing (TS5) the provided system parameters (18), building data (20), environmental data (22) and control data to generate a training data set (60, 62, 64, 66) for the neural network (12; 112), wherein the training data set (60, 62, 64, 66) comprises the preprocessed system parameters (18), the preprocessed building data (20) and the preprocessed environmental data (22) as input data (24) and the preprocessed control data as output data; and - Training (TS6) the neural network (12; 112) with the training data set (60, 62, 64, 66) in order to learn a relationship between the input data (24) and the output data.

2. A method for training a neural network (12; 112) according to claim 1, wherein the provided system parameters (18), building data (20), environmental data (22) and / or control data are historical values ​​and / or are synthetically generated.

3. A method for training a neural network (12; 112) according to claim 1 or 2, wherein the provided system parameters (18), building data (20), environmental data (22) and / or control data relate to sub-areas (42, 44, 46, 48) of the building (40), sub-areas of the supply technology, a plurality of buildings and / or the respective supply technology of a plurality of buildings.

4. A method for training a neural network (12; 112) according to one of claims 1 to 3, wherein the preprocessing of the provided system parameters (18), building data (20), environmental data (22) and control data comprises indexing and / or reducing.

5. A method for training a neural network (12; 112) according to claim 4, wherein the indexing is performed with a cycle index depending on time information of the system parameters (18) or the respective data (20, 22); wherein the cycle index comprises a month, a day of the week, a time of day, and / or information about building usage.

6. A method for training a neural network (12; 112) according to one of claims 4 or 5, wherein the reducing comprises determining a correlation coefficient of the input data with the output data, wherein the training of the neural network (12; 112) is carried out only with input data whose correlation coefficient with the output data is above a threshold value.

7. A method for using a neural network (12; 112) to control a supply technology (30; 50) of a building (40), the method comprising the following steps: - receiving (BS1) system parameters (18) of the supply technology (30; 50); - inputting (BS2) building data (20) of the building (40), environmental data (22) of the building (40) and the received system parameters (18) into the neural network (12; 112); - generating (BS3) a control data set (26; 70) which can be used to control the supply technology (30; 50) of the building (40), based on a relationship learned by the neural network (12; 112) between the system parameters (18), the building data (20) and the environmental data (22) with control data which can be used to control the supply technology (30; 50).

8. A method for using a neural network (12; 112) according to claim 7, wherein the system parameters (18) are representative of an energy consumption, Energy costs and / or an actual performance of the supply technology (30; 50) of the building (40), wherein the control data set (26; 70) is generated by the neural network (12; 112) with regard to an optimization condition of the energy consumption and / or the energy costs or with regard to a target performance of the supply technology (30; 50).

9. A method for using a neural network (12; 112) according to claim 7 or 8, wherein the method is applied to sub-areas (42, 44, 46, 48) of the building (40), sub-areas of the supply technology (30; 50) of the building, several buildings and / or the respective supply technology (30; 50) of several buildings.

10. A method for using a neural network (12; 112) according to one of claims 7 to 9, wherein the control data set (26; 70) generated by the neural network (12; 112) is used as an input value for a classical mathematical model for controlling the supply technology (30; 50) of the building (40).

11. A method for using a neural network (12; 112) according to any one of claims 7 to 10, wherein the neural network (12; 112) is trained by means of the method for training a neural network (12; 112) according to any one of claims 1 to 6.

12. A method for training a neural network (12; 112) according to any one of claims 1 to 6 or a method for using a neural network (12; 112) according to any one of claims 7 to 11, wherein - the system parameters (18) of the supply technology (30; 50) are selected from at least one of the following: a temperature within the building (40), an air humidity within the building (40), a load of a power source of the supply technology (30; 50), a number of people within the building (40), a number of electrical consumers within the building (40), a mass flow through at least one heating element of a heating circuit, a flow temperature through at least one heating element of a heating circuit, a temperature of an air flow of a ventilation system, an air humidity of an air flow of a ventilation system, a mass flow through a ventilation system, a dimensioning of at least one supply element of the supply technology (30; 50), and a predetermined change in at least one of the aforementioned parameters; - the building data (20) of the building (40) are selected from at least one of the following: a type of building (40), geometric dimensions of the building (40), a wall thickness of the building (40), a material composition of the building (40), a number of windows of the building (40), and an orientation of the building (40) with respect to a north-south axis; - the environmental data (22) of the building (40) are selected from at least one of the following: historical weather data for an environment of the building (40), historical climate data for the environment of the building (40), forecast weather data for the environment of the building (40), forecast climate data for the environment of the building (40), and an electricity price in an environment of the building (40); and / or - the control data are selected from at least one of the following: a setting of a control element of a heating system of the building (40), a setting of a control element of a cooling system of the building (40), a setting of a control element of an air conditioning system of the building (40), and a setting of a control element of a ventilation system of the building (40).

13. Computer program product for controlling the supply technology (30; 50) of a building (40) by means of a control data set (26; 70) which is generated according to the method according to one of claims 7 to 12 and / or which is generated by a neural network (12; 112) trained with the method according to one of claims 1 to 6, wherein the control data set (26; 70) is used in particular for controlling a heating system, a cooling system, an air conditioning system and / or a ventilation system of the building (40).

14. Control device (10) for controlling a supply technology (30; 50) of a building (40) by means of a control data set (26; 70), comprising a computer-readable, non-volatile storage medium (14) containing computer program code which, when executed on a computer, generates the control data set (26; 70) according to the method according to one of claims 7 to 12.

15. A trained neural network (12; 112) which is designed to be used to control a supply technology (30; 50) of a building (40) according to one of claims 7 to 12, wherein the neural network (12; 112) has been trained according to the method according to one of claims 1 to 6.

Images