Building energy-saving scheme intelligent generation method based on two-stage agent-assisted evolution

Abstract

The invention discloses a building energy-saving scheme intelligent generation method based on two-stage agent-assisted evolution, and belongs to the field of building energy conservation, and the method comprises the steps: generating a uniform feasible solution through a rejection sampling method according to a residential building model decision variable and range, obtaining a real target value through EnergyPlus simulation, and storing the real target value in a sample set DB; uniformly constructing a reference vector division space in a target space, dividing a DB sample into a corresponding vector file according to a nearest distance, and constructing an initial RBF global proxy model; in the first stage, selecting a non-dominated solution with the minimum congestion density from each vector file as an initial population, evaluating individuals by using a global agent model, selecting a filling solution according to a rule, updating a local agent model, entering the second stage when a hypervolume increment value is greater than a threshold value, otherwise, directly entering the step S5, and dynamically updating a reference vector according to the quality of the solution in the second stage. By adopting the method, the efficiency and the precision of building energy-saving optimization are remarkably improved by combining the agent model and the staged evolution strategy.

Description

Technical Field

[0001] The present invention relates to the technical field of building energy conservation, and in particular to an intelligent generation method for building energy conservation schemes based on two-stage agent-assisted evolution. Background Art

[0002] In actual situations, building energy-saving design is essentially a typical multi-objective optimization problem, which has many conflicting performance indicators such as building energy consumption and user discomfort.

[0003] Evolutionary algorithms (EAs) are a class of global search methods based on the principles of biological evolution. Currently, EAs have demonstrated strong performance in multi-objective optimization, constrained optimization, and dynamic optimization. Since the 1990s, new technologies combining building energy simulation with multi-objective evolutionary optimization algorithms have gradually gained widespread attention among scholars. These technologies have significantly reduced the complex manual trial-and-error process, improved the quality of optimal solutions, and achieved high-performance building design. Some methods, such as Asadi et al., proposed a multi-objective optimization model that optimizes energy consumption, investment costs, and thermal comfort using the Chebyshev decomposition method. They also verified the feasibility of the proposed multi-objective model in practical situations using an existing residential building requiring renovation as an example. Malatji et al. developed a multi-objective optimization model formulated as a multi-objective optimization problem constrained by net present value, initial investment, energy target, and payback period. The model was solved using a genetic algorithm to help decision makers make optimal decisions when investing in energy-efficient building renovations. Han et al. designed a fast and convenient multi-objective optimization scheme for the early planning of eco-cities and communities. This scheme takes building energy consumption, carbon emissions, and initial costs as three sub-objectives, sets different weights for the three sub-objectives according to the decision maker's plan, and uses the BIN method to calculate building energy consumption. Harkouss et al. used the NSGA-II algorithm to minimize heat and electricity demand and life cycle costs to improve the design of net-zero energy buildings for the energy performance issues of specific buildings. Fesanghary et al. proposed a multi-objective optimization method that uses the Harmony Search (HS) algorithm and selects several structural parameters as design variables. Bui et al. studied a building energy-saving design algorithm based on multi-objective firefly intelligent optimization based on the design of an adaptive facade system, which can effectively respond to changes in the temperature environment. Ferdinando et al. considered the renovation of existing buildings and designed a building energy-saving design method that integrates the non-dominated sorting genetic algorithm and EnergyPlus, and applied it to the renovation of typical local residential buildings.

[0004] The above-mentioned evolutionary optimization methods have achieved promising results in building energy-saving design. However, these algorithms still suffer from high runtime costs, as they require computationally intensive building energy consumption software to continuously evaluate individual fitness values. To address this time-consuming evaluation issue, researchers in the late 20th century began researching surrogate-assisted evolutionary algorithms (SAEAs). The idea is to construct a surrogate model of the original problem based on historical solution data and use it to replace the computationally expensive original optimization problem to predict the target or constraint values ​​of candidate solutions. Because SAEAs can significantly reduce algorithm runtime, they have become an important approach for solving building energy-saving optimization problems. For example, Magnier and Haghighat used TRNSYS simulation to train a neural network and coupled the trained neural network with a genetic algorithm to optimize thermal comfort and energy consumption. This method significantly shortened the optimization time for building energy-saving design problems. Chen et al. used a feedforward neural network with a single hidden layer to identify temperature in intelligent buildings and proposed a particle swarm optimization-based method for designing building structural parameters. Khan et al. used the moving least squares method to construct a surrogate model from extremely expensive CFD simulations and used this surrogate model to optimize the design of a hospital ward ventilation system. Gengembre et al. combined the Kriging surrogate model with the particle swarm optimization algorithm to optimize the life cycle cost of a single-area building model.

[0005] The above methods are primarily designed for unconstrained building energy-saving design problems. However, actual building energy-saving design faces numerous constraints, such as the ratio of window area to wall area, building orientation angle restrictions, and the coordination of window areas between thermal zones. Therefore, building energy-saving design problems are essentially expensive constrained multi-objective optimization problems (ECMOPs). To solve ECMOPs, Li et al. proposed a data-driven CMOEA that effectively balances convergence, diversity, and feasibility by employing a feasible region localization strategy and a performance improvement exploration mechanism. Gu et al. proposed a random forest-assisted adaptive multi-objective particle swarm optimization algorithm. This algorithm leverages a random forest model to accelerate population search speed and an adaptive random ranking strategy and particle state update rules to balance population search performance. Wang et al. proposed an agent-assisted global optimization algorithm that utilizes a constraint improvement strategy and multiple metrics to build a two-layer selection framework. The adaptive selection framework dynamically adjusts the population's search direction, bringing the population closer to the feasible region. Long et al. proposed a data-driven co-evolutionary algorithm, constructed a co-evolutionary exploration framework, and improved the accuracy of the agent model through an adaptive selection strategy. Sun et al. proposed an optimization algorithm for constrained mixed problems based on a multi-objective adaptive agent model, using multiple evolutionary operators to process different types of decision variables, and constructed a constraint model using a classification agent model, and optimized the generation process of the filling solution through the Kriging model. The above method has shown good performance in solving ECMOPs, but it has not been applied to building energy-saving optimization problems. In addition, the research results of many literatures show that this type of problem has an irregular PF, which is rarely considered by existing methods. Therefore, if the existing method for solving ECMOPs is directly used in the building energy-saving optimization design problem, the resulting design solution may not be ideal. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for intelligently generating building energy-saving solutions based on two-stage agent-assisted evolution to solve the problems existing in the background technology.

[0007] To achieve the above objectives, the present invention provides a method for intelligently generating building energy-saving solutions based on a two-stage agent-assisted evolution, comprising the following steps:

[0008] S1. Based on the decision variables and their value ranges of the residential building model, a rejection-based sampling method is used to generate several uniformly distributed feasible solutions. The solutions are then introduced into the constructed EnergyPlus simulation model to obtain the true target values ​​of the solutions and store them in the sample set DB.

[0009] S2. Uniformly construct N in the target space W The reference vectors divide the target space; the initial sample in DB is divided into the reference vector closest to the sample in the target space, forming the vector file V of the vector i And use the samples in DB to build the initial RBF global surrogate model of M objective functions;

[0010] S3, enter the first stage, select the non-dominated solution with the minimum crowding density from each vector archive as the initial population, then execute the evolution operator based on the fixed vector, use the global agent model to evaluate the individual to obtain the target prediction value of the offspring, select the filling solution according to the first stage model management method, and update the local agent model according to the current sample, and calculate the hypervolume increment ΔHV value of the current sample every h generations. When ΔHV is greater than the threshold α, the algorithm enters S4, otherwise it directly enters S5;

[0011] S4. Enter the second stage, dynamically update the reference vector based on the quality of the solution obtained in each generation; use the samples in each vector archive to build an RBF local proxy model for each reference vector; after building the local proxy model, execute the vector classification strategy, and execute the search strategy adaptive selection mechanism based on the classification results; after obtaining the offspring according to the selected search strategy, select the filling solution according to the second stage model management strategy, dynamically update the reference vector based on the quality of the solution obtained in each generation, and update the local proxy model based on the current sample;

[0012] S5. Output several non-dominated feasible solutions.

[0013] Preferably, the steps of dividing the target space with reference vectors and constructing the proxy model in S2 are as follows:

[0014] S21. Use uniformly distributed fixed reference vectors to guide population evolution, thereby achieving cross-regions of infeasibility caused by constraints, ensuring that the population covers the feasible solution space as early as possible and gradually approaches the Pareto frontier. Specifically, uniformly distributed reference vectors are generated in the M-dimensional target space, dividing the target space into several subregions. Each reference vector corresponds to a subspace, and the population evolves based on these fixed reference vectors.

[0015] S22. Based on the sample information in the database, radial basis functions are used to obtain fitting functions for the building's annual energy consumption and user discomfort hours, respectively. These fitting functions are defined as proxy models. In subsequent evolutionary processes, when a new solution is generated, the proxy model can be used to predict the new solution's annual energy consumption and user discomfort hours.

[0016] Preferably, the steps of S3 are as follows:

[0017] S31, based on the individuals assigned to each vector, respectively calculate the APD value of each individual, and obtain a non-dominated solution according to the obtained APD value;

[0018] S32, using the non-dominated solution as the initial population and the agent model obtained in S22 as the objective function, execute the RVEA evolution operation to obtain multiple new individual solutions;

[0019] S33. Calculate the constraint violation values ​​of all new solutions, select NM individuals with the smallest constraint violation values, and use the proxy model to predict the objective function values ​​of these individuals. Select the individual that contributes most to HV improvement as the filling solution, truly evaluate the objective function value of the filling solution, and update the proxy model.

[0020] Preferably, the content of S31 is as follows:

[0021] RVEA uses angle penalized distance (APD) to select the appropriate individual for each reference vector. For the current reference vector, the smaller the APD value of an individual, the higher the probability of its selection. The specific expression of APD is as follows:

[0022] APd j =(1+Pe(θ j ))·||f j || (1)

[0023] Among them, ||f j || is the distance from the jth individual to the origin in the target space; ||f j The smaller the || value, the better the convergence of the individual; θ j is the angle between the jth individual and the reference vector to which it is assigned; Pe(θ j ) is the penalty function.

[0024] Preferably, the content of S33 is as follows:

[0025] For a candidate solution set HS, assume that the corresponding reference point is Individual x i The objective function value of ∈HS is (f1(x i ),f2(x i ),...,f m (x i ), then the HV contribution of the individual relative to the reference point Z is:

[0026]

[0027] in, Represents the maximum point of the objective function of the set HS in the jth dimension; set the reference point Z to 1.1 times of .

[0028] Preferably, the content of S4 is as follows:

[0029] S41. Dynamically update the reference vector based on the quality of the solution obtained in each generation. The goal of the second stage is to accurately locate the true PF of the problem and find more solutions located on the PF. Based on the feasible solution information obtained in the first stage, the reference vector of the population is dynamically adjusted to further guide the population to evolve towards the most promising feasible area. In this process, in order to effectively utilize computing resources, the algorithm will conduct a detailed search of the explored feasible area, add new reference vectors in the area to enhance the algorithm's search density for the area; update the vectors that have been explored as infeasible areas, and delete the reference vectors corresponding to the infeasible areas;

[0030] S42. Allocate the existing architectural schemes in the DB according to the newly generated vector to obtain a new vector file, and use the architectural schemes in each vector file to construct an RBF local proxy model for each reference vector;

[0031] S43. Introduce a search strategy adaptive selection mechanism driven by reference vector type, classify each reference vector, and adaptively select the next generation population update mechanism based on the obtained type;

[0032] S44. Randomly obtain an initial population, and use the search strategy selected in S43 and the proxy model obtained in S42 to predict the newly generated individuals.

[0033] Preferably, the reference vector updating strategy in S41 is as follows:

[0034] For any reference vector V i And the corresponding vector file VS i , if within h iterations VS i If no feasible solution is included in V, then V is deleted from the reference vector set. i ; If within h iterations VS i If the improvement of the non-inferior solutions included is always less than the threshold ε, then the reference vector is adjusted according to the position of the high-quality solutions in the population, and the equation (3) is used to calculate the value of V i With the reference vector V j Generate a new reference vector V k :

[0035]

[0036] Among them, V j For the adjacent V i And contains a reference vector with a large number of feasible solutions; λ is an interpolation coefficient;

[0037] At the same time, VS is redistributed according to the distance between the individual and the reference vector i and VS j ; If VS i The number of feasible non-dominated solutions exceeds the preset threshold N th , calculate V i Two adjacent reference vectors (assuming V i-1 and V i+1 ) and select the one with the least number of feasible solutions, assuming it is V i+1 ; Then, a new reference vector is generated according to formula (3) and the non-dominated feasible solutions are redistributed.

[0038] Preferably, the content of S43 is as follows:

[0039] For any reference vector V i and its corresponding files VS i , the correlation between the constraint violation minimization direction and the target optimization direction is calculated as follows:

[0040] First, use formula (4) to calculate the file VS i The constraint violation value of each individual in , and sort all individuals from small to large according to the constraint violation value, and get the sorting rank CVN(x j ); at the same time, calculate VS i The average rank of all individuals in

[0041] Then, according to the objective function value, VS i The individuals in the non-dominated sorting are sorted, that is, the lower the sorting order, the better the target optimization performance of the solution, and the non-dominated sorting rank MN(x i ) and the average rank

[0042] Finally, the correlation coefficient ρ between the constraint violation direction and the target convergence direction is calculated according to formula (5): i .

[0043]

[0044] Where n is the vector file VS i The number of candidate solutions in g; i (x) represents the constraint value of the solution x;

[0045] Based on formula (5), we can get all N W Correlation coefficient ρ corresponding to the reference vector i ,i=1,2,…,N w ; For any reference vector, if its correlation coefficient ρ iIf μ > μ, the target optimization direction of the reference vector is considered to be consistent with the direction of minimizing the constraint violation value. The population has not yet entered the optimal feasible region, and individuals can continue to search along the optimization direction of the reference vector. Therefore, individuals are selected from the vector archive according to the non-dominated sorting as the initial population, and then the DE operator is executed for exploration.

[0046] If ρ i If the vector is less than -μ, the target optimization direction of the reference vector is considered to be opposite to the direction of minimizing the constraint violation value. In this case, the individuals contained in the vector may have crossed the feasible region of the constraint and entered the infeasible region. Therefore, the exploration position is readjusted through the local search method. The feasible solution with the optimal target value in the vector archive is used as the current optimal point, and the subsequent local search strategy is executed.

[0047] If -μ<ρ i <μ, the target optimization direction of the reference vector is considered to be weakly correlated with the direction of minimizing the constraint violation value, and the population on this vector has not yet reached the constraint feasible region. For such vectors, a flexible search strategy can be adopted, so the DE operator or subsequent local search operator can be executed randomly.

[0048] Preferably, the content of S44 is as follows:

[0049] According to S43, the vector classification and search adaptive mechanism are executed, and the corresponding candidate solution is obtained for each vector file and stored in HS. For positively correlated vectors, the DE operator is executed to obtain the candidate solution. All candidate solutions are sorted according to the predicted value of the surrogate model, and the N solutions with the best predicted value are selected. Their constraint violation values ​​are calculated, and the solution with the smallest constraint violation value is selected as the filling solution. For negatively correlated vectors, the local search operator TR-LS is executed to obtain the new individual x c Then, calculate the constraint violation value of the individual. If the individual is a feasible solution, use the surrogate model to calculate x c The predicted value of Better than the solution in the existing archive, then choose x c As filling samples, a true evaluation is performed. For weakly correlated vectors, the constraint violation value of each new candidate solution is calculated, and the solutions with the smallest constraint violation value are saved. The target value of the saved candidate solutions is evaluated using the surrogate model, and the solution with the best predicted target value is selected as the filling sample. The vector archive and its corresponding surrogate model are updated.

[0050] Therefore, the present invention adopts the above-mentioned two-stage agent-assisted evolution-based intelligent generation method for building energy-saving solutions, which has the following beneficial effects:

[0051] (1) A two-stage evolutionary strategy is adopted, which uses a uniformly distributed fixed reference vector to guide the population, quickly covers the feasible solution space and approaches the Pareto frontier (PF), crosses the infeasible area caused by constraints, ensures the diversity of the initial population, encrypts the search in the feasible area, deletes the infeasible area vector, accurately locates the true PF, and improves the quality and density of the solution. The phased strategy balances global exploration and local development, preventing traditional algorithms from falling into local optimality too early.

[0052] (2) The adaptive mechanism enables the algorithm to dynamically switch strategies according to the population position, adaptively switch DE / local search for positive / negative / weakly correlated vectors, solve the problem of population deviation from the feasible region, and effectively handle complex PF structures;

[0053] (3) In view of the engineering constraints in building design (such as window-to-wall ratio, orientation angle, heat transfer coefficient combination, etc.), individuals are screened by constraint violation values, and feasible solutions are retained first to improve the engineering practicality of the solutions; at the same time, the two conflicting goals of building annual energy consumption (BEC) and user discomfort hours (UDH) are optimized, and a variety of choices are provided through the Pareto frontier solution set to support decision makers in balancing energy consumption and comfort.

[0054] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 Schematic diagram of the process of the intelligent generation method of building energy-saving solutions based on two-stage agent-assisted evolution of the present invention;

[0056] Figure 2 This is a flowchart of the overall process of the intelligent generation method of building energy-saving solutions based on two-stage agent-assisted evolution of the present invention;

[0057] Figure 3 This is a diagram of the second phase of vector updating in an embodiment of the present invention, where (a) is a schematic diagram showing a vector set after the last iteration; (b) is a schematic diagram showing the vectors after being re-sorted and named according to their target values;

[0058] Figure 4 is the relationship between the population position and the feasible region in the embodiment of the present invention;

[0059] Figure 5 This is a schematic diagram of the appearance of a residential building in an embodiment of the present invention;

[0060] Figure 6 This is a schematic diagram of a framework for solving the building energy-saving problem according to an embodiment of the present invention;

[0061] Figure 7 Schematic diagram of PF curves obtained by using eight algorithms in the embodiment of the present invention for the building energy-saving design problem. DETAILED DESCRIPTION

[0062] The following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without inventive effort shall fall within the scope of protection of the present invention.

[0063] This example uses EnergyPlus simulation software to describe the mathematical model of the problem. Nineteen parameters are used as decision variables, including building orientation, window length and width in each thermal zone, window heat transfer coefficient and solar heat gain coefficient, wall insulation thickness, exterior wall solar absorptivity, lighting power density, and heating and cooling setpoints of the air conditioning system. The building energy-saving optimization model can be described as follows:

[0064] minF=(BEC(X),UDE(X))

[0065] stX=(x or ,x tolw ,x srar ,x ghtc ,x shgc ,x lwl ,x lww ,x bwl ,x bww ,x kwl , (6)x kww ,x wwl ,x www ,x lepd ,x bepd ,x kepd ,x wepd ,x hst ,x cst )

[0066] Among them, BEC is the annual energy consumption of the building, and UDH is the number of user discomfort hours.

[0067] Based on the above description, the expressions of the five cost constraints that need to be considered in this embodiment are:

[0068] Constraint 1: Window area to wall area ratio limit:

[0069] Project basis: According to the "Civil Building Energy Saving Design Standard" (JGJ26-2018), the upper limit of the window-to-wall ratio of residential buildings in each direction is: ≤25% for north, ≤35% for south, and ≤30% for east-west. Mathematical expression: For each thermal zone (living room, bedroom, etc.), the window parameter x lwl ,x lww (length, width) and corresponding wall area A wall , the constraints are:

[0070]

[0071] Where γ is the orientation-related threshold (e.g., γ = 0.35 for south).

[0072] Constraint 2: Building orientation angle restriction:

[0073] Project basis: According to the Beijing Residential Building Energy Saving Design Standard

[0074] (DB11 / 891-2020), in order to improve the efficiency of solar energy utilization in winter, the orientation of the main facade should be controlled within 15° south-east / west.

[0075] The constraints are specifically stated as follows:

[0076] g2(x)=|x or -180°|-15°≤0 (due south is 180°) (8)

[0077] Constraint 3: Combined constraint of window heat transfer coefficient and solar heat gain coefficient:

[0078] Project Basis: Referring to the "Energy-Saving Performance Standard for Building Doors and Windows" (GB / T 8484-2020), high heat transfer coefficient windows must be combined with low solar heat gain coefficient to balance energy consumption.

[0079] The constraint is specifically expressed as (x ghtc is the heat transfer coefficient, x shgc is the solar heat gain coefficient):

[0080]

[0081] Constraint 4: Window area coordination constraint between thermal zones:

[0082] Project basis: According to the "Architectural Lighting Design Standard" (GB 50033-2013), the differences in lighting requirements for different functional areas of residential buildings need to be coordinated and balanced through window area to avoid excessive lighting in the living room and insufficient lighting in the bedroom.

[0083] Mathematical expression:

[0084]

[0085] Effect: Prevents lighting imbalance from causing user behavior energy consumption compensation (such as turning on extra lights in the bedroom), and reduces invalid layouts requiring simulation verification by approximately 22%.

[0086] Constraint 5: Heat transfer-heat gain product constraint:

[0087] Engineering Basis: Based on the principle of thermal performance balance, high heat transfer coefficient windows must be combined with low solar heat gain coefficient to reduce fluctuations in total energy consumption throughout the year.

[0088] Mathematical expression:

[0089] g4=x ghtc ·x shgc -0.9≤0 (11)

[0090] Function: Eliminates the combination of heat transfer and heat gain deterioration (such as ordinary single-layer glass), and improves screening efficiency by 31%.

[0091] The ranges of variables for problem (6) are shown in Table 1. Some initial parameters for this problem are set as follows: the initial length and width of the windows are 1.8 m and 1.2 m, respectively. Following the recommendations in the EnergyPlus software manual, the return air volume coefficient is set to 0, the emissivity is set to 0.37, the visible light coefficient is set to 0.18, and the coefficient of heat transfer from light to the area air is set to 0.40.

[0092] Table 1 Decision variables and their value ranges of the residential building model

[0093]

[0094]

[0095] See also Figure 1 ,The intelligent generation method of building energy saving scheme based on two-stage agent assisted evolution ,includes the following steps:

[0096] S1. According to the decision variables and their value ranges of the residential building model, a rejection-based sampling method is used to generate several uniformly distributed feasible solutions, and the solutions are brought into the constructed EnergyPlus simulation model to obtain the true target value of the solution and put it into the sample set DB.

[0097] S2. Uniformly construct N in the target space W The reference vectors divide the target space; the initial sample in DB is divided into the reference vector closest to the sample in the target space, forming the vector file V of the vector i And use the samples in DB to build the initial RBF global surrogate model of M objective functions.

[0098] S21. Use uniformly distributed fixed reference vectors to guide population evolution to achieve crossing the infeasible area caused by the constraints, ensuring that the population covers the feasible solution space as early as possible and gradually approaches the Pareto frontier. Specifically, generate uniformly distributed reference vectors in the M-dimensional target space, and divide the target space into several sub-regions. Each reference vector corresponds to a subspace, and the population evolves based on these fixed reference vectors. Taking the two-objective optimization problem as an example, Figure 2 A schematic diagram of the individual assignment of reference vectors is shown.

[0099] S3. Enter the first stage, select the non-dominated solution with the smallest crowding density from each vector archive as the initial population, then execute the evolution operator based on the fixed vector, use the global agent model to evaluate the individual to obtain the target prediction value of the offspring, select the filling solution according to the first stage model management method, and update the local agent model according to the current sample, calculate the hypervolume increment ΔHV value of the current sample every h generations, when ΔHV is greater than the threshold α, the algorithm enters S4, otherwise it directly enters S5.

[0100] S31. Based on the individuals assigned to each vector, the APD value of each individual is calculated respectively, and a non-dominated solution is obtained according to the obtained APD value.

[0101] RVEA uses angle penalized distance (APD) to select the appropriate individual for each reference vector. For the current reference vector, the smaller the APD value of an individual, the higher the probability of its selection. The specific expression of APD is as follows:

[0102] APd j =(1+Pe(θ j ))·||f j || (1)

[0103] Among them, ||f j || is the distance from the jth individual to the origin in the target space; ||f j The smaller the || value, the better the convergence of the individual; θ j is the angle between the jth individual and the reference vector to which it is assigned; Pe(θ j ) is the penalty function.

[0104] S32. Using the non-dominated solution as the initial population and the agent model obtained in S22 as the objective function, perform the RVEA evolution operation to obtain multiple new individual solutions.

[0105] S33. Calculate the constraint violation values ​​of all new solutions, select NM individuals with the smallest constraint violation values, and use the proxy model to predict the objective function values ​​of these individuals. Select the individual that contributes most to HV improvement as the filling solution, truly evaluate the objective function value of the filling solution, and update the proxy model.

[0106] For a candidate solution set HS, assume that the corresponding reference point is Individual x i The objective function value of ∈HS is (f1(x i ),f2(x i ),...,f m (x i ), then the HV contribution of the individual relative to the reference point Z is:

[0107]

[0108] in, Represents the maximum point of the objective function of the set HS in the jth dimension; set the reference point Z to 1.1 times of .

[0109] During the algorithm execution, the hypervolume increment ΔHV of the current sample is calculated every h generations. When ΔHV is less than or greater than the threshold α, it indicates that the algorithm needs to explore additional constraint areas. At this time, the algorithm executes S4.

[0110] S4. Enter the second stage and dynamically update the reference vector according to the quality of the solution obtained in each generation; use the samples in each vector archive to build an RBF local proxy model for each reference vector; after building the local proxy model, execute the vector classification strategy, and execute the search strategy adaptive selection mechanism based on the classification results; after obtaining the offspring according to the selected search strategy, select the filling solution according to the second stage model management strategy, dynamically update the reference vector according to the quality of the solution obtained in each generation, and update the local proxy model based on the current sample.

[0111] S41. Dynamically update the reference vector based on the quality of the solution obtained in each generation. The goal of the second stage is to accurately locate the true PF of the problem and find more solutions located on the PF. Based on the feasible solution information obtained in the first stage, the reference vector of the population is dynamically adjusted to further guide the population to evolve towards the most promising feasible area. In this process, in order to effectively utilize computing resources, the algorithm will conduct a detailed search of the explored feasible area, add new reference vectors to the area to enhance the algorithm's search density for the area; update the vectors of the explored infeasible area and delete the reference vectors corresponding to the infeasible area.

[0112] The reference vector update strategy is as follows:

[0113] For any reference vector V i And the corresponding vector file VS i , if within h iterations VS i If no feasible solution is included in V, then V is deleted from the reference vector set. i ; If within h iterations VS i If the improvement of the non-inferior solutions included is always less than the threshold ε, then the reference vector is adjusted according to the position of the high-quality solutions in the population, and the equation (3) is used to calculate the value of V i With the reference vector V j Generate a new reference vector V k :

[0114]

[0115] Among them, V j For the adjacent V i And contains a reference vector with a large number of feasible solutions; λ is an interpolation coefficient;

[0116] At the same time, VS is redistributed according to the distance between the individual and the reference vector i and VS j ; If VS i The number of feasible non-dominated solutions exceeds the preset threshold N th , calculate V i Two adjacent reference vectors (assuming V i-1 and V i+1 ) and select the one with the least number of feasible solutions, assuming it is V i+1 ; Then, a new reference vector is generated according to formula (3) and the non-dominated feasible solutions are redistributed.

[0117] Figure 3 A specific example of reference vector update is shown, where Figure 3 (a) V i (i=1,2,3,4) is the vector set after the last iteration. It can be seen that the density of feasible non-dominated solutions in V2 exceeds the set threshold N th At the same time, the number of feasible solutions for V1 among its adjacent vectors V1 and V3 is small, so a new reference vector V is generated based on V1 new . V new After being placed in the reference vector set, the vectors are re-sorted and named according to their target values, such as Figure 3 (b) shown.

[0118] S42. According to the newly generated vector, the existing building plans in the DB are allocated to obtain a new vector file, and the building plans in each vector file are used to construct an RBF local proxy model for each reference vector.

[0119] S43. Introduce a reference vector type-driven search strategy adaptive selection mechanism to classify each reference vector and adaptively select the next generation population update mechanism based on the obtained type.

[0120] The relationship between the optimal Pareto frontier (UPF) of an unconstrained optimization problem and the optimal Pareto frontier (CPF) of an optimization problem under constraints is complex. Figure 4 As shown, the following situations may occur in the algorithm population: ① The population has crossed the optimal feasible region and entered the infeasible region; ② The population is in the optimal feasible region; ③ The population has not yet entered the optimal feasible region. For the above three situations, the evolutionary algorithm based on the fixed search strategy is difficult to accurately guide the search direction of the population. Therefore. This embodiment proposes a reference vector type driven search strategy adaptive selection mechanism to adaptively determine the evolutionary search strategy adopted by each individual in the population. The mechanism classifies the reference vectors by analyzing the correlation between the constraint violation value minimization direction and the target optimization direction, and applies different evolutionary search strategies to different types of reference vectors to effectively guide the population to converge to CPF.

[0121] For any reference vector V i and its corresponding files VS i , the correlation between the constraint violation minimization direction and the target optimization direction is calculated as follows:

[0122] First, use formula (4) to calculate the file VS i The constraint violation value of each individual in , and sort all individuals from small to large according to the constraint violation value, and get the sorting rank CVN(x j ); at the same time, calculate VS i The average rank CVN(x) of all individuals in ;

[0123] Then, according to the objective function value, VS i The individuals in the non-dominated sorting are sorted, that is, the lower the sorting order, the better the target optimization performance of the solution, and the non-dominated sorting rank MN(x i ) and the mean rank MN(x);

[0124] Finally, the correlation coefficient ρ between the constraint violation direction and the target convergence direction is calculated according to formula (5): i .

[0125]

[0126] Where n is the vector file VS i The number of candidate solutions in g; i (x) represents the constraint value of the solution x;

[0127] Based on formula (5), we can get all N W Correlation coefficient ρ corresponding to the reference vector i ,i=1,2,…,N w ; For any reference vector, if its correlation coefficient ρ i If μ > μ, the target optimization direction of the reference vector is considered to be consistent with the direction of minimizing the constraint violation value. The population has not yet entered the optimal feasible region, and individuals can continue to search along the optimization direction of the reference vector. Therefore, individuals are selected from the vector archive according to the non-dominated sorting as the initial population, and then the DE operator is executed for exploration.

[0128] If ρ i If the vector is less than -μ, the target optimization direction of the reference vector is considered to be opposite to the direction of minimizing the constraint violation value. In this case, the individuals contained in the vector may have crossed the feasible region of the constraint and entered the infeasible region. Therefore, the exploration position is readjusted through the local search method. The feasible solution with the optimal target value in the vector archive is used as the current optimal point, and the subsequent local search strategy is executed.

[0129] If -μ<ρ i <μ, the target optimization direction of the reference vector is considered to be weakly correlated with the direction of minimizing the constraint violation value, and the population on this vector has not yet reached the constraint feasible region. For such vectors, a flexible search strategy can be adopted, so the DE operator or subsequent local search operator can be executed randomly.

[0130] This embodiment adopts a proxy-assisted trust region-based local search (TR-LS) as the local search operator, and constructs an approximate proxy model within the local trust region to improve the algorithm's ability to handle discontinuous PF.

[0131] S44. Randomly obtain an initial population, and use the search strategy selected in S43 and the proxy model obtained in S42 to predict the newly generated individuals. According to S43, execute the vector classification and search adaptive mechanism, and obtain the corresponding candidate solution for each vector file and store it in HS. For positively correlated vectors, execute the DE operator to obtain the candidate solution, sort all candidate solutions according to the predicted value of the proxy model, select the N solutions with the best predicted value, calculate their constraint violation value, and select the solution with the smallest constraint violation value as the filling solution. For negatively correlated vectors, execute the local search operator TR-LS to obtain the new individual x c Then, calculate the constraint violation value of the individual. If the individual is a feasible solution, use the surrogate model to calculate x c The predicted value of Better than the solution in the existing archive, then choose x cAs filling samples, a true evaluation is performed. For weakly correlated vectors, the constraint violation value of each new candidate solution is calculated, and the solutions with the smallest constraint violation value are saved. The target value of the saved candidate solutions is evaluated using the surrogate model, and the solution with the best predicted target value is selected as the filling sample. The vector archive and its corresponding surrogate model are updated.

[0132] S5. Output several non-dominated feasible solutions.

[0133] Examples:

[0134] Taking the rural residential building design in a certain area as an example, Figure 5 The basic shape of a single-room residential building is given. According to different functions, the residential building is divided into four thermal zones: living room, bedroom, kitchen, and bathroom.

[0135] Figure 6 The application framework for this type of problem in this example is presented. Similar to compressor aerodynamic optimization design, this framework is divided into three modules: evolutionary optimization, surrogate model construction, and simulation. In the simulation module, when evaluating a new sample point or candidate solution, the SketchUP building model parameters are first modified based on the sample input. Then, combined with the weather problem, the target value is obtained by applying the model to the EnergyPlus simulation software.

[0136] For this example, some initial parameters for this problem are set as follows: the initial length and width of the windows are 1.8m and 1.2m, respectively. Following the recommendations in the EnergyPlus software manual, their return air volume coefficient is set to 0, the emissivity is set to 0.37, the visible light coefficient is set to 0.18, and the coefficient of heat transfer from light to the area air is set to 0.40.

[0137] In order to verify the effectiveness, seven agent-assisted multi-objective evolutionary algorithms, including ABSAEA, CSEA, EDN-ARMOEAD, HeEMOEAD, MCEAD, ParEGO, and REMO, are used as comparison algorithms.

[0138] Table 2. HV values ​​obtained by 8 methods for building energy-saving design

[0139]

[0140] Table 2 shows the HV values ​​obtained by the eight algorithms. It can be seen that the HV value obtained by this embodiment is much higher than that of the seven selected comparison algorithms, and the variance of the results obtained is also smaller than that of the seven comparison algorithms. Figure 7The PF curves obtained by eight algorithms for the building energy-saving design problem are given. It can be seen that although the optimal solution of VD-SAMOEA is dominated by the partial solution obtained by ABSAEA in some areas, the results obtained by VD-SAMOEA dominate the results of the other six comparison algorithms in most areas. In summary, the proposed VD-SAMOEA algorithm can obtain a better non-dominated solution set within a limited number of true evaluations. Compared with existing methods, VD-SAMOEA is more competitive in solving building energy-saving design problems with complex PFs. In summary, when solving building energy-saving design problems with low-cost constraints, the algorithm of this embodiment can obtain a better non-dominated solution set within a limited number of true evaluations.

[0141] Therefore, the present invention adopts the above-mentioned intelligent generation method of building energy-saving solutions based on two-stage agent-assisted evolution, and significantly improves the efficiency and accuracy of building energy-saving optimization by combining the agent model with the staged evolution strategy.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit the same. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solutions of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for intelligent generation of building energy-saving solutions based on two-stage agent-assisted evolution, characterized in that: The following steps are involved: S1. Based on the decision variables and their value ranges of the residential building model, a rejection-based sampling method is used to generate several uniformly distributed feasible solutions. The solutions are then introduced into the constructed EnergyPlus simulation model to obtain the true target values ​​of the solutions and store them in the sample set DB. S2. Uniformly construct N in the target space W The reference vectors divide the target space; the initial sample in DB is divided into the reference vector closest to the sample in the target space, forming the vector file V of the vector i And use the samples in DB to build the initial RBF global surrogate model of M objective functions; S3, enter the first stage, select the non-dominated solution with the minimum crowding density from each vector archive as the initial population, then execute the evolution operator based on the fixed vector, use the global agent model to evaluate the individual to obtain the target prediction value of the offspring, select the filling solution according to the first stage model management method, and update the local agent model according to the current sample, and calculate the hypervolume increment ΔHV value of the current sample every h generations. When ΔHV is greater than the threshold α, the algorithm enters S4, otherwise it directly enters S5; S4. Enter the second stage, dynamically update the reference vector based on the quality of the solution obtained in each generation; use the samples in each vector archive to build an RBF local proxy model for each reference vector; after building the local proxy model, execute the vector classification strategy, and execute the search strategy adaptive selection mechanism based on the classification results; after obtaining the offspring according to the selected search strategy, select the filling solution according to the second stage model management strategy, dynamically update the reference vector based on the quality of the solution obtained in each generation, and update the local proxy model based on the current sample; S5. Output several non-dominated feasible solutions.

2. The intelligent generation method of building energy-saving scheme based on two-stage agent-assisted evolution according to claim 1 is characterized in that: The steps for dividing the target space and building the proxy model with reference vectors in S2 are as follows: S21. Use a uniformly distributed fixed reference vector to guide the population evolution, cross the infeasible region caused by the constraints, make the population cover the feasible solution space and gradually approach the Pareto frontier; S22. Based on the sample information in the DB, the radial basis function is used to obtain the fitting functions of the building's annual energy consumption and the number of user discomfort hours, and they are defined as the proxy model.

3. The intelligent generation method of building energy-saving scheme based on two-stage agent-assisted evolution according to claim 2 is characterized in that: The steps for S3 are as follows: S31, based on the individuals assigned to each vector, respectively calculate the APD value of each individual, and obtain a non-dominated solution according to the obtained APD value; S32, using the non-dominated solution as the initial population and the agent model obtained in S22 as the objective function, execute the RVEA evolution operation to obtain several new individual solutions; S33. Calculate the constraint violation values ​​of all new solutions, select NM individuals with the smallest constraint violation values, and use the proxy model to predict the objective function value of the individuals. Select the individual that contributes the most to HV improvement as the filling solution, truly evaluate the objective function value of the filling solution, and update the proxy model.

4. The intelligent generation method of building energy-saving scheme based on two-stage agent-assisted evolution according to claim 3 is characterized in that: S31 content is as follows: RVEA uses the penalty angle distance (APD) to select the corresponding individual for each reference vector. For the current reference vector, the smaller the APD value of an individual, the higher the probability of its selection. The specific expression of APD is as follows: APd j =(1+Pe(θ j ))·||f j || (1) Among them, ||f j || is the distance from the jth individual to the origin in the target space; θ j is the angle between the jth individual and the reference vector to which it is assigned; Pe(θ j ) is the penalty function.

5. The intelligent generation method of building energy-saving scheme based on two-stage agent-assisted evolution according to claim 4 is characterized in that: S33 content is as follows: For a candidate solution set HS, assume that the corresponding reference point is Individual x i The objective function value of ∈HS is (f1(x i ),f2(x i ),...,f m (x i ), then the HV contribution of the individual relative to the reference point Z is: in, Represents the maximum point of the objective function of the set HS in the jth dimension; set the reference point Z to 1.1 times of .

6. The intelligent generation method of building energy-saving scheme based on two-stage agent-assisted evolution according to claim 5 is characterized in that: The content of S4 is as follows: S41. In the second stage, the reference vector is dynamically updated according to the quality of the solution obtained in each generation. Based on the feasible solution information obtained in the first stage, the reference vector of the population is dynamically adjusted to guide the population to evolve towards the most promising feasible area. S42. Allocate the existing architectural schemes in the DB according to the newly generated vector to obtain a new vector file, and use the architectural schemes in each vector file to construct an RBF local proxy model for each reference vector; S43. Introduce a search strategy adaptive selection mechanism driven by reference vector type, classify each reference vector, and adaptively select the next generation population update mechanism based on the obtained type; S44. Randomly obtain an initial population, predict the newly generated individuals based on the search strategy selected in S43 and the proxy model obtained in S42.

7. The intelligent generation method of building energy-saving scheme based on two-stage agent-assisted evolution according to claim 6 is characterized in that: The reference vector update strategy in S41 is as follows: For any reference vector V i And the corresponding vector file VS i , if within h iterations VS i If no feasible solution is included in V, then V is deleted from the reference vector set. i ; If within h iterations VS i If the improvement of the non-inferior solutions included is always less than the threshold ε, then the reference vector is adjusted according to the position of the high-quality solutions in the population, and the equation (3) is used to calculate the value of V i With the reference vector V j Generate a new reference vector V k : Among them, V j For the adjacent V i And contains a large number of reference vectors with feasible solutions; λ is an interpolation coefficient; At the same time, VS is redistributed according to the distance between the individual and the reference vector i and VS j ; If VS i The number of feasible non-dominated solutions exceeds the preset threshold N th , calculate V i Two adjacent reference vectors V i-1 and V i+1 The number of non-dominated feasible solutions, and select the one with the least number of feasible solutions, assuming it is V i+1 ; Then, a new reference vector is generated according to formula (3) and the non-dominated feasible solutions are redistributed.

8. The intelligent generation method of building energy-saving scheme based on two-stage agent-assisted evolution according to claim 6 is characterized in that: S43 content is as follows: For any reference vector V i and its corresponding files VS i , the correlation between the constraint violation minimization direction and the target optimization direction is calculated as follows: First, use formula (4) to calculate the file VS i The constraint violation value of each individual in , and sort all individuals from small to large according to the constraint violation value, and get the sorting rank CVN(x j ); at the same time, calculate VS i The average rank of all individuals in Then, according to the objective function value, VS i The individuals in the non-dominated sort are sorted, and the non-dominated sort rank MN(x i ) and the average rank Finally, the correlation coefficient ρ between the constraint violation direction and the target convergence direction is calculated according to formula (5): i : Where n is the vector file VS i The number of candidate solutions in g; i (x) represents the constraint value of the solution x; Based on formula (5), we can get all N W Correlation coefficient ρ corresponding to the reference vector i ,i=1,2,…,N w ; For any reference vector, if its correlation coefficient ρ i >μ, then it is considered that the target optimization direction of the reference vector is consistent with the direction of minimizing the constraint violation value, and the population has not yet entered the optimal feasible region. Individuals continue to search along the optimization direction of the reference vector; individuals are selected from the vector archive according to the non-dominated sorting as the initial population, and then the DE operator is executed for exploration; If ρ i <-μ, the target optimization direction of the reference vector is considered to be opposite to the direction of minimizing the constraint violation value; at this time, the individuals contained in the vector have crossed the feasible region of the constraint and reached the infeasible region, and the exploration position is readjusted through the local search method; the feasible solution with the optimal target value in the vector archive is used as the current optimal point, and the subsequent local search strategy is executed; If -μ<ρ i <μ, it is considered that the target optimization direction of the reference vector is weakly correlated with the direction of minimizing the constraint violation value, and the population on this vector has not yet reached the constraint feasible region; a flexible search strategy is adopted to randomly execute the DE operator or subsequent local search operator.

9. The intelligent generation method of building energy-saving scheme based on two-stage agent-assisted evolution according to claim 6 is characterized in that: S44 content is as follows: According to S43, the vector classification and search adaptive mechanism are executed, and the corresponding candidate solution is obtained for each vector file and stored in HS. For positively correlated vectors, the DE operator is executed to obtain the candidate solution, and all candidate solutions are sorted according to the predicted value of the surrogate model. The N solutions with the best predicted value are selected, and their constraint violation values ​​are calculated. The solution with the smallest constraint violation value is selected as the filling solution. For negatively correlated vectors, the local search operator TR-LS is executed to obtain the new individual x c Then, calculate the constraint violation value of the individual; if the individual is a feasible solution, use the agent model to calculate x c The predicted value of Better than the solution in the existing archive, then choose x c As filling samples, perform real evaluation; for weakly correlated vectors, calculate the constraint violation value of each new candidate solution and save the solution with the smaller constraint violation value; Use the proxy model to evaluate the target values ​​of the saved candidate solutions, select the solution with the best predicted target value as the filling sample, and update the vector archive and its corresponding proxy model.

Images