KR20200034023A - A method for creating demand response determination model for hvac system and a method for demand response - Google Patents

Abstract

Provided is a method for generating a demand response (DR) determining model for an HVAC system of a building. According to the present invention, the method comprises the steps of: generating a zone temperature determining model for outputting a temperature of a building by receiving input power of an HVAC system and a thermal state of a building by training a first artificial neural network on the basis of a plurality of first training data; generating an objective function for a power supply schedule having an optimal solution varied in accordance with an electricity price and a thermal state while including a linear equation for simulating the generated zone temperature determining model; determining each optimal solution of the objective function for a plurality of electricity price profiles and thermal state profiles; and generating a DR determining model for outputting a power supply schedule for the HVAC system by receiving the electricity price profiles and the thermal state profiles by training a second artificial neural network on the basis of a plurality of second training data including the electricity price profiles, the thermal state profiles, and the determined optimal solution.

Description

A method for generating a demand response decision model for a heat and air conditioning system and a method for performing a demand response {A METHOD FOR CREATING DEMAND RESPONSE DETERMINATION MODEL FOR HVAC SYSTEM AND A METHOD FOR DEMAND RESPONSE}

The present invention relates to a demand response, and more particularly, to a method of using supervised learning for an optimal demand response to a building's HVAC system.

The large thermal capacity, a characteristic of buildings, enables the building's heating and air conditioning (HVAC) systems to be used for demand response (DR). However, it is particularly difficult to apply the demand response to the HVAC system in a multi-zone building. In order to use the HVAC system as a demand response resource, a temperature model for each area of the building's thermal conditions and facility operation is required. This is because the model needs to reflect the physical characteristics in detail.

Korean Registered Patent Publication No. 10-1739271 ("Smart Portfolio Optimization Modeling Forming Apparatus and Method for Optimal Configuration of Demand Response Resources", Next Generation Convergence Technology Research Institute)

The object of the present invention for solving the above-mentioned problems is to learn an artificial neural network based on the data accumulated under normal building operating conditions using machine learning, and to simulate it as an equation using a partial linear equation to optimize price-based demand response. It is to provide a method for generating a demand response decision model for a heat and air conditioning system that can derive an optimal scheduling of input power of an HVAC system simply and quickly by applying it to a scheduling problem.

Another object of the present invention for solving the above-described problem is to learn an artificial neural network based on the data accumulated under normal building operating conditions using machine learning, and simulate it as an equation using a partial linear equation to determine the price-based demand response. It is to provide a method for performing a demand response that can easily and quickly derive an optimal scheduling of input power of an HVAC system by applying to an optimal scheduling problem.

However, the problem to be solved of the present invention is not limited to this, and may be variously extended without departing from the spirit and scope of the present invention.

Method for generating a demand response (DR) determination model for a heating, air conditioning (HVAC) system of a building according to an embodiment of the present invention for achieving the above object is , By training a first artificial neural network based on a plurality of first training data, generating an area temperature determination model that receives HVAC system input power and a thermal state of the building and outputs the building temperature; Generating an objective function for a power supply schedule that includes a linear equation that simulates the zone temperature determination model, and that an optimal solution varies according to electricity price and thermal conditions; Determining respective optimal solutions of the objective function for a plurality of electricity price profiles and a thermal state profile; And training the second artificial neural network based on a plurality of second training data each including the electric price profile, the thermal condition profile, and the determined optimal solution, to receive the electric price profile and the thermal condition profile and enter the HVAC system. And generating a demand response determination model outputting a power supply schedule for the power supply.

According to one aspect, the building includes a plurality of zones, and the zone temperature determination model can output temperatures for each of the plurality of zones.

According to an aspect, each of the plurality of first training data may include information regarding a conventional HVAC system input power and a thermal state of a building, and each of the plurality of zones according to the conventional HVAC system input power and a thermal condition of a building. It may include a temperature for.

According to one aspect, the thermal state of the building may include at least one of air temperature, sunshine, wind power, humidity, thermal load of the building, and building usage schedule.

According to one aspect, information on the input power of the HVAC system and the thermal state of the building may be obtained from a building energy management system (BEMS).

According to one aspect, the input layer of the first artificial neural network includes: a HVAC system input power from a first predetermined time point to a current time point, a thermal state from a second predetermined time point to a current time point, and a second predetermined time point. Temperatures for each of the plurality of zones from the time point 3 to the time point before the current time point may be included as an input neuron.

According to one aspect, the first artificial neural network may be implemented in the form of a deep nonlinear auto-regressive network (D-NARX).

According to an aspect, the first artificial neural network includes a plurality of hidden layers, and at least one of the plurality of hidden layers includes a sigmoid function or a rectified linear unit, ReLU) function can be used as an active function.

According to one aspect, the first artificial neural network may include a pre-processor for normalizing input data and a post-processor for reverting normalization for output data.

According to one aspect, the first artificial neural network includes one or more weighting factors and one or more bias values, and the weighting factors and biasing values are based on normalized mean squared errors (NMSE). It can be determined by.

According to an aspect, the weighting factor and the deflection value may be respectively determined for each of the plurality of zones.

According to an aspect, the linear equation simulating the zone temperature determination model may include a partial linear equation generated through partial linearization of an active function corresponding to a plurality of hidden layers included in the first artificial neural network. .

According to one aspect, the power supply schedule may be determined to minimize the total electricity cost for the electricity price changing over time while maintaining the temperature of each of the plurality of zones within a predetermined range.

According to one aspect, the objective function is to optimize the power supply schedule for the HVAC system and the temperature for each of the plurality of zones to minimize the sum of excess amounts for the total electricity cost and a predetermined boundary temperature. As may be to determine.

According to one aspect, the boundary condition for the temperature of each of the plurality of zones of the objective function allows a first offset to the lower limit of the boundary temperature and a second offset to the upper limit of the boundary temperature, and the first The offset and the second offset may be set differently for each of the plurality of zones.

According to one aspect, the first offset may not be set to exceed the first enhanced offset, and the second offset may be set not to exceed the second enhanced offset.

According to one aspect, in the objective function, the HVAC system input power for a time when there is no occupant in the predetermined building may be set to zero.

According to an aspect, the electricity price profile may include information on electricity prices that change over time, and the thermal status profile may include information on thermal conditions of buildings that change over time.

According to an aspect, the step of determining the optimal solutions of the power schedule objective function may determine an optimal solution of the power schedule objective function using mixed-integer linear programming (MILP).

In order to solve the above-described problem, a demand response (DR) for a heating, air conditioning (HVAC) system of a building including a plurality of zones according to another embodiment of the present invention (Demand Response, DR) The method for performing is obtained a prediction profile (Profile) of the electricity price, which includes information about the electricity price, which changes over time, and a thermal condition prediction profile, which includes information about the thermal condition of the building, which changes over time. To do; And determining a power supply schedule for the HVAC system of the building based on the information on the electricity price and the information on the thermal condition.

According to an aspect, the step of determining the power supply schedule may be configured to determine the power supply schedule using a demand response determination model according to an embodiment of the present invention described above.

The disclosed technology can have the following effects. However, since the specific embodiment does not mean to include all of the following effects or only the following effects, the scope of rights of the disclosed technology should not be understood as being limited thereby.

According to the method for generating a demand response determination model for the HVAC system and the method for demand response according to the above-described exemplary embodiment of the present invention, an artificial neural network is based on data accumulated under normal building operating conditions using machine learning. The optimal scheduling of the input power of the HVAC system can be derived simply and quickly by learning and applying it to the optimal scheduling problem of price-based demand response by simulating it as a formula using a partial linear equation.

That is, the demand response determination according to an embodiment of the present invention can be applied immediately to various types of HVAC systems and commercial buildings by applying machine learning techniques, not physically-based modeling methods.

In addition, an artificial neural network featuring a feedback loop, before and after processors of data, and multiple hidden layers can be simulated, and by using this, an optimization problem can be constructed with a linear equation to obtain a global solution using a mixed integer linear programming method.

Furthermore, the super prediction learning meta prediction method according to an embodiment of the present invention can immediately derive the optimal scheduling of the input power of the HVAC system for the thermal condition of the building for 24 hours and the variable electricity price. This can significantly shorten the calculation time and help the distribution system operator by adjusting the total load demand of the HVAC system in the distribution network through the electricity price signal.

1 is a flowchart of an algorithm for scheduling an optimal demand response according to an embodiment of the present invention.
2 is a flowchart of a method for generating a demand response determination model for an HVAC system according to an embodiment of the present invention.
3 is a flowchart of a method for performing demand response of an HVAC system according to an embodiment of the present invention.
4 shows the structure of a first artificial neural network according to an embodiment of the present invention.
5 is a detailed view of the J-th hidden layer of FIG. 4.
6 is a detailed view of the K-th concealment layer in FIG. 4.
7 is a detailed view of the L-th concealment layer in FIG. 4.
8 is an example of partial linear approximation of an active function using a sigmoid function.
9 is an exemplary diagram of partial linear approximation of an active function using a ReLU function.
10 shows the processing of a data set relating to the optimal solution of the objective function.
11 shows processing of a training data set for learning a second artificial neural network.
12 shows an optimal demand response scheduling decision using the learned second artificial neural network.
13 shows an optimal demand response algorithm according to an embodiment of the present invention.

The present invention can be applied to various changes and can have various embodiments, and specific embodiments will be illustrated in the drawings and described in detail.

However, this is not intended to limit the present invention to specific embodiments, and should be understood to include all modifications, equivalents, and substitutes included in the spirit and scope of the present invention.

Terms such as first and second may be used to describe various components, but the components should not be limited by the terms. The terms are used only for the purpose of distinguishing one component from other components. For example, the first component may be referred to as a second component without departing from the scope of the present invention, and similarly, the second component may be referred to as a first component. The term and / or includes a combination of a plurality of related described items or any one of a plurality of related described items.

When an element is said to be "connected" or "connected" to another component, it is understood that other components may be directly connected to or connected to the other component, but other components may exist in the middle. It should be. On the other hand, when a component is said to be "directly connected" or "directly connected" to another component, it should be understood that no other component exists in the middle.

The terms used in this application are only used to describe specific embodiments, and are not intended to limit the present invention. Singular expressions include plural expressions unless the context clearly indicates otherwise. In this application, the terms "include" or "have" are intended to indicate the presence of features, numbers, steps, actions, components, parts or combinations thereof described herein, one or more other features. It should be understood that the existence or addition possibilities of fields or numbers, steps, operations, components, parts or combinations thereof are not excluded in advance.

Unless defined otherwise, all terms used herein, including technical or scientific terms, have the same meaning as commonly understood by a person skilled in the art to which the present invention pertains. Terms, such as those defined in a commonly used dictionary, should be interpreted as having meanings consistent with meanings in the context of related technologies, and should not be interpreted as ideal or excessively formal meanings unless explicitly defined in the present application. Does not.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings. In order to facilitate the overall understanding in describing the present invention, the same reference numerals are used for the same components in the drawings, and duplicate descriptions for the same components are omitted.

summary

According to an aspect of the present invention, a method is provided in which an HVAC system in a multi-zone building can be used for optimal demand response using machine learning.

The method disclosed in accordance with an aspect of the present invention, after learning a first artificial neural network based on data accumulated under normal building operating conditions, uses the learned first artificial neural network using a piecewise linear equation The equation can be simulated and applied to the optimal scheduling problem of price-based demand response. The optimal scheduling problem is solved for various electricity prices and thermal conditions of the building, and the resulting optimal solution is a second neural network (e.g., deep neural network) to be used to determine optimal demand response scheduling. network)). According to an aspect of the present invention, the algorithm may be referred to as supervised-learning-aided meta-prediction (SLAMP) using supervised learning. The machine learning-based strategy according to an embodiment of the present invention can be practically applied without compromising the thermal preference and cost-effective operation of the occupant and can be efficiently operated in terms of calculation time.

More specifically, according to one aspect of the present invention, a new method is disclosed in which a HVAC system in a multi-zone commercial building can optimally participate in demand response using machine learning. This is to continuously schedule the input power of the HVAC system over time, which is optimally determined for the electricity price that changes over time while keeping the temperature in the zone within a certain range.

A first artificial neural network (ANN) model with feedback loops, time delayed inputs and multiple hidden layers to predict the temperature change of each zone for input power changes in the HVAC system Can be implemented. Using the supervised learning algorithm, the above first artificial neural network model is trained with building operation data in a steady state, and then simulated as a set of partial linear equations. For optimal demand response scheduling, the optimization problem can be composed of linear equations, and global optimal solutions can be derived using mixed-integer linear programming (MILP). The optimal solution derived by considering the profile of various electricity prices and the thermal state of the building is used to train the second artificial neural network (eg, deep neural network) model, and then the second artificial neural network model is an optimization problem. It is used to instantly derive an optimal demand response schedule without solving the problem. This can be referred to as a supervised-learning-aided meta-prediction (SLAMP) algorithm.

According to the SLAMP algorithm according to an aspect of the present invention, the proposed algorithm has an advantageous effect that can be applied immediately to various types of HVAC systems and commercial buildings because it applies machine learning techniques, not physically-based modeling methods. In addition, a simulation of the first artificial neural network, which features a feedback loop, before and after processors of data, and multiple hidden layers can be included as the main technical features of the algorithm. The method can be used to find the global optimal solution. Furthermore, the super prediction learning meta prediction method according to an aspect of the present invention can immediately derive the optimal scheduling of the input power of the HVAC system for the thermal condition of the building for 24 hours and the variable electricity price. This can significantly shorten the calculation time and help the distribution system operator by adjusting the total load demand of the HVAC system in the distribution network through the electricity price signal.

Supervised learning of the thermal response of multi-zone buildings to the operation of HVAC systems

1 is a flowchart of an algorithm for scheduling an optimal demand response according to an embodiment of the present invention. That is, FIG. 1 is a schematic view of a machine learning-based demand response strategy algorithm according to an aspect of the present invention. The algorithm according to an aspect of the present invention can be roughly divided into three steps. First, (a) the first artificial neural network can be implemented for the operation of the HVAC system and modeling of the thermal state of the building. Subsequently, (b) an optimization problem may be constructed and an optimal solution may be derived using a simulation method as a linear equation for the first artificial neural network. Subsequently, (c) a meta prediction algorithm for supervised learning is implemented using the optimal solution derived above in order to obtain an optimal demand response scheduling.

Referring to FIG. 1, in order to perform an algorithm according to an aspect of the present invention, first, a building energy management system relating to HVAC system operation and thermal conditions (weather, zone temperature, etc.) of a target multi-zone building ( Building Energy Management System (BEMS) (step 10).

For estimating zone temperature under given HVAC system operating conditions (e.g. power supply profile) and thermal conditions (e.g. values for the building environment such as air temperature, sunlight, etc.) obtained from BEMS The first artificial neural network model training (see FIG. 4) may be performed (step 20). Subsequently, the first artificial neural network model can be simulated using linear and partial linear equations (using pre- and post-processors, see Equations 2 to 11) (step 30), and using this, regarding price-based demand response scheduling The optimization problem can be formulated (see equations 5-25) (step 40). The optimal solution of equations 5 to 25 can be derived using mixed integer linear programming (MILP) (step 50). The above steps 20 to 50 may be referred to as an artificial neural network simulation method (1) for estimating the zone temperature according to the operating and thermal conditions of the HVAC system.

Referring again to FIG. 1, it is possible to calculate optimal solutions of each objective function according to a plurality of thermal conditions and electricity price profiles. Specifically, deriving an optimal solution for a standardized optimization objective function for price-based demand response scheduling of equations 5 to 25 using mixed integer linear programming (MILP) (step 50), building thermal conditions and electricity price Iterate over a variety of profiles (see Algorithm 1 -Part 1 of Figure 13) (step 60). Learning a second artificial neural network (e.g., deep neural network) for the optimal demand response schedule of the HVAC system based on the building thermal conditions, various profiles of electricity prices, and the derived optimal solutions, respectively (Algorithm 1-Part of Figure 13) 2) (step 70). The above steps 50 to 70 may be referred to as a SLAMP method.

Then, based on the trained second artificial neural network, it is possible to quickly and simply derive the optimal demand response scheduling of the HVAC system reflecting the time-varying electricity price and building thermal conditions (step 80).

According to another aspect of the present invention, the above-described algorithm may be divided into and defined as a method for generating a demand response determination model for the HVAC system and a method for performing a demand response for the HVAC system based on the generated model. .

2 is a flowchart of a method for generating a demand response determination model for an HVAC system according to an embodiment of the present invention. As illustrated in FIG. 2, a method for generating a demand response determination model for an HVAC system according to an embodiment of the present invention includes a heating, ventilating, and air-conditioning (HVAC) system for a building. As a method of generating a demand response (DR) decision model, the building may include a plurality of zones. However, the method according to an embodiment of the present invention can be applied to a building having a single zone.

First, by training a first artificial neural network based on a plurality of first training data, a zone temperature determination model that receives input power of the HVAC system and a thermal state of a building and outputs temperatures for each of the plurality of zones is generated. Can (step 210).

Subsequently, it includes a linear equation that simulates the zone temperature determination model generated as above, and can generate an objective function for the power supply schedule in which the optimal solution varies depending on the electricity price and thermal conditions (step 220). , Determine the respective optimal solutions of the objective function generated above for the plurality of electricity price profiles and the thermal state profile (step 230).

Then, by training a second artificial neural network based on a plurality of second training data each including an electric price profile, a thermal condition profile, and a determined optimal solution, the electric price profile and the thermal condition profile are input and power for the HVAC system is received. A demand response determination model that outputs a supply schedule may be generated (step 240).

3 is a flowchart of a method for performing demand response of an HVAC system according to an embodiment of the present invention. As shown in Figure 3, the method for performing the demand response of the HVAC system according to an embodiment of the present invention, the heating and air conditioning of buildings (Heating, ventilating, and air-conditioning, HVAC) system demand response As a method for performing (Demand Response, DR), the building may include a plurality of zones. However, the method according to an embodiment of the present invention can be applied to a building having a single zone.

First, a prediction profile (Profile) of an electricity price including information about an electricity price changing over time and a thermal condition prediction profile including information on a thermal condition of a building changing over time are obtained (step 310). , It is possible to determine the power supply schedule for the HVAC system of the building based on the information on the electricity price and the thermal condition (step 320). This determination can be made based on the demand response decision model described above.

Hereinafter, each step of the algorithm and / or method according to the above-described embodiments of the present invention will be described in more detail.

Artificial neural network structure and learning

As described above with reference to FIG. 2, a method of generating a demand response determination model for an HVAC system according to an embodiment of the present invention includes first training a first artificial neural network based on a plurality of first training data. By this, it is possible to generate an area temperature determination model that receives the input power of the HVAC system and the thermal state of the building and outputs the building temperature (step 210). When a building includes a plurality of zones, the zone temperature determination model can output temperatures for each of the plurality of zones.

In relation to this, first, the structure and learning of the first artificial neural network for generating the zone temperature determination model will be described in more detail. Figure 4 shows the structure of the first artificial neural network according to an embodiment of the present invention, Figure 5 is a detailed view of the J-th hidden layer of Figure 4, Figure 6 is a detailed view of the K-th hidden layer of Figure 4, Figure 7 Is a detailed view of the L-th concealment layer in FIG. 4.

) 는 해당 시점 이전 (t-τ) 부터 원래 시점 (t) 까지의 다중 구역 빌딩의 열적 상태 (E^t) 와 HVAC 시스템의 입력 전력 (P^t) 에 의해 결정될 수 있다. 만약 빌딩의 열적 상태 (E^t) 와 설비의 입력 전력 (P^t) 이 동일하다면 이전 시점에서의 실내 온도(

)가 현재 온도(

)에 직접적으로 영향을 준다. 즉, 현재 온도(

)는 빌딩의 열 동적 설계 (thermal dynamics) 의 상태 공간 모델 (state-space model) 에서 출력 결과일 뿐만 아니라 상태 변수로 사용된다. 이를 통해 열 동적 설계에 있어서, 과거의 입력 전력과 열적 상태를 입력으로 받고, 실내 온도에 대해 되먹임 루프로 구성된 인공 신경망 모델이 적합하다는 것을 알 수 있다. Room temperature in each zone (z) over an arbitrary time (t)

) Can be determined by the thermal state (E ^t ) of the multi-zone building from the time point (t-τ) to the original time point ( ^t ) and the input power (P ^t ) of the HVAC system. If the thermal condition of the building (E ^t ) and the input power of the facility (P ^t ) are the same, the room temperature at the previous point (

) Is the current temperature (

) Directly. That is, the current temperature (

) Is used as a state variable as well as an output result from the state-space model of the building's thermal dynamics. Through this, it can be seen that, in the thermal dynamic design, an artificial neural network model composed of a feedback loop for the room temperature receiving the input power and the thermal state of the past is suitable.

In relation, each of the plurality of first training data for training the first artificial neural network may include information regarding the conventional HVAC system input power and the thermal state of the building and the conventional HVAC system input power and thermal state of the building. It may include a temperature for each of the plurality of zones.

More specifically, as illustrated in FIGS. 4 to 7, the input layer of the first artificial neural network includes: HVAC system input power from a first predetermined time point to a current time point, and a second current time point from a predetermined second time point. A thermal state up to and a temperature for each of the plurality of zones from a predetermined third time point to a time point before the current time point may be included as an input neuron.

Further, in order to increase the accuracy of the model, multiple hidden layers using a sigmoid function or a rectified linear unit (ReLU) function as an active function may be used. As a result, as shown in FIGS. 4 to 7, the first artificial neural network model may be implemented in the form of a deep nonlinear auto-regressive network (D-NARX) with exogenous input. In addition, it is possible to modify some of the multi-layer networks with different forms and apply them to the algorithm.

According to one aspect, the first artificial neural network model may include a pre-processor for normalizing input data and a post-processor for reverting normalization for output data.

) 를 정규화 (normalization) 하기 위한 전 처리기 (pre-processor) 를 포함할 수 있다. 이는 학습 알고리즘 내 미분값 (gradient) 이 매우 작아져 인공 신경망의 학습 속도가 느려지는 것을 방지할 수 있다. 빌딩 에너지 관리 시스템 (Building Energy Management System, BEMS) 의 빌딩 운영 데이터 유효성을 기반으로, 빌딩의 열적 상태 (E^t) 는 대기 온도, 일조량, 풍력, 습도, 빌딩의 열적 부하 그리고 건물 사용 스케쥴 중 적어도 하나 이상을 포함할 수 있고, 각 항목들은 서로 다른 범위의 값을 갖는다. 후 처리기 (post-processor) 는 정규화된 실내 온도 (

) 를 원래 온도 (

) 의 범위로 역 변환하는 데에 사용된다. Specifically, the first artificial neural network model includes an input data set (

) May include a pre-processor for normalization. This can prevent the learning speed of the artificial neural network from slowing down because the gradient in the learning algorithm is very small. Based on the validity of the building operation data of the Building Energy Management System (BEMS), the thermal state of the building (E ^t ) is at least one of air temperature, sunshine, wind, humidity, thermal load on the building, and building usage schedule. It may contain the above, and each item has a different range of values. The post-processor is a normalized room temperature (

) To the original temperature (

) Is used to inverse transform to a range.

,

,

) 과 편향 (bias) 값 (

,

) 은 정규화 평균 제곱 오차 (normalized mean squared errors, NMSE) 를 기반으로 모든 뉴런 (입력, 은닉, 출력 뉴런) 에 대해 결정될 수 있다. 4 to 7, the first artificial neural network may include one or more weighting factors and one or more bias values. Weighting factors during the learning process (

,

,

) And bias values (

,

) Can be determined for all neurons (input, concealment, output neurons) based on normalized mean squared errors (NMSE).

은 실내 온도 (

) 의 모델 예측 값이며,

는 전체 학습 (혹은 테스트) 데이터 세트의 개수이다. 네트워크의 복잡성을 고려하여 가중 계수와 편향 값들은 스케줄링 기간 (

) 동안 일정하도록 설정될 수 있다. 제 1 인공 신경망의 가중 계수 및 편향 값은, 건물에 포함된 복수의 구역들마다 각각 결정될 수 있고, 따라서 각 구역에 대해 제 1 인공 신경망 모델은 상이한 가중 계수와 편향 값을 가질 수 있다. 이전 실내 온도 (

) 를 제외한 입력과 활성 함수, 네트워크 구조는 모든 구역에 대해 동일하게 설정될 수 있다. In Equation 1 above

Silver room temperature (

) Is the model predicted value,

Is the total number of training (or test) data sets. Considering the complexity of the network, the weighting factors and bias values are used for the scheduling period (

). The weighting coefficient and bias value of the first artificial neural network may be determined for each of a plurality of zones included in the building, and thus, for each zone, the first artificial neural network model may have different weighting coefficients and bias values. Previous room temperature (

Except for), the input, active function, and network structure can be set the same for all zones.

Partial linear simulation of a trained artificial neural network model

Referring back to FIG. 2, a method of generating a demand response determination model for an HVAC system according to an embodiment of the present invention derives a linear equation that simulates a zone temperature determination model generated by learning a first artificial neural network. You can (step 220).

As described above with reference to FIGS. 4 to 7, the first artificial neural network according to an aspect of the present invention includes a plurality of hidden layers, and at least one of the plurality of hidden layers is sigmoid ) Function or Rectified Linear Unit (ReLU) function can be used as the active function. According to one aspect, the linear equation simulating the zone temperature determination model may include a partial linear equation generated through partial linearization of an active function corresponding to a plurality of hidden layers included in the first artificial neural network.

)은 아래의 수학식 2 와 같이 계산될 수 있다. 여기서

는 특정 구역 (Z) , 특정 시간 (t) 에서 i 번째 입력 뉴런의 정규화된 입력 데이터를 의미한다.Specifically, the model for determining the zone temperature of the learned first artificial neural network is applied to the optimization problem addressed in step 220 of FIG. 2 by simulating a linear equation and a partial linear equation. 4 to 7, for example, the output of the j-th hidden neuron (hidden neuron) present in the first hidden layer (

) Can be calculated as in Equation 2 below. here

Denotes normalized input data of the i-th input neuron at a specific zone (Z) and at a specific time (t).

) 은 이전 은닉 뉴런 (j, j = k - 1) 의 활성 함수의 출력값 (

) 을 이용하여 아래의 수학식 3 과 같이 계산될 수 있다.Similarly, the output of the kth hidden neuron (

) Is the output of the active function of the previous hidden neuron (j, j = k-1) (

) Can be calculated as in Equation 3 below.

) 은

로 표현될 수 있으며, 활성 함수의 종류에 따라 하기의 수학식 4 의 첫 행 또는 두번째 행과 같이 계산될 수 있다. 시그모이드 함수와 ReLu 함수에 대하여 각각 수학식 4 의 첫 행과 두번째 행과 같이 나타낼 수 있다.Output of j-th active function in Equation (3)

) Silver

It can be represented by, and can be calculated as the first or second row of Equation 4 below according to the type of the active function. The sigmoid function and the ReLu function can be expressed as the first row and the second row of Equation 4, respectively.

8 is an exemplary diagram of partial linear approximation of an active function using a sigmoid function, and FIG. 9 is an exemplary diagram of partial linear approximation of an active function using a ReLU function. 8 to 9, the activity function of all hidden neurons can be represented using partial linearization as shown in Equations 5 to 8 below.

는 부분 선형 블록의 수를 나타내며,

는

의 최솟값,

는

의 s번째 선형 구분에 해당하는 활성 함수의 출력 값 (

) 의 기울기를 부분 선형화한 결과이다. 수학식 6 내지 8 에서

은

의 s번째 선형 구간에 해당하는 활성 함수 입력 값을 의미하며,

은 구분 선형화를 완성하기 위한 이진 변수를 의미한다. 비슷하게, 출력 뉴런의 출력은 아래의 수학식 9 와 같다. 여기서

은 마지막 은닉층에 존재하는 뉴런의 개수이다.In Equation 5

Denotes the number of partial linear blocks,

The

The minimum value of,

The

The output value of the active function corresponding to the s-th linear division of (

) Is the result of partial linearization of the slope. In equations 6 to 8

silver

Means the input value of the active function corresponding to the s-th linear interval of

Denotes a binary variable to complete classification linearization. Similarly, the output of the output neuron is shown in Equation 9 below. here

Is the number of neurons present in the last hidden layer.

) 는 선형 식별 함수 (linear identity function) 로 설정할 수 있다. 여기서 제 1 인공 신경망 모델이 다중 은닉층에 의해 빌딩의 열 동적 설계를 성공적으로 반영한다는 가정이 존재한다.For simplicity, the active function of the output layer (

) Can be set as a linear identity function. Here, the assumption exists that the first artificial neural network model successfully reflects the thermal dynamic design of the building by multiple hidden layers.

와

는 학습 데이터

의 최대값과 최소값을,

와

는 학습 데이터 내 실내 온도 (

) 의 최대값과 최소값을 의미한다. 보편성 손실의 문제 없이 수학식 2 내지 11 은 다른 구조의 인공 신경망 혹은 활성 함수에 적용될 수 있다.Equations 10 and 11 are expressions of the aforementioned pre- and post-processors. here

Wow

Is learning data

The maximum and minimum values of

Wow

Is the room temperature in the learning data (

) Means the maximum and minimum values. Equations 2 to 11 can be applied to artificial neural networks or active functions of other structures without problems of universality loss.

Demand response for supervised learning support of HVAC systems in multi-zone buildings

Formulating optimization problems using the trained artificial neural network (Formulation)

Referring back to FIG. 2, a method of generating a demand response determination model for an HVAC system according to an embodiment of the present invention includes a linear equation that simulates a zone temperature determination model generated by learning a first artificial neural network. Then, an objective function for the power supply schedule in which the optimal solution varies according to the electricity price and the thermal condition may be generated (step 220). That is, the optimization problem using the trained artificial neural network can be formulated.

Algorithm and / or method according to an embodiment of the present invention, using the simulated first artificial neural network model in order to optimally operate the HVAC system of a multi-zone building for a price-based demand response, is expressed as Equation 12 You can derive an optimal solution for the function.

Here, limit conditions to be considered can be set, and each exemplary item is as follows.

)에 대한 제한 조건 (수학식 13 내지 15) -Room temperature (

Restriction condition for (Equation 13 to 15)

)와 HVAC 시스템 입력 전력(P ^t ) 간 제한 조건 -Room temperature (

) And HVAC system input power (P ^t )

In addition to Equations 5 to 11, the constraints of Equations 16 to 22 below may be premised.

)에 관한 제한 조건 (수학식 23 내지 25) -Time delayed input power (

) Restrictions (Equations 23 to 25)

The power supply schedule of the algorithm and / or method according to an embodiment of the present invention minimizes the total electricity cost for electricity prices that change over time while maintaining the temperature of each of a plurality of zones included in the building within a predetermined range. It may be decided to let. Thus, according to one aspect, the objective function, which is the basis for achieving the optimal demand response, is to provide a power supply schedule for the HVAC system and to minimize the sum of excess amounts for the total electricity cost and a predetermined boundary temperature. It may be to determine the temperature for each of the plurality of zones as an optimal solution.

) 에 24 시간 동안의 HVAC 시스템의 입력 전력 (P^t) 을 곱하여 합산한 것으로, 설비 운영 비용을 낮추는 것을 목적으로 할 수 있다. More specifically, referring to Equation 12, the objective function of Equation 12 is largely divided into two terms, and the first term is an electricity price that changes with time (

) Multiplied by the input power (P ^t ) of the HVAC system for 24 hours, which can be aimed at lowering the operation cost of the facility.

) 가 최대 한도 (

) 혹은 최소 한도 (

) 를 넘어가게 될 경우, 한도 대비 초과된 온도 차이 (

혹은

) 에 대한 불이익 (penalty) 을 나타내는 항이다. 이는 각각 수학식 13 과 14 에 해당한다. 다시 말해서, 온도 경계 조건 (temperature boundary condition) 은 수학식 12 에서 온도 차 (

혹은

) 를 포함시킴으로 인해 완화될 수 있으며, 이는 신뢰성 있게 해당 최적화 문제를 해결할 수 있도록 한다. 환언하면, 본 발명의 일 측면에 따른 알고리즘 및/또는 방법의 목적 함수에 있어서, 복수의 구역 각각의 온도에 대한 경계 조건은, 경계 온도의 하한에 대한 제 1 오프셋 및 경계 온도의 상한에 대한 제 2 오프셋을 허용하도록 구성되어 온도 경계 조건을 완화할 수 있다. The second term of Equation 12 is the room temperature corresponding to the input power (

) Is the maximum limit (

) Or minimum (

), The temperature difference exceeded the limit (

or

) Is a term that represents a penalty. This corresponds to Equations 13 and 14, respectively. In other words, the temperature boundary condition is the temperature difference (

or

) Can be mitigated, which reliably solves the optimization problem. In other words, in the objective function of the algorithm and / or method according to an aspect of the present invention, the boundary condition for the temperature of each of the plurality of zones includes: a first offset to the lower limit of the boundary temperature and a limit to the upper limit of the boundary temperature. It is configured to allow 2 offsets to relieve temperature boundary conditions.

Here, based on the thermal preference of residents existing in each zone, the maximum and minimum limits, or the first offset and the second offset may be set differently for a plurality of zones included in the building. However, in order to prevent excessive increase or decrease in temperature, an enhanced boundary condition such as Equation 15 may be set. Accordingly, the first offset may not be set to exceed the first enhanced offset, and the second offset may be set not to exceed the second enhanced offset. The first enhanced offset and the second enhanced offset may be set to have the same size as in Equation 15, or may be set differently from each other.

) 간의 관계를 의미한다. 특히 수학식 2 는 수학식 16 과 등가적으로 표현되며, 여기서 입력 변수 (

) 는 제어 가능한 변수 (P^t), 되먹임 변수 (

), 환경 입력 변수 (E^t) 로 구성될 수 있다. 수학식 17 은, 수학식 16 의 선형 구간 s 에 해당하는 활성 함수 입력 (

) 이 도 8 내지 도 9 에서와 같이 해당 구간의 길이 (

) 의 합과 임의의 큰 값을 갖는 음수 r₀ 에 대해 등가적으로 표현됨을 나타내며, 이와 비슷하게, 다른 층에 존재하는 활성 함수 입력 (

) 의 선형 표현 (

) 을 이용하여 t수학식 3 이 등가적으로 수학식 18 과 같이 표현될 수 있음을 알 수 있다. The second restriction condition of Equations 16 to 22 is the input power (P ^t ), which is one of the input variables of the learned first artificial neural network model, and the temperature for each zone, which is the output variable (

). In particular, Equation 2 is equivalent to Equation 16, where the input variable (

) Is a controllable variable (P ^t ), a feedback variable (

), And an environment input variable (E ^t ). Equation 17 is the input of the active function corresponding to the linear interval s of Equation 16 (

) As in FIGS. 8 to 9, the length of the section (

) And represents an equivalent expression for a negative r ₀ with an arbitrary large value, similarly, inputting the active function present in another layer (

) Linear representation of (

), It can be seen that t equation 3 can be equivalently expressed as equation 18.

) 의 되먹임 루프를 반영한 것으로, 이 중 수학식 19 는 학습된 인공 신경망 모델의 되먹임 입력 변수 (

,

) 와 출력 변수 (

) 와의 관계를 의미한다. 운영 비용 계산 시 시간의 흐름에 따라 변하는 전기 가격 (C_E) 을 적용하고, 본 발명의 일 실시예에서 한 시간 별로 전기 가격이 변동될 때, 단위 샘플링 시간 (unit sampling time,

) 은 한 시간으로 설정될 수 있다. 수학식 20 에서와 같이,

(

) 는 하루 전에 예측된 온도 (

) 에 대하여 정해질 수 있다. 나아가, 수학식 21 은 수학식 10 에서 얻어진 제어 가능 변수와 되먹임 입력 변수가 -1 과 1 사이의 값을 갖는다는 것을 의미한다. 이는 입력 전력에 해당하는 변수 (P^t) 가 0 보다 크거나 같으며, 정격 전력 (

) 이하의 값을 가짐을 의미한다. 이와 같이 수학식 22 에서도 정규화된 변수 (

) 의 최대값과 최소값을 확인할 수 있다. 여기서 정규화된 변수 (

) 는 수학식 5 내지 9 를 이용하여 수학식 18 에 연결되고, 수학식 11 을 이용하여 실내 온도로 역 변환될 수 있다. Equation 19 and Equation 20 represent the room temperature (

), And Equation 19 is the feedback input variable of the trained artificial neural network model (

,

) And output variables (

). When calculating the operating cost, the electric price (C _E ) that changes according to the passage of time is applied, and when the electric price changes for each hour in an embodiment of the present invention, the unit sampling time (unit sampling time,

) Can be set to one hour. As in Equation 20,

(

) Is the temperature predicted one day ago (

). Furthermore, Equation 21 means that the controllable variable and the feedback input variable obtained in Equation 10 have a value between -1 and 1. This means that the variable (P ^t ) corresponding to the input power is greater than or equal to 0, and the rated power (

) It means to have the following values. Likewise, in Equation 22, the normalized variable (

) Of the maximum and minimum values. Where the normalized variable (

) Is connected to Equation 18 using Equations 5 to 9, and may be inversely converted to room temperature using Equation 11.

) 에는 HVAC 시스템을 정지시키고, 예비 냉방 (pre-cooling) 을 위해 사람들이 일을 하기 위해 빌딩으로 모이기 전 이른 아침에 HVAC 시스템을 작동시키는 것을 반영할 수 있다. 따라서, 목적 함수에서, 설정에 따라 미리 결정된, 건물에 거주자가 존재하지 않는 시간에 대한 HVAC 시스템 입력 전력은 0 으로 설정될 수 있다. In order to apply the trained artificial neural network to the optimization problem, Equation 23 and Equation 24 refer to the constraint condition of the input neuron with respect to the time delayed input power of the HVAC system. Moreover, in Equation 25, the time zone during which there are few residents in the building after working hours (

) May reflect stopping the HVAC system and operating the HVAC system early in the morning before people come to the building to do work for pre-cooling. Thus, in the objective function, the HVAC system input power for a time when there is no occupant in the building, predetermined according to the setting, may be set to zero.

,

,

,

,

)가 달라질 수 있다. 수학식 5 내지 25 는 선형 방정식과 이진 변수들을 포함하므로, 본 발명의 일 측면에 따른 알고리즘 및/또는 방법의 최적화 문제는 혼합 정수 선형 프로그래밍 (Mixed-integer linear programming, MILP) 을 이용을 이용하여 해결될 수 있다. 이를 통해 최적의 수요 반응 일정을 도출할 수 있다. The optimization problems of Equations 5 to 25 can be applied to various building models without significant modification. However, depending on the thermal dynamic design of the building and the load characteristics of the HVAC system, for example, five variables within the first artificial neural network (

,

,

,

,

) May vary. Since Equations 5 to 25 include linear equations and binary variables, the optimization problem of the algorithm and / or method according to an aspect of the present invention is solved by using mixed-integer linear programming (MILP). Can be. Through this, it is possible to derive an optimal demand response schedule.

Referring again to FIG. 2, a method of generating a demand response determination model for an HVAC system according to an embodiment of the present invention may determine respective optimal solutions of objective functions for a plurality of electricity price profiles and thermal state profiles (Step 230). Here, the electricity price profile may include information on electricity prices that change over time, and the thermal status profile may include information on thermal conditions that change over time. The interval at which the electricity price and / or the thermal state changes may be different depending on the implementation, and may include information, for example, that the electricity price is changed in units of one hour. The interval between changes in the electricity price and the thermal state may be the same or may be different from each other. Meanwhile, the step of determining the optimal solutions of the power schedule objective function of FIG. 2 (step 230) may be configured to determine the optimal solution of the power schedule objective function using mixed-integer linear programming (MILP). have.

Optimal demand response through super prediction based meta prediction method

Referring again to FIG. 2, respective optimal solutions derived according to a plurality of different electricity price profiles and thermal state profiles can be obtained (step 230). Power for the HVAC system by receiving the electrical price profile and the thermal state profile by training the second artificial neural network based on a plurality of second training data each including the electrical price profile, the thermal state profile, and the determined optimal solution A demand response determination model that outputs a supply schedule can be generated (step 240).

) 과 환경 변수 (E^t) (예를 들어, 열적 조건) 가 주어졌을 때, SLAMP 방법 (도 10 내지 12 와 도 13 의 알고리즘 1 참조) 을 통해 수학식 5 내지 25 의 최적 해를 즉시 구할 수 있다. 이는 최적 수요 반응 일정에 대한 계산 시간을 현저하게 감소시키며, 배전 설비 운영자로 하여금 배전 망 내 여러 버스에서의 전기 가격 (

) 을 최적으로 설정하는 데에 영향을 줄 수 있다. 이는 배전 망 내에 넓게 분포하는 HVAC 시스템의 전체 부하 수요 프로파일을 예측할 수 있도록 하기 때문이며, SLAMP 방법은 앞서 언급된 바와 같이, 빌딩 운영 데이터를 이용한 인공 신경망 모사 방법으로 구현될 수 있다. Specifically, electricity prices (

) And the environment variable (E ^t ) (e.g., thermal conditions), the SLAMP method (see Algorithm 1 in FIGS. 10 to 12 and 13) can be used to immediately obtain the optimal solution of equations 5 to 25 have. This significantly reduces the calculation time for the optimal demand response schedule, and allows the distribution facility operator the electricity prices (on multiple buses in the distribution network).

) Can be optimally set. This is because the entire load demand profile of the HVAC system widely distributed in the distribution network can be predicted, and the SLAMP method can be implemented by an artificial neural network simulation method using building operation data as mentioned above.

) 를 도출하고, 현재 입력 (

) 의 과거 데이터에 대하여 이에 상응하는 운영 비용 (

) 과, 경계 온도와의 차이와 연관된 온도 패널티 항 (

) 을 계산할 수 있다. 도 10 에서와 같이, 예를 들어 BEMS 데이터베이스 (1010) 로부터의 입력 (m) 은 SLAMP 알고리즘 (1020) 에 의해 제 2 인공 신경망 (예를 들어, 심층 신경망) 모델의 학습을 위한 데이터 (

) 로 확장되어 입력 데이터 ( I ) 와 출력 데이터 ( O ) 로 재 가공될 수 있다. 심층 신경망 모델의 성능 향상을 위해 입력 데이터 ( I ) 는 학습 데이터 ( M ) 의 시간 지연된 데이터를 포함함으로써 확장되고, 도 11 에 도시된 바와 같이 데이터베이스 (1110) 로부터의 데이터로부터 획득될 수 있는 입력 데이터 ( I ) 와 출력 데이터 ( O ) 는 무작위로 섞여 심층 신경망 모델 (1120) 을 위한 학습 데이터 (

) 와 테스트 데이터 (

) 가 다양한 전기 가격 (

) 및 환경 변수 (E^t) 프로파일, 그리고 이에 해당하는 실내 온도 (

) 와 입력 전력 (P^t) 의 최적 스케줄 결과를 포함하도록 한다. In particular, referring to Part 1 in Algorithm 1 of FIG. 13, the optimization problems of Equations 5 to 25 may be solved through iteration offline. As a result, the optimal solution (

), And the current input (

) Corresponding operating costs for historical data (

) And the temperature penalty term associated with the difference between the boundary temperature (

). As in FIG. 10, for example, input m from the BEMS database 1010 is data for training a second artificial neural network (eg, deep neural network) model by the SLAMP algorithm 1020 (

) And can be reprocessed as input data ( I ) and output data ( O ). In order to improve the performance of the deep neural network model, the input data I is expanded by including time-delayed data of the training data M , and the input data can be obtained from data from the database 1110 as shown in FIG. 11. ( I ) and output data ( O ) are randomly mixed to learn data for the deep neural network model 1120 (

) And test data (

) Various electricity prices (

) And environmental variable (E ^t ) profiles, and corresponding room temperature (

) And input power (P ^t ).

) 와 테스트 데이터 (

) 를 이용하여 심층 신경망의 학습 및 테스트를 각각 진행한다. 이 때, 네트워크 매개 변수들 (

) 에 대해 진행되며, 매개 변수들은 입력 데이터 ( I ) 의 시간 지연 ( τ ) 와 은닉 층의 수 ( G ), 각각의 층들의 활성 함수 ( F ) 와 은닉 뉴런 ( U ) 을 포함한다. 최적의 입력 전력 (P^t) 이 메타 예측 (meta-predicted) 되면, 상응하는 실내 온도 (

) 는 제 1 인공 신경망 모델 (도 12 의 1240, 도 4 내지 7 참조) 을 학습시킨 구역 온도 결정 모델을 통해 도출될 수 있다. Referring back to FIG. 13, in Part 2, learning data (

) And test data (

) To learn and test deep neural networks, respectively. At this time, the network parameters (

), The parameters include the time delay ( τ ) of the input data ( I ) and the number of hidden layers ( G ), the active function of each layer ( F ) and the hidden neurons ( U ). When the optimal input power (P ^t ) is meta-predicted, the corresponding room temperature (

) Can be derived through a zone temperature determination model trained on a first artificial neural network model (see 1240 in FIG. 12, FIGS. 4 to 7).

), 운영 비용 (E_C), 온도 패널티 항 (T_V) 들의 가중치 합 (

, 수학식 1) 을 이용하여 평가될 수 있다. 매개 변수 세트 (

) 중 정규화 평균 제곱 오차 (NMSE) 의 최대값 (

) 을 도출하는 세트 (

) 를 선택하여 심층 신경망 모델 (

) 을 구현할 수 있다. 최종적으로, 다음 스케줄링 기간에 해당하는 전기 가격 (

) 과 환경 변수 (E^t) 의 예측 값에 대해 입력 전력 (P^t), 실내 온도 (

) 의 최적 수요 반응 스케줄을 도출할 수 있다 (도 12 참조). 도 12 에 도시된 바와 같이, 가격 프로파일 (1210) 및 환경 변수 (1220) 가 입력되면, 심층 신경망 (1230) 및 인공 신경망들 (1240) 을 기반으로 최적의 공급 전력 및 각각의 구역에 대한 최적의 온도가 도출될 수 있다. Performance of the deep neural network In one aspect of the invention, the input data (D) every day input power value for the profile (P ^t), room temperature in the (

), Operating cost (E _C ), weighted sum of temperature penalty terms (T _V ) (

, It can be evaluated using Equation 1). Parameter set (

), The maximum value of the normalized mean squared error (NMSE) (

Set to derive ()

Select the deep neural network model ()

) Can be implemented. Finally, the posting price for the next scheduling period (

) And environmental variables (the input power for the prediction value (P ^t), a room temperature of E ^t) (

), An optimal demand response schedule can be derived (see FIG. 12). As shown in FIG. 12, when the price profile 1210 and the environmental variable 1220 are input, the optimal supply power and the optimal for each zone based on the deep neural network 1230 and artificial neural networks 1240 Temperature can be derived.

As described above, according to an aspect of the present invention, a new algorithm and / or method based on machine learning for optimal demand response of an HVAC system in a multi-zone building is disclosed. The first artificial neural network model in the algorithm and / or method is trained based on a supervised learning algorithm to reflect the complex thermal dynamic design of the building and can be simulated using partial linear equations. This can be used to formalize the optimization problem, so that an optimal solution can be derived using mixed integer linear programming. Here, the optimal solution is obtained based on the past data on the electrical price and the thermal state of the building, and the corresponding optimal solutions on the plurality of electrical prices and the thermal state of the building are second artificial neural networks (eg, deep neural networks) models in the SLAMP method. It is used to train, and this model can derive the optimal input power schedule of the HVAC system. The SLAMP method is an effective method for scheduling an optimal demand response when considering practical applicability, operating cost, and calculation time. In addition, this SLAMP method can ensure the load shifting function of the HVAC system by reflecting the electric price changing over time and the thermal preference of the occupant.

The method according to the present invention described above can be embodied as computer readable codes on a computer readable recording medium. The computer-readable recording medium includes any type of recording medium that stores data that can be read by a computer system. For example, there may be a read only memory (ROM), a random access memory (RAM), a magnetic tape, a magnetic disk, a flash memory, and an optical data storage device. In addition, the computer-readable recording medium may be distributed over computer systems connected by a computer communication network, and stored and executed as code readable in a distributed manner.

Although described above with reference to the drawings and examples, the protection scope of the present invention is not meant to be limited by the drawings or the examples, and those skilled in the art of the present invention described in the claims below It will be understood that various modifications and changes can be made to the present invention without departing from the spirit and scope.

Specifically, the described features can be implemented in digital electronic circuitry, or computer hardware, firmware, or combinations thereof. The features can be implemented in a computer program product implemented in storage in a machine-readable storage device, eg, for execution by a programmable processor. And the features can be performed by a programmable processor executing a program of instructions for performing the functions of the described embodiments by operating on input data and generating output. The described features include at least one programmable processor, at least one input device, and at least one output device coupled to receive data and directives from a data storage system and to transmit data and directives to the data storage system. It can be executed in one or more computer programs that can be executed on a programmable system comprising a. A computer program includes a set of directives that can be used directly or indirectly within a computer to perform a specific action on a given result. A computer program is written in any form of programming language, including compiled or interpreted languages, and is included as a module, element, subroutine, or other unit suitable for use in other computer environments, or as a standalone program. Can be used in any form.

Suitable processors for the execution of the program of directives include, for example, both general purpose and special purpose microprocessors, and either a single processor or multiple processors of other types of computers. Also suitable for implementing computer program instructions and data embodying the described features are storage devices suitable for example, semiconductor memory devices such as EPROM, EEPROM, and flash memory devices, magnetic devices such as internal hard disks and removable disks. Devices, magneto-optical disks and all forms of non-volatile memory including CD-ROM and DVD-ROM disks. The processor and memory may be integrated within application-specific integrated circuits (ASICs) or added by ASICs.

The present invention described above is described based on a series of functional blocks, but is not limited by the above-described embodiments and the accompanying drawings, and various substitutions, modifications, and changes without departing from the spirit of the present invention. It will be apparent to those skilled in the art that the present invention is possible.

Combinations of the above-described embodiments are not limited to the above-described embodiments, and various forms of combinations may be provided as well as the above-described embodiments according to implementation and / or needs.

In the above-described embodiments, the methods are described based on a flowchart as a series of steps or blocks, but the present invention is not limited to the order of steps, and some steps may occur in a different order than the steps described above or simultaneously. have. In addition, those skilled in the art may recognize that the steps shown in the flowchart are not exclusive, other steps are included, or one or more steps in the flowchart may be deleted without affecting the scope of the present invention. You will understand.

The foregoing embodiments include examples of various aspects. It is not possible to describe all possible combinations for representing various aspects, but a person skilled in the art will recognize that other combinations are possible. Accordingly, the present invention will be said to include all other replacements, modifications and changes that fall within the scope of the following claims.

Claims (22)

As a method of generating a demand response (DR) decision model for a building's heating, air conditioning (HVAC) system,
Generating a zone temperature determination model that receives the input power of the HVAC system and the thermal state of the building and outputs the temperature of the building by training the first artificial neural network based on the plurality of first training data;
Generating an objective function for a power supply schedule that includes a linear equation that simulates the zone temperature determination model, and that an optimal solution varies according to electricity price and thermal conditions;
Determining respective optimal solutions of the objective function for a plurality of electricity price profiles and a thermal state profile; And
By training a second artificial neural network based on a plurality of second training data each including the electric price profile, the thermal condition profile, and the determined optimal solution, an electric price profile and a thermal condition profile are input to the HVAC system. A method of generating a demand response determination model for an HVAC system, comprising generating a demand response determination model outputting a power supply schedule.
According to claim 1,
The building includes a plurality of areas,
The zone temperature determination model is a method for generating a demand response determination model for an HVAC system, outputting temperatures for each of the plurality of zones.
According to claim 2,
Each of the plurality of first training data,
A demand response determination model for an HVAC system, comprising information about a conventional HVAC system input power and thermal condition of a building and temperature for each of the plurality of zones according to the conventional HVAC system input power and a building thermal condition. How to generate. According to claim 1,
The thermal condition of the building,
A method for generating a demand response determination model for an HVAC system, comprising at least one of air temperature, sunshine, wind, humidity, thermal load of the building, and building usage schedule.
According to claim 1,
A method for generating a demand response determination model for an HVAC system, wherein the information on the input power of HVAC system and thermal condition of a building is obtained from a Building Energy Management System (BEMS).
According to claim 2,
The input layer of the first artificial neural network,
Enter the HVAC system input power from the first predetermined time point to the current time point, the thermal state from the second predetermined time point to the current time point, and the temperature for each of the plurality of zones from the third predetermined time point to the time point before the current time point A method for generating a demand response decision model for an HVAC system, including as a neuron.
The method of claim 6,
The first artificial neural network,
A method of generating a demand response decision model for an HVAC system, implemented in the form of a deep nonlinear auto-regressive network (D-NARX).
According to claim 1,
The first artificial neural network includes a plurality of hidden layers,
A method of generating a demand response determination model for an HVAC system, wherein at least one of the plurality of hidden layers uses a sigmoid function or a rectified linear unit (ReLU) function as an active function.
According to claim 1,
The first artificial neural network comprises a pre-processor for normalizing input data and a post-processor for reverting normalization for output data. How to generate.
According to claim 2,
The first artificial neural network includes one or more weighting factors and one or more bias values,
The weighting coefficients and bias values are determined based on a normalized mean squared error (NMSE), a method of generating a demand response determination model for an HVAC system.
The method of claim 10,
The method of generating a demand response determination model for an HVAC system, wherein the weighting coefficient and the bias value are respectively determined for each of the plurality of zones.
According to claim 1,
The linear equation simulating the zone temperature determination model,
And a partial linear equation generated through partial linearization of an active function corresponding to a plurality of hidden layers included in the first artificial neural network.
According to claim 2,
The power supply schedule,
A method of generating a demand response determination model for an HVAC system that is determined to minimize the total electricity cost for a time-varying electricity price while maintaining the temperature of each of the plurality of zones within a predetermined range.
The method of claim 13,
The objective function is,
To determine the power supply schedule for the HVAC system and the temperature for each of the plurality of zones as the optimal solution, to minimize the sum of excess amounts for the total electricity cost and a predetermined boundary temperature. How to generate a demand response decision model for
According to claim 2,
The boundary condition for the temperature of each of the plurality of zones of the objective function is:
The first offset for the lower limit of the boundary temperature and the second offset for the upper limit of the boundary temperature, the first offset and the second offset can be set differently for each of the plurality of zones, HVAC system How to generate a demand response decision model for
The method of claim 15,
A method for generating a demand response determination model for an HVAC system, wherein the first offset does not exceed a first enhanced offset and the second offset does not exceed a second enhanced offset.
According to claim 1,
In the objective function, a method for generating a demand response determination model for an HVAC system, wherein the HVAC system input power for a time when there is no occupant in the predetermined building is set to zero.
According to claim 1,
The electricity price profile includes information on electricity prices that change over time,
The thermal condition profile includes a method for generating a demand response determination model for an HVAC system, which includes information about a thermal condition of a building that changes over time.
According to claim 1,
Determining the optimal solutions of the power schedule objective function, respectively,
A method of generating a demand response determination model for an HVAC system that determines an optimal solution of the power schedule objective function using mixed-integer linear programming (MILP).
As a method for performing a demand response (DR) to the building's heating and air conditioning (Heating, ventilating, and air-conditioning, HVAC) system,
Obtaining a prediction profile (Profile) of the electricity price including information on the electricity price changing over time and a thermal condition prediction profile including information on the thermal condition of the building changing over time; And
And determining a power supply schedule for the HVAC system in the building based on the information on the electricity price and the information on the thermal condition.
The method of claim 20,
The building comprises a plurality of zones, a method for performing a demand response of an HVAC system.
The method of claim 20,
Determining the power supply schedule,
A method for performing a demand response of an HVAC system, wherein the power supply schedule is determined using a demand response determination model generated by any one of claims 1 to 14.

Images